A wearable motion assistance device control method, device, and storage medium

Abstract

The application provides a wearable motion auxiliary device control method, device and storage medium, and belongs to the technical field of wearable device control. The method comprises the following steps: obtaining multiple original human motion data, original electromyography data, original pressure data and original device running data from sensors arranged on a wearable motion auxiliary device; and performing feature analysis on the original human motion data, original electromyography data, original pressure data and original device running data respectively to obtain original human motion features, original electromyography features, original pressure features and original device running features. The application solves the problems of poor generalization ability and response lag of existing models, can adapt to multiple motion scenes such as walking, running and going up and down stairs, solves the problems of high generalization degree and insufficient personalization of existing models, further improves the control precision and stability, and has high practicality and promotional value.

Description

Technical Field

[0001] This invention relates to the field of wearable device control technology, specifically to a wearable sports assistive device control method, device, and storage medium. Background Technology

[0002] With the increasing aging population and rising demand for sports rehabilitation and fitness, wearable motion assistive devices (such as lower limb exoskeletons and assistive belts) have gradually become a research hotspot. Their core function is to sense the user's movement state and output appropriate assistive torque to help users reduce exercise burden, improve postural stability, or achieve rehabilitation training goals. Currently, the control core of wearable motion assistive devices largely relies on motion intention recognition models and assistive torque control models. Existing technologies have the following key shortcomings: (1) Existing control models suffer from poor generalization ability and slow response: Most mainstream control models currently use a single deep learning model (such as CNN, LSTM) or a traditional machine learning model (such as SVM, finite state machine FSM), which can only be trained for specific motion scenarios (such as walking on flat ground) or specific groups of people (such as healthy young people). It is difficult to adapt to the differences in motion characteristics of different motion modes (walking, running, going up and down stairs) and different groups of people (elderly, rehabilitation patients, athletes). Moreover, the model responds slowly to sudden changes in motion state, and the auxiliary torque is prone to being out of sync with the rhythm of human movement, which affects the human-machine collaboration effect. For example, the algorithm based on RSVM and FSM developed by Wu et al. of ETH Zurich, Switzerland, has improved the gait phase detection capability, but there is still a response delay when switching motion modes, which is difficult to adapt to complex motion scenarios.

[0003] (2) Existing loss function design is unreasonable and control accuracy is insufficient: The loss function used in the training of existing models is mostly the single mean square error (MSE), which only focuses on the deviation between the model's predicted value and the actual value, without considering the human-machine collaboration characteristics, motion safety and the real-time requirements of wearable devices. For example, some control methods based on electromyography prediction models only use the root mean square error as the loss function, which leads to an excessive pursuit of prediction accuracy during model training, ignoring the smoothness of the auxiliary torque, and easily generating torque jitter, which not only affects the user's wearing comfort, but may also cause additional impact on the human joints; at the same time, the error weight when the motion state changes suddenly is not considered, resulting in insufficient prediction accuracy of the model at key motion nodes (such as posture switching and center of gravity shift), which cannot meet the requirements of precise assistance.

[0004] (3) Poor model adaptability, making it difficult to achieve personalized control: Existing models mostly adopt generalized training methods, which do not fully integrate the personalized characteristics of human movement (such as muscle strength, posture habits, and joint range of motion), and lack adaptation to dynamic factors such as muscle fatigue and changes in exercise intensity. As a result, the same control parameters are difficult to meet the needs of different users, and there may even be cases of over-assistance or under-assistance. For example, although the existing CEINMS-RT open-source framework has achieved real-time control, the calibration time is long and it is difficult to dynamically adapt to changes in the user's muscle state during exercise.

[0005] In summary, existing control methods for wearable motion assistive devices suffer from deficiencies in control accuracy, human-machine collaboration, and scenario adaptability due to unreasonable structural design of the control model and a single optimization objective of the loss function, making it difficult to meet diverse motion assistance and rehabilitation needs. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a control method, device and storage medium for wearable sports assistive devices, which are in line with the shortcomings of the prior art.

[0007] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: A control method for a wearable sports assistive device, comprising the following steps: Multiple raw human motion data, raw electromyography data corresponding to each raw human motion data, raw pressure data corresponding to each raw human motion data, and raw device operation data corresponding to each raw human motion data are obtained from sensors installed in the wearable sports assist device. Feature analysis is performed on each of the original human motion data, the original electromyography data corresponding to each of the original human motion data, the original pressure data corresponding to each of the original human motion data, and the original device operation data corresponding to each of the original human motion data to obtain the original human motion features, the original electromyography features, the original pressure features, and the original device operation features corresponding to each of the original human motion data. A training model is constructed, and the training model is analyzed using all the original human motion features, all the original electromyographic features, the original pressure features, and all the original device operation features to obtain the target control model. The wearable motion assist device obtains the human motion data to be controlled, the electromyography data to be controlled, the pressure data to be controlled, and the operating data of the device to be controlled from the sensors installed on the wearable motion assist device. The target control model predicts the human motion data, electromyography data, pressure data, and device operation data to be controlled, thereby obtaining the target human motion intention and the target auxiliary torque value. The wearable motion assist device is then controlled based on the target human motion intention and the target auxiliary torque value.

[0008] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: A wearable sports assistive device control device, comprising: The first data acquisition module is used to acquire multiple raw human motion data, raw electromyography data corresponding to each of the raw human motion data, raw pressure data corresponding to each of the raw human motion data, and raw device operation data corresponding to each of the raw human motion data from the sensors installed in the wearable sports assist device. The feature analysis module is used to perform feature analysis on each of the original human motion data, the original electromyography data corresponding to each of the original human motion data, the original pressure data corresponding to each of the original human motion data, and the original device operation data corresponding to each of the original human motion data, to obtain the original human motion features, the original electromyography features, the original pressure features, and the original device operation features corresponding to each of the original human motion data. The model analysis module is used to construct a training model. It performs model analysis on the training model using all the original human motion features, all the original electromyographic features, the original pressure features, and all the original device operation features to obtain the target control model. The second data acquisition module is used to acquire human motion data to be controlled, electromyographic data to be controlled, pressure data to be controlled, and operating data of the device to be controlled from the sensors installed in the wearable sports assist device; The control module is used to predict the human motion data, electromyography data, pressure data, and device operation data to be controlled through the target control model, to obtain the target human motion intention and the target auxiliary torque value, and to control the wearable motion assist device according to the target human motion intention and the target auxiliary torque value.

[0009] Based on the above-mentioned control method for wearable sports assistive devices, the present invention also provides a control system for wearable sports assistive devices.

[0010] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: a wearable sports assistive device control system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the wearable sports assistive device control method as described above.

[0011] Based on the above-described control method for a wearable sports assistive device, the present invention also provides a computer-readable storage medium.

[0012] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the wearable sports assistive device control method as described above.

[0013] The beneficial effects of this invention are as follows: By analyzing the features of original human motion data, original electromyography (EMG) data, original pressure data, and original device operation data, original human motion characteristics, original EMG characteristics, original pressure characteristics, and original device operation characteristics are obtained. The target control model is obtained through model analysis of the training model based on these original human motion characteristics, EMG characteristics, original pressure characteristics, and original device operation characteristics. The target control model is then used to predict the target human motion intention and target auxiliary torque value based on the target human motion data, EMG data, pressure data, and device operation data. This invention controls the wearable motion assist device based on the target human motion intention and target auxiliary torque value, solving the problems of poor generalization ability and lag in existing models. It is adaptable to various exercise scenarios such as walking, running, and climbing stairs, as well as different groups such as the elderly, rehabilitation patients, and athletes. It reduces the model's response time, ensures smooth and safe auxiliary torque output, reduces impact on human joints, significantly improves wearing comfort and exercise safety, and also solves the problems of high generalization and insufficient personalization in existing models. Furthermore, it improves control accuracy and stability, and has high practicality and promotional value. Attached Figure Description

[0014] Figure 1 A flowchart illustrating the control method for a wearable sports assistive device provided in an embodiment of the present invention; Figure 2 This is a block diagram of a wearable sports assistive device control device provided in an embodiment of the present invention. Detailed Implementation

[0015] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0016] Figure 1This is a flowchart illustrating a control method for a wearable sports assistive device provided in an embodiment of the present invention.

[0017] like Figure 1 As shown, a control method for a wearable motion assistive device includes the following steps: S1: Obtain multiple raw human motion data, raw electromyography data corresponding to each of the raw human motion data, raw pressure data corresponding to each of the raw human motion data, and raw device operation data corresponding to each of the raw human motion data from the sensors installed in the wearable motion assist device; S2: Perform feature analysis on each of the original human motion data, the original electromyography data corresponding to each of the original human motion data, the original pressure data corresponding to each of the original human motion data, and the original device operation data corresponding to each of the original human motion data to obtain the original human motion features, the original electromyography features, the original pressure features, and the original device operation features corresponding to each of the original human motion data. S3: Construct a training model, and perform model analysis on the training model using all the original human motion features, all the original electromyographic features, all the original pressure features, and all the original device operation features to obtain the target control model; S4: Obtain the human motion data to be controlled, the electromyographic data to be controlled, the pressure data to be controlled, and the operating data of the device to be controlled from the sensors installed in the wearable sports assist device; S5: The target control model is used to predict the human motion data to be controlled, the electromyographic data to be controlled, the pressure data to be controlled, and the operating data of the device to be controlled, so as to obtain the target human motion intention and the target auxiliary torque value, and the wearable motion assist device is controlled according to the target human motion intention and the target auxiliary torque value.

[0018] It should be understood that the wearable motion assist device collects human motion data and device operation data (i.e., raw human motion data, raw electromyography data, raw pressure data, and raw device operation data) in real time through the sensor module (i.e., sensor) integrated on the device. The sensor module (i.e., sensor) includes an inertial measurement unit (IMU), a surface electromyography sensor (sEMG), a pressure sensor, and a torque sensor. The collected data (i.e., raw human motion data, raw electromyography data, raw pressure data, and raw device operation data) includes: human joint angles, joint angular velocities, electromyography signals (EMG), plantar pressure distribution, device output auxiliary torque, and motor operating parameters.

[0019] Specifically, inertial measurement units (IMUs) are installed at the hip, knee, and ankle joints of the wearable lower limb exoskeleton to collect joint angles and angular velocities (i.e., raw human motion data); surface electromyography (sEMG) sensors are attached to the front, back, and back of the thigh and calf to collect electromyographic signals of the rectus femoris, biceps femoris, and gastrocnemius muscles (i.e., raw electromyographic data); pressure sensors are installed on the sole of the foot to collect plantar pressure distribution (i.e., raw pressure data); and torque sensors are installed at the actuator (motor) to collect the auxiliary torque output by the device (i.e., raw device operating data). The sampling frequency of all sensors is set to 100Hz, and data is collected in real time and transmitted to the controller.

[0020] Specifically, the real-time collected and preprocessed motion data (i.e., the human motion data to be controlled, the electromyography data to be controlled, the pressure data to be controlled, and the operating data of the device to be controlled) are input into the trained improved control model (i.e., the target control model). The model outputs the human motion intention recognition result and the optimal auxiliary torque command (i.e., the target human motion intention and the target auxiliary torque value). The actuator (motor, reducer) of the wearable motion assist device outputs the appropriate auxiliary torque according to the auxiliary torque command (i.e., the target auxiliary torque value) to achieve motion assistance. At the same time, the device operating data is fed back to the model (i.e., the target control model) in real time to achieve closed-loop regulation.

[0021] In the above embodiments, the original human motion characteristics, original electromyography (EMG) characteristics, original pressure characteristics, and original device operation characteristics are obtained through feature analysis of the original human motion data, original EMG data, original pressure data, and original device operation data. The target control model is obtained through model analysis of the training model based on these original human motion characteristics, EMG characteristics, original pressure characteristics, and original device operation characteristics. The target control model is then used to predict the target human motion intention and target auxiliary torque value based on the target human motion data, EMG data, pressure data, and device operation data. The wearable motion assist device is then controlled based on the target human motion intention and target auxiliary torque value. This solves the problems of poor generalization ability and lag in existing models, and is adaptable to various sports scenarios such as walking, running, and climbing stairs, as well as different groups such as the elderly, rehabilitation patients, and athletes. It reduces the model's response time, ensures smooth and safe auxiliary torque output, reduces impact on human joints, significantly improves wearing comfort and sports safety, and also solves the problems of high generalization and insufficient personalization in existing models. Furthermore, it improves control accuracy and stability, and has high practicality and promotional value.

[0022] Optionally, as an embodiment of the present invention, the process of performing feature analysis on each of the original human motion data, the original electromyographic data corresponding to each of the original human motion data, the original pressure data corresponding to each of the original human motion data, and the original device operation data corresponding to each of the original human motion data to obtain the original human motion features, the original electromyographic features, the original pressure features, and the original device operation features corresponding to each of the original human motion data includes: The original electromyography (EMG) data are filtered using a wavelet filtering algorithm to obtain filtered EMG data corresponding to each of the original human motion data. The original human motion data, the original pressure data corresponding to each original human motion data, and the original device operation data corresponding to each original human motion data are filtered using a moving average filtering algorithm to obtain the filtered human motion data, the filtered pressure data, and the filtered device operation data corresponding to each original human motion data. Normalization processing is performed on each of the filtered electromyography data, the filtered human motion data corresponding to each of the original human motion data, the filtered pressure data corresponding to each of the original human motion data, and the filtered device operation data corresponding to each of the original human motion data to obtain normalized electromyography data, normalized human motion data, normalized pressure data, and normalized device operation data corresponding to each of the original human motion data. Feature extraction is performed on each of the normalized electromyography data, the normalized human motion data corresponding to each of the original human motion data, the normalized pressure data corresponding to each of the original human motion data, and the normalized device operation data corresponding to each of the original human motion data to obtain the original human motion features, the original electromyography features, the original pressure features, and the original device operation features corresponding to each of the original human motion data.

[0023] Understandably, the collected raw data (i.e., raw human motion data, raw electromyography (EMG) data, raw pressure data, and raw device operation data) are filtered, denoised, normalized, and feature extracted to obtain feature vectors for model training and inference. Among them, the EMG signal (i.e., raw EMG data) is filtered using wavelet filtering to remove power frequency interference and EMG noise, and the joint angle and angular velocity (i.e., raw human motion data) are smoothed using moving average filtering. Feature extraction includes time-domain features (integrated EMG value and root mean square value of EMG signal), frequency-domain features (power spectral density of EMG signal), and kinematic features (joint angle change rate and center of gravity transfer velocity).

[0024] Specifically, the electromyography (EMG) signal (i.e., the raw EMG data) is filtered using a 5-50Hz wavelet filter to remove power frequency interference (50Hz) and EMG noise; the joint angle, joint angular velocity and auxiliary torque data (i.e., the raw human motion data and the raw device operation data) are filtered using a moving average filter (window size of 5) to smooth data fluctuations.

[0025] It should be understood that all collected data (i.e., filtered electromyography data, filtered human motion data, filtered pressure data, and filtered device operation data) should be normalized to the [0,1] interval to eliminate dimensional differences, using the min-max normalization method: Where x is the original data, The minimum value of the data. This represents the maximum value of the data.

[0026] Specifically, the time-domain features (integrated electromyography value, root mean square value) and frequency-domain features (peak power spectral density) of electromyography signals are extracted; the time-domain features of joint motion (joint angle change rate, peak angular velocity) are extracted; and the plantar pressure features (coordinates of pressure center, maximum pressure value) are extracted. All features are integrated to form a feature vector with a dimension of 32 (i.e., original human motion features, original electromyography features, original pressure features, and original device operation features).

[0027] In the above embodiments, feature analysis is performed on the original human motion data, original electromyography data, original pressure data, and original device operation data to obtain the original human motion characteristics, original electromyography characteristics, original pressure characteristics, and original device operation characteristics, thereby removing power frequency interference and electromyography noise and smoothing data fluctuations.

[0028] Optionally, as an embodiment of the present invention, the training model includes a batch normalization layer, a first fully connected network, a residual connection layer, a temporal feature extraction layer, an attention mechanism layer, a softmax activation function layer, and a linear activation function layer; The process of analyzing the training model using all the original human motion characteristics, all the original electromyographic characteristics, all the original pressure characteristics, and all the original device operation characteristics to obtain the target control model includes: The batch normalization layer normalizes each of the original human motion features, the original electromyographic features corresponding to each of the original human motion data, the original pressure features corresponding to each of the original human motion data, and the original device operation features corresponding to each of the original human motion data, respectively, to obtain normalized human motion features, normalized electromyographic features, normalized pressure features, and normalized device operation features corresponding to each of the original human motion data. The first fully connected network is used to fuse the normalized human motion features, the normalized electromyographic features corresponding to the original human motion data, the normalized pressure features corresponding to the original human motion data, and the normalized device operation features corresponding to the original human motion data to obtain the original fused features corresponding to the original human motion data. Each of the original fusion features is mapped to obtain the mapped fusion features corresponding to each of the original human motion data. The residual connection layer performs feature superposition on each of the normalized human motion features, the normalized electromyographic features corresponding to each of the original human motion data, the normalized pressure features corresponding to each of the original human motion data, the normalized device operation features corresponding to each of the original human motion data, and the mapped fusion features corresponding to each of the original human motion data, to obtain the target fusion features corresponding to each of the original human motion data. The temporal feature extraction layer extracts features from each of the target fusion features to obtain target temporal features corresponding to each of the original human motion data. The attention mechanism layer performs attention analysis on each of the target temporal features to obtain target attention features corresponding to each of the original human motion data. The attention features of each target are predicted by the softmax activation function layer to obtain human motion intention prediction labels corresponding to each of the original human motion data. The linear activation function layer is used to predict the attention features of each target, thereby obtaining the auxiliary torque prediction value corresponding to each of the original human motion data. Import the real human motion intention labels corresponding to each of the original human motion data, the real values ​​of the auxiliary torque corresponding to each of the original human motion data, the user's personalized motion data corresponding to each of the original human motion data, and the user's reference trajectory corresponding to each of the original human motion data. The user-personalized motion data and the user reference trajectory corresponding to each of the original human motion data are fused to obtain the user-personalized auxiliary torque target value corresponding to each of the original human motion data. A loss function analysis is performed on all the predicted human motion intention labels, all the predicted auxiliary torque values, all the actual human motion intention labels, all the user-personalized auxiliary torque target values, and all the actual auxiliary torque values ​​to obtain the target loss function; The parameters of the trained model are updated according to the target loss function to obtain the target control model.

[0029] It should be understood that a hybrid deep learning control model (Attention-LSTM-GRU) based on an attention mechanism (i.e., the training model) is constructed, and an improved loss function (i.e., the target loss function) is used to train the model (i.e., the training model) to obtain a trained control model (i.e., the target control model); the improved control model is used to predict human motion intention (motion pattern, motion intensity) (i.e., human motion intention prediction label) and optimal auxiliary torque (i.e., auxiliary torque prediction value) in real time based on the preprocessed feature vector.

[0030] It should be understood that, in order to address the problems of poor generalization ability and slow response of existing models, this invention improves the control model structure by integrating LSTM, GRU and attention mechanisms to construct a hybrid deep learning model (i.e., target control model).

[0031] Specifically, the system receives preprocessed feature vectors (i.e., original human motion features, original electromyographic features, original pressure features, and original device operation features), including human kinematic features, electromyographic features, pressure features, and device operation features. The feature vector dimension is determined according to the data type collected by the sensor. The input is processed using batch normalization (i.e., batch normalization layer) to accelerate the model training convergence speed and reduce the impact of data distribution differences on the model.

[0032] Specifically, a fully connected network (i.e., the first fully connected network) is used to fuse the multi-dimensional features of the input (i.e., normalized human motion features, normalized electromyography features, normalized pressure features, and normalized device operation features), mapping different types of features to the same feature space. At the same time, residual connections (i.e., residual connection layers) are introduced to avoid the gradient vanishing problem caused by the increase in model depth and to retain the effective information of the original features.

[0033] It should be understood that a dual-output structure is adopted, which outputs the motion intention recognition result (through the softmax activation function) (i.e., human motion intention prediction label) and the optimal auxiliary torque prediction value (through the linear activation function) (i.e., auxiliary torque prediction value). Among them, the motion intention recognition (i.e., human motion intention prediction label) includes the recognition of motion mode (walking, running, climbing stairs) and motion intensity (low intensity, medium intensity, high intensity), and the auxiliary torque prediction outputs the corresponding optimal auxiliary torque value (i.e., auxiliary torque prediction value) for different joints (hip joint, knee joint, ankle joint).

[0034] Specifically, by using the parallel design of LSTM and GRU, the extraction of both long-term temporal features and short-term abrupt features is taken into account, thus solving the problem of response lag in existing models. By introducing an attention mechanism, the model’s attention to key features and key temporal nodes is enhanced, thereby improving the model’s generalization ability and adapting it to the motion characteristics of different motion scenarios and different groups of people. The dual-output structure can simultaneously realize motion intention recognition and auxiliary torque prediction, simplifying the control process and improving control efficiency.

[0035] In the above embodiments, the training model is analyzed using all original human motion features, all original electromyographic features, all original pressure features, and all original device operation features to obtain the target control model. This accelerates the convergence speed of model training, reduces the impact of data distribution differences on the model, avoids the gradient vanishing problem caused by the increase in model depth, preserves the effective information of the original features, solves the problem of lag in the response of existing models, enhances the generalization ability of the model, and can adapt to the motion features of different motion scenarios and different groups of people.

[0036] Optionally, as an embodiment of the present invention, the temporal feature extraction layer includes an LSTM network and a GRU network; The process of extracting features from each of the target fusion features through the temporal feature extraction layer to obtain the target temporal features corresponding to each of the original human motion data includes: The LSTM network is used to capture the fused features of each target to obtain long-time features corresponding to each of the original human motion data. The GRU network is used to capture the features of each target fusion feature to obtain short-term mutation features corresponding to each of the original human motion data. Each of the long-term temporal features and the short-term mutation features corresponding to each of the original human motion data are concatenated to obtain the target temporal features corresponding to each of the original human motion data.

[0037] It should be understood that the system is composed of LSTM units (i.e., LSTM networks) and GRU units (i.e., GRU networks) in parallel. The LSTM units (i.e., LSTM networks) are used to capture long-term dependencies in motion data (such as the temporal characteristics of attitude cycles), while the GRU units (i.e., GRU networks) are used to capture short-term abrupt changes in motion data (such as motion mode switching and sudden shifts in the center of gravity). The temporal characteristics output by the two are concatenated to balance long-term stability and short-term response speed.

[0038] In the above embodiments, the temporal feature extraction layer extracts features from each target fusion feature to obtain target temporal features, which takes into account both long-term stability and short-term response speed, avoids the gradient vanishing problem caused by the increase of model depth, retains the effective information of the original features, and solves the problem of response lag in existing models.

[0039] Optionally, as an embodiment of the present invention, the attention mechanism layer includes a global average pooling layer, a global max pooling layer, a second fully connected network, a first sigmoid activation function layer, a convolutional layer, and a second sigmoid activation function layer; The process of performing attention analysis on each of the target temporal features through the attention mechanism layer to obtain the target attention features corresponding to each of the original human motion data includes: The global average pooling layer is used to perform average pooling on each of the target temporal features to obtain the first pooled temporal features corresponding to each of the original human motion data. The global max pooling layer is used to perform max pooling on each of the target temporal features to obtain the second pooled temporal features corresponding to each of the original human motion data. The second fully connected network is used to fuse each of the first pooled temporal features and the second pooled temporal features corresponding to each of the original human motion data to obtain the fused features to be processed corresponding to each of the original human motion data. The first Sigmoid activation function layer is used to map each of the fusion features to be processed to obtain the channel attention weight vector corresponding to each of the original human motion data. The channel attention features corresponding to each of the original human motion data are obtained by multiplying each of the target temporal features and the channel attention weight vector corresponding to each of the original human motion data element by element. Each of the target temporal features is subjected to channel-dimensional average pooling to obtain a third pooled temporal feature corresponding to each of the original human motion data. Each of the target temporal features is subjected to channel-dimensional max pooling to obtain a fourth pooled temporal feature corresponding to each of the original human motion data. The convolutional layer is used to fuse the third pooling time-series features and the fourth pooling time-series features corresponding to the original human motion data to obtain the fused features corresponding to the original human motion data. The second Sigmoid activation function layer is used to map each of the processed fused features to obtain a spatial attention weight vector corresponding to each of the original human motion data. Each of the spatial attention weight vectors and the channel attention features corresponding to each of the original human motion data are multiplied element-wise to obtain the target attention features corresponding to each of the original human motion data.

[0040] It should be understood that channel attention mechanism and spatial attention mechanism are introduced to assign weights to the features output by the hybrid temporal feature extraction layer (i.e., the target temporal features).

[0041] Specifically, firstly, the temporal feature matrix (i.e., the target temporal feature) output by the hybrid temporal feature extraction layer is received, denoted as F∈R^(C×T), where C is the number of feature channels and T is the number of temporal nodes. For the channel attention mechanism, the feature matrix F (i.e., the target temporal feature) is first processed by global average pooling and global max pooling to obtain two 1×C channel feature vectors (i.e., the first pooled temporal feature and the second pooled temporal feature). The two vectors (i.e., the first pooled temporal feature and the second pooled temporal feature) are input into a fully connected network containing a single hidden layer (i.e., the second fully connected network). After passing through the Sigmoid activation function (i.e., the first Sigmoid activation function layer), the channel attention weight vector W_c∈R^C is output. This weight vector (i.e., the channel attention weight vector) is then multiplied element-wise by the original feature matrix F (i.e., the target temporal feature) according to the channel dimension to achieve weight allocation for different feature channels. For the spatial attention mechanism, the feature matrix F (i.e., the target temporal feature) is first processed by channel-dimensional average pooling and max pooling. The model undergoes max pooling to obtain two 1×T temporal feature vectors (i.e., the temporal features after the third and fourth pooling). These two vectors are concatenated and input into a convolutional layer (with a 3×3 kernel). After passing through the Sigmoid activation function (i.e., the second Sigmoid activation function layer), a spatial attention weight vector W_s∈R^T is output. This weight vector (spatial attention weight vector) is then multiplied element-wise with the channel-weighted feature matrix (channel attention features) according to temporal nodes to allocate weights to key temporal nodes. Finally, the feature matrix (target attention features) after dual channel and spatial weighting is output to the output layer. The channel attention mechanism is used to highlight feature channels related to motion intention and auxiliary torque (such as electromyography feature channels and joint angle feature channels) and suppress interference from irrelevant features. The spatial attention mechanism is used to focus on key temporal nodes of motion state change (such as posture switching moments) to improve the model's response speed to motion changes.

[0042] In the above embodiments, attention analysis is performed on each target temporal feature through the attention mechanism layer to obtain target attention features, which suppresses the interference of irrelevant features and improves the model's response speed to sudden motion changes.

[0043] Optionally, as an embodiment of the present invention, the process of performing loss function analysis on all the predicted human motion intention labels, all the predicted auxiliary torque values, all the actual human motion intention labels, all the user-personalized auxiliary torque target values, and all the actual auxiliary torque values ​​to obtain the target loss function includes: The target loss function is obtained by calculating all the predicted human motion intention labels, all the predicted auxiliary torque values, all the actual human motion intention labels, all the user-personalized auxiliary torque target values, and all the actual auxiliary torque values ​​using the first formula. The first formula is: , in, , , , , , in, Let be the target loss function. , , , as well as All are weighting coefficients. , Let be the first prediction error loss function. This is the second prediction error loss function. Let the torque smoothness loss function be... Let the motion safety loss function be... To personalize the loss function, This represents the total number of original human motion data. For the first The first dynamic weighting factor corresponding to each original human motion data point. For the first The actual value of the auxiliary torque corresponding to each original human motion data. For the first The predicted auxiliary torque value corresponding to each original human motion data. For the first The second dynamic weighting factor corresponding to each original human motion data. For the first The true label of human movement intention corresponding to each piece of raw human movement data For the first Human movement intention prediction labels corresponding to each piece of raw human movement data For the first The predicted auxiliary torque value corresponding to each original human motion data. To preset the minimum safe threshold for auxiliary torque, To preset the maximum safe threshold for auxiliary torque, For the first User personalization coefficients corresponding to each set of raw human motion data. For the first The user-personalized auxiliary torque target value corresponding to each original human motion data.

[0044] Specifically, addressing the problem that existing loss functions are singular and fail to consider multi-dimensional optimization objectives, this invention designs an improved loss function that integrates prediction error loss, torque smoothness loss, motion safety loss, and personalized adaptation loss to achieve multi-dimensional optimization of the model. The specific expression is as follows: , in, The total loss value of the improved loss function (i.e., the target loss function); These are the weighting coefficients, and It can be dynamically adjusted according to the actual exercise scenario and user needs (e.g., increasing size in rehabilitation scenarios). The weighting, increased in fitness scenarios (weight).

[0045] It should be understood that the prediction error loss (The first or second prediction error loss function) is used to ensure that the deviation between the model's predicted value and the actual value is minimized. An improved mean squared error loss is adopted, and a dynamic weighting factor is introduced to give higher weight to the prediction error of nodes with abrupt changes in motion state, thereby improving the prediction accuracy of key nodes. The expression is as follows: , Where N is the number of samples; The actual value of the i-th sample (actual value of auxiliary torque or motion intention label); This is the model's predicted value for the i-th sample (i.e., the human motion intention prediction label and the auxiliary torque prediction value). As a dynamic weighting factor, when a sudden change in motion state (such as a change in posture or motion intensity) is detected, ,otherwise The dynamic weighting factor is determined in real time through a motion state change detection algorithm.

[0046] Specifically, torque smoothness loss (i.e., the torque smoothness loss function) is used to suppress the chattering of the auxiliary torque, ensure the smoothness of the torque output, and improve wearing comfort. It is calculated using the sum of the squares of the differences between the predicted auxiliary torque values ​​at adjacent times, and the expression is as follows: , in, Let be the predicted value of the auxiliary torque at time i (i.e., the predicted value of the auxiliary torque). This is the predicted value of the auxiliary torque at time i-1. This loss term can prevent abrupt changes in the auxiliary torque output by the model, thus reducing the impact on human joints.

[0047] It should be understood that sports safety loss (i.e., the motion safety loss function) is used to ensure that the auxiliary torque output is within a safe range, avoiding motion injuries caused by over-assistance or under-assistance. A piecewise loss function is used, and a penalty loss is generated when the predicted auxiliary torque value exceeds the safe threshold range. The expression is as follows: , in, This is the minimum safe threshold for the auxiliary torque (i.e., the preset minimum safe threshold for the auxiliary torque). The maximum safe threshold for the assist torque (i.e., the preset maximum safe threshold for the assist torque) is dynamically set based on the user's weight, muscle strength, and activity scenario (e.g., for elderly people). (Lower than the athlete's); this loss factor ensures that the auxiliary torque is always within a safe range, improving sports safety.

[0048] Specifically, personalized adaptation loss (The personalized adaptation loss function) is used to adapt to the different exercise habits and muscle states of different users, reducing control bias caused by individual differences. It is calculated using the deviation between the user's personalized features and the model's predicted values, and the expression is as follows: , in, The user-personalized coefficient is pre-calibrated based on the user's muscle strength, posture habits, and range of motion of joints. Let $\frac{i}{i}$ be the user-personalized auxiliary torque target value corresponding to the $i$-th sample. This loss term allows the model to adapt to the personalized needs of different users and improve the targeting of control.

[0049] It should be understood that the advantages of this improved loss function are: it breaks through the limitations of the existing single loss function, and by integrating multiple loss terms, it takes into account the model's prediction accuracy, torque smoothness, motion safety and personalized adaptation needs, and solves the problem of "emphasizing accuracy, neglecting experience and lacking safety" in the training of existing models, so that the trained model is more in line with the actual wear and motion needs.

[0050] In the above embodiments, loss function analysis is performed on all human motion intention prediction labels, all auxiliary torque prediction values, all human motion intention real labels, all user-personalized auxiliary torque target values, and all auxiliary torque real values ​​to obtain the target loss function. This breaks through the limitations of the existing single loss function. By fusing multiple sub-losses, it takes into account the needs of model prediction accuracy, torque smoothness, motion safety, and personalized adaptation. It solves the problems of emphasizing accuracy, neglecting user experience, and lacking safety in the training of existing models, making the trained model more in line with actual wearable and exercise needs.

[0051] Alternatively, as another embodiment of the present invention, the present invention relates to the field of wearable device control technology, specifically to a wearable sports assistive device control method, which is particularly applicable to wearable sports assistive devices such as lower limb exoskeletons and sports assistive clothing, and can realize accurate recognition of movement intentions and adaptive adjustment of assistive torque, thereby improving human-machine coordination and exercise comfort.

[0052] Optionally, as another embodiment of the present invention, in view of the technical defects of the control model of wearable sports assistive devices in the prior art, such as poor generalization ability, slow response, unreasonable loss function design, and insufficient control accuracy, the present invention provides a control method for wearable sports assistive devices. By improving the control model structure, the adaptability and response speed of the model to different sports scenarios and different groups of people are improved; at the same time, the loss function of model training is improved to take into account prediction accuracy, torque smoothness, sports safety and personalized needs, so as to achieve precise and adaptive control of wearable sports assistive devices, and improve human-machine collaboration effect and wearing comfort.

[0053] Optionally, as another embodiment of the present invention, the present invention periodically collects user motion data and performs online fine-tuning of the improved control model to adapt to changes in user exercise habits and muscle state, thereby further improving control accuracy and personalized adaptation capabilities.

[0054] Alternatively, as another embodiment of the present invention, the beneficial effects of the present invention compared with the prior art are as follows: (1) Strong generalization ability and fast response speed of the control model: The improved Attention-LSTM-GRU hybrid model, through the parallel design of LSTM and GRU, takes into account the extraction of long-term time-series features and short-term mutation features, and combines the attention mechanism to highlight key features and key time-series nodes, which solves the problems of poor generalization ability and slow response of existing models. It can be adapted to various sports scenarios such as walking, running, and going up and down stairs, as well as different groups such as the elderly, rehabilitation patients, and athletes. The model response time is shortened by more than 30% compared with existing technologies.

[0055] (2) High control accuracy and good wearing comfort: The improved multi-objective loss function integrates prediction error, torque smoothness, motion safety and personalized adaptation multi-dimensional objectives, which not only improves the prediction accuracy of auxiliary torque (prediction error is reduced by more than 25%), but also suppresses torque chattering, ensuring smooth and safe output of auxiliary torque, reducing the impact on human joints, and significantly improving wearing comfort and motion safety.

[0056] (3) Strong personalization and adaptation capability: Through personalized adaptation of loss terms and online adaptive calibration steps, the model can dynamically adapt to the user's exercise habits and muscle state changes, which solves the problem of high generalization and insufficient personalization of existing models. Different users can obtain the auxiliary effect that is adapted to themselves, and the scope of application is wider.

[0057] (4) Achieve closed-loop adaptive control: The device operation data is collected in real time by the sensor and fed back to the improved control model to achieve closed-loop adjustment of the auxiliary torque, further improving the control accuracy and stability. It can be widely used in sports rehabilitation, fitness, elderly mobility assistance and other fields, and has high practicality and promotion value.

[0058] Optionally, as another embodiment of the present invention, traditional analytical methods rely on fixed material parameters, which cannot adapt to different flexible materials and the dynamic changes in material mechanical properties with deformation, resulting in poor adaptability. The present invention innovatively designs an adaptive material parameter dynamic matching algorithm, pre-constructing a flexible material parameter database covering the mechanical parameters and stress-strain response models of mainstream flexible materials such as polydimethylsiloxane, thermoplastic polyurethane elastomer rubber, polyimide, and ionogel-based materials. Simultaneously, it captures the deformation data of the sensor array in real time, dynamically updating key parameters such as the material's elastic modulus and Poisson's ratio in the model, achieving real-time matching between the analytical model and the actual working state of the sensor. This eliminates the need to reconstruct analytical models for different sensor materials, adapts to various flexible substrates and sensing layer materials, and expands the adaptability range by more than 60% compared to traditional methods. It also solves the analytical error problem caused by performance drift due to repeated bending and stretching of flexible materials, improving long-term analytical stability.

[0059] Optionally, as another embodiment of the present invention, existing technologies suffer from problems such as single error correction and slow dynamic response, and cannot balance data processing efficiency and analytical accuracy, making them unsuitable for real-time monitoring scenarios. The present invention innovatively adopts a multi-source error collaborative correction mechanism, integrating correction logic for three types of core errors: first, by comparing the dual signals of dual conductive patterns, errors caused by nonlinear material deformation are corrected; second, based on a correlation model between environmental parameters and stress analysis results, deviations caused by temperature and humidity interference are corrected; and third, by using a stress transfer model between adjacent units, coupling effect errors are corrected. Simultaneously, event-driven neuromorphic coding logic is introduced to perform sparse coding processing on the sensor data, reducing data bandwidth by more than 90%. While ensuring correction accuracy, the analytical latency is shortened to less than 50ms, achieving a balance between high accuracy and high real-time performance. Furthermore, by combining the stress dispersion characteristics of gradient modulus structures, the error correction weight allocation is optimized, improving the linearity R² of the analytical results to above 0.995, making it suitable for scenarios requiring both real-time performance and accuracy, such as real-time human motion monitoring and precision operation of flexible robots.

[0060] Alternatively, as another embodiment of the present invention, existing analytical methods mostly output only a single stress value, lacking targeted visualization processing and multi-parameter integration, making it difficult to meet the application needs of different scenarios. The present invention constructs an integrated process from data acquisition, analysis, correction to visualization, not only outputting the stress values ​​of each sensing unit, but also automatically calculating key parameters such as stress peak value, stress gradient, and uniformity, generating an intuitive stress distribution cloud map, facilitating rapid identification of stress concentration areas and distribution patterns. Simultaneously, it adapts to different array densities (from low-density to high-density arrays with 1728 sensing points) and different application scenarios (flexible robots, human motion monitoring, precise motion detection, etc.), allowing for customized output parameters and visualization formats according to scenario requirements without additional adaptation processing, thus improving the practicality and versatility of the method. Furthermore, it supports multimodal mechanical parameter fusion analysis, simultaneously processing data such as pressure, shear force, and torque, further broadening application scenarios.

[0061] Alternatively, as another embodiment of the present invention, the beneficial effects of the present invention compared with the prior art are as follows: 1. Significantly improved analytical accuracy: Through a multi-field coupled analytical model and a multi-source error collaborative correction mechanism, the errors caused by stress coupling between adjacent units, nonlinear deformation of materials, and environmental interference are effectively eliminated, reducing analytical error by more than 35%, linearity R²≥0.995, and stress detection accuracy can reach ±2%. It can accurately capture local stress abrupt changes and stress gradient changes, meeting the requirements of high-precision detection.

[0062] 2. Significantly expanded adaptability: Based on the adaptive material parameter dynamic matching algorithm, it is compatible with a variety of mainstream flexible materials and sensors with different array densities. There is no need to rebuild the analytical model. It is also compatible with small deformation and large deformation (maximum strain range can reach 3.3%) scenarios. The adaptability range is more than 60% wider than the traditional method, making it more adaptable.

[0063] 3. Balancing real-time performance and stability: Through sparse coding and optimized computational logic, the parsing latency is reduced to less than 50ms, enabling real-time analysis of dynamic stress. At the same time, the multi-source error correction and dynamic material parameter update mechanism improve the stability of the analysis results, maintaining excellent parsing performance even after 30,000 cycles of testing, and significantly enhancing anti-interference capabilities.

[0064] 4. Enhanced practicality: The integrated analysis and visualization process can output multi-dimensional stress parameters and intuitive stress distribution cloud maps, adapting to various scenarios such as flexible robot operation, human motion monitoring, precise assembly, and pipeline inspection. No additional data processing tools are required, lowering the application threshold. At the same time, it can support closed-loop control of intelligent equipment and customized solutions.

[0065] Alternatively, as another embodiment of the present invention, the real-time control of the present invention is as follows: After elderly people wear lower limb exoskeletons, sensors collect motion data in real time. After preprocessing, the data is input into an improved control model that has been trained. The model outputs motion intention recognition results and optimal auxiliary torque commands for the three joints in real time. The actuators (motors) output auxiliary torques according to the commands. For example, when walking on flat ground, the auxiliary torque for the hip joint is 10-15 N·m, the auxiliary torque for the knee joint is 8-12 N·m, and the auxiliary torque for the ankle joint is 5-8 N·m. When going up and down stairs, the auxiliary torques for the hip and knee joints are increased to 15-20 N·m to meet the force requirements of walking up stairs. At the same time, the torque sensors feed back the actual output auxiliary torques to the model. The model adjusts the predicted values ​​in real time based on the feedback error to achieve closed-loop control.

[0066] Optionally, as another embodiment of the present invention, the present invention collects the user's exercise data (approximately 10 minutes) every 7 days of use, performs online fine-tuning of the model (training rounds of 10, learning rate reduced to 0.0001), and updates the personalization coefficients. By adjusting to safety thresholds and adapting to changes in user exercise habits and muscle state, control precision is further improved.

[0067] Figure 2 This is a block diagram of a wearable sports assistive device control device provided in an embodiment of the present invention.

[0068] Alternatively, as another embodiment of the present invention, such as Figure 2 As shown, a wearable sports assistive device control device includes: The first data acquisition module is used to acquire multiple raw human motion data, raw electromyography data corresponding to each of the raw human motion data, raw pressure data corresponding to each of the raw human motion data, and raw device operation data corresponding to each of the raw human motion data from the sensors installed in the wearable sports assist device. The feature analysis module is used to perform feature analysis on each of the original human motion data, the original electromyography data corresponding to each of the original human motion data, the original pressure data corresponding to each of the original human motion data, and the original device operation data corresponding to each of the original human motion data, to obtain the original human motion features, the original electromyography features, the original pressure features, and the original device operation features corresponding to each of the original human motion data. The model analysis module is used to construct a training model. It performs model analysis on the training model using all the original human motion features, all the original electromyographic features, the original pressure features, and all the original device operation features to obtain the target control model. The second data acquisition module is used to acquire human motion data to be controlled, electromyographic data to be controlled, pressure data to be controlled, and operating data of the device to be controlled from the sensors installed in the wearable sports assist device; The control module is used to predict the human motion data, electromyography data, pressure data, and device operation data to be controlled through the target control model, to obtain the target human motion intention and the target auxiliary torque value, and to control the wearable motion assist device according to the target human motion intention and the target auxiliary torque value.

[0069] Optionally, as an embodiment of the present invention, the feature analysis module is specifically used for: The original electromyography (EMG) data are filtered using a wavelet filtering algorithm to obtain filtered EMG data corresponding to each of the original human motion data. The original human motion data, the original pressure data corresponding to each original human motion data, and the original device operation data corresponding to each original human motion data are filtered using a moving average filtering algorithm to obtain the filtered human motion data, the filtered pressure data, and the filtered device operation data corresponding to each original human motion data. Normalization processing is performed on each of the filtered electromyography data, the filtered human motion data corresponding to each of the original human motion data, the filtered pressure data corresponding to each of the original human motion data, and the filtered device operation data corresponding to each of the original human motion data to obtain normalized electromyography data, normalized human motion data, normalized pressure data, and normalized device operation data corresponding to each of the original human motion data. Feature extraction is performed on each of the normalized electromyography data, the normalized human motion data corresponding to each of the original human motion data, the normalized pressure data corresponding to each of the original human motion data, and the normalized device operation data corresponding to each of the original human motion data to obtain the original human motion features, the original electromyography features, the original pressure features, and the original device operation features corresponding to each of the original human motion data.

[0070] Optionally, another embodiment of the present invention provides a wearable sports assistive device control system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the wearable sports assistive device control method as described above. This system can be a computer or similar system.

[0071] Optionally, another embodiment of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the wearable motion assist device control method as described above.

[0072] 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.

[0073] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described apparatus and unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0074] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0075] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention, depending on actual needs.

[0076] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0077] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0078] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A control method for a wearable motion assistive device, characterized in that, Includes the following steps: Multiple raw human motion data, raw electromyography data corresponding to each raw human motion data, raw pressure data corresponding to each raw human motion data, and raw device operation data corresponding to each raw human motion data are obtained from sensors installed in the wearable sports assist device. Feature analysis is performed on each of the original human motion data, the original electromyography data corresponding to each of the original human motion data, the original pressure data corresponding to each of the original human motion data, and the original device operation data corresponding to each of the original human motion data to obtain the original human motion features, the original electromyography features, the original pressure features, and the original device operation features corresponding to each of the original human motion data. A training model is constructed, and the training model is analyzed using all the original human motion features, all the original electromyographic features, the original pressure features, and all the original device operation features to obtain the target control model. The wearable motion assist device obtains the human motion data to be controlled, the electromyography data to be controlled, the pressure data to be controlled, and the operating data of the device to be controlled from the sensors installed on the wearable motion assist device. The target control model predicts the human motion data, electromyography data, pressure data, and device operation data to be controlled, thereby obtaining the target human motion intention and the target auxiliary torque value. The wearable motion assist device is then controlled based on the target human motion intention and the target auxiliary torque value.

2. The control method for wearable sports assistive devices according to claim 1, characterized in that, The process of performing feature analysis on each of the original human motion data, the corresponding original electromyographic data, the corresponding original pressure data, and the corresponding original device operation data to obtain the original human motion features, the corresponding original electromyographic features, the corresponding original pressure features, and the corresponding original device operation features corresponding to each of the original human motion data includes: The original electromyography (EMG) data are filtered using a wavelet filtering algorithm to obtain filtered EMG data corresponding to each of the original human motion data. The original human motion data, the original pressure data corresponding to each original human motion data, and the original device operation data corresponding to each original human motion data are filtered using a moving average filtering algorithm to obtain the filtered human motion data, the filtered pressure data, and the filtered device operation data corresponding to each original human motion data. Normalization processing is performed on each of the filtered electromyography data, the filtered human motion data corresponding to each of the original human motion data, the filtered pressure data corresponding to each of the original human motion data, and the filtered device operation data corresponding to each of the original human motion data to obtain normalized electromyography data, normalized human motion data, normalized pressure data, and normalized device operation data corresponding to each of the original human motion data. Feature extraction is performed on each of the normalized electromyography data, the normalized human motion data corresponding to each of the original human motion data, the normalized pressure data corresponding to each of the original human motion data, and the normalized device operation data corresponding to each of the original human motion data to obtain the original human motion features, the original electromyography features, the original pressure features, and the original device operation features corresponding to each of the original human motion data.

3. The control method for wearable motion assistive devices according to claim 1, characterized in that, The training model includes a batch normalization layer, a first fully connected network, a residual connection layer, a temporal feature extraction layer, an attention mechanism layer, a softmax activation function layer, and a linear activation function layer. The process of analyzing the training model using all the original human motion characteristics, all the original electromyographic characteristics, all the original pressure characteristics, and all the original device operation characteristics to obtain the target control model includes: The batch normalization layer normalizes each of the original human motion features, the original electromyographic features corresponding to each of the original human motion data, the original pressure features corresponding to each of the original human motion data, and the original device operation features corresponding to each of the original human motion data, respectively, to obtain normalized human motion features, normalized electromyographic features, normalized pressure features, and normalized device operation features corresponding to each of the original human motion data. The first fully connected network is used to fuse the normalized human motion features, the normalized electromyographic features corresponding to the original human motion data, the normalized pressure features corresponding to the original human motion data, and the normalized device operation features corresponding to the original human motion data to obtain the original fused features corresponding to the original human motion data. Each of the original fusion features is mapped to obtain the mapped fusion features corresponding to each of the original human motion data. The residual connection layer performs feature superposition on each of the normalized human motion features, the normalized electromyographic features corresponding to each of the original human motion data, the normalized pressure features corresponding to each of the original human motion data, the normalized device operation features corresponding to each of the original human motion data, and the mapped fusion features corresponding to each of the original human motion data, to obtain the target fusion features corresponding to each of the original human motion data. The temporal feature extraction layer extracts features from each of the target fusion features to obtain target temporal features corresponding to each of the original human motion data. The attention mechanism layer performs attention analysis on each of the target temporal features to obtain target attention features corresponding to each of the original human motion data. The attention features of each target are predicted by the softmax activation function layer to obtain human motion intention prediction labels corresponding to each of the original human motion data. The linear activation function layer is used to predict the attention features of each target, thereby obtaining the auxiliary torque prediction value corresponding to each of the original human motion data. Import the real human motion intention labels corresponding to each of the original human motion data, the real values ​​of the auxiliary torque corresponding to each of the original human motion data, the user's personalized motion data corresponding to each of the original human motion data, and the user's reference trajectory corresponding to each of the original human motion data. The user-personalized motion data and the user reference trajectory corresponding to each of the original human motion data are fused to obtain the user-personalized auxiliary torque target value corresponding to each of the original human motion data. A loss function analysis is performed on all the human motion intention prediction labels, all the auxiliary torque prediction values, all the human motion intention real labels, all the user-personalized auxiliary torque target values, and all the auxiliary torque real values ​​to obtain the target loss function; The parameters of the trained model are updated according to the target loss function to obtain the target control model.

4. The control method for wearable motion assistive devices according to claim 3, characterized in that, The temporal feature extraction layer includes an LSTM network and a GRU network; The process of extracting features from each of the target fusion features through the temporal feature extraction layer to obtain the target temporal features corresponding to each of the original human motion data includes: The LSTM network is used to capture the fused features of each target to obtain long-time features corresponding to each of the original human motion data. The GRU network is used to capture the features of each target fusion feature to obtain short-term mutation features corresponding to each of the original human motion data. Each of the long-term temporal features and the short-term mutation features corresponding to each of the original human motion data are concatenated to obtain the target temporal features corresponding to each of the original human motion data.

5. The control method for a wearable motion assistive device according to claim 3, characterized in that, The attention mechanism layer includes a global average pooling layer, a global max pooling layer, a second fully connected network, a first sigmoid activation function layer, a convolutional layer, and a second sigmoid activation function layer. The process of performing attention analysis on each of the target temporal features through the attention mechanism layer to obtain the target attention features corresponding to each of the original human motion data includes: The global average pooling layer is used to perform average pooling on each of the target temporal features to obtain the first pooled temporal features corresponding to each of the original human motion data. The global max pooling layer is used to perform max pooling on each of the target temporal features to obtain the second pooled temporal features corresponding to each of the original human motion data. The second fully connected network is used to fuse each of the first pooled temporal features and the second pooled temporal features corresponding to each of the original human motion data to obtain the fused features to be processed corresponding to each of the original human motion data. The first Sigmoid activation function layer is used to map each of the fusion features to be processed, thereby obtaining the channel attention weight vector corresponding to each of the original human motion data. The channel attention features corresponding to each of the original human motion data are obtained by multiplying each of the target temporal features and the channel attention weight vector corresponding to each of the original human motion data element by element. Each of the target temporal features is subjected to channel-dimensional average pooling to obtain a third pooled temporal feature corresponding to each of the original human motion data. Each of the target temporal features is subjected to channel-dimensional max pooling to obtain a fourth pooled temporal feature corresponding to each of the original human motion data. The convolutional layer is used to fuse the third pooling time-series features and the fourth pooling time-series features corresponding to the original human motion data to obtain the fused features corresponding to the original human motion data. The second Sigmoid activation function layer is used to map each of the processed fused features to obtain a spatial attention weight vector corresponding to each of the original human motion data. Each of the spatial attention weight vectors and the channel attention features corresponding to each of the original human motion data are multiplied element-wise to obtain the target attention features corresponding to each of the original human motion data.

6. The control method for a wearable motion assistive device according to claim 3, characterized in that, The process of performing loss function analysis on all the predicted human motion intention labels, all the predicted auxiliary torque values, all the actual human motion intention labels, all the user-personalized auxiliary torque target values, and all the actual auxiliary torque values ​​to obtain the target loss function includes: The target loss function is obtained by calculating all the predicted human motion intention labels, all the predicted auxiliary torque values, all the actual human motion intention labels, all the user-personalized auxiliary torque target values, and all the actual auxiliary torque values ​​using the first formula. The first formula is: ， in, , , , , , in, Let be the target loss function. , , , as well as All are weighting coefficients. , Let the first prediction error loss function be... This is the second prediction error loss function. Let the torque smoothness loss function be... Let the motion safety loss function be... To personalize the loss function, This represents the total number of original human motion data. For the first The first dynamic weighting factor corresponding to each original human motion data point. For the first The actual value of the auxiliary torque corresponding to each original human motion data. For the first The predicted auxiliary torque value corresponding to each original human motion data. For the first The second dynamic weighting factor corresponding to each original human motion data. For the first The true label of human movement intention corresponding to each piece of raw human movement data For the first Human movement intention prediction labels corresponding to each piece of raw human movement data For the first The predicted auxiliary torque value corresponding to each original human motion data. To preset the minimum safe threshold for auxiliary torque, To preset the maximum safe threshold for auxiliary torque, For the first User personalization coefficients corresponding to each set of raw human motion data. For the first The user-personalized auxiliary torque target value corresponding to each original human motion data.

7. A control device for a wearable sports assistive device, characterized in that, include: The first data acquisition module is used to acquire multiple raw human motion data, raw electromyography data corresponding to each of the raw human motion data, raw pressure data corresponding to each of the raw human motion data, and raw device operation data corresponding to each of the raw human motion data from the sensors installed in the wearable sports assist device. The feature analysis module is used to perform feature analysis on each of the original human motion data, the original electromyography data corresponding to each of the original human motion data, the original pressure data corresponding to each of the original human motion data, and the original device operation data corresponding to each of the original human motion data, to obtain the original human motion features, the original electromyography features, the original pressure features, and the original device operation features corresponding to each of the original human motion data. The model analysis module is used to construct a training model. It performs model analysis on the training model using all the original human motion features, all the original electromyographic features, the original pressure features, and all the original device operation features to obtain the target control model. The second data acquisition module is used to acquire human motion data to be controlled, electromyographic data to be controlled, pressure data to be controlled, and operating data of the device to be controlled from the sensors installed in the wearable sports assist device; The control module is used to predict the human motion data, electromyography data, pressure data, and device operation data to be controlled through the target control model, to obtain the target human motion intention and the target auxiliary torque value, and to control the wearable motion assist device according to the target human motion intention and the target auxiliary torque value.

8. The wearable sports assistive device control device according to claim 7, characterized in that, The feature analysis module is specifically used for: The original electromyography (EMG) data are filtered using a wavelet filtering algorithm to obtain filtered EMG data corresponding to each of the original human motion data. The original human motion data, the original pressure data corresponding to each original human motion data, and the original device operation data corresponding to each original human motion data are filtered using a moving average filtering algorithm to obtain the filtered human motion data, the filtered pressure data, and the filtered device operation data corresponding to each original human motion data. Normalization processing is performed on each of the filtered electromyography data, the filtered human motion data corresponding to each of the original human motion data, the filtered pressure data corresponding to each of the original human motion data, and the filtered device operation data corresponding to each of the original human motion data to obtain normalized electromyography data, normalized human motion data, normalized pressure data, and normalized device operation data corresponding to each of the original human motion data. Feature extraction is performed on each of the normalized electromyography data, the normalized human motion data corresponding to each of the original human motion data, the normalized pressure data corresponding to each of the original human motion data, and the normalized device operation data corresponding to each of the original human motion data to obtain the original human motion features, the original electromyography features, the original pressure features, and the original device operation features corresponding to each of the original human motion data.

9. A wearable sports assistive device control device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the wearable motion assist device control method as described in any one of claims 1 to 6.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the wearable motion assist device control method as described in any one of claims 1 to 6.

Images