Non-invasive Techniques for Muscle Fatigue Monitoring: A Comprehensive Survey

Muscle fatigue represents a complex physiological and psychological phenomenon that impairs physical performance and increases the risks of injury. It is important to continuously monitor fatigue levels for early detection and management of fatigue. The detection and classification of muscle fatigue also provide important information in human-computer interactions (HMI), sports injuries and performance, ergonomics, and prosthetic control. With this purpose in mind, this review first provides an overview of the mechanisms of muscle fatigue and its biomarkers and further enumerates various non-invasive techniques commonly used for muscle fatigue monitoring and detection in the literature, including electromyogram (EMG), which records the muscle electrical activity during muscle contractions, mechanomyogram (MMG), which records vibration signals of muscle fibers, near-infrared spectroscopy (NIRS), which measures the amount of oxygen in the muscle, ultrasound (US), which records signals of muscle deformation during muscle contractions. This review also introduces the principle and mechanism, parameters used for fatigue detection, application in fatigue detection, and advantages and disadvantages of each technology in detail. To conclude, the limitations/challenges that need to be addressed for future research in this area are presented.


INTRODUCTION
In medicine science, sports science, and other related fields, "fatigue" is used to describe a decrease in physical performance associated with an increase in the actual and/or perceived difficulty of a task [ 1 ].Fatigue can be divided into perceived fatigue and performance fatigue (Figure 1 shows the different types of fatigue and physiological systems involved) [ 2 ].Perceived fatigue derives from the individual's self-reported subjective sensations based on homeostasis and the psychological state.In contrast, performance fatigue depends on the peripheral contractile function and the central activation.More specifically, in exercise physiology, "muscle fatigue" is broadly defined as the failure to maintain the required or expected force (or power output) that reduces the capacity of muscles to complete task overtime at a specific load [ 3 -5 ] and is accompanied by an increase in feelings of tiredness or even exhaustion [ 6 ].The definition of muscle fatigue combines the physiological function and exercise capacity during fatigue to analyze the occurrence and development of fatigue, which is helpful to select objective measurements to evaluate muscle fatigue.
Muscle fatigue encompasses a host of complex and comprehensive physiological and psychological phenomena.With the progress of the exercise, at some point, the sensations of fatigue and exhaustion will occur.The physiological effect of these sensations is to protect the exercising subject from the deleterious effects of exercise, such as cramps, muscle strains, and so on.As a result of these sensations, the subject would adjust his or her exercise strategy to avoid sports injuries [ 7 ].From the perspective of physiological mechanism, muscle fatigue is due to a series of changes in neuromusculoskeletal metabolism, energy, and structure, reduced oxygen, and nutrient supply by the blood circulation system, as well as a result of alterations in neuromuscular system efficiency [ 8 , 9 ].The neuromuscular mechanisms of muscle fatigue are related to changes in various components of the neuromuscular system, including fatigue in supraspinal and spinal (central fatigue) and fatigue in the neuromuscular junction and the muscle (peripheral fatigue) [ 10 ], and Figure 2 shows the neuromuscular mechanisms of performance fatigue.Central fatigue can be defined as a decrease in the voluntary activation of muscles (i.e., the recruitment of muscle fibers during a voluntary contraction effort), which is directly related to the reduction in the discharge rate of motoneurons and a reduced drive from the motor cortex [ 1 , 11 ].Peripheral fatigue refers to a reduction in muscular output resulting from the changes in the electrochemical mechanisms (e.g., neuromuscular transmission, Action Potential (AP) propagation, Excitation-contraction (EC) coupling) and mechanic means (e.g., altered viscoelasticity and stiffness of the contractile apparatus) downstream of the neuromuscular junction [ 12 ].It is worth noting that peripheral fatigue and central fatigue are not independent, and the interactions between the two are complex.In general, the relative contributions of the two types of fatigue to overall fatigue development are believed  to be task-related [ 13 ].For example, a study demonstrated that, during fatiguing contractions at sub-maximal levels, peripheral fatigue dominates central fatigue [ 14 ].But ultra-endurance activities have a significant component of central fatigue compared to shorter athletic activities [ 15 ].And in the same exercise, the ratio between the two changes at different stages of the physical exercise [ 1 ].
Currently, in the literature, muscle fatigue is considered a two-stage process: non-fatigue and fatigue [ 16 , 17 ].However, because the predictive nature of the measurement is inherently lost by the dichotomy, this classification does not contribute to predicting the onset of fatigue.To remove this limitation, researchers proposed adding an intermediate state, namely, the transitionto-fatigue state [ 18 ].In studies on muscle fatigue, the joint angle and oscillations are recorded as indicators of biomechanical fatigue.For example, Al-Mulla et al. [ 19 ] applied the goniometer to segment the signal during muscle contraction into non-fatigue state (above a certain angle), transition-to-fatigue state (between a certain angle boundary), and fatigue state (below a certain angle).When in the non-fatigue state, the Central Nervous System (CNS) can recruit Motor Units (MUs) unimpaired, and all the muscle fibers can exert their maximum strength.When in the transition-to-fatigue state, new recruitment of MUs emerges, where there is a sudden increase in MUs firing rate until the onset of the fatigue state [ 20 ].As such, identifying the transition-tofatigue state will contribute to monitoring and even predicting the onset of muscle fatigue.
Biomarkers produced during the physiological process of muscle fatigue development can be used for diagnosis, monitoring, or risk reduction.Based on the characteristics of the biomarkers, the biomarkers of muscle fatigue can be classified into "wet biomarkers" and "dry biomarkers" [ 21 , 22 ].Most wet biomarkers are biological substances measured in body fluids such as blood, saliva, or urine.They usually reflect the modulation of an endogenous substance in these fluids.For example, lactate [ 23 ], ammonia [ 24 ], isoprostanes [ 25 ], glutathione [ 26 ], Interleukin-6 (IL-6) [ 27 ], Tumor Necrosis Factor-α (TNF-α ) [ 28 ], and so on.Among the investigated wet biomarkers, serum lactate and IL-6 are the most accurate.The main limitation of wet biomarkers is that they cannot monitor the development of fatigue in real-time [ 29 ].In addition, they are usually obtained invasively, and the acquisition methods are relatively more complicated than non-invasive methods [ 30 ].
Dry biomarkers are usually non-invasive and easy to acquire.Currently, dry biomarkers of muscle fatigue are mainly based on the following measuring principles: subjective measures, performance-related methods, and physiological signal-based methods [ 22 ].Subjective measures consist of assessing self-reported muscle fatigue through questionnaires and scales [ 31 , 32 ].A major disadvantage of these methods is that they rely on subjective information.Performance-related methods rely on subjects' exercise performance on specific tasks (e.g., peak torque test, reaction time test).Although easy to standardize and widely used in practice, performance-related methods are not suitable for wearable devices to monitor the development of muscle fatigue in real-time while participating in physical exercise so preventive measures can be taken before a muscle injury happens.Physiological signal-based methods detect the onset of fatigue based on changes in subjects' physiological responses.Mechanomyogram (MMG), near-infrared spectroscopy (NIRS), ultrasound (US) , and electromyogram (EMG) are among the most investigated techniques for muscle fatigue detection.The MMG can record the vibration of muscle during muscle fiber movement and then observe the mechanical signals on the surface of the contracted muscle [ 33 ].The NIRS detects fatigue by measuring the amount of oxygen in the muscle [ 34 ].The US describes the structural and morphological changes of skeletal muscle [ 35 ].In various evaluation methods of human physiological information, EMG has proven to be a mature and highly reliable non-invasive measurement method compared with other techniques; it can non-invasively measure the activation state of muscles [ 36 ].Taking physiological signals as measurements of muscle fatigue enables objective, real-time fatigue monitoring at the individual level [ 20 , 37 , 38 ].Because muscle fatigue is an ongoing and dynamic process, rather than an instantaneous and static event, it is important to monitor the temporal changes of the physiological variables as the fatigue develops [ 39 ].This is why a continuous monitoring of the state of fatigue is particularly important in theoretical analysis and attractive in real-world applications.

CONTRIBUTION OF EMG TO MUSCLE FATIGUE MONITORING 3.1 Principles and Mechanisms of EMG
Muscle contraction is the physiological process by which muscle fibers shorten due to the sliding of the myofilaments [ 40 ].The smallest neuromuscular functional unit responding to the descending neural drive is called a Motor Unit (MU) (Figure 4 shows the diagram of MUs).It consists of an alpha motoneuron and all the muscle fibers it innervates [ 41 ].Activated by its efferent neural drive, a motoneuron generates a sequence of AP, each of which propagates down to the muscle fibers along its axon.When a motoneuron AP arrives at the neuromuscular junction (a.k.a.motor endplate), neural transmitters are released from the axon terminal to the muscle fibers, resulting in an instantaneous reversal of the electrical potential across the plasma membrane of the muscle fibers.This change of transmembrane potential then generates a muscle fiber AP that propagates along the axial direction of the fiber and eventually terminates at the tendon of the muscle.This generation process of the muscle fiber AP is called the innervation of muscle fibers.
The propagating of a muscle fiber AP generates a changing electromagnetic field that can be detected by an electrode placed in the vicinity of the membrane, which is called the Single Fiber AP (SFAP) .Because all fibers of a motor unit are innervated simultaneously (at the temporal scale of the SFAPs), the temporal and spatial summation of the SFAPs from all the fibers of a motor unit would form a Motor Unit Action Potential (MUAP) [ 42 -44 ].The characteristics of the MUAP depend on a large number of parameters, including but not limited to the diameter of the muscle fibers, the Conduction Velocity (CV) , and its position relative to the electrode  that acquires the signal [ 45 ].In most cases, one motoneuron AP would innervate one MUAP.Therefore, a sequence of motoneuron APs would generate a train of MUAPs, and the detected electric activity is called a MUAP train (MUAPt) .When there are multiple active MUs within the detection volume of the electrode, multiple MUAPts can be seen.The composite signal is usually called electromyography (EMG) [ 46 ], and Figure 5 shows the generation of EMG signals.

Acquisition of EMG Signals
According to the position of the detection, EMG electrodes can be divided into intramuscular electrodes and surface electrodes (Table 1 displays the different categories of EMG electrodes and their advantages and disadvantages).Intramuscular EMG (iEMG) signals are generally acquired by fine-wire electrodes or needle electrodes and can only detect a relatively small number of MUAPs.Usually, the purpose of recording iEMG is to obtain a MUAP with high temporal precision, with as limited interfering signals from other MUAPs as possible.However, the invasive nature of iEMG is a major limitation.As such, applications of iEMG are largely limited to neurology laboratory settings [ 47 , 48 ].
Surface EMG (sEMG) measures the changing electrical field generated by a muscle through electrodes placed on the skin surface just above the target muscle.sEMG usually consists of spatial and temporal summation of anywhere between several to hundreds of simultaneously active MUs of the target muscle, but often contains MUs from other muscles in close vicinity [ 49 -52 ].Surface electrodes can be divided into wet electrodes and dry electrodes.Wet electrodes contain a gelled electrolytic substance as a medium between skin and electrodes.Silver-silver chloride is the most common composite for wet electrodes.Wet electrodes are usually disposable.Dry electrodes allow direct contact with the skin surface, such as bar electrodes and array electrodes.Dry electrodes are wearable and can be reused.There are three main types of sEMG electrode montage: monopolar, bipolar, and multipolar.The monopolar montage directly measures the potential difference between the detection site (usually directly above the muscle) and an electrically neutral site, i.e., the reference.Bipolar montage measures the differentiation between two sites over the target muscle [ 53 ].Compared to monopolar, bipolar montage acts as a spatial high-pass filter, which suppresses common-mode noise.But at the same time, the filtering process inevitably removes potentially useful information and the montage lacks setup flexibility.Multipolar electrodes, such as high-density surface EMG (HD-sEMG) , form a grid of electrodes that could acquire the electric potential arising from a much wider area of the target muscle over the skin compared with the first two montages [ 52 ].The biggest advantage of HD-sEMG is the ability to acquire rich spatial information.HD-sEMG overcomes the limitation that ordinary sEMG can simply provide the summed activity of all MUs and can differentiate the output of multiple independent MUs.From this, more detailed information can be extracted from HD-sEMG, such as identifying the innervation zone of the muscle, measuring the CV of the AP propagation, and even extracting individual MUAPts through HD-sEMG decomposition algorithms [ 52 , 54 ].Regardless of the electrodes montage, the  properties of EMG signals are further influenced by many other factors, including the number of active MUs within the detection electrodes, the composition of MUs fiber type, tissues thickness between the electrode and the target muscle, electrodes' placement, external noise (motion artifacts, power line interference, etc.), the properties of amplifiers, and more [ 55 ].
Acquiring high-quality data is the precondition to ensure the accuracy of muscle fatigue prediction and recognition.First, selecting a reasonable electrode placement is crucial for obtaining a high-quality EMG signal.The electrode placement is usually referred by the European SE-NIAM recommendations [ 56 ].Second, careful skin preparation before the electrode placement, such as skin cleansing, skin shaving, and skin abraded, is also essential to improving EMG signal quality [ 57 ].

Signals Pre-processing for EMG
The raw EMG signals acquired by the front-end acquisition equipment are generally noisy, and they are usually pre-processed before subsequent analyses, including when used for fatigue analysis.Generally, the processing steps consist of digital filtering, segmentation, rectification (e.g., full-wave rectification and half-wave rectification), smoothing (e.g., Average Rectified Value (ARV) and Root Mean Square (RMS) ), amplitude normalization [ 55 ].The exact steps and the order in which they are taken are application-dependent.
It is known that the energetic distribution of the sEMG signal is approximately within the 0-500 Hz range in the frequency domain, with the dominant components in the 50-150 Hz range [ 58 ].Filtering is one of the most effective methods to reduce the noise of a signal, improving its fidelity.Band-pass filters are used to process EMG signals to attenuate the low-and high-frequency noise and artifacts [ 46 ].It is generally accepted that the most commonly used EMG signal filter is essentially a finite impulse response bandpass filter with a cutoff point of 20 Hz and 500 Hz [ 59 -61 ].However, the other filtering ranges can be selected according to the difference between the measuring position and the measuring item.For example, a 20-350 Hz filter might accurately display an EMG signal and is sensitive to muscle fatigue [ 62 , 63 ].Temporal windowing or segmentation is usually used to "slice" continuous sEMG data into short segments for subsequent analyses.Generally, the properties of the signal decide the window length when analyzing the EMG data.Huang et al. [ 64 ] suggested that the analysis window length should not exceed 200 ms, which is an ideal upper limit to control the signal variation.Smith et al. [ 65 ] suggested that the windows of 150-250 ms as the optimal tradeoff between the classification accuracy and the delay.Ashraf et al. [ 66 ] suggested that the optimum window size ranges are from 200-300 ms for the disjoint and 225-300 ms for the overlap windowing technique, respectively.

Time-domain Analysis.
The time-domain features regard the EMG signal as a function of time, and the two most commonly used functions for muscle fatigue are the RMS of EMG envelope and integrated EMG (IEMG) , which is to be distinguished from iEMG.The RMS of EMG is defined as the RMS value of the windowed sEMG [ 67 ].And the IEMG is defined as the area under the curve of the rectified raw sEMG signals, i.e., the mathematical integral of the absolute value of the raw sEMG signals [ 68 ].The changes in the temporal and spectral parameters of the sEMG during sustained muscle contractions are often referred to as "myoelectric manifestations of fatigue" [ 61 ].Previous studies have confirmed that the IEMG gradually increases with the progress of muscle fatigue [ 50 , 69 ].Therefore, the IEMG can be used to detect the progress of fatigue by comparing its current value with its initial value.However, the (temporal) amplitude of the EMG signal is rarely used as a parameter of muscle fatigue alone, and it is usually used in conjunction with frequency-domain parameters [ 70 ].

Frequency-domain Analysis.
The frequency-domain analysis of EMG signals transforms the acquired temporal EMG signal into a frequency-domain representation by methods such as the Fast Fourier Transform (FFT) and then extracts power spectrum characteristics.In the frequency domain, the Mean Power Frequency (MPF) and Median Frequency (MDF) are the most important parameters for fatigue monitoring or detection.The MPF is the average frequency at which the sum of the product of the EMG power spectrum and frequency is divided by the total, referring to Equation ( 1 ): (1) Here, f n is the frequency variable, N is the number of samples, and PSD is the Power Spectral Density of the sEMG signal, FT is the Fourier Transform.However, the MDF is the frequency at Here, N is the number of samples and PSD is the Power Spectral Density of the sEMG signal.During fatiguing isometric muscle contractions, the decreases in MPF and MDF are typically accompanied by an increase in the EMG's temporal amplitude.MUs synchronization is the tendency of separate MUs to discharge at approximately the same time (within 1-5 ms of each other) more frequently than chance level, which quantifies the amount of correlation between APs discharged by MUs.The MPF and the MDF of the EMG power spectral density are also sensitive to MUs synchronization.MUs synchronization usually increases during fatiguing contractions, which indicates more MUs are activated [ 50 , 71 , 72 ].

Other Methods.
Even in the case of constant-level isometric muscle contractions, the frequency spectrum of sEMG signals exhibits a non-stationary nature due to changes related to muscle fatigue processes.Short-time Fourier Transform (STFT) provides the time-localized frequency information for situations in which frequency components of a signal vary over time [ 73 ].By selecting appropriate window lengths and overlaps, it can be used to assess biological signals and support decisions between fatigue or non-fatigue states [ 74 ].Studies have confirmed that a combination of 250 ms with 50% of overlap reduced the dispersion both for MPF and MDF [ 73 ].Continuous Wavelet Transform (CWT) allows the free choice of wavelet scales corresponding to frequency and time values, thus providing fine control over frequency resolution.It ensures good time resolution at high frequencies and good frequency resolution at low frequencies in muscle fatigue conditions [ 75 ].After the sEMG signal is preprocessed, the features for evaluating muscle fatigue need to be extracted.The quality of the features has a crucial influence on the classification and prediction of muscle fatigue.Regarding nonlinear analysis, extracting entropy from sEMG is an effective method for analyzing muscle fatigue [ 76 , 77 ].The entropy feature represents the uncertainty and confusion of the signal.Fatigue reduced the complexity of muscle activity during the contraction [ 78 ].
The above analysis method is also applicable to HD-sEMG, but the difference between HD-sEMG and ordinary sEMG is that there are spatial features.The HD-sEMG signals usually contain a lot of artifacts and redundant information, so spatial filtering algorithms such as Principal Component Analysis (PCA), Independent Component Analysis (ICA) , and Common Spatial Pattern (CSP) are needed to separate the useless information and reduce the processing amount of data [ 79 ]. sEMG decomposition methods based on Blind Source Separation (BSS) are used to accurately identify a large number of MUs, such as the Convolution Kernel Compensation (CKC) algorithm [ 80 ], fast ICA (fICA) [ 81 ], and K-means clustering [ 82 ], have been developed and validated.However, real-time HD-sEMG decomposition is still in its infancy and not yet ready for applications such as real-time fatigue monitoring.Meanwhile, the topographic map of the signal amplitude intensity obtained by the high-density electrode matrix analysis method can determine the electrode location in the region with relatively strong muscle activity [ 83 ].

Application of EMG in Muscle Fatigue Monitoring
3.5.1 Isometric Contractions.In most experimental designs on fatigue, subjects were required to perform subject-specific maximum voluntary contraction (MVC) .The EMG signals are affected by the form of muscle contraction, such as static contraction, dynamic contraction, or electrically evoked contraction.To date, the majority of EMG research in the context of muscle fatigue has focused on isometric contraction to establish typical EMG readings in controlled settings.This may be due to the non-stationarity of the EMG signal during dynamic contraction [ 84 , 85 ], which makes spectral analysis (e.g., FFT) difficult to apply.However, the EMG signals during isometric contraction can be assumed to be stationary between short-time intervals of 0.5 s-2 s [ 47 ].
In terms of changes in time-domain features, Gawda et al. [ 86 ] found that the mean EMG amplitude of vastus medialis muscles increased significantly during isometric activity in a 60second squatting position.According to Komi et al. [ 87 ], the relationship between IEMG and force shifted to the left during isometric contraction, and the IEMG increase in the submaximal isometric contraction (60% MVC), while the maximum strength and the IEMG decreased during the MVC.Studies suggested that the IEMG is closely related to the muscle fiber composition [ 88 ] and the discharge frequency of MUs [ 89 ] during voluntary contraction.Furthermore, the mean EMG amplitude also has been reported to increase during submaximal isometric contractions (50%-70%MVC) and decrease during maximal ones [ 90 , 91 ].In 1912, Piper et al. [ 92 ] observed a progressive "slowing" of the EMG during isometric voluntary sustained contractions.Given the random nature of voluntary EMG, this "slowing" cannot easily be quantified in the time domain.But it is easier to describe it in the frequency domain using spectral characteristics.
In terms of changes in frequency-domain features, Wellems et al. [ 93 ] found that during the 20% MVC isometric contraction of the quadriceps femoris to failure, the MDF and MPF first increased, then decreased, and finally fell below the initial state.Mehra et al. [ 94 ] found that the shape of the EMG spectral distribution of deltoid muscle did not change when performing isometric lateral raise at 60% MVC till the endurance limit but shifted towards lower frequency with an increase of magnitude at characteristic mode frequency.Furui et al. [ 95 ] proposed a continuous estimation method for variance distribution parameters using a sliding window, enabling the evaluation of the time-varying stochastic properties of EMG signals.The results showed that with the endurance time, the EMG MDF decreased.Using the HD-sEMG, Kimura et al. [ 96 ] detected a significant decrease in muscle fibers CV and MDF during isometric contraction of the trapezius at 30% MVC.It was shown that the rates of change of spectral variables and global CV during sustained isometric constant force contractions are indicative of muscle fatigue and may be related to MUs type.And it was also shown, both theoretically [ 97 ] and experimentally [ 98 ], that during fatiguing isometric constant force contractions, CV and MPF or MDF of the EMG signal are highly correlated, with MPF and MDF reflecting mainly the changes in the CV of the active MUs.The MUs synchronization is generally assumed to represent the activities of spinal nerves, and the HD-sEMG can identify when there is an increase in neural drive to muscle.Liu et al. [ 99 ] compared the MUs synchronization in delta (1-4 Hz), alpha (8-12 Hz), beta (15)(16)(17)(18)(19)(20)(21)(22)(23)(24)(25)(26)(27)(28)(29)(30), and gamma (30-60 Hz) frequency bands during the fatigue condition induced by isometric contraction (sustain the 2 Kg dumbbell until muscle exhausted), and the results showed that MU synchronization increased significantly in all frequency bands.
In some isometric fatigue protocols, the researchers also proposed models and algorithms to predict, identify, and classify muscle fatigue.Qassim et al. [ 62 ] proposed an algorithm that applies the difference between the Instantaneous Mean Amplitude (IMA) values of the low-frequency sub-signal and high-frequency sub-signal to represent a novel muscle fatigue index.The results showed that the proposed algorithm had an accuracy of 94.66% in distinguishing between conditions of muscle fatigue and non-fatigue.Convolutional Neural Networks (CNN) and Support Vector Machines (SVM) are also prevalent.Moniri et al. [ 100 ] compared the performance of the shallow model and CNN in forecasting the sEMG feature of trunk muscle fatigue.The results show that although sEMG features are non-stationary, adaptive learning and prediction, especially using CNN, can provide accurate and precise muscle fatigue predictions for a range of physical activities.Wang et al. [ 101 ] proposed a neural architecture search framework based on reinforcement learning to autogenerate neural networks and combined CNN with an SVM, K-Nearest Neighbor (KNN) , and Random Forest (RF) to improve the performance of muscle fatigue detection.The results showed that the accuracy of CNN combined with SVM is 96.5%.

Dynamic Contractions.
Although the research on EMG is dominated by isometric contraction, EMG has also been applied to dynamic contraction.Dynamic contraction generates force by changing the length of the muscle (it can be concentric, eccentric, or isokinetic), and represents a realistic view of many daily activities as well as sports.Masuda et al. [ 102 ] studied changes in sEMG patterns during static and dynamic fatiguing contractions by the CV of MUAPs and MDF in the vastus lateralis.The CV of MUAPs appears to be influenced by the muscle metabolic status, as it significantly decreased in isometric contractions, while it remained constant during dynamic contractions.More importantly, studies have shown that the low-frequency band is a reliable indicator of muscle fatigue during dynamic contraction [ 103 ].
For the dynamic muscle contraction, Thongpanja et al. [ 104 ] found that up to 80% of the signals measured during a subject's elbow flexion and extension task (0-150 °, duration of 5 s) could be regarded as a static contraction (i.e., stationary signals).Moreover, the percentage of stationary signals decreased as the window size increased.The authors also suggested that the size of the analysis window of 250 ms is suited for both static contractions and dynamic contractions.However, Farfán et al. [ 105 ] during dynamic contraction (0-90 °, about 4.5 °/sec) and static contraction (duration of 6 s, for 6 trials) induced by arm abduction and adduction movements, the optimal window length in static contractions is 200 ms, but the optimal window length in dynamic contractions is 300 ms.As is mentioned above, when muscles contract dynamically, the EMG signal generated by the muscle is no longer considered to be a stationary process [ 84 , 85 ], because the neural drive descending on the muscle is not constant, but after windowing processing, the descending drive could be considered as approximately "constant."Second, the muscle movement itself adds motion artifacts to the sEMG.Finally, mechanical processes may influence signal attenuation, resulting in decreased sEMG [ 106 ].For example, to study the EMG signal under the condition of dynamic contraction, a spectrum estimation technique suitable for the analysis of non-stationary processes must be adopted.Research has shown that Cohen's class time-frequency distributions and wavelet analysis may be more appropriate to study the non-stationary signals during dynamic contractions [ 107 ], which is consistent with the findings of Smale et al. [ 108 ].
In recent years, some innovative signal analysis techniques have been used to study muscle fatigue.Makaram et al. [ 109 ] analyzed sEMG signals in non-fatigue and fatigue conditions using the degree distribution of visibility graphs, which can convert a time series into a graph to capture the presence of nonlinear correlations of a time series and found that the non-fatigue condition has minimal connectivity.During fatigue conditions, the increased connectivity may be attributed to MUs synchronization.The above authors also extracted the entropy features (such as symbolic entropy, network entropy, etc.) of EMG signals in biceps brachii during the dynamic contraction and proved that the signal complexity decreased during fatigue.This might be attributed to the reduction in the CV of muscle fibers and MUs synchronization during fatigue [ 110 ].During dynamic contraction, the sEMG signals are not only considered to be nonstationary, but also cyclostationary.Cyclostationary refers to if the cyclic properties of the signal are reflected in the statistical characteristics, but cannot be directly observed in the signal [ 111 ].Bharathi et al. [ 112 ] developed an automated muscle fatigue detection system using cyclostationary-based geometric features (such as perimeter, area, circularity, eccentricity, inertia, etc.) of sEMG signals.The results show that the cyclostationarity increases in the fatigue state, which may be attributed to enhanced neural drive or increased firing rate of motor neurons.Boyer et al. [ 113 ] applied the MDF of the sEMG power spectrum obtained with the CWT as an indicator of the muscle fatigue level and compared the results of CWT to those of the traditional STFT.The results showed that the CWT performs better than the STFT in both static and dynamic contraction conditions.Entropy measures quantify the complexity of a time series.Murillo-Escobar et al. [ 76 ] found that the Permutation Entropy (PE) value decreased around by 50% from non-fatigue to fatigue contractions, indicating a clear reduction in sEMG complexity in the presence of fatigue.The authors also found that PE could distinguish non-fatigue, transition-to-fatigue, and fatigue more effectively than MPF of sEMG signal (mean ROC of PE: 0.77, mean ROC of MPF: 0.54).Duan et al. [ 77 ] proposed a quantifiable fatigue method called Wavelet Packet Energy Entropy (WPEE) to quantitatively express the dynamic muscle physical fatigue, which is also better compared with the traditional method MPF.Karthick et al. [ 114 ] proposed the analysis based on high-resolution time-frequency methods, namely, Stockwell Transforms (S-transform), B-distribution (BD) , and Extended Modified B-distribution (EMBD) , to distinguish the dynamic muscle non-fatigue and fatigue conditions.The results demonstrated that the combination of EMBD-polynomial kernel-based SVM is found to be the most accurate (91% accuracy) in classifying the conditions with the features selected using a Genetic Algorithm (GA) .Wang et al. [ 115 ] proposed the Long Short-Term Memory (LSTM) network for the recognition of muscle fatigue and confirmed that the classification performance of LSTM is better than that of CNN and SVM.

Electrically Elicited Contractions.
Differences in muscle fibers' recruitment patterns and functions elicited by voluntary contraction are different from those by electrically elicited contractions, following the different recruitment principles and patterns [ 116 ].When conventional electrical stimulation is applied over a muscle belly, it preferentially recruits superficial MUs, with progressively deeper MUs recruited as stimulus amplitude increases [ 90 , 117 ].Consequently, recruitment order during conventional electrical stimulation is random to MU type, contrary to voluntary contractions, and does not follow the "size principle" [ 118 , 119 ].As a result, conventional electrical stimulation recruits fewer fatigue-resistant MUs, and more fast-fatigable MUs compared with voluntary contractions of similar amplitude, resulting in greater fatigability.
Merletti et al. [ 120 ] studied the time course of muscle fibers' CV and sEMG signal parameters during sustained isometric voluntary or electrically elicited contractions (20, 25, 30, 35, and 40 Hz for the 20 s) and found that sEMG signal variables obtained from electrically elicited contractions show fluctuations smaller than those observed in voluntary contractions.Hamada et al. [ 121 ] found that the RMS of sEMG increased and the mean MUs spike amplitude was significantly reduced, while those MUs with small spike amplitude increased their firing rate during the 40% MVC test contraction after the electrical stimulation (20 Hz).Watanabe et al. measured the CV during electrical stimulation (CV-EC) and MVC (CV-VC) before and immediately, 30 min, 60 min, 120 min, and 24 h after exhaustive leg-pedaling exercise.The results showed that the CV-EC significantly increased, but the CV-VC significantly decreased immediately after the fatiguing exercise.Stratton et al. [ 122 ] compared the EMG characteristics of two muscle groups with different proportions of fast and slow muscle fibers during voluntary contraction (20%, 50%, and 75% MVC) and voluntary activation plus electrical stimulation (10, 35, and 50 Hz).Higher stimulation frequency (35 and 50 Hz) was found to induce higher electrical output at 25% of the MVC, suggesting more recruitment at higher frequencies.Electrical stimulation may stimulate muscle strength by activating more fatigued fast-acting fibers.These findings indicated differential MU activation patterns in terms of MU recruitment and rate coding characteristics.It also strongly suggests the possibility of "an inverse size principle" of MU recruitment during electrical stimulation [ 123 ].

Strengths and Limitations
From the discussion above, the application of EMG has multiple advantages in fatigue detection.First, the EMG signals contain information that is pertinent to the physiological state during movements; in particular, it can identify the changes related to MUs recruitment and discharge rate, which means that EMG can recognize the transition-to-fatigue state.Therefore, it is suitable for early fatigue prediction and detection.Second, it is a non-invasive technique that can provide real-time, continuous, and quantitative data and it is suitable for developing the wearable devices to adapt to various scenarios, such as sports exercise and pattern recognition [ 49 , 50 ].
However, there are also several limitations associated with EMG in fatigue analysis.First, EMG signals are susceptible to factors that are unrelated to the exercise set by the experimental protocol, such as contraction speed.Therefore, changes in muscle activation are not necessarily reflected in the EMG signal features currently used.Although it is commonly believed that activation of the CNS produces higher EMG amplitudes, this assumption does not hold in many conditions [ 124 ].Second, in the study of muscle fatigue, the relationship between CV and spectral frequency is studied to determine the recruitment of MUs.However, the EMG analysis of muscle fatigue during dynamic contraction is complicated, because several factors, such as changes in the number of active MUs, changes in force/power, changes in muscle fibers length, and changes in muscle fibers CV caused by muscle fatigue, can increase the non-stationarity of the EMG signals, which increases the difficulty of signal analysis [ 84 , 85 ].Studies also suggested that the sEMG provides a more appropriate measure of the change in muscle activation during a fatiguing isometric contraction [ 86 , 87 ].This suggests that EMG may be better suited to monitoring muscle fatigue during isometric contractions.

Mechanisms of MMG
In 1993, Orizio discovered that the change in intramuscular pressure caused by the muscle fiber contraction could produce a subtle vibration on the surface of the muscle, and the vibration signal could be amplified, detected, and recorded quantitatively [ 125 ].He called such signal mechanomyography (MMG) .Other terms have also been used to describe MMG, such as phonomyography (PMG) [ 126 ], acoustic myography (AMG) [ 127 ], vibromyography (VMG) [ 128 ], and so on, but finally evolved to use MMG to summarize this type of muscular signal.
The MMG signal generation is the mechanical activity of the MUs [ 129 ].When a muscle contracts, the length of muscle fibers will change, and the changes in the volume of muscle fibers further generate pressure waves.Due to the asynchronous activity of the muscle fibers, transverse movement and oscillation are generated between muscle fibers, eventually generating MMG signals [ 125 ].The MMG signal can be recorded by specific mechanical sensors placed at the skin surface to analyze muscle surface movements or changes by detecting muscle fibers oscillation, muscle fibers displacement, contraction velocity, and sound.Therefore, the MMG signal is a mechanical phenomenon of transverse oscillation during muscle contraction [ 130 ].The most widely used types of sensors for MMG are piezoelectric contact sensors, condenser microphones, accelerometers, and laser distance sensors (Figure 6 shows the most widely used types of sensors for MMG) [ 125 , 131 ].It has been suggested that accelerometers may be more appropriate than condenser microphones and piezoelectric contact sensors when recording MMG signals from small muscles, such as the first dorsal interosseous.For larger muscles, such as quadriceps femoris, however, condenser microphones and piezoelectric contact sensors can be used [ 125 ].Generally, the amplitude of the MMG signal reflects the change in mechanical properties of recruited MUs, and the frequency of the MMG signal is qualitatively related to the discharge rate of the activated MUs [ 132 ].In addition, since the muscle thickness, muscle fibers length, and angle will be influenced by the muscle contraction, this means that the MMG potentially carries important information from the mechanical state of muscle fibers and the neuromuscular function.Therefore, MMG is believed to be more sensitive to the physiological condition of muscle and potentially more robust than EMG in some cases [ 133 , 134 ].

MMG Signal Parameters of Muscle Fatigue
Similar to those of EMG, MMG features include both time-domain and frequency-domain.The most commonly used time-domain feature in muscle fatigue monitoring is RMS.Generally, the greater the muscle activation, the higher the MMG amplitude [ 135 ].The RMS of the MMG signal is calculated from oscillation during the force plateau, which indirectly reflects the number of recruited MUs [ 131 ] and the mechanical characteristics of the contractile and viscoelastic components during muscle contraction [ 135 ], and more importantly, it can reflect the mechanical status of muscle fibers.The frequency-domain features of the MMG signal indirectly reflect the discharge rate of the activated MUs [ 131 ].The two most commonly used frequency-domain features for muscle fatigue are MPF and MDF.Overall, analysis of MMG signals indirectly reveals the neuromuscular strategies adopted by contractile muscles to activate and regulate force output during muscle contraction on the macroscopic level, and the mechanical properties of active muscle fibers on the microscopic level.

Application of MMG in Muscle Fatigue Monitoring
4.3.1 Isometric Contractions.MMG and force relationship in different muscle groups during isometric contractions have been the subject of muscle physiology studies.Mohamad et al. [ 136 ] compared the mean difference before and after repetitive submaximal (60% MVC) grip muscle contraction-induced muscle fatigue and found that the MMG RMS and the MMG MPF were significantly decreased.Although Camic et al. [ 137 ] found that there were decreases during the maximal isometric contraction in EMG amplitude, the EMG MPF, and the MMG MPF, no change in MMG amplitude.However, more detailed studies have shown that the MMG amplitude and its frequency content increased continuously during low-intensity exhaustive exercise, which may be related to the recruitment of additional MUs and the synchronization of active MUs [ 138 , 139 ].Xie et al. [ 33 , 140 ] used the Volterra-Wiener-Korenberg model and noise titration approach to detect deterministic chaotic character in the MMG signal measured from the biceps brachii during fatigue of isometric contraction at 80% MVC.The results indicated that MMG is a high-dimensional chaotic signal and supported the use of nonlinear dynamic theory for the analysis and modeling of fatigue MMG signals.Okkesim et al. [ 141 ] applied a new feature called frequency radio change (FRC) to evaluate muscle fatigue induced by 90 °-isometric elbow flexion (holding a 2.5 Kg-5 Kg dumbbell for 3 min) and proved the reliability of the new feature by making correlation analyses between this with MPF and MDF.The results showed that there was a high correlation between features, and the FRC can be used to quantitatively evaluate muscle fatigue.In addition, the MMG MPF Fatigue Threshold (MPF-FT) can be used to examine the global motor unit discharge rate during isometric contraction [ 71 , 142 ].

Dynamic Contractions.
In comparison to isometric muscle action models, studies on dynamic contractions can provide additional insights into motor control strategies used during concentric fatiguing tasks.One potential advantage of MMG over EMG during dynamic contraction is that it might be less affected by the relative positional change between the sensor and the signal source.Camic et al. [ 137 ] found that there was a reduction during the concentric contraction for EMG amplitude, the EMG MPF, the MMG MPF, and the MMG amplitude.Ebersole et al. [ 143 ] reported that with an increase in velocity from 60 to 300 °/s, there was a significant decrease in leg extension peak torque and an increase in MMG amplitude.However, there was no difference between the MMG MPF at 60 °/s and 300 °/s.Dinyer et al. [ 144 ] showed that during the performance of leg extension repetitions to muscle fatigue, MMG MPF decreased while MMG amplitude, and there was a positive relationship between EMG amplitude and MMG amplitude, suggesting the fatigueinduced changes in muscle excitation and MUs recruitment.Ebersole et al. [ 145 ] also found that the EMG amplitude increased while the MMG MPF decreased during isokinetic contractions to fatigue.Thus, the progressive decrease in MMG amplitude and MPF may suggest a cessation of MUs recruitment and a decrease in the firing rate of the recruited MUs, which along with a slowing in the CV resulted in diminished mechanical changes of the active muscle fibers.Beck et al. [ 146 ] examined the MMG amplitude and MPF responses from the biceps brachii during 50 consecutive maximal concentric isokinetic muscle actions of the forearm flexors at a velocity of 180 °/s.It was found that linear reduction in both MMG amplitude and MPF over the course of 50 repetitions.It was suggested that the decrease in MMG amplitude and MPF may have been due to fatigue-induced decreases in MU firing rates and/or reduced compliance in the muscle [ 147 ].
As mentioned above, the MMG signal is high-dimensionally chaotic [ 140 ] and should therefore be analyzed using nonlinear dynamics.Al-Mulla et al. [ 148 ] proved that the evolved pseudowavelet approach is an efficient method of classifying fatigue content in MMG signals.Qi et al. [ 149 ] describe and examine the variations in recruitment patterns of MUs in biceps brachii through a range of joint motion during dynamic eccentric and concentric contractions.And the MMG signals were decomposed into their intensities in time-frequency space using a wavelet technique.The MMG spectrum was then compared using principal component analysis.The results suggested that the MMG total intensity was lower during concentric than during eccentric contractions.

Electrically Elicited Contractions.
Dzulkifli et al. [ 150 ] found that MMG RMS gradually decreased as the stimulation continues (30 Hz) and its Zero Crossing Rate (ZCR) increased at the end of the session.Mohamad et al. [ 151 , 152 ] found that MMG RMS percentage (%RMS-MMG) declined after repetitive electrically evoked contraction (30 Hz) in both healthy people and people with tetraplegia and demonstrated that repetitive electrically evoked contraction can lead to accelerated muscle fatigue [ 151 ].Ibitoye et al. [ 153 ] applied two different electrically evoked contractions (high frequency, 35 Hz; low frequency, 20 Hz) to maintain upright posture in Spinal Cord Injury (SCI) patients.The results suggested that although the standing duration was generally longer for the low-frequency strategy, a faster muscle fatigue onset during high-frequency stimulation.The authors also suggested that, since MMG can distinguish between different stimulation frequencies, it is speculated that the MMG can quantify muscle fatigue during prolonged electrically evoked contractions.Based on MMG, Naeem et al. [ 154 ] used SVM to detect electrically evoked (30 Hz) muscle fatigue.The results showed that the average classification accuracy of non-fatigue and fatigue was 90.7% using Mel Frequency Cepstral Coefficient (MFCC) , 74.5% using RMS, and 88.8% with combined MFCC and RMS.

Electromechanical Delay.
It was suggested that if MMG is to be used as a means of monitoring force during functional activities, its relationship to force during dynamic activation must be considered [ 155 ].There is a delay between the onset of muscle activation and that of force development; this time lag is known as Electromechanical Delay (EMD) [ 155 , 156 ], more precisely, excitation EMD (E-EMD) .Figure 7 (a) shows the schematic diagram of the E-EMD.Calculations starting from the surface EMG and force signals were used to determine E-EMD, i.e., the onset of the EMG signal to the onset of force production (E-EMD E-F ) .The application of MMG signals serves to divide E-EMD into electrochemical and mechanical components [ 155 -159 ].The electrochemical component includes the events linked to EC coupling and pressure wave transmission to the skin surface detected at the start of the MMG signal, i.e., the onset of the EMG signal to the onset of the MMG signal (E-EMD E-M ) .The mechanical component reflects the time needed to take up the muscle-tendon unit slack before force transmission becomes effective at the tendon insertion point, i.e., the onset of the MMG signal to the onset of force production (E-EMD M-F ) [ 160 ].Thus, E-EMD E-F can be decomposed into E-EMD E-M and E-EMD M-F .With this decomposition, MMG can be used as a mechanical counterpart to EMG in analyzing fatigue [ 142 , 161 , 162 ].It has been proved that during supramaximal stimulation of the tibial nerve, E-EMD E-F increased significantly with advancing age and with muscle fatigue [ 163 ].
Conversely, the time difference between the end of muscle activation and the return of the force output toward pre-activation values is defined as the relaxation EMD (R-EMD) [ 164 ], which represents the electrochemical processes associated with the reversal of EC coupling and the mechanical events related to the relaxation of the series elastic component returning to a resting state after a contraction (Figure 7   signal to the cessation of force production (R-EMD E-F ) .Smith et al. [ 164 ] found that, with the deepening of muscle fatigue, R-EMD E-M and R-EMD M-F decreased.Cè et al. [ 165 ] recorded EMG, force, and MMG during biceps brachii MVC before and after fatigue.The results showed that, after fatigue, R-EMD E-M , R-EMD M-F , the R-EMD E-F , and the MMG peak amplitude all increased.The MMG peak amplitude was the main contributor to R-EMD.In addition, there are greater relative contributions from the R-EMD M-F than R-EMD E-M to R-EMD E-F .

Strengths and Limitations
The MMG technique is relatively easier to use than EMG, because it requires a less complex preparation procedure, without steps such as special skin preparation, coupling gel, direct skin contact, and so on [ 12 ].Most importantly, the MMG amplitude more tracks the fatigue-induced decline in torque production at each velocity closer than the EMG amplitude does [ 166 ], indicating that MMG amplitude may be more suitable for estimating muscle fatigue during dynamic contractions.This ability is complementary to that of EMG.
However, several factors affecting the signal have been identified during the analysis of dynamic contraction, including muscle length [ 167 ], muscle temperature [ 156 , 168 ], intramuscular pressure [ 169 , 170 ], and the effect of the volume conductor due to the thickness between the target muscle and the MMG sensors [ 171 ].Another important consideration of using MMG is that it can be obtained through various sensors, and each has specific limitations in different types of muscle contraction.For example, a piezoelectric contact sensor is suitable for large muscle groups such as quadriceps, but it is difficult to fix and is susceptible to motion artifacts.Thus, the choice of MMG sensor has been suggested to significantly affect the interpretation of muscle fatigue [ 172 ].It is necessary to investigate the sensors' responses during different muscle activities including different muscle contractions.

CONTRIBUTION OF NIRS TO MUSCLE FATIGUE MONITORING 5.1 Principles and Mechanisms of NIRS
Near-infrared (NIR) light (760 nm-1,100 nm) is the section of electromagnetic radiation wavelengths nearest to the normal range but just longer than what we can see (Figure 8 shows the different light wavelength ranges) [ 173 ].NIR spectroscopy (NIRS) helps determine the chemical composition of a compound or solution by measuring how much near-infrared radiation the compound or solution absorbs.The main principle behind the NIRS is the Beer-Lambert Law [ 174 ].According to this law, the concentration of a certain chemical compound in a solution determines how much light, whether visible or infrared, this solution will absorb.The higher the concentration, the more radiation energy of a specific wavelength will be absorbed [ 175 ].Because of this mechanism, NIRS is commonly referred to as vibrational absorption spectroscopy.Oxygen-related molecules are highly active in interacting with NIR light and based on the oxygen-dependent characteristics of NIR light, the changes in NIR absorption reflect the concentration of different these oxygen-related molecules, therefore NIRS can be used for real-time monitoring of oxygen availability in tissues such as skin, subcutaneous fat, muscle, and so on, inferring oxygenation and hemodynamic processes in these tissues [ 176 -178 ].

Sources of NIRS Signals
In skeletal muscle, the primary heme compounds are vascular hemoglobin (Hb) , intracellular myoglobin (Mb) , and cytochrome oxidase (cytox) [ 179 ].At 760 nm, Hb and Mb occur in the deoxygenated state (deoxy-Hb, deoxy-Mb), whereas at 850 nm, these two chromophores occur primarily in the oxygenated form (oxyHb and oxyMb) [ 180 ].Considering that the contribution of capillaries exceeds 90% of the total blood volume in muscle, under normal conditions, all regions of the muscle receive nearly fully oxygenated arterial blood.And the relative concentration of cytox is quantitatively small, thus its contribution to the NIRS signals is typically ignored.Therefore, oxygenation changes detected by NIRS would mostly reflect changes in capillary (Hb-related) and intracellular (Mb-related) oxygen levels [ 181 ].However, it should be recognized that the total (Hb +Mb) reflects the concentration of total heme in the field of view.When NIR light with a continuously changing frequency irradiates the muscle, the NIR passing through the muscle will weaken in some wavelength ranges as the result of selective absorption of different NIR wavelengths according to the hemodynamic processes occurring in the tissue [ 182 ].In the context of muscular fatigue, the proportion of the concentration of oxyHb and Hb in the muscular tissue will keep changing during muscle contraction, and these changes can be detected and used as an indicator of fatigue.
NIRS detects the light passing through the tissue as a consequence of the amount of light absorbed by the tissue under investigation (Figure 9 shows the functioning principle of NIRS).Relevant variables or parameters include oxy-Hb and deoxy-Hb, as well as derivatives such as total hemoglobin (tHb) and Muscle Oxygen Saturation (SmO 2 , which is calculated as [oxyHb/(oxyHb +deoxyHb)] × 100%), among others [ 176 ].More recent work using quantitative NIRS instruments and protocols [ 183 , 184 ] or morphometric modeling [ 185 ] has concluded that Mb likely contributes 60%-90% to the NIRS signals coming from skeletal muscle.Average values for Mb in human skeletal muscle range from ∼300 to 700 μm, with type I fibers reported having more Mb than type II fibers [ 186 ], potentially providing a new way for assessing the fiber composition of a muscle.

NIRS Technique and Its Measurable Parameters
Some characteristics, such as muscle oxygen consumption, and the ratio of oxygenated and deoxygenated from Hb and Mb in the muscle, can be determined by analyzing the optical density of transmitted or reflected light through the detector to detect the degree of fatigue [ 187 , 188 ].Continuous Wave NIRS (CW-NIRS) is the most widely used NIRS method currently and has already been employed successfully in fatigue assessment alone or together with EMG [ 189 ].The transmitted light source of CW-NIRS is of constant intensity, and the detected light intensity changes in light attenuation due to tissue absorption.Output variables typically are a "difference signal," representing tissue deoxygenation, and a "blood volume" signal, reflecting total signal strength (total [Hb +Mb]).The spatially resolved spectrometers systems provide a quantitative measure of tissue oxygenation: SmO 2 , Tissue Oxygen Saturation (StO 2 ), Tissue Oxygenation Index (TOI) , or Tissue Saturation Index (TSI) , all calculated as oxy[heme]*100/total [Heme] and expressed as percent [ 189 -191 ].However, CW-NIRS assesses the overall light attenuation inside the tissue and cannot differentiate the effects of scattering and absorption [ 192 , 193 ].Due to scattering in the tissues, the resulting signal is delayed, broadened, and attenuated [ 194 ].Time-domain NIRS (TD-NIRS) allows the assessment of the tissue absorption and scattering properties.In TD-NIRS, knowledge of these three variables permits the calculation of absolute deoxy [Hb +Mb], oxy [Hb +Mb], and their sum, total [Hb +Mb] in micromolar units.Frequency-domain NIRS (FD-NIRS) is sensitive to scattering changes and provides data with richer information content than CW-NIRS.One example is the ability of FD-NIRS to measure the absorption and reduced scattering coefficients of tissue, and thus the absolute concentrations of oxyHb and deoxy-Hb in muscle tissue.

Isometric Contractions.
When the muscle contracts, intramuscular pressure inhibits blood flow, resulting in a considerable reduction in blood volume and oxygenation.Sustained muscle contraction, however, would require an increase in blood flow so more oxygen to be delivered to the tissue.This is a hemodynamic change during sustained muscular contraction [ 195 ].Callewaert et al. [ 196 ] explored the muscle oxygen metabolism after performing 90-s submaximal (30%-40% MVC) isometric knee extension.The results indicated deoxy [Hb +Mb] increased significantly after muscle fatigue.McNeil et al. [ 197 ] explored how blood flow and muscle oxygenation respond to isometric contractions at 30%, 60%, and 100% MVC when performing 1-min dorsiflexion contractions.The results indicated that TOI of the tibialis anterior decreased with the deepening of muscle fatigue.Paiziev et al. [ 198 ] assessed muscle oxygenation during sustained isometric low (30% MVC), moderate (60% MVC), and submaximal (90% MVC) contraction of the dorsiflexor muscle.The results indicated that limited muscle blood flow due to vasoconstriction and intramuscular pressure during sustained isometric contraction leads to fatigue due to a lack of oxygen and nutrients.Scano et al. [ 199 ] applied TD-NIRS and sEMG to record on the deltoid lateralis muscle and found that StO 2 and deoxy-Hb are slightly better descriptors of sustained fatigue than oxyHb, since they showed a higher correlation with MDF, while tHb correlation with MDF was lower.Pethick et al. [ 200 ] explored the muscle oxygen consumption of the knee extensor during intermittent isometric contractions (50% MVC) at 30 °, 60 °, and 90 °of knee flexion to task failure.Over time, the complexity of the sEMG signal decreased at both 90 °and 60 °knee flexion conditions, but not at the 30 °knee flexion situation.Further, in the same way, while sustained muscle contraction above ∼25% MVC would restrict the circulation supply of contracting muscle, more extended joint angles would slow the fatigue-induced increase in muscle oxygen consumption.

Dynamic
Contractions.Bailey et al. [ 201 ] found that after low-load, isokinetic, concentric/eccentric elbow flexion, biceps brachii, and brachialis MUAP CV, the biceps oxyHb decreased.And it is suggested that adjustments in biceps brachii oxygenation were linked to changes in MUAP CV more local to the site of fatigue with older age.Matsuura et al. [ 202 ] compared the muscle NIRS responses during static and dynamic contractions and found that, during all contractions, there was a rapid decline in muscle blood volume and muscle oxygenation during the early stages of the contractions, with a plateau or slight increase toward the end.It suggested that muscle fatigue was most likely mediated by peripheral factors.Denis et al. [ 203 ] found that the decrease in TOI observed between the beginning and post-fatigue values were significantly greater during knee eccentric extension than during knee concentric extension, while no significant change was found for tHb volume.Yoon et al. [ 204 ] found that the oxyHb and TOI decreased while deoxy-Hb increased in upper trapezius muscles during a repetitive arm motion-induced fatiguing task.Cherouveim et al. [ 205 ] compared skeletal muscle oxygenation levels during and after submaximal concentric and eccentric isokinetic exercise (60% MVC, 60 °/s).The results demonstrated that oxyHb and TSI decreased while deoxy-Hb and tHb increased in vastus lateralis muscles, but there was no difference between the two exercise modes.

Electrically Evoked Contractions.
Muthalib et al. [ 206 ] compared the changes in biceps brachii muscle oxygen consumption during voluntary and electrically evoked (30 Hz) isometric contractions (30% MVC).The results suggested that the metabolic demand for oxygen in electrically evoked contraction is greater than voluntary contraction at the same torque level.This is consistent with the findings of McNeil et al. [ 207 ].Muthalib et al. [ 152 ] also revealed that TSI was proportional to pre-fatigue and inversely proportional to the MMG RMS during post-fatigue.Mohamad et al. [ 151 ] assessed electrically evoked (30 Hz) fatiguing muscle with changes to the MMG RMS and the TSI in the extensor carpi radialis.Regression analysis showed that TSI was inversely proportional to the MMG RMS during post-fatigue.Quantifying the transient changes in muscle microcirculation and metabolism during electrically evoked contraction before muscle contraction occurs is crucial, because it enables the current intensity to be optimally tuned, thereby reducing the electrical stimulation-induced muscle pain and fatigue.Huang et al. [ 34 ] conclude that NIRS can be used to determine the optimal current intensity by monitoring oxygenation changes when performing electrically evoked contraction.Szczyglowski et al. [ 208 ] used electrically stimulated (100 Hz) eccentric knee extension to produce the Exercise-induced Muscle Damage (EIMD) and found that a more rapid depletion of anaerobic energy stores and/or accumulation of metabolic by-products leads to fatigue following EIMD.However, McCully et al. [ 209 ] found no differences in oxy-Hb values between the resting and the electrical stimulation (2 Hz, 4 Hz, and 6 Hz).This may be because the stimulation protocol showed no evidence of limiting oxygen delivery, and thus the test should reflect oxidative metabolism rather than contraction-induced limited blood flow.

Strengths and Limitations
NIRS allows continuous, non-invasive measurements of tissue oxygenation.It provides real-time trends in tissue oxygenation during muscle fatigue from a hemodynamic perspective.Such monitoring provides a sensitive signal, which can help prevent muscle damage caused by acute hypoxia [ 210 , 211 ].One of the key advantages of using NIRS over other measurement modalities is that its signals can yield acceptable signal-to-noise ratios even during dynamic exercise.
It is important to note that non-invasive monitoring of muscle oxygenation using NIRS presents several limitations.First, the adipose tissue thickness, the skin with varying concentrations of melanin, complex tissue arrangements, the presence of large blood vessels, and so on, may interfere with the measurement.And the muscle will often change its shape during contraction, which can alter the scattering of light traveling through the tissues [ 196 ].Furthermore, due to the presence of blood circulation, the changes in muscle oxygen caused by muscle fatigue have a time lag compared to MMG signals and EMG signals [ 212 ].This means that the signals of muscle oxygen might not be suitable for the early prediction of muscle fatigue as EMG is.NIRS in combination with other techniques, such as EMG, can interpret muscle fatigue phenomena from multiple dimensions [ 213 ].However, this combination also provides some issues.The EMG electrodes are usually smaller than the NIRS probes, meaning that both do not guarantee that the same volume of tissue is mapped.This means that, in general, the mapping of the muscle with EMG electrodes might not be completely consistent with the NIRS probes [ 214 ].

CONTRIBUTION OF US TO MUSCLE FATIGUE MONITORING 6.1 Principles and Mechanisms of US
The US consists of mechanical sound waves with frequencies above the audible range, ( ≥20 kHz) [ 215 ].Typical frequencies used in the medical US are 2-10 MHz [ 35 ].Currently, commercial US devices usually use piezoelectric crystal material to generate US waves [ 216 ].In these transducers, the piezoelectric crystals transform electric energy into mechanical energy, and the generated US waves are sent from the transducer, propagate through tissues under investigation, and part of the acoustic energy would return to the transducer.The crystals inside the transducer convert the received ultrasonic beam returned from the body into electrical impulses, which are further processed to form inferences about different properties of the tissues under investigation [ 217 ].
As the US beam passes through the body, its intensity and amplitude decrease, a process known as attenuation.The degree of attenuation depends on the type of tissue the sound wave is passing through.Where the tissue is densely packed (such as bone), attenuation will be much greater than in less densely packed tissue (such as fat).There are three main causes of attenuation: reflection, scatter, and absorption (Figure 10 shows the main causes of ultrasound beam attenuation).Reflection is defined as the amount of sound returned to the transducer, which occurs as a result of the interaction of sound at an interface of two types of medium.Scatter is the redirection of sound in many directions when the tissue interface is smaller than the wavelength of the sound wave and is the reason different echogenic structures can be seen within a particular tissue.Absorption occurs when acoustic energy is converted into heat, a major cause of attenuation.The number of absorption increases as both transducer frequency and increasing scanning depth.Especially, the US was used to probe the structural and morphological properties of skeletal muscles.The US wave will be reflected at the interface of muscle-muscle, muscle-soft tissue, and muscle-skeleton interface with different acoustic impedance in the process of passing through the human muscle tissue, returning detailed muscle information such as muscle thickness change, muscle fiber angle, muscle fascicle length, muscle cross-sectional area, and so on.This detailed information can be used to detect and predict the level of fatigue much more accurately [ 218 , 219 ].Fig. 10.The main causes of ultrasound beam attenuation.Reflection is defined as the amount of sound returned to the transducer.Scatter is the redirection of sound in many directions when the tissue interface is smaller than the wavelength of the sound wave.Absorption occurs when acoustic energy is converted into heat, a major cause of attenuation.Fig. 11.Ultrasound (US) mode and its imaging.(a) Amplitude mode (A-mode) US.A-mode US measures the distance by displaying the amplitude of the wave versus time (i.e., depth) for the return of that pulse and plots it as a one-dimensional image.(b) Brightness mode (B-mode) US.B-mode US is displayed in the form of a gray scale or with color.In the B-mode US, the varying amplitudes are converted into dots of varying intensity used to generate two-dimensional images.Weaker reflections appear as darker gray dots, whereas stronger reflections appear as brighter white dots.(c) Elastography (E-mode) US.A color-coded map represents the tissue hardness with qualitative color scales (white arrow).In the US strain image, the red color represents soft tissue, while the blue color indicates hard tissue.

US Technique for Muscle Fatigue Detection
There are three main US techniques for muscle fatigue detection, including amplitude mode (A-mode), brightness mode (B-mode) , and elastography (E-mode) [ 35 ] (Figure 11 shows the different US modes and their imaging).
A-mode displays the echo signals of the US in the form of waves.It measures the distance by displaying the amplitude of the wave versus time (i.e., depth) for the return of that pulse and plots it as a one-dimensional image [ 215 ].A-mode is ideal for measuring distances between the various structures.This method is very sensitive to changes in various parameters of echo, and its volume is very small.In recent years, it has received greater attention for motion intention recognition [ 220 -222 ] and fatigue detection [ 223 -225 ].Depth, that is, muscle thickness, is the most commonly used A-mode US parameter for detecting muscle fatigue.
B-mode US is also known as US imaging, displayed in the form of grayscale or with color.In the B-mode US, the varying amplitudes are converted into dots of varying intensity used to generate two-dimensional images.The intensity of light dots reflects the intensity of ultrasonic dots reflected and attenuated by the echo interface [ 226 ].Weaker reflections appear as darker gray dots, whereas stronger reflections appear as brighter white dots.A recent systematic review lists 17 studies making use of B-Mode US for biomonitoring muscle and tendon dynamics during locomotion [ 227 ].Muscle thickness [ 201 , 228 ], muscle cross-section area [ 229 ], pennation angle [ 230 ], and fiber length [ 231 ] are the most commonly used B-mode US parameters for detecting muscle fatigue.
Muscle fatigue is known to transiently increase muscle stiffness [ 232 ].A relatively new imaging technique is E-mode US, which adds value by assessing tissue elasticity.E-mode US is an imaging technology sensitive to tissue stiffness [ 233 ].It is used for imaging tissue stiffness (e.g., muscle) in either passive or active conditions [ 12 ].There are more types of elastography, but the two most used are Strain Elastography (SE) and Shear-wave Elastography (SWE) .SE evaluates the stiffness by applying external pressure, which deforms the tissue.The deformation is named strain.To measure the stiffness, SWE uses shock waves generated by the machine.Muscle shear elastic modulus is the most commonly used E-mode US parameter for detecting muscle fatigue.

Application of US in Muscle Fatigue Monitoring
6.3.1 Isometric Contractions.Shi et al. [ 234 ] found that, during a 30-second 80% MVC isometric contraction to fatigue, the biceps brachii thickness detected by B-mode US images and the RMS of sEMG increased, while the MDF of sEMG decreased.It is suggested that the morphological changes of muscles detected by the US are consistent with those obtained from sEMG during muscle fatigue processes.Image entropy is usually used to represent the texture of an image.It is further introduced into information theory to describe the uncertainty of a system.Zimmer et al. [ 235 ] demonstrated that the local entropy as a textural parameter of the muscle can measure the homogeneity of the tissues in US images.Li et al. [ 236 ] investigated the B-mode US image entropy and sEMG signals during a static sustained isometric contraction at various intensities (20%, 30%, 40%, and 50% MVC).The results showed that US image entropy decreased and the RMS of sEMG increased along with the fatigue, suggesting it may have a superior performance in evaluating the muscle fatigue state.Muscle Fascicle Length (FL) is defined as the linear distance between the points at which the fascicle intercepted the extrapolated lines of the superficial and deep aponeurosis.Muanjai et al. [ 231 ] found by the B-mode US that although FL was shortened following eccentric contraction, the development of long-lasting fatigue following eccentric contraction is not due to the change in muscle length.Recently, A-mode US has gained attention in the context of fatigue analysis.Muscle deformation under different movements will lead to changes in features of A-mode US signals.Therefore, the A-mode US echo signal on the muscles can reflect the interface of the muscle-bone and muscle deformation.By analyzing the echo signal of muscles, different movements can be decoded, which can be used to classify the muscle fatigue.Sun et al. [ 223 ] used one micro-size single-element US transducer, i.e., A-mode US, to continuously monitor the thickness change of biceps brachii during elbow isometric contraction at 50% MVC.The result shows that the muscle thickness of the biceps brachii continues to increase during fatigue.And this is the first study to detect muscle fatigue using A-mode US.Linear Discriminant Analysis (LDA) is a supervised machine learning algorithm that is utilized to find a new feature space to project the data to maximize inter-class separability and thereby predict the class label of the test feature [ 237 ].Zeng et al. [ 224 ] experimented with eight types of gestures, with a range of 0-60% MVC for each motion.The gesture recognition accuracies of sEMG and A-mode US under the non-fatigue state and fatigue state are compared through an LDA classifier.The experimental results demonstrated that the fatigue robustness of A-mode US signals is better than that of sEMG signals.A-mode US signals have advantages over sEMG signals in terms of both accuracy and muscle fatigue sensitivity.In addition, Brausch et al. [ 225 ] found that comparatively simple machine learning models, such as SVM or logistic regression (LR) , the model based on A-mode US yielded the best performance in the accuracy of muscle fatigue classification.By applying SWE, some studies reported a decrease in shear elastic modulus during [ 238 , 239 ] and immediately after [ 240 , 241 ] a muscle fatigue protocol based on isometric contractions.

Dynamic Contractions.
Although several studies have used the US in dynamic contractions [ 242 , 243 ], there is comparatively less explored than in the static contraction condition.Recently, many studies reported the simultaneous collection of EMG signals and the US.But it was noted that the placement of the US probe on the skin would introduce external motion artifacts to the sEMG measurements during dynamic contractions.Huang et al. [ 244 ] introduced a new method to synchronize B-mode US images, EMG signals, joint angles, and other related signals (e.g., force and velocity signals) in real time.Preliminary experiments demonstrated that the sEMG signals were not significantly affected by the amount of the US gel.The system is being used for the study of muscle fatigue.Gonzalez-Izal et al. [ 242 ] found that the echogenic intensity detected by the US after eccentric contractions is higher than that detected by concentric contractions, which may explain the fact that eccentric contractions lead to greater muscle damage.More recently, Wang et al. [ 243 ] collected the B-mode US images and the corresponding reflected ultrasonic signals of the biceps in the rest, dynamic contraction, and fatigue state and applied the ultrasonic attenuation coefficient as an indicator for assessing muscle fatigue.The results demonstrated that the muscle thickness and the ultrasonic attenuation coefficient increased to the maximum in the contraction state and decreased in the fatigue state.

Electrically Evoked Contractions.
Witte et al. [ 245 ] applied SE imaging to capture the elastic and viscoelastic changes in the 3rd flexor digitorum superficialis muscle after maximum isometric contraction to fatigue.The results indicated that higher strains occurred near the muscletendon junction during an isometric contraction, revealing the effects of a sustained maximal contraction dominated by low-frequency fatigue (2 Hz).During muscle contraction, muscle fibers are deformed to generate force.Sheng et al. [ 246 ] investigated a quantifiable muscle displacement adaptive speckle tracking algorithm to determine the axial strain changes in the quadriceps muscle during the electrically evoked (35 Hz) muscle fatigue isometric knee extensions.The results showed that there was a reduction in the strain peaks, a change in the strain peak distribution, and a decrease in an area occupied by the large positive strain.Several studies indicated that the proposed measurements from US images are promising to quantify the muscle contractility changes during electrical stimulation [ 246 , 247 ].In future studies, this methodology will be applied to improve the control performance of an electrical stimulation-based neuro-prosthesis in a way that the assessed degree of contractility is used as feedback information to modulate the delivered electrical stimulation dosage.Zhang et al. [ 248 ] proposed to use US imaging-derived echogenicity signal as an indicator of electrically evoked (33 Hz) muscle fatigue.The results showed an exponential reduction in the US echogenicity relative change as the fatigue progressed under the isometric and dynamic conditions.The results also implied a strong linear relationship between US echogenicity relative change and muscle fatigue.In a 2022 study, Sheng et al. [ 249 ]  demonstrated experimentally for the first time that real-time measurement of the muscle state due to the induced muscle fatigue by ultrasonic strain measures and its integration into feedback can be used for the hybrid exoskeleton controller's decision-making.

Strengths and Limitations
Using the US as the detection signal for fatigue detection has the following advantages: the detection is non-invasive; it can provide a wealth of information of a targeted muscle irrespective of its depth in real time; what is more, it allows direct visualization of the target muscle.Due to its relatively high temporal and spatial resolution, two-dimensional anatomical images of the region of interest can be obtained and further processed to provide comprehensive muscle contraction.
However, it also has some limitations.First, since the sizes of human muscles are in general too large when compared to the US probe, muscle contraction could not be observed with the whole muscle in the field of view due to technological limitations [ 250 ].Second, the noise in US images increases with the increase of depth, so the initial detection sensitivity of deep muscle is also low [ 250 ].Finally, only type A-mode US may be more appropriate for real-time, wearable fatigue monitoring.

DISCUSSION AND CONCLUSION
This review provided an overview of the literature on various non-invasive techniques of muscle fatigue monitoring and detection.We extensively described muscle fatigue, its neuromuscular mechanism, and commonly used biomarkers.And then, the principle and mechanism, parameters used for fatigue detection, application in fatigue detection, as well as advantages and disadvantages of each technology are discussed in detail for EMG, MMG, NIRS, and the US, respectively.Non-invasive fatigue monitoring techniques are briefly summarized in Table 2 .
Muscle fatigue impairs physical performance and/or cognition, causing changes in muscle activation patterns, which in turn leads to an increase in the risks for injury and fatality [ 251 ].Studies suggested that lower extremity muscle fatigue affects postural control ability and dynamic balance in elderly people [ 252 , 253 ] and exercise performance in young adults [ 254 ].In prosthetic control, muscle fatigue can deteriorate the effectiveness of the sEMG-based control in the long term, resulting in less accurate results in sEMG-based control [ 255 , 256 ].Mahoney et al. [ 257 ] found that individuals with SCI are more susceptible to low-frequency fatigue than able-bodied subjects.Muscle fatigue increases the time of E-EMD E-F , but understanding how much muscle fatigue affects the E-EMD E-F in people with SCI could help in the design of closed-loop neuro-prostheses that compensate for this delay [ 258 ].Duncan et al. [ 259 ] showed that muscle fatigue may interfere with physiotherapy and can negatively affect stroke survivors' physical and psychological function.Qing et al. [ 260 ] demonstrated that the decoding accuracy of the classification model decreased by an average of 7% considering muscle fatigue.Thus, monitoring the physiological information in the process of muscle fatigue to quantify muscle fatigue and proposing new algorithms to compensate for the effects of muscle fatigue on physiological signals are of crucial relevance in areas such as myoelectric control and sports science [ 54 , 261 ].
A typical experiment in muscle fatigue research involves a subject performing a specific task while sensors attached to the skin detect changes in signals arising from the movement.These tasks typically include static contraction (isometric) [ 86 , 197 ], dynamic contraction (concentric, eccentric, or isokinetic) [ 76 , 262 ], and electrically elicited contraction [ 116 , 122 ].Another method is to recognize movement patterns, like gestures [ 224 ].Non-stationary signal parameters of dynamic contractions [ 84 , 85 ] increase the difficulty of signal analysis and feature extraction.Conversely, signals during static contraction are relatively stable [ 86 ].Therefore, most muscle fatigue studies were focused on isometric contraction.In most studies, the acquired signals were then recorded and post-processed to reveal the characteristics of the muscle during that particular exercise.Surprisingly, most of the muscle fatigue studies have been mainly short-term studies, conducted in laboratory settings, likely due to the long-term robustness of measurement modalities for fatigue monitoring.This limitation affected the practical use of wearable fatigue monitoring devices in the real world.At present, most studies on fatigue monitoring only focus on a single physiological signal, because using physiological signals as indicators of muscle fatigue can enable objective, real-time, wearable muscle fatigue monitoring at the individual level.However, individuals have different responses to fatigue, and physiological signals are susceptible to other factors (environmental conditions, emotions, pathophysiology, etc.) [ 262 ], so the sensitivity and reliability of fatigue detection/prediction based on a single physiological signal are still unclear.A possible solution is to concurrently acquire the information from different modalities and integrate them for further analysis to complement each other.However, as mentioned above, sEMG signals are weak and easily affected by many factors [ 263 , 264 ], so it is difficult to recognize slight muscle fatigue only by using sEMG signals, and the detection accuracy is strongly influenced by environmental factors [ 54 ].To address the above issues, multi-modal sensing technology centered on sEMG may contribute to the accuracy of muscle fatigue detection and new mechanistic interpretation of peripheral fatigue.In the discussion of the mechanisms of muscle fatigue, Guo et al. [ 265 ] developed a multi-channel compact-size wireless hybrid sEMG/NIRS acquisition system and demonstrated that combining the advantages of sEMG and NIRS in terms of spatial and temporal resolution can monitor muscle activation from the perspective of electrophysiology, hemodynamics, and oxidative metabolism that reveal fatigue mechanisms.Ding et al. [ 266 ] fused sEMG, NIRS, and MMG into a compact sensor.The hybrid sensor system can be used to monitor muscle motion and explain muscle fatigue from the modalities of electrophysiology, optics, and acoustics.Yoshitake et al. [ 170 ] recorded the signals of sEMG, MMG, and NIRS during isometric back extensions for a period of 60 s.The results demonstrate that the restriction of blood flow due to high intramuscular mechanical pressure is one of the most important factors in muscle fatigue.Blangsted et al. [ 267 ] recorded the signals of biceps brachii during static elbow flexion at 10% MVC and found that MMG RMS increased, MMG MPF, EMG MPF, and muscle tissue oxygenation decreased, but intramuscular pressure remained constant.The author believed that decreased muscle tissue oxygenation did not underlie either acute or long-term muscle fatigue development evidenced.Moreover, in the discussion of the classification accuracy of muscle fatigue, Sarillee et al. [ 268 ] used KNN to classify the extracted features.The results showed that the mean accuracy of EMG, MMG, and AMG was 87.10%, 81.40%, and 67.23%, respectively.But the mean accuracy of the multimodal system was improved to 92.07%.Sheng et al. [ 213 ] presented a multi-modal sensing system that can collect signals from sEMG, NIRS, and MMG simultaneously.LDA classifier was constructed for classifying the extracted features.The results convincingly demonstrated a significantly improved classification accuracy by using multi-modal features.The classification accuracy is compensated by 3.6%-22.9%in the presence of muscle fatigue.It is proven that it is feasible to improve the performance of muscle fatigue detection and classification through sensor fusion.In a recent study, Sheng et al. [ 269 ] proposed a Multimodal Multilevel Converged Attention Network (MMCANet) model for multisource signals composed of sEMG and A-mode US and compared MMCANet's performance with that of traditional machine learning algorithms for sEMG and A-mode US features.The experimental results showed that the proposed model improved by 14.31% and 3.80% over the CNN method with single sEMG and A-mode US, respectively.Potentially, the appropriate fusion of sEMG, MMG, NIRS, US, and other signals could address limitations of single sensor modality, as the modalities provide complementary information of the muscular activities to each other.
There are still limitations in terms of accuracy, robustness, data volume, and flexibility when the current multi-modal methods are applied to practical applications.First, high-quality data acquisition is the premise of high-precision multi-sensor muscle fatigue recognition.For multi-sensor fusion, variability in the subject's muscle and physical condition is large, making it difficult to obtain high-quality multi-sensor data.Given that the quality of data in each dimension can notably affect the final recognition accuracy, the data quality of measurement signals stands out as the most important factor.As mentioned above, muscle fatigue is a complex physiological phenomenon.Therefore, explaining the comprehensive changes in muscle fatigue from multiple physiological perspectives is helpful to improve the accuracy of muscle fatigue recognition, such as electrophysiology (i.e., EMG), biomechanics (i.e., MMG), blood oxygen metabolism (i.e., NIRS), morphology (i.e., US), and so on.Second, advanced detection techniques could be possible solutions to the problem of low-quality data, such as HD-sEMG, more sensitive inertial sensors, light detectors, high-intensity linear array US transducers, and so on.The difficulty in trying to attach multiple sensors on a muscle, especially to measure a small muscle group (such as the muscles of the hand).Even when measuring a large muscle group (such as the quadriceps), there is limited space to apply too many measurement techniques simultaneously.Future research should focus on integrating multiple sensors into a smaller sensor without compromising the data quality of each dimension.In addition, most of the current research on muscle fatigue detection involved recruiting a relatively small size of volunteers to collect signals, such as 5-20 healthy individuals [ 224 , 245 , 268 ], or no more than five patients [ 213 ].It may be difficult to know how well fatigue prediction models are applied in real life.Thus, it is particularly important to establish more open-source databases with high data quality, complete muscle measurement sites, and rich application scenarios.Finally, to improve the identification and classification accuracy of muscle fatigue models with multi-sensors, more advanced algorithms for signal analysis should be proposed and verified.

Fig. 1 .
Fig. 1.Different types of fatigue and physiological systems involved.Perceived fatigue is derived from the individual's self-reported subjective sensations that are based on homeostasis and the psychological state.Performance fatigue depends on the peripheral contractile function, as well as the central activation.

Fig. 2 .
Fig. 2. The neuromuscular mechanisms of performance fatigue.Performance fatigue includes fatigue in supraspinal and spinal (central fatigue) and fatigue in the neuromuscular junction and the muscle (peripheral fatigue).Central fatigue is directly related to a decrease in reduced efferent drive from the motor cortex and consequently the reduced discharge rate of motoneurons.Peripheral fatigue refers to a reduction in muscular output resulting from the changes in the electrochemical mechanisms (e.g., neuromuscular transmission and excitation-contraction (EC) coupling) and mechanical mechanisms (e.g., altered viscoelasticity and stiffness of the contractile apparatus) downstream of the neuromuscular junction.

Fig. 4 .
Fig. 4. Diagram of motion units.A motor unit consists of an α motoneuron and all the muscle fibers it innervates.

Fig. 5 .
Fig. 5.The generation of electromyography (EMG) signals.All muscle fibers of a motor unit (MU) are innervated simultaneously by a discharge of the α motoneuron, so all the fibers of a MU would form a MU action potential (MUAP).One motoneuron action potential (AP) would elicit one MUAP.Therefore, a sequence of motoneuron APs would generate a train of MUAPs, i.e., MUAPt.When there are multiple active MUs within the detection volume of the electrode, multiple MUAPts can be seen.The composite signal is usually called EMG.

221 : 10 N
. Li et al. which the EMG power spectrum is divided equally into two regions, referring to Equation ( 2 ):

Fig. 7 .
Fig. 7.The schematic diagram of electromechanical delay (EMD).(a) Excitation EMD (E-EMD).Calculations starting from the onset of the EMG signal to the onset of force production (E-EMD E-F ).The application of MMG signals serves to divide E-EMD into the electrochemical and mechanical component.The electrochemical component includes the onset of the EMG signal to the onset of the MMG signal (E-EMD E-M ).The mechanical component reflects the onset of the MMG signal to the onset of force production (E-EMD M-F ).(b) Relaxation EMD (R-EMD).Similar to E-EMD, R-EMD can be divided into E-EMD into the cessation of the EMG signal to the cessation of the MMG signal (R-EMD E-M ), the cessation of the MMG signal to cessation of force production (R-EMD M-F ), and the cessation of the EMG signal to the cessation of force production(R-EMD E-F ).
(b) shows the schematic diagram of the R-EMD).Similar to the E-EMD, the assessment of the EMG, MMG, and force signals allows for the identification of the cessation of the EMG signal to the cessation of the MMG signal (R-EMD E-M ) , the cessation of the MMG signal to cessation of force production (R-EMD M-F ) , and the cessation of the EMG

Fig. 8 .
Fig. 8.Light wavelength ranges.Violet light is said to have the shortest form of wavelength, whereas red light is said to have the longest wavelength.The wavelength of visible light ranges from 400 nm to 700 nm.Near-infrared (NIR) light (760 nm-1,100 nm) is the section of electromagnetic radiation wavelengths nearest to the normal range but just longer than what we can see.Moreover, wavelength is inversely proportional to the energy.

Fig. 9 .
Fig. 9.The diagram of the functioning principle of NIRS.The light source on the skin surface (right) emits light to the tissue through the skin.The lights are scattered by and absorbed in the tissue, and a part of the scattered lights is detected by the light detector on the skin surface (left), usually placed at a distance of a few cm from the light source.
integrated a real-time US image acquisition and processing framework into a switching-based feedback control of the hybrid knee exoskeleton to monitor muscle responses to electrically evoked contractions.The results 221:26 N. Li et al.

Table 1 .
Categories of EMG Electrodes 1. Quick, easy to apply 2. No medical supervision 3. Minimal discomfort 4. Wireless acquisition can be realized 1.Only for superficial muscles 2. Cross-talk 3. Secretion of oil or sweat from the skin can affect signals quality Array/Grid Multipolar

Table 2 .
(2)-invasive Techniques for Muscle Fatigue MonitoringThe adipose tissue thickness, the skin with varying concentrations of melanin, complex tissue arrangements, and the presence of large blood vessels during contraction may interfere with the measurement; (2) Due to the presence of blood circulation, the changes in muscle oxygen caused by muscle fatigue have a time lag compared to MMG signals and EMG signals.Muscle contraction could not be observed with the whole muscle in the field of view;(2)The noise in US images increases with the increase of depth, so the initial detection sensitivity of deep muscle is also low; (3) Only type A-mode US may be more appropriate for real-time, wearable fatigue monitoring.