Exoskeleton robot control method and system based on brain-computer interface

Abstract

The invention belongs to the technical field of robot control, and discloses an exoskeleton robot control method and system based on a brain-computer interface, and the method comprises the steps: obtaining brain-computer interface data, carrying out the preprocessing of the brain-computer interface data, obtaining the preprocessed brain-computer interface data, extracting motion features, and obtaining a motion feature vector; performing motion intention recognition on the motion feature vector by using an integrated learning model, generating a motion control instruction, and performing motion control according to the motion control instruction; and based on the obtained motion state feedback data, motion state adjustment is carried out by using a dynamic adjustment algorithm, a motion control instruction is dynamically adjusted according to a motion state adjustment result, and motion control optimization is carried out through the adjusted motion control instruction. The problems that in the prior art, the signal processing precision is low, the intention recognition accuracy is insufficient, and the control strategy is rigid are solved, and efficient cooperation of the exoskeleton robot and the motion intention of the user is achieved.

Description

Technical Field

[0001] The present invention relates to the field of robot control technology, and in particular to an exoskeleton robot control method and system based on a brain-computer interface. Background Art

[0002] As an important device for assisting human movement, exoskeleton robots typically consist of a mechanical structure, drive system, sensors, control system, and power supply. They can coordinate movement with the human body, enhancing or assisting the body's motor abilities and physiological functions. This has led to their widespread application in medical rehabilitation and assistance for people with disabilities. However, traditional control methods often rely on mechanical sensors or manual operation, which can lead to high response latency and difficulty accurately capturing user intent.

[0003] At the same time, in the existing brain-computer interface-based control technology, EEG signals are easily interfered by environmental noise and physiological artifacts, resulting in insufficient feature extraction accuracy; the motion intention recognition model is single and has weak generalization ability, making it difficult to adapt to the signal characteristics of different users; and there is a lack of effective feedback and adjustment mechanism, and control instructions are easily disconnected from actual motion needs.

[0004] In summary, there is an urgent need for an exoskeleton robot control method and system that can improve signal processing robustness, intention recognition accuracy and dynamic adjustment capabilities.

[0005] Therefore, how to provide an exoskeleton robot control method and system based on brain-computer interface is a problem that needs to be solved urgently. Summary of the Invention

[0006] The embodiments of the present invention provide an exoskeleton robot control method and system based on a brain-computer interface to solve the above-mentioned technical problems in the prior art.

[0007] To provide a basic understanding of some aspects of the disclosed embodiments, the following is a brief summary. This summary is not intended to be an extensive review, identify key or critical elements, or delineate the scope of these embodiments. Its sole purpose is to present some concepts in a simplified form as a prelude to the detailed description that follows.

[0008] According to a first aspect of an embodiment of the present invention, a method for controlling an exoskeleton robot based on a brain-computer interface is provided.

[0009] In one embodiment, a method for controlling an exoskeleton robot based on a brain-computer interface includes: Acquiring brain-computer interface data, preprocessing the brain-computer interface data to obtain preprocessed brain-computer interface data, and extracting motion features using the preprocessed brain-computer interface data to obtain a motion feature vector; Using an integrated learning model to identify motion intentions from motion feature vectors, generating motion control instructions based on the motion intention identification results, and performing motion control according to the motion control instructions; Based on the acquired motion state feedback data, the motion state is adjusted using a dynamic adjustment algorithm, and the motion control instructions are dynamically adjusted according to the motion state adjustment results, and motion control optimization is performed through the adjusted motion control instructions.

[0010] According to a second aspect of an embodiment of the present invention, an exoskeleton robot control system based on a brain-computer interface is provided.

[0011] In one embodiment, an exoskeleton robot control system based on a brain-computer interface includes: a motion feature extraction module, a motion control module, and a motion control optimization module; A motion feature extraction module is used to obtain brain-computer interface data, preprocess the brain-computer interface data to obtain preprocessed brain-computer interface data, and extract motion features using the preprocessed brain-computer interface data to obtain a motion feature vector; A motion control module is used to identify motion intentions of motion feature vectors using an integrated learning model, generate motion control instructions based on the motion intention identification results, and perform motion control according to the motion control instructions; The motion control optimization module is used to adjust the motion state based on the acquired motion state feedback data using a dynamic adjustment algorithm, and dynamically adjust the motion control instructions according to the motion state adjustment results, and perform motion control optimization through the adjusted motion control instructions.

[0012] According to a third aspect of embodiments of the present invention, a computer device is provided.

[0013] In some embodiments, the computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above-mentioned brain-computer interface-based exoskeleton robot control method when executing the computer program.

[0014] According to a fourth aspect of embodiments of the present invention, a computer-readable storage medium is provided.

[0015] In one embodiment, a computer program is stored on a computer-readable storage medium, and when the computer program is executed by a processor, the steps of the above-mentioned brain-computer interface-based exoskeleton robot control method are implemented.

[0016] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: 1. The present invention avoids the interference of environmental noise and physiological artifacts on the extracted EEG signals by extracting motion features, and ensures the accuracy of feature extraction. Under the action of the integrated learning model for motion intention recognition, the problems of the singleness and weak generalization ability of the motion intention recognition model are avoided, thereby adapting to the signal characteristics of different users. Finally, under the action of the dynamic adjustment algorithm for motion state adjustment, effective feedback adjustment is performed to avoid the disconnection between control instructions and actual motion requirements.

[0017] 2. The present invention uses motion feature vectors to identify motion intentions, solving the problems of low signal processing accuracy, insufficient intention recognition accuracy, and rigid control strategies in the existing technology, achieving efficient coordination between the exoskeleton robot and the user's motion intentions, and avoiding the problems of high response delays and difficulty in accurately capturing user intentions in traditional control methods.

[0018] 3. The data preprocessing and integrated learning model of the present invention improves the signal processing accuracy and intention recognition accuracy, and at the same time combines the motion state feedback data for dynamic adjustment, solving the problems of response delay and poor adaptability in traditional exoskeleton robot control, and significantly improves the human-machine collaboration efficiency and control safety, and is suitable for various scenarios such as medical rehabilitation and assisted action.

[0019] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0021] Figure 1 is a flow chart of a method for controlling an exoskeleton robot based on a brain-computer interface according to an exemplary embodiment; Figure 2 is a structural diagram of a system according to an exemplary embodiment; Figure 3 is a structural diagram of a computer device according to an exemplary embodiment; Figure 4 is an overall framework diagram of an exoskeleton robot control method based on a brain-computer interface according to an exemplary embodiment; Figure 5 It is a structural diagram of an integrated learning model in a brain-computer interface-based exoskeleton robot control method according to an exemplary embodiment.

[0022] Reference numerals: 201. Motion feature extraction module; 202. Motion control module; 203. Motion control optimization module. DETAILED DESCRIPTION

[0023] The following description and accompanying drawings sufficiently illustrate the specific embodiments herein to enable those skilled in the art to practice them. Portions and features of some embodiments may be included in or substituted for portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims, including all available equivalents thereof. Herein, the terms "first," "second," and the like are used solely to distinguish one element from another and do not require or imply any actual relationship or order between these elements. In practice, the first element can also be referred to as the second element, and vice versa. Furthermore, the terms "comprise," "comprising," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a structure, device, or apparatus comprising a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such structure, device, or apparatus. Without further limitation, an element defined by the phrase "comprising a..." does not preclude the presence of other identical elements in the structure, device, or apparatus comprising the element. The various embodiments herein are described in a progressive manner, with each embodiment focusing on its differences from the other embodiments. Similar or identical parts between the various embodiments can be referenced to each other.

[0024] The terms "longitudinal", "transverse", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like used herein to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are intended only to facilitate the description of this document and simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on the present invention. In the description herein, unless otherwise specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, they can be mechanical or electrical connections, or they can be internal connections between two elements, they can be directly connected, or they can be indirectly connected through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to the specific circumstances.

[0025] As used herein, unless otherwise specified, the term "plurality" means two or more.

[0026] In this document, the character " / " indicates that the preceding and following objects are in an "or" relationship. For example, A / B means: A or B.

[0027] In this article, the term "and / or" describes the association relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0028] It should be understood that, although the various steps in the flowchart are shown in sequence as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps may be performed in other orders. Moreover, at least a portion of the steps in the figure may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily performed at the same time, but may be performed at different times. The execution order of these sub-steps or stages is not necessarily to be performed in sequence, but may be performed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0029] Each module in the device or system of the present application can be implemented in whole or in part by software, hardware, or a combination thereof. The above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software so that the processor can call and execute the operations corresponding to the above modules.

[0030] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0031] Figure 1 An embodiment of the exoskeleton robot control method based on brain-computer interface of the present invention is shown.

[0032] In this optional embodiment, the exoskeleton robot control method based on the brain-computer interface includes: Step S101, acquiring brain-computer interface data, preprocessing the brain-computer interface data to obtain preprocessed brain-computer interface data, and extracting motion features using the preprocessed brain-computer interface data to obtain a motion feature vector; Step S102, using an integrated learning model to perform motion intention recognition on the motion feature vector, generating a motion control instruction based on the motion intention recognition result, and performing motion control according to the motion control instruction; Step S103 , based on the acquired motion state feedback data, a dynamic adjustment algorithm is used to adjust the motion state, and the motion control instructions are dynamically adjusted according to the motion state adjustment result, and motion control optimization is performed using the adjusted motion control instructions.

[0033] It should be noted that, in order to obtain the brain-computer interface data, the brain-computer interface data is subjected to three-level dynamic preprocessing, that is, it is first subjected to adaptive filtering, wherein the step size factor qDynamically adjust according to the proportion of power frequency interference, with a value range of 0.01-0.1; remove power frequency interference and electromyographic noise, and then use wavelet transform, where the number of decomposition layers is dynamically adjusted according to the proportion of electrooculogram artifacts, and the value range is 5-7 layers, and then perform data decomposition and frequency band screening, and finally maximize the difference in motion intention based on the common spatial pattern (CSP) algorithm to extract motion features and obtain motion feature vectors; use a heterogeneous ensemble learning model (i.e., ensemble learning model) to perform motion intention recognition on the motion feature vectors, and the heterogeneous ensemble learning model includes an optimized convolutional neural network model, an optimized support vector machine model, and an optimized random forest model, and the optimized convolutional neural network model is used to extract motion features. The network model contains 3 layers of convolution and 2 layers of residual connection. The optimized support vector machine model adopts radial basis kernel function, and the kernel parameter γ1 is dynamically adjusted to 0.1-1. The optimized random forest model contains 100 decision trees and adopts feature random subspace sampling. Through dynamic weighted voting, the weight of weighted voting is allocated in real time according to the sub-model confidence, and the initial weight ratio is 0.35:0.3:0.35. The motion intention recognition result is output, and the motion control instruction is generated and the motion control is executed based on the recognition result; based on the acquired motion state feedback data, the double closed-loop dynamic adjustment algorithm is used to adjust the motion state, that is, the inner loop uses the reinforcement learning algorithm to update the state value function formula V '( s )= V ( s )+ α [ r + γV ( s ')- V ( s )] to adjust the joint angle deviation, the outer loop uses the impedance control algorithm, and its formula is To adjust the human-computer interaction force and dynamically optimize the motion control instructions based on the adjustment results.

[0034] In this optional embodiment, brain-computer interface data is acquired, the brain-computer interface data is preprocessed to obtain preprocessed brain-computer interface data, and motion features are extracted using the preprocessed brain-computer interface data to obtain a motion feature vector including: Acquire brain-computer interface data, use adaptive filtering to remove power frequency interference and myoelectric noise from the brain-computer interface data, and obtain denoised brain-computer interface data; The denoised brain-computer interface data is decomposed using wavelet transform, and the decomposed brain-computer interface data is combined with the preset frequency band to perform data screening to obtain preprocessed brain-computer interface data; Based on the common space pattern algorithm, the motion features of the preprocessed brain-computer interface data are extracted to maximize the difference in motion intentions and obtain the motion feature vector.

[0035] In this optional embodiment, the motion feature extraction that maximizes the difference in motion intention is performed on the pre-processed brain-computer interface data based on the common space pattern algorithm, and the obtained motion feature vector includes: The covariance of the movement intention is calculated based on the preprocessed brain-computer interface data to obtain the movement intention covariance matrix, and the spatial filter matrix is ​​solved for the movement intention covariance matrix to obtain the movement intention spatial filter matrix; The motion intention spatial filter matrix is ​​optimized using the pre-built objective function, and the variance difference of different motion intention signals is maximized according to the optimization result of the motion intention spatial filter matrix to obtain the variance difference results of different motion intention signals; Motion features are extracted based on the variance difference results of different motion intention signals, and a motion feature vector is generated based on the motion feature extraction results.

[0036] It should be noted that adaptive filtering is used to remove power frequency interference and myoelectric noise from brain-computer interface data, that is, to calculate the power ratio of power frequency interference in brain-computer interface data in real time. P 工频 ,like P 工频 >30%, then adjust the step size factor of the adaptive filter q =0.01+0.09×( P 工频 -30%) / 20%, and its value range is 0.01-0.1, and through the formula w ( n+1 ) =w ( n ) +2qe ( n ) x ( n )Update the filter weights so that the power frequency interference suppression ratio is ≥45dB.

[0037] It should be noted that the wavelet transform is used to decompose the denoised brain-computer interface data, that is, to calculate the power ratio of the electrooculogram artifact in the brain-computer interface data in real time. P 眼电 ,like P 眼电 >20%, then increase the number of wavelet decomposition layers from 5 to 5+2×( P 眼电 -20%) / 10%, and the value range is 5-7 layers, and the modal entropy threshold is introduced T entropy =0.8-0.2× P 眼电 / 50%, perform threshold filtering on the EOG signals in the 3-8 Hz frequency band to make the EOG artifact removal rate ≥96%.

[0038] In this optional embodiment, using an integrated learning model to identify motion intentions on motion feature vectors, generating motion control instructions based on the motion intention identification results, and performing motion control according to the motion control instructions includes: Construct an integrated learning model and use the pre-acquired brain-computer interface data to train the integrated learning model to obtain a movement intention recognition model; The motion intention recognition model is combined with the preset motion intention category to identify the motion intention of the motion feature vector and obtain the motion intention recognition result; The motion intention recognition result is subjected to motion intention confidence verification, and the motion parameters are calculated based on the motion intention confidence verification result combined with the preset motion parameter mapping rules and the pre-acquired human motion data, and motion control instructions are generated according to the motion parameters, and motion control is performed using the motion control instructions.

[0039] In this optional embodiment, an integrated learning model is constructed and trained using pre-acquired brain-computer interface data to obtain a movement intention recognition model including: Perform feature extraction on the pre-acquired brain-computer interface data to obtain a brain-computer interface feature dataset, and divide the brain-computer interface feature dataset into a training set, a validation set, and a test set based on a preset ratio; The convolutional neural network model, the support vector machine model and the random forest model are trained and optimized using the training set to obtain the optimized convolutional neural network model, the optimized support vector machine model and the optimized random forest model; The optimized convolutional neural network model, the optimized support vector machine model and the optimized random forest model are integrated in parallel to obtain a motion intention recognition model.

[0040] It should be noted that the optimized convolutional neural network model, the optimized support vector machine model and the optimized random forest model are integrated in parallel to obtain the motion intention recognition model, that is, the accuracy of each sub-model on the validation set is updated every 500 samples. , through the formula ; Calculate dynamic weight, where , so that the sub-model weight is positively correlated with the recognition accuracy.

[0041] In this optional embodiment, the convolutional neural network model, the support vector machine model, and the random forest model are trained and optimized using the training set, and the optimized convolutional neural network model, the optimized support vector machine model, and the optimized random forest model are obtained, including: The convolutional neural network model is initialized using the normal distribution to obtain an initial convolutional neural network model, the initial convolutional neural network model is iteratively trained according to the training set, and the convolutional neural network model is updated through the cross entropy loss function and back propagation to obtain an optimized convolutional neural network model; A support vector machine model with motion intention classification is constructed through kernel function selection and multi-classification strategy. The support vector machine model is trained based on the training set, and the model objective function is optimized by solving a convex quadratic programming problem to obtain an optimized support vector machine model. The training set is sampled with replacement to obtain a training subset, and a decision tree is constructed using the classification and regression tree algorithm. The splitting features and thresholds of the decision tree are optimized according to the Gini impurity. The random forest model is trained based on the optimized decision tree combined with the training subset to obtain an optimized random forest model.

[0042] In this optional embodiment, the motion intention recognition model is combined with a preset motion intention category to perform motion intention recognition on the motion feature vector, and the obtained motion intention recognition results include: Based on the convolution structure of the motion intention recognition model and the preset motion intention category, the motion feature vector is subjected to motion intention recognition with temporal correlation, and the motion intention classification probability with temporal correlation is obtained; The radial basis kernel function of the motion intention recognition model is used in combination with the preset motion intention category to perform motion intention recognition with statistical regularity on the motion feature vector, and obtain the motion intention classification probability with statistical regularity; The motion intention recognition model is based on a decision tree and a preset motion intention category to identify the motion intention with boundary features on the motion feature vector, and the motion intention classification probability with boundary features is obtained; The weighted voting method is used to integrate the results of motion intention classification probability with temporal correlation, motion intention classification probability with statistical regularity and motion intention classification probability with boundary features to obtain the motion intention recognition result.

[0043] It should be noted that the integration of classification probabilities using the weighted voting method also includes conflict resolution of sub-model probabilities. If the highest probability categories of the three sub-models are different, the difference between the highest probability and the second highest probability of each model is calculated. Δp = p 1- p 2. Prioritize adoption Δp The largest model result, where Δp When ≥0.5, it is judged as a high confidence result and output directly; Δp When <0.5, secondary recognition is triggered, and the sample collection time is extended to three seconds before recalculation.

[0044] In this optional embodiment, based on the acquired motion state feedback data, a dynamic adjustment algorithm is used to adjust the motion state, and the motion control instructions are dynamically adjusted according to the motion state adjustment result. The motion control optimization using the adjusted motion control instructions includes: Calculating the motion deviation of the acquired motion state feedback data, and using the motion deviation calculation result in combination with a preset threshold to perform motion adjustment judgment to obtain a motion adjustment judgment result; Based on the motion adjustment judgment result, a dynamic adjustment algorithm is used to adjust the motion state parameters of the acquired motion state feedback data to obtain a motion state adjustment result; The motion control instructions are dynamically adjusted according to the motion state adjustment result, and the motion control optimization is performed through the adjusted motion control instructions.

[0045] It should be noted that before adjusting the motion state based on the acquired motion state feedback data, user adaptation is also included, that is, collecting three minutes of motion imagery signals from new users, aligning the feature sequence of the new user with the preset standardized template through the dynamic time warping (DTW) algorithm, that is, aligning with the offline library containing more than 200 user features, and calculating the sequence similarity. Sim ,like Sim ≥0.75, weighted transfer learning is used to make the weight Sim The ensemble learning model parameters are initialized and the adaptation time is ≤ 10 minutes.

[0046] In this optional embodiment, based on the motion adjustment judgment result, the motion state parameters of the acquired motion state feedback data are adjusted using a dynamic adjustment algorithm, and the obtained motion state adjustment result includes: Based on the motion adjustment judgment result, the state value function and parameters of the dynamic adjustment algorithm are initialized to obtain an initialized dynamic adjustment algorithm; The initial dynamic adjustment algorithm is combined with the obtained motion state feedback data to select actions based on the exploration probability greedy strategy to obtain the execution action; Update the motion state adjustment parameters by executing the action, obtain the motion state feedback data of the executed action, calculate the instant reward based on the motion state feedback data of the executed action, and use the instant reward calculation result to update the state value function; The updated state value function and value iteration optimization strategy are used to adjust the motion state and obtain the motion state adjustment result.

[0047] It should be noted that the dynamic adjustment algorithm is used to adjust the motion state parameters, and also includes scene adaptation logic. If an outdoor scene is detected, that is, the ambient noise is ≥ 65dB, the exploration probability of the algorithm will be dynamically adjusted. e From 0.2 to 0.05, by the formula Correct the joint linkage delay time so that the motion synchronization error is ≤220ms.

[0048] Figure 2 An embodiment of the exoskeleton robot control system based on brain-computer interface of the present invention is shown.

[0049] In this optional embodiment, the exoskeleton robot control system based on the brain-computer interface includes: a motion feature extraction module 201, a motion control module 202 and a motion control optimization module 203; A motion feature extraction module 201 is used to obtain brain-computer interface data, preprocess the brain-computer interface data to obtain preprocessed brain-computer interface data, and extract motion features using the preprocessed brain-computer interface data to obtain a motion feature vector; A motion control module 202 is configured to identify motion intentions of motion feature vectors using an integrated learning model, generate motion control instructions based on the motion intention identification results, and perform motion control according to the motion control instructions; The motion control optimization module 203 is used to adjust the motion state using a dynamic adjustment algorithm based on the acquired motion state feedback data, dynamically adjust the motion control instructions according to the motion state adjustment result, and perform motion control optimization using the adjusted motion control instructions.

[0050] It should be noted that the motion feature extraction module is used to obtain brain-computer interface data, and performs adaptive filtering based on dynamic adjustment of the step size factor through a three-level dynamic preprocessing unit, as well as wavelet transform and common space pattern (CSP) feature extraction based on dynamic adjustment of the number of layers, and outputs a motion feature vector; the motion control module includes a heterogeneous integrated learning unit and an instruction generation unit, the heterogeneous integrated learning unit contains a convolutional neural network sub-module with 3 layers of convolution and 2 layers of residual, a support vector machine sub-module with a radial basis kernel function γ1=0.1-1, and a random forest sub-module based on 100 decision trees, through a dynamic weighted voting unit, where the weight is allocated according to the confidence of the sub-model, and the motion intention recognition result is output; the instruction generation unit is based on the recognition result and the preset motion parameter mapping rule, where the mapping rule contains the user height and weight correction formula Generate motion control instructions; the motion control optimization module includes a dual closed-loop adjustment unit, that is, the inner loop adjusts the joint angle deviation through the state value function update logic of the reinforcement learning subunit, and the outer loop adjusts the human-computer interaction force through the impedance control subunit, and dynamically optimizes the motion control instructions according to the adjustment results.

[0051] It should be noted that the brain-computer interface data includes raw EEG signals. The raw EEG signals of the user are obtained through EEG acquisition equipment, which is preprocessed to remove noise and artifacts. Feature vectors related to movement intention are extracted from the preprocessed EEG signals to obtain movement feature vectors. For EEG signal acquisition, dry electrode EEG acquisition equipment is used. The electrodes are arranged according to the international 10-20 system. The EEG signals of the motor cortex C3, C4 area and the prefrontal lobe Fz area are collected. The sampling frequency is set to 250Hz to ensure that the signal covers movement-related m Waves in the 8-13Hz frequency band β The wave is in the 13-30Hz frequency band. Signal preprocessing is done by using the adaptive filtering formula: y ( n ) =w T ( n ) x ( n ); e ( n ) =d ( n ) -y ( n ); w ( n+1 ) =w ( n ) +2qe ( n ) x ( n ). Where, y ( n ) represents the filter output; w ( n ) represents the weight vector; x ( n ) represents the input signal vector; e ( n ) represents an error signal; d ( n ) represents the expected signal; q Indicates the step size factor, the value range is 0.01-0.1; w ( n+1 ) represents the updated filter weight; the filter weight is updated in real time through the adaptive filtering formula to remove the power frequency 50Hz interference and myoelectric noise, and the signal-to-noise ratio of the filtered signal is improved to more than 30dB; the wavelet transform can use the db4 wavelet basis function to decompose the signal into 5 layers, decompose the EEG signal into sub-bands in the 8-30Hz frequency band, and obtain the brain-computer interface data after decomposition, retain the sub-band signal in the 3-30Hz frequency band (i.e., the preset frequency band), further suppress the electrooculogram and electrocardiogram artifacts, and the artifact removal rate reaches 92%, and the pre-processed brain-computer interface data is obtained. This frequency band covers the frequency band related to motor imagery. m wave andβ waves to effectively extract motion-related features.

[0052] Motion feature extraction uses the Common Spatial Pattern (CSP) algorithm to optimize the spatial filter through an objective function (i.e., a pre-built objective function) to maximize the variance difference of different motion intention signals. For example, for the two intentions of "left hand extension" and "right hand fisting," the feature vector extracted by the CSP algorithm can improve the differentiation between the two types of signals by 40%. The motion feature vector extracted using the Common Spatial Pattern algorithm is formulated as follows: ; Where, J ( W ) represents the objective function; W represents the spatial filter matrix; ∑1 represents the signal covariance matrix corresponding to the two types of motion intentions; ∑2 represents; tr Represents the matrix trace operation.

[0053] It should be noted that an integrated learning model is used to classify feature vectors, identify the user's movement intention, generate control instructions for the exoskeleton robot based on the movement intention, and send them to the actuator. An integrated learning model is constructed, which includes a convolutional neural network model (CNN), a support vector machine model (SVM), and a random forest model (RF). Among them, the CNN adopts a three-layer convolution structure and extracts local signal features through a 3×3 convolution kernel; the SVM uses a radial basis kernel function to classify high-dimensional features; and the RF contains 100 decision trees, which integrate the results through a voting method to determine the movement intention recognition result. The formula is: ; Where Γ represents the integrated voting result of the final motion intention recognition, and the output is the index of the specific motion intention category, such as "elbow extension" and "elbow flexion"; v i ( k ) indicates the i The classifier is k Voting value of motion-like intention; argmax k Indicates the index corresponding to the maximum value.

[0054] A multi-source EEG signal dataset (i.e., pre-acquired BCI data) was collected, encompassing at least 200 individuals of varying age groups (18-65 years old) and motor abilities, including healthy individuals, individuals with mild motor impairments, and individuals with moderate motor impairments. EEG signals were collected for each individual for six basic motor intentions, such as hand and leg raising, bending, fist clenching, stretching, and resting. For each intention, 100 valid samples were collected, each lasting 2 seconds and sampled at 250 Hz. After preprocessing, 128-dimensional feature vectors (i.e., the BCI feature dataset) were extracted. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio. The training set was used for model parameter learning, the validation set for hyperparameter tuning, and the test set for evaluating the final model performance. The training set was augmented using a ±10ms time axis offset and ±5% amplitude scaling to increase the sample size to 1.5 times the original to prevent model overfitting. Using a parallel integrated architecture, the three sub-models, CNN, SVM, and RF, each process the input 128-dimensional feature vector and output their own motion intent classification probabilities, namely 6-dimensional vectors corresponding to the probabilities of the six types of intent. Finally, a weighted voting method is used to fuse the final classification results (i.e., motion intent recognition results). Weight allocation can dynamically assign voting weights based on the accuracy of each sub-model on the validation set. The initial weight is set to 1. During training, it is adjusted every 5 epochs based on the accuracy of the validation set. The formula is: ; Where, w i Indicates the i The weight of each sub-model; acc i Indicates the i The accuracy of each sub-model on the validation set; acc j Indicates the jThe accuracy of each sub-model on the validation set. The total number of training rounds is set to 50 epochs, and each epoch contains a complete traversal of the training set. If the accuracy of the validation set does not improve for 10 consecutive epochs and the fluctuation is less than 0.5%, the training is terminated early and the current optimal model parameters are saved. The construction of the CNN model includes the input layer receiving a 128-dimensional EEG feature vector, which is converted into a 16×8 two-dimensional feature matrix through a reshape operation to simulate the spatial distribution; the convolution layer 1 uses a 3×3 convolution kernel, 32 in number, a stride of 1, a padding method of "same", an activation function of ReLU, and an output feature map size of 16×8×32; the pooling layer 1 uses a 2×2 maximum pooling, a stride of 2, and an output feature map size of 8×4×32; the convolution layer 2 uses a 3×3 convolution kernel, 64 in number, a stride of 1, a padding method of "same", an activation function of ReLU, and an output feature map size of The size of the feature map is 8×4×64; the pooling layer 2 uses 2×2 maximum pooling with a stride of 2, and the output feature map size is 4×2×64; the fully connected layer 1 flattens the features output by the pooling layer into a 512-dimensional vector, and connects it to a 512-dimensional fully connected layer with an activation function of ReLU and a dropout rate of 0.3; the output layer is a 6-dimensional fully connected layer with an activation function of softmax, which outputs the probability distribution of 6 types of motion intentions; the training parameter settings include the use of the Adam optimizer, an initial learning rate of 0.001, and a learning rate decay of 0.5 times the original every 10 epochs; the cross entropy loss function formula is: ; Where, L Represents the cross entropy loss function value, which measures the difference between the model prediction and the true label. The smaller the value, the more accurate the prediction. y i The true label i Components are encoded using one-hot encoding, such as "1" indicates that the category is a true value, and "0" indicates a false value; Indicates the model i The output probability for each category, ranging from [0 to 1], reflects the confidence of the model in predicting that category. The batch size is 32, meaning that 32 samples are input each time for parameter update. Model parameters are initialized using the He normal distribution. Training is performed iteratively on the training set, calculating the loss value at each iteration and updating the model parameters through backpropagation until the preset training rounds are reached or the early stopping mechanism is triggered. The model is saved, meaning the model parameters with the highest validation set accuracy are saved as the optimized convolutional neural network model.

[0055] The kernel function used in the construction of the SVM model is the radial basis kernel function (RBF), and the formula is: ; Where, K (u , y ) represents the output value of the radial basis kernel function, which is used to measure the similarity between two samples. The larger the value, the higher the similarity. u Represents the sample feature vector of the current input, such as EEG signal features, EMG signal features, etc., and its dimension is consistent with the sample feature space; y Represents the reference sample feature vector in the support vector set, that is, the support vector selected during the SVM model training process; g represents the kernel function parameter, also known as the "bandwidth parameter", which controls the range of the radial basis function and determines the optimal value through grid search; || u - y || 2 The Euclidean distance squared between the feature vectors of two samples measures the degree of difference between the samples in the original feature space. The multi-classification strategy uses a "one-vs-one" strategy to build 15 binary SVM sub-models between the six types of intents, and finally determines the multi-classification result through voting. The grid search range and penalty coefficient in parameter selection C The search range for 1 is [0.1, 1, 10, 100], g The search scope is [0.001, 0.01, 0.1, 1], the optimal parameter combination is determined by 5-fold cross validation on the validation set (the optimal C 1=10, g =0.1). The input 128-dimensional feature vector is normalized so that the mean of each feature is 0 and the standard deviation is 1. The formula is: ; Where, x ' represents the normalized feature vector, which has the same dimension as the original feature vector, 128 dimensions, with a mean of 0 and a standard deviation of 1; f The feature vector representing the original input, specifically 128 dimensions, such as multimodal features extracted from EEG or EMG signals; m Represents the mean of the original feature vector, which is calculated through the training set, that is, ,in, N is the sample size, f i Indicates the i Original feature vectors; s Represents the standard deviation of the feature. The SVM model is trained using the training set, and the model objective function is optimized by solving the convex quadratic programming problem. The model objective function formula is: ; Where, N Represents the total number of training samples, that is, the number of samples participating in model training;w Represents the weight vector of the classification hyperplane, which determines the direction and tilt of the hyperplane, and its dimension is consistent with the feature space; x i Indicates the i The non-negative slack variable of the sample is used to allow some samples to not meet the strict classification constraints, that is, the soft margin SVM. The larger the value, the further the sample deviates from the ideal classification boundary. C 1 represents the penalty coefficient, and the optimal value has been determined to be 10 through grid search, which is used to balance "maximization of classification interval" and "misclassification penalty". C The larger the value of 1, the heavier the penalty for misclassified samples, that is, the model tends to be more strict in classification and may overfit; C The smaller the value of 1, the more misclassification is allowed, that is, the model focuses more on maximizing the margin and may underfit. Evaluate the model performance on the validation set, adjust the parameters, and save the optimized support vector machine model.

[0056] The constraints are: ; Where, h i Indicates the i The label of a sample is usually +1 or -1 in binary classification, representing two categories respectively; x i Indicates the i The feature vector of the training sample is a 128-dimensional feature that has been standardized; b Represents the bias term of the classification hyperplane, which determines the position of the hyperplane in the feature space, that is, it defines the hyperplane together with the weight vector. The core of this objective function is to find a balance between maximizing the classification interval and minimizing the classification error. represents the maximum classification interval, and the interval size is related to || w || 2 Inversely proportional; Represents minimizing the total slack variable, that is, minimizing the penalty for classification errors; the constraint ensures that the samples are classified correctly as much as possible and allows a small number of samples to be fault-tolerant by the slack variable. By solving this convex quadratic programming problem, the optimal weight vector can be obtained w and bias b , thereby determining the optimal classification hyperplane.

[0057] The RF model is constructed to determine the number of decision trees, specifically 100 decision trees, which is determined by the validation set. That is, when the number of decision trees exceeds 100, the model performance improvement is not obvious. The parameters of a single decision tree include the maximum depth set to 10 to avoid overfitting, and the number of features randomly selected when each node is split. , the minimum number of samples per leaf node is 5. Bootstrap sampling is used to sample the training set with replacement, generating an independent training subset for each decision tree. The sample size is the same as the original training set. Each decision tree is constructed using the Classification and Regression Tree (CART) algorithm, selecting the optimal splitting feature and splitting threshold based on the Gini impurity formula: ; Where, G Gini Impurity is used to quantify the degree of confusion of sample categories in a node. If all samples in a node belong to the same category, the purity is the highest. G =0; if the samples in the node are evenly distributed K The highest degree of confusion among the categories is G =1-1 / K , such as the maximum value in the binary classification is 0.5. r i Indicates the node i The proportion of samples in the class, that is, the number of samples in the node that belong to i The ratio of the number of samples of the class to the total number of samples of the node satisfies ,in T is the total number of sample categories. 100 decision trees classify the input samples separately, and the output of the RF model is determined by majority voting. The contribution of each feature in all decision tree splits is calculated, that is, the sum of the Gini impurity reduction of the feature, and the feature importance is evaluated. Features with a contribution of less than 0.1% are removed to simplify the model while improving the generalization ability, and the model is optimized. Post-prune each decision tree to remove branches that do not improve the accuracy of the validation set, reduce the complexity of the model, and obtain an optimized random forest model. Through the above specific steps, the integrated learning model and each sub-model can be completely constructed and trained to ensure the reproducibility of the model and avoid the problem of insufficient patent disclosure. Experimental verification shows that the accuracy of the model trained according to the above steps on the test set can reach 93.5%, which meets the actual needs of exoskeleton robot control.

[0058] Testing showed that the model achieved a 93.5% accuracy rate in recognizing six basic motor intentions, such as raising a hand, raising a leg, and bending, with a response time of ≤200ms. The testing hardware included an EEG acquisition device, a 32-channel dry-electrode EEG recorder with a sampling frequency of 250Hz, an input impedance ≥100MΩ, and a noise level ≤1μV. The computing equipment consisted of a workstation equipped with an Intel Xeon E5-2690 processor and an NVIDIA Tesla V100 graphics card, specifically 32GB of video memory, ensuring a model inference latency of ≤200ms. The exoskeleton's lower limb hip and knee joints had a range of motion of ±90°, and the torque sensor had an accuracy of ±0.1N·m. The test data set contains 200 testers, specifically 120 healthy people, 50 patients with mild movement disorders, and 30 patients with moderate movement disorders. The male-to-female ratio is 1:1, the age distribution is 18-65 years old, and the average age is 38.5 years old. Six types of movement intention samples are collected for each person category, including hand raising, i.e. shoulder flexion, leg raising, i.e. hip flexion, bending, i.e. lumbar flexion, fist clenching, i.e. finger flexion, extension, i.e. whole body extension, and stillness. 100 groups of valid samples are collected for each type of intention, i.e. samples with severe blinking and myoelectric interference are eliminated. After preprocessing, the sample features are 128-dimensional feature vectors, including time domain features such as mean, variance, and peak, a total of 32 dimensions, frequency domain features of 8-30Hz power spectrum, a total of 64 dimensions, and spatial domain features, such as C3 / C4 electrode coherence, a total of 32 dimensions. Load the trained CNN, SVM, and RF weight files in the integrated model and the mean and standard deviation matrices in the standardized parameters; initialize the exoskeleton control interface, set the communication baud rate to 115200bps, and ensure that the command transmission delay is ≤50ms. The single-sample test process includes the tester wearing the EEG device and completing the specified movement intention such as "lifting the leg" according to the command, collecting 2 seconds of EEG signals; then the signal is adjusted by the step size factor q =0.05 adaptive filtering, db4 wavelet transform decomposition to 8-30Hz; then extract 128-dimensional feature vectors, normalize them and input into the integrated model; finally, CNN, SVM, and RF output classification probabilities respectively, and assign weights according to the weighted voting method, such as CNN weight 0.35, SVM weight 0.3, and RF weight 0.35 to fuse and output the final intention (i.e., motion intention recognition result); record the matching between the model recognition result and the actual intention, and calculate the single sample recognition accuracy. Batch testing is a batch test of 200 people × 6 categories × 10 groups = 12,000 samples in the test set. The accuracy of each type of intention and each tester is calculated separately; calculate the overall accuracy, that is, the number of correctly identified samples / total number of samples, the category accuracy, that is, the number of correct intentions of each category / total number of samples of that category, and the 6×6 confusion matrix, element ( i , j ) indicates the i Class was misclassified as jThe test results showed an overall accuracy of 93.5%, meaning 11,220 out of 12,000 samples were correctly identified. The category accuracy rates were 95.2% for hand raising, 94.8% for leg raising, 92.3% for bending, 96.1% for fist clenching, 91.5% for stretching, and 92.6% for stillness. Key confusion matrix values ​​included a 3.2% misclassification rate for "leg raising" and "bending," as both involve torso movement. The misclassification rates for other categories were ≤2%. The inference latency averaged 185ms, meaning ≤250ms from signal acquisition to intent output, meeting real-time control requirements. This is shown in Table 1.

[0059] Table 1 Comparison example with a single model ; The comparative analysis in Table 1 shows that the integrated model improves the accuracy by 6.2%-7.9% through complementarity, especially for patients with moderate movement disorders with poor signal quality, the accuracy is improved by 10.4%-16.6%, because SVM is sensitive to noise, while CNN and RF can make up for its shortcomings. Compared with the existing technology, the accuracy of single CNN is 86.2%, and patients with movement disorders are not considered; the accuracy of SVM+RF integration is 89.7%, and the inference delay is 280ms, which exceeds the real-time requirement; the accuracy of the present invention is improved by 3.8%-7.3%, the delay is reduced by 34%-95ms, and it is the first time to achieve adaptation for patients with different degrees of movement disorders. The key technical reason for the accuracy is the multi-model complementary mechanism. CNN is good at capturing the local temporal correlation of features, such as m The instantaneous change of wave attenuation, for the fine movement of "clenching fist", the signal characteristics are clear and the recognition rate is 96.1%; RF is good at processing the global statistical laws of high-dimensional features, and the stable recognition rate of "stationary" signals is 92.6%; SVM performs well when the feature boundaries are clear. After fusion with the first two, it solves the fuzzy classification problem of "lifting legs" and "bending over", and the misjudgment rate is reduced from 8.7% to 3.2%. Among them, the data enhancement and generalization design simulates the reaction time differences of different testers through time axis offset, and the amplitude scaling simulates the fluctuation of EEG signal intensity, so that the overfitting coefficient of the model on the test set, that is, the training accuracy - test accuracy ≤2.3%; for patients with movement disorders, through weighted sampling, the weight of patient samples in the training set is increased by 1.5 times, which enhances the model's recognition ability for weak signals. Dynamic weight optimization is achieved by using CNN on the validation set at 8-13Hz m When the wave feature is significant, the weight is increased to 0.4, and SVM is at 30Hz. βWhen wave features are clear, the weight is increased to 0.35, ensuring that each model maximizes its effectiveness in its preferred scenarios. Stability testing results, based on 10 consecutive days of testing, with 1,200 samples repeated daily, showed an accuracy fluctuation range of 93.5% ± 0.8%, with no significant drift. After replacing three exoskeleton devices from different batches, the accuracy remained at 93.2%, validating robustness due to the decoupling of the model from the hardware interface. Through this testing process and comparative analysis, it is clear that the integrated model's 93.5% accuracy is reproducible, outperforming existing technologies, and fully meeting the control requirements of exoskeleton robots.

[0060] Based on the identified motion intent, it is mapped to parameters such as the exoskeleton's joint angles and movement speeds. The mapping rules for motion intent to exoskeleton joint parameters are as follows: a kinematic parameter mapping table (preset motion parameter mapping rules) is pre-established for each of the six types of motion intent (i.e., pre-set motion intent categories). This table clearly defines the target joints for each type of intent, including the hip, knee, and shoulder joints, as well as their angle range, speed range, and motion timing. The specific rules are shown in Table 2.

[0061] Table 2 Basic mapping table ; It should be noted that the angles in the table are based on the neutral position of the human body as 0°, flexion as positive, and extension as negative; the speed is the angular velocity of the joint movement. The dynamic parameter calculation includes the intention confidence check, that is, extracting the probability distribution of the movement intention output by the integrated model. When the highest probability P max ≥0.85, directly use the basic parameters corresponding to the intention; when 0.7≤ P max When <0.85, fuzzy correction is activated and weighted calculation is performed based on the parameters of the second highest probability intention. The formula is: ,in, P max The confidence level of the highest probability intention, i.e., the probability value, ranges from [0, 1], indicating the degree of trust the model has in the "most likely motion intention", such as the highest probability output by SVM or decision tree. i target Represents the target control parameters after fuzzy correction, such as the joint angle, movement speed and other execution parameters of the exoskeleton robot, as the final output control instruction; i 1 is the angle of the highest probability intention, and the control parameters corresponding to the highest probability intention, such as the joint angle target value corresponding to the "elbow extension" intention, i 2 is the angle of the second highest probability intention. The control parameters corresponding to the second highest probability intention, such as the joint angle target value corresponding to the "elbow flexion" intention, have a probability second only to the highest probability; that is, the weight is distributed according to the probability. The dynamic calculation of joint angle includes the correction of the user's height and weight in the pre-acquired human motion data. i base Correct the range and angle i adj , the correction formula is: ; ( H is height in cm, W =bodyweight in kg, 170cm and 65kg are the benchmark values); the upper limit of the angle is adjusted based on the user's athletic ability level (healthy person / mild disability / moderate disability) in the pre-acquired human motion data, i.e., healthy people use 100% of the basic angle; mild disability uses 80% of the basic angle (e.g., the leg lifting angle is reduced from 60° to 48°); moderate disability uses 60% of the basic angle (e.g., the leg lifting angle is reduced to 36°); the final target angle (i.e., motion parameter) ( k ability is the capacity coefficient, 1 / 0.8 / 0.6).

[0062] The dynamic calculation of movement speed is based on the positive correlation between speed and EEG signal intensity. The calculation formula is: ;in, v final is the final motion speed; v base As the basic movement speed; S eeg is the normalized intensity of the EEG signal (0-1, calculated by the power of the 8-30 Hz frequency band). For example, when S eeg =0.8 (strong signal), leg lifting speed ;when S eeg =0.3 (weak signal), leg lifting speed The upper limit of the speed shall not exceed the maximum value of the basic range, and the lower limit shall not be lower than the minimum value. The determination of the motion timing parameters includes the correction of the joint linkage delay time according to the spinal nerve conduction velocity. The correction formula is: ; v nerve The spinal nerve conduction velocity is m / s, with 80m / s as the benchmark and the normal range being 60-120m / s. t delay is the corrected joint linkage delay time; t base is the initial joint linkage delay time, for example, when v nerve =100m / s (fast conduction), knee joint lag time when raising the leg t =50ms×(1-0.01×20)=40ms; when v nerve =60m / s (slow conduction), lag time t=50ms×(1-0.01×(-20))=60ms. Format the parameters and output them. i final 、v final 、t delay The motion parameters are converted into an instruction format recognizable by the exoskeleton control system (i.e., motion control instructions). For example, the angle instruction is 16-bit binary code (range -128° to +128°, accuracy 0.5°); the speed instruction is 8-bit binary code (range 0° / s to 100° / s, accuracy 1° / s); and the timing instruction is 10-bit binary code (range 0ms to 1023ms, accuracy 1ms).

[0063] Sensors collect information about the exoskeleton's motion state (i.e., acquired motion state feedback data), build a feedback mechanism (i.e., a dynamic adjustment algorithm), and dynamically adjust control instructions. The feedback mechanism is implemented through a reinforcement learning algorithm, and the state value function update formula is: ; Where, V '( s ) represents the updated state value function; V ( s ) represents the current state value; α represents the learning rate; r Indicates the reward value; c represents the discount factor; s ' indicates the next state. Perform feedback verification, that is, after the exoskeleton executes the command, the actual angle is collected in real time through the joint encoder i real , calculate the deviation: e θ =| i final - i real |. When e θ When the angle is >3°, the trigger parameter is corrected twice, and the correction amount is: ; "sign" is the sign function, until the deviation is ≤3°. For example, the parameter mapping process of the "lift leg" intention recognizes the intention as "lift leg". P max =0.92 (≥0.85), call the basic parameters (hip joint 0°-60°, speed 20° / s-40° / s); based on the user's height of 180cm and weight of 70kg, perform angle correction, specifically: i adj=60°×(1+0.005×10)×(1+0.003×5)=60°×1.05×1.015≈63.9°; the user has a mild disability. k ability =0.8, i final =63.9°×0.8≈51.1°; EEG signal intensity S eeg =0.6, speed v final =40° / s×(0.5+0.5×0.6)=32° / s; spinal cord conduction velocity v nerve =90m / s, knee lag time t =50ms×(1-0.01×10)=45ms; after the exoskeleton is executed, the actual angle is 50.8°, with a deviation of 0.3° (≤3°), and the parameters are valid. Through these steps, precise mapping of movement intent to exoskeleton joint parameters is achieved, ensuring that the exoskeleton's movements closely match the human's intended movements for different users and motion states. The "leg raise" intention corresponds to a target hip joint angle of 45° and a movement speed of 0.3m / s. Control commands are transmitted to the actuator via the CAN bus with a transmission delay of ≤50ms. Testing has shown that the average error in parameter mapping is ≤2.5° (angle) and ≤2° / s (speed), meeting the accuracy requirements for exoskeleton control.

[0064] The force sensor and inertial measurement unit (IMU) are used to collect information such as the joint torque and motion trajectory of the exoskeleton (i.e., the obtained motion state feedback data), and the control parameters are dynamically adjusted using the reinforcement learning algorithm (i.e., dynamic adjustment algorithm). If the deviation between the actual motion angle and the target angle exceeds 5°, the control instructions are updated through the state value function to quickly converge the deviation to within ±2°. The core elements of the reinforcement learning algorithm include the state space (S), which is defined as the characteristic vector of the current motion state of the exoskeleton, including the joint angle deviation e θ =| i final - i real |(unit:°); joint angular velocity deviation e v =| v target - v real | (unit: ° / s); e v Indicates the joint angular velocity deviation, v Velocity is a non-negative value that reflects the difference between the target angular velocity and the actual angular velocity. The smaller the value, the higher the control accuracy.v target The target angular velocity of the joint, i.e. the expected angular velocity, is calculated by the control system according to the motion intention, such as the target parameter after fuzzy correction. i target Derived; v real Indicates the actual angular velocity of the joint, which is the robot joint movement speed measured in real time by sensors such as encoders and IMUs; joint torque t real (Unit: N·m); EEG signal intensity S eeg (normalized to 0-1); the state vector is represented as s = e θ , e v , t real , S eeg , all dimensions are normalized (mapped to the interval [-1,1]); the action space (A) is defined as the adjustable control parameter increment, including the angle correction Dth (Range -5°-+5°, step size 0.5°); speed correction Δv (range -5° / s-+5° / s, step size 0.5° / s); torque correction coefficient k τ (range 0.8-1.2, step size 0.05); the action vector is represented as a =[ Dth , Δv , k τ The reward function (R) is designed based on motion deviation and safety, and the formula is: ;in, R ( s , a ) represents the reward value, which is the state s and actions a The function of , the larger the value, the more actions are performed in state s a The better the effect, the more comprehensive the motion accuracy, speed control and safety; s Represents the current state, such as multi-dimensional state vectors such as robot joint angles, angular velocities, and patient biosignals; a Represents the actions performed by the agent, such as control instructions such as robot joint torque adjustment and motion trajectory correction; e θ Indicates the joint angle deviation, that is, the difference between the target angle and the actual angle. The smaller the value, the more precise the angle control. e vIndicates the joint angular velocity deviation. The smaller the value, the smoother the speed control. t real Indicates the actual output torque of the robot joint, measured by the torque sensor; t max Indicates the maximum safe torque of the joint at the current angle, referring to the dynamic boundary in the "Safety Constraint Working Domain", such as , q 1 indicates the current angle; I ( t real > t max ) represents the indicator function, also known as the "indicator function", if t real > t max , that is, the actual torque exceeds the safety threshold, then ;otherwise , used to impose penalties for unsafe behavior; is an indicator function. When the torque exceeds the safety threshold t max When , the value is 1, otherwise it is 0; when e θ ≤2° and e v When ≤2° / s, an additional 5 points will be awarded (to encourage convergence to the target range); the state value function ( V ( s )), which means that in the state s Under this circumstance, the expected long-term cumulative reward that can be obtained by following the current strategy is: V ( s )=E[ G t | s t = s ]=E[ R t+1 +γ R t+2 +γ 2 R t+3 +...| s t = s ];in, V ( s ) represents the state value function, referred to as "state value", that is, in the current state s The expected long-term cumulative reward that can be obtained by following the optimal strategy under the same circumstances, that is, the average value. The larger the value, the more "favorable" the state; Represents the mathematical expectation, that is, the mean operator, which is the average value of a random variable. Due to the randomness of the environment and actions, it is used here to quantify the average level of long-term cumulative rewards; G t Indicates from time t The initial cumulative reward, also known as the "reward", is the weighted sum of all future rewards, i.e. G t = R t+1 +γ R t+2 +γ 2 R t+3 +...; s t = s represents a conditional expression, i.e. "at time t the system is in state s", s t For the moment t status, s For a specific state; R t+z Indicates time t + z The immediate reward obtained, such as the reward function mentioned above R ( s , a ) calculated value, z =1,2,… means the next z The reward of the next step; γ represents the discount factor, which has been set to 0.9 and ranges from [0,1], and is used to “decay” the weight of future rewards, γ z Follow z When γ=0.9, it pays more attention to the recent rewards, but also considers the long-term benefits, which is used to weigh the immediate rewards and future rewards. G t From the moment t The initialization phase of reinforcement learning to dynamically adjust the control parameters is the state value function V ( s ) is initialized to a 0 matrix (covering all possible states); the learning rate α =0.1, exploration probability (20% probability of randomly selecting an action, 80% probability of selecting the optimal action); set the maximum number of iterations T =100 (maximum number of steps for a single adjustment). Real-time adjustment includes state perception, which means collecting data through force sensors and IMU every 50ms to calculate the current state. s t =[ e θ , e v , t real , S eeg ],when e θ >5° triggers the reinforcement learning adjustment mechanism; action selection, that is, based on The greedy strategy selects actions based on the probability Randomly select actions a t ; with probability Choose the action that maximizes the immediate reward a t =argmax a R( s t ,a). a t Indicates at time t The specific action selected is achieved by maximizing the immediate reward R( s t ,a) determine; for example, when e θ =6° (deviation exceeds the standard), preferred Dth = +1° (positive correction) action; execute action and state transfer, execute action through exoskeleton a t , update the control parameters: ;in, i target ' represents the updated joint target angle, that is, the corrected angle instruction; i target Represents the joint target angle before updating, that is, the original angle command; Dth Represents the angle correction amount, that is, the angle adjustment value calculated according to the control strategy or feedback; v target ' represents the updated joint target angular velocity, that is, the corrected velocity instruction; v target Indicates the joint target angular velocity before updating, that is, the original velocity instruction; Δv Represents the speed correction, that is, the speed adjustment value calculated according to the control strategy or feedback; t target ' represents the updated joint target torque, that is, the corrected torque command; t target Represents the joint target torque before updating, that is, the original torque command; k τ Represents the torque correction factor, which is used to proportionally adjust the target torque, such as calculated based on safety constraints or load changes; collects the new state after execution (i.e., the motion state feedback data of the execution action), and the new state iss t+1 =[ e θ ', e v ', t real ', S eeg '];in, s t+1 Indicates time t +1 is the new state vector, which is the state of the system after the action is performed; e θ ' represents the updated value of the angle deviation, that is, the deviation between the corrected target angle and the actual angle; e v ' represents the updated value of the angular velocity deviation, that is, the deviation between the corrected target angular velocity and the actual angular velocity; t real 'Indicates the actual output torque of the joint after the update is executed, which can be measured by the sensor; S eeg ' represents the updated EEG signal features, that is, the EEG signal processing results collected after the action is performed; calculate the immediate reward based on the new state R t = R ( s t , a t ), and update the state value function. The specific update formula is: V ( s t )+ α [ R t +γ V ( s t+1 )- V ( s t )];in, V ( s t ) indicates the time t state s t The value function of , that is, the estimated value before the update; α Represents the learning rate, which ranges from [0,1] and controls the update amplitude of the value function, such as α =0.1 is the deviation of 10% for each update; R t Indicates time t The immediate reward is based on the reward function R ( s t ,a t ) calculation, reflected in the state s t Execute an action a t effect; V ( s t+1 ) indicates the time t +1 New Status s t+1 the value function (estimate) of R t +γ V ( s t+1 ) represents the target value, also known as the TD target, and also represents the ideal estimate of the current state value; R t +γ V ( s t+1 )- V ( s t ) represents the temporal difference error, or TD error, which measures the deviation between the current value estimate and the ideal estimate and is used to drive the update of the value function; for example, if V ( s t )=5, R t =8, V ( s t+1 )=6, then the updated result is 5+0.1×(8+0.9×6-5)=5+0.1×(8+5.4-5)=5.84; if e θ If the error is less than 2° and the error is stable for three consecutive cycles, the adjustment is terminated; otherwise, the state perception is returned until the maximum number of iterations is reached. T Among them, the strategy optimization is to optimize the strategy through the value iteration method every time 100 adjustment cycles are completed, so that the action selection is more inclined to the high-value state: in, Indicates status s The optimal strategy under The asterisk in the box indicates the optimal output. s The optimal action to be chosen under argmax a Indicates action a The maximum value operation is to select the action that maximizes the value of the subsequent expression from all possible actions; It represents the mathematical expectation operator, which represents the average value of a random variable, that is, there is randomness due to environmental feedback; R Indicates execution of an action aThe immediate reward obtained after R ( s , a )calculate; V ( s ') indicates execution of an action a The new state after transfer s 'The state value function also represents the long-term value of the new state; s 'Indicates execution of an action a The new state after transfer is the same as the original state symbol s To distinguish, identify the subsequent state by apostrophe; | s , a Indicates a conditional expression, also indicates that in the current state s Next action a Under the condition of gradually reducing the exploration probability (Decrease by 0.02 every 100 cycles, to a minimum of 0.05), improving the stability of the strategy. State value function V ( s ) is the core basis for action selection. s Next, give priority to the V ( s ') Maximized action a For example, when V ( s 1)=10(corresponding action a 1), V ( s 2)=7(Corresponding action a 2) When, V ( s 1) Indicates execution of an action a 1 after the transfer to the new state s 1 state value function; V ( s 2) Indicates execution of an action a 2 after the transfer to the new state s 2 state value function; the algorithm will choose a 1 in pursuit of higher long-term rewards. Value update drives parameter convergence by continuously updating V ( s ), the algorithm can learn from historical adjustment experience, when a certain action a Can quickly reduce the deviation ( e θ from 6° to 2°), corresponding to R t Increase, V ( s t) increases, the probability of the action being selected in similar future states increases; when the action causes the torque to exceed the limit ( t real > t max ), R t becomes negative, V ( s t ) decreases, the algorithm will avoid choosing this action again. The rate of change of the state value function ΔV = V ( s t+1 )- V ( s t ) is positively correlated with the convergence speed of the angle deviation. ΔV >0, the deviation convergence speed is accelerated (an average reduction of 0.8° per step); when ΔV When ≤0, the algorithm will trigger a strategy correction to find a better action by increasing the exploration probability. The deviation adjustment process of the "lift leg" intention includes the initial state e θ =7° (exceeding the standard), e v =3° / s, t real =12N·m (within the limit), s =[7,3,12,0.6], V ( s )=3; Action selection is selection a =[+1°,+0.5° / s,1.0] (increase angle target and increase speed); the new state is e θ '=5.5°, e v '=2.2° / s, R t =10-2×5.5-1×2.2=10-11-2.2=-3.2; the value is updated to ; The iterative result is after 8 steps of adjustment, e θ =1.8° (converges to within ±2°), at this time V ( s ) = 12 (high-value state). Through the above steps, the reinforcement learning algorithm uses the state-value function to dynamically optimize the control parameters, ensuring that the exoskeleton quickly converges when deviations exceed the limit. Tests show that the average adjustment time is ≤ 0.5s, significantly better than traditional PID control (1.2s).

[0065] In summary, the use of a multi-level signal preprocessing algorithm significantly improves the quality of EEG signals, providing a reliable data basis for subsequent recognition; the integrated learning model improves the accuracy and generalization ability of motion intention recognition, and adapts to the signal characteristics of different users; the feedback adjustment mechanism realizes dynamic optimization of control instructions, reduces motion deviation, and improves the safety of human-computer collaboration; the control response delay is ≤250ms, which meets real-time control requirements and is suitable for high-precision scenarios such as rehabilitation training.

[0066] Taking lower limb rehabilitation training for patients with spinal cord injury as an example, the patient wears an EEG acquisition device with electrodes attached to the C3, C4, and Fz areas. After system initialization, the system collects 5 minutes of resting-state EEG signals as a baseline. The patient then imagines the movement of "raising the left leg." The device collects raw EEG signals, adaptively filters to remove 50Hz power frequency interference, and then uses a wavelet transform to remove eye contact artifacts. The CSP algorithm extracts feature vectors and inputs them into an ensemble learning model. The CNN layer outputs a probability of "raising the left leg" of 0.91, the SVM outputs 0.88, and the RF outputs 0.93. Voting confirms the intention. Control commands are generated: a target hip angle of 30° and a target knee angle of 20°, driving the exoskeleton to raise the left leg at a speed of 0.2 m / s. The force sensor detects an abnormal hip torque (exceeding a preset threshold of 15 N·m) and feeds it back to the control system. Reinforcement learning adjusts the command to reduce the torque to 12 N·m, keeping the angle deviation within 3°. During training, the system updates model parameters every 10 minutes to adapt to changes in the patient's signal characteristics and ensure long-term control stability. When applied to different users, the dynamic time warping (DTW) algorithm is used to align feature sequences, allowing the model to quickly adapt to new users, shortening the adaptation time to less than 10 minutes. The dynamic time warping (DTW) algorithm aligns feature sequences, including the sequence definition of feature sequences, where the feature sequences to be aligned are template sequences and test sequences. The template sequence T The standard motion intention feature sequence from the training set is of length n , expressed as T =[ t 1, t 2,..., t n ],in For the i 128-dimensional feature vector at the moment (corresponding to 8-30Hz EEG features); test sequence X is the feature sequence to be identified collected in real time, with a length of m ( m and n may not be equal), expressed as X =[ x 1, x 2,..., x m ],in For the j The difference in sequence length comes from the difference in movement execution speed of different users (for example, a healthy person takes 1.2 seconds to complete a "leg lift", while a person with movement disorders takes 1.8 seconds), which leads to the length of the feature sequence. n =300 (1.2s×250Hz), m =450 (1.8s×250Hz). Normalize the time axis of the features of each dimension in the sequence to eliminate amplitude differences: , where Indicates the first k Under the feature dimension, i The original feature value of each time step. The template feature sequence here can be understood as a predefined feature sequence used as a reference, such as the standard feature sequence when a healthy person completes the "leg lifting" action; Indicates that after the test feature sequence is normalized by time axis, k Under the feature dimension, j The test feature sequence is the feature sequence to be analyzed and compared with the template sequence, such as the feature sequence of a patient with movement disorders completing the "leg lift" action. The purpose of normalization is to eliminate the amplitude differences caused by factors such as different movement execution speeds among different users, so that subsequent comparison with the template sequence can be carried out at the same scale. Indicates the first k Under the feature dimension, j The original eigenvalues ​​of the time steps are the original data without normalization. k =1,2,...,128 is the index of the feature dimension, m k 、 s k For the training set k The mean and standard deviation of the dimensional features are used to ensure that the template and the test sequence are compared at the same scale. The construction of the DTW distance matrix includes local distance calculation, specifically calculating the Euclidean distance between any two feature vectors in the template sequence and the test sequence as the elements of the local distance matrix D. The calculation formula is: Where, Indicates the first i The normalized feature vector of each time step. In addition, the template sequence is a predetermined feature sequence used as a reference standard, such as the brain-computer interface feature sequence when a healthy person completes a certain action. It has been normalized on the time axis, that is, the feature vector after the amplitude difference is eliminated. Indicates the test sequence jThe normalized feature vector of time steps is obtained. The test sequence is the feature sequence to be analyzed and compared with the template sequence, such as the brain-computer interface feature sequence of a patient with movement disorders completing the same action, which has also been normalized on the time axis. Indicates the first i In the feature vector of the time step, k The standardized eigenvalues ​​of feature dimensions; Indicates the test sequence j In the feature vector of the time step, k The standardized eigenvalues ​​of feature dimensions. i =1,2,..., n , j =1,2,..., m ,matrix D The size is n × m Optimization processing is D ( i , j )> i dist (distance threshold, set to 15), forced D ( i , j )=+∞ (avoid matching of irrelevant points); the cumulative distance matrix is ​​constructed to define the cumulative distance matrix C ,in C ( i , j ) means t i and x j The minimum cumulative distance during alignment is calculated as follows: C ( i , j )= D ( i , j )+min{ C ( i -1, j ), C ( i , j -1), C ( i -1, j -1)}; its boundary conditions are: C (1,1)= D (1,1); C ( i ,1)= D ( i ,1)+ C (i -1,1) (the first column can only be accumulated from top to bottom); C (1, j )= D (1, j )+ C (1, j -1) (the first row can only be accumulated from left to right). The search for the optimal alignment path includes setting path constraints to avoid excessive path distortion and setting slope constraints (Sakoe-Chiba belt): in α =0.1 is the constraint coefficient, that is, the alignment point ( i , j ) must fall within 10% of the bandwidth on both sides of the main diagonal to reduce the amount of calculation while ensuring the rationality of alignment. Backtracking method is used to find the optimal path, that is, from the lower right corner of the matrix C ( n , m ) starts backtracking; n yes i The maximum value of m yes j The maximum value of , each time choose the direction with the smallest cumulative distance from the previous step (up, left, left-up), record the path point ( i , j );like C ( i , j )- D ( i , j )= C ( i -1, j ), then the previous step is ( i -1, j ) (vertical movement, stretching template sequence); if C ( i , j )- D ( i , j )= C ( i , j -1), then the previous step is ( i , j -1) (horizontal movement, stretch test sequence); if C ( i , j )- D ( i , j )=C( i -1, j -1), then the previous step is ( i -1, j-1) (diagonal movement, synchronous alignment). The termination condition is backtracking to C Stop at (1,1) to get the optimal alignment path P =[( i 1, j 1),( i 2, j 2),...,( i L , j L )],in L is the path length ( L ≥max( n , m )). Path smoothing is the process of smoothing the path. P Perform sliding average filtering to eliminate local jitter: ; The first and last points are not processed, among which, Indicates the optimal alignment path after sliding average filtering. P Middle k The time step index of the template sequence corresponding to each path point. The purpose of the sliding average filter is to eliminate the local jitter of the path and make the aligned path smoother. This is achieved by averaging the time step index of the current and adjacent template sequences. Indicates the optimal alignment path after sliding average filtering. P Middle k The time step index of the test sequence corresponding to each path point is also used to eliminate local jitter of the path. A smoother result is obtained by averaging the time step index of the current and adjacent test sequences, ensuring the continuity of the aligned path. The time regularization of the aligned sequence is achieved by P , the length is m Test sequence X Regularized to template sequence T Equal length ( n ) sequence ; For the template sequence i points t i , find all the i k = i Corresponding test points ; If it corresponds to multiple test points, take the average: ;in, K 1 is the number of matching points; if it corresponds to a single test point, assign the value directly The alignment effect evaluation includes calculating the sequence similarity after alignment: ;in, C ( n , m) in the lower right corner of the matrix, which is also the total cumulative distance, For example, the sequence alignment of the "lift leg" intention, the template sequence T ( n =300), that is, the EEG characteristic sequence of "leg lifting" in healthy people, the key time point i =150 corresponds to a hip joint angle of 30°; test sequence X ( m =450); the characteristic sequence of the patient's "leg lifting" is slow, and the 30° angle corresponds to j =250; the alignment result is the path P (150,250) is the key matching point. The length is 300, and T Similarity Sim =0.82, ensuring the timing consistency during subsequent intention recognition. Through the above steps, the DTW algorithm can effectively solve the problem of feature sequence length mismatch caused by differences in movement rhythms of different users. After testing, the recognition accuracy of the integrated model after alignment increased by 4.2% (from 89.3% to 93.5%), especially for patients with movement disorders. The intention recognition effect is significantly improved. Dynamic cascade data preprocessing breaks through the bottleneck of the signal-to-noise ratio of EEG signals. Compared with traditional preprocessing methods, adaptive filtering and wavelet transform are mostly executed independently and serially, and the dynamic changes of noise characteristics are not considered. The preprocessing mechanism of the present invention analyzes the noise type in the signal (such as the proportion of power frequency interference, electromyographic noise, and electrooculographic artifacts) in real time through the noise feature monitoring module, and automatically adjusts the weight distribution of the two-stage processing. When the proportion of power frequency interference exceeds 30%, the step size factor of the adaptive filter is enhanced. q (The value range is dynamically adjusted from 0.01 to 0.1). Using an adaptive filtering formula, the filter convergence speed is accelerated, increasing the 50Hz interference rejection ratio to over 45dB. When the proportion of electrooculogram artifacts exceeds 20%, the number of wavelet transform decomposition layers is optimized (dynamically increased from 5 to 7). A modal entropy threshold, based on the db4 wavelet basis function, is introduced to specifically suppress electrooculogram signals in the 3-8Hz frequency band, increasing the artifact removal rate from 92% to 96%. This approach dynamically matches noise type with processing intensity, addressing the limited adaptability of a single preprocessing algorithm in complex scenarios. It provides high-purity signals (stable signal-to-noise ratio of 35dB ± 2dB) for subsequent feature extraction, making it particularly suitable for high-noise environments such as hospitals and homes. A heterogeneous ensemble learning model achieves super-resolution recognition of motion intent. Existing ensemble learning methods often fuse homogeneous models. This invention constructs a CNN-SVM-RF heterogeneous ensemble learning model, improving recognition accuracy through feature complementarity.

[0067] CNN-SVM-RF heterogeneous integrated learning model and voting mechanism adopts a three-layer architecture of "parallel input-feature splitting-result fusion" (such as Figure 5 As shown in the figure, the input layer receives a 128-dimensional feature vector (including 32 dimensions in the time domain + 64 dimensions in the frequency domain + 32 dimensions in the spatial domain) and transmits it to the three sub-models of CNN, SVM, and RF at the same time; the sub-models of the processing layer calculate independently, CNN focuses on capturing the local spatiotemporal correlation of features, SVM is good at dividing the boundaries of high-dimensional features, and RF focuses on global statistical laws; the fusion layer integrates the output probabilities of the three sub-models through a weighted voting mechanism to generate the final motion intention. The feature vector of the data interaction interface adopts a standardized format (float32 type, range [-1,1]), and data transmission between sub-models is achieved through a shared memory buffer, with a transmission delay of ≤5ms; the output of each sub-model is unified into a probability distribution vector of 6 categories of motion intentions ( p 1, p 2,..., p 6), where and . The CNN model execution includes feature reconstruction, that is, reshaping the 128-dimensional vector into a two-dimensional feature matrix of 16×8×1 (simulating the spatial distribution of EEG signals) as CNN input; feature extraction, that is, convolution layer 1 (3×3 kernels×32) performs sliding convolution on the input matrix and outputs a 16×8×32 feature map, with the activation function ReLU; pooling layer 1 (2×2 maximum pooling) downsamples to 8×4×32 feature maps; convolution layer 2 (3×3 kernels×64) outputs 8×4×64 feature maps, with the activation function ReLU; pooling layer 2 (2×2 maximum pooling) downsamples to 4×2×64 feature maps; classification output, that is, flattening to a 512-dimensional vector and compressing it through the fully connected layer (512-64-6); and outputting the probability distribution through the softmax function. , the inference takes about 150ms.

[0068] The SVM model execution includes kernel function mapping, which uses the RBF kernel function to map the 128-dimensional features to a high-dimensional space; multi-classification calculation, which constructs 15 binary SVM sub-classifiers (six types of intent are combined in pairs); each sub-classifier outputs a category judgment (+1 / -1), and the category with the highest cumulative votes is assigned the highest probability; probability calibration, which converts the decision value into a probability distribution through Platt scaling , inference takes about 220ms. RF model execution includes feature sampling, i.e. random sampling of 128-dimensional features (each sampling dimensional), generating an independent feature set for each decision tree; decision tree voting, that is, 100 CART trees are calculated in parallel, and each tree is split according to the minimum Gini impurity principle; a single tree outputs a category prediction (1-6), and the category vote rate of the 100 trees is counted; probability generation, that is, the vote rate is normalized to generate a probability distribution , the inference takes about 200ms.

[0069] The weighted voting mechanism includes dynamic weight calculation, that is, the performance evaluation of the sub-models, and the accuracy of each model on the validation set is updated every 500 samples. acc CNN , acc SVM , acc RF ; Then perform weight normalization and calculate the weight coefficient: , ; In actual test, it is stable w CNN =0.35, w SVM =0.3, w RF =0.35; where w CNN Represents the weight coefficient of the CNN sub-model, reflecting the importance of the model in the fusion decision. The larger the weight, the stronger the impact on the final result; acc CNN The accuracy of the convolutional neural network sub-model on the validation set reflects the recognition accuracy of the model for motion intention, ranging from [0, 1]. acc SVM represents the accuracy of the support vector machine sub-model on the validation set; acc RF Indicates the accuracy of the random forest sub-model on the validation set; w SVM Represents the weight coefficient of the SVM sub-model; w RF Represents the weight coefficient of the RF sub-model; the probability fusion calculation is performed through weighted summation, and the probabilities of the six types of intentions are weighted and accumulated respectively. The weighted accumulation expression is: ;in, p i final Indicates the i The final fusion probability of the class motion intention is used as the basis for multi-classification decision-making. The larger its value is, the higher the credibility of the class intention is. p i CNN Indicates that the CNN sub-model predicts i The probability of motion intention, ranging from [0,1]; p i SVM Indicates that the SVM sub-model predicts i The probability of motion intention, ranging from [0,1]; p i RF Indicates that the RF sub-model predicts i The probability of motion intention is in the range of [0,1];i =1,2,...,6 (corresponding to 6 types of motion intentions); the result is to select the category with the highest probability as the final output; intention = argmax i { p i final If the maximum probability , triggering secondary recognition (increasing sample collection time to 3s); among them, argmax i Represents the index of motion intention category i Take the maximum value operation, that is, traverse the final fusion probability corresponding to all 6 types of intents and select the category index with the largest probability i , and use the intent corresponding to the index as the final output The conflict resolution mechanism is to start confidence backtracking when the highest probability categories of the three sub-models are all different (for example, CNN selects "lifting legs", SVM selects "bending", and RF selects "standing still"): calculate the difference between the second highest probability category and the highest probability of each model Δp = p 1- p 2. Priority adoption Δp The largest model result (highest confidence), for example, CNN Δp =0.82-0.1=0.72; SVM Δp =0.6-0.3=0.3; RF Δp =0.55-0.2=0.35 Finally, the CNN "leg raising" result is adopted. For example, the output of the integrated recognition sub-model of the "fist clenching" intention is CNN's P CNN =[0.02,0.01,0.01,0.92,0.03,0.01] (the fourth category “clenched fist” has the highest probability); SVM P SVM =[0.03,0.02,0.02,0.85,0.05,0.03] (the fourth category “clenched fist” has the highest probability); RF P RF =[0.01,0.01,0.02,0.88,0.06,0.02] (the probability of the fourth category "clenched fist" is the highest). Weighted fusion is p 4 final =0.35×0.92+0.3×0.85+0.35×0.88=0.322+0.255+0.308=0.885 Other categories p i final<0.1, and finally outputs the intention of "clenching a fist". Performance comparison: The accuracy of a single model in recognizing "clenching a fist" is 92.1%, the accuracy of SVM is 89.3%, and the accuracy of RF is 91.5%; the accuracy of the integrated model is 96.1%, and the error rate is reduced by 43%-57%, reflecting the advantages of heterogeneous integration. Through the above steps, the CNN-SVM-RF heterogeneous integrated learning model realizes the complementary advantages of different learning mechanisms, and the weighted voting mechanism ensures the reliability of the results. After verification by 12,000 test samples, the average recognition accuracy of the integrated model reached 93.5%, and its robustness to noise interference is significantly better than that of a single model. The CNN branch adopts a 3-layer convolution + 2-layer residual connection structure, and 3×3 convolution kernels and 1×1 convolution kernels are used alternately to specifically capture the local temporal characteristics of EEG signals (such as m The output feature map is pooled globally to obtain a 256-dimensional feature vector. The SVM branch uses a radial basis kernel function (γ1 is dynamically adjusted to 0.1-1) to perform nonlinear classification on the spatial features extracted by CSP, solving the boundary ambiguity problem under small sample sizes. The 100 decision trees in the RF branch use a random feature subspace (each node randomly selects 20% of the features) to focus on learning the statistical characteristics of the signal (such as variance and peak frequency) and suppress outlier interference. The voting mechanism introduces confidence weighting, the formula is: ( c i is the classifier confidence, ranging from 0.8 to 1.0), to prevent low confidence results from affecting the final decision.

[0070] The three models cover temporal, spatial, and statistical feature dimensions, respectively. The recognition accuracy for six basic motor intentions has increased to 95.8%. Even for spinal cord injury patients with weak signal strength (EEG amplitude <50μV), the recognition accuracy remains above 90%, surpassing the feature coverage limitations of traditional single models. Dual closed-loop feedback, integrating reinforcement learning with impedance control, enables human-machine collaborative and compliant control.

[0071] The present invention can construct a dual closed-loop feedback mechanism including an inner loop (angle closed loop), which is to collect joint angles through IMU i real , and the target angle i cmd Deviation Dth = i cmd - i real Input reinforcement learning module, using state value function, when Dth When >5°, dynamically adjust the reward value r (positive reward is inversely proportional to the square of the deviation), which increases the angle convergence speed by 40%; the outer loop (force closed loop), which is the human-computer interaction force collected by the force sensor Fand expectation F 0 (Dynamic setting based on user weight, for example, if the weight is 50kg F 0 =10N) deviation ΔF , through the impedance control formula ; ( K 2 is the stiffness coefficient, B is the damping coefficient), where Indicates the target joint angular velocity, that is, the target angle i cmd The rate of change over time, that is, the desired angular velocity of the joint, where i represents the joint angle; Indicates the actual joint angular velocity, that is, the actual rotational angular velocity of the joint collected by sensors such as IMU (Inertial Measurement Unit), which is a measurement of the actual joint movement speed; adjusts the joint torque in real time t , so that the interaction force fluctuation is controlled within ±2N. It solves the contradiction of "precise angle but large force impact" in traditional control. In the rehabilitation training of patients with spinal cord injury, it can not only ensure the accuracy of the motion trajectory (deviation ≤ 2°), but also avoid secondary injuries caused by excessive interaction force, and achieve the dual control goals of "precision + safety". User-adaptive feature alignment network to achieve rapid cross-user adaptation. In view of the differences in EEG signal characteristics of different users (such as the motor cortex activation position offset of 1-3cm), an adaptive network based on dynamic time warping (DTW) and feature transfer learning is designed. In the offline stage, a multi-user feature library is constructed (including EEG features of 200+ users of different age groups and injury levels). The DTW algorithm is used to align the feature sequences of similar motion intentions on the time axis to generate standardized feature templates; the DTW algorithm is used to generate standardized feature templates for similar motion intentions. The screening and preprocessing of samples of similar motion intentions include sample set construction, that is, for a single motion intention (such as "lifting the leg"), feature sequences of N=50 healthy subjects are collected. Three groups of valid samples are collected from each person at an interval of 24 hours, for a total of 150 original sequences (denoted as X 1, X 2,..., X 150 Each sequence X k =[ x k,1 ,x k,2 ,..., x k,mk ],in m k is the sequence length (due to individual movement speed differences, , corresponding to a sampling time of 0.8-2.0s). is a 128-dimensional feature vector.

[0072] Abnormal sample removal is to calculate the kinematic characteristics of each sequence, specifically the average angular velocity ( Δt =4ms is the sampling interval), characteristic variance ;in, Indicates the k In the feature sequence, t time step, d The original eigenvalues ​​of feature dimensions; i max Indicates the maximum angle; i min Indicates the minimum angle; culling Exceeds [20° / s, 50° / s] or s k 2 For samples that exceed the mean ± 2 times the standard deviation, 120 valid sequences are retained. Feature normalization is to perform global normalization on all retained sequences: ;in, Indicates the k In the feature sequence, t time step, d The feature value of each feature dimension after global standardization; m d 、 s d For the d The total mean and standard deviation of the dimensional features in 120 sequences; that is, m d Indicates the d The total mean of the dimension feature in all the 120 valid sequences retained is used for feature standardization to eliminate the scale differences of the dimension feature between different sequences; s d Indicates the d The total standard deviation and total mean of the dimension features in all the 120 valid sequences retained m d In coordination, the feature normalization is completed, that is, the feature is mapped to a scale with a mean of 0 and a standard deviation of 1, ensuring that the feature scale is consistent across samples. The reference sequence in the DTW-based multi-sequence hierarchical alignment is selected to calculate the average DTW distance of each sequence with all other sequences: ;in, l Represents the sequence index, used to traverse all sequences except the k All other sequences except the first sequence to calculate the k The average DTW distance between the sequence and other sequences; Indicates the lThe feature sequence is one of the comparison objects when calculating the average DTW distance. k bar sequence X k Perform DTW distance calculation; select The smallest sequence is used as the initial reference sequence X ref (i.e., the overall similarity with similar sequences is the highest), and its length is m ref Among them, the first-level alignment is to align all sequences with the reference sequence, and for each sequence X k , execute the DTW algorithm with X ref Alignment (refer to the above DTW implementation steps), by calculating the local distance matrix ;in, Indicates the k Among the sequences to be aligned, i The normalized feature vector of the time step contains multiple feature dimensions and is the feature representation after the initial normalization process; Represents the initial reference sequence X ref In j The normalized feature vector of the time step is used to calculate the local distance with the feature vector of the sequence to be aligned to achieve dynamic time warping; the cumulative distance matrix is ​​constructed C , set the Sakoe-Chiba bandwidth α =0.15; backtrack to get the optimal path P k =[( i 1, j 1),( i 2, j 2),...,( i L , j L )];Will X k Regularized to X ref Sequences of equal length (length m ref ),in ;in, Indicates the k The sequences were aligned to the initial reference sequence. X ref After alignment and regularization, j The feature mean of time steps, in addition, for mapping to the same j All iThe corresponding eigenvalues ​​are averaged to normalize the sequence to be aligned to the same length as the reference sequence; the secondary alignment includes the consistency optimization of the aligned sequence, that is, calculating the point-by-point mean of the 120 normalized sequences. ,in ;by For the new reference sequence, repeat the first-level alignment step and Reorganize into , eliminate the accumulated error of the first alignment; calculate the sequence consistency after the second alignment: ;in, Indicates the k The sequence is aligned to the second level, i.e. the point-wise mean sequence After re-aligning the new reference sequence, j time step, d The eigenvalue of the feature dimension; when Consistency The iteration stops when the value is ≥0.85, otherwise repeat the secondary alignment (2 iterations are sufficient in actual testing). The standardized feature template is generated for the 120 sequences after secondary alignment. , calculate the point-by-point mean as the normalization template ,in, m T = m ref (Uniform length, such as 300 frames corresponding to 1.2s); calculate the point-by-point standard deviation of the template ,in, Indicates the k In the sequence after secondary alignment, j The feature vector of each time step is the feature representation of the sequence at the corresponding time step after completing the secondary alignment operation; Represents a standardized feature template T In j The point-by-point mean of the time step is the 120 sequences after secondary alignment. j The average of the characteristic values ​​of the time steps is used as the typical characteristic representation of the standardized template at this time step; the confidence interval of the template (±2 s j The key time points are marked by combining kinematic data (joint angles) and marking three key frames in the template T, specifically the starting frame. j s , that is, the difference between the feature vector and the rest state exceeds the threshold for the first time ( );in, Indicates the start frame j sThe corresponding eigenvector is used to determine whether the difference between the eigenvector and the static state exceeds the threshold for the first time, that is, by calculating the difference between the eigenvector and the static state eigenvector t rest The norm of is used to measure the starting time of the action; Represents the eigenvector in the static state, which is used to determine whether the eigenvector has returned to the vicinity of the static state, that is, by calculating the eigenvector and t rest The norm of the peak frame is used to measure the difference; j p , which are the feature dimensions most relevant to movement intention (e.g. β The wave power) reaches the maximum value; the frame is terminated j e , that is, the eigenvector returns to the vicinity of the stationary state ;in, Indicates the termination frame j e The corresponding feature vector is the feature vector at the end frame when the feature vector returns to the static state. For example, in the "leg lift" template, j s =50 (action started), j p =150 (maximum hip angle), j e =250 (action completed). The test set is verified by using 20 new subjects’ similar intention sequences to perform DTW matching with the template and calculating the average similarity , where Sim is required to be ≥ 0.75; cross-category discrimination, that is, the average distance between the template and the other five categories of intent templates, must be greater than twice the distance between the same sequence and the template (to ensure category uniqueness).

[0073] The generation process of the standardized template for the “leg lift” intention includes the following: the initial data is 150 “leg lift” feature sequences with a length of 220-480 frames, 120 of which are retained after screening; the reference sequence is the sequence with the smallest average DTW distance. X ref (length 300 frames); the first-level alignment is 120 sequences are all regularized into 300 frames, the consistency index Consistency =0.79; the secondary alignment is to use the mean sequence as the new reference, and then align again Consistency =0.88; the template is generated as T is the mean sequence of 300 frames, key frame j s =48, j p =145, j e=242; the verification result is the average similarity between the new sample and the template Sim =0.81, and the distance from the "bending" template is 2.3 times the distance of the same type, meeting the effectiveness requirements. The standardized feature template generated by the above steps solves the problem of sequence inconsistency caused by individual differences in the same type of movement intentions, and provides a reliable benchmark for subsequent DTW real-time matching. After testing, the accuracy of intention recognition based on this template increased by 5.3%, especially the recognition stability of patients with movement disorders was significantly improved (the fluctuation range was reduced from ±8% to ±3%). During the online adaptation stage, that is, after the new user wears the device, 3 minutes of motor imagery signals are collected. Through feature distance measurement (calculating the Euclidean distance between the new feature and the template), the three most similar templates are automatically matched from the library, and weighted transfer learning (weight is inversely proportional to the distance) is used to initialize the parameters of the ensemble learning model. During the adaptation process, 10% of the model parameters are updated every 5 minutes through incremental learning to avoid catastrophic forgetting.

[0074] The Transformer-based multimodal fusion attention mechanism, that is, based on the original heterogeneous integrated learning model, introduces the Transformer architecture to build a multimodal fusion attention mechanism to achieve deep association between EEG signals and auxiliary modal data. The data input and output of the Transformer architecture to build a multimodal fusion attention mechanism include the preprocessing of the multimodal data input layer, that is, the data modality division and feature dimension unification. The input data contains three types of modalities, specifically, EEG feature sequence E 128 dimensions / frame, duration 2s (500 frames), from preprocessed 8-30Hz frequency band signal; electromyographic feature sequence M 64 dimensions / frame, duration 2s (500 frames), extracted from surface electromyographic signals of 6 muscles including quadriceps femoris; motion posture sequence P 32 dimensions / frame, duration 2s (500 frames), composed of acceleration and angular velocity data collected by IMU; standardization of myoelectric and posture sequences (same as EEG features) m 、 s Parameters), ensuring that the scales of the three modal features are consistent. Modal embedding and position encoding involves adding a dedicated embedding vector for each modality: E embed ( i )= E ( i )+ e eeg , M embed ( i )= M ( i )+ e emg , P embed (i )= P ( i )+ e pose ;in, E embed ( i ) indicates that after the EEG feature sequence is modally embedded, i The feature vector of the time step is the original EEG feature vector E ( i ), an EEG-specific embedding vector is added e eeg The obtained ones are used for subsequent fusion and processing of multimodal features; E ( i ) indicates that in the EEG feature sequence, i The original feature vector of time steps, that is, the EEG feature without modality embedding operation; M embed ( i ) indicates that after the myoelectric feature sequence is modally embedded, i The feature vector of the time step is the original electromyographic feature vector M ( i ) based on the EMG-specific embedding vector e emg obtained; M ( i ) indicates that in the electromyographic feature sequence, i The original feature vector of the time step, that is, the electromyographic feature without modal embedding operation; P embed ( i ) indicates that after the motion posture sequence is modally embedded, i The feature vector of the time step is the feature vector of the original motion posture P ( i ) based on the motion posture-specific embedding vector e pose obtained; P ( i ) indicates that in the motion posture sequence, i The original feature vector of the time step, that is, the motion posture feature without modal embedding operation; e eeg Represents a unique learnable embedding vector of EEG modality, with dimension R 128 , used to add modality-specific embedded information to EEG features, facilitating the subsequent differentiation and fusion of multimodal features; e emg Dedicated learnable embedding vector representing the electromyographic modality, with dimension R 128, which is used to add myoelectric modality-specific embedded information to the myoelectric features; e pose A unique learnable embedding vector representing the motion posture modality, with dimension R 128 , used to add motion posture modality-specific embedded information to the motion posture feature; wherein, For learnable modal embedding (initially randomly generated and optimized through training); add sinusoidal position encoding to capture timing information: , ,in, d =128 is the feature dimension, pos is the frame index, k =0,1,...,63, and finally we get the input sequence of three types of modes The cross-modal attention calculation of the Transformer multimodal fusion attention layer includes constructing query (Q), key (K), and value (V) matrices; using the EEG sequence as the query source and the EMG and posture sequences as the key-value sources, ;in ) is the projection matrix, compressing the 128-dimensional features to 64 dimensions, [ X m ; X p ] represents the concatenation of EMG and posture sequences in the frame dimension (1000×128). Calculate the attention weight ;in M is a mask matrix (to prevent future frame information leakage), the diagonal and lower triangle are 0, and the upper triangle is -∞; output EEG-multimodal fusion features F eeg =Linear( Attention ( Q , K , V ))+ X e (Residual connection + LayerNorm); the dimension remains 500×128. Intra-modal self-attention is strengthened, that is, the fused EEG features are F eeg Perform self-attention calculation: , F eeg '=Linear( Attention ( Q ', K ', V '))+ F eeg ;in, Q ' represents the query matrix in the intra-modal self-attention calculation, which is composed of the fused EEG features F eeg With the projection matrixW q 'Multiplied together, used to query relevant information in the self-attention mechanism; K ' represents the key matrix in the self-attention calculation within the modality, which is also composed of the fused EEG features F eeg With the projection matrix W q 'Multiplying together, here, K '= Q ', is a common setting in the self-attention mechanism where the key and query are calculated in the same way, and is used to match the query matrix to calculate the attention weight; V 'Represents the value matrix in the self-attention calculation within the modality, which is also composed of the fused EEG features F eeg With the projection matrix W q 'The multiplication results in the feature matrix used for weighted summation in the attention mechanism; W q ' represents the projection matrix in the intra-modal self-attention calculation, which is used to transform the fused EEG features F eeg Project to a specific dimension to generate query, key, and value matrices to participate in the self-attention calculation process; F eeg ' represents the EEG feature after the self-attention enhancement within the modality. It is achieved by linearly transforming the self-attention calculation result (Linear operation), combining it with the residual connection, and adding the original F eeg The layer normalization (LayerNorm) is performed to strengthen the EEG feature representation after the attention association within the feature; enhance the temporal correlation within the EEG sequence and highlight the feature peaks related to movement intention (such as m The interaction with the original heterogeneous integration model includes feature splitting and sub-model adaptation, that is, the fusion feature of Transformer output F eeg 'Processing, specifically taking the last frame feature (128 dimensions) and inputting it into the SVM and RF models (maintaining compatibility with the original architecture); retaining the complete sequence (500×128) and inputting it into the improved CNN model (with a new temporal convolution layer); temporal adaptation of the CNN model, i.e. adding a 1D convolution layer (kernel size 3, output channels 32) before the original structure, converting the 500×128 sequence into 500×32, and then obtaining a 128-dimensional feature vector through global average pooling. Attention weight-guided sub-model weighting, i.e. extracting the weight matrix of the Transformer cross-modal attention layer The attention degree of EEG frame to EMG / gesture frame); Calculate the modal confidence: , c emg + pose )=1- c eeg ,in, pose Represents the motion posture (Pose) modality, which is one of the three types of input data modalities, namely EEG, EMG, and motion posture. It is used to describe motion-related posture information and consists of acceleration, angular velocity and other data collected by the IMU; dynamically adjust the sub-model weights: ;in, w CNN ' represents the weight of the convolutional neural network (CNN) sub-model after dynamic adjustment, which is the weight of the original CNN sub-model w CNN Based on the EEG modality confidence c eeg The dynamic adjustment is used to weight the output of the CNN sub-model in the subsequent multi-model probability fusion; w SVM ' represents the weight of the support vector machine (SVM) sub-model after dynamic adjustment, which is the weight of the original SVM sub-model w SVM Based on this, the confidence of electromyography and motion posture modality is combined c emg + pose as well as w RF 'The relevant balance items are dynamically adjusted and used to weight the output of the SVM sub-model during multi-model probability fusion; w RF ' represents the weight of the random forest (RF) sub-model. In the process of dynamically adjusting the weight of the sub-model, w RF 'Remain unchanged as a balancing term to ensure that the sum of all sub-model weights is 1, so as to reasonably integrate the outputs of each sub-model; remain unchanged (as a balancing term) to ensure that the sum of the weights is 1. The multi-model probability fusion of the output layer includes the probability distribution of the sub-model outputs: ; Weighted fusion formula: ; Decision result output and verification; The final intention is argmax( P final ), and output the modal consistency score: ;when Score When the value is ≥0.85, the result is directly output; otherwise, multimodal re-acquisition is triggered (extended to 3s) to ensure decision reliability.

[0075] The multimodal fusion process of the “leg lift” intention includes the input data, namely the EEG sequence E 500 frames, 128 dimensions (including m wave suppression characteristics); EMG sequence M 500 frames, 64 dimensions (quadriceps activation features are obvious); posture sequence P 500 frames, 32 dimensions (increasing hip joint angle). Transformer processing for cross-modal attention shows the attention paid by EEG frames 150-250 (corresponding to the middle of the leg lift) to EMG frames 180-280. α =0.8 (highly correlated); fusion features F eeg 'Highlight the feature peaks of frames 150-250. The sub-model output is P CNN =[0.02,0.93,0.01,0.01,0.02,0.01] (Class 2 “Leg lift”); P SVM =[0.03,0.89,0.02,0.02,0.03,0.01]; P RF =[0.01,0.91,0.01,0.02,0.03,0.02]. The weights and fusion are c eeg =0.7, so w CNN '=0.35×(0.5+0.35)=0.35×0.85=0.2975, w SVM '=0.3×(0.5+0.15)=0.195, w RF '=0.35; P final(2)=0.2975×0.93+0.195×0.89+0.35×0.91≈0.276+0.174+0.319=0.769, outputting the intention of "lifting the leg". By introducing the multimodal fusion attention mechanism of Transformer, the model's adaptability to complex scenarios (such as noise interference and action variations) is significantly improved. Compared with the original heterogeneous integration model, the average recognition accuracy is increased from 93.5% to 96.2%, and the cross-user generalization error is reduced by 40%, fully demonstrating the technical advantages of multimodal information complementarity. In addition to the 256-dimensional feature vector extracted by CSP from the EEG signal, two types of auxiliary modal data are integrated into the modal input: the 16-channel muscle activity signal collected by the electromyography sensor is decomposed into 64-dimensional features by wavelet packet decomposition; the 3-axis acceleration and 3-axis angular velocity recorded by the inertial measurement unit (IMU) are statistically analyzed by sliding windows to obtain 32-dimensional kinematic features. The three modal data are mapped to a 128-dimensional vector space through a linear projection layer to form a modality-specific feature sequence (the length of the EEG sequence is L1=30, the EMG sequence is L2=20, and the kinematic sequence is L3=15). The intra-modality self-attention mechanism is that each modality branch is equipped with an independent self-attention layer to calculate the dependencies within the feature sequence; the EEG branch, that is, for m Wave, β The temporal correlation of the wave is determined by using causal masked self-attention to ensure that only historical signals are focused on at the current moment. The attention weight calculation formula is: ;in M is the causal mask matrix (the lower triangle is 0, the upper triangle is -∞), d k =128 is the feature dimension. The myoelectric and kinematics branch uses unmasked self-attention to capture global feature associations, such as the synchronization between muscle activity intensity and joint movement speed. The cross-modal cross-attention mechanism specifically constructs three groups of cross-attention layers to achieve information interaction between modalities; using EEG features as the query ( Q ), the myoelectric characteristics are key ( K ) and value ( V ), learn the mapping relationship between “brain wave intention and muscle response”; take kinematic features as Q , the EEG characteristics are K / V , strengthen the timing alignment of "movement state-EEG command"; design bidirectional cross attention to make electromyography and kinematic features mutually Q / K / V, to make up for the single modality delay defect. The cross-attention output is layer-normalized (LayerNorm) and connected with the residual, retaining the original features while incorporating cross-modal information. The global fusion attention layer aggregates the cross-attention outputs of the three modalities to form a mixed feature sequence of length 65 (30+20+15), which is input into the global self-attention layer; a multi-head attention mechanism (number of heads) is used. h =8), each head independently calculates the attention distribution of different subspaces, the formula is: MultiHead ( Q , K , V )= Concat ( head 1,..., head h ) W O in head i = Attention ( QW i Q , KW i K , VW i V ), W i Q , W i K , W i V , W Ois a learnable parameter matrix. Attention weights are visualized to dynamically adjust the contribution of each modality. For example, EEG weighting is 60% when movement intent is clear, while myoelectrical weighting is increased to 50% during movement execution. A classification head design uses a feedforward neural network (FFN) to map the [CLS] tag vector output by global attention to a six-category movement intent space. The FFN consists of two linear layers (with an intermediate dimension of 512 and a GELU activation function) and a dropout layer (with a probability of 0.3) to mitigate overfitting. Comparative experiments show that incorporating multimodal fusion improves the recognition accuracy of complex intentions such as "fine grasp" from 89.2% to 94.7%, and improves the tolerance to signal interruptions (<200ms) by 30%. The multimodal fusion experiment tested 20 healthy adults (10 men and 10 women, aged 22-45 years, with no history of movement disorders) and 10 patients with mild hand dysfunction (post-stroke sequelae, who can perform simple grasping but have poor fine control). The experimental equipment included EEG acquisition, specifically a 64-channel wet-electrode EEG (500Hz sampling rate, 0.1-500Hz bandwidth); EMG acquisition, specifically an 8-channel surface EMG sensor (1000Hz sampling rate, attached to the forearm flexor and extensor muscles); posture tracking, specifically an inertial measurement unit (IMU, 200Hz sampling rate, measuring the three-dimensional acceleration and angular velocity of the hand); and motion capture, specifically an optical motion capture system (0.1mm accuracy, recording finger joint angles). The test intent categories included complex intentions such as fine grasping (pinching a 5mm diameter steel ball), multi-finger coordination (unscrewing a bottle cap), and force-controlled grasping (picking an egg); simple intentions included clenching a fist, extending a palm, and remaining still (as a control). Data collection for the specific experiment involved three sets of experiments per subject, each consisting of 10 repetitions of each of the six intent categories, separated by a one-minute interval. Each intent category lasted 3 seconds, with the first 2 seconds representing the action execution period and the last 1 second representing the resting recovery period. Signal interruptions were simulated by randomly inserting 200ms of EEG signal interruptions (replaced with baseline noise) in 10% of the samples, simulating signal loss in real-world use. The model testing process included a control group using the original heterogeneous integration model (CNN+SVM+RF) based solely on EEG features, while the experimental group introduced a Transformer multimodal fusion model (EEG+EMG+gesture). The test set was divided into "external-subject groups" (80% training and 20% testing) to ensure model generalization to new users. Performance evaluation metrics included accuracy (number of correctly recognized samples divided by total number of samples); error tolerance (accuracy during signal interruption / accuracy during normal signal × 100%); and average recognition latency (the time from signal input to intent output, required to be ≤300ms). Comparative experimental results are shown in Tables 3 and 4.

[0076] Table 3 Comparison of complex intent recognition accuracy ; As shown in Table 3, the recognition of complex intentions is more significantly improved because multimodal fusion makes up for the ambiguity of EEG in representing fine movements (for example, electromyography can provide details of finger force, and posture can reflect the spatial trajectory of the hand).

[0077] Table 4 Comparison of signal interruption tolerance ; As shown in Table 4, when the signal is interrupted, the experimental group can temporarily replace EEG with EMG and posture features to make decisions (for example, the EMG peak pattern during grasping has strong recognition), while the control group's accuracy rate drops sharply due to its sole reliance on EEG. In the typical sample comparison example, the control group had 5 error cases due to EEG. β The wave (13-30Hz) features were not distinct and were misidentified as "force-controlled grasps." The experimental group's correction mechanism correctly identified all five samples using the EMG signal "abductor digiti minimi-flexor pollicis brevis coordination pattern" (feature vector cosine similarity greater than 0.85). In the control group, seven "multi-finger coordination" samples were misidentified as "fists" when the signal was interrupted for 200ms. The experimental group retained the correct identification of six of these samples using the gesture sensor's "wrist pronation angle greater than 30°" feature.

[0078] In the delay performance comparison, the average delay of the control group was 245ms; the average delay of the experimental group was 280ms (due to the increased computational complexity due to multimodal fusion), which still met the real-time control requirements (≤300ms).

[0079] In patient testing, the accuracy of "fine grasp" recognition for patients with mild impairments was 76.4% in the control group (poor EEG signal quality) and 89.7% in the experimental group (myography and posture compensated for EEG deficiencies), a 13.3% improvement, significantly exceeding the improvement seen in healthy controls. These experiments demonstrate that multimodal fusion significantly improves performance in complex intentions and signal instability scenarios (p < 0.01), validating the effectiveness of the Transformer architecture in complementing cross-modal information and avoiding the limitations of a single EEG feature. Experimental data demonstrates that this improvement significantly enhances the exoskeleton's robustness in recognizing complex motion intentions while maintaining real-time performance. Breaking through the limitations of traditional single modality or simple splicing and fusion, the Transformer's attention mechanism automatically mines the implicit correlation between EEG, EMG, and kinematic data. That is, when the EEG signal quality is poor (such as signal-to-noise ratio <25dB), the complementary information of EMG and kinematic modalities is used to maintain a recognition accuracy of more than 90%. Cross-attention is used to solve the delay difference in multimodal data acquisition (the EEG leads the EMG by about 150ms), reducing the intention-action response synchronization error to ±50ms. It provides richer decision-making basis for the "intention-execution-feedback" closed loop in rehabilitation training. For example, the execution strength of EEG intention can be corrected by the intensity of muscle activity to avoid excessive movement.

[0080] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 3 As shown. The computer device includes a processor, a memory and a network interface connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store static information and dynamic information data. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the steps in the above-mentioned embodiment of the exoskeleton robot control method based on the brain-computer interface are implemented.

[0081] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention and does not constitute a limitation on the computer device to which the solution of the present invention is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0082] In addition, the present invention also provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps in the above-mentioned embodiment of the exoskeleton robot control method based on the brain-computer interface are implemented.

[0083] In addition, the present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned embodiment of the exoskeleton robot control method based on the brain-computer interface are implemented.

[0084] Those skilled in the art will appreciate that all or part of the processes in the above-described brain-computer interface-based exoskeleton robot control method can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described brain-computer interface-based exoskeleton robot control method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include at least non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0085] The present invention is not limited to the structures described above and shown in the drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. An exoskeleton robot control method based on brain-computer interface, characterized in that: include: Acquiring brain-computer interface data, preprocessing the brain-computer interface data to obtain preprocessed brain-computer interface data, and extracting motion features using the preprocessed brain-computer interface data to obtain a motion feature vector; Using an integrated learning model to identify motion intentions from motion feature vectors, generating motion control instructions based on the motion intention identification results, and performing motion control according to the motion control instructions; Based on the acquired motion state feedback data, the motion state is adjusted using a dynamic adjustment algorithm, and the motion control instructions are dynamically adjusted according to the motion state adjustment results, and motion control optimization is performed through the adjusted motion control instructions.

2. The exoskeleton robot control method based on brain-computer interface according to claim 1 is characterized in that: The acquiring of brain-computer interface data, preprocessing of the brain-computer interface data to obtain preprocessed brain-computer interface data, and extraction of motion features using the preprocessed brain-computer interface data to obtain a motion feature vector include: Acquire brain-computer interface data, use adaptive filtering to remove power frequency interference and myoelectric noise from the brain-computer interface data, and obtain denoised brain-computer interface data; The denoised brain-computer interface data is decomposed using wavelet transform, and the decomposed brain-computer interface data is combined with the preset frequency band to perform data screening to obtain preprocessed brain-computer interface data; Based on the common space pattern algorithm, the motion features of the preprocessed brain-computer interface data are extracted to maximize the difference in motion intentions and obtain the motion feature vector.

3. The exoskeleton robot control method based on brain-computer interface according to claim 2 is characterized in that: The motion feature extraction based on the common space pattern algorithm for maximizing the difference in motion intentions on the pre-processed brain-computer interface data to obtain the motion feature vector includes: The covariance of the movement intention is calculated based on the preprocessed brain-computer interface data to obtain the movement intention covariance matrix, and the spatial filter matrix is ​​solved for the movement intention covariance matrix to obtain the movement intention spatial filter matrix; The motion intention spatial filter matrix is ​​optimized using the pre-built objective function, and the variance difference of different motion intention signals is maximized according to the optimization result of the motion intention spatial filter matrix to obtain the variance difference results of different motion intention signals; Motion features are extracted based on the variance difference results of different motion intention signals, and a motion feature vector is generated based on the motion feature extraction results.

4. The exoskeleton robot control method based on brain-computer interface according to claim 1, characterized in that: The method of using an integrated learning model to identify motion intentions of motion feature vectors, generating motion control instructions based on the motion intention identification results, and performing motion control according to the motion control instructions includes: Construct an integrated learning model and use the pre-acquired brain-computer interface data to train the integrated learning model to obtain a movement intention recognition model; The motion intention recognition model is combined with the preset motion intention category to identify the motion intention of the motion feature vector and obtain the motion intention recognition result; The motion intention recognition result is subjected to motion intention confidence verification, and the motion parameters are calculated based on the motion intention confidence verification result combined with the preset motion parameter mapping rules and the pre-acquired human motion data, and motion control instructions are generated according to the motion parameters, and motion control is performed using the motion control instructions.

5. The exoskeleton robot control method based on brain-computer interface according to claim 4 is characterized in that: The construction of the integrated learning model and the use of pre-acquired brain-computer interface data to train the integrated learning model to obtain a movement intention recognition model include: Perform feature extraction on the pre-acquired brain-computer interface data to obtain a brain-computer interface feature dataset, and divide the brain-computer interface feature dataset into a training set, a validation set, and a test set based on a preset ratio; The convolutional neural network model, the support vector machine model and the random forest model are trained and optimized using the training set to obtain the optimized convolutional neural network model, the optimized support vector machine model and the optimized random forest model; The optimized convolutional neural network model, the optimized support vector machine model and the optimized random forest model are integrated in parallel to obtain a motion intention recognition model.

6. The exoskeleton robot control method based on brain-computer interface according to claim 5 is characterized in that: The convolutional neural network model, the support vector machine model and the random forest model are trained and optimized using the training set to obtain the optimized convolutional neural network model, the optimized support vector machine model and the optimized random forest model, including: The convolutional neural network model is initialized using the normal distribution to obtain an initial convolutional neural network model, the initial convolutional neural network model is iteratively trained according to the training set, and the convolutional neural network model is updated through the cross entropy loss function and back propagation to obtain an optimized convolutional neural network model; A support vector machine model with motion intention classification is constructed through kernel function selection and multi-classification strategy. The support vector machine model is trained based on the training set, and the model objective function is optimized by solving a convex quadratic programming problem to obtain an optimized support vector machine model. The training set is sampled with replacement to obtain a training subset, and a decision tree is constructed using the classification and regression tree algorithm. The splitting features and thresholds of the decision tree are optimized according to the Gini impurity. The random forest model is trained based on the optimized decision tree combined with the training subset to obtain an optimized random forest model.

7. The exoskeleton robot control method based on brain-computer interface according to claim 4 is characterized in that: The motion intention recognition model is combined with a preset motion intention category to perform motion intention recognition on the motion feature vector, and the motion intention recognition result obtained includes: Based on the convolution structure of the motion intention recognition model and the preset motion intention category, the motion feature vector is subjected to motion intention recognition with temporal correlation, and the motion intention classification probability with temporal correlation is obtained; The radial basis kernel function of the motion intention recognition model is used in combination with the preset motion intention category to perform motion intention recognition with statistical regularity on the motion feature vector, and obtain the motion intention classification probability with statistical regularity; The motion intention recognition model is based on a decision tree and a preset motion intention category to identify the motion intention with boundary features on the motion feature vector, and the motion intention classification probability with boundary features is obtained; The weighted voting method is used to integrate the results of motion intention classification probability with temporal correlation, motion intention classification probability with statistical regularity and motion intention classification probability with boundary features to obtain the motion intention recognition result.

8. The exoskeleton robot control method based on brain-computer interface according to claim 1, characterized in that: The method of adjusting the motion state based on the acquired motion state feedback data using a dynamic adjustment algorithm, dynamically adjusting the motion control instructions according to the motion state adjustment result, and optimizing the motion control using the adjusted motion control instructions includes: Calculating the motion deviation of the acquired motion state feedback data, and using the motion deviation calculation result in combination with a preset threshold to perform motion adjustment judgment to obtain a motion adjustment judgment result; Based on the motion adjustment judgment result, a dynamic adjustment algorithm is used to adjust the motion state parameters of the acquired motion state feedback data to obtain a motion state adjustment result; The motion control instructions are dynamically adjusted according to the motion state adjustment result, and the motion control optimization is performed through the adjusted motion control instructions.

9. The exoskeleton robot control method based on brain-computer interface according to claim 8, characterized in that: The method of adjusting the motion state parameters of the acquired motion state feedback data using a dynamic adjustment algorithm based on the motion adjustment judgment result to obtain the motion state adjustment result includes: Based on the motion adjustment judgment result, the state value function and parameters of the dynamic adjustment algorithm are initialized to obtain an initialized dynamic adjustment algorithm; The initial dynamic adjustment algorithm is combined with the obtained motion state feedback data to select actions based on the exploration probability greedy strategy to obtain the execution action; Update the motion state adjustment parameters by executing the action, obtain the motion state feedback data of the executed action, calculate the instant reward based on the motion state feedback data of the executed action, and use the instant reward calculation result to update the state value function; The updated state value function and value iteration optimization strategy are used to adjust the motion state and obtain the motion state adjustment result.

10. The exoskeleton robot control system based on brain-computer interface is characterized in that: The exoskeleton robot control system based on brain-computer interface includes: a motion feature extraction module, a motion control module and a motion control optimization module; The motion feature extraction module is used to obtain brain-computer interface data, preprocess the brain-computer interface data to obtain preprocessed brain-computer interface data, and extract motion features using the preprocessed brain-computer interface data to obtain a motion feature vector; The motion control module is used to identify motion intentions of motion feature vectors using an integrated learning model, generate motion control instructions based on the motion intention identification results, and perform motion control according to the motion control instructions; The motion control optimization module is used to adjust the motion state based on the acquired motion state feedback data using a dynamic adjustment algorithm, and dynamically adjust the motion control instructions according to the motion state adjustment result, and perform motion control optimization through the adjusted motion control instructions.

Images