Muscle movement fatigue degree evaluation system based on wireless dry electrode signal

Abstract

The application discloses a muscle movement fatigue degree evaluation system based on a wireless dry electrode signal, and comprises the following modules: an electromyography acquisition module, which is used for constructing electromyography movement time series data; a signal processing module, which is used for signal conditioning processing and generating a periodic electromyography sequence; a baseline construction module, which is used for constructing an individual baseline sample set and generating a time series embedding sequence; a time series modeling module, which is used for inputting the time series embedding sequence into an improved DeepAR model, introducing a phase sequence shift clustering mechanism in a time series recursion layer, and outputting an electromyography probability sequence and a reference distribution parameter; a topological analysis module, which is used for extracting topological structure information of a periodic evolution point cloud by using a Persistent Homology algorithm and generating a topological barcode sequence; and a fatigue control module, which is used for outputting a training limit instruction or a termination instruction. The application realizes high-precision evaluation and real-time regulation and control of the muscle movement fatigue degree.

Description

Technical Field

[0001] This invention relates to the field of bioelectric signal processing technology, and in particular to a muscle fatigue assessment system based on wireless dry electrode signals. Background Technology

[0002] With the increasing demand for smart wearable devices and sports health monitoring, technologies for analyzing surface electromyography (EMG) signals generated during human muscle activity and assessing muscle fatigue have received widespread attention. Existing muscle fatigue assessment methods mostly rely on wet electrode acquisition or signal analysis based on single statistical characteristics for fatigue determination, but these methods generally suffer from the following problems in practical applications: Acquired surface electromyography (EMG) signals are susceptible to motion artifacts, baseline drift, and electromagnetic interference. The effective frequency band components and low-frequency perturbation components overlap, leading to poor signal stability and significant fluctuations in feature extraction results, thus reducing the reliability of fatigue assessment results. There is a time asynchrony between multi-channel EMG signals and motion state parameters such as movement rhythm and joint displacement. Existing methods have limited accuracy in aligning multi-source signals, making it difficult to guarantee the accuracy of movement cycle segmentation and the consistency of periodic EMG sequences. For EMG signals with significant nonlinear and periodic variation characteristics, traditional methods based on mean, variance, or frequency... Analysis methods for domain energy distribution struggle to characterize structural changes and phase shifts between different action cycles, leading to the loss of crucial structural information during fatigue evolution. In terms of sequence modeling, existing methods based on conventional time series models typically only utilize amplitude variation information, lacking the ability to model the evolutionary relationships of periodic structures and failing to reflect phase sequence changes and rhythmic disturbances during muscle fatigue. Furthermore, existing methods lack the ability to characterize the overall evolutionary morphology of electromyographic sequences at the topological level, failing to describe the trend of muscle state changes from a multi-cycle perspective, resulting in insufficient adaptability of fatigue assessment results to complex motion scenarios.

[0003] Therefore, how to provide a muscle fatigue assessment system based on wireless dry electrode signals is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a muscle fatigue assessment system based on wireless dry electrode signals. This invention collects surface electromyography (EMG) signals and motion state parameters from multiple muscle groups using wireless dry electrode clothing, constructs time-series EMG data, generates periodic EMG sequences through a signal processing module, forms an individual baseline sample set and generates a time-series embedded sequence through a baseline construction module, inputs the time-series embedded sequence into an improved DeepAR model through a time-series modeling module, introduces a phase-shift clustering mechanism in the time-series recursive layer to generate EMG probability sequences and reference distribution parameters, constructs a periodic evolution point cloud and extracts a topology barcode sequence through a topology analysis module, and calculates probability offsets and topology offsets through a fatigue control module, combining them with drift parameters to generate a fatigue score, thus achieving muscle fatigue assessment. This system possesses advantages such as strong anti-interference capability, strong periodic structure characterization capability, and high fatigue assessment accuracy.

[0005] A muscle fatigue assessment system based on wireless dry electrode signals according to an embodiment of the present invention includes: The electromyography (EMG) acquisition module is used to acquire surface EMG signals from multiple muscle groups through a wireless dry electrode garment worn on the body surface of the test subject, and simultaneously acquire the motion state parameters of the test subject to construct EMG motion time series data. The signal processing module is used to perform signal conditioning on the surface electromyography (EMG) signals in the EMG motion time series data, determine the start and end positions of the action cycle based on the motion state parameters, perform periodic segmentation and length alignment on the surface EMG signals, and generate periodic EMG sequences. The baseline construction module is used to acquire the periodic electromyography (EMG) sequence of the test subject in a non-fatigue state, construct an individual baseline sample set, extract feature parameters from the individual baseline sample set, encode the periodic EMG sequence and the feature parameters, and generate a time-series embedding sequence. The temporal modeling module is used to improve the DeepAR model by embedding temporal sequences into the input sequence. It introduces a phase drift clustering mechanism in the temporal recursive layer, divides the periodic electromyography sequence into phase segments and maps them into phase cluster chains, performs drift calculation on the phase cluster chains of adjacent action cycles, and outputs the electromyography probability sequence and reference distribution parameters. The topology analysis module is used to construct a high-dimensional feature point set based on the periodic electromyography sequence and reconstruct the phase space to generate a periodic evolution point cloud. The Persistent Homology algorithm is used to extract the topological structure information of the periodic evolution point cloud and generate a topological barcode sequence. The fatigue control module is used to calculate probability offset and topological offset based on electromyographic probability sequence, reference distribution parameters and topological barcode sequence, and generate local fatigue index by combining drift parameters. It performs cumulative analysis on local fatigue index of multiple action cycles to generate fatigue score. If the fatigue score reaches the first preset threshold, it outputs training restriction instruction. If the fatigue score exceeds the second preset threshold, it outputs termination instruction.

[0006] Optionally, the electromyography (EMG) acquisition module specifically comprises: The wireless dry electrode garment is attached to the corresponding muscle group on the body surface of the test subject. The target muscle group is divided into channels according to the preset channel distribution and a channel number sequence is established. The surface electromyography (EMG) signal at each channel location was continuously sampled to obtain the multi-channel raw EMG sequence. While acquiring surface electromyography signals, motion state parameters are recorded, including movement rhythm parameters, joint displacement parameters, movement repetition count parameters, and movement load parameters. The raw multi-channel electromyography sequences and motion state parameters are processed with unified time stamping to generate a timestamp sequence; The multi-channel electromyography (EMG) raw sequences and motion state parameters are aligned based on the timestamp sequence to form EMG motion temporal data arranged in chronological order.

[0007] Optionally, the signal processing module specifically comprises: The surface electromyography (EMG) signals in the EMG motion time series data were subjected to power frequency suppression, bandpass filtering, motion artifact removal, baseline drift correction, rectification, and envelope smoothing to obtain a preprocessed EMG sequence. Time sequence analysis is performed on the motion state parameters, including the action beat parameters, joint displacement parameters, and action repetition number parameters, to extract the start and end times of the action and determine the start and end positions of the action cycle. The preprocessed electromyographic sequence was periodically truncated according to the start and end positions of the action cycle to obtain multiple action cycle segments; The length of multiple motion cycle segments is normalized, mapping the multiple motion cycle segments to a uniform sampling length; Multiple motion cycle segments with normalized lengths are arranged in chronological order to generate periodic electromyographic sequences.

[0008] Optionally, the step of acquiring the periodic electromyography sequences of the test subject in a non-fatigue state to construct an individual baseline sample set specifically involves: When the subject is in a non-fatigue state, surface electromyography (EMG) signals from multiple muscle groups are collected through wireless dry electrode clothing. Simultaneously, action beat parameters, joint displacement parameters, action repetition parameters, and motion load parameters are collected to generate raw EMG sequences and corresponding motion state parameters. The original electromyography (EMG) sequences were subjected to power frequency suppression, bandpass filtering, motion artifact removal, baseline drift correction, rectification, and envelope smoothing to obtain preprocessed EMG sequences. Based on the action beat parameters, joint displacement parameters, and action repetition parameters, the start and end positions of the action cycle are determined. The preprocessed electromyography sequence is then segmented into cycles and aligned in length to generate a periodic electromyography sequence. Multiple consecutive motion cycle segments are selected from the periodic electromyography sequence and arranged according to the channel number sequence and motion cycle order to form a baseline cycle set; The amplitude of the motion cycle segments in the baseline cycle set is made uniform and stored according to the muscle group channel and motion cycle number to construct an individual baseline sample set.

[0009] Optionally, the step of extracting feature parameters from the individual baseline sample set, encoding the periodic electromyography sequence and the feature parameters, and generating a temporal embedding sequence specifically involves: Baseline cycle sets are read from the individual baseline sample set, and electromyographic amplitude parameters, cycle position parameters, movement phase parameters, channel position parameters, and exercise load parameters are extracted according to muscle group channel and movement cycle number; The periodic electromyography (EMG) sequence is unfolded into a time step sequence according to the action cycle. The EMG amplitude parameters are combined with the corresponding time step's cycle position parameters, action phase parameters, channel position parameters, and motion load parameters to form a time step feature group. The parameters in the time step feature group are numerically mapped to obtain the time step encoding vector; the time step encoding vectors are arranged in the order of the time steps to form a time sequence encoding matrix. The temporal coding matrix is ​​subjected to embedding transformation to generate a temporal embedding sequence.

[0010] Optionally, the improved DeepAR model specifically includes an input encoding layer, a temporal recursive layer, a probability output layer, and an autoregressive feedback layer; The input encoding layer expands the temporal embedding sequence into a time step sequence according to the action cycle. At time step t, it reads the corresponding electromyographic amplitude parameters, cycle position parameters, action phase parameters, channel position parameters, and motion load parameters, and concatenates them into the input vector at time step t according to a preset arrangement. The temporal recursive layer establishes an initial hidden state at the first time step, reads the input vector at the t-th time step and the hidden state at the (t-1)-th time step to obtain the hidden state at the t-th time step, and introduces a phase sequence drift clustering mechanism in the temporal recursive layer. The phase sequence drift clustering mechanism divides the sampling point sequence of each action cycle into an initiation phase, a peak holding phase, and a release phase according to the rising, peak, and falling segments of the electromyographic amplitude in the periodic electromyographic sequence. The starting sampling position, ending sampling position, phase length, and peak sampling position are extracted for the initiation phase, peak holding phase, and release phase, respectively, and arranged according to the action cycle sequence to form a phase sequence cluster chain. Establish a correspondence between the same phase segment types for the phase sequence cluster chains in the k-th and k+1-th operation cycles. Calculate the center position difference between the initiating phase segment, peak holding phase segment, and release phase segment to obtain the cluster center displacement. Calculate the length difference between the initiating phase segment, peak holding phase segment, and release phase segment to obtain the phase segment length change. Calculate the difference between the peak sampling positions in the k-th and k+1-th operation cycles to obtain the peak position offset. Based on the change in the arrangement order of the initiating phase segment, peak holding phase segment, and release phase segment in the k-th and k+1-th operation cycles, obtain the phase sequence arrangement change. Combine the cluster center displacement, phase segment length change, peak position offset, and phase sequence arrangement change to form the drift parameter. At time step t, the drift parameter is concatenated with the electromyographic amplitude parameter, periodic position parameter and action phase parameter in the input vector at time step t to obtain the extended input vector. The extended input vector and the hidden state at time step t-1 are written into the state update unit to obtain the hidden state at time step t containing the drift information. The probability output layer performs a linear mapping on the hidden state at time step t to obtain the distribution mean parameter and distribution scale parameter, and generates the electromyographic probability sequence and reference distribution parameter corresponding to time step t based on the distribution mean parameter and distribution scale parameter; The autoregressive feedback layer extracts sampled values ​​from the electromyographic probability sequence at time step t, and then concatenates the sampled values ​​with the input vector at time step t+1 before inputting them into the temporal recursive layer.

[0011] Optionally, the step of constructing a high-dimensional feature point set based on periodic electromyography sequences and reconstructing the phase space to generate a periodic evolution point cloud specifically involves: Multiple action cycle segments in the periodic electromyography sequence are read in the order of action cycle, and the sampled value sequence in each action cycle segment is extracted according to the muscle group channel number. In each action cycle segment, the sampled value sequence is sampled at a preset delay interval, and the sampled values ​​at different delay positions are arranged into a state vector according to a preset embedding dimension. Multiple state vectors in the same action cycle segment are combined according to the sampling time order to form a single-cycle state point set; The single-cycle state point sets corresponding to multiple action cycle segments are collected according to the action cycle order and muscle group channel number to form a high-dimensional feature point set; Each state vector in the high-dimensional feature point set is assigned a coordinate position according to the time sequence to generate discrete state points in phase space; The discrete state points in phase space are arranged in the order of action cycles to form a periodic evolution point cloud.

[0012] Optionally, the step of using the Persistent Homology algorithm to extract the topological structure information of the periodically evolving point cloud and generate a topological barcode sequence specifically involves: A distance matrix is ​​constructed based on the distance between points in the periodically evolving point cloud, and a filter radius sequence is established based on different distance values ​​in the distance matrix; Under each filtering radius, discrete state points whose distance between points is no greater than the corresponding filtering radius are connected to form a simplex, thus forming a simplex complex under the corresponding filtering radius; Multiple simple complexes are arranged sequentially according to the filter radius sequence to form a filter complex sequence; Extract the generation and vanishing radii of connected components from the filtered complex sequence, and extract the generation and vanishing radii of loop structures; Based on the generation and disappearance radii of connected components and the generation and disappearance radii of loop structures, persistent intervals for connected components and persistent intervals for loop structures are established respectively. The topological barcode sequence is generated by arranging the persistent intervals of connected components and persistent intervals of ring structures according to their start and end positions.

[0013] Optionally, the step of calculating probability offset and topological offset based on electromyographic probability sequence, reference distribution parameters, and topological barcode sequence, and generating local fatigue index by combining drift parameters, and performing cumulative analysis on local fatigue indexes across multiple action cycles to generate a fatigue score, specifically: Read the electromyographic probability sequence and reference distribution parameters in the order of the action cycle, and generate a probability offset based on the difference between the electromyographic probability sequence and the reference distribution parameters; Read the topology barcode sequence according to the action cycle order, and generate the topology offset based on the difference between the persistent interval of the connected component and the persistent interval of the ring structure between adjacent action cycles. The drift parameters are read sequentially according to the action cycle, and the probability offset, topology offset and drift parameters are combined to form a local fatigue feature group; Local fatigue characteristic groups are accumulated in sequence according to the action cycle to generate local fatigue index; local fatigue indexes from multiple action cycles are accumulated to generate fatigue score.

[0014] Optionally, if the fatigue score reaches a first preset threshold, a training restriction command is output; if the fatigue score exceeds a second preset threshold, a termination command is output. Specifically: Set a first preset threshold and a second preset threshold, wherein the second preset threshold is greater than the first preset threshold; If the fatigue score reaches the first preset threshold and does not exceed the second preset threshold, a training restriction instruction is generated. If the fatigue score exceeds the second preset threshold, a termination command will be generated.

[0015] The beneficial effects of this invention are: A phase-order drift clustering mechanism is introduced into the temporal modeling module. By dividing the periodic electromyography sequence into phase segments and constructing a phase-order cluster chain, drift parameters are generated by combining the cluster center displacement, phase segment length change, peak position offset, and phase order arrangement change between adjacent action cycles. This enables the improved DeepAR model to simultaneously characterize amplitude changes and periodic structure changes during the hidden state update process, thereby improving the ability to depict the phase-order change characteristics during muscle fatigue. In the topology analysis module, a periodic evolution point cloud is constructed based on the periodic electromyography sequence, and the topology barcode sequence is extracted by the PersistentHomology algorithm. The structural change characteristics in the periodic evolution process are described by the persistent intervals of connected components and the persistent intervals of loop structures, so as to realize the multi-scale representation of the global morphology of electromyography signals and improve the ability to identify the fatigue evolution trend under complex motion conditions. The fatigue control module combines electromyographic probability sequences, reference distribution parameters, topological barcode sequences, and drift parameters to generate local fatigue indicators. It also generates fatigue scores through cumulative analysis of multiple action cycles. Combined with the first and second preset thresholds, it outputs training restriction and termination commands to achieve continuous quantitative assessment and graded control of muscle fatigue, thereby improving the accuracy and controllability of fatigue assessment results. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the muscle fatigue assessment system based on wireless dry electrode signals proposed in this invention. Figure 2 This is a schematic diagram of the improved DeepAR model proposed in this invention; Figure 3 This is a data flow diagram of the muscle fatigue assessment system based on wireless dry electrode signals proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A muscle fatigue assessment system based on wireless dry electrode signals includes: The electromyography (EMG) acquisition module is used to acquire surface EMG signals from multiple muscle groups through a wireless dry electrode garment worn on the body surface of the test subject, and simultaneously acquire the motion state parameters of the test subject to construct EMG motion time series data. The signal processing module is used to perform signal conditioning on the surface electromyography (EMG) signals in the EMG motion time series data, determine the start and end positions of the action cycle based on the motion state parameters, perform periodic segmentation and length alignment on the surface EMG signals, and generate periodic EMG sequences. The baseline construction module is used to acquire the periodic electromyography (EMG) sequences of the test subjects in a non-fatigue state, construct an individual baseline sample set, extract feature parameters from the individual baseline sample set, encode the periodic EMG sequences and feature parameters, and generate a temporal embedding sequence. The temporal modeling module is used to improve the DeepAR model by embedding temporal sequences into the input sequence. It introduces a phase drift clustering mechanism in the temporal recursive layer, divides the periodic electromyography sequence into phase segments and maps them into phase cluster chains, performs drift calculation on the phase cluster chains of adjacent action cycles, and outputs the electromyography probability sequence and reference distribution parameters. The topology analysis module is used to construct a high-dimensional feature point set based on the periodic electromyography sequence and reconstruct the phase space to generate a periodic evolution point cloud. The Persistent Homology algorithm is used to extract the topological structure information of the periodic evolution point cloud and generate a topological barcode sequence. The fatigue control module is used to calculate probability offset and topological offset based on electromyographic probability sequence, reference distribution parameters and topological barcode sequence, and generate local fatigue index by combining drift parameters. It performs cumulative analysis on local fatigue index of multiple action cycles to generate fatigue score. If the fatigue score reaches the first preset threshold, it outputs training restriction instruction. If the fatigue score exceeds the second preset threshold, it outputs termination instruction.

[0019] In this embodiment, the electromyography (EMG) acquisition module specifically comprises: The wireless dry electrode garment is attached to the corresponding muscle group on the body surface of the test subject. The position of the electrode attachment area in the wireless dry electrode garment is marked according to the anatomical distribution of the target muscle group on the body surface. Each electrode attachment area corresponds to a collection position, and a preset channel distribution is formed according to the spatial arrangement order of the electrode attachment areas on the body surface. The target muscle group is divided into channels according to the preset channel distribution, and a channel numbering sequence is established based on the spatial arrangement of the electrode attachment area. The surface electromyography (EMG) signals at each channel position are read synchronously at a uniform sampling frequency. The signal values ​​of multiple channels at the same sampling time are arranged according to the channel number sequence to form a multi-channel sampling group, and then arranged according to the sampling time order to form a multi-channel EMG raw sequence. While reading surface electromyography signals, parameters of movement rhythm, joint displacement, number of repetitions of movement, and motion load are collected and a parameter sequence is formed according to the order of recording time. Each sampling time marker is assigned to each sampling group in the multi-channel electromyography raw sequence, and a recording time marker is assigned to the action beat parameter, joint displacement parameter, action repetition number parameter and motion load parameter. The sampling time marker and the recording time marker are uniformly mapped to the same time base to form a timestamp sequence. The multi-channel electromyography (EMG) raw sequences and parameter sequences are aligned according to the timestamp sequence. Sample groups with consistent times are arranged to correspond with parameter records. Parameter records with times between adjacent sampling moments are inserted into their corresponding time positions and reordered to form EMG motion time series data arranged in chronological order.

[0020] In this embodiment, the signal processing module specifically comprises: Surface electromyography (EMG) signals are extracted from EMG motion time-series data according to channel number sequence, and the signal values ​​corresponding to the same channel at consecutive sampling times are arranged in chronological order to form channel EMG sequences. The power frequency component in the channel electromyography sequence is suppressed, and the periodic interference component corresponding to the power supply frequency is removed from the channel electromyography sequence; The channel electromyography (EMG) sequence after removing the power frequency component retains the signal components within the effective frequency band of EMG. The slowly changing components below the lower limit of the effective frequency band of EMG and the high-frequency perturbation components above the upper limit of the effective frequency band of EMG are separated from the channel EMG sequence to form a band-limited EMG sequence. The amplitude abrupt change segment and short-term abnormal fluctuation segment in the band-limited electromyography sequence are identified, and the signal values ​​corresponding to the amplitude abrupt change segment and short-term abnormal fluctuation segment are replaced with the continuous transition values ​​of the adjacent normal segment to form an artifact correction sequence. The low-frequency drift component in the artifact correction sequence is separated, and the low-frequency drift component is subtracted from the artifact correction sequence to form the baseline correction sequence; The negative signal values ​​in the baseline correction sequence are converted into their corresponding absolute values ​​to form a rectified sequence; The signal values ​​of the rectified sequence within a preset time window are averaged to form an envelope sequence, which is then used as a preprocessed electromyographic sequence. The motion rhythm parameters, joint displacement parameters, and motion repetition count parameters are extracted from the motion state parameters and arranged in the order of recording time to form the motion state time sequence. In the motion state timing, the moment when the action beat parameter changes from a static state to an increasing state is marked as the start time of the action, and the moment when the action beat parameter changes from a decreasing state to a stable state is marked as the end time of the action. The position of the initial change of displacement in the joint displacement parameters is matched with the start time of the movement, and the position of the return of displacement in the joint displacement parameters is matched with the end time of the movement. Position corrections are made for the start time and end time of the movement. The time interval between two consecutive count changes in the action repetition parameter is defined as a complete action cycle, and the start and end positions of the action cycle are determined based on the start and end times of the action. Signal values ​​within the corresponding time intervals are extracted from the preprocessed electromyography sequence according to the start and end positions of the action cycle to form action cycle segments; The number of sampling points in the action cycle segments is counted. Action cycle segments with different numbers of sampling points are resampled according to a unified sampling length so that each action cycle segment corresponds to the same number of sampling points. The action cycle segments with unified sampling length are arranged in the order of the action cycles to form a periodic electromyography sequence.

[0021] In this embodiment, periodic electromyography sequences of the test subject in a non-fatigue state are acquired to construct an individual baseline sample set, specifically as follows: When the subject completes the preset action preparation and keeps the motion load parameters in a stable load range, the surface electromyography signals, action beat parameters, joint displacement parameters, action repetition number parameters and motion load parameters of multiple muscle groups are collected synchronously through the wireless dry electrode clothing to form the original electromyography sequence and corresponding motion state parameters. The preprocessed electromyography (EMG) sequence is formed by removing periodic interference components corresponding to the power supply frequency from the original EMG sequence and retaining signal components with frequencies in the range of 20Hz to 450Hz. Sampling segments with amplitude changes exceeding twice the average amplitude of adjacent sampled values ​​are replaced. Abnormal sampled values ​​are replaced with linear transition values ​​of adjacent stable sampled values. Signal components with frequencies below 10Hz are separated from the original EMG sequence and subtracted from the signal sequence. Negative values ​​in the signal sequence are converted into their corresponding absolute values. The continuous sampled values ​​are then averaged according to a time window of 50ms to 200ms to form a preprocessed EMG sequence. The start and end times of the action are determined by the rising and falling positions of the action beat parameters, and the start and end times of the action are corrected by combining the displacement range of the joint displacement parameters. The start and end positions of the action cycle are determined by the time interval between two adjacent count changes in the action repetition parameter. According to the start and end positions of the action cycle, the signal values ​​of the corresponding time intervals are extracted from the preprocessed electromyography sequence. The number of sampling points in different action cycle segments is uniformly resampled so that each action cycle segment corresponds to the same number of sampling points, thus forming a periodic electromyography sequence. Multiple consecutive action cycle segments are selected from the periodic electromyography sequence. The action cycle segments corresponding to the same muscle group channel are arranged in the action cycle order, and the action cycle segments corresponding to different muscle group channels are arranged in the channel number sequence to form a baseline cycle set. Extract the maximum amplitude value for each motion cycle segment in the baseline cycle set, and divide each sampled value of the motion cycle segment by the corresponding maximum amplitude value to obtain the motion cycle segment with consistent amplitude. The amplitude-homogenized motion cycle segments are stored according to the muscle group channel number and the motion cycle number, and multiple consecutive motion cycle segments are combined to form an individual baseline sample set.

[0022] In this embodiment, feature parameters are extracted from the individual baseline sample set, and the periodic electromyography sequence and the feature parameters are encoded to generate a temporal embedding sequence, specifically: Baseline cycle sets are read from the individual baseline sample set. Sample value sequences are extracted from the corresponding action cycle segments according to the muscle group channel number and action cycle number. Electromyographic amplitude parameters corresponding to each sampling position are extracted from the sample value sequences. Each sampling position is assigned a sequential number according to the total number of sampling points in the motion cycle segment, and the ratio of the sequential number to the total number of sampling points in the motion cycle segment is used as the cycle position parameter. Based on the rising segment, peak segment, and falling segment of the sampled value in the action cycle segment, a corresponding action stage label is assigned to each sampling position, and the action stage label is used as the action stage parameter. Based on the arrangement of the muscle group channel number in the channel number sequence, each muscle group channel is assigned a corresponding channel position mark, and the channel position mark is used as the channel position parameter; Read the load level value from the motion load parameters corresponding to the motion cycle number, and map the load level value to each sampling position in the motion cycle segment to form the motion load parameters; The periodic electromyography sequence is unfolded into a time step sequence according to the action cycle, and the corresponding electromyography amplitude parameters, cycle position parameters, action phase parameters, channel position parameters and motion load parameters are read at each time step. The electromyographic amplitude parameters in each time step are retained as continuous values, the periodic position parameters are retained as continuous values, the action phase parameters are converted into phase numbers, the channel position parameters are converted into position numbers, and the motion load parameters are converted into load numbers. The electromyographic amplitude parameters, cycle position parameters, action phase number, channel position number and load number in the same time step are placed in a fixed order to form a time step feature group. The continuous values ​​in the time step feature group are scaled to a uniform value range. The stage number, position number and load number are mapped to the corresponding value sequence. The scaled continuous values ​​and the mapped value sequence are then connected in a fixed order to form the time step encoding vector. Arrange the multiple time step encoding vectors row by row according to the time step order to form a time sequence encoding matrix; Each row of the encoding vector in the temporal encoding matrix is ​​multiplied with a preset embedding parameter matrix. The embedding parameter matrix is ​​a parameter matrix set during the model initialization phase. Its elements are generated according to a uniform or normal distribution and remain fixed during model operation or are adjusted as the model parameters are updated. The product result is used as the embedding vector for the corresponding time step. Multiple embedding vectors are arranged sequentially according to the time step order to form a temporal embedding sequence.

[0023] In this embodiment, the improved DeepAR model specifically includes an input encoding layer, a temporal recursive layer, a probability output layer, and an autoregressive feedback layer; The input encoding layer expands the temporal embedding sequence into a time step sequence according to the action cycle order, and reads the corresponding electromyographic amplitude parameters, cycle position parameters, action phase parameters, channel position parameters and motion load parameters at time step t. The electromyographic amplitude parameters, cycle position parameters, action phase parameters, channel position parameters and motion load parameters are arranged from left to right in a fixed order to form the input vector at time step t. The temporal recursive layer establishes an initial hidden state at the first time step. Each state value in the initial hidden state is set to zero or to the initial parameter value given during the model initialization phase. At time step t, the input vector at time step t and the hidden state at time step (t-1) are read. The input vector at time step t and the hidden state at time step (t-1) are input into the state update unit respectively. The input gate, forget gate and output gate in the state update unit read the values ​​at the corresponding positions of the input vector at time step t and the hidden state at time step (t-1) respectively, and form the gating result through the gating parameter matrix and the bias parameter vector. The gating result and the candidate state together generate the hidden state at time step t. Based on the amplitude slope of the electromyographic amplitude changing with the sampling position in the periodic electromyographic sequence, the continuously rising sampling segment is determined as the initiation phase segment, the sampling segment with amplitude change less than the preset change threshold near the peak is determined as the peak holding phase segment, and the continuously falling sampling segment is determined as the release phase segment. In each action cycle, the starting sampling position, ending sampling position, phase length, and peak sampling position of the initiation phase segment, peak holding phase segment, and release phase segment are recorded respectively. The oscillation phase segment, peak holding phase segment, and release phase segment in the same action cycle are arranged in the order of sampling position, and the arranged phase segment sequence is recorded as the phase sequence cluster chain; Establish a one-to-one correspondence between the phase sequence cluster chain of the kth action cycle and the phase sequence cluster chain of the (k+1)th action cycle according to the same phase segment type. Read the starting sampling position and the ending sampling position of the oscillation phase segment, the peak holding phase segment, and the release phase segment respectively. Divide the sum of the starting sampling position and the ending sampling position by two to obtain the phase segment center position. Subtract the phase segment center position of the kth action cycle from the phase segment center position of the (k+1)th action cycle to obtain the cluster center displacement. Read the phase segment lengths of the oscillation phase segment, peak holding phase segment, and release phase segment in the k-th and k+1-th action cycles respectively, and subtract the phase segment length of the k-th action cycle from the phase segment length of the k+1-th action cycle to obtain the change in phase segment length. Read the peak sampling positions in the k-th action cycle and the (k+1)-th action cycle respectively, and subtract the peak sampling position of the k-th action cycle from the peak sampling position of the (k+1)-th action cycle to obtain the peak position offset; Read the order of the oscillation phase segment, peak holding phase segment, and release phase segment in the k-th and k+1-th action cycles respectively, and subtract the order numbers of different phase segment types in the two action cycles to obtain the change in phase sequence arrangement. The cluster center displacement, phase segment length change, peak position offset, and phase sequence arrangement change are connected sequentially in a fixed order to form drift parameters; At time step t, the cluster center displacement, phase segment length change, peak position offset, and phase sequence arrangement change in the drift parameters are read. The cluster center displacement, phase segment length change, peak position offset, and phase sequence arrangement change are then placed sequentially after the electromyographic amplitude parameter, period position parameter, and action phase parameter in the input vector at time step t to form an extended input vector. The extended input vector and the hidden state at time step t-1 are input into the state update unit. The values ​​in the extended input vector are multiplied by the corresponding elements of the gating parameter matrix one by one and summed according to position. Then, the values ​​in the extended input vector and the hidden state at time step t-1 are used together to generate the hidden state at time step t containing drift information. The probability output layer multiplies each state value in the hidden state at time step t with the corresponding element of the output parameter matrix and sums them by position. Then, it adds the mean bias parameter and the value scale bias parameter to obtain the distribution mean parameter and the distribution scale parameter. The distribution mean parameter is used as the center value of the electromyography probability sequence at time step t, and the distribution scale parameter is used as the dispersion parameter of the electromyography probability sequence at time step t. The reference distribution parameter is determined based on the distribution mean parameter and the distribution scale parameter. The autoregressive feedback layer reads sampled values ​​from the electromyography (EMG) probability sequence at time step t, places the sampled values ​​at the end of the input vector at time step t+1, and arranges the sampled values ​​in sequence with the EMG amplitude parameters, periodic position parameters, action phase parameters, channel position parameters, and motion load parameters at time step t+1 to form a feedback input sequence. The feedback input sequence is then input into the temporal recursive layer to obtain the hidden state at time step t+1, and is repeatedly updated according to the time step order to form the EMG probability sequence and reference distribution parameters corresponding to the periodic EMG sequence.

[0024] In this embodiment, both the improved DeepAR model and the DeepAR model adopt an autoregressive probabilistic modeling structure based on time series. Both convert the time-series embedded sequence into time-step input vectors through an input encoding layer. Both generate time-step hidden states by combining the current time-step input vector with the hidden state of the previous time step in a time-series recursive layer. They output the distribution mean parameter and distribution scale parameter through a probability output layer to characterize the probability distribution characteristics of the sequence. Simultaneously, both use an autoregressive feedback layer to introduce the sampled value of the current time step into subsequent time steps to form a recursive input. The improved DeepAR... The model introduces a phase-order drift clustering mechanism in the temporal recursive layer, dividing the periodic electromyography sequence into initiation phases, peak-holding phases, and release phases according to the amplitude variation trend. Phase-order cluster chains are constructed using the initiation and end sampling positions of each phase segment, the phase segment length, and the peak sampling position, expanding the periodic structure of the time series from a single time dimension to a phase-order structure. A correspondence between phase-order cluster chains is established between adjacent action cycles. Cluster center displacement is generated by the difference in phase segment center positions, phase segment length change is generated by the difference in phase segment lengths, and peak sampling position change is generated by the difference in peak sampling positions. The peak position offset is calculated, and the phase sequence arrangement change is generated by changing the phase segment arrangement order. This multi-dimensional structural change is combined to form a drift parameter, which is then introduced into the time step input vector. This allows the hidden state update process to simultaneously include amplitude change information and periodic structural evolution information. During the state update process, the drift parameter is aligned with the electromyographic amplitude parameter, periodic position parameter, and action phase parameter, and each parameter participates in the gating calculation, enabling the hidden state to respond to periodic structural changes. In the probability output process, the distribution mean parameter and distribution scale parameter are generated based on the hidden state containing phase sequence drift information, so that the probability distribution reflects not only amplitude fluctuation characteristics but also periodic structural change trends. The improved DeepAR model enhances the ability to represent phase changes, structural shifts, and rhythmic instability features in periodic electromyographic sequences by introducing a phase sequence drift clustering mechanism. This allows the model to maintain structural consistency when dealing with morphological changes between different action cycles, improves the sensitivity to subtle changes in electromyographic signals under fatigue conditions, and enhances the stability and continuity of the probability distribution output. It also reduces prediction bias caused by periodic misalignment, thereby improving the accuracy and robustness of fatigue assessment.

[0025] In this embodiment, a high-dimensional feature point set is constructed based on the periodic electromyography sequence, and the phase space is reconstructed to generate a periodic evolution point cloud, specifically: Multiple action cycle segments in the periodic electromyography sequence are read in the order of action cycle, and the sampled value sequence in each action cycle segment is extracted according to the muscle group channel number. The sampled value sequences corresponding to the same muscle group channel are arranged in the order of action cycle. In each sample value sequence, a delay interval and an embedding dimension are set, wherein the delay interval is an integer value between 1 and 5 sample points, and the embedding dimension is an integer value between 2 and 6 dimensions. Starting from the i-th sampling position in the sampled value sequence, read the i-th sampled value, the sampled value corresponding to the i-th plus 1 delay interval, the sampled value corresponding to the i-th plus 2 delay interval, until the sampled value corresponding to the i-th plus embedding dimension minus 1 delay interval, and arrange the read sampled values ​​from left to right according to the reading order to form a state vector; The starting position is moved point by point along the sampling position of the sampling value sequence, and the sampling values ​​at different delay positions are read repeatedly to form multiple state vectors corresponding to the same action cycle segment; Arrange multiple state vectors in the same action cycle segment according to the order of their initial sampling positions to form a single-cycle state point set; The single-cycle state point sets corresponding to multiple action cycle segments are arranged sequentially according to the action cycle order, and the single-cycle state point sets corresponding to different muscle group channels are arranged sequentially according to the muscle group channel number order to form a high-dimensional feature point set. Each state vector in the high-dimensional feature point set is assigned a time position marker and an action cycle position marker, where the time position marker corresponds to the starting sampling position of the state vector and the action cycle position marker corresponds to the action cycle number to which the state vector belongs. The first-dimensional sampled value, the second-dimensional sampled value, and so on up to the embedded dimension sampled value in the state vector are respectively used as coordinate components on the phase space coordinate axis. The time position mark and the action cycle position mark are used as additional positioning information of the state vector in the phase space to form discrete state points in the phase space. Discrete state points within the same action cycle are arranged in ascending order of time position markers, and discrete state points corresponding to different action cycles are arranged in ascending order of action cycle position markers, forming a periodic evolution point cloud.

[0026] In this embodiment, the Persistent Homology algorithm is used to extract the topological structure information of the periodically evolving point cloud and generate a topological barcode sequence, specifically: According to the coordinate position of discrete state points in phase space in the periodic evolution point cloud, the difference of coordinate components of any two discrete state points on each coordinate axis is read point by point. The difference of coordinate components on each coordinate axis is squared and summed. Then the square root of the sum is taken to obtain the distance between corresponding point pairs. Fill in the distances between all point pairs into the corresponding row and column positions according to the order of the discrete state points to form a distance matrix; Read the distance values ​​between different points from the distance matrix, arrange them in ascending order, and use the increasing distance between adjacent distance values ​​as the filtering step size to form a filtering radius sequence; For each filter radius in the filter radius sequence, read the distance between points in the distance matrix that is not greater than that filter radius, and establish a connection edge between the corresponding two discrete state points; For the case where the distance between any two pairs of discrete state points is no greater than the same filtering radius, the closed region enclosed by the three connecting edges is filled with a two-dimensional simplex. For cases where the distance between any two points among the four discrete state points is no greater than the same filtering radius, the spatial region enclosed by the four discrete state points is filled with a three-dimensional simplex. By combining the connecting edges, two-dimensional simplex and three-dimensional simplex formed under the same filtering radius according to the connection relationship of discrete state points, a simplex complex under the corresponding filtering radius is formed. Arrange the simplexes corresponding to different filter radii in the filter radius sequence in ascending order of filter radius to form a filter complex sequence; In the filtered complex sequence, the filtering radius corresponding to the first appearance of a connected component is recorded as the generation radius of the connected component, and the filtering radius corresponding to the merged connected component when two connected components are merged into one connected component is recorded as the disappearance radius of the connected component. In the filtered complex sequence, the filtering radius corresponding to the first formation of the closed loop structure is recorded as the generation radius of the loop structure, and the filtering radius corresponding to the inside of the closed loop structure being filled with a two-dimensional simplex and no longer maintaining a closed cavity is recorded as the disappearance radius of the loop structure. Arrange the generation radius and disappearance radius of each connected component in a corresponding order to form a persistent interval of the connected component; arrange the generation radius and disappearance radius of each loop structure in a corresponding order to form a persistent interval of the loop structure. Arrange the persistent intervals of connected components and persistent intervals of ring structures in ascending order of generation radius, and assign the starting position of each persistent interval to the generation radius and the ending position to the disappearance radius to form a topological barcode sequence.

[0027] In this embodiment, probability offset and topological offset are calculated based on electromyographic probability sequences, reference distribution parameters, and topological barcode sequences. These are then combined with drift parameters to generate local fatigue indices. Accumulated analysis of local fatigue indices across multiple motion cycles is performed to generate a fatigue score. Specifically: Read the sampled values ​​in the electromyographic probability sequence according to the action cycle, and read the mean and scale parameters of the distribution in the reference distribution parameters corresponding to the action cycle; Within the same action cycle, the sampled value corresponding to each sampling position is subtracted from the distribution mean parameter of the corresponding sampling position to obtain the amplitude difference sequence. Each difference in the amplitude difference sequence is divided by the distribution scale parameter of the corresponding sampling position to obtain the normalized difference sequence. The absolute values ​​in the normalized difference sequence within the same action cycle are added in order of sampling position, and the sum is divided by the total number of sampling positions within the action cycle to generate the probability offset of the corresponding action cycle. Read the persistent intervals of connected components and persistent intervals of ring structures in the topology barcode sequence according to the action cycle, and read the start position, end position, start position, and end position of the persistent interval of connected components, the persistent interval of ring structures, and the persistent interval of ring structures corresponding to two adjacent action cycles. The difference between the start positions of the connected component and the start position of the persistent interval of the connected component in the previous action cycle is obtained by subtracting the start position of the persistent interval of the connected component in the next action cycle from the start position of the persistent interval of the connected component in the previous action cycle. The difference between the end positions of the connected component and the end positions of the persistent interval of the connected component in the next action cycle is obtained by subtracting the end position of the persistent interval of the connected component in the previous action cycle. Subtract the starting position of the persistent interval of the loop structure in the previous action cycle from the starting position of the persistent interval of the loop structure in the next action cycle to obtain the difference in the starting position of the loop structure. Subtract the ending position of the persistent interval of the loop structure in the previous action cycle from the ending position of the persistent interval of the loop structure in the next action cycle to obtain the difference in the ending position of the loop structure. The absolute values ​​of the differences in the starting and ending positions of connected components, the differences in the starting and ending positions of the loop structure, and the differences in the ending positions of the loop structure are summed to generate the topology offset for the corresponding action cycle. Read the cluster center displacement, phase segment length change, peak position offset, and phase sequence arrangement change in the drift parameters according to the action cycle sequence, and arrange the probability offset, topological offset, cluster center displacement, phase segment length change, peak position offset, and phase sequence arrangement change from left to right in a fixed arrangement order to form a local fatigue feature group. Subtract the mean of the corresponding feature in the individual baseline sample set from each value in the local fatigue feature set, and divide the result of the subtraction by the standard deviation of the corresponding feature in the individual baseline sample set to obtain the normalized local fatigue feature set. The absolute values ​​of all values ​​in the normalized local fatigue feature group are summed, and the sum is divided by the number of features in the local fatigue feature group to generate the local fatigue index corresponding to the action cycle. Read the local fatigue indicators corresponding to multiple consecutive action cycles in the order of action cycles, add the local fatigue indicators corresponding to multiple consecutive action cycles in sequence to obtain the cumulative fatigue value, and divide the cumulative fatigue value by the number of action cycles to generate a fatigue score.

[0028] In this embodiment, if the fatigue score reaches a first preset threshold, a training restriction command is output; if the fatigue score exceeds a second preset threshold, a termination command is output. Specifically: A first preset threshold and a second preset threshold are set, with the second preset threshold being greater than the first preset threshold. The first and second preset thresholds are determined based on the fatigue score distribution corresponding to the individual baseline sample set, including: collecting fatigue scores corresponding to multiple action cycles in a non-fatigue state to form a baseline fatigue score sequence; statistically analyzing the values ​​in the baseline fatigue score sequence to obtain the mean and standard deviation; adding one standard deviation to the mean as the first preset threshold, and adding two standard deviations to the mean as the second preset threshold; the first preset threshold is used to mark the initial stage of fatigue, and the second preset threshold is used to mark the stage of fatigue aggravation. If the fatigue score reaches the first preset threshold but does not exceed the second preset threshold, a training restriction instruction is generated; if the fatigue score exceeds the second preset threshold, a termination instruction is generated.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a combined upper limb flexion-extension and lower limb extension training scenario at a rehabilitation training center. The subjects were 24 adults aged 22 to 41 years who had long participated in strength training and sports rehabilitation training, with 12 men and 12 women. During training, the subjects wore a data acquisition device with wireless dry electrodes attached to the corresponding body surface locations of the vastus lateralis, rectus femoris, biceps brachii, and triceps brachii muscles. The sampling frequency was set to 1000 Hz, the movement rhythm was set to 18 repetitions per minute, the training duration was set to 6 minutes, and the exercise load parameters were set to four levels: self-load, 5 kg, 10 kg, and 15 kg. Before training began, surface electromyography (EMG) signals were collected for 10 consecutive movement cycles under resting and low-load conditions to form an individual baseline sample set. After training began, surface EMG signals, movement rhythm parameters, joint displacement parameters, movement repetition count parameters, and exercise load parameters were collected simultaneously to construct EMG temporal data. Subsequently, a periodic electromyography (EMG) sequence is obtained through a signal processing module, and a temporal embedding sequence is generated through a baseline construction module. This sequence is then input into an improved DeepAR model, and drift parameters between the initiation phase, peak holding phase, and release phase are extracted using a phase sequence drift clustering mechanism. Simultaneously, a periodic evolution point cloud is constructed from the periodic EMG sequence, and a topological barcode sequence is generated using the Persistent Homology algorithm. The fatigue control module outputs a fatigue score based on the EMG probability sequence, reference distribution parameters, topological barcode sequence, and drift parameters. A training limit instruction is output when the fatigue score reaches 0.68, and a termination instruction is output when the fatigue score exceeds 0.85. Field applications have shown that traditional evaluation methods relying solely on root mean square values ​​or median frequencies are prone to misjudging movement imbalance as increased fatigue when there are fluctuations in movement rhythm and instability in joint displacement. This invention, however, can simultaneously combine probability offset, topological offset, and drift parameters to continuously analyze multiple movement cycles, thereby distinguishing between true fatigue and movement rhythm disturbances.

[0030] Table 1 shows the comparison of the fatigue assessment results of the present invention and comparative method A under different load conditions, where comparative method A uses the traditional time-domain and frequency-domain feature joint threshold determination method. The sample size is 48.

[0031] Table 1 Comparison of Fatigue Assessment Results under Different Load Conditions

[0032] As shown in Table 1, under four load conditions, the fatigue recognition accuracy of this invention consistently remained between 94.6% and 96.7%, significantly higher than the 86.9% to 89.4% of the comparative method A. Particularly under 10kg and 15kg load conditions, the changes in motion amplitude and rhythm drift were more pronounced. The false alarm rates of the traditional method reached 8.9% and 11.2%, respectively, while the false alarm rates of this invention were still controlled at 3.1% and 3.5%, indicating that the phase sequence drift clustering mechanism can better identify structural changes within the motion cycle. Regarding the fluctuation amplitude of fatigue scores, this invention consistently remained below 0.11, indicating continuous and stable output results, and is less prone to score jumps due to instantaneous noise. The average response time only increased by 0.03s to 0.07s compared to the comparative method A, a small increase that can meet the real-time evaluation needs in actual training scenarios.

[0033] Table 2 shows a comparison between the output results of this invention and the effects of manual annotation and training control during continuous training. The manual annotation was completed jointly by two rehabilitation trainers, considering movement quality, subjective fatigue level, and heart rate recovery. Comparison method B was a scoring scheme based solely on electromyographic probability sequences. The sample size was 60 periods.

[0034] Table 2 Comparison of Control Effects During Continuous Training

[0035] As shown in Table 2, the present invention demonstrates more stable performance in continuous training control scenarios, achieving a consistency rate of 93.8% to 96.2% with manual annotation, significantly higher than the 84.8% to 86.1% of the comparative method B. The accuracy rates for both training constraint triggering and termination triggering remain at high levels. Especially in the latter half of the second group, when fatigue accumulation is most pronounced, the present invention still achieves 95.7% and 97.1% respectively, indicating that the fatigue control module, by combining probability offset, topology offset, and drift parameters, can make more accurate judgments on the fatigue aggravation trend during training. The number of invalid constraints and invalid terminations are significantly reduced, indicating that the system can avoid excessive intervention in the normal action cycle and output termination commands in a timely manner when the second preset threshold is reached, thereby improving training safety and control effectiveness.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A muscle fatigue assessment system based on wireless dry electrode signals, characterized in that, include: The electromyography (EMG) acquisition module is used to acquire surface EMG signals from multiple muscle groups through a wireless dry electrode garment worn on the body surface of the test subject, and simultaneously acquire the motion state parameters of the test subject to construct EMG motion time series data. The signal processing module is used to perform signal conditioning on the surface electromyography (EMG) signals in the EMG motion time series data, determine the start and end positions of the action cycle based on the motion state parameters, perform periodic segmentation and length alignment on the surface EMG signals, and generate periodic EMG sequences. The baseline construction module is used to acquire the periodic electromyography (EMG) sequence of the test subject in a non-fatigue state, construct an individual baseline sample set, extract feature parameters from the individual baseline sample set, encode the periodic EMG sequence and the feature parameters, and generate a time-series embedding sequence. The temporal modeling module is used to improve the DeepAR model by embedding temporal sequences into the input sequence. It introduces a phase drift clustering mechanism in the temporal recursive layer, divides the periodic electromyography sequence into phase segments and maps them into phase cluster chains, performs drift calculation on the phase cluster chains of adjacent action cycles, and outputs the electromyography probability sequence and reference distribution parameters. The topology analysis module is used to construct a high-dimensional feature point set based on the periodic electromyography sequence and reconstruct the phase space to generate a periodic evolution point cloud. The Persistent Homology algorithm is used to extract the topological structure information of the periodic evolution point cloud and generate a topological barcode sequence. The fatigue control module is used to calculate probability offset and topological offset based on electromyographic probability sequence, reference distribution parameters and topological barcode sequence, and generate local fatigue index by combining drift parameters. It performs cumulative analysis on local fatigue index of multiple action cycles to generate fatigue score. If the fatigue score reaches the first preset threshold, it outputs training restriction instruction. If the fatigue score exceeds the second preset threshold, it outputs termination instruction.

2. The muscle fatigue assessment system based on wireless dry electrode signals according to claim 1, characterized in that, The electromyography acquisition module is specifically: The wireless dry electrode garment is attached to the corresponding muscle group on the body surface of the test subject. The target muscle group is divided into channels according to the preset channel distribution and a channel number sequence is established. The surface electromyography (EMG) signal at each channel location was continuously sampled to obtain the multi-channel raw EMG sequence. While acquiring surface electromyography signals, motion state parameters are recorded, including movement rhythm parameters, joint displacement parameters, movement repetition count parameters, and movement load parameters. The raw multi-channel electromyography sequences and motion state parameters are processed with unified time stamping to generate a timestamp sequence; The multi-channel electromyography (EMG) raw sequences and motion state parameters are aligned based on the timestamp sequence to form EMG motion temporal data arranged in chronological order.

3. The muscle fatigue assessment system based on wireless dry electrode signals according to claim 1, characterized in that, The signal processing module is specifically: The surface electromyography (EMG) signals in the EMG motion time series data were subjected to power frequency suppression, bandpass filtering, motion artifact removal, baseline drift correction, rectification, and envelope smoothing to obtain a preprocessed EMG sequence. Time sequence analysis is performed on the motion state parameters, including the action beat parameters, joint displacement parameters, and action repetition number parameters, to extract the start and end times of the action and determine the start and end positions of the action cycle. The preprocessed electromyographic sequence was periodically truncated according to the start and end positions of the action cycle to obtain multiple action cycle segments; The length of multiple motion cycle segments is normalized, mapping the multiple motion cycle segments to a uniform sampling length; Multiple motion cycle segments with normalized lengths are arranged in chronological order to generate periodic electromyographic sequences.

4. The muscle fatigue assessment system based on wireless dry electrode signals according to claim 1, characterized in that, The acquisition of periodic electromyographic sequences of the test subject in a non-fatigue state to construct an individual baseline sample set specifically involves: When the subject is in a non-fatigue state, surface electromyography (EMG) signals from multiple muscle groups are collected through wireless dry electrode clothing. Simultaneously, action beat parameters, joint displacement parameters, action repetition parameters, and motion load parameters are collected to generate raw EMG sequences and corresponding motion state parameters. The original electromyography (EMG) sequences were subjected to power frequency suppression, bandpass filtering, motion artifact removal, baseline drift correction, rectification, and envelope smoothing to obtain preprocessed EMG sequences. Based on the action beat parameters, joint displacement parameters, and action repetition parameters, the start and end positions of the action cycle are determined. The preprocessed electromyography sequence is then segmented into cycles and aligned in length to generate a periodic electromyography sequence. Multiple consecutive motion cycle segments are selected from the periodic electromyography sequence and arranged according to the channel number sequence and motion cycle order to form a baseline cycle set; The amplitude of the motion cycle segments in the baseline cycle set is made uniform and stored according to the muscle group channel and motion cycle number to construct an individual baseline sample set.

5. The muscle fatigue assessment system based on wireless dry electrode signals according to claim 1, characterized in that, The step of extracting feature parameters from the individual baseline sample set, encoding the periodic electromyography sequence and the feature parameters, and generating a time-series embedding sequence specifically involves: Baseline cycle sets are read from the individual baseline sample set, and electromyographic amplitude parameters, cycle position parameters, movement phase parameters, channel position parameters, and exercise load parameters are extracted according to muscle group channel and movement cycle number; The periodic electromyography (EMG) sequence is unfolded into a time step sequence according to the action cycle. The EMG amplitude parameters are combined with the corresponding time step's cycle position parameters, action phase parameters, channel position parameters, and motion load parameters to form a time step feature group. The parameters in the time step feature group are numerically mapped to obtain the time step encoding vector; the time step encoding vectors are arranged in the order of the time steps to form a time sequence encoding matrix. The temporal coding matrix is ​​subjected to embedding transformation to generate a temporal embedding sequence.

6. The muscle fatigue assessment system based on wireless dry electrode signals according to claim 1, characterized in that, The improved DeepAR model specifically includes an input encoding layer, a temporal recursive layer, a probability output layer, and an autoregressive feedback layer; The input encoding layer expands the temporal embedding sequence into a time step sequence according to the action cycle. At time step t, it reads the corresponding electromyographic amplitude parameters, cycle position parameters, action phase parameters, channel position parameters, and motion load parameters, and concatenates them into the input vector at time step t according to a preset arrangement. The temporal recursive layer establishes an initial hidden state at the first time step, reads the input vector at the t-th time step and the hidden state at the (t-1)-th time step to obtain the hidden state at the t-th time step, and introduces a phase sequence drift clustering mechanism in the temporal recursive layer. The phase sequence drift clustering mechanism divides the sampling point sequence of each action cycle into an initiation phase, a peak holding phase, and a release phase according to the rising, peak, and falling segments of the electromyographic amplitude in the periodic electromyographic sequence. The starting sampling position, ending sampling position, phase length, and peak sampling position are extracted for the initiation phase, peak holding phase, and release phase, respectively, and arranged according to the action cycle sequence to form a phase sequence cluster chain. Establish a correspondence between the same phase segment types for the phase sequence cluster chain in the k-th action cycle and the phase sequence cluster chain in the (k+1)-th action cycle, calculate the center position difference of the oscillating phase segment, the peak holding phase segment, and the release phase segment to obtain the cluster center displacement; calculate the length difference of the oscillating phase segment, the peak holding phase segment, and the release phase segment to obtain the phase segment length change. The difference between the peak sampling positions in the k-th and (k+1)-th action cycles is calculated to obtain the peak position offset. The phase sequence arrangement change is obtained based on the change in the arrangement order of the oscillation phase segment, peak holding phase segment, and release phase segment in the k-th and (k+1)-th action cycles. The cluster center displacement, phase segment length change, peak position offset, and phase sequence arrangement change are combined to form the drift parameter. At time step t, the drift parameter is concatenated with the electromyographic amplitude parameter, periodic position parameter and action phase parameter in the input vector at time step t to obtain the extended input vector. The extended input vector and the hidden state at time step t-1 are written into the state update unit to obtain the hidden state at time step t containing the drift information. The probability output layer performs a linear mapping on the hidden state at time step t to obtain the distribution mean parameter and distribution scale parameter, and generates the electromyographic probability sequence and reference distribution parameter corresponding to time step t based on the distribution mean parameter and distribution scale parameter; The autoregressive feedback layer extracts sampled values ​​from the electromyographic probability sequence at time step t, and then concatenates the sampled values ​​with the input vector at time step t+1 before inputting them into the temporal recursive layer.

7. The muscle fatigue assessment system based on wireless dry electrode signals according to claim 1, characterized in that, The process of constructing a high-dimensional feature point set based on periodic electromyography sequences and reconstructing the phase space to generate a periodic evolution point cloud is as follows: Multiple action cycle segments in the periodic electromyography sequence are read in the order of action cycle, and the sampled value sequence in each action cycle segment is extracted according to the muscle group channel number. In each action cycle segment, the sampled value sequence is sampled at a preset delay interval, and the sampled values ​​at different delay positions are arranged into a state vector according to a preset embedding dimension. Multiple state vectors in the same action cycle segment are combined according to the sampling time order to form a single-cycle state point set; The single-cycle state point sets corresponding to multiple action cycle segments are collected according to the action cycle order and muscle group channel number to form a high-dimensional feature point set; Each state vector in the high-dimensional feature point set is assigned a coordinate position according to the time sequence to generate discrete state points in phase space; The discrete state points in phase space are arranged in the order of action cycles to form a periodic evolution point cloud.

8. The muscle fatigue assessment system based on wireless dry electrode signals according to claim 1, characterized in that, The method of extracting topological structure information of periodically evolving point clouds using the Persistent Homology algorithm and generating topological barcode sequences is as follows: A distance matrix is ​​constructed based on the distance between points in the periodically evolving point cloud, and a filter radius sequence is established based on different distance values ​​in the distance matrix; Under each filtering radius, discrete state points whose distance between points is no greater than the corresponding filtering radius are connected to form a simplex, thus forming a simplex complex under the corresponding filtering radius; Multiple simple complexes are arranged sequentially according to the filter radius sequence to form a filter complex sequence; Extract the generation and vanishing radii of connected components from the filtered complex sequence, and extract the generation and vanishing radii of loop structures; Based on the generation and disappearance radii of connected components and the generation and disappearance radii of loop structures, persistent intervals for connected components and persistent intervals for loop structures are established respectively. The topological barcode sequence is generated by arranging the persistent intervals of connected components and persistent intervals of ring structures according to their start and end positions.

9. The muscle fatigue assessment system based on wireless dry electrode signals according to claim 1, characterized in that, The method involves calculating probability offsets and topological offsets based on electromyographic probability sequences, reference distribution parameters, and topological barcode sequences, and combining these with drift parameters to generate local fatigue indices. Accumulated analysis of these local fatigue indices across multiple motion cycles is then performed to generate a fatigue score. Specifically: Read the electromyographic probability sequence and reference distribution parameters in the order of the action cycle, and generate a probability offset based on the difference between the electromyographic probability sequence and the reference distribution parameters; Read the topology barcode sequence according to the action cycle order, and generate the topology offset based on the difference between the persistent interval of the connected component and the persistent interval of the ring structure between adjacent action cycles. The drift parameters are read sequentially according to the action cycle, and the probability offset, topology offset and drift parameters are combined to form a local fatigue feature group; Local fatigue characteristic groups are accumulated in sequence according to the action cycle to generate local fatigue index; local fatigue indexes from multiple action cycles are accumulated to generate fatigue score.

10. The muscle fatigue assessment system based on wireless dry electrode signals according to claim 1, characterized in that, If the fatigue score reaches a first preset threshold, a training restriction command is output; if the fatigue score exceeds a second preset threshold, a termination command is output. Specifically: Set a first preset threshold and a second preset threshold, wherein the second preset threshold is greater than the first preset threshold; If the fatigue score reaches the first preset threshold and does not exceed the second preset threshold, a training restriction instruction is generated. If the fatigue score exceeds the second preset threshold, a termination command will be generated.

Images