Muscle movement abnormity analysis method for rehabilitation process

Abstract

The invention provides a muscle motion anomaly analysis method for a rehabilitation process, and relates to the field of muscle motion anomaly detection.The muscle motion anomaly analysis method comprises the steps that firstly, the asymmetric cooperative offset rate between channels is extracted based on a muscle cooperative offset tensor, and a normalized cooperative offset matrix is constructed in combination with a time weight factor and a channel pair weight factor; deriving time sequence curve enhanced offset trend perception; then, dynamically integrating abnormal response by adopting an energy disturbance matrix and an accumulated deviation embedding structure, constructing an asymmetric collaborative modeling path based on a self-adaptive adjacency matrix and a disturbance operator matrix, and fusing a tensor interaction structure and channel deviation mapping to generate a deviation field coupling matrix; and finally, a deviation sensitivity response vector and a deviation intensity tensor are generated in combination with the standard template, a Gaussian disturbance kernel function, a sign function and a channel direction weight are introduced to calculate an abnormal scoring index, and model training and optimization updating are completed through a multi-layer sensing network and a cross entropy loss function.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of muscle movement abnormality detection, and in particular to a muscle movement abnormality analysis method for a rehabilitation process. BACKGROUND

[0002] With the development of rehabilitation medicine and intelligent detection technology, the demand for muscle movement function evaluation of rehabilitation populations such as stroke, brain injury, and muscle atrophy is increasing. Existing rehabilitation evaluation relies on the experience of doctors and the subjective feelings of patients, lacks objective and quantitative parameter support, and electromyographic signals, as an important electrophysiological indicator reflecting muscle activity, can provide a basis for movement abnormality analysis during rehabilitation. However, existing methods have poor anti-interference ability, single analysis dimension, and difficulty in adapting to individualized rehabilitation needs. Therefore, it is necessary to propose a muscle movement abnormality analysis method based on electrical measurement to provide support for subsequent rehabilitation process evaluation and training plan adjustment.

[0003] Publication No. CN119564236A discloses a muscle movement analysis method, device, equipment and storage medium based on electromyography data, which uses electromyography data for rehabilitation detection. First, self-adaptive filtering is used for denoising, then a sliding window is used to extract energy and zero-crossing rate, and time-frequency features are obtained by combining wavelet threshold denoising. Finally, these features are fused into a vector and input into a trained random forest model for analysis, thereby identifying muscle movement state in real time and detecting rehabilitation process. Publication No. CN112329640A discloses a facial nerve palsy disease rehabilitation detection system based on eye muscle movement analysis, which uses multi-dimensional data such as eyelid area, eyeball movement speed, eyebrow and forehead wrinkle features, and facial symmetry during the rehabilitation process. The parameters are extracted through semantic segmentation and three-dimensional face reconstruction, a disease severity prediction and rehabilitation evaluation model is established, the features are integrated and the rehabilitation degree is quantified based on an exponential function, thereby achieving objective detection of rehabilitation progress.

[0004] In the prior art, CN119564236A mainly relies on energy features and zero-crossing rate features of electromyography for movement state analysis. The feature extraction method has certain manual dependence and dimension fragmentation, making it difficult to accurately depict the weak disturbance differences in muscle coordination during the rehabilitation process and the dynamic coupling features between multiple channels. At the same time, the use of random forest as a classification model has static structure and is difficult to dynamically adapt to the non-stationarity of abnormal distribution in the training process. CN112329640A focuses more on the fusion modeling of facial geometric feature parameters and explicit posture information, lacks time sequence dynamic perception ability for muscle movement neural control level deviation, and has problems such as insufficient global sensitivity and insufficient dynamic normalization in the way of scoring rehabilitation degree using an exponential function, making it difficult to reflect abnormal mutation trends under instantaneous disturbance and limiting the accuracy and timeliness of rehabilitation abnormality recognition. SUMMARY

[0005] The application provides a muscle movement anomaly analysis method for a rehabilitation process, which comprises the following steps: first, extracting an asymmetric collaborative shift rate between channels based on a muscle collaborative shift tensor, combining a time weight factor and a channel pair weight factor to construct a normalized collaborative shift matrix, and further deriving a time sequence curvature to enhance the shift trend perception; then, describing a disturbance behavior through an energy disturbance matrix, and realizing dynamic integration of abnormal responses by using a cumulative deviation embedding structure; establishing an asymmetric collaborative modeling path between multidimensional channels by using an adaptive adjacency matrix and a disturbance operator matrix, fusing a tensor interaction structure and a channel shift mapping to generate a shift field coupling matrix, and constructing an abnormal performance feature in a multi-channel space; finally, constructing a deviation sensitivity response vector based on the shift field coupling matrix and a standard template, generating a multidimensional deviation intensity tensor, combining a Gaussian disturbance kernel function with bandwidth adjustment capability, a sign function and a channel direction weight to complete abnormal score index calculation, and realizing model training and optimization update by using the abnormal score index through a multilayer perception network and a probability discrimination loss function based on cross entropy, so as to realize multidimensional modeling and accurate abnormal analysis of muscle movement anomalies in the rehabilitation process.

[0006] A muscle movement anomaly analysis method for a rehabilitation process, which comprises the following steps: S1, collecting a muscle movement data set in a rehabilitation detection process, and preprocessing the collected data set to construct a muscle movement data set; S2, dividing the muscle movement data set into muscle movement vectors, calculating a dynamic adjacency matrix sequence according to a collaborative relationship between muscle groups, and obtaining a muscle collaborative shift tensor according to a structure change rate matrix; S3, according to the muscle collaborative shift tensor, weighting and normalizing the asymmetric collaborative shift rate to construct a collaborative shift matrix, and extracting a continuous difference to obtain a time sequence curvature, weighting and superimposing the collaborative shift matrix and the time sequence curvature, introducing a channel coupling weight and a disturbance adjustment factor, obtaining an energy disturbance matrix and cumulatively calculating, and constructing a collaborative embedding feature; S4, based on the collaborative embedding feature, introducing a disturbance weight vector to obtain an adaptive adjacency matrix and constructing a disturbance operator matrix with an asymmetric adjustment factor, fusing matrix information at a historical time, generating a tensor interaction superposition result, introducing a channel shift factor for difference coupling, and forming a shift field coupling matrix; S5, performing exponential mapping of a channel shift intensity through the shift field coupling matrix and a Gaussian disturbance kernel function, and generating a deviation sensitivity response vector by fusing a channel difference; combining a channel direction weight coefficient and a residual direction projection component, generating a deviation intensity tensor and jointly modeling, and obtaining an abnormal score index; S6, inputting the muscle movement data set, and constructing a muscle movement anomaly analysis model in combination with a loss function, and iteratively training until convergence, to complete training of the muscle movement anomaly analysis model.

[0007] Preferably, the construction process of the muscle movement dataset includes data acquisition and preprocessing. In the data acquisition stage, patients at different rehabilitation stages are selected as detection objects, and four types of standard rehabilitation actions, i.e., looking down, looking up, flexion and extension, and fist clenching, are designed. A multi-channel surface electromyography sensor array and an inertial measurement unit are used to collect original electromyography signals, three-axis acceleration, three-axis angular velocity and joint angle change information of target muscle groups, respectively. Synchronization labeling is achieved through a unified time stamp to construct an initial dataset containing multi-source time series data.

[0008] In the preprocessing stage, first, the original electromyography signals are baseline corrected by adjusting the starting value to zero through the whole segment mean offset method. Then, the sliding average method is used to take the mean value of each sampling point and the previous and subsequent sampling points to suppress sudden noise and smooth the waveform. The acceleration and angular velocity signals are synchronously smoothed, and static calibration is completed through zero repositioning. The whole signal is divided into fixed-length time slices, and adjacent time slices are set to overlap to enhance continuity. In each time slice, electromyography features such as electromyography energy, root mean square value, zero-crossing rate, average frequency and median frequency are extracted, the average value, maximum value, peak value and standard deviation of three-axis acceleration and three-axis angular velocity are counted, and the joint angle change and its rate of change between adjacent time slices are calculated. All feature data are uniformly subjected to minimum-maximum normalization to form a muscle movement dataset, which is used as the input of the subsequent abnormal analysis model and rehabilitation evaluation module.

[0009] Preferably, in step S2, the data collected at each time point is divided into muscle movement vectors, and combined according to a fixed time step to form a muscle movement signal matrix. In each time step, the multi-channel muscle movement data features in the muscle movement vector are paired, the cooperative activation intensity between the channel pairs is calculated, and a stability constant is introduced in the calculation process to avoid numerical operation errors. Finally, the cooperative activation intensity of each channel pair is spliced according to the index position to obtain the muscle group adjacency matrix of the current time step, and the dynamic adjacency matrix sequence is formed through iterative calculation in the whole sampling window. Based on the dynamic adjacency matrix sequence, the difference value of the cooperative activation intensity at different times is calculated to obtain the structure change rate. The difference between adjacent time points is calculated through the structure change rate to obtain the cooperative offset change difference value of each channel pair. Based on the cooperative offset change difference value and the consistency sign function, an asymmetric cooperative offset rate for distinguishing the cooperative change direction is obtained. The asymmetric cooperative offset rates at each time point are spliced in turn to form a complete muscle cooperative offset tensor.

[0010] Further, through the establishment of a dynamic muscle group adjacency matrix, the synergistic activation intensity of all channel pairs can be recorded in matrix form at each time step, and a dynamic adjacency graph sequence is formed within the entire sampling window. The structural change rate matrix is used to calculate the difference in synergistic activation intensity between adjacent time steps, and the resulting structural change rate can refine the instantaneous fluctuations of muscle group synergy. Subsequently, based on the construction process of the synergistic offset rate, the structural change rate at different time steps is further decomposed to extract the offset information of the channel pairs during the evolution process. On this basis, the consistency sign function is introduced to distinguish the direction of the synergistic offset rate, and then the asymmetric synergistic offset rate is constructed, and the muscle synergistic offset tensor is generated by time iteration. This step as a whole ensures the organic connection between the muscle group adjacency matrix, the structural change rate matrix, the asymmetric synergistic offset rate, and the muscle synergistic offset tensor, so that the muscle movement signal is completely modeled in terms of inter-channel synergistic activation intensity, time series difference, directional offset, and asymmetric evolution trend.

[0011] Preferably, in step S3, based on the muscle synergistic offset tensor, the asymmetric synergistic offset rate between channels is extracted, the channel pair weight factor is introduced in combination with the influence degree of different muscle groups during the rehabilitation training process, and the time weight factor is set in combination with the time decay to jointly weight the asymmetric synergistic offset rate; the weighted result is normalized according to the global maximum and minimum values of the muscle synergistic offset tensor, and a synergistic offset matrix is constructed; Based on the synergistic offset matrix, the continuous difference between adjacent time steps is calculated to the second order, and normalized to construct a time sequence curvature sequence. According to the synergistic offset matrix and the time sequence curvature, the offset difference item of adjacent time steps is extracted, the time sequence curvature is taken as a calculation component, and the offset difference item is weighted and superimposed according to a preset weight relationship; in the weighting process, the channel coupling weight factor is introduced to distinguish the sensitivity of different channel pairs in the calculation, and the disturbance adjustment factor is combined to balance the relative proportion of the curvature component and the difference component, and an energy disturbance matrix is obtained. An accumulated deviation embedding method is constructed through the energy disturbance matrix, and the energy disturbance of each channel pair is weighted and accumulated within a fixed time window W based on the historical average energy disturbance level, and the accumulated deviation result is obtained and converged as a synergistic embedding feature.

[0012] Further, the muscle group synergy relationship in the rehabilitation training process is dynamically characterized at different levels. On the one hand, by combining the synergy offset matrix and the time curvature, the change of the muscle group synergy mode in the time dimension is captured, which can sensitively reflect the acceleration characteristics of abnormal offset. On the other hand, by introducing the channel pair weight factor, the time weight factor and the disturbance adjustment factor, the ability to distinguish the characteristics of different muscle groups and different time periods is enhanced, effectively avoiding the distortion caused by a single index. At the same time, the joint construction of the energy disturbance matrix and the cumulative deviation embedding can take into account both the instantaneous fluctuation and the long-term cumulative effect, so as to describe the stability and offset trend of muscle synergy in the rehabilitation process at multiple scales. This method improves the robustness and comprehensiveness of feature extraction, and provides discriminative embedded feature support for subsequent anomaly detection and rehabilitation evaluation.

[0013] Preferably, in step S4, the synergy feature relationship between any channel pair is jointly modeled, and a disturbance weight vector is introduced to dynamically regulate the response degree between channels. Combined with a set of adjustment mechanisms for enhancing numerical stability, a weight value for representing the relationship strength between channels is formed, and an adaptive adjacency matrix is constructed. Based on the adaptive adjacency matrix, the bidirectional synergy strength between channels is differentially characterized, and an asymmetric adjustment factor is introduced to regulate the relative weight between different calculation components to form a disturbance operator matrix for representing asymmetric synergy disturbance. Through the disturbance operator matrix, combined with the multi-level matrix information at the historical moment, the tensor outer product operation and the element-by-element multiplication operation are introduced at the matrix level to respectively establish the interaction structure and the historical association path across time points. At the same time, the historical memory weight is set to regulate the joint degree between the current and past operators, and the tensor interaction superposition result is constructed. Based on the channel dimension of the tensor interaction superposition result, the channel offset factor is introduced for difference coupling operation on the corresponding elements at consecutive time points. For the change trend of the tensor in each channel dimension, combined with the tensor mapping results of the current and the previous moment, relying on multiple independent offset factors, the response degree mapping across dimensions is completed to generate an offset field coupling matrix.

[0014] Further, the adaptive adjacency matrix, the perturbation operator matrix, the tensor interaction superposition result and the offset field coupling matrix are constructed to realize dynamic modeling of the asymmetric collaborative relationship between the multi-channel electromyography signals in the rehabilitation training process. This process effectively captures the collaborative response difference and perturbation propagation characteristics between channels, fuses historical state information, and enhances the ability to describe the nonlinear interaction path between channels. By introducing the perturbation weight vector, the asymmetric adjustment factor, the historical memory weight and the channel offset factor, multi-scale modeling and time sequence sensitivity regulation of the evolution of the relationship between channels are realized, thereby improving the recognition ability and robustness of the muscle movement abnormality analysis model in complex muscle group collaborative scenarios, and providing more discriminative collaborative embedded structure representation for subsequent abnormality detection and rehabilitation process evaluation.

[0015] Preferably, in step S5, the perturbation performance of each channel dimension is extracted based on the offset field coupling matrix, a Gaussian perturbation kernel function with bandwidth adjustment capability is introduced to perform exponential mapping on the channel offset strength, and the activation difference between the standard template activation mode and the actual activation mode of the patient on the corresponding channel dimension is combined to generate a bias sensitivity response vector dimension by dimension; The channel dimension fusion is performed on the bias sensitivity response vector and the channel direction weight coefficient, then the difference direction between the reference template activation value and the patient activation value is biased and coded by combining the sign function, the multi-channel direction projection component is further superimposed, the weighted normalization operation of the residual direction tensor is completed, and finally the bias strength tensor of the multi-dimensional channel bias response strength is generated; Based on the bias strength value at the current moment, a normalization construction strategy with local logarithmic amplitude variation characteristics is introduced, and an abnormal score index is formed in combination with the bias strength value at the previous moment.

[0016] Further, the joint mapping mechanism of the Gaussian perturbation kernel function and the channel activation difference is introduced based on the offset field coupling matrix, which can realize fine modeling of the perturbation response in the channel dimension and improve the resolution capability of the bias sensitivity response vector. Further, the multi-dimensional direction projection relationship is established by combining the channel direction weight coefficient and the sign function, so that the residual direction tensor can accurately reflect the intensity difference of the bias between channels. At the same time, by introducing the normalization construction strategy with local logarithmic amplitude variation characteristics, the bias strength values at the current and previous moments are dynamically associated to form a time sequence continuous abnormal score index, thereby realizing dynamic description of the channel offset behavior and accurate quantification of the abnormal response, and providing higher time sequence sensitivity and discrimination accuracy for multi-channel collaborative bias recognition.

[0017] Preferably, in step S6, the abnormal score index is used to realize muscle movement abnormality analysis through two-layer multilayer perception. The muscle movement abnormality analysis model is trained based on the cross-entropy loss function of probability discrimination to realize updating of the muscle movement abnormality analysis model hyperparameters.

[0018] Further, by inputting the final abnormal score index into a two-layer multi-layer perception machine to realize nonlinear mapping analysis of muscle movement abnormalities, the representation fit between the score index and the actual abnormal pattern can be effectively enhanced, and the recognition and analysis capability of the model for complex abnormal patterns can be improved; the cross-entropy loss function based on probability discrimination is used to supervise and optimize the abnormal analysis model, which can realize the discrete alignment between the abnormal probability prediction result and the real abnormal label, ensure the convergence stability of the model, and help improve its generalization performance and discrimination robustness in the muscle coordination abnormal recognition task, so as to realize the automatic optimization and structure adaptive update of the key hyperparameters of the muscle movement abnormal analysis model.

[0019] In the above technical solution, the present application has the following technical effects and advantages: 1. The present application proposes to extract the asymmetric coordination offset rate based on the muscle coordination offset tensor, and construct a normalized coordination offset matrix combined with a time weight factor and a channel pair weight factor, and further deduce the time sequence curvature to strengthen the dynamic modeling capability of the coordination relationship, which can realize the continuous tracking and accurate perception of the change trend of the coordination relationship between muscle groups during rehabilitation training, thereby enhancing the recognition accuracy and time sensitivity of abnormal activation patterns; 2. The present application introduces an energy disturbance matrix to express the disturbance behavior and construct a cumulative deviation embedding structure to realize dynamic aggregation of abnormal response information, which can capture long-time-scale muscle group activation abnormal accumulation performance while maintaining the response speed of local disturbance changes, thereby realizing effective modeling and analysis of implicit abnormal states in the rehabilitation process of multiple time periods; 3. The present application establishes an asymmetric coordination modeling path by constructing an adaptive adjacency matrix and a disturbance operator matrix, and introduces a tensor interaction structure and a channel offset mapping to construct an offset field coupling matrix, which can strengthen the spatial interaction feature expression between muscle groups across channels, realize the identification and modeling of coordination abnormal paths in the channel dimension, and ensure the accurate recovery of the dynamic conduction relationship between channels and the complete analysis of the disturbance transmission mechanism; 4. The present application introduces a Gaussian disturbance kernel function with bandwidth adjustment capability to exponentially map the channel offset strength based on the offset field coupling matrix, and constructs a deviation sensitivity response vector and a deviation intensity tensor combined with a sign function and a channel direction weight, which realizes the multi-dimensional modeling process of the abnormal score index with high numerical resolution and direction discrimination ability, effectively distinguishing between edge abnormal states and normal fluctuations in muscle activation; 5. The present application inputs the abnormal score index into a multi-layer perception network and introduces a probability discrimination loss function based on cross-entropy for model training, which enhances the convergence stability and generalization capability of the muscle movement abnormal analysis model in high-dimensional nonlinear space, thereby realizing muscle group activation abnormal analysis in the whole rehabilitation training process. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a flow chart of a muscle movement abnormality analysis method for a rehabilitation process provided by the present application.

[0021] Figure 2 is a structural diagram of a muscle synergy shift tensor constructed by the present application.

[0022] Figure 3 is a structural diagram of a synergy embedding feature constructed by the present application.

[0023] Figure 4 is a structural diagram of a shift field coupling matrix constructed by the present application.

[0024] Figure 5 is a structural diagram of an abnormality scoring indicator constructed by the present application.

[0025] Figure 6 is a loss convergence curve diagram in a muscle movement abnormality analysis model training process provided by the present application.

[0026] Figure 7 is a visualization dimension reduction diagram of a high-dimensional deviation intensity tensor provided by the present application.

[0027] Figure 8 is a comparison curve diagram of true values and predicted values provided by the present application. DETAILED DESCRIPTION

[0028] The present application provides a muscle movement abnormality analysis method for a rehabilitation process. First, the inter-channel asymmetric synergy shift rate is extracted based on a muscle synergy shift tensor. A normalized synergy shift matrix is constructed by combining a time weight factor and a channel pair weight factor, and a time sequence curvature is further derived to enhance the shift change trend perception. Then, the disturbance behavior is described by an energy disturbance matrix, and the dynamic integration of abnormal responses is realized by using a cumulative deviation embedding structure. An asymmetric synergy modeling path in a multi-dimensional channel is established by using an adaptive adjacency matrix and a disturbance operator matrix, an offset field coupling matrix is generated by fusing a tensor interaction structure and a channel shift mapping, and an abnormality performance feature in a multi-channel space is constructed. Finally, a deviation sensitivity response vector is constructed based on the offset field coupling matrix and a standard template, a multi-dimensional deviation intensity tensor is generated, an abnormality scoring indicator is calculated by combining a Gaussian disturbance kernel function with bandwidth adjustment capability, a sign function and a channel direction weight, and model training and optimization update are realized by using the abnormality scoring indicator through a multi-layer perception network and a probability discriminant loss function based on cross entropy, so as to realize multi-dimensional modeling and accurate abnormality analysis of muscle movement abnormalities in a rehabilitation process.

[0029] Please see Figure 1As shown, the muscle movement abnormality analysis method for the rehabilitation process in the embodiment of the present application includes the following specific steps.

[0030] S1, collect the muscle movement data set in the rehabilitation detection process, and pre-process the collected data set to construct the muscle movement data set.

[0031] In the embodiment, the construction process of the muscle movement data set includes data collection and pre-processing. In the data collection stage, patients in different rehabilitation stages are selected as detection objects, four types of standard rehabilitation actions, i.e., looking down, looking up, flexing and clenching, are designed to cover various typical action types. A multi-channel surface electromyography sensor array is used as the collection device, the sensors are arranged at the key positions on the surface of the target muscle group, the sampling rate of each channel is set to 2000 Hz, the sampling accuracy is set to 16 bit, and the time-domain waveform of muscle contraction and relaxation is recorded completely. The kinematic parameters of the muscle group are synchronously collected, including three-axis acceleration, three-axis angular velocity and joint angle change, the sampling frequency of the inertial measurement unit is set to 200 Hz. The start and end time of the action is labeled by writing a unified time base signal channel through a wired trigger signal, and all collection channels are synchronized based on a unified timestamp. The collected muscle movement data set in the rehabilitation detection process includes the following: original electromyography signal, electromyography energy, electromyography root mean square value, electromyography zero-crossing rate, electromyography average frequency, electromyography median frequency, acceleration component X, acceleration component Y, acceleration component Z, acceleration synthetic value, angular velocity component X, angular velocity component Y, angular velocity component Z, angular velocity average value, angular velocity peak-to-peak value, angular velocity standard deviation, joint angle change and joint angular velocity, which are used to represent muscle contraction intensity, contraction rhythm, action amplitude and action speed. In the preprocessing stage, the collected data is baseline corrected, outliers removed, signal cut, statistical calculation and numerical standardization steps; first, the average value of each collected electromyography signal is calculated, and the average value is subtracted from the signal point by point, so that the starting value of the signal is close to zero; then each sample point in the whole signal is averaged with the adjacent two points, and the original value is replaced point by point, which is used to weaken the mutation peak and make the waveform smoother; the acceleration and angular velocity data are processed in the same way, each sample point is replaced by the average value of the adjacent two points, which eliminates the jitter and stabilizes the fluctuation; all channel signals are divided into equal length short segments in time sequence, each segment lasts about 0.2 seconds, and the adjacent segments overlap half the time length; in each time segment, the energy of the electromyography signal is calculated by summing the square of each sample value, the root mean square value is calculated by averaging the square and then taking the square root, the zero crossing rate is calculated by dividing the number of times the signal crosses zero by the number of sample points, and the average frequency and median frequency are calculated by using fast Fourier transform to calculate the frequency spectrum of each segment; at the same time, the average value and maximum value of the sample values in the three directions of the acceleration signal are calculated, and the synthetic acceleration is obtained by calculating the square sum and square root of the three directions; the average value and maximum value of the three axes of the angular velocity data are also calculated, and the peak-peak value is calculated by calculating the difference between the maximum and minimum of each segment, and the stability index is calculated by calculating the standard deviation of the fluctuation amplitude; the joint angle change is obtained by comparing the synthetic acceleration and angular velocity of the previous and next time slices, and the position difference of the two time points is approximated as the angle change; the change rate of angular velocity is calculated by dividing the angle difference of the adjacent two time slices by the time interval; all the calculation results are arranged in time sequence, and each data is linearly scaled to [-1, 1] interval by the minimum and maximum value, so that the data from different sources has a unified scale, and finally the muscle movement data set is constructed; the processed muscle movement data set is arranged in time sequence, the row index corresponds to the time sequence, the column index corresponds to the feature dimension, and the category information and start and end time of the action are also retained, which is used as the input of the subsequent rehabilitation muscle abnormality analysis model.

[0032] S2, divide the muscle movement data set into muscle movement vectors, calculate the synergistic relationship between muscle groups to obtain a dynamic adjacency matrix sequence, and obtain a muscle synergistic offset tensor according to a structural change rate matrix.

[0033] Further, in step S2, a muscle synergistic offset tensor is constructed, and the process is as shown in Figure 2 The specific steps of constructing the muscle synergistic offset tensor are as follows.

[0034] S21, divide the data collected at each time point into muscle movement vectors, and combine them according to a fixed time step to form a muscle movement signal matrix.

[0035] In this embodiment, the mathematical model of the sample data at time t is: ; wherein C is the number of channels, and is 18, is the data value of the Cth channel at time t, is the sample at time t; the data is divided by a fixed time step T, T is 50, each sample is combined according to the fixed time step to form a muscle movement signal matrix, and the mathematical model is: ; wherein, is the muscle movement signal matrix, that is, a muscle movement data sample.

[0036] S22, in each time step, the multi-channel muscle movement data features in the muscle movement vector are paired, the synergistic activation strength between the channel pairs is calculated, and a stability constant is introduced in the calculation process to avoid numerical operation errors, and finally the synergistic activation strength of each channel pair is spliced according to the index position to obtain the muscle group adjacency matrix of the current time step, and the dynamic adjacency matrix sequence is formed through iteration calculation in the whole sampling window.

[0037] In this embodiment, in order to extract the synergistic activation strength of multi-channel electromyography signal in the process of executing rehabilitation action, a muscle group adjacency matrix based on each time t is constructed, which can describe the synergistic activation degree between any two muscle groups in the channel at this time, and the mathematical model is: ; wherein, , is the data value of channel i and channel j at time t; is a numerical stability constant, which is used to avoid division by zero error, and is preferably set to 10e-8, is the norm operation of , is the norm operation of , is the synergistic activation strength of channel i and j at time t, and the synergistic activation strength value obtained by calculating each channel is spliced into a muscle group adjacency matrix ; through iterative calculation of sliding time step, a complete dynamic adjacency graph sequence is obtained, and the mathematical model is: ; wherein, is the dynamic adjacency graph sequence, is the synergistic activation strength value obtained by calculating each channel spliced into a muscle group adjacency matrix at time t.

[0038] S23. Based on the dynamic adjacency matrix sequence, calculate the difference in co-activation intensity at different times to obtain the structural change rate. Calculate the difference between adjacent times using the structural change rate to obtain the co-offset change difference of each channel pair. Based on the co-offset change difference combined with the consistency sign function, obtain the asymmetric co-offset rate used to distinguish the direction of co-offset change. Concatenate the asymmetric co-offset rates at each time point to form a complete muscle co-offset tensor.

[0039] In this embodiment, the structural change rate matrix is ​​defined based on the difference in the adjacency matrix of each muscle group in the dynamic adjacency graph sequence, and its mathematical model is as follows: ; in, The structural change rate matrix consists of the structural change rate at each time step. The internal calculation of the structural change rate represents the cooperative activation intensity of channels a and b at different times. Its mathematical model is as follows: ; in, Let be the cooperative activation intensity of channels a and b at time t. Let the cooperative activation strength of channels a and b at time t-1 be denoted as . Let a and b be the rates of structural change of channels at time t; To describe the asymmetric trend in the evolution of muscle group synergistic structures, the difference in synergistic offset is defined, and its mathematical model is as follows: ; in, Let be the difference in coordinated offset changes between muscle group pairs a and b at time t. Then, to determine the consistency of the perturbation direction, a direction consistency sign function is introduced, and an asymmetric coordinated offset rate is defined. Its mathematical model is as follows: ; in, This represents the asymmetric cooperative offset rate of muscle group pairs a and b at time t. A value less than 0 indicates a sudden change in the cooperative direction. The `sign` function is used to determine whether the cooperative directions are consistent. This structure, while maintaining the perturbation amplitude, highlights the asymmetric offset characteristics caused by inconsistent directions. After calculating the asymmetric cooperative offset rate for each channel, the muscle cooperative offset tensor is finally constructed, and its mathematical model is as follows: ; in, For muscle coordinating offset tensor, Let be the muscle coordination offset rate at time T in the sample.

[0040] S3. Based on the muscle cooperative offset tensor, the asymmetric cooperative offset rate is weighted and normalized to construct a cooperative offset matrix, and the continuous difference is extracted to obtain the temporal curvature. The cooperative offset matrix and the temporal curvature are weighted and superimposed, and channel coupling weights and perturbation adjustment factors are introduced to obtain the energy perturbation matrix and accumulate calculation to construct cooperative embedding features.

[0041] Furthermore, in step S3, collaborative embedding features are constructed, the process of which is as follows: Figure 3 As shown, the specific steps for constructing collaborative embedding features are as follows.

[0042] S31. Based on the muscle co-location tensor, extract the asymmetric co-location rate between channels, introduce channel pair weighting factors in combination with the influence of different muscle groups during rehabilitation training, and combine the time weighting factor set with time decay to perform joint weighting processing on the asymmetric co-location rate; perform normalization operation on the weighting result based on the global maximum and minimum values ​​of the muscle co-location tensor to construct the co-location matrix.

[0043] In this embodiment, the muscle co-shift tensor After initial construction, significant differences in amplitude and uneven distribution over time were observed between different muscle group channels. Therefore, a channel-weighting factor was introduced. With time weighting factor The asymmetric cooperative offset rate is weighted and normalized to obtain the cooperative offset matrix. The channel is weighted by factors The time weighting factor is set based on the importance of the dominant muscle groups in rehabilitation training; An exponential decay function is used to highlight recent abnormal contributions during rehabilitation training. The mathematical model for the synergistic offset matrix is ​​as follows: ; in, The channel-to-weighting factor is used to differentiate the influence of different muscle groups in anomaly analysis. During training, the channel-to-weighting factor with a larger influence value is set to 1, while the influence value for auxiliary muscle groups is set to 0.6. To obtain the maximum value in the muscle co-shift tensor, To obtain the minimum value in the muscle co-shift tensor, Let c and d be the asymmetric cooperative offset rates of muscle groups at time t. The time-weighted factor has the following mathematical model: ,in This is the attenuation parameter, with a value of 0.01. Let be the normalized cooperative offset matrix for the c-th and d-th pairs of muscle groups at time t.

[0044] S32. Based on the cooperative offset matrix, perform second-order calculations on the continuous differences of adjacent time steps and normalize them to construct a time-series curvature sequence.

[0045] In this embodiment, based on the normalized cooperative offset matrix The second-order difference between adjacent time steps is calculated to obtain the temporal curvature of the channel pair, highlighting the acceleration of the shift value, thereby capturing the inflection point characteristics of muscle group synergy patterns during rehabilitation movement execution. The mathematical model of the temporal curvature of the channel pair is as follows: ; in, Let be the normalized cooperative offset matrix for the c and d pairs of muscle groups at time t+1; Let be the normalized cooperative offset matrix for the c and d pairs of muscle groups at time t-1; Let be the temporal curvature of channel c and channel d at time t. By introducing the temporal curvature operator, not only can the direction of the shift change be reflected, but the acceleration effect of abnormal inflection points can also be amplified. When a patient experiences movement imbalance or a sudden excessive contraction of a muscle group during rehabilitation training, the curvature value will change significantly, thus providing an early signal for abnormal detection.

[0046] S33. Based on the cooperative offset matrix and temporal curvature, firstly, the offset difference terms of adjacent time steps are extracted. The temporal curvature is used as a calculation component and is weighted and superimposed with the offset difference terms according to a preset weight relationship. During the weighting process, a channel coupling weight factor is introduced to distinguish the sensitivity of different channels in the calculation. At the same time, a perturbation adjustment factor is combined to balance the relative weight of the curvature component and the difference component, thus obtaining the energy perturbation matrix.

[0047] In this embodiment, to characterize the abnormality of muscle group coordinated migration from an energy perspective, a coordinated energy perturbation field modeling method is introduced. Specifically, the temporal curvature and the migration difference are weighted and superimposed to construct an energy perturbation matrix. Its mathematical model is: ; in, This is the normalized cooperative offset difference term, whose function is to preserve the instantaneous change information of the offset amplitude. This is a perturbation adjustment factor, adaptively learned during training, with an initial value of 0.5. Its role is to balance the relative values ​​of the curvature term and the difference term in energy modeling. These are channel coupling weights, whose values ​​are updated during training. These weights are used to adjust the sensitivity of different channels to temporal curvature. Let be the energy perturbation matrix of channel c and channel d at time t.

[0048] S34. Construct a cumulative deviation embedding method based on the energy perturbation matrix. Using the historical average energy perturbation level as a benchmark, perform weighted accumulation of the energy perturbation of each channel pair within a fixed time window W to obtain the cumulative deviation result, and converge it into a collaborative embedding feature.

[0049] In this embodiment, within the time window W, the cumulative deviation of channel pairs c and d is calculated. The energy disturbance at each moment is compared with its historical average value, and a time decay factor is introduced. The results at different time points are weighted and summed to obtain the cumulative deviation. The mathematical model of the cumulative deviation is as follows: ; in, Let be the accumulated deviation of channel pairs c and d within window W at time t. This reflects the cumulative disturbance of the channel pairs within window W, where W is the length of the time window, with a value of 40. The historical average energy perturbation level of the channel with respect to c and d. For a moment The energy disturbance value, This is the time decay factor, used to assign weights to the perturbation values ​​within window W. The initial value is set to 0.5, and its value during training is adjusted accordingly. and descent parameters The related exponential decay, The initial value is set to 0.05; After obtaining the cumulative deviation of all channel pairs, the cooperative embedding feature at time t is formed by summing their absolute values. The mathematical model is as follows: ; in, Let be the collaborative embedding feature at time t.

[0050] S4. Based on the collaborative embedding feature, a perturbation weight vector is introduced to obtain an adaptive adjacency matrix and a perturbation operator matrix is ​​constructed with an asymmetric adjustment factor. The matrix information of historical moments is fused to generate a tensor interactive superposition result. A channel offset factor is introduced for differential coupling to form an offset field coupling matrix.

[0051] Furthermore, in step S4, the offset field coupling matrix is ​​constructed, and the process is as follows: Figure 4 As shown, the specific steps for constructing the offset field coupling matrix are as follows.

[0052] S41. By jointly modeling the collaborative feature relationship between any channel pairs and introducing a perturbation weight vector to dynamically adjust the response degree between channels, combined with a set of adjustment mechanisms to enhance numerical stability, a weight value is formed to characterize the strength of the relationship between channels, and an adaptive adjacency matrix is ​​constructed.

[0053] In this embodiment, to characterize the dynamically changing muscle group synergy relationships during rehabilitation training, an adaptive adjacency matrix is ​​constructed based on synergistic embedding features. Each element in the adaptive adjacency matrix is ​​an adaptive adjacency weight. During the calculation process, a perturbation weight vector is introduced to progressively weight and correct the embedding dimensions, and a stability constant is added to the denominator to ensure numerical stability. At time t, the adaptive adjacency weights of channel c and channel d are obtained by multiplying their embedding features and perturbation weights dimension by dimension and then taking the inner product. The mathematical model of the adaptive adjacency weights is as follows: ; in, Let be the collaborative embedding feature of channel c at time t. Let be the collaborative embedding feature of channel d at time t. This is a stability constant, taking the value 10e−6, used to avoid the denominator being zero. This is a perturbation weight vector, initially set to all 1s. Its function is to adjust the importance of feature dimensions and is updated during training based on the energy perturbation signal. The adaptive adjacency weight represents the dynamic cooperative strength between channels c and d. for The second normal form operation, for The second normal form operation.

[0054] S42. Based on the adaptive adjacency matrix, the perturbation operator matrix is ​​formed to characterize the bidirectional cooperative strength between channels by differentiating the strength of the cooperative strength and introducing an asymmetric adjustment factor to regulate the relative weights between different computational components.

[0055] In this embodiment, after obtaining the adaptive adjacency weight Subsequently, to further characterize the asymmetric activation relationship between different muscle groups during rehabilitation training, an asymmetric perturbation operator was constructed. This asymmetric perturbation operator, by maintaining directional differences and introducing a product term to enhance bidirectional interaction, can characterize the imbalance between channels during the co-activation process. The mathematical model of the asymmetric perturbation operator is as follows: ; in, The strength of the cooperative relationship from channel c to channel d at time t is derived from the adaptive adjacency matrix of S41, reflecting the degree of cooperative activation of the two channels at that time. Let be the reverse cooperative strength from channel d to channel c at time t, and . They are independent of each other, reflecting the asymmetry inherent in bidirectional coupling; This is an asymmetric adjustment factor used to balance the contributions of the difference and product terms. The initial value is set to 0.3, and it is automatically updated during training through gradient descent. The range is [0,1]. The larger the value, the stronger the influence of the product term. The result of the asymmetric perturbation operator represents the asymmetric cooperative offset of the channel pair (c,d) at time t. The asymmetric perturbation operator results for each channel pair constitute the perturbation operator matrix. .

[0056] S43. By perturbating the operator matrix and combining multi-level matrix information at historical moments, tensor outer product operation and element-wise multiplication operation are introduced at the matrix level to establish cross-time interaction structure and historical association path respectively. At the same time, historical memory weight is set to regulate the degree of joint between current and past operators, and tensor interaction superposition result is constructed.

[0057] In this embodiment, to characterize the temporal continuity and high-order dynamic evolution of muscle group synergy during rehabilitation, a tensor interaction superposition mechanism is proposed based on the aforementioned asymmetric perturbation operator. The electromyographic signals of rehabilitation training typically exhibit a multi-layered progressive relationship in the time series: changes in adjacent time steps reflect short-term muscle activation responses, while changes across time steps reveal long-term adjustment trends between training phases. Traditional first-order difference methods can only capture transient changes and cannot simultaneously consider multi-temporal dependencies. Therefore, this step introduces a combined operation of tensor outer product and element-wise product at the matrix level to form a multi-order structure with historical memory characteristics. The mathematical model of the tensor interaction superposition mechanism is as follows: ; in, Let be the asymmetric perturbation operator matrix at time t. , These are the asymmetric perturbation operator matrices for the previous time step and the two time steps prior, respectively. The tensor outer product operation maps two matrices to a higher-order tensor structure, and its purpose is to characterize nonlinear interactions across time steps. For element-wise multiplication, it emphasizes the coupling relationship between the current matrix and the historical matrix at the item-by-item level. The historical memory weights are used to control the influence of element-wise product terms. The initial value is set to 0.5, and it is updated during training, ranging from [0,1]. This is the result of tensor interaction overlay, containing interaction information from multiple time points.

[0058] S44. Based on the channel dimension of the tensor interaction superposition result, a channel offset factor is introduced to the corresponding elements at consecutive time steps to perform differential coupling operation; for the change trend of the tensor in each channel dimension, combined with the tensor mapping result of the current time step and the previous time step, the cross-dimensional responsivity mapping is completed by relying on multiple independent offset factors to generate the offset field coupling matrix.

[0059] In this embodiment, to convert the dynamic features in the high-order tensor structure into an energy representation that can be used for anomaly detection, this step, based on the difference operation of the tensor channel dimension of the tensor cross-superposition result and the introduction strategy of the channel offset factor, obtains the offset field coupling matrix through offset field coupling mapping. The muscle group synergy pattern in rehabilitation training has temporal evolution, representing changes in energy coupling between channels. Different channels exhibit complex responses such as synchronous enhancement, phase lag, and signal attenuation at specific movement stages. To capture multi-channel offset patterns, this step generates the offset field coupling matrix by performing layer-by-layer analysis of the channel dimension of the tensor cross-superposition result. The offset field coupling matrix is ​​composed of offset field coupling elements; the mathematical model of the offset field coupling elements is: ; Where k is the tensor channel dimension index, ranging from [1, 18], and each k corresponds to an independent channel offset factor. ; This is the channel offset factor, used to measure the sensitivity of different dimensions to anomalies. The initial value is randomly distributed in the range of [0.1, 0.9] and adaptively adjusted. The element values ​​of the tensor interaction superposition result in the channel pair (c,d) and dimension k are derived from S43 and reflect the interaction strength of the channel pair in dimension k. Given the tensor interaction values ​​in the same dimension at the previous time step, extract the rate of change of interaction by differing from the current time step; The element values ​​of the offset field coupling matrix represent the energized offset response of the channel pair (c,d) at time t. The element values ​​of the offset field coupling matrix for each channel pair constitute the offset field coupling matrix. .

[0060] S5. By exponentially mapping the channel offset intensity to the offset field coupling matrix and the Gaussian perturbation kernel function, and by fusing the channel differences, a deviation sensitivity response vector is generated. Combined with the channel direction weight coefficient and the residual direction projection component, a deviation intensity tensor is generated and jointly modeled to obtain the anomaly scoring index.

[0061] Furthermore, in step S5, anomaly scoring metrics are constructed, the process of which is as follows: Figure 5 As shown, the specific steps for constructing anomaly scoring indicators are as follows.

[0062] S51. Based on the offset field coupling matrix, the perturbation performance of each channel dimension is extracted. A Gaussian perturbation kernel function with bandwidth adjustment capability is introduced to perform exponential mapping on the channel offset intensity. Combined with the activation difference between the standard template activation mode and the patient's actual activation mode in the corresponding channel dimension, the deviation sensitivity response vector is generated dimension by dimension.

[0063] In this embodiment, when quantitatively measuring the response difference between the standard template activation mode and the patient's actual activation mode, a Gaussian perturbation kernel function is introduced to exponentially modulate the channel response in order to accurately express the perturbation performance of different channel dimensions in the offset field coupling matrix. Specifically, a deviation sensitivity response vector is constructed by combining the offset field coupling matrix and the channel activation difference. The mathematical model of the deviation sensitivity response vector is as follows: ; in, To control the response sensitivity to channel perturbations, the core width is adjusted, preferably to 0.8. The activation value of the standard template activation mode in the k-th dimension. The activation value of the patient's actual activation mode in the k-th dimension. Let be the bias sensitivity response vector in the k-th channel dimension, representing the bias response in the k-th channel dimension. It is an exponential function.

[0064] S52. Channel dimension fusion is performed by combining the bias sensitivity response vector and the channel direction weight coefficient. Then, the difference direction between the reference template activation value and the patient activation value is biased and encoded by combining the sign function. The multi-channel direction projection components are further superimposed to complete the weighted normalization operation of the residual direction tensor, and finally the bias intensity tensor of the multi-dimensional channel bias response intensity is generated.

[0065] In this embodiment, to further describe the direction of patient deviation from multiple dimensions, a residual weighting mechanism is used to weight the deviation sensitivity response vector generated in the previous step. By fusing the channel direction weights, a multidimensional deviation direction tensor is constructed, and the deviation direction is specified using a sign function, resulting in a deviation intensity tensor. The mathematical model of the deviation intensity tensor is as follows: ; in, The value of the deviation intensity after projection of the residual direction at time t. These are the channel direction weight coefficients, initially set to 1, and iteratively updated during training. Let k be the bias sensitivity response vector in the k-th channel dimension. For a sign function, when When less than 0, the sign function value is -1; when... When equal to 0, the sign function value is 0; when... When the value is greater than 0, the sign function value is 1, and K is the total number of feature channels with a value of 18.

[0066] S53. Based on the deviation intensity value at the current moment, a normalization construction strategy with local logarithmic amplitude characteristics is introduced, which is combined with the deviation intensity value at the previous moment to form an anomaly scoring index.

[0067] In this embodiment, to comprehensively assess the degree of abnormal deviation of the patient's activation behavior at the temporal level, this step introduces a time-sensitive distribution structure entropy comparison mechanism based on the deviation intensity scalar, to achieve dynamic detection of deviation across time. This distribution structure entropy comparison mechanism constructs an exponential deviation scoring function based on local logarithmic increase, fusing the current deviation intensity value with the deviation intensity value at the previous time step to form an abnormality scoring index. This index measures the magnitude of the temporal deviation abrupt change between the patient's activation pattern and the standard template. Its mathematical model is as follows: ; in, The constant term is introduced and has a value of 10e−6 to prevent the denominator from being zero and to ensure calculation stability; the logarithmic term is used to measure the rate of change of the current deviation value relative to the historical deviation, thereby amplifying and normalizing the abnormal amplitude. This is the final anomaly scoring indicator; the larger the value, the more significant the deviation. This represents the deviation intensity value after projection of the residual direction at time t-1.

[0068] S6. Input the muscle movement dataset, combine it with the loss function, construct a muscle movement anomaly analysis model, and iteratively train it until convergence to complete the training of the muscle movement anomaly analysis model.

[0069] S61. The abnormal scoring index is used to analyze abnormal muscle movement through a two-layer multilayer perceptron.

[0070] In this embodiment, a muscle movement anomaly analysis model is constructed. The model is input into a muscle movement dataset and sequentially progresses through multiple stages: muscle co-external offset tensor, co-embedding features, offset field coupling matrix, and anomaly scoring index. Based on this, channel coupling weights, perturbation adjustment factors, perturbation weight vectors, asymmetric adjustment factors, channel offset factors, and channel direction weight coefficients are introduced. To achieve a nonlinear discriminative mapping from the channel scoring vector to the anomaly probability, this step constructs a multilayer perceptron with two-layer mapping capabilities. Embedding space projection is achieved by combining ReLU linear transformation with a Sigmoid nonlinear activation function, ultimately yielding the results of the muscle movement anomaly analysis. .

[0071] S62. The muscle movement abnormality analysis model is trained based on the cross-entropy loss function of probability discrimination to update the hyperparameters of the muscle movement abnormality analysis model.

[0072] In this embodiment, during the training process after the abnormality scoring index is mapped to the deep layer, to ensure the consistency between the predicted output and the actual abnormal state, a cross-entropy loss function based on probability discrimination is adopted. Specifically, the probability value output by the muscle movement abnormality analysis model is compared with the actual label sample by sample, and the mathematical model of the loss function is constructed as follows: ; in, The output of the muscle movement abnormality analysis model. This is the label of the actual abnormal state corresponding to the current sample. The loss function value is used as the target for backpropagation to update the hyperparameters.

[0073] Furthermore, the muscle movement abnormality analysis model proposed in this invention is implemented using the Python programming language and modeled and trained based on the PyTorch deep learning framework. During model training, the Adam optimizer is used to iteratively update the model parameters, with the initial learning rate set to 0.001. A fixed step size decay strategy is used to dynamically adjust the learning rate to adapt to the model's convergence trend. The training batch size is set to 64, and the total number of training epochs is 2000. The loss function is a probabilistic discriminant loss structure based on cross-entropy, and a residual modulation term is embedded on the abnormality scoring index to enhance the training sensitivity to edge samples. At the same time, the abnormality scoring index is mapped at the output end through two layers of multilayer perceptron, combined with the Sigmoid activation function to achieve a stable output of the abnormality probability, effectively improving the model's training robustness and classification ability for multi-channel deviation responses during rehabilitation.

[0074] Furthermore, the collected muscle movement dataset is input into the constructed muscle movement anomaly analysis model for training and validation. The convergence trend of the joint loss during model training is as follows: Figure 6 As shown, with the increase of training rounds, the loss value of the muscle movement anomaly analysis model continuously decreases from an initial high level and gradually stabilizes, indicating that the proposed muscle movement anomaly analysis model structure has good convergence characteristics during training and can effectively complete the modeling of perturbation features and accurate estimation of anomaly scores. To further verify the model's discriminative ability in anomaly identification, the t-SNE algorithm is used to perform visualization dimensionality reduction on the high-dimensional bias intensity tensor, and the results are shown below. Figure 7 As shown in the figure, different categories of samples form clear boundaries in the two-dimensional embedding space, demonstrating the significant advantage of the constructed offset field coupling mechanism and Gaussian perturbation response joint modeling capability in sample discriminability; the prediction results are as follows. Figure 8As shown in the figure, the horizontal axis represents the time step, and the vertical axis represents the anomaly score. In the figure, circles represent the anomaly level corresponding to the real label, and triangles represent the model prediction results. It can be observed from the figure that the overall consistency between the output of the muscle movement anomaly analysis model and the real label is high, and the constructed anomaly score index can accurately locate edge samples. Compared with existing methods, the proposed analysis model shows superior performance in terms of anomaly detection accuracy, response speed, and collaborative discrimination ability between cross-channel muscle groups. It can adapt to the diverse activation pattern perturbations in different rehabilitation stages and ensure the reliability and consistency of anomaly monitoring and evaluation results in the rehabilitation process.

[0075] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements are all within the scope of protection of the present invention.

Claims

1. A method for analyzing muscle movement abnormalities during the rehabilitation process, characterized in that, Includes the following steps: Collect muscle movement datasets during rehabilitation testing, and preprocess the collected datasets to construct a muscle movement dataset. The muscle motion dataset is divided into muscle motion vectors, the cooperative relationship between muscle groups is calculated to obtain a dynamic adjacency matrix sequence, and the muscle cooperative offset tensor is obtained based on the structural change rate matrix. Based on the muscle cooperative offset tensor, the asymmetric cooperative offset rate is weighted and normalized to construct a cooperative offset matrix, and the temporal curvature is obtained by extracting continuous differences. The cooperative offset matrix and the temporal curvature are weighted and superimposed, and channel coupling weights and perturbation adjustment factors are introduced to obtain the energy perturbation matrix and accumulate calculations to construct cooperative embedding features. Based on the collaborative embedding feature, a perturbation weight vector is introduced to obtain an adaptive adjacency matrix, which is then used to construct a perturbation operator matrix with an asymmetric adjustment factor. The matrix information from historical moments is fused to generate a tensor interactive superposition result. A channel offset factor is introduced for differential coupling to form an offset field coupling matrix. By exponentially mapping the channel offset intensity to the offset field coupling matrix and the Gaussian perturbation kernel function, and fusing the channel differences, a deviation sensitivity response vector is generated. By combining the channel direction weight coefficient and the residual direction projection component, a deviation intensity tensor is generated and jointly modeled to obtain an anomaly scoring index. Input a muscle movement dataset and combine it with a loss function to construct a muscle movement anomaly analysis model. Iterate the training until convergence to complete the training of the muscle movement anomaly analysis model.

2. The method for analyzing muscle movement abnormalities in the rehabilitation process according to claim 1, characterized in that, During the data acquisition phase, patients at different stages of rehabilitation were selected as the subjects of the test, and data including raw electromyography signals, electromyography energy, root mean square value, zero crossing rate, average frequency, median frequency, triaxial acceleration, triaxial angular velocity, and joint angle characteristics were generated. In the data preprocessing stage, baseline correction is performed on the collected data; noise reduction is performed on the electromyography (EMG) signals; smoothing and zero-point calibration are performed on the acceleration and angular velocity signals; time synchronization is performed on the collected data and the synchronized signals are divided into time slices; statistical features of EMG, acceleration, and angular velocity are extracted in each time slice. The processed features are organized into a muscle motion dataset according to time sequence and feature dimensions.

3. The method for analyzing muscle movement abnormalities in the rehabilitation process according to claim 2, characterized in that, The data collected at each time point is divided into muscle motion vectors, which are then combined according to fixed time steps to form a muscle motion signal matrix. Within each time step, the multi-channel muscle motion data features within the muscle motion vector are paired, the co-activation intensity between the channel pairs is calculated, and a stability constant is introduced during the calculation process to avoid numerical calculation errors. Finally, the co-activation intensity of each channel pair is concatenated according to the index position to obtain the muscle group adjacency matrix of the current time step, and a dynamic adjacency matrix sequence is formed through iterative calculation within the entire sampling window. Based on the dynamic adjacency matrix sequence, the difference in co-activation intensity at different times is calculated to obtain the structural change rate. The difference component between adjacent times is calculated through the structural change rate to obtain the co-offset change difference of each channel pair. Based on the co-offset change difference and the consistency sign function, the asymmetric co-offset rate used to distinguish the direction of co-offset change is obtained. The asymmetric co-offset rates at each time are concatenated in sequence to form a complete muscle co-offset tensor.

4. The method for analyzing muscle movement abnormalities in the rehabilitation process according to claim 3, characterized in that, Based on the muscle synergistic offset tensor, the asymmetric synergistic offset rate between channels is extracted. The influence of different muscle groups during rehabilitation training is combined with the channel pair weighting factor, and the asymmetric synergistic offset rate is jointly weighted by the time weighting factor set with time decay. The weighted results are normalized based on the global maximum and minimum values ​​of the muscle co-offset tensor to construct the co-offset matrix; Based on the cooperative offset matrix, second-order calculations are performed on the continuous differences of adjacent time steps, and then normalized to construct a time-series curvature sequence. Based on the collaborative offset matrix and temporal curvature, the offset difference terms of adjacent time steps are extracted. The temporal curvature is used as a calculation component and is weighted and superimposed with the offset difference terms according to a preset weight relationship. During the weighting process, a channel coupling weight factor is introduced to distinguish the sensitivity of different channels in the calculation. At the same time, a perturbation adjustment factor is combined to balance the relative weight of curvature components and differential components, thus obtaining the energy perturbation matrix. A cumulative bias embedding method is constructed by using the energy perturbation matrix. Based on the historical average energy perturbation level, the energy perturbation of each channel pair is weighted and accumulated within a fixed time window W to obtain the cumulative bias result, which is then aggregated into a collaborative embedding feature.

5. The method for analyzing muscle movement abnormalities in the rehabilitation process according to claim 4, characterized in that, By jointly modeling the collaborative feature relationship between any channel pairs and introducing a perturbation weight vector to dynamically adjust the response degree between channels, combined with a set of adjustment mechanisms to enhance numerical stability, a weight value is formed to characterize the strength of the relationship between channels, and an adaptive adjacency matrix is ​​constructed. Based on the adaptive adjacency matrix, the perturbation operator matrix is ​​formed to characterize the bidirectional cooperative strength between channels by differentiating the strength of the cooperative strength and introducing an asymmetric adjustment factor to regulate the relative weights between different computational components. By perturbating the operator matrix and combining multi-level matrix information from historical moments, tensor outer product operations and element-wise multiplication operations are introduced at the matrix level to establish cross-time interaction structures and historical association paths. At the same time, historical memory weights are set to regulate the degree of jointness between current and past operators, thus constructing tensor interaction superposition results. Based on the channel dimension of the tensor interaction superposition result, a channel offset factor is introduced for the corresponding elements at consecutive time steps to perform differential coupling operation. Based on the changing trend of tensors in each channel dimension, and combined with the tensor mapping results of the current and previous time steps, cross-dimensional responsivity mapping is completed by relying on multiple independent offset factors to generate an offset field coupling matrix.

6. The method for analyzing muscle movement abnormalities in the rehabilitation process according to claim 5, characterized in that, Based on the offset field coupling matrix, the perturbation performance of each channel dimension is extracted. A Gaussian perturbation kernel function with bandwidth adjustment capability is introduced to perform exponential mapping of the channel offset intensity. Combined with the activation difference between the standard template activation mode and the patient's actual activation mode in the corresponding channel dimension, the deviation sensitivity response vector is generated dimension by dimension. Channel dimension fusion is performed by combining the bias sensitivity response vector and the channel direction weight coefficient. Then, the difference direction between the reference template activation value and the patient activation value is biased and encoded by combining the sign function. Multi-channel direction projection components are further superimposed to complete the weighted normalization operation of the residual direction tensor, and finally the bias intensity tensor of multi-dimensional channel bias response intensity is generated. Based on the deviation intensity value at the current moment, a normalization construction strategy with local logarithmic amplitude characteristics is introduced, which is combined with the deviation intensity value at the previous moment to form an anomaly scoring index.

7. The method for analyzing muscle movement abnormalities in the rehabilitation process according to claim 6, characterized in that, The abnormality scoring index is used to analyze abnormal muscle movement through a two-layer multilayer perceptron. The muscle movement anomaly analysis model is trained using a cross-entropy loss function based on probability discrimination, thereby updating the hyperparameters of the muscle movement anomaly analysis model.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the muscle movement abnormality analysis method for rehabilitation process as described in any one of claims 1 to 7.

Images