An intelligent assistance method, system, device and medium for rehabilitation care

Abstract

The application relates to an intelligent auxiliary method, system, device and medium for rehabilitation nursing, the method comprising: acquiring original sensing data of key parts of a patient's body, obtaining multi-modal time series data through time synchronization calibration and signal reliability evaluation; constructing a compensation-sensitive feature set and outputting a compensation action category probability and a comprehensive compensation score through a double-flow neural network; fusing physiological and facial expression data to generate dynamic safety constraint parameters; predicting a motion intention through a long short-term memory network and combining the safety constraint parameters to generate impedance control instructions; acquiring actual motion state data to obtain a time series deviation index; constructing an implicit non-compliance feature vector and classifying to obtain a non-compliance behavior type; and generating intervention instructions by matching intervention strategies through a reinforcement learning framework. The method can guarantee the intelligence and adaptability of the patient's rehabilitation training process.

Description

Technical Field

[0001] This invention belongs to the field of intelligent assistive technology for rehabilitation nursing, and in particular relates to an intelligent assistive method, system, device and medium for rehabilitation nursing. Background Technology

[0002] With the development of rehabilitation medicine and technologies in the fields of intelligent sensing and artificial intelligence, intelligent assistive technologies for rehabilitation nursing have emerged.

[0003] Intelligent rehabilitation nursing assistance technology can integrate multiple types of sensing devices and intelligent algorithms to achieve automated assistance in rehabilitation training. It can effectively make up for the shortcomings of uneven distribution of rehabilitation medical resources, promote the extension of rehabilitation training from professional medical institutions to home scenarios, and become an important technological direction for addressing social problems such as the increasing aging population and the growing number of patients with chronic diseases.

[0004] However, current intelligent assistive technologies for rehabilitation and nursing still have many technical shortcomings and application limitations when adapted to home rehabilitation scenarios, making it difficult to meet the actual rehabilitation needs of home settings: First, insufficient action recognition capabilities. Computer vision-based pose estimation methods struggle to distinguish subtle differences between target and compensatory movements. Due to the high temporal overlap and spatial coupling between compensatory and target movements, traditional action segmentation algorithms cannot effectively separate minute compensatory movements, easily leading to the solidification of abnormal movement patterns, muscle overuse, and even secondary injury to the patient. Second, the adjustment of training intensity lacks scientific rigor. Existing technologies only assess patient fatigue through physiological indicators such as heart rate and electromyography signals, neglecting psychological endurance and physiological fatigue. The nonlinear coupling relationship between labor and rehabilitation makes it impossible to dynamically adjust training parameters according to the patient's psychological state, which can easily lead to insufficient or excessive training intensity and affect rehabilitation results. Third, the model lacks adaptive updating ability and often uses static models that remain unchanged after training. This makes it impossible to adapt to the data distribution drift caused by the patient's rehabilitation progress, resulting in prominent model aging problems. Models in the early stages of rehabilitation are prone to prediction bias in the later stages, and the output training parameters may be too conservative or aggressive. Fourth, the lack of adaptability, such as the unstructured environment of the home setting and sensor wearing deviation, further exacerbates the above technical defects. Moreover, the lack of real-time guidance and intervention from professionals in the home setting makes it difficult to guarantee the accuracy of the system's perception, assessment, and control. Summary of the Invention

[0005] Therefore, it is necessary to address the aforementioned technical issues by providing an intelligent auxiliary method, system, device, and medium for rehabilitation care that enables precise perception, intelligent assessment, and adaptive control of the training process in a home rehabilitation setting, thereby improving the effectiveness and safety of rehabilitation training.

[0006] In a first aspect, this application provides an intelligent assistive method for rehabilitation nursing, comprising:

[0007] S101. Acquire raw sensor data from key parts of the patient's body, and perform time-series synchronization calibration and dynamic evaluation of signal reliability on the raw sensor data to obtain multimodal time-series data; wherein, the multimodal time-series data includes electromyographic signal data, inertial measurement unit data, physiological signal data and facial expression data;

[0008] S102. Based on multimodal time series data, a compensation-sensitive feature set is constructed, and the compensation-sensitive feature set is input into a two-stream neural network based on spatiotemporal attention mechanism to output the probability of compensation action category and comprehensive compensation score at each time step.

[0009] S103. Based on physiological signal data and facial expression data, generate dynamic safety constraint parameters; among which, dynamic safety constraint parameters are used to characterize the patient's current physiological fatigue limit and psychological endurance boundary;

[0010] S104. Input electromyography signal data and inertial measurement unit data into long short-term memory network to obtain motion intention prediction value and prediction confidence interval corresponding to the motion intention prediction value. Based on the motion intention prediction value, the prediction confidence interval corresponding to the motion intention prediction value and dynamic safety constraint parameters, generate impedance control command. The impedance control command is used to instruct the exoskeleton robot to provide auxiliary torque according to the predicted motion intention.

[0011] S105. Obtain the actual motion state data of the exoskeleton robot assisting the patient in rehabilitation nursing according to the impedance control command, extract the action timing features from the actual motion state data, obtain the action timing features, and compare the action timing features with the preset rehabilitation nursing prescription standard timing template to obtain the timing deviation index.

[0012] S106. Based on the temporal deviation index, the probability of compensatory action category and the comprehensive compensatory score, a latent nonconformity feature vector is constructed, and the latent nonconformity feature vector is input into a pre-trained classifier to obtain the latent nonconformity behavior type.

[0013] S107. Using a reinforcement learning framework, select behavioral intervention strategies that match the implicit nonconformity behavior type from a pre-built intervention strategy library and generate intervention instructions; wherein, the intervention instructions are used to instruct the exoskeleton robot to adjust the assistance mode.

[0014] Secondly, this application also provides an intelligent assistive system for rehabilitation care, comprising:

[0015] The data preprocessing module is used to acquire raw sensor data from key parts of the patient's body, and to perform time-series synchronization calibration and dynamic evaluation of signal reliability on the raw sensor data to obtain multimodal time-series data. The multimodal time-series data includes electromyographic signal data, inertial measurement unit data, physiological signal data, and facial expression data.

[0016] The compensatory action recognition module is used to construct a compensatory sensitive feature set based on multimodal time series data, and input the compensatory sensitive feature set into a two-stream neural network based on a spatiotemporal attention mechanism to output the compensatory action category probability and comprehensive compensatory score at each time step.

[0017] The safety constraint generation module is used to generate dynamic safety constraint parameters based on physiological signal data and facial expression data; among them, the dynamic safety constraint parameters are used to characterize the patient's current physiological fatigue limit and psychological endurance boundary;

[0018] The impedance control command generation module is used to input electromyography signal data and inertial measurement unit data into the long short-term memory network to obtain the motion intention prediction value and the prediction confidence interval corresponding to the motion intention prediction value. Based on the motion intention prediction value, the prediction confidence interval corresponding to the motion intention prediction value and the dynamic safety constraint parameters, the impedance control command is generated. The impedance control command is used to instruct the exoskeleton robot to provide auxiliary torque according to the predicted motion intention.

[0019] The compliance assessment module is used to obtain the actual motion state data of the exoskeleton robot assisting the patient in rehabilitation care according to the impedance control command. The module extracts the action timing features from the actual motion state data to obtain the action timing features, and compares the action timing features with the preset rehabilitation care prescription standard timing template to obtain the timing deviation index.

[0020] The nonconformity behavior identification module is used to construct a latent nonconformity feature vector based on the temporal deviation index, the probability of the compensatory action category, and the comprehensive compensatory score. The latent nonconformity feature vector is then input into a pre-trained classifier to obtain the latent nonconformity behavior type.

[0021] The instruction generation module is used to select behavioral intervention strategies that match the implicit noncompliance behavior type from a pre-built intervention strategy library using a reinforcement learning framework, and generate intervention instructions; wherein, the intervention instructions are used to instruct the exoskeleton robot to adjust the assistance mode.

[0022] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0023] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0024] The aforementioned intelligent assistive methods, systems, devices, and media for rehabilitation nursing, by acquiring and calibrating electromyographic signals, inertial measurement units, physiological signals, and facial expression data from key parts of the patient's body, can lay a data foundation for accurate perception and intelligent assessment in home rehabilitation scenarios. By constructing a compensatory sensitive feature set and achieving accurate identification and scoring of compensatory movements based on a two-stream neural network with a spatiotemporal attention mechanism, it can effectively distinguish subtle differences between target movements and compensatory movements, solving the problem of blind spots in the recognition of minute compensatory movements. By fusing physiological signals and facial expression data to generate dynamic safety constraint parameters characterizing the patient's physiological fatigue limits and psychological tolerance boundaries, it can achieve comprehensive fatigue assessment in both physiological and psychological dimensions, overcoming the limitations of assessment based on a single physiological indicator. This method addresses several key limitations of existing technologies. First, by predicting movement intentions using a long short-term memory network and generating impedance control commands based on dynamic safety constraint parameters, the exoskeleton robot can accurately match the patient's movement intentions and provide appropriate assistive torques, achieving adaptive control in rehabilitation training. Second, by extracting the temporal features of actual movements and comparing them with standardized templates to obtain a temporal deviation index, and combining this with compensatory movement recognition results to construct a latent nonconformity feature vector and accurately classify behavior types, it can effectively identify latent nonconformity behaviors in patient rehabilitation training and promptly detect training quality deviations. Third, by using a reinforcement learning framework to match and generate targeted intervention commands from a pre-built intervention strategy library, the exoskeleton robot can dynamically adjust its assistive mode to achieve intelligent intervention for latent nonconformity behaviors. This approach overcomes the technical shortcomings of existing technologies, such as insufficient movement recognition capabilities, unscientific training intensity adjustment, lack of adaptive control, and inadequate intervention for nonconformity behaviors, while ensuring the intelligence and adaptability of the rehabilitation training process. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A flowchart illustrating an intelligent assistive method for rehabilitation nursing provided in one embodiment of this application;

[0027] Figure 2 This is a schematic diagram of the structure of an intelligent assistive system for rehabilitation nursing provided in one embodiment of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0029] In one embodiment, such as Figure 1 As shown, an intelligent assistive method for rehabilitation nursing is provided. This embodiment illustrates the application of this method to an intelligent assistive terminal. It is understood that this method can also be applied to an intelligent assistive server, and further to a system including both an intelligent assistive terminal and an intelligent assistive server, and is implemented through the interaction between the intelligent assistive terminal and the intelligent assistive server. In this embodiment, the method includes the following steps:

[0030] S101. Obtain raw sensor data of key parts of the patient's body, and perform time-series synchronization calibration and dynamic evaluation of signal reliability on the raw sensor data to obtain multimodal time-series data.

[0031] Optionally, multimodal time-series data may include, but is not limited to, electromyographic signal data, inertial measurement unit data, physiological signal data, and facial expression data.

[0032] Specifically, intelligent assistive terminals can acquire raw sensor data through various types of sensing devices deployed in key parts of the patient's body.

[0033] Furthermore, the intelligent auxiliary terminal can use a unified system timestamp as a benchmark to eliminate timing discrepancies caused by sampling frequency and transmission delays in different sensing devices, thereby obtaining raw sensing data aligned in the time dimension. Secondly, the intelligent auxiliary terminal can perform real-time calculations on indicators such as signal-to-noise ratio, data missing rate, and outlier ratio of the raw sensing data aligned in the time dimension, eliminating invalid and distorted sensing data, and correcting low-reliability data to integrate multi-dimensional, high-reliability multimodal time-series data.

[0034] S102. Based on multimodal time series data, a compensation-sensitive feature set is constructed, and the compensation-sensitive feature set is input into a two-stream neural network based on a spatiotemporal attention mechanism to output the probability of the compensation action category and the comprehensive compensation score at each time step.

[0035] Optionally, the compensatory sensitivity feature set can be used to characterize the core physiological motor features related to compensatory behavior during the patient's rehabilitation training process, as well as the deviation of the core physiological motor features from the standard features of the rehabilitation movements, the dynamic trend of change, and the timing, form, and severity of the compensatory behavior.

[0036] Optionally, the probability of compensatory movement categories can be used to characterize the confidence level of a patient in the occurrence of compensatory movements at each time step of the rehabilitation training movement execution.

[0037] Optionally, the comprehensive compensation score can be used to characterize the overall severity, frequency, and overall impact of compensatory behaviors during a complete rehabilitation training cycle.

[0038] Optionally, the construction method of a two-stream neural network based on a spatiotemporal attention mechanism may include:

[0039] S1021. Construct the basic structure of a two-stream neural network based on a spatiotemporal attention mechanism.

[0040] Preferably, the basic network of the two-stream neural network based on the spatiotemporal attention mechanism may include, but is not limited to, a spatial feature learning branch, a temporal feature learning branch, and a spatiotemporal feature fusion layer.

[0041] Optionally, the spatial feature learning branch can extract spatial dimension correlation features from the compensation-sensitive feature set by stacking convolutional layers and pooling layers; the temporal feature learning branch can extract temporal dimension evolution features from the compensation-sensitive features by recurrent neural network layers; a spatiotemporal attention mechanism module can be embedded in the spatial feature learning branch and the temporal feature learning branch respectively to strengthen the weights of key feature regions related to compensation behavior recognition, and the output features of the spatial feature learning branch and the temporal feature learning branch can be spliced ​​and fused through a spatiotemporal feature fusion layer to obtain a feature representation that integrates spatiotemporal information.

[0042] S1022. Construct the network training dataset.

[0043] Optionally, multimodal time-series data of rehabilitation training from multiple patients can be collected and labeled with the corresponding compensatory movement categories, time step labels, and comprehensive compensatory scores. Secondly, labeled training, validation, and test sets can be constructed based on the multimodal time-series data of rehabilitation training.

[0044] S1023. Complete the training and performance optimization of the two-stream neural network.

[0045] Optionally, the training set can be input into the constructed two-stream neural network, using the weighted sum of classification and regression losses as the loss function. Next, the gradient descent algorithm can be used to iteratively update the weights and bias parameters of the two-stream neural network. The training effect of the two-stream neural network can be monitored in real time using a validation set, and the hyperparameters of the two-stream neural network can be adjusted. After the two-stream neural network has been trained to convergence, the accuracy of compensatory action recognition and the scoring prediction effect of the two-stream neural network can be verified using a test set. Through multiple iterative optimizations, a two-stream neural network based on a spatiotemporal attention mechanism whose performance meets the preset requirements can be obtained.

[0046] Specifically, the intelligent assistive terminal can extract highly correlated feature information with compensatory behavior from different dimensions of data, such as electromyography signal data and inertial measurement unit data, when acquiring multimodal time-series data, to obtain multi-dimensional initial compensatory features. Secondly, the intelligent assistive terminal can perform feature validity verification and feature screening on the multi-dimensional initial compensatory features to construct a compensatory sensitive feature set.

[0047] Furthermore, the intelligent auxiliary terminal can perform standardized preprocessing on the compensatory sensitive feature set. By eliminating the dimensional differences between different feature dimensions in the compensatory sensitive feature set, the feature data of the compensatory sensitive feature set can be matched with the input specification requirements of a two-stream neural network based on a spatiotemporal attention mechanism.

[0048] Furthermore, the intelligent auxiliary terminal can input the pre-processed compensatory sensitive feature set into a pre-trained two-stream neural network based on a spatiotemporal attention mechanism. The spatial feature learning branch and the temporal feature learning branch of the two-stream neural network based on the spatiotemporal attention mechanism extract the spatial dimension correlation features and temporal dimension evolution features from the pre-processed compensatory sensitive feature set, respectively. The spatiotemporal attention mechanism module embedded in the two-stream neural network based on the spatiotemporal attention mechanism is used to strengthen the weights of key feature regions related to compensatory behavior recognition. Then, the spatiotemporal feature fusion layer splices and fuses the output features of the spatial feature learning branch and the temporal feature learning branch to obtain a feature representation that integrates spatiotemporal information.

[0049] Furthermore, after being processed sequentially by subsequent computational layers of a two-stream neural network based on a spatiotemporal attention mechanism, the system outputs the category probabilities corresponding to different compensatory actions at each time step during the patient's rehabilitation training. Based on the category probabilities of compensatory actions at each time step and the feature correlation, a comprehensive quantitative calculation is performed to obtain a comprehensive compensation score.

[0050] S103. Generate dynamic safety constraint parameters based on physiological signal data and facial expression data.

[0051] Optionally, dynamic safety constraint parameters can be used to characterize the patient's current physiological fatigue limits and psychological tolerance boundaries.

[0052] Specifically, the intelligent assistive terminal can eliminate noise interference and remove invalid data fragments from physiological signal data and facial expression data to obtain standardized physiological signal data and facial expression data. Secondly, based on the standardized physiological signal data, the intelligent assistive terminal can mine core quantitative indicators to characterize the patient's level of physical fatigue, obtaining quantitative results of physiological state. Furthermore, it can perform feature analysis and state recognition on the standardized facial expression data to extract key features characterizing the patient's psychological emotions and tolerance level, obtaining quantitative results of psychological state.

[0053] Furthermore, the intelligent auxiliary terminal can call the built-in analysis model to complete the coupled evaluation of the quantitative results of physiological state and the quantitative results of psychological state, and generate dynamic safety constraint parameters.

[0054] S104. Input the electromyography signal data and the inertial measurement unit data into the long short-term memory network to obtain the predicted value of the motion intention and the prediction confidence interval corresponding to the predicted value of the motion intention. Based on the predicted value of the motion intention, the prediction confidence interval corresponding to the predicted value of the motion intention, and the dynamic safety constraint parameters, generate the impedance control command.

[0055] Alternatively, impedance control commands can be used to instruct the exoskeleton robot to provide auxiliary torque in accordance with the predicted motion intention.

[0056] Optionally, the Long Short-Term Memory (LSTM) network may include, but is not limited to, an input layer, a hidden layer, and an output layer; the hidden layer may include, but is not limited to, a forget gate, an input gate, a cell state, and an output gate.

[0057] Preferably, the processing flow of the long short-term memory network for electromyographic signal data and inertial measurement unit data may include:

[0058] S1041, the intelligent auxiliary terminal can fuse the preprocessed electromyographic signal data and inertial measurement unit data into a normalized temporal feature sequence, and input the normalized temporal feature sequence into the input layer of the long short-term memory network. The input layer completes the mapping and adaptation of the feature dimensions, and the input layer transmits the mapped and adapted data to the hidden layer.

[0059] S1042. The forget gate of the hidden layer can selectively filter the cell state information of historical time steps according to preset weights, and remove redundant temporal features that are irrelevant to the current movement trend; the input gate can weight and encode the feature information of the current time step and update it to the cell state. The cell state serves as the core of feature storage and completes the continuous transfer of feature information across time steps.

[0060] S1043. The output gate of the hidden layer can generate the hidden layer features of the current time step based on the current cell state and input features. Through iterative calculations of multiple hidden layers, the deep encoding of the full-time features of electromyography signal data and inertial measurement unit data is completed, resulting in a deep feature representation that characterizes the motion trend.

[0061] S1044. The output layer of the Long Short-Term Memory network receives deep feature representations, completes the calculation of motion intention-related values ​​through preset calculation rules, outputs motion intention prediction values, and completes quantization calculation based on the feature output probability distribution of each layer of the Long Short-Term Memory network and the confidence features of the encoded features to obtain the prediction confidence interval corresponding to the motion intention prediction value.

[0062] Specifically, the intelligent assistive terminal can perform standardized preprocessing on electromyography (EMG) signal data and inertial measurement unit (IMU) data, sequentially completing noise filtering, missing temporal completion, and feature dimension normalization to obtain temporal data adapted to the input specifications of the Long Short-Term Memory (LSTM) network. This adapted temporal data is then fused into a normalized temporal feature sequence, which is input into the pre-trained LSM network. Through the collaborative operation of the LSM network's input, hidden, and output layers, deep encoding and decomposition of the full temporal features of the data are completed. The intelligent assistive terminal then obtains the network's output motion intention prediction value and the corresponding prediction confidence interval.

[0063] Furthermore, the intelligent auxiliary terminal can perform multi-dimensional fusion analysis of motion intention prediction values, prediction confidence intervals, and dynamic safety constraint parameters. Relying on a preset impedance control algorithm, it can complete the calculation of motion intention prediction values, prediction confidence intervals, and dynamic safety constraint parameters, as well as the construction of instruction logic, and generate impedance control instructions.

[0064] S105. Obtain the actual motion state data of the exoskeleton robot assisting the patient in rehabilitation nursing according to the impedance control command, extract the action timing features from the actual motion state data, obtain the action timing features, and compare the action timing features with the preset rehabilitation nursing prescription standard timing template to obtain the timing deviation index.

[0065] Optionally, the timing deviation index can be used to characterize the overall degree of deviation, specific dimensions of deviation, and quantitative magnitude of deviation of the timing sequence of the rehabilitation nursing actions performed by the patient relative to the standard timing sequence set by the clinical rehabilitation standards.

[0066] Optionally, the actual motion state data may include, but is not limited to, real-time temporal monitoring data of the exoskeleton robot's joint angles, angular velocities, output torque, and patient limb movements.

[0067] Specifically, the intelligent assistive terminal can establish a real-time data interaction link with the exoskeleton robot to obtain actual motion state data collected during the process of the exoskeleton robot assisting the patient in rehabilitation care according to impedance control commands. Secondly, the intelligent assistive terminal can sequentially perform noise filtering, temporal completion, and standardization processing on the actual motion state data to obtain preprocessed actual motion state data.

[0068] Furthermore, the intelligent assistive terminal can extract core features representing the temporal characteristics of the movement from pre-processed actual motion state data, such as the rhythm of movement execution, the duration of each stage, and the continuity of limb movements, and integrate them to obtain the temporal features. Subsequently, the intelligent assistive terminal can compare the extracted temporal features with the temporal template of the rehabilitation nursing prescription standard in multiple dimensions using preset feature distance calculation rules to obtain the temporal deviation index.

[0069] S106. Based on the temporal deviation index, the probability of compensatory action category, and the comprehensive compensatory score, a latent nonconformity feature vector is constructed, and the latent nonconformity feature vector is input into a pre-trained classifier to obtain the latent nonconformity behavior type.

[0070] Optionally, the implicit noncompliance feature vector can be used to characterize the comprehensive quantitative features of implicit noncompliance behavior of patients during the rehabilitation training cycle, manifested by deviations from the training sequence and the occurrence of compensatory behaviors.

[0071] Optionally, implicit noncompliance behavior types can be used to characterize the specific categories and core manifestations of implicit noncompliance behaviors exhibited by patients during rehabilitation training.

[0072] Specifically, the intelligent assistive terminal can integrate the temporal deviation index, the probability of compensatory action categories at each time step, and the comprehensive compensation score obtained from the aforementioned steps, and perform feature extraction and standardization preprocessing on the temporal deviation index, the probability of compensatory action categories at each time step, and the comprehensive compensation score. Secondly, the intelligent assistive terminal can perform dimensional concatenation, fusion, and reconstruction of the three types of preprocessed features to construct an implicit noncompliance feature vector.

[0073] Furthermore, the intelligent assistive terminal can then input the implicit noncompliance feature vector into a classifier that has been pre-trained based on multi-patient rehabilitation training labeled data. After the classifier completes feature matching and category determination, the implicit noncompliance behavior type corresponding to the patient's current rehabilitation training is obtained.

[0074] S107. Using a reinforcement learning framework, select behavioral intervention strategies that match the implicit noncompliance behavior type from a pre-built intervention strategy library and generate intervention instructions.

[0075] Optionally, intervention commands can be used to instruct the exoskeleton robot to adjust its assistance mode.

[0076] Optionally, the reinforcement learning framework can take the adaptability of rehabilitation training intervention as the optimization goal. The state space of the reinforcement learning framework is constructed with implicit noncompliance behavior type as the core and integrates the real-time state characteristics of the patient's current rehabilitation training. The action space is mapped to the strategy types in the pre-built intervention strategy library.

[0077] Specifically, the intelligent assistive terminal can input implicit nonconformity behavior types into a pre-built reinforcement learning framework. Secondly, the reinforcement learning framework can calculate the matching reward value of each candidate strategy based on the reward function, and select a behavioral intervention strategy highly suitable for the current implicit nonconformity behavior type from a pre-built intervention strategy library. Furthermore, the intelligent assistive terminal can perform instruction parsing and logical reconstruction on the selected intervention strategy, transforming the intervention strategy into standardized intervention instructions that the exoskeleton robot can recognize and execute.

[0078] The aforementioned intelligent assistive method for rehabilitation nursing, by acquiring and calibrating electromyographic signals, inertial measurement units, physiological signals, and facial expression data from key parts of the patient's body, lays a data foundation for accurate perception and intelligent assessment in home rehabilitation scenarios. By constructing a compensatory sensitive feature set and using a two-stream neural network based on a spatiotemporal attention mechanism to accurately identify and score compensatory movements, it can effectively distinguish subtle differences between target movements and compensatory movements, solving the problem of blind spots in the recognition of minute compensatory movements. By fusing physiological signals and facial expression data to generate dynamic safety constraint parameters characterizing the patient's physiological fatigue limits and psychological tolerance boundaries, it can achieve a comprehensive assessment of fatigue in both physiological and psychological dimensions, overcoming the limitations of assessment based on a single physiological indicator. By using a long short-term memory network to predict movement intentions and combining them with dynamic safety constraint parameters to generate impedance control commands, the exoskeleton robot can accurately match the patient's movement intentions and provide appropriate assistive torques, achieving adaptive control in rehabilitation training. By extracting the temporal features of actual movements and comparing them with standardized templates to obtain a temporal deviation index, and combining this with compensatory movement recognition results to construct a latent non-compliance feature vector and achieve accurate behavior type classification, the robot can effectively identify latent non-compliance behaviors in patient rehabilitation training and promptly detect training quality deviations. Furthermore, by using a reinforcement learning framework to match and generate targeted intervention commands from a pre-built intervention strategy library, the exoskeleton robot can dynamically adjust its assistive mode, achieving intelligent intervention for latent non-compliance behaviors. This method addresses the technical shortcomings of existing technologies, such as insufficient movement recognition capabilities, unscientific training intensity adjustment, lack of adaptive control, and inadequate intervention for non-compliance behaviors, while ensuring the intelligence and adaptability of the rehabilitation training process.

[0079] In one embodiment, constructing a compensation-sensitive feature set based on multimodal time-series data may include the following steps:

[0080] S201. Input the inertial measurement unit data into the multi-rigid-body kinematics model to obtain the joint contribution feature vector.

[0081] Among them, the joint contribution feature vector can be used to characterize the deviation of the degree of participation of each joint from the standard movement during the execution of the action.

[0082] For example, the intelligent assistive terminal can input the extracted inertial measurement unit data into a multi-rigid-body kinematics model. The multi-rigid-body kinematics model can calculate the actual motion participation of each joint during the patient's rehabilitation movements through forward and inverse kinematics operations. The actual motion participation of each joint during the patient's rehabilitation movements is then compared with the standard joint participation corresponding to the standardized rehabilitation movements in multiple dimensions. After feature extraction and dimension reconstruction, a joint contribution feature vector is obtained.

[0083] Optionally, the multi-rigid-body kinematics model can be a multi-rigid-body kinematics model based on the human skeletal structure. Optionally, the multi-rigid-body kinematics model may include, but is not limited to, a rigid body modeling layer, a kinematic constraint layer, a forward and inverse kinematics operation layer, and a feature calculation layer.

[0084] Optionally, the processing flow of inertial measurement unit data for the multi-rigid-body kinematics model may include:

[0085] S2011, the rigid body modeling layer can decompose the human body into multiple interconnected rigid links based on the anatomical parameters of the human skeleton, set the structural parameters of each rigid body segment, establish a one-to-one mapping relationship between each rigid body segment and the inertial measurement unit data of the corresponding acquisition part, complete the association and matching of data with the rigid body segments of the human body, and output the original motion data with rigid body segment labels.

[0086] S2012, the kinematic constraint layer can pre-set parameters such as the type of kinematic pairs, the direction of the rotation axis, and the physiological range of motion constraints of each joint of the human body. It applies kinematic constraints to the original motion data with rigid body segment labels, removes abnormal data that exceeds the physiological range of motion, constructs spatial motion correlation equations between rigid body segments, and outputs constrained rigid body motion time series data.

[0087] S2013, the forward and inverse kinematics operation layer can solve rigid body motion data through forward kinematics algorithm to obtain core kinematic parameters such as the actual motion participation of each joint, joint angle, and angular velocity during the execution of rehabilitation movements. It also retrieves the standard kinematic parameters of the joints for the corresponding rehabilitation movements from the standard movement database. The forward and inverse kinematics operation layer completes the dimensional alignment and initial quantization comparison of the actual parameters and standard parameters, and outputs the motion deviation of each joint from the original data.

[0088] S2014, the feature solution layer can extract features, remove redundant information and normalize dimensions from motion deviations from the original data. It obtains the deviation quantization value of the degree of participation of each joint relative to the standard movement through quantization calculation, and arranges and reconstructs the vector of each deviation quantization value according to the joint dimension order, and outputs the joint contribution feature vector.

[0089] S202. Extract the muscle synergy primitive matrix from the electromyographic signal data, and generate the electromyographic synergy deviation feature by calculating the reconstruction error of the current electromyographic signal under the muscle synergy primitive matrix and the deviation between the synergy activation coefficient and the expected activation coefficient.

[0090] Among them, the electromyographic synergy deviation feature can be used to characterize the degree of deviation between the current muscle activation mode and the normative muscle synergy mode.

[0091] For example, the intelligent assistive terminal can retrieve multi-channel surface electromyography (EMG) signals collected when the patient performs standardized rehabilitation movements as reference data. Secondly, the intelligent assistive terminal can perform matrix decomposition on the reference data using a non-negative matrix factorization algorithm to obtain a muscle coordination primitive matrix.

[0092] Furthermore, the intelligent auxiliary terminal can project the effective electromyographic signal data at the current moment onto the muscle coordination primitive matrix, and complete the reconstruction of the current electromyographic signal by solving a non-negative least squares problem, thereby obtaining the reconstruction error vector and the current coordination activation coefficient vector.

[0093] Furthermore, the intelligent assistive terminal can retrieve the expected co-activation coefficient vector that matches the phase of the patient's current action from the standardized action database, and calculate the quantization deviation value between the current co-activation coefficient vector and the expected co-activation coefficient vector. Subsequently, the intelligent assistive terminal can perform weighted fusion of the L2 norm squared sum of the quantization deviation values ​​of the reconstructed error vector to generate electromyographic co-activation deviation features.

[0094] S203. Based on inertial measurement unit data, motion smoothness features are generated by calculating the curvature of the motion trajectory and the acceleration index of the joint angular velocity of the end effector in three-dimensional space.

[0095] Among them, motion smoothness features can be used to characterize the smoothness and continuity of movements.

[0096] For example, the intelligent auxiliary terminal can perform kinematic calculations based on inertial measurement unit data and the correlation of human limb movements to obtain the real-time pose coordinate sequence of the end effector in three-dimensional space, as well as the angular velocity time-series data of each joint movement. Secondly, based on the three-dimensional coordinate sequence of the end effector, the intelligent auxiliary terminal can construct a complete three-dimensional motion trajectory using a curve fitting algorithm, and solve for the curvature value of the motion trajectory at each time step according to the calculation rules of trajectory curvature in differential geometry.

[0097] Furthermore, the intelligent auxiliary terminal can perform derivative calculations on the time-series data of the angular velocity of each joint to obtain the joint angular acceleration data, and extract the core statistical features of the joint angular acceleration data as the acceleration index of the joint angular velocity.

[0098] Furthermore, the intelligent auxiliary terminal can standardize the trajectory curvature-related quantitative values ​​and acceleration indicators, and then weight and integrate the standardized trajectory curvature-related quantitative values ​​and acceleration indicators to generate motion smoothness features.

[0099] S204. The joint contribution feature vector, electromyographic coordination deviation feature, and motion fluency feature are concatenated and fused along the feature dimension to generate a compensation-sensitive feature set.

[0100] For example, the intelligent assistive terminal can eliminate the dimensional differences and numerical range deviations among the joint contribution feature vector, electromyographic coordination deviation feature, and motion smoothness feature by using a preset normalization algorithm, and remove redundant values ​​that have no actual representational meaning from each feature, while retaining the core effective feature information.

[0101] Furthermore, the intelligent assistive terminal can vectorize the preprocessed electromyographic coordination deviation features and motion smoothness features, and then concatenate them with the original joint contribution feature vector according to the feature dimension arrangement rules to construct a joint feature matrix.

[0102] Furthermore, the intelligent auxiliary terminal can perform deep fusion of the joint feature matrix to obtain multimodal compensation fusion features, and perform dimensional regularization and validity verification on the multimodal compensation fusion features to generate a compensation-sensitive feature set.

[0103] In this embodiment, the intelligent auxiliary terminal extracts and quantifies highly correlated features of joints, electromyography, motion smoothness, and compensatory behavior from multimodal time-series data using algorithms such as multi-rigid-body kinematics models and non-negative matrix factorization. After standardization, splicing and fusion, and validity verification, a compensatory sensitive feature set is generated, which can achieve comprehensive integration and accurate representation of multi-dimensional core physiological and motor features of compensatory behavior in rehabilitation training, providing a high-quality feature foundation for the subsequent recognition of compensatory movements.

[0104] In one embodiment, extracting a muscle coordination primitive matrix from electromyographic signal data and generating an electromyographic coordination deviation feature by calculating the reconstruction error of the current electromyographic signal under the muscle coordination primitive matrix and the deviation between the coordination activation coefficient and the expected activation coefficient may include the following steps:

[0105] S301. Acquire multi-channel surface electromyography (EMG) signals when the patient performs standardized movements. Use the multi-channel EMG signals as training data and decompose the training data using a non-negative matrix factorization algorithm to obtain the muscle coordination primitive matrix.

[0106] Optionally, the muscle synergy primitive matrix can be used to characterize the core synergistic activation patterns among multiple muscles when a patient performs standard rehabilitation movements.

[0107] For example, the intelligent assistive terminal can sequentially perform noise filtering, baseline drift correction, full-wave rectification, and temporal smoothing on the acquired multi-channel surface electromyography (EMG) signals to obtain clean and standardized EMG timing data, which is then used as training data. Secondly, the intelligent assistive terminal can employ a non-negative matrix factorization algorithm to perform high-dimensional matrix factorization on the training data.

[0108] Furthermore, the nonnegative matrix factorization algorithm can decompose a high-dimensional electromyography data matrix into the product of two nonnegative matrices, extract primitive components that can characterize the muscle core co-activation features from the product of the two nonnegative matrices, and perform feature integration and matrix reconstruction on the primitive components to obtain the muscle co-activation primitive matrix.

[0109] S302. Project the electromyographic signal data collected at the current moment onto the muscle coordination primitive matrix, and reconstruct the current electromyographic signal by solving a non-negative least squares problem to obtain the coordination activation coefficient vector and reconstruction error vector at the current moment.

[0110] For example, the intelligent assistive terminal can project clean, standardized electromyographic timing data onto a pre-extracted matrix of muscle coordinating primitives in the feature space.

[0111] Furthermore, the intelligent auxiliary terminal can use the non-negative least squares method to construct the optimization objective function for signal reconstruction, take minimizing the reconstruction error as the solution criterion, solve the optimization problem under non-negativity constraints, obtain the co-activation coefficient vector at the current moment, and calculate the numerical difference between the current actual electromyographic signal data and the reconstructed signal data to obtain the reconstruction error vector at the current moment.

[0112] S303. Based on the patient's current action execution phase, retrieve the corresponding expected co-activation coefficient vector from the standardized action database, and obtain the activation coefficient deviation value by calculating the Euclidean distance between the current co-activation coefficient vector and the expected co-activation coefficient vector.

[0113] Optionally, the activation coefficient deviation value can be used to characterize the degree of quantitative deviation of the muscle synergistic activation coefficient vector from the expected synergistic activation coefficient vector of the normative action in the same phase of the patient's current action execution phase.

[0114] Optionally, the standardized movement database may include, but is not limited to, standardized electromyographic synergistic characteristic parameters of different execution phases of various rehabilitation movements.

[0115] For example, the intelligent assistive terminal can perform action phase segmentation and identification on the real-time execution process of the patient's current rehabilitation action to determine the specific action execution phase the patient is currently in. Secondly, the intelligent assistive terminal can retrieve the pre-labeled expected co-activation coefficient vector under the corresponding execution phase of the rehabilitation action from a standardized action database.

[0116] Furthermore, the intelligent assistive terminal can perform feature dimension consistency verification between the currently calculated co-activation coefficient vector and the retrieved expected co-activation coefficient vector. Secondly, based on the Euclidean distance calculation rules, the intelligent assistive terminal can calculate the Euclidean distance between the co-activation coefficient vector and the expected co-activation coefficient vector in the electromyographic co-activation feature space, and use the calculated Euclidean distance between the two vectors as the activation coefficient deviation value.

[0117] S304. Based on the weighted sum of the squared values ​​of the second norm and the deviation values ​​of the activation coefficients of the reconstruction error vector, the electromyographic coordination deviation feature is obtained.

[0118] Preferably, the expression for the electromyographic coordination deviation characteristic can be:

[0119]

[0120] In the formula, express Characteristics of electromyographic coordination deviation at any given time. express The reconstruction error vector at time step 1. express The co-activation coefficient vector at time step 1. express The expected co-activation coefficient vector at time step 1. This represents the weighting coefficients for balancing the reconstruction error term and the activation coefficient deviation term.

[0121] For example, the intelligent auxiliary terminal can access The reconstruction error vector and activation coefficient deviation at time step are used to perform a L2 norm square operation on the reconstruction error vector. This operation quantifies the overall error level of the current electromyographic signal after reconstruction under the muscle coordination primitive matrix, thus obtaining the quantized value of the reconstruction error term.

[0122] Furthermore, the intelligent auxiliary terminal can retrieve preset weighting coefficients. Based on the fundamental calculation of the activation coefficient deviation value, the difference between the current co-activation coefficient vector and the expected co-activation coefficient vector is subjected to a L2 norm square operation. Subsequently, the intelligent auxiliary terminal can combine this result with the weight coefficients... Perform a product operation to obtain the weighted activation coefficient deviation term quantization value.

[0123] Furthermore, the intelligent auxiliary terminal can sum the quantized value of the reconstruction error term and the quantized value of the weighted activation coefficient deviation term, and obtain the result through this weighted summation. Characteristics of electromyographic coordination deviation at any given time.

[0124] In this embodiment, the intelligent auxiliary terminal uses algorithms such as non-negative matrix factorization, non-negative least squares solution, and Euclidean distance calculation to perform hierarchical processing and quantitative analysis of electromyographic signal data of the standard movement and the current movement. It extracts the muscle coordination primitive matrix and calculates the electromyographic coordination deviation feature, which can achieve a comprehensive and quantitative characterization of the degree of deviation of the patient's current muscle activation mode from the standard coordination mode, providing electromyographic dimension feature support for the identification of subsequent compensatory behaviors.

[0125] In one embodiment, generating dynamic safety constraint parameters based on physiological signal data and facial expression data may include the following steps:

[0126] S401. The physiological fatigue index is obtained by weighted fusion of the time-domain index of heart rate variability, the frequency-domain index of heart rate variability, the median frequency of electromyography signal, and the skin conductance response level.

[0127] Optionally, physiological signal data may include, but are not limited to, time-domain indices of heart rate variability, frequency-domain indices of heart rate variability, median frequency of electromyography signals, and skin conductance response levels.

[0128] For example, the intelligent auxiliary terminal can extract heart rate variability time-domain indicators, heart rate variability frequency-domain indicators, median frequency of electromyography signals, and skin conductance response levels from multimodal time-series data.

[0129] Furthermore, the intelligent auxiliary terminal can eliminate the dimensional differences and numerical range deviations between the time-domain indicators of heart rate variability, the frequency-domain indicators of heart rate variability, the median frequency of electromyography signals, and the level of skin conductance response, and remove outliers and noise interference in the data to obtain the pure quantitative values ​​of each indicator.

[0130] Furthermore, the intelligent auxiliary terminal can retrieve weighting coefficients set based on the results of clinical research in rehabilitation medicine, and multiply the pure quantified values ​​of the time-domain index of heart rate variability, the frequency-domain index of heart rate variability, the median frequency of electromyography signal, and the skin conductance response level with the corresponding weighting coefficients to obtain the calculation results. The calculation results are then summed to obtain the physiological fatigue index.

[0131] S402. Extract the activation intensity of facial action units from facial expression data, and use the activation intensity of facial action units as facial expression features. Extract the high-frequency power of heart rate variability and the level of skin conductance response from physiological signal data, and use the high-frequency power of heart rate variability and the level of skin conductance response as autonomic nerve features.

[0132] Optionally, facial expression features can be used to characterize the patient's real-time facial emotional expression state, psychological stress level, and emotional arousal level during rehabilitation training.

[0133] Optionally, autonomic features can be used to characterize the real-time activity state of the autonomic nervous system during patient rehabilitation training.

[0134] For example, the intelligent assistive terminal can perform feature analysis on the activation intensity of facial motion units, extract the real-time activation intensity of each facial motion unit during the patient's training process, and complete the quantification and grading. Secondly, the intelligent assistive terminal can perform temporal normalization and feature normalization on the quantified real-time activation intensity to obtain standardized facial expression features and complete the unified normalization of feature dimensions.

[0135] Furthermore, the intelligent auxiliary terminal can filter out raw heart rate variability data and raw skin conductance response data from the physiological signal data of multimodal time series data, perform frequency domain analysis on the raw heart rate variability data to obtain the high-frequency power of heart rate variability, perform baseline drift correction and noise filtering on the raw skin conductance response data to retain its real-time quantization level value, and integrate and align the processed raw heart rate variability data and raw skin conductance response data to construct autonomic neural features.

[0136] S403. Input facial expression features and autonomic nerve features into the support vector regression model to obtain the psychological endurance index.

[0137] Optionally, the psychological resilience index can be used to characterize the patient's current psychological resilience.

[0138] Optionally, the support vector regression model may include, but is not limited to, an input layer, a kernel function mapping layer, a loss function calculation layer, an optimization solution layer, and a regression output layer.

[0139] Optionally, the processing flow of the support vector regression model for facial expression features and autonomic neural features may include:

[0140] S4031. The feature input layer receives facial expression features and autonomic neural features. It can perform dimensional concatenation and feature fusion on the facial expression features and autonomic neural features, and eliminate the dimensional differences between features through the built-in normalization operation, remove redundant feature information, and output a fused feature vector with unified dimensions and effective features.

[0141] S4032. The kernel function mapping layer can nonlinearly map the fused feature vector from the low-dimensional feature space to the high-dimensional feature space through the kernel function, explore the complex intrinsic relationship between facial expression features and autonomic neural features, solve the problem of linearly inseparable features in the low-dimensional space, and output a high-dimensional feature representation vector.

[0142] S4033, the loss function calculation layer can be based on The insensitive loss function is based on a high-dimensional feature representation vector. It calculates the error loss between the predicted feature value and the actual labeled value, constructs a regression optimization objective function with the goal of minimizing the loss, and introduces a regularization term to constrain the model complexity, outputting a constrained optimization objective function.

[0143] S4034, the optimization solution layer can solve the constrained optimization objective function, determine the relevant parameters of the optimal regression hyperplane in the high-dimensional feature space, complete the support vector regression model's learning and fitting of feature patterns, and output the optimal regression hyperplane model.

[0144] S4035. The regression output layer can substitute the high-dimensional feature representation vector into the optimal regression hyperplane model to perform regression prediction calculation, output quantitative prediction results, and perform validity verification and range adjustment on the prediction results to obtain the psychological endurance index.

[0145] S404. Input the physiological fatigue index and psychological endurance index into the physiological-psychological coupling model based on the nonlinear weighted coupling function structure to obtain the comprehensive fatigue tolerance.

[0146] Optionally, the overall fatigue tolerance can be used to characterize the patient's overall fatigue tolerance to rehabilitation training under the current physiological and psychological dimensions.

[0147] Optionally, the physiological-psychological coupling model based on the nonlinear weighted coupling function structure may include, but is not limited to, a feature input layer, a coupling weight adaptation layer, a nonlinear function operation layer, and a comprehensive solution layer.

[0148] For example, the processing flow of the physiological fatigue index and the psychological endurance index by the physiological-psychological coupling model based on the nonlinear weighted coupling function structure may include:

[0149] S4041. The feature input layer receives the physiological fatigue index and the psychological endurance index. The feature input layer can perform standardized secondary verification and outlier removal on the physiological fatigue index and the psychological endurance index to eliminate the dimensional bias remaining in the previous calculation and obtain the standardized physiological fatigue index and the psychological endurance index.

[0150] S4042. The coupling weight adaptation layer can retrieve the preset coupling weight coefficients based on the physiological-psychological interaction law of rehabilitation medicine, and perform weight allocation and weighted calculation on the standardized physiological fatigue index and psychological endurance index respectively to obtain the weighted physiological fatigue component and psychological endurance component.

[0151] S4043. The nonlinear function operation layer can perform nonlinear coupling operations on the weighted physiological fatigue component and psychological endurance component through the built-in nonlinear weighted coupling function, and quantify the nonlinear correlation and interaction between the physiological fatigue component and the psychological endurance component, and output the coupling fusion value.

[0152] S4044. The comprehensive solution layer can perform numerical normalization on the coupled fusion value, map the numerically normalized coupled fusion value to a preset fatigue tolerance quantification range, verify the validity of the mapping result, remove outliers that exceed the reasonable range, and output the quantified comprehensive fatigue tolerance.

[0153] S405. Based on the numerical range of comprehensive fatigue tolerance, dynamic safety constraint parameters are constructed by querying the training intensity adjustment coefficient and joint movement safety boundary threshold corresponding to the comprehensive fatigue tolerance from the preset safety constraint mapping table.

[0154] Optionally, the training intensity adjustment coefficient can be used to characterize the adjustment range of the auxiliary torque or drag torque of the exoskeleton robot.

[0155] Optionally, the joint motion safety boundary threshold can be used to characterize the maximum permissible range of motion of the exoskeleton robot's joint angles.

[0156] Optionally, the safety constraint mapping table can be constructed based on the clinical training standards of rehabilitation medicine and the physiological and psychological fatigue tolerance data of multi-patient rehabilitation training. It can be divided into several continuous and non-overlapping numerical intervals according to the numerical range of comprehensive fatigue tolerance. Each interval uniquely corresponds to a set of training intensity adjustment coefficients and joint movement safety boundary thresholds. The parameter values ​​are adapted to the rehabilitation training safety requirements of different fatigue tolerance levels.

[0157] For example, the intelligent auxiliary terminal can retrieve the comprehensive fatigue tolerance quantified value obtained from the aforementioned steps and retrieve the safety constraint mapping table. Secondly, the intelligent auxiliary terminal can perform a numerical range matching judgment on the comprehensive fatigue tolerance quantified value to determine the specific range to which the comprehensive fatigue tolerance belongs.

[0158] Furthermore, the intelligent auxiliary terminal can retrieve the training intensity adjustment coefficient and joint movement safety boundary threshold corresponding to the comprehensive fatigue tolerance from the safety constraint mapping table according to the unique mapping rules of intervals and parameters.

[0159] Furthermore, the intelligent auxiliary terminal can verify the validity of the retrieved training intensity adjustment coefficient and joint motion safety boundary threshold, eliminate abnormal matching values ​​caused by misjudgment of interval boundaries, and integrate and encapsulate the verified training intensity adjustment coefficient and joint motion safety boundary threshold to construct dynamic safety constraint parameters.

[0160] In this embodiment, the intelligent auxiliary terminal obtains the physiological fatigue index by weighted fusion of multiple physiological signals, calculates the psychological endurance index by combining facial expression features and autonomic nerve features through a support vector regression model, and then obtains the comprehensive fatigue tolerance by nonlinear coupling of facial expression features and autonomic nerve features through a physiological-psychological coupling model. Based on the safety constraint mapping table, dynamic safety constraint parameters are accurately constructed, which can provide core parameter support for the safety intensity control of rehabilitation training that is accurate and adapted to the real-time status of patients.

[0161] In one embodiment, a latent noncompliance feature vector is constructed based on the temporal deviation index, the probability of compensatory action categories, and the comprehensive compensation score, which may include the following steps:

[0162] S501. Extract the temporal feature components of the patient in the current training cycle from the temporal deviation index, extract the compensation category feature components from the compensation action category probability, and use the value of the comprehensive compensation score in the current training cycle as the score feature components.

[0163] The compensatory category feature component may include, but is not limited to, the compensatory action category and confidence level corresponding to the maximum probability at each time step.

[0164] For example, the intelligent assistive terminal can extract the core temporal statistical features and deviation trend features of the temporal deviation index within the current training cycle, and integrate the generated temporal feature components. Secondly, for the probability of compensatory action categories, the intelligent assistive terminal can filter out the compensatory action category with the highest probability value at each time step, and extract the confidence value of the compensatory action category matching. The compensatory action categories and the corresponding confidence values ​​of the compensatory action categories at each time step of the entire training cycle are then systematically integrated to generate compensatory category feature components.

[0165] Furthermore, the intelligent auxiliary terminal can use the original quantitative value of the comprehensive compensation score after validity verification within the current training cycle as the score feature component.

[0166] S502. Normalize the temporal feature components to obtain normalized temporal features. By converting the compensatory action categories with confidence levels greater than a preset confidence threshold in the compensatory category feature components into binary encoding vectors, the compensatory behavior encoding features are obtained.

[0167] For example, the intelligent assistive terminal can retrieve a preset confidence threshold based on the accuracy requirements for recognizing compensatory behaviors in rehabilitation training, and then judge and filter the compensatory action categories and corresponding confidence levels for each time step in the compensatory category feature components, retaining only the compensatory action categories with confidence values ​​greater than the confidence threshold.

[0168] Furthermore, the intelligent auxiliary terminal can perform vector conversion on the filtered compensatory action categories using a binary encoding method, and assign exclusive encoding dimensions to different types of compensatory actions according to the category encoding rules to generate binary encoded vectors.

[0169] S503. The normalized time series features, compensatory behavior coding features and scoring feature components are concatenated along the feature dimension to generate an initial multidimensional feature vector. Principal component analysis is then performed on the initial multidimensional feature vector to reduce its dimensionality, resulting in a set of principal components.

[0170] For example, the intelligent auxiliary terminal can sequentially concatenate the normalized temporal features, compensatory behavior encoding features, and scoring feature components to integrate the core representation information of training temporal deviation, high-confidence compensatory behavior, and overall compensatory severity, and generate an initial multidimensional feature vector.

[0171] Furthermore, the intelligent auxiliary terminal can perform decentralization preprocessing on the initial multidimensional feature vector to eliminate data mean bias, solve for the feature covariance matrix of the initial multidimensional feature vector, and calculate the eigenvalues ​​and eigenvectors corresponding to the feature covariance matrix.

[0172] Furthermore, the intelligent auxiliary terminal can sort the obtained feature vectors in descending order of feature values, then project the sorted feature vectors and the initial multidimensional feature vectors to extract principal components that can represent the main information of the original features. By integrating all the extracted principal components, a set of principal components is obtained.

[0173] S504. Extract the principal components in the principal component set whose cumulative contribution rate reaches the preset contribution rate threshold, and construct the feature subspace.

[0174] For example, the intelligent auxiliary terminal can calculate the contribution rate of each principal component in the principal component set in descending order of eigenvalues.

[0175] Furthermore, the intelligent auxiliary terminal can perform a step-by-step summation of the contribution rates of single principal components to obtain the cumulative contribution rate values ​​for each stage. Subsequently, the intelligent auxiliary terminal can compare the cumulative contribution rate of each stage with a preset contribution rate threshold, selecting all principal components whose cumulative contribution rate reaches or is not lower than the preset threshold, and removing principal components with low contribution rates and redundant representation information. Finally, the intelligent auxiliary terminal performs an ordered integration of the selected core principal components according to feature dimensions to construct a feature subspace.

[0176] S505. Project the initial multidimensional feature vectors onto the feature subspace to obtain the latent non-compliance feature vectors.

[0177] Among them, the implicit noncompliance feature vector can be used to characterize the comprehensive features of the degree of deviation of the patient's training quality and the severity of compensatory behavior in the current training cycle.

[0178] For example, the intelligent assistive terminal can linearly map the initial multidimensional feature vector from the original feature space to the core feature subspace, retain the core feature information that is highly correlated with the patient's implicit non-compliance behavior, and eliminate redundant features that have no representational value.

[0179] Furthermore, the intelligent auxiliary terminal can normalize the feature values ​​after projection calculation and perform validity verification, eliminating abnormal values ​​generated during the calculation process. Finally, the intelligent auxiliary terminal can reconstruct the verified projection feature results into vectors to generate implicit non-compliance feature vectors.

[0180] In this embodiment, the intelligent auxiliary terminal extracts and processes the feature components corresponding to the temporal deviation index, compensatory action category probability, and comprehensive compensation score from the temporal deviation index, compensatory action category probability, and comprehensive compensation score. After splicing to generate an initial multidimensional feature vector, it constructs a feature subspace through principal component analysis to reduce the dimensionality. Finally, the initial multidimensional feature vector is projected onto the feature subspace to obtain the implicit nonconformity feature vector. This enables a quantitative representation of the degree of deviation in training quality and the severity of compensatory behavior in the current training cycle of the patient, providing core and effective feature support for the subsequent classification and identification of implicit nonconformity behavior types.

[0181] In one embodiment, the construction of the intervention strategy library may include the following steps:

[0182] S601. Obtain historical rehabilitation training data from multiple patients and use the historical rehabilitation training data as a training sample set.

[0183] The training sample set may include, but is not limited to, effective interventions labeled by rehabilitation physicians.

[0184] For example, the intelligent assistive terminal can acquire historical rehabilitation training data for different rehabilitation stages, different types of limb injuries, and different training frequencies. Secondly, the intelligent assistive terminal can screen the acquired historical rehabilitation training data for validity, removing invalid data segments with excessively high data loss rates, distorted sensor signals, or unnatural interruptions in the training process, thus obtaining the filtered, valid historical rehabilitation training data.

[0185] Furthermore, the intelligent auxiliary terminal can perform structured processing on the filtered original and valid historical rehabilitation training data, and classify and archive it according to dimensions such as patient basic information, rehabilitation training type, multimodal time-series data of the training process, compensation behavior identification results, and implicit noncompliance behavior type.

[0186] Furthermore, the intelligent auxiliary terminal can integrate effective intervention measures that rehabilitation physicians have labeled and standardized on-site for different rehabilitation problems and different types of implicit non-compliance behaviors, and associate and match various historical rehabilitation training data with the corresponding labeled effective intervention measures to construct a training sample set.

[0187] S602. Based on the training sample set, extract patient static features from multimodal time series data and use patient static features as patient profile labels. Extract behavior pattern labels from implicit noncompliance behavior types and extract intervention action codes from effective intervention measures.

[0188] For example, the intelligent assistive terminal can retrieve the constructed training sample set and filter out the core information representing the patient's inherent attributes from the multimodal time-series data in the training sample set. Secondly, the intelligent assistive terminal can standardize the encoding of various static features and normalize the feature dimensions, using the processed set of static features as patient profile labels.

[0189] Furthermore, the intelligent auxiliary terminal can extract key information from implicit nonconformity behavior types labeled in the training sample set, refine the labels according to preset behavior classification rules, generate behavior pattern labels, and establish a unique correspondence between implicit nonconformity behavior types and behavior pattern labels; among which, the key information in implicit nonconformity behavior types may include, but is not limited to, core behavioral manifestations and occurrence scenarios.

[0190] Furthermore, the intelligent assistive terminal can break down the core intervention measures, execution logic, and exoskeleton parameter adjustment rules of the effective intervention measures based on the effective intervention measures labeled by rehabilitation physicians in the training sample set, and digitize and structure them according to the instruction execution specifications of the exoskeleton robot to generate standardized intervention action codes.

[0191] S603. Link and store patient profile tags, behavior pattern tags, and intervention action codes to construct a case library.

[0192] For example, the intelligent assistive terminal can establish association mapping rules for patient profile tags, behavior pattern tags, and intervention action codes based on the clinical logic of rehabilitation training. It uses behavior pattern tags as the core retrieval index, associates the core attributes of the patient profile tags corresponding to the behavior pattern, and binds the intervention action codes labeled by rehabilitation physicians that are suitable for this type of patient profile and behavior pattern, thereby realizing the corresponding association between patient profile tags, behavior pattern tags, and intervention action codes.

[0193] Furthermore, the intelligent assistive terminal can build a structured case library storage architecture, setting up multi-level classification directories according to the patient's rehabilitation stage, limb injury type, and implicit non-compliance behavior pattern, and assigning a unique case identifier to each group of associated tag-coded data.

[0194] Furthermore, the intelligent auxiliary terminal can supplement the data by recording auxiliary information such as the rehabilitation training scenario and the background of intervention implementation corresponding to the combination of patient profile tags, behavioral pattern tags, and intervention action codes for each group of data that has completed the association mapping. Finally, the intelligent auxiliary terminal can store the standardized case data that has completed association and identification in an orderly manner according to a multi-level classification directory into the architecture to build a case library.

[0195] S604. Effective intervention measures are functionally classified to obtain the intervention functions corresponding to each type of implicit noncompliance behavior. Based on the behavior change technology classification system, behavior change technologies are matched for each intervention function to construct a rule base.

[0196] For example, the intelligent assistive terminal can retrieve all effective intervention measures annotated by rehabilitation physicians in the training sample set, classify and collect the effective intervention measures according to the type of implicit non-compliance behavior, so that each type of implicit non-compliance behavior corresponds to its own set of effective intervention measures.

[0197] Secondly, the intelligent assistive terminal can decompose and summarize the core functions of effective intervention measures within each set of effective intervention measures. Based on the implementation purpose, mode of action, and actual rehabilitation effect of the effective intervention measures, the effective intervention measures can be divided into different intervention functions such as movement standardization guidance, dynamic adjustment of training parameters, real-time correction of compensatory behavior, and adaptation and adjustment of training rhythm. This completes the association mapping between each type of implicit noncompliance behavior and the corresponding intervention function, and standardizes the naming and definition of each intervention function.

[0198] Furthermore, the intelligent assistive terminal can match the corresponding behavior modification technology to each intervention function based on the execution logic and behavior intervention principle of each intervention function, and establish a unique correspondence between intervention functions and behavior modification technologies.

[0199] Furthermore, intelligent auxiliary terminals can structure and encode all related data according to the hierarchical logic of implicit noncompliance behavior type - intervention function - behavior change technology, assign unique identification information to each level of data, and build a rule base.

[0200] S605, integrating case libraries and rule libraries, to construct an intervention strategy library.

[0201] The intervention strategy library can store intervention strategies in key-value pairs.

[0202] For example, the intelligent auxiliary terminal can access the pre-built case library and rule library, and match the behavioral pattern tags in the case library with the implicit non-compliance behavior types in the rule library.

[0203] Furthermore, the intelligent auxiliary terminal establishes a unique association mapping between the intervention action codes in the case library and the intervention functions in the rule library, unifying the identification coding rules, field naming formats, and data storage dimensions of the two types of libraries.

[0204] Furthermore, the intelligent assistive terminal can take the clinical intervention logic of rehabilitation training as its core, and use the hierarchical rules of implicit noncompliance behavior type-intervention function-behavior change technology in the rule base as the basic framework of the intervention strategy, and embed specific intervention case data in the case base that are adapted to different patient profile labels into the corresponding framework level of the rule base.

[0205] Furthermore, the intelligent assistance terminal can construct an intervention strategy library according to the key-value pair storage specification, using the combination of implicit non-compliance behavior type and patient profile label as the unique search key, and the matching complete intervention strategy as the corresponding value. Finally, the intelligent assistance terminal can perform validity verification and redundant information removal on the fused strategy data, and establish a fast search index for each key-value pair to generate the intervention strategy library.

[0206] In this embodiment, the intelligent auxiliary terminal constructs a structured training sample set from historical rehabilitation training data of multiple patients, extracts patient profile tags, behavioral pattern tags, and intervention action codes, and completes the associated storage based on the clinical logic of rehabilitation training to build a case library. It also categorizes effective intervention measures by function and matches them with behavior modification techniques to build a hierarchical rule library. Finally, it integrates personalized intervention case data from the case library with standardized intervention rules from the rule library, unifying the data format and constructing an intervention strategy library in key-value pair form. This provides high-quality, rapidly searchable strategy support for reinforcement learning frameworks to quickly match and select suitable behavioral intervention strategies.

[0207] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0208] Based on the same inventive concept, this application also provides an intelligent assistive system for rehabilitation nursing described above. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the intelligent assistive system for rehabilitation nursing provided below can be found in the limitations of the intelligent assistive method for rehabilitation nursing described above, and will not be repeated here.

[0209] In one exemplary embodiment, such as Figure 2 As shown, an intelligent assistive system 700 for rehabilitation care is provided, comprising:

[0210] The data preprocessing module 701 is used to acquire raw sensor data of key parts of the patient's body, and to perform time-series synchronization calibration and dynamic evaluation of signal reliability on the raw sensor data to obtain multimodal time-series data; wherein, the multimodal time-series data includes electromyographic signal data, inertial measurement unit data, physiological signal data and facial expression data;

[0211] The compensatory action recognition module 702 is used to construct a compensatory sensitive feature set based on multimodal time series data, and input the compensatory sensitive feature set into a two-stream neural network based on a spatiotemporal attention mechanism, and output the compensatory action category probability and comprehensive compensatory score at each time step.

[0212] The safety constraint generation module 703 is used to generate dynamic safety constraint parameters based on physiological signal data and facial expression data; wherein, the dynamic safety constraint parameters are used to characterize the patient's current physiological fatigue limit and psychological endurance boundary;

[0213] The impedance control command generation module 704 is used to input electromyography signal data and inertial measurement unit data into a long short-term memory network to obtain the predicted value of motion intention and the prediction confidence interval corresponding to the predicted value of motion intention. Based on the predicted value of motion intention, the prediction confidence interval corresponding to the predicted value of motion intention, and dynamic safety constraint parameters, an impedance control command is generated. The impedance control command is used to instruct the exoskeleton robot to provide auxiliary torque according to the predicted motion intention.

[0214] The compliance assessment module 705 is used to acquire the actual motion state data of the exoskeleton robot assisting the patient in rehabilitation nursing according to the impedance control command, extract the action timing features from the actual motion state data, obtain the action timing features, and compare the action timing features with the preset rehabilitation nursing prescription standard timing template to obtain the timing deviation index.

[0215] The nonconformity behavior identification module 706 is used to construct a latent nonconformity feature vector based on the temporal deviation index, the probability of the compensatory action category and the comprehensive compensatory score, and input the latent nonconformity feature vector into a pre-trained classifier to obtain the latent nonconformity behavior type.

[0216] The instruction generation module 707 is used to select behavioral intervention strategies that match the implicit noncompliance behavior type from a pre-built intervention strategy library through a reinforcement learning framework, and generate intervention instructions; wherein, the intervention instructions are used to instruct the exoskeleton robot to adjust the assistance mode.

[0217] In one embodiment, the compensatory action recognition module includes:

[0218] The contribution feature extraction unit is used to input the data from the inertial measurement unit into the multi-rigid-body kinematics model to obtain the joint contribution feature vector; wherein, the joint contribution feature vector is used to characterize the deviation of the degree of participation of each joint from the standard movement during the execution of the action;

[0219] The co-operational deviation feature generation unit is used to extract the muscle co-operational primitive matrix from the electromyographic signal data, and generate the electromyographic co-operational deviation feature by calculating the reconstruction error of the current electromyographic signal under the muscle co-operational primitive matrix and the deviation between the co-activation coefficient and the expected activation coefficient; wherein, the electromyographic co-operational deviation feature is used to characterize the degree of deviation between the current muscle activation mode and the normative muscle co-operation mode.

[0220] The smoothness feature construction unit is used to generate motion smoothness features based on inertial measurement unit data by calculating the curvature of the motion trajectory of the end effector in three-dimensional space and the acceleration index of the joint angular velocity; among which, motion smoothness features are used to characterize the smoothness and continuity of the motion.

[0221] The sensitive feature set fusion unit is used to concatenate and fuse the joint contribution feature vector, electromyographic coordination deviation feature, and motion fluency feature in the feature dimension to generate a compensatory sensitive feature set.

[0222] In one embodiment, the collaborative deviation feature generation unit includes:

[0223] The coordinating primitive matrix extraction subunit is used to acquire multi-channel surface electromyography (EMG) signals when the patient performs standardized movements. The multi-channel EMG signals are used as training data, and the training data is decomposed by a non-negative matrix factorization algorithm to obtain the muscle coordinating primitive matrix.

[0224] The coefficient solving subunit is used to project the electromyographic signal data collected at the current moment onto the muscle coordination primitive matrix, and reconstruct the current electromyographic signal by solving the non-negative least squares problem to obtain the coordination activation coefficient vector and reconstruction error vector at the current moment.

[0225] The coefficient deviation calculation subunit is used to retrieve the corresponding expected co-activation coefficient vector from the standardized action database according to the current action execution phase of the patient, and obtain the activation coefficient deviation value by calculating the Euclidean distance between the current co-activation coefficient vector and the expected co-activation coefficient vector.

[0226] The deviation feature generation sub-unit is used to obtain the electromyographic coordination deviation feature by weighted sum of the squared values ​​of the L2 norm and the deviation values ​​of the activation coefficients based on the reconstruction error vector.

[0227] In one embodiment, the security constraint generation module includes:

[0228] The fatigue index calculation unit is used to perform weighted fusion of heart rate variability time-domain index, heart rate variability frequency-domain index, median frequency of electromyography signal and skin conductance response level to obtain the physiological fatigue index.

[0229] The body and mind feature extraction construction unit is used to extract the activation intensity of facial action units from facial expression data and use the activation intensity of facial action units as facial expression features. It also extracts the high-frequency power of heart rate variability and the level of skin conductance response from physiological signal data and uses the high-frequency power of heart rate variability and the level of skin conductance response as autonomic nerve features.

[0230] The psychological endurance index generation unit is used to input facial expression features and autonomic nerve features into the support vector regression model to obtain the psychological endurance index; the psychological endurance index is used to characterize the patient's current psychological endurance capacity.

[0231] The tolerance coupling unit is used to input the physiological fatigue index and the psychological endurance index into the physiological-psychological coupling model based on the nonlinear weighted coupling function structure to obtain the comprehensive fatigue tolerance.

[0232] The constraint parameter construction unit is used to construct dynamic safety constraint parameters based on the numerical range of comprehensive fatigue tolerance by querying the training intensity adjustment coefficient and joint motion safety boundary threshold corresponding to the comprehensive fatigue tolerance from a preset safety constraint mapping table. Among them, the training intensity adjustment coefficient is used to characterize the adjustment range of the auxiliary torque or resistance torque of the exoskeleton robot; the joint motion safety boundary threshold is used to characterize the maximum allowable range of motion of the joint angle of the exoskeleton robot.

[0233] In one embodiment, the non-compliance behavior recognition module includes:

[0234] The feature component extraction unit is used to extract the temporal feature components of the patient in the current training cycle from the temporal deviation index, extract the compensation category feature components from the probability of compensation action categories, and use the value of the comprehensive compensation score in the current training cycle as the scoring feature components; wherein, the compensation category feature components include the compensation action category and confidence level corresponding to the maximum probability of each time step.

[0235] The feature encoding conversion unit is used to normalize the temporal feature components to obtain normalized temporal features. By converting the compensation action categories with confidence scores greater than a preset confidence threshold in the compensation category feature components into binary encoding vectors, the compensation behavior encoding features are obtained.

[0236] The principal component analysis unit is used to concatenate the normalized time series features, compensatory behavior coding features and scoring feature components along the feature dimension to generate an initial multidimensional feature vector, and then perform principal component analysis to reduce the dimensionality of the initial multidimensional feature vector to obtain the principal component set.

[0237] The feature subspace construction unit is used to extract principal components in the principal component set whose cumulative contribution rate reaches a preset contribution rate threshold, and construct the feature subspace.

[0238] The implicit noncompliance feature vector generation unit is used to project the initial multidimensional feature vector onto the feature subspace to obtain the implicit noncompliance feature vector; wherein, the implicit noncompliance feature vector is used to characterize the comprehensive features of the degree of deviation of the patient's training quality and the severity of compensatory behavior in the current training cycle.

[0239] In one embodiment, the instruction generation module includes:

[0240] The sample set construction unit is used to acquire historical rehabilitation training data from multiple patients and use this historical rehabilitation training data as the training sample set; the training sample set includes effective intervention measures labeled by rehabilitation physicians.

[0241] The label encoding extraction unit is used to extract patient static features from multimodal time-series data based on the training sample set, and use the patient static features as patient profile labels, extract behavioral pattern labels from implicit noncompliance behavior types, and extract intervention action codes from effective intervention measures.

[0242] The case library building unit is used to associate and store patient profile tags, behavior pattern tags, and intervention action codes to build the case library;

[0243] The rule base construction unit is used to classify effective intervention measures by function, obtain the intervention functions corresponding to each type of implicit noncompliance behavior, and match behavior modification technologies to each intervention function according to the behavior modification technology classification system, thus constructing the rule base.

[0244] The strategy library generation unit is used to integrate the case library and the rule library to build an intervention strategy library; the intervention strategy library stores intervention strategies in key-value pair format.

[0245] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of an intelligent assistive method for rehabilitation care as described above.

[0246] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0247] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and 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 modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0248] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. An intelligent assistive method for rehabilitation nursing, characterized in that, The method includes: S101. Acquire raw sensor data of key parts of the patient's body, and perform time-series synchronization calibration and dynamic evaluation of signal reliability on the raw sensor data to obtain multimodal time-series data; wherein, the multimodal time-series data includes electromyographic signal data, inertial measurement unit data, physiological signal data and facial expression data; S102. Based on the multimodal time series data, a compensation-sensitive feature set is constructed, and the compensation-sensitive feature set is input into a two-stream neural network based on a spatiotemporal attention mechanism to output the probability of the compensation action category and the comprehensive compensation score at each time step. S103. Based on the physiological signal data and the facial expression data, generate dynamic safety constraint parameters; wherein, the dynamic safety constraint parameters are used to characterize the patient's current physiological fatigue limit and psychological endurance boundary; S104. Input the electromyography signal data and the inertial measurement unit data into the long short-term memory network to obtain the motion intention prediction value and the prediction confidence interval corresponding to the motion intention prediction value. Based on the motion intention prediction value, the prediction confidence interval corresponding to the motion intention prediction value, and the dynamic safety constraint parameters, generate an impedance control command. The impedance control command is used to instruct the exoskeleton robot to provide auxiliary torque according to the predicted motion intention. S105. Obtain actual motion state data of the exoskeleton robot assisting the patient in rehabilitation nursing according to the impedance control command, extract the action timing features from the actual motion state data to obtain action timing features, and compare the action timing features with the preset rehabilitation nursing prescription standard timing template to obtain the timing deviation index. S106. Based on the temporal deviation index, the probability of the compensatory action category, and the comprehensive compensatory score, a latent nonconformity feature vector is constructed, and the latent nonconformity feature vector is input into a pre-trained classifier to obtain the latent nonconformity behavior type. S107. Using a reinforcement learning framework, select behavioral intervention strategies that match the implicit nonconformity type from a pre-built intervention strategy library and generate intervention instructions; wherein, the intervention instructions are used to instruct the exoskeleton robot to adjust its assistance mode.

2. The method according to claim 1, characterized in that, The compensation-sensitive feature set constructed based on the multimodal time-series data includes: S201. Input the data from the inertial measurement unit into the multi-rigid-body kinematics model to obtain the joint contribution feature vector; wherein, the joint contribution feature vector is used to characterize the deviation of the degree of participation of each joint from the standard action during the execution of the action; S202. Extract the muscle coordination primitive matrix from the electromyographic signal data, and generate an electromyographic coordination deviation feature by calculating the reconstruction error of the current electromyographic signal under the muscle coordination primitive matrix and the deviation between the coordination activation coefficient and the expected activation coefficient; wherein, the electromyographic coordination deviation feature is used to characterize the degree of deviation between the current muscle activation mode and the standard muscle coordination mode. S203. Based on the data from the inertial measurement unit, motion smoothness features are generated by calculating the curvature of the motion trajectory and the acceleration index of the joint angular velocity of the end effector in three-dimensional space; wherein, the motion smoothness features are used to characterize the smoothness and continuity of the motion. S204. The joint contribution feature vector, the electromyographic coordination deviation feature, and the motion fluency feature are concatenated and fused along the feature dimension to generate the compensation-sensitive feature set.

3. The method according to claim 2, characterized in that, The step of extracting a muscle coordination primitive matrix from the electromyographic signal data, and generating an electromyographic coordination deviation feature by calculating the reconstruction error of the current electromyographic signal under the muscle coordination primitive matrix and the deviation between the coordination activation coefficient and the expected activation coefficient, includes: S301. Acquire multi-channel surface electromyography (EMG) signals when the patient performs standardized movements, use the multi-channel EMG signals as training data, and decompose the training data using a non-negative matrix factorization algorithm to obtain a muscle coordination primitive matrix. S302. Project the electromyographic signal data collected at the current moment onto the muscle coordination primitive matrix, and reconstruct the current electromyographic signal by solving a non-negative least squares problem to obtain the coordination activation coefficient vector and reconstruction error vector at the current moment. S303. Based on the patient's current action execution phase, retrieve the corresponding expected co-activation coefficient vector from the standardized action database, and obtain the activation coefficient deviation value by calculating the Euclidean distance between the current co-activation coefficient vector and the expected co-activation coefficient vector. S304. Based on the squared L2 of the reconstructed error vector and the weighted sum of the activation coefficient deviation values, the electromyographic coordination deviation feature is obtained. The expression for the electromyographic coordination deviation characteristic is as follows: In the formula, express The electromyographic coordination deviation characteristics at time points. express The reconstruction error vector at time t, express The cooperative activation coefficient vector at time t, express The expected co-activation coefficient vector at time t, This represents the weighting coefficients for balancing the reconstruction error term and the activation coefficient deviation term.

4. The method according to claim 1, characterized in that, The physiological signal data includes time-domain indices of heart rate variability, frequency-domain indices of heart rate variability, median frequency of electromyography signals, and skin conductance response levels. The generation of dynamic safety constraint parameters based on the physiological signal data and the facial expression data includes: S401. The time-domain index of heart rate variability, the frequency-domain index of heart rate variability, the median frequency of electromyography signal, and the skin conductance response level are weighted and fused to obtain the physiological fatigue index. S402. Extract the facial motion unit activation intensity from the facial expression data, and use the facial motion unit activation intensity as a facial expression feature. Extract the heart rate variability high-frequency power and skin conductance response level from the physiological signal data, and use the heart rate variability high-frequency power and skin conductance response level as autonomic nerve features. S403. Input the facial expression features and the autonomic nerve features into a support vector regression model to obtain a psychological endurance index; wherein, the psychological endurance index is used to characterize the patient's current psychological endurance. S404. Input the physiological fatigue index and the psychological endurance index into a physiological-psychological coupling model based on a nonlinear weighted coupling function structure to obtain the comprehensive fatigue tolerance. S405. Based on the numerical range of the comprehensive fatigue tolerance, the dynamic safety constraint parameters are constructed by querying the training intensity adjustment coefficient and joint motion safety boundary threshold corresponding to the comprehensive fatigue tolerance from a preset safety constraint mapping table; wherein, the training intensity adjustment coefficient is used to characterize the adjustment range of the auxiliary torque or resistance torque of the exoskeleton robot; and the joint motion safety boundary threshold is used to characterize the maximum allowable range of motion of the joint angle of the exoskeleton robot.

5. The method according to any one of claims 1 to 4, characterized in that, The implicit noncompliance feature vector is constructed based on the temporal deviation index, the probability of the compensatory action category, and the comprehensive compensatory score, including: S501. Extract the temporal feature components of the patient in the current training cycle from the temporal deviation index, extract the compensation category feature components from the compensation action category probability, and use the value of the comprehensive compensation score in the current training cycle as the scoring feature components; wherein, the compensation category feature components include the compensation action category and confidence level corresponding to the maximum probability of each time step. S502. Normalize the time series feature components to obtain normalized time series features. Convert the compensation action categories with confidence scores greater than a preset confidence threshold in the compensation category feature components into binary encoding vectors to obtain compensation behavior encoding features. S503. The normalized time series features, the compensatory behavior encoding features, and the scoring feature components are concatenated along the feature dimension to generate an initial multidimensional feature vector, and the initial multidimensional feature vector is subjected to principal component analysis to reduce the dimensionality and obtain a set of principal components. S504. Extract the principal components in the principal component set whose cumulative contribution rate reaches a preset contribution rate threshold, and construct the feature subspace. S505. Project the initial multidimensional feature vector onto the feature subspace to obtain the latent noncompliance feature vector; wherein, the latent noncompliance feature vector is used to characterize the comprehensive features of the degree of deviation of the patient's training quality and the severity of compensatory behavior in the current training cycle.

6. The method according to claim 1, characterized in that, The construction method of the intervention strategy library includes: S601. Obtain historical rehabilitation training data from multiple patients and use the historical rehabilitation training data as a training sample set; wherein, the training sample set includes effective intervention measures labeled by rehabilitation physicians. S602. Based on the training sample set, extract patient static features from the multimodal time series data, and use the patient static features as patient profile labels. Extract behavior pattern labels from the implicit noncompliance behavior types, and extract intervention action codes from the effective intervention measures. S603. The patient profile tags, the behavior pattern tags, and the intervention action codes are associated and stored to construct a case library; S604. The effective intervention measures are functionally classified to obtain the intervention functions corresponding to each type of implicit noncompliance behavior, and behavior change technologies are matched to each intervention function according to the behavior change technology classification system to construct a rule base. S605. The case library and the rule library are integrated to construct the intervention strategy library; wherein the intervention strategy library stores intervention strategies in key-value pair format.

7. An intelligent assistive system for rehabilitation nursing, characterized in that, The system includes: The data preprocessing module is used to acquire raw sensor data of key parts of the patient's body, and to perform time-series synchronization calibration and dynamic evaluation of signal reliability on the raw sensor data to obtain multimodal time-series data; wherein, the multimodal time-series data includes electromyographic signal data, inertial measurement unit data, physiological signal data and facial expression data; The compensatory action recognition module is used to construct a compensatory sensitive feature set based on the multimodal time series data, and input the compensatory sensitive feature set into a two-stream neural network based on a spatiotemporal attention mechanism, and output the compensatory action category probability and comprehensive compensatory score at each time step. A safety constraint generation module is used to generate dynamic safety constraint parameters based on the physiological signal data and the facial expression data; wherein, the dynamic safety constraint parameters are used to characterize the patient's current physiological fatigue limit and psychological endurance boundary; An impedance control command generation module is used to input the electromyography signal data and the inertial measurement unit data into a long short-term memory network to obtain a predicted value of motion intention and a prediction confidence interval corresponding to the predicted value of motion intention, and to generate an impedance control command based on the predicted value of motion intention, the prediction confidence interval corresponding to the predicted value of motion intention, and the dynamic safety constraint parameters; wherein, the impedance control command is used to instruct the exoskeleton robot to provide auxiliary torque according to the predicted motion intention; The compliance assessment module is used to acquire the actual motion state data of the exoskeleton robot assisting the patient in rehabilitation nursing according to the impedance control command, extract the action timing features from the actual motion state data to obtain the action timing features, and compare the action timing features with the preset rehabilitation nursing prescription standard timing template to obtain the timing deviation index. The nonconformity behavior identification module is used to construct a latent nonconformity feature vector based on the temporal deviation index, the probability of the compensatory action category, and the comprehensive compensatory score, and input the latent nonconformity feature vector into a pre-trained classifier to obtain the latent nonconformity behavior type. The instruction generation module is used to select behavioral intervention strategies that match the implicit noncompliance behavior type from a pre-built intervention strategy library using a reinforcement learning framework, and generate intervention instructions; wherein, the intervention instructions are used to instruct the exoskeleton robot to adjust its assistance mode.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images