Intelligent closed-loop rehabilitation method and system based on multi-source information fusion fatigue evolution prediction and adaptive regulation

Abstract

This invention discloses an intelligent closed-loop rehabilitation method and system based on multi-source information fusion, fatigue evolution prediction, and adaptive regulation. It collects inertial motion signals and electromyographic signals through multimodal wearable sensors. After signal preprocessing and time synchronization, a multidimensional correlation-sensory attention coding mechanism and a gated recurrent network are used to perform feature encoding and future signal prediction on the multidimensional time-series signals. Based on the predicted signals, a multidimensional fatigue index system is constructed to achieve continuous, feedforward assessment of fatigue states during training. Furthermore, a fuzzy mapping method is used to map the multidimensional fatigue states into four risk levels, and training parameters are dynamically adjusted according to the risk level to form real-time, intelligent closed-loop training control. This invention can improve the scientific, precise, and individualized level of rehabilitation training while ensuring training safety, and is particularly suitable for rehabilitation training of patients with neurological diseases and those with impaired motor function.

Description

Technical Field

[0001] This invention belongs to the field of intelligent health monitoring technology, and in particular relates to an intelligent closed-loop control method and system that integrates wearable motion sensing and surface electromyography signals, training fatigue evolution prediction based on deep learning, and adaptive training regulation. It is applicable to rehabilitation training assessment and safe intervention for patients with nervous system diseases and people with impaired motor function. Background Technology

[0002] In rehabilitation training and functional reconstruction, rationally controlling training intensity and load is crucial for ensuring rehabilitation effectiveness and preventing sports injuries. This is especially true for patients with stroke, traumatic brain injury, and other neurological injuries, whose muscle endurance and fatigue recovery abilities are significantly lower than healthy individuals. During training, they are highly susceptible to fatigue accumulation, compensatory movements, and even secondary injuries. Current rehabilitation training primarily relies on the therapist's experience or the patient's subjective fatigue feedback, lacking objective, continuous, and predictable assessment methods for muscle fatigue. Some studies attempt to assess fatigue using single kinematic or electromyographic signals, but these largely remain at the post-hoc analysis stage, making it difficult to predict fatigue evolution trends in real time and intervene early during training. Furthermore, existing technologies mostly focus on single-moment assessments of fatigue, lacking systematic modeling of the fatigue evolution process, cumulative effects, and their impact on training risks. An integrated intelligent closed-loop control mechanism of "monitoring-prediction-regulation" has not yet been established, failing to meet the practical needs of precise rehabilitation and personalized training. Summary of the Invention

[0003] The purpose of this invention is to address the problems mentioned in the background technology by proposing an intelligent closed-loop rehabilitation method and system based on multi-source information fusion for fatigue evolution prediction and adaptive regulation. By fusing multi-source wearable signals, a fatigue evolution prediction model is constructed to achieve multi-dimensional assessment of instantaneous fatigue state, fatigue change trend, and fatigue accumulation risk during training. Based on the prediction results, training parameters are dynamically adjusted, thereby improving the scientific nature, accuracy, and individualization of rehabilitation training while ensuring training safety.

[0004] To achieve the objectives of this invention, a smart closed-loop rehabilitation method based on multi-source information fusion for fatigue evolution prediction and adaptive regulation is disclosed, comprising the following steps:

[0005] Step 1: Multimodal wearable signal acquisition. Using wearable inertial motion sensors and surface electromyography sensors, real-time acquisition of human kinematic signals and electromyographic signals during training is performed.

[0006] Step 2: Multi-source signal preprocessing. The acquired multi-modal signals are denoised, normalized, time-synchronized, and segmented using a sliding window to construct standardized time-series input data.

[0007] Step 3: Signal prediction. Construct a multidimensional related perception attention encoding mechanism and a gated recurrent network model to model the current multimodal time-series signal and predict the motion and electromyographic signals in the future.

[0008] Step 4: Multidimensional fatigue calculation. Based on the predicted future signals, calculate the instantaneous fatigue value, fatigue change rate, and fatigue accumulation index at future times to form a multidimensional fatigue state vector.

[0009] Step 5: Construct fatigue risk classification and adaptive training control rules. Use fuzzy mapping to classify multidimensional fatigue states into different risk levels and dynamically adjust training parameters according to the risk level to achieve intelligent closed-loop control of the training process.

[0010] Furthermore, in step 1, the multimodal wearable signal acquisition specifically includes the following steps:

[0011] Step 1-1, Wearing the device; The wearable device is worn by installing c1 inertial sensors on the corresponding parts of the patient's body and c2 surface electromyography sensors on the surface of the main muscle groups related to the target training.

[0012] Steps 1-2: Data Acquisition; Turn on the wearable device, perform sensor calibration, check accuracy and correct any possible offsets; Start the data acquisition program, the patient wears the wearable device to begin training or exercise, and the system records the motion data of various movements during the patient's rehabilitation training.

[0013] Steps 1-3: Data storage and transmission: The collected multimodal data can be stored in the local storage unit of the wearable device or in a cloud database; in some application scenarios, the data is transmitted to the data analysis unit in real time via wireless communication for online fatigue assessment and training regulation.

[0014] The dataset for building the deep learning model was also collected following the above process, and included labels for multidimensional fatigue levels assessed by doctors.

[0015] Furthermore, in step 2, the multi-source signal preprocessing specifically includes the following steps:

[0016] Step 2-1: Inertial Motion Signal IMU Preprocessing; The attitude angle, acceleration, and angular velocity signals acquired by the inertial sensor are preprocessed as follows:

[0017] Missing and outlier handling: When a local time segment is missing in the signal, it is filled by linear interpolation or mean compensation of the adjacent signals before and after the missing segment; when the entire signal of a single sensor is missing within a certain time period, historical samples with the same training action and the same fatigue level are selected from the database for replacement; outliers that exceed the reasonable physiological range are detected and corrected by neighborhood smoothing.

[0018] Signal denoising and smoothing: To reduce the impact of random noise in inertial signals on model modeling, the original IMU signal is filtered, specifically including: using moving average filtering or low-pass filtering on acceleration and angular velocity signals to suppress high-frequency noise; and smoothing the attitude angle signal to avoid interference from instantaneous jitter on timing modeling.

[0019] Time synchronization and standardization: To ensure the consistency of signals acquired by different sensors on the time axis, time alignment processing is performed on the signals of each IMU channel; at the same time, normalization or standardization processing is performed on the signals of each channel to eliminate the impact of differences in units and amplitudes on model training.

[0020] Step 2-2: Preprocessing of surface electromyography (sEMG) signals; Considering the high-frequency and noise-sensitive characteristics of surface electromyography signals, the following preprocessing operations are performed:

[0021] Bandpass filtering and power frequency interference suppression: The acquired raw electromyography (EMG) signals are bandpass filtered to remove low-frequency motion artifacts and high-frequency noise, and power frequency interference is suppressed to obtain stable and effective EMG signals.

[0022] Signal amplitude normalization: The amplitude of the electromyographic signal is normalized to make the electromyographic signals of different subjects and different training stages comparable;

[0023] Steps 2-3: Multimodal signal alignment and sliding window segmentation; After completing the independent preprocessing of inertial motion signals and surface electromyography signals, the multimodal signals undergo preprocessing for fusion, specifically including:

[0024] The inertial motion signal and the surface electromyography signal are synchronized in time according to a unified time reference, and the signals with different sampling rates are resampled to keep the two types of signals consistent in time axis and sampling frequency.

[0025] The synchronized multimodal signal is divided into a sliding window of fixed length n, with an overlap rate of 50% between adjacent windows.

[0026] Each sliding window contains complete temporal features of motion and electromyography, which serve as input samples for subsequent fatigue evolution prediction models.

[0027] Steps 2-4: Preprocessing output data format; After the above preprocessing, the data within a single time window is represented as follows:

[0028]

[0029] Among them, X imu This represents the preprocessed inertial motion signal feature matrix, where c1*9 represents c1 IMU sensors, each with 9 channels, and X... emg This represents the preprocessed electromyographic signal feature matrix; the above multimodal features are used together as input to the fatigue evolution prediction model for subsequent fatigue state modeling and adaptive training regulation.

[0030] Furthermore, in step 3, the signal prediction module completes signal prediction, that is, it uses the existing signal X. imu or X emg Predicting future time series signal Y imu and Y emg And the predicted future signal length is m, i.e., Y imu ={y1,y2,...,y m}∈R (c1*9)×m Y emg ={y1,y2,...,y m}∈R c2×m To simplify the expression, the two types of signals are uniformly represented as Y={y1,y2,...,y...} m}∈R p×m , where y t ∈R p , where p represents the dimension of the signal; a gated recurrent signal prediction network with a multidimensional correlation-sensory attention encoding mechanism is constructed, which is applicable to both surface electromyography (EMG) signals and motion signals prediction. The specific steps include:

[0031] Step 3-1: Multidimensional correlation-based perceptual attention encoding; By constructing a sequence-time correlation attention and channel contribution perceptual attention mechanism, the key temporal dependencies and important channel features in the signal are adaptively weighted and encoded to improve feature representation ability;

[0032] Step 3-2: Gated Recurrent Unit Extracts Spatiotemporal Features; The attention-encoded temporal signal is input into the Gated Recurrent Unit (GRU) network, and the temporal dynamic change features and long-term dependencies in the signal are extracted through the gating update mechanism;

[0033] Step 3-3: Multi-step prediction; Based on the global time series features extracted by the GRU network, an autoregressive iterative prediction method is used to gradually generate signals for multiple future time steps, thereby achieving forward-looking prediction of the signals.

[0034] Furthermore, in step 3-1, for the multimodal data within a single time window after preprocessing... For any modality of data in the dataset, we simplify it to represent X:

[0035]

[0036] Where, x r ∈R p This represents the multidimensional feature vector at the r-th time step, where n represents the length of the time series and p is the feature dimension; specifically, it includes the following steps:

[0037] Step 3-1-1: Calculation of temporal correlation attention weights in sequences;

[0038] First, a sequential temporal correlation attention mechanism is constructed for the input signal X to characterize the influence of different historical time steps on the evolution of the current fatigue state; then, a query vector Q is generated through linear mapping. T and bond vector K T The calculation method is as follows:

[0039]

[0040] Among them, W QT ∈R d×p and W KT ∈R d×p This is a trainable spatial attention weight matrix;

[0041] To enhance the ability to perceive signal change trends, a time step variation ΔX∈R is introduced. p×n :

[0042]

[0043] And construct the trend consistency term C∈R n×n :

[0044]

[0045] Among them, W Δ ∈R d×p The feature mapping matrix is ​​a variable.

[0046] Taking into account the above factors, the sequence association attention weight matrix A T Defined as:

[0047]

[0048] Where α is the trend consistency weight coefficient, d represents the feature dimension, and softmax is used to normalize the attention weights.

[0049] Step 3-1-2: Calculation of channel contribution perceived attention weights;

[0050] Based on sequence association modeling, a channel contribution perception attention mechanism is introduced to further distinguish the differences in contributions of different sensor channels and feature dimensions in the fatigue evolution process.

[0051] By performing a linear mapping of the input features along the channel dimension, a query vector Q along the channel dimension is generated. S and bond vector K S :

[0052]

[0053] Among them, W QS ∈R d×d W KS ∈R d×d W is a trainable spatial attention weight matrix. VT ∈R d×p Let be a trainable dimensionality transformation matrix; considering the different noise levels of electromyography and motion signals, to reduce the impact of high-noise channels on attention calculation, channel uncertainty weights are introduced, and the confidence factor for each channel is defined as:

[0054]

[0055] MLP stands for Multilayer Perceptron, which is used to estimate the noise level of each time step and each channel.

[0056] Based on this, the channel contribution perception attention weight matrix A S Defined as:

[0057]

[0058] Where ⊗ represents element-wise multiplication;

[0059] Step 3-1-3: Multi-dimensional attention feature fusion;

[0060] After obtaining the sequence association attention weight matrix A T And channel contribution perception attention weight matrix A S Then, the input signals are weighted and fused; first, a linear transformation is performed on the input signals to obtain the value vector V:

[0061]

[0062] Among them, W VT ∈R d×p is a trainable weight matrix;

[0063] Subsequently, the sequence association attention weight matrix A TAnd channel contribution perception attention weight matrix A S Together, they act on the value vector to construct the feature representation after multidimensional relevance-aware attention encoding:

[0064]

[0065] The final input signal obtained after multidimensional correlation perceptual attention encoding is:

[0066]

[0067] Where, x r ′∈R d Let represent the r-th time-step feature vector fused with multidimensional related perceptual attention.

[0068] Furthermore, in step 3-2, the input sequence X after multidimensional related perceptual attention encoding... ST The input is fed into a bidirectional gated recurrent unit network (GRU1) for historical temporal feature extraction to establish a bidirectional temporal dependency between historical motion and electromyographic signals. The input feature dimension of the GRU1 unit is d, the hidden layer dimension is h, and the gated recurrent unit operation is represented as follows:

[0069]

[0070] Among them, h f ∈R h Indicates the hidden state at the current time step; h f-1 This is the hidden state from the previous time step; x f ′ ∈R d This represents the feature vector at the f-th time step that incorporates multidimensional related perceptual attention.

[0071] For an input timing signal X of length n ST After passing through the bidirectional GRU network GRU1, the forward hidden state sequence H1 and the reverse hidden state sequence H2 are obtained respectively:

[0072]

[0073]

[0074] In this context, superscripts (1) and (2) represent networks propagating in the forward and backward directions, respectively.

[0075] The hidden states at the last time step are fused to obtain the comprehensive feature representation h of the sequence. n :

[0076]

[0077] Among them, hn (1) This represents the hidden state at the last time step of the forward GRU; h n (2) Represents the hidden state at the last time step of the reverse GRU; [;] represents vertical concatenation; Wc∈R h×2h b is the fusion weight matrix; c ∈R h This is the fusion bias; h is the length of the hidden layer; h n The global time series feature representation corresponding to the historical time window is used for multi-step prediction of subsequent future signals.

[0078] Furthermore, in step 3-3, based on the historical global temporal features h extracted in step 3-2... n A unidirectional autoregressive future signal prediction network, GRU2, is constructed to progressively generate and predict motion and electromyographic signals within future time windows, thereby achieving feedforward modeling of fatigue evolution trends. The specific steps include:

[0079] Step 3-3-1, Multi-step prediction initialization;

[0080] In the multi-step prediction process, an iterative prediction strategy based on autoregression is adopted; the historical global time series feature vector h n As the initial hidden state s0 of the future prediction network GRU2:

[0081]

[0082] The first prediction for the future time step is generated through a fully connected layer:

[0083]

[0084] Among them, W fc ∈R p×h Let b represent the weight matrix of the fully connected layer. fc ∈R p Indicates the corresponding bias term;

[0085] Step 3-3-2: Prediction Result Feedback and Iterative Generation

[0086] The prediction result y1 at the current time step is used as the input for the next time step and fed into the one-way GRU prediction network GRU2 to recursively and dynamically generate the future sequence:

[0087]

[0088] Subsequently, the predicted value for the next time step is generated through the same fully connected layer:

[0089]

[0090] Repeat the above prediction process to gradually generate prediction results for multiple future time steps, namely:

[0091]

[0092]

[0093] Where m represents the prediction time step length;

[0094] Step 3-3-3: Multi-step prediction output;

[0095] The GRU2 one-way autoregressive future signal prediction network is used to progressively generate a predicted sequence Y for the next m time steps:

[0096]

[0097] It is used to characterize the evolution trend of the signal and provide forward-looking input for subsequent fatigue evolution assessment and training regulation.

[0098] Furthermore, in step 4, the fatigue calculation module is used to assess the fatigue state of the future motion and electromyographic signals obtained from multi-step predictions, thereby constructing a multi-dimensional fatigue index system that reflects the fatigue evolution characteristics during training. Based on the predicted signals, the fatigue calculation module calculates the instantaneous fatigue value, fatigue change rate, and fatigue accumulation trend, respectively, for subsequent training risk assessment and adaptive control, specifically including the following steps:

[0099] Step 4-1: Calculation of instantaneous fatigue;

[0100] At time step k, based on the preprocessed inertial motion signal feature matrix X imu and the preprocessed electromyographic signal feature matrix X emg Each signal prediction network is used independently to predict the motion signal and electromyographic signal with a length of n at time step k in the future, including features from m consecutive time steps as the input window for fatigue calculation.

[0101]

[0102]

[0103] Where m represents the length of the time window used for fatigue estimation; this input window is used to reflect the comprehensive changes in motion and electromyographic signals over a future period of time, thereby avoiding the random influence of a single time point signal on fatigue judgment;

[0104] The two are concatenated to obtain the complete representation Y of the predicted signal. k :

[0105]

[0106] The future multimodal prediction signal Y k The data is input into a fatigue mapping network to establish a mapping relationship between future motion, electromyographic changes, and fatigue state. The fatigue mapping network GRU3 models the dynamic changes within a future time window step-by-step.

[0107]

[0108] Among them, u i Let u0 represent the hidden state vector corresponding to the i-th future time step, i∈{1,2,...,m}, u0 is a unit vector, and h2 represents the dimension of the GRU3 hidden layer;

[0109] Then, the hidden state u i The input is fed into a fully connected layer, and the instantaneous fatigue value at time step k+i is calculated:

[0110]

[0111] Among them, W g ∈R 1×h2 Let b be the weight matrix of the fully connected layer. g ∈R is the bias term, F k+i This represents the estimated instantaneous fatigue level at time step k+i in the future.

[0112] By repeating the above calculation process for each time step within the prediction time window, a future fatigue sequence F containing m time points is gradually generated:

[0113]

[0114] This fatigue sequence is used for calculating the subsequent fatigue change rate and fatigue accumulation trend, as well as for training risk assessment and adaptive control.

[0115] Step 4-2: Calculation of fatigue change rate;

[0116] The fatigue change rate is used to characterize the trend of fatigue state over time, and is used to determine whether fatigue is in a rapid accumulation, stable, or recovery phase; the fatigue change rate is defined as follows:

[0117] Given a prediction step size s, the fatigue change rate G at time a a Defined as:

[0118]

[0119] Among them, F a+s This represents the fatigue value at the predicted time step a+s;

[0120] When G a When G > 0, it indicates that the fatigue level is on the rise; when G a When G < 0, it indicates a downward trend in fatigue level; when G a When the value is approximately 0, it indicates that the fatigue state is relatively stable.

[0121] Step 4-3: Calculation of fatigue accumulation trend;

[0122] The cumulative fatigue tendency (AFI) is used to describe the cumulative effect of fatigue over a complete training cycle, reflecting the overall load level on the body caused by continuous training. The cumulative fatigue index (AFI) at time a is defined as:

[0123]

[0124] Among them, w i The weighting coefficients for the corresponding time steps are used to adjust the influence of fatigue in the cumulative calculation at different time periods. The weights are set according to the training stage or clinical experience. The larger the fatigue accumulation index, the higher the fatigue load in the training cycle, and the greater the risk of overtraining or sports injury.

[0125] Step 4-4: Constructing a multidimensional fatigue state vector;

[0126] Based on the above calculations, a single-point multidimensional fatigue state vector can be constructed at any future time point t:

[0127]

[0128] This multidimensional fatigue state vector serves as the core input to the fatigue risk assessment and adaptive training control module, enabling dynamic monitoring and intelligent intervention of the training process.

[0129] Furthermore, in step 5, the multi-point-dimensional fatigue state vector obtained from the fatigue calculation module is as follows:

[0130]

[0131] By constructing a fatigue risk grading mechanism and formulating adaptive training control rules accordingly, intelligent closed-loop control of the training process is achieved, specifically including the following steps:

[0132] Step 5-1: Construct a fatigue risk classification mechanism;

[0133] Training risks are assessed in a graded manner based on instantaneous fatigue values, fatigue change rates, and fatigue accumulation trends.

[0134] The following steps are used to fuzzily map the multidimensional fatigue state vector to four fatigue risk levels: safe zone, warning zone, risk zone, and prohibited zone:

[0135] Step 5-1-1: Indicator normalization;

[0136] Linear normalization is performed on each indicator to unify its range to [0,1], resulting in the normalized index. :

[0137]

[0138] in This represents the normalized instantaneous fatigue value, fatigue rate of change, and fatigue accumulation trend.

[0139] Step 5-1-2: Membership function design;

[0140] Fatigue instantaneous value F k The membership functions are as follows:

[0141]

[0142]

[0143]

[0144]

[0145] in, The membership function representing the instantaneous fatigue value under the safety level; The membership function representing the instantaneous fatigue value under the warning level; The membership function representing the instantaneous fatigue value under the risk level; The membership function representing the instantaneous fatigue value under the prohibition level;

[0146] Fatigue change rate G k The membership functions are as follows:

[0147]

[0148]

[0149]

[0150]

[0151] in, The membership function representing the fatigue rate of change under the safety level; The membership function representing the fatigue rate of change under the warning level; The membership function representing the fatigue rate of change under the risk level; The membership function representing the fatigue rate of change under the prohibition level;

[0152] AFI (Accumulated Fatigue Index) k The membership functions are as follows:

[0153]

[0154]

[0155]

[0156]

[0157] in, The membership function representing the fatigue accumulation index under the safety level; The membership function representing the fatigue accumulation index under the warning level; The membership function representing the fatigue accumulation index under the risk level; The membership function representing the fatigue accumulation index under the prohibition level;

[0158] Step 5-1-3: Weighted summarization of multiple indicators;

[0159] To synthesize the fuzzy membership degrees of each indicator, the weighted comprehensive membership degree S(e) for each level is calculated:

[0160]

[0161] Where e∈{safety, vigilance, risk, prohibition}, w F ,w G ,w A ∈[0,1] are weight coefficients, satisfying w F +w G +w A =1, used to balance the contribution of different indicators to risk assessment;

[0162] Step 5-1-4: Determine the maximum membership degree;

[0163] Finally, the risk level E corresponding to the maximum value is selected based on the overall membership degree:

[0164]

[0165] Wherein, argmax represents obtaining an optimal value of e such that S(e) reaches its maximum value; thereby realizing the fuzzy mapping from the multidimensional fatigue state vector to the four-level risk level;

[0166] Step 5-1-5: Calculation of final fatigue risk level;

[0167] For the m future points starting from time step k, calculate the fatigue risk level for each point sequentially to obtain a risk level sequence:

[0168]

[0169] Among them: E k+i This represents the fatigue risk level corresponding to the i-th time step in the future; m represents the length of the future prediction time window; the risk levels include: Level 1, i.e., the safe zone; Level 2, i.e., the warning zone; Level 3, i.e., the risk zone; Level 4, i.e., the prohibited zone; to ensure safety during training, the principle of prioritizing the highest risk is adopted, and the risk levels within the future prediction window are comprehensively evaluated, and the highest risk level in the future time window is taken as the final fatigue risk level:

[0170]

[0171] Among them: E final This indicates the final fatigue risk level; max indicates taking the maximum value. In this way, the system can identify the highest fatigue risk state that may occur within a future time window in advance, thereby providing a basis for the feedforward adaptive adjustment of subsequent training parameters.

[0172] Step 5-2: Construct an adaptive adjustment strategy for training parameters;

[0173] For different fatigue risk levels, training parameters are adjusted in stages, including but not limited to: training movement amplitude, number of repetitions, movement execution rhythm, training load intensity, and training interval time.

[0174] Step 5-3: Feedforward regulation based on prediction results;

[0175] By utilizing the future fatigue state obtained through multi-step prediction, the system implements control strategies in advance before fatigue actually reaches a high-risk level, thus achieving feedforward training control. By identifying the risk of rapid fatigue growth or accumulation in advance, the system can adjust training parameters before fatigue is fully formed, thereby effectively reducing the risk of overtraining.

[0176] Step 5-4: Train and regulate the closed-loop update mechanism;

[0177] During training, the system continuously collects real-time motion and electromyographic signals and constantly updates fatigue prediction results and risk levels; the training control module dynamically adjusts training parameters based on the latest fatigue status, forming a closed-loop control process of "monitoring-prediction-evaluation-control".

[0178] To achieve the objectives of this invention, this invention also discloses an intelligent closed-loop rehabilitation system based on multi-source information fusion for fatigue evolution prediction and adaptive regulation, comprising the following modules:

[0179] The multimodal wearable signal acquisition module utilizes wearable inertial motion sensors and surface electromyography sensors to acquire human kinematic signals and electromyography signals in real time during training.

[0180] The multi-source signal preprocessing module performs noise reduction, normalization, time synchronization, and sliding window segmentation on the acquired multi-modal signals to construct standardized time-series input data.

[0181] The signal prediction module constructs a multidimensional related perception attention encoding mechanism and a gated recurrent network model to model the current multimodal time-series signal and predict the motion and electromyographic signals in the future.

[0182] The multidimensional fatigue calculation module calculates the instantaneous fatigue value, fatigue change rate, and fatigue accumulation index at future time based on the predicted future signals, forming a multidimensional fatigue state vector.

[0183] A fatigue risk classification and adaptive training control rule module is constructed. The multidimensional fatigue state is divided into different risk levels by using the fuzzy mapping method, and the training parameters are dynamically adjusted according to the risk level to achieve intelligent closed-loop control of the training process.

[0184] Compared with existing technologies, the significant advancements of this invention are: 1) Constructing a multidimensional correlation-sensory attention encoding mechanism combined with a gated recurrent network model; innovatively constructing a multidimensional correlation-sensory attention encoding mechanism to extract the temporal correlation and channel contribution of multimodal signals, and combining it with a gated recurrent unit for temporal feature extraction, thereby achieving accurate characterization of fatigue evolution by the predicted signal; 2) Integrating temporal prediction and fatigue prediction modules; this invention integrates a temporal signal prediction module and a fatigue calculation module within the same framework. The temporal prediction module predicts future motion signals through a GRU network, while the fatigue calculation module calculates the corresponding fatigue level based on the predicted future signals. This allows for early prediction of future fatigue levels before fatigue occurs, thus avoiding overtraining and injury; 3) Constructing a multidimensional fatigue index system; quantifying fatigue state into three-dimensional indicators: instantaneous fatigue value Ft; fatigue change rate Gt; and fatigue accumulation index AFI. A sliding time window GRU model is used to map the predicted signal into a multidimensional fatigue sequence, achieving continuous, feedforward fatigue monitoring; 4) Risk level determination using fuzzy mapping. Using fuzzy logic algorithms, multidimensional fatigue state vectors are mapped to four levels of fatigue risk (safe zone, warning zone, risk zone, and prohibited zone). Combined with adjustable weights, individualized risk assessment is achieved, providing a quantitative basis for training intervention; 5) Adaptive training closed-loop control: Based on the feedforward regulation strategy of predicting fatigue and risk levels, dynamic adjustment of training movement amplitude, repetition count, load resistance, and interval time is achieved, forming a closed-loop control of training-monitoring-prediction-regulation, which significantly improves the safety and personalization of rehabilitation training.

[0185] To more clearly illustrate the functional characteristics and structural parameters of the present invention, further explanation is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description

[0186] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0187] Figure 1 This is a schematic diagram of the framework of an intelligent closed-loop rehabilitation system based on multi-source information fusion for fatigue evolution prediction and adaptive regulation.

[0188] Figure 2 This is a schematic diagram of the signal prediction network and instantaneous fatigue calculation. Detailed Implementation

[0189] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0190] A smart closed-loop rehabilitation method based on multi-source information fusion for fatigue evolution prediction and adaptive regulation includes the following steps:

[0191] Step 1: Multimodal wearable signal acquisition. Using wearable inertial motion sensors and surface electromyography sensors, real-time acquisition of human kinematic signals and electromyographic signals during training is performed.

[0192] Step 2: Multi-source signal preprocessing. The acquired multi-modal signals are denoised, normalized, time-synchronized, and segmented using a sliding window to construct standardized time-series input data.

[0193] Step 3: Signal prediction. Construct a multidimensional related perception attention encoding mechanism and a gated recurrent network model to model the current multimodal time-series signal and predict the motion and electromyographic signals in the future.

[0194] Step 4: Multidimensional fatigue calculation. Based on the predicted future signals, calculate the instantaneous fatigue value, fatigue change rate, and fatigue accumulation index at future times to form a multidimensional fatigue state vector.

[0195] Step 5: Construct fatigue risk classification and adaptive training control rules. Use fuzzy mapping to classify multidimensional fatigue states into different risk levels and dynamically adjust training parameters according to the risk level to achieve intelligent closed-loop control of the training process.

[0196] Specifically, in step 1, the multimodal wearable signal acquisition includes the following steps:

[0197] Step 1-1: Device Wearing; The wearable device is worn by installing c1 inertial sensors on corresponding parts of the patient's body (including key locations such as the forearm and upper arm of the upper limb, the thigh and lower leg of the lower limb, and the waist and navel), and c2 surface electromyography sensors, which are placed on the surface of the main muscle groups related to the target training (key muscles of the upper limb including but not limited to the deltoid, biceps brachii, triceps brachii, brachioradialis, and extensor digitorum muscles; key muscles of the lower limb including but not limited to the gluteus maximus, rectus femoris, biceps femoris, tibialis anterior, and gastrocnemius muscles).

[0198] Steps 1-2: Data Acquisition; Turn on the wearable device, perform sensor calibration, check accuracy and correct any possible offsets to ensure the generated data is reliable and accurate; Start the data acquisition program, the patient wears the wearable device to begin training or exercise, and the system records the motion data of various movements during the patient's rehabilitation training.

[0199] The acquired motion signals (IMU) include, but are not limited to: attitude angle signals: pitch angle, yaw angle, roll angle; acceleration signals: X, Y, Z axes; angular velocity signals: X, Y, Z axes. Each inertial sensor acquires a total of 9 channels of data.

[0200] The collected electromyographic signals include: raw electromyographic time-domain signals.

[0201] The final acquired motion signals and electromyographic signals are represented in matrix form. The motion signals acquired by the inertial sensor are represented using X... imu This indicates that its data dimensions are:

[0202]

[0203] This is achieved by c1 inertial sensors, each containing 9 motion signal types with a signal length of L1. Combining the sensor dimension and the signal type dimension, it can be simplified as follows:

[0204]

[0205] Surface electromyography signals using X emg This indicates that its data dimensions are:

[0206]

[0207] That is, it is acquired by c2 electromyographic channels, and the signal length is L2;

[0208] Steps 1-3: Data storage and transmission: The collected multimodal data can be stored in the local storage unit of the wearable device or in a cloud database to ensure data security, integrity and accessibility; in some application scenarios, the data is transmitted to the data analysis unit in real time via wireless communication for online fatigue assessment and training regulation.

[0209] The dataset for building the deep learning model was also collected following the above process, and included labels for multidimensional fatigue levels assessed by doctors.

[0210] Specifically, in step 2, the multi-source signal preprocessing includes the following steps:

[0211] Step 2-1: Inertial Motion Signal IMU Preprocessing; The attitude angle, acceleration, and angular velocity signals acquired by the inertial sensor are preprocessed as follows:

[0212] Missing and outlier handling: During data acquisition, signal loss or anomalies may occur due to temporary sensor disconnection, wireless transmission interference, etc. This invention addresses different situations: When a local time segment is missing in the signal, linear interpolation or mean compensation of adjacent signals before and after the missing segment is used to fill it; when the entire signal of a single sensor is missing within a certain time period, historical samples with the same training action and fatigue level are selected from the database for replacement; outliers exceeding the reasonable physiological range are detected and corrected using neighborhood smoothing.

[0213] Signal denoising and smoothing: To reduce the impact of random noise in inertial signals on model modeling, the original IMU signal is filtered, specifically including: using moving average filtering or low-pass filtering on acceleration and angular velocity signals to suppress high-frequency noise; and smoothing the attitude angle signal to avoid interference from instantaneous jitter on timing modeling.

[0214] Time synchronization and standardization: To ensure the consistency of signals acquired by different sensors on the time axis, time alignment processing is performed on the signals of each IMU channel; at the same time, normalization or standardization processing is performed on the signals of each channel to eliminate the impact of differences in units and amplitudes on model training.

[0215] Step 2-2: Preprocessing of surface electromyography (sEMG) signals; Considering the high-frequency and noise-sensitive characteristics of surface electromyography signals, the following preprocessing operations are performed:

[0216] Bandpass filtering and power frequency interference suppression: The acquired raw electromyography (EMG) signals are bandpass filtered to remove low-frequency motion artifacts and high-frequency noise, and power frequency interference is suppressed to obtain stable and effective EMG signals.

[0217] Signal amplitude normalization: To reduce the impact of individual differences and electrode attachment differences on the amplitude of electromyographic signals, the amplitude of electromyographic signals is normalized to make the electromyographic signals of different subjects and different training stages comparable.

[0218] Steps 2-3: Multimodal signal alignment and sliding window segmentation; After completing the independent preprocessing of inertial motion signals and surface electromyography signals, the multimodal signals undergo preprocessing for fusion, specifically including:

[0219] The inertial motion signal and the surface electromyography signal are synchronized in time according to a unified time reference, and the signals with different sampling rates are resampled to keep the two types of signals consistent in time axis and sampling frequency.

[0220] The synchronized multimodal signal is divided into a sliding window of fixed length n, with an overlap rate of 50% between adjacent windows.

[0221] Each sliding window contains complete temporal features of motion and electromyography, which serve as input samples for subsequent fatigue evolution prediction models.

[0222] Steps 2-4: Preprocessing output data format; After the above preprocessing, the data within a single time window is represented as follows:

[0223]

[0224] Among them, X imu This represents the preprocessed inertial motion signal feature matrix, where c1*9 represents c1 IMU sensors, each with 9 channels, and X... emg This represents the preprocessed electromyographic signal feature matrix; the above multimodal features are used together as input to the fatigue evolution prediction model for subsequent fatigue state modeling and adaptive training regulation.

[0225] Specifically, in step 3, the signal prediction module completes signal prediction, that is, it uses the existing signal X. imu or X emg Predicting future time series signal Y imu and Y emg And the predicted future signal length is m, i.e., Y imu ={y1,y2,...,y m}∈R (c1*9)×m Y emg ={y1,y2,...,y m}∈R c2×m To simplify the expression, the two types of signals are uniformly represented as Y={y1,y2,...,y...} m}∈R p×m , where y t ∈R pp represents the dimension of the signal (p=c2 for electromyography signals, p=9*c1 for motor signals); for example Figure 2 As shown, a gated recurrent signal prediction network based on a multidimensional related perceptual attention encoding mechanism is constructed. This network is applicable to both surface electromyography (EMG) signals and motion signals. The specific steps include:

[0226] Step 3-1: Multidimensional correlation-based perceptual attention encoding; By constructing a sequence-time correlation attention and channel contribution perceptual attention mechanism, the key temporal dependencies and important channel features in the signal are adaptively weighted and encoded to improve feature representation ability;

[0227] Step 3-2: Gated Recurrent Unit Extracts Spatiotemporal Features; The attention-encoded temporal signal is input into the Gated Recurrent Unit (GRU) network, and the temporal dynamic change features and long-term dependencies in the signal are extracted through the gating update mechanism;

[0228] Step 3-3: Multi-step prediction; Based on the global time series features extracted by the GRU network, an autoregressive iterative prediction method is used to gradually generate signals for multiple future time steps, thereby achieving forward-looking prediction of the signals.

[0229] Specifically, in step 3-1, for the preprocessed multimodal data within a single time window... For any modality of data in the dataset, we simplify it to represent X:

[0230]

[0231] Where, x r ∈R p This represents the multidimensional feature vector at the r-th time step, where n represents the length of the time series, and p is the feature dimension (p=c2 for electromyographic signals, and p=9*c1 for motor signals); specifically, it includes the following steps:

[0232] Step 3-1-1: Calculation of temporal correlation attention weights in sequences;

[0233] First, a sequential temporal correlation attention mechanism is constructed for the input signal X to characterize the influence of different historical time steps on the evolution of the current fatigue state; then, a query vector Q is generated through linear mapping. T and bond vector K T The calculation method is as follows:

[0234]

[0235] Among them, W QT ∈R d×p and W KT ∈R d×p This is a trainable spatial attention weight matrix;

[0236] To enhance the ability to perceive signal change trends, a time step variation ΔX∈R is introduced. p×n :

[0237]

[0238] And construct the trend consistency term C∈R n×n :

[0239]

[0240] Among them, W Δ ∈R d×p The feature mapping matrix is ​​a variable.

[0241] Taking into account the above factors, the sequence association attention weight matrix A T Defined as:

[0242]

[0243] Where α is the trend consistency weight coefficient, d represents the feature dimension, and softmax is used to normalize the attention weights.

[0244] Step 3-1-2: Calculation of channel contribution perceived attention weights;

[0245] Based on sequence association modeling, a channel contribution perception attention mechanism is introduced to further distinguish the differences in contributions of different sensor channels and feature dimensions in the fatigue evolution process.

[0246] By performing a linear mapping of the input features along the channel dimension, a query vector Q along the channel dimension is generated. S and bond vector K S :

[0247]

[0248] Among them, W QS ∈R d×d W KS ∈R d×d W is a trainable spatial attention weight matrix. VT ∈R d×p Let be a trainable dimensionality transformation matrix; considering the different noise levels of electromyography and motion signals, to reduce the impact of high-noise channels on attention calculation, channel uncertainty weights are introduced, and the confidence factor for each channel is defined as:

[0249]

[0250] MLP stands for Multilayer Perceptron, which is used to estimate the noise level of each time step and each channel.

[0251] Based on this, the channel contribution perception attention weight matrix A S Defined as:

[0252]

[0253] in, This represents element-wise multiplication;

[0254] Step 3-1-3: Multi-dimensional attention feature fusion;

[0255] After obtaining the sequence association attention weight matrix A T And channel contribution perception attention weight matrix A S Then, the input signals are weighted and fused; first, a linear transformation is performed on the input signals to obtain the value vector V:

[0256]

[0257] Among them, W VT ∈R d×p is a trainable weight matrix;

[0258] Subsequently, the sequence association attention weight matrix A T And channel contribution perception attention weight matrix A S Together, they act on the value vector to construct the feature representation after multidimensional relevance-aware attention encoding:

[0259]

[0260] The final input signal obtained after multidimensional correlation perceptual attention encoding is:

[0261]

[0262] Where, x r ′∈R d Let represent the r-th time-step feature vector fused with multidimensional related perceptual attention.

[0263] Specifically, in step 3-2, the input sequence X after multidimensional correlation-sensory attention encoding... ST The input is fed into a bidirectional gated recurrent unit network (GRU1) for historical temporal feature extraction to establish a bidirectional temporal dependency between historical motion and electromyographic signals. The input feature dimension of the GRU1 unit is d, the hidden layer dimension is h, and the gated recurrent unit operation is represented as follows:

[0264]

[0265] Among them, h f ∈R h Indicates the hidden state at the current time step; hf-1 This is the hidden state from the previous time step; x f ′ ∈R d This represents the feature vector at the f-th time step that incorporates multidimensional related perceptual attention.

[0266] For an input timing signal X of length n ST After passing through the bidirectional GRU network GRU1, the forward hidden state sequence H1 and the reverse hidden state sequence H2 are obtained respectively:

[0267]

[0268]

[0269] In this context, superscripts (1) and (2) represent networks propagating in the forward and backward directions, respectively.

[0270] The hidden states at the last time step are fused to obtain the comprehensive feature representation h of the sequence. n :

[0271]

[0272] Among them, h n (1) This represents the hidden state at the last time step of the forward GRU; h n (2) Represents the hidden state at the last time step of the reverse GRU; [;] represents vertical concatenation; Wc∈R h×2h b is the fusion weight matrix; c ∈R h This is the fusion bias; h is the length of the hidden layer; h n The global time series feature representation corresponding to the historical time window is used for multi-step prediction of subsequent future signals.

[0273] Specifically, in step 3-3, based on the historical global temporal features h extracted in step 3-2... n A unidirectional autoregressive future signal prediction network, GRU2, is constructed to progressively generate and predict motion and electromyographic signals within future time windows, thereby achieving feedforward modeling of fatigue evolution trends. The specific steps include:

[0274] Step 3-3-1, Multi-step prediction initialization;

[0275] In the multi-step prediction process, an iterative prediction strategy based on autoregression is adopted; the historical global time series feature vector h n As the initial hidden state s0 of the future prediction network GRU2:

[0276]

[0277] The first prediction for the future time step is generated through a fully connected layer:

[0278]

[0279] Among them, W fc ∈R p×h Let b represent the weight matrix of the fully connected layer. fc ∈R p Indicates the corresponding bias term;

[0280] Step 3-3-2: Prediction Result Feedback and Iterative Generation

[0281] The prediction result y1 at the current time step is used as the input for the next time step and fed into the one-way GRU prediction network GRU2 to recursively and dynamically generate the future sequence:

[0282]

[0283] Subsequently, the predicted value for the next time step is generated through the same fully connected layer:

[0284]

[0285] Repeat the above prediction process to gradually generate prediction results for multiple future time steps, namely:

[0286]

[0287]

[0288] Where m represents the prediction time step length;

[0289] Step 3-3-3: Multi-step prediction output;

[0290] The GRU2 one-way autoregressive future signal prediction network is used to progressively generate a predicted sequence Y for the next m time steps:

[0291]

[0292] It is used to characterize the evolution trend of signals over a future period of time, providing forward-looking input for subsequent fatigue evolution assessment and training regulation.

[0293] Specifically, in step 4, the fatigue calculation module is used to assess the fatigue state of the future motion and electromyographic signals obtained from multi-step predictions, thereby constructing a multi-dimensional fatigue index system that reflects the fatigue evolution characteristics during training. Based on the predicted signals, the fatigue calculation module calculates the instantaneous fatigue value, fatigue change rate, and fatigue accumulation trend, respectively, for subsequent training risk assessment and adaptive control, specifically including the following steps:

[0294] Step 4-1: Calculation of instantaneous fatigue;

[0295] At time step k, based on the preprocessed inertial motion signal feature matrix X imu and the preprocessed electromyographic signal feature matrix X emg Each signal prediction network is used independently to predict the motion signal and electromyographic signal with a length of n at time step k (inclusive) in the future at time step k, including features from m consecutive time steps, as the input window for fatigue calculation.

[0296]

[0297]

[0298] Where m represents the length of the time window used for fatigue estimation; this input window is used to reflect the comprehensive changes in motion and electromyographic signals over a future period of time, thereby avoiding the random influence of a single time point signal on fatigue judgment;

[0299] The two are concatenated to obtain the complete representation Y of the predicted signal. k :

[0300]

[0301] like Figure 2 As shown, the future multimodal prediction signal Y k The input is fed into a fatigue mapping network to establish a mapping relationship between future motion and electromyographic changes and fatigue state; the fatigue mapping network GRU3 (each unit of the network is GRU3, the input feature dimension is 9*c1+c2, and the hidden state length is h2) models the dynamic changes within the future time window step by step:

[0302]

[0303] Among them, u i Let u0 represent the hidden state vector corresponding to the i-th future time step, i∈{1,2,...,m}, u0 is a unit vector, and h2 represents the dimension of the GRU3 hidden layer;

[0304] Then, the hidden state u i The input is fed into a fully connected layer, and the instantaneous fatigue value at time step k+i is calculated:

[0305]

[0306] Among them, W g ∈R 1×h2 Let b be the weight matrix of the fully connected layer. g ∈R is the bias term, Fk+i This represents the estimated instantaneous fatigue level at time step k+i in the future.

[0307] By repeating the above calculation process for each time step within the prediction time window, a future fatigue sequence F containing m time points is gradually generated:

[0308]

[0309] This fatigue sequence is used for calculating the subsequent fatigue change rate and fatigue accumulation trend, as well as for training risk assessment and adaptive control.

[0310] Step 4-2: Calculation of fatigue change rate;

[0311] The fatigue change rate is used to characterize the trend of fatigue state over time, and is used to determine whether fatigue is in a rapid accumulation, stable, or recovery phase; the fatigue change rate is defined as follows:

[0312] Given a prediction step size s, the fatigue change rate G at time a a Defined as:

[0313]

[0314] Among them, F a+s This represents the fatigue value at the predicted time step a+s;

[0315] When G a When G > 0, it indicates that the fatigue level is on the rise; when G a When G < 0, it indicates a downward trend in fatigue level; when G a When the value is approximately 0, it indicates that the fatigue state is relatively stable. This indicator is used to identify risk periods of rapid fatigue growth in advance, providing a basis for feedforward adjustment of training parameters.

[0316] Step 4-3: Calculation of fatigue accumulation trend;

[0317] The cumulative fatigue tendency (AFI) is used to describe the cumulative effect of fatigue over a complete training cycle, reflecting the overall load level on the body caused by continuous training. The cumulative fatigue index (AFI) at time a is defined as:

[0318]

[0319] Among them, w i The weighting coefficients for the corresponding time steps are used to adjust the influence of fatigue in the cumulative calculation at different time periods. The weights are set according to the training stage or clinical experience. The larger the fatigue accumulation index, the higher the fatigue load in the training cycle, and the greater the risk of overtraining or sports injury.

[0320] Step 4-4: Constructing a multidimensional fatigue state vector;

[0321] Based on the above calculations, a single-point multidimensional fatigue state vector can be constructed at any future time point t (t∈{k,k+1,...,k+m}, where k is the current time point):

[0322]

[0323] This multidimensional fatigue state vector serves as the core input to the fatigue risk assessment and adaptive training control module, enabling dynamic monitoring and intelligent intervention of the training process.

[0324] Specifically, in step 5, the multi-point-dimensional fatigue state vector obtained from the fatigue calculation module is as follows:

[0325]

[0326] By constructing a fatigue risk grading mechanism and formulating adaptive training control rules accordingly, intelligent closed-loop control of the training process is achieved, specifically including the following steps:

[0327] Step 5-1: Construct a fatigue risk classification mechanism;

[0328] Training risk is assessed in a graded manner based on instantaneous fatigue value, fatigue change rate, and fatigue accumulation trend; the fatigue risk level includes, but is not limited to, the following four levels:

[0329] Level 1: Safety Level: Current fatigue level is low, fatigue changes are stable, and training risks are controllable;

[0330] Level 2: Alert Level: Fatigue level is moderate or the rate of fatigue increase is accelerating, requiring early warning adjustments to training intensity;

[0331] Level 3: Risk Level: High fatigue level or fatigue is accumulating rapidly, posing a risk of overtraining;

[0332] Level 4: Prohibition Level: Fatigue level or cumulative load exceeds the safety threshold, training must be terminated immediately.

[0333] Fatigue risk level determination: The multidimensional fatigue state vector is fuzzily mapped to four fatigue risk levels through the following steps: safe zone, warning zone, risk zone, and prohibited zone.

[0334] Step 5-1-1: Indicator normalization;

[0335] Linear normalization is performed on each indicator to unify its range to [0,1], resulting in the normalized index. :

[0336]

[0337] in This represents the normalized instantaneous fatigue value, fatigue rate of change, and fatigue accumulation trend.

[0338] Step 5-1-2: Membership function design;

[0339] Fatigue instantaneous value F k The membership functions are as follows:

[0340]

[0341]

[0342]

[0343]

[0344] in, The membership function representing the instantaneous fatigue value under the safety level; The membership function representing the instantaneous fatigue value under the warning level; The membership function representing the instantaneous fatigue value under the risk level; The membership function representing the instantaneous fatigue value under the prohibition level;

[0345] Fatigue change rate G k The membership functions are as follows:

[0346]

[0347]

[0348]

[0349]

[0350] in, The membership function representing the fatigue rate of change under the safety level; The membership function representing the fatigue rate of change under the warning level; The membership function representing the fatigue rate of change under the risk level; The membership function representing the fatigue rate of change under the prohibition level;

[0351] AFI (Accumulated Fatigue Index) k The membership functions are as follows:

[0352]

[0353]

[0354]

[0355]

[0356] in, The membership function representing the fatigue accumulation index under the safety level; The membership function representing the fatigue accumulation index under the warning level; The membership function representing the fatigue accumulation index under the risk level; The membership function representing the fatigue accumulation index under the prohibition level;

[0357] Step 5-1-3: Weighted summarization of multiple indicators;

[0358] To synthesize the fuzzy membership degrees of each indicator, the weighted comprehensive membership degree S(e) for each level is calculated:

[0359]

[0360] Where e∈{safety, vigilance, risk, prohibition}, w F ,w G ,w A ∈[0,1] are weight coefficients, satisfying w F +w G +w A =1, used to balance the contribution of different indicators to risk assessment;

[0361] Step 5-1-4: Determine the maximum membership degree;

[0362] Finally, based on the comprehensive membership degree, the risk level E corresponding to the maximum value is selected (Level 1 (Safe), Level 2 (Alert), Level 3 (Risk), Level 4 (Prohibited)):

[0363]

[0364] Wherein, argmax represents obtaining an optimal value of e such that S(e) reaches its maximum value; thereby realizing the fuzzy mapping from the multidimensional fatigue state vector to the four-level risk level;

[0365] Step 5-1-5: Calculation of final fatigue risk level;

[0366] For the m future points starting from time step k, calculate the fatigue risk level for each point sequentially to obtain a risk level sequence:

[0367]

[0368] Among them: E k+iThis represents the fatigue risk level corresponding to the i-th time step in the future; m represents the length of the future prediction time window; the risk levels include: Level 1, i.e., the safe zone; Level 2, i.e., the warning zone; Level 3, i.e., the risk zone; Level 4, i.e., the prohibited zone; to ensure safety during training, the principle of prioritizing the highest risk is adopted, and the risk levels within the future prediction window are comprehensively evaluated, and the highest risk level in the future time window is taken as the final fatigue risk level:

[0369]

[0370] Among them: E final This indicates the final fatigue risk level; max indicates taking the maximum value. In this way, the system can identify the highest fatigue risk state that may occur within a future time window in advance, thereby providing a basis for the feedforward adaptive adjustment of subsequent training parameters.

[0371] Step 5-2: Construct an adaptive adjustment strategy for training parameters;

[0372] For different fatigue risk levels, training parameters are adjusted in stages, including but not limited to: training movement amplitude, number of repetitions, movement execution rhythm, training load intensity, and training interval time.

[0373] The specific control rules are as follows:

[0374] ① Safe Zone Control Rules: When in the safe zone, the system maintains the predetermined training prescription parameters, allowing the normal training process to continue.

[0375] ② Warning Zone Control Rules: When in the warning zone, the system automatically reduces the training intensity, including but not limited to: proportionally reducing the range of motion; reducing the number of consecutive movements; and appropriately extending the training interval.

[0376] ③ Risk Zone Control Rules: When in the risk zone, the system takes further protective control measures, including but not limited to: significantly reducing training load or resistance; forcibly inserting rest cycles; and restricting the execution of high-intensity movements.

[0377] ④ Prohibited Zone Control Rules: When in a prohibited zone, the system immediately terminates the current training task and issues a warning to the user or therapist to prevent sports injuries.

[0378] Step 5-3: Feedforward regulation based on prediction results;

[0379] By utilizing the future fatigue state obtained through multi-step prediction, the system implements control strategies in advance before fatigue actually reaches a high-risk level, thus achieving feedforward training control. By identifying the risk of rapid fatigue growth or accumulation in advance, the system can adjust training parameters before fatigue is fully formed, thereby effectively reducing the risk of overtraining.

[0380] Step 5-4: Train and regulate the closed-loop update mechanism;

[0381] During training, the system continuously collects real-time motion and electromyographic signals and constantly updates fatigue prediction results and risk levels; the training control module dynamically adjusts training parameters based on the latest fatigue status, forming a closed-loop control process of "monitoring-prediction-evaluation-control".

[0382] like Figure 1 As shown, an intelligent closed-loop rehabilitation system based on multi-source information fusion for fatigue evolution prediction and adaptive regulation includes the following modules:

[0383] The multimodal wearable signal acquisition module utilizes wearable inertial motion sensors and surface electromyography sensors to acquire human kinematic signals and electromyography signals in real time during training.

[0384] The multi-source signal preprocessing module performs noise reduction, normalization, time synchronization, and sliding window segmentation on the acquired multi-modal signals to construct standardized time-series input data.

[0385] The signal prediction module constructs a multidimensional related perception attention encoding mechanism and a gated recurrent network model to model the current multimodal time-series signal and predict the motion and electromyographic signals in the future.

[0386] The multidimensional fatigue calculation module calculates the instantaneous fatigue value, fatigue change rate, and fatigue accumulation index at future time based on the predicted future signals, forming a multidimensional fatigue state vector.

[0387] We construct fatigue risk classification and adaptive training control rules, divide multidimensional fatigue states into different risk levels through fuzzy mapping, and dynamically adjust training parameters according to the risk level to achieve intelligent closed-loop control of the training process.

[0388] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0389] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart closed-loop rehabilitation method based on multi-source information fusion for fatigue evolution prediction and adaptive regulation, characterized in that, Includes the following steps: Step 1: Multimodal wearable signal acquisition. Using wearable inertial motion sensors and surface electromyography sensors, real-time acquisition of human kinematic signals and electromyographic signals during training is performed. Step 2: Multi-source signal preprocessing. The acquired multi-modal signals are denoised, normalized, time-synchronized, and segmented using a sliding window to construct standardized time-series input data. Step 3: Signal prediction. Construct a multidimensional related perception attention encoding mechanism and a gated recurrent network model to model the current multimodal time-series signal and predict the motion and electromyographic signals in the future. Step 4: Multidimensional fatigue calculation. Based on the predicted future signals, calculate the instantaneous fatigue value, fatigue change rate, and fatigue accumulation index at future times to form a multidimensional fatigue state vector. Step 5: Construct fatigue risk classification and adaptive training control rules. Use fuzzy mapping to classify multidimensional fatigue states into different risk levels and dynamically adjust training parameters according to the risk level to achieve intelligent closed-loop control of the training process.

2. The intelligent closed-loop rehabilitation method based on multi-source information fusion fatigue evolution prediction and adaptive regulation according to claim 1, characterized in that, Step 1, the multimodal wearable signal acquisition specifically includes the following steps: Step 1-1, Wearing the device; The wearable device is worn by installing c1 inertial sensors on the corresponding parts of the patient's body and c2 surface electromyography sensors on the surface of the main muscle groups related to the target training. Steps 1-2: Data Acquisition; Turn on the wearable device, perform sensor calibration, check accuracy and correct any possible offsets; Start the data acquisition program, the patient wears the wearable device to begin training or exercise, and the system records the motion data of various movements during the patient's rehabilitation training. Steps 1-3: Data storage and transmission: The collected multimodal data can be stored in the local storage unit of the wearable device or in a cloud database; in some application scenarios, the data is transmitted to the data analysis unit in real time via wireless communication for online fatigue assessment and training regulation. The dataset for building the deep learning model was also collected according to steps 1-1, 1-2, and 1-3, and included labels for multidimensional fatigue as assessed by doctors.

3. The intelligent closed-loop rehabilitation method based on multi-source information fusion fatigue evolution prediction and adaptive regulation according to claim 1, characterized in that, Step 2, the multi-source signal preprocessing specifically includes the following steps: Step 2-1: Inertial Motion Signal IMU Preprocessing; The attitude angle, acceleration, and angular velocity signals acquired by the inertial sensor are preprocessed as follows: Missing and outlier handling: When a local time segment is missing in the signal, it is filled by linear interpolation or mean compensation of the adjacent signals before and after the missing segment; when the entire signal of a single sensor is missing within a certain time period, historical samples with the same training action and the same fatigue level are selected from the database for replacement; outliers that exceed the reasonable physiological range are detected and corrected by neighborhood smoothing. Signal denoising and smoothing: To reduce the impact of random noise in inertial signals on model modeling, the original IMU signal is filtered, specifically including: using moving average filtering or low-pass filtering on acceleration and angular velocity signals to suppress high-frequency noise; and smoothing the attitude angle signal to avoid interference from instantaneous jitter on timing modeling. Time synchronization and standardization: To ensure the consistency of signals acquired by different sensors on the time axis, time alignment processing is performed on the signals of each IMU channel; at the same time, normalization or standardization processing is performed on the signals of each channel to eliminate the impact of differences in units and amplitudes on model training. Step 2-2: Preprocessing of surface electromyography (sEMG) signals; Considering the high-frequency and noise-sensitive characteristics of surface electromyography signals, the following preprocessing operations are performed: Bandpass filtering and power frequency interference suppression: The acquired raw electromyography (EMG) signals are bandpass filtered to remove low-frequency motion artifacts and high-frequency noise, and power frequency interference is suppressed to obtain stable and effective EMG signals. Signal amplitude normalization: The amplitude of the electromyographic signal is normalized to make the electromyographic signals of different subjects and different training stages comparable; Steps 2-3: Multimodal signal alignment and sliding window segmentation; After completing the independent preprocessing of inertial motion signals and surface electromyography signals, the multimodal signals undergo preprocessing for fusion, specifically including: The inertial motion signal and the surface electromyography signal are synchronized in time according to a unified time reference, and the signals with different sampling rates are resampled to keep the two types of signals consistent in time axis and sampling frequency. The synchronized multimodal signal is divided into a sliding window of fixed length n, with an overlap rate of 50% between adjacent windows. Each sliding window contains complete temporal features of motion and electromyography, which serve as input samples for subsequent fatigue evolution prediction models. Steps 2-4: Preprocessing output data format; After the above preprocessing, the data within a single time window is represented as follows: Among them, X imu This represents the preprocessed inertial motion signal feature matrix, where c1*9 represents c1 IMU sensors, each with 9 channels, and X... emg This represents the preprocessed electromyographic signal feature matrix; the motion signal feature matrix and the electromyographic signal feature matrix together serve as inputs to the fatigue evolution prediction model, used for subsequent fatigue state modeling and adaptive training regulation.

4. The intelligent closed-loop rehabilitation method based on multi-source information fusion fatigue evolution prediction and adaptive regulation according to claim 1, characterized in that, In step 3, the signal prediction module completes signal prediction, that is, it uses the existing signal X. imu or X emg Predicting future time series signal Y imu and Y emg And the predicted future signal length is m, i.e., Y imu ={y1,y2,...,y m }∈R (c1 *9)×m Y emg ={y1,y2,...,y m }∈R c2×m To simplify the expression, the two types of signals are uniformly represented as Y={y1,y2,...,y...} m }∈R p×m , where y t ∈R p , where p represents the dimension of the signal; a gated recurrent signal prediction network with a multidimensional correlation-sensory attention encoding mechanism is constructed, which is applicable to both surface electromyography (EMG) signals and motion signals prediction. The specific steps include: Step 3-1: Multidimensional correlation-based perceptual attention encoding; By constructing a sequence-time correlation attention and channel contribution perceptual attention mechanism, the key temporal dependencies and important channel features in the signal are adaptively weighted and encoded to improve feature representation ability; Step 3-2: Gated Recurrent Unit Extracts Spatiotemporal Features; The attention-encoded temporal signal is input into the Gated Recurrent Unit (GRU) network, and the temporal dynamic change features and long-term dependencies in the signal are extracted through the gating update mechanism; Step 3-3: Multi-step prediction; Based on the global time series features extracted by the GRU network, an autoregressive iterative prediction method is used to gradually generate signals for multiple future time steps, thereby achieving forward-looking prediction of the signals.

5. The intelligent closed-loop rehabilitation method based on multi-source information fusion fatigue evolution prediction and adaptive regulation according to claim 4, characterized in that, In step 3-1, for the multimodal data within a single time window after preprocessing... Data from any modality in the dataset can be simplified and represented as X: Where, x r ∈R p This represents the multidimensional feature vector at the r-th time step, where n represents the length of the time series and p is the feature dimension; specifically, it includes the following steps: Step 3-1-1: Calculation of temporal correlation attention weights in sequences; First, a sequential temporal correlation attention mechanism is constructed for the input signal X to characterize the influence of different historical time steps on the evolution of the current fatigue state; then, a query vector Q is generated through linear mapping. T and bond vector K T The calculation method is as follows: Among them, W QT ∈R d×p and W KT ∈R d×p This is a trainable spatial attention weight matrix; To enhance the ability to perceive signal change trends, a time step variation ΔX∈R is introduced. p×n : And construct the trend consistency term C∈R n×n : Among them, W Δ ∈R d×p The feature mapping matrix is ​​a variable. Combining query vector, key vector, and trend consistency term, the sequence association attention weight matrix A T Defined as: Where α is the trend consistency weight coefficient, d represents the feature dimension, and softmax is used to normalize the attention weights. Step 3-1-2: Calculation of channel contribution perceived attention weights; Based on sequence association modeling, a channel contribution perception attention mechanism is introduced to further distinguish the differences in contributions of different sensor channels and feature dimensions in the fatigue evolution process. By performing a linear mapping of the input features along the channel dimension, a query vector Q along the channel dimension is generated. S and bond vector K S : Among them, W QS ∈R d×d W KS ∈R d×d W is a trainable spatial attention weight matrix. VT ∈R d×p Let be a trainable dimensionality transformation matrix; considering the different noise levels of electromyography and motion signals, to reduce the impact of high-noise channels on attention calculation, channel uncertainty weights are introduced, and the confidence factor for each channel is defined as: MLP stands for Multilayer Perceptron, which is used to estimate the noise level of each time step and each channel. Based on this, the channel contribution perception attention weight matrix A S Defined as: in, This represents element-wise multiplication; Step 3-1-3: Multi-dimensional attention feature fusion; After obtaining the sequence association attention weight matrix A T And channel contribution perception attention weight matrix A S Then, the input signals are weighted and fused; first, a linear transformation is performed on the input signals to obtain the value vector V: Among them, W VT ∈R d×p is a trainable weight matrix; Subsequently, the sequence association attention weight matrix A T And channel contribution perception attention weight matrix A S Together, they act on the value vector to construct the feature representation after multidimensional relevance-aware attention encoding: The final input signal obtained after multidimensional correlation perceptual attention encoding is: Where, x r ′∈R d Let represent the r-th time-step feature vector fused with multidimensional related perceptual attention.

6. The intelligent closed-loop rehabilitation method based on multi-source information fusion fatigue evolution prediction and adaptive regulation according to claim 4, characterized in that, In step 3-2, the input sequence X after multidimensional correlation-sensory attention encoding ST The input is fed into a bidirectional gated recurrent unit network (GRU1) for historical temporal feature extraction to establish a bidirectional temporal dependency between historical motion and electromyographic signals. The input feature dimension of the GRU1 unit is d, the hidden layer dimension is h, and the gated recurrent unit operation is represented as follows: Among them, h f ∈R h Indicates the hidden state at the current time step; h f-1 This is the hidden state from the previous time step; x f ′ ∈R d This represents the feature vector at the f-th time step that incorporates multidimensional related perceptual attention. For an input timing signal X of length n ST After passing through the bidirectional GRU network GRU1, the forward hidden state sequence H1 and the reverse hidden state sequence H2 are obtained respectively: In this context, superscripts (1) and (2) represent networks propagating in the forward and backward directions, respectively. The hidden states at the last time step are fused to obtain the comprehensive feature representation h of the sequence. n : Among them, h n (1) This represents the hidden state at the last time step of the forward GRU; h n (2) Represents the hidden state at the last time step of the reverse GRU; [;] represents vertical concatenation; Wc∈R h×2h b is the fusion weight matrix; c ∈R h This is the fusion bias; h is the length of the hidden layer; h n The global time series feature representation corresponding to the historical time window is used for multi-step prediction of subsequent future signals.

7. The intelligent closed-loop rehabilitation method based on multi-source information fusion fatigue evolution prediction and adaptive regulation according to claim 4, characterized in that, In step 3-3, based on the historical global temporal features h extracted in step 3-2... n A unidirectional autoregressive future signal prediction network, GRU2, is constructed to progressively generate and predict motion and electromyographic signals within future time windows, thereby achieving feedforward modeling of fatigue evolution trends. The specific steps include: Step 3-3-1, Multi-step prediction initialization; In the multi-step prediction process, an iterative prediction strategy based on autoregression is adopted; the historical global time series feature vector h n As the initial hidden state s0 of the future prediction network GRU2: The first prediction for the future time step is generated through a fully connected layer: Among them, W fc ∈R p×h Let b represent the weight matrix of the fully connected layer. fc ∈R p Indicates the corresponding bias term; Step 3-3-2: Prediction Result Feedback and Iterative Generation The prediction result y1 at the current time step is used as the input for the next time step and fed into the one-way GRU prediction network GRU2 to recursively and dynamically generate the future sequence: Subsequently, the predicted value for the next time step is generated through the same fully connected layer: Repeat the above prediction process to gradually generate prediction results for multiple future time steps, namely: Where m represents the prediction time step length; Step 3-3-3: Multi-step prediction output; The GRU2 one-way autoregressive future signal prediction network is used to progressively generate a predicted sequence Y for the next m time steps: It is used to characterize the evolution trend of the signal and provide forward-looking input for subsequent fatigue evolution assessment and training regulation.

8. The intelligent closed-loop rehabilitation method based on multi-source information fusion fatigue evolution prediction and adaptive regulation according to claim 1, characterized in that, In step 4, the fatigue calculation module is used to assess the fatigue state of the future motion and electromyographic signals obtained from multi-step prediction, thereby constructing a multi-dimensional fatigue index system that reflects the fatigue evolution characteristics during training. The fatigue calculation module, based on the predicted signal, calculates the instantaneous fatigue value, fatigue change rate, and fatigue accumulation trend, which are used for subsequent training risk assessment and adaptive adjustment. Specifically, it includes the following steps: Step 4-1: Calculation of instantaneous fatigue; At time step k, based on the preprocessed inertial motion signal feature matrix X imu and the preprocessed electromyographic signal feature matrix X emg Each signal prediction network is used independently to predict the motion signal and electromyographic signal with a length of n at time step k in the future, including features from m consecutive time steps as the input window for fatigue calculation. Where m represents the length of the time window used for fatigue estimation; this input window is used to reflect the comprehensive changes in motion and electromyographic signals over a future period of time, thereby avoiding the random influence of a single time point signal on fatigue judgment; The two are concatenated to obtain the complete representation Y of the predicted signal. k : The future multimodal prediction signal Y k The data is input into a fatigue mapping network to establish a mapping relationship between future motion, electromyographic changes, and fatigue state. The fatigue mapping network GRU3 models the dynamic changes within a future time window step-by-step. Among them, u i Let u0 represent the hidden state vector corresponding to the i-th future time step, i∈{1,2,...,m}, u0 is a unit vector, and h2 represents the dimension of the GRU3 hidden layer; Then, the hidden state u i The input is fed into a fully connected layer, and the instantaneous fatigue value at time step k+i is calculated: Among them, W g ∈R 1×h2 Let b be the weight matrix of the fully connected layer. g ∈R is the bias term, F k+i This represents the estimated instantaneous fatigue level at time step k+i in the future. By performing time-step modeling and fully connected layer operations at each time step within the prediction time window, a future fatigue sequence F containing m time steps is gradually generated: This fatigue sequence is used for calculating the subsequent fatigue change rate and fatigue accumulation trend, as well as for training risk assessment and adaptive control. Step 4-2: Calculation of fatigue change rate; The fatigue change rate is used to characterize the trend of fatigue state over time, and is used to determine whether fatigue is in a rapid accumulation, stable, or recovery phase; the fatigue change rate is defined as follows: Given a prediction step size s, the fatigue change rate G at time a a Defined as: Among them, F a+s This represents the fatigue value at the predicted time step a+s; When G a When G > 0, it indicates that the fatigue level is on the rise; when G a When G < 0, it indicates a downward trend in fatigue level; when G a When the value is approximately 0, it indicates that the fatigue state is relatively stable. Step 4-3: Calculation of fatigue accumulation trend; The cumulative fatigue tendency (AFI) is used to describe the cumulative effect of fatigue over a complete training cycle, reflecting the overall load level on the body caused by continuous training. The cumulative fatigue index (AFI) at time a is defined as: Among them, w i The weighting coefficients for the corresponding time steps are used to adjust the influence of fatigue in the cumulative calculation at different time periods. The weights are set according to the training stage or clinical experience. The larger the fatigue accumulation index, the higher the fatigue load in the training cycle, and the greater the risk of overtraining or sports injury. Step 4-4: Constructing a multidimensional fatigue state vector; Based on the above calculations, a single-point multidimensional fatigue state vector can be constructed at any future time point t: This multidimensional fatigue state vector serves as the core input to the fatigue risk assessment and adaptive training control module, enabling dynamic monitoring and intelligent intervention of the training process.

9. The intelligent closed-loop rehabilitation method based on multi-source information fusion fatigue evolution prediction and adaptive regulation according to claim 1, characterized in that, In step 5, the multi-point-dimensional fatigue state vector obtained from the fatigue calculation module is as follows: By constructing a fatigue risk grading mechanism and formulating adaptive training control rules accordingly, intelligent closed-loop control of the training process is achieved, specifically including the following steps: Step 5-1: Construct a fatigue risk classification mechanism; Training risks are assessed in a tiered manner based on instantaneous fatigue values, fatigue change rates, and fatigue accumulation trends. The multidimensional fatigue state vector is then fuzzily mapped to four fatigue risk levels: safe zone, warning zone, risk zone, and prohibited zone, through the following steps: Step 5-1-1: Indicator normalization; Linear normalization is performed on each indicator to unify its range to [0,1], resulting in the normalized index. : in This represents the normalized instantaneous fatigue value, fatigue rate of change, and fatigue accumulation trend; Step 5-1-2: Membership function design; Fatigue instantaneous value F t The membership functions are as follows: in, The membership function representing the instantaneous fatigue value under the safety level; The membership function representing the instantaneous fatigue value under the warning level; The membership function representing the instantaneous fatigue value under the risk level; The membership function representing the instantaneous fatigue value under the prohibition level; Fatigue change rate G t The membership functions are as follows: in, The membership function representing the fatigue rate of change under the safety level; The membership function representing the fatigue rate of change under the warning level; The membership function representing the fatigue rate of change under the risk level; The membership function representing the fatigue rate of change under the prohibition level; AFI (Accumulated Fatigue Index) t The membership functions are as follows: in, The membership function representing the fatigue accumulation index under the safety level; The membership function representing the fatigue accumulation index under the warning level; The membership function representing the fatigue accumulation index under the risk level; The membership function representing the fatigue accumulation index under the prohibition level; Step 5-1-3: Weighted summarization of multiple indicators; To synthesize the fuzzy membership degrees of each indicator, the weighted comprehensive membership degree S(e) for each level is calculated: Where e∈{safety, vigilance, risk, prohibition}, w F ,w G ,w A ∈[0,1] are weight coefficients, satisfying w F +w G +w A =1, used to balance the contribution of different indicators to risk assessment; Step 5-1-4: Determine the maximum membership degree; Finally, the risk level E corresponding to the maximum value is selected based on the overall membership degree: Wherein, argmax represents obtaining an optimal value of e such that S(e) reaches its maximum value; thereby realizing the fuzzy mapping from the multidimensional fatigue state vector to the four-level risk level; Step 5-1-5: Calculation of final fatigue risk level; For the m future points starting from time step k, calculate the fatigue risk level for each point sequentially to obtain a risk level sequence: Among them: E k+i This represents the fatigue risk level corresponding to the i-th time step in the future; m represents the length of the future prediction time window; the risk levels include: Level 1, i.e., the safe zone; Level 2, i.e., the warning zone; Level 3, i.e., the risk zone; Level 4, i.e., the prohibited zone; to ensure safety during training, the principle of prioritizing the highest risk is adopted, and the risk levels within the future prediction window are comprehensively evaluated, and the highest risk level in the future time window is taken as the final fatigue risk level: Among them: E final The value represents the final fatigue risk level; max represents the maximum value. By calculating the highest risk level in the future time window, the system can identify the highest fatigue risk state that may occur in the future time window in advance, thus providing a basis for the feedforward adaptive adjustment of subsequent training parameters. Step 5-2: Construct an adaptive adjustment strategy for training parameters; For different fatigue risk levels, training parameters are adjusted in stages, including but not limited to: training movement amplitude, number of repetitions, movement execution rhythm, training load intensity, and training interval time. Step 5-3: Feedforward regulation based on prediction results; By utilizing the future fatigue state obtained through multi-step prediction, the system implements control strategies in advance before fatigue actually reaches a high-risk level, thus achieving feedforward training control. By identifying the risk of rapid fatigue growth or accumulation in advance, the system can adjust training parameters before fatigue is fully formed, thereby effectively reducing the risk of overtraining. Step 5-4: Train and regulate the closed-loop update mechanism; During training, the system continuously collects real-time motion and electromyographic signals and constantly updates fatigue prediction results and risk levels; the training control module dynamically adjusts training parameters based on the latest fatigue status, forming a closed-loop control process of "monitoring-prediction-evaluation-control".

10. An intelligent closed-loop rehabilitation system based on multi-source information fusion for fatigue evolution prediction and adaptive regulation, characterized in that, Includes the following modules: The multimodal wearable signal acquisition module utilizes wearable inertial motion sensors and surface electromyography sensors to acquire human kinematic signals and electromyography signals in real time during training. The multi-source signal preprocessing module performs noise reduction, normalization, time synchronization, and sliding window segmentation on the acquired multi-modal signals to construct standardized time-series input data. The signal prediction module constructs a multidimensional related perception attention encoding mechanism and a gated recurrent network model to model the current multimodal time-series signal and predict the motion and electromyographic signals in the future. The multidimensional fatigue calculation module calculates the instantaneous fatigue value, fatigue change rate, and fatigue accumulation index at future time based on the predicted future signals, forming a multidimensional fatigue state vector. A fatigue risk classification and adaptive training control rule module is constructed. The multidimensional fatigue state is divided into different risk levels by using the fuzzy mapping method, and the training parameters are dynamically adjusted according to the risk level to achieve intelligent closed-loop control of the training process.

Images