A rehabilitation intention recognition feedback training method based on electroencephalogram signal analysis

Abstract

The application relates to the technical field of upper limb rehabilitation, and discloses an upper limb rehabilitation intention recognition feedback training method based on electroencephalogram signal analysis. The method integrates upper limb motor imagination electroencephalogram signals of multiple people, constructs a data set containing multiple effective electroencephalogram features of actions through quality screening, trains an initial electroencephalogram decoding model and a feature template through transfer learning, and stores clinical data of patients; an acquisition unit is deployed to screen effective signals, and a report containing time sequence and clinical information is generated through decoding; subsequently, a multi-modal feedback and exoskeleton assistance guiding training are performed, an evaluation matrix is generated, and a personalized rehabilitation path is dynamically adjusted; finally, qualified signals are screened, the model and the template parameters are updated through incremental transfer learning, and the threshold is corrected and adapted, and then stored. The application effectively improves the accuracy, adaptability, rehabilitation effect and efficiency of upper limb rehabilitation training.

Description

Technical Field

[0001] This invention relates to the field of upper limb rehabilitation technology, and more specifically, to an upper limb rehabilitation intention recognition feedback training method based on electroencephalogram (EEG) signal analysis. Background Technology

[0002] Brain-computer interface (BCI) combined with upper limb rehabilitation training technology is a novel rehabilitation technology that identifies rehabilitation intentions by analyzing the patient's motor imagination EEG signals, and then drives rehabilitation devices to assist in the execution of movements. It does not rely on peripheral neuromuscular transmission and can directly realize the linkage between "thought and action". It provides non-invasive and precise rehabilitation intervention for patients with upper limb dysfunction caused by stroke, spinal cord injury and other conditions. At the same time, it can promote neural remodeling and has become an important development direction in the field of upper limb rehabilitation.

[0003] However, existing upper limb rehabilitation training methods based on EEG signals still face several technical bottlenecks: First, the accuracy of intent recognition is insufficient. Traditional models often rely on EEG datasets from a single population for training, failing to fully consider individual differences such as patient injury type, muscle strength level, and EEG feature intensity. Furthermore, the signal quality screening mechanism is imperfect, improperly handling issues such as artifact residue and low signal-to-noise ratio, leading to significant decoding errors. Second, personalized adaptation is lacking. Rehabilitation paths are mostly fixed templates, failing to dynamically adjust based on patient clinical data and real-time training feedback, making it difficult to match the rehabilitation foundation and progress pace of different patients. Third, the feedback and training loop is incomplete. The form of neural feedback is singular, and the correlation between feedback intensity and intent execution completion and patient attention state is insufficient, making it impossible to guide patients to adjust their training state in a timely manner. Fourth, the model has poor dynamic adaptability, lacking an effective incremental update mechanism. When EEG features change with the recovery of neural function during the patient's rehabilitation process, the model cannot be optimized synchronously, limiting long-term training effectiveness.

[0004] Therefore, it is necessary to design an upper limb rehabilitation intention recognition feedback training method based on EEG signal analysis to solve the problems of low intention recognition accuracy, poor personalization adaptability, imperfect feedback loop and insufficient dynamic optimization of the model in traditional methods, so as to improve the accuracy, adaptability and long-term rehabilitation effect of upper limb rehabilitation training. Summary of the Invention

[0005] In view of this, the present invention proposes an upper limb rehabilitation intention recognition feedback training method based on EEG signal analysis, which aims to solve the problems of low intention recognition accuracy, poor personalization adaptability, imperfect feedback loop and insufficient dynamic optimization of the model in traditional methods.

[0006] This invention proposes a method for upper limb rehabilitation intention recognition and feedback training based on electroencephalogram (EEG) signal analysis, comprising: A dataset is constructed by integrating upper limb motor imagery EEG signals from different populations and screening the signal quality. The dataset contains effective EEG features corresponding to various upper limb movements. A feature extraction module and an intention classification unit are constructed. An initial EEG decoding model and feature template are trained and generated using a transfer learning algorithm. Patient clinical data are entered and the parameter adjustment threshold of the initial EEG decoding model is set. The determined initial EEG decoding model, feature template and clinical information are associated and stored in the storage unit. The signal acquisition unit is deployed to cover the corresponding scalp area of ​​the brain. After acquiring EEG signals, the signal quality is screened simultaneously. Valid signal segments that are synchronized with the motor imagination period and pass the quality inspection are input into the initial EEG decoding model to extract features and generate action intention vectors. The feature template is matched to decode information and information quality is checked. The time series data, clinical information and information quality inspection results are associated to form a decoding report. Configure a multimodal neurofeedback unit to output feedback information, deploy an exoskeleton rehabilitation device to adjust the assist force according to the decoding report to guide the execution of the movement, record the patient's feedback response behavior, extract qualified decoding data and response data and combine them with initial clinical information to construct a rehabilitation assessment matrix, set training period division standards, generate personalized rehabilitation paths and dynamically adjust parameters; Select EEG signal segments that meet the quality inspection standards and are intended to be executed in accordance with the requirements. Combine feedback response data and rehabilitation pathway execution status, use incremental transfer learning algorithm to update the parameters of the initial EEG decoding model and feature template. The personalized adaptation unit simultaneously corrects the adaptation threshold and evaluation matrix judgment criteria of the initial EEG decoding model. The updated initial EEG decoding model, adaptation parameters and rehabilitation pathway are stored in the storage unit.

[0007] Furthermore, an initial EEG decoding model and feature template are generated using a transfer learning algorithm, including: The upper limb movement type, EEG signal acquisition scenario, signal quality level, and EEG characteristic intensity level of different populations were determined as the model training feature group, and the injury type, muscle strength level, and joint range of motion in the patient's clinical data were determined as the model adaptation feature group. Calculate the similarity coefficients between the model training feature group, the model adaptation feature group, and the historical training feature group and historical adaptation feature group in the historical model library in the dataset; A preset similarity coefficient threshold is set to filter out all historical training feature groups and historical adaptation feature groups whose similarity coefficients are greater than the preset similarity coefficient threshold, and training selection feature groups and adaptation selection feature groups are constructed respectively. Based on the training and screening feature group, an effective subset of EEG features in the dataset is determined. Then, the EEG features of healthy individuals are transferred to the patient data training process using a transfer learning algorithm to train and generate an initial EEG decoding model. Based on the adaptive filtering feature group, the core feature dimensions and matching thresholds of the feature template are determined to form an initial feature template.

[0008] Further, the step of determining the effective EEG feature subset of the dataset based on the training and screening feature group and training to generate an initial EEG decoding model, and determining the core feature dimension and matching threshold of the feature template based on the adaptation and screening feature group, includes: The training feature set and the fitting feature set with the maximum similarity coefficient are obtained from the training feature set and the fitting feature set, respectively. If the maximum similarity coefficient training feature set contains only a single historical feature set, and the maximum similarity coefficient fitting feature set contains only a single historical feature set, then the EEG feature subset corresponding to the historical training feature set is directly used as the effective feature subset of the dataset, and the initial EEG decoding model is generated by combining the transfer learning algorithm; the feature dimension and threshold corresponding to the historical fitting feature set are directly used as the core feature dimension and matching threshold of the initial feature template. If there are multiple historical feature groups in the maximum similarity coefficient training feature group set and the maximum similarity coefficient fitting feature group set, then the EEG feature dimensions of each historical training feature group are aligned first, and then the weighted fusion value of the corresponding EEG feature is calculated using the similarity coefficient of each historical training feature group as the weight, forming an effective EEG feature subset of the dataset, and then the initial EEG decoding model is generated by combining the transfer learning algorithm. Simultaneously, using the similarity coefficient of each historical adaptation feature group as the weight, the weighted average of the corresponding feature dimension parameters and matching threshold is calculated as the core feature dimension and matching threshold of the initial feature template.

[0009] Furthermore, the data used for signal quality screening include the signal-to-noise ratio, artifact persistence rate, and feature integrity of the EEG signal. The feature completeness refers to the coverage ratio of effective EEG feature dimensions; Preset quality standard data to determine whether the EEG signal passes quality inspection, including: When comparing the signal quality judgment data with the preset quality standard data, if the signal-to-noise ratio is greater than or equal to the preset signal-to-noise ratio threshold, the artifact residue rate is less than or equal to the preset residue rate threshold, and the feature integrity is greater than or equal to the preset integrity threshold, then the EEG signal is judged to be qualified for quality inspection and retained as an effective signal segment for synchronization during the motor imagery period. When the signal-to-noise ratio is less than a preset signal-to-noise ratio threshold, the artifact persistence rate is greater than a preset persistence rate threshold, or the feature integrity is less than a preset integrity threshold, the EEG signal is determined to be unqualified for quality inspection. The signal segment is removed, the reason for the abnormality is recorded, and the corresponding acquisition time sequence information is marked synchronously.

[0010] Furthermore, the judgment data for the information quality inspection includes the type matching degree, strength stability, and completion reliability of the decoded information; Preset information quality standard data to determine whether the decoded information passes quality inspection, including: When the type matching degree is greater than or equal to the preset matching degree threshold, the strength stability is greater than or equal to the preset stability threshold, and the completion reliability is greater than or equal to the preset reliability threshold, the decoded information is determined to be qualified for quality inspection, and is directly included in the decoding report and associated with the corresponding time series data and clinical information; If the type matching degree is less than a preset matching degree threshold, the strength stability is less than a preset stability threshold, or the completion reliability is less than a preset reliability threshold, then the decoded information is determined to be unqualified, and a second decoding process is initiated. If the information quality inspection meets the standards after secondary decoding, the secondary decoding result will be included in the decoding report and marked as a secondary correction record. If the standard is still not met after secondary decoding, the decoding information is discarded, the reason for the decoding abnormality is recorded, the corresponding valid signal segment and acquisition timing information are marked simultaneously, and it is not included in the subsequent rehabilitation assessment data.

[0011] Furthermore, the feedback types of the multimodal neural feedback unit include visual feedback, auditory feedback, and tactile feedback, and the feedback intensity is positively correlated with the intention execution completion degree and intensity stability in the decoding report; The dynamic adjustment process for the output feedback information includes: Two preset feedback intensity thresholds, high and medium, are used to compare the intent execution completion rate and intensity stability in the decoding report with the preset thresholds: When the completion rate and intensity stability of the intention execution are both greater than or equal to the advanced feedback threshold, the three-level collaborative feedback mode is activated: visual feedback displays the complete motion trajectory animation, auditory feedback plays high-frequency excitation sound effects, and tactile feedback outputs strong vibration prompts. When the completion rate and intensity stability of the intent execution are both between the medium and high feedback thresholds, the second-level collaborative feedback mode is activated, retaining the combination of visual and auditory feedback, adjusting the sound effect frequency and animation clarity to a medium level, and turning off haptic feedback. When the completion rate and intensity stability of the intention are both less than or equal to the intermediate feedback threshold, the basic feedback mode is activated, retaining only visual feedback, displaying simplified action guidance icons, and simultaneously reducing the frequency of feedback triggering to avoid overstimulation. The system collects real-time data on the patient's attention concentration during feedback responses. If the attention concentration is less than the preset focus threshold, the system automatically increases the current feedback intensity level by one gradient, and then returns to the original level after three action cycles.

[0012] Furthermore, the adjustment of the assist force of the exoskeleton rehabilitation device is based on the intensity level in the decoding report, the muscle strength level in the patient's clinical data, and the real-time movement execution deviation. The precise control of the adjustment of the assist force includes: A preset muscle strength matching coefficient table and deviation compensation coefficient are used. The basic assistance coefficient is matched from the matching coefficient table according to the patient's muscle strength level, and the initial assistance force is determined by combining the intensity level in the decoding report. Real-time acquisition of position and speed deviation data during the execution of the action, and calculation of the comprehensive deviation value: when the comprehensive deviation value is less than or equal to the preset deviation threshold, the initial assist force remains unchanged; When the overall deviation value is greater than the preset deviation threshold but less than or equal to twice the preset deviation threshold, the assist force is increased by 10%-30% according to the deviation compensation coefficient, with a focus on compensating for joints with greater resistance to the action. When the overall deviation value exceeds twice the preset deviation threshold, the current motion guidance is paused, deviation prompts are sent through the multimodal neurofeedback unit, and the assistance intensity is reduced to 50% of the base value. Motion guidance is restarted after the patient adjusts their posture, and abnormal deviation data is recorded simultaneously for subsequent rehabilitation pathway optimization.

[0013] Furthermore, the core indicators of the rehabilitation assessment matrix include the degree of intention execution completion and the rate of achievement of intensity level in the decoded data, the frequency of action correction and coordination consistency in the feedback response data, and the muscle strength level and range of motion in the initial clinical data. The process of setting training period division criteria and generating personalized rehabilitation pathways includes: The system presets threshold values ​​for the basic training period, intensive training period, and consolidation training period, including a first preset threshold, a second preset threshold, a correction threshold, and a synergistic threshold. The first preset threshold is lower than the second preset threshold. The core indicators of the rehabilitation assessment matrix are then compared with these preset thresholds. When the completion rate of intention execution and the rate of achievement of intensity level are less than or equal to the first preset threshold, and the frequency of action correction is greater than the preset correction threshold, it is determined that the basic training period has been entered. The path parameters are set as basic action training combination, low frequency, fewer number of times per set of actions, gentle intensity increase gradient and long training cycle. When the completion rate of intent execution and the rate of achievement of intensity level are between the first preset threshold and the second preset threshold, and the coordination consistency is greater than the preset coordination threshold, it is determined that the intensive training period is entered, and the path parameters are adjusted to compound action training combination, medium frequency, moderate number of actions per set, medium intensity incremental gradient and regular training cycle. When the completion rate of intention execution and the rate of achievement of intensity level are greater than or equal to the second preset threshold, and the frequency of action correction is less than or equal to the preset correction threshold, it is determined that the training period is to be entered. The path parameters are set as comprehensive action training combination, high frequency, more number of times per set of actions, steep gradient of intensity and short training cycle. Each week, the updated rehabilitation assessment matrix of intention execution completion rate and intensity level achievement rate is extracted and compared with the corresponding training period threshold. If both key upgrade conditions are met at the same time, the path parameters are adjusted upward according to the gradient. If either intention execution completion rate or intensity level achievement rate falls back to the threshold range of the previous training period, the parameters of the previous training period are reverted. Simultaneously, the training plan and movement demonstration for the next day are updated through the AR device.

[0014] Furthermore, the criteria for determining whether the selected EEG signal segments meet the quality inspection standards and the intended execution requirements include the signal quality inspection pass rate, the degree of matching of the intended execution, and the accuracy of the action completion. The method of updating the parameters of the initial EEG decoding model and feature template using incremental transfer learning algorithm includes: The system presets thresholds for the number of valid signal segments, a threshold for full compliance, and a threshold for partial compliance. It then compares the filtered signal segment judgment data with these preset thresholds. When the number of signal segments that pass the quality inspection is greater than or equal to the preset threshold, and the intention execution matching degree and action completion accuracy are both greater than or equal to the full compliance threshold, the EEG features of a single batch of signal segments are extracted as incremental training data and integrated into the original training dataset according to the first weight ratio. The network parameters of the initial EEG decoding model and the core feature dimension matching threshold of the feature template are updated through the incremental transfer learning algorithm. When the number of signal segments that pass the quality inspection is between 50% and 100% of the preset threshold, and the intention execution matching degree and action completion accuracy are greater than or equal to the partial pass threshold, the sub-signal segments in the batch of signal segments that fully meet the intention execution are extracted as incremental training data and integrated into the original training dataset according to the second weight ratio. Only the key layer parameters and feature template matching thresholds of the initial EEG decoding model are updated, and the core feature dimensions are retained. When the number of qualified signal segments is less than or equal to 50% of the preset threshold, or when the intention execution matching degree or action completion accuracy is less than the partial qualification threshold, the model parameters will not be updated on a large scale. Instead, the personalized adaptation unit will fine-tune the adaptation threshold and the fault tolerance range of the initial EEG decoding model and the evaluation matrix judgment standard based on the deviation data of the rehabilitation path, without changing the core parameters and feature template structure.

[0015] Furthermore, before storing the updated initial EEG decoding model, adaptation parameters, and rehabilitation path into the storage unit, the effectiveness of the update needs to be verified by comparing historical data. The comparison criteria include model decoding accuracy, path execution compliance rate, and patient rehabilitation effect feedback data. The verification and final storage process includes: Preset accuracy deviation threshold and target achievement deviation threshold, calculate the first deviation value between the decoding accuracy of the updated model and the historical training best decoding accuracy, and the second deviation value between the execution achievement rate of the updated rehabilitation path and the target achievement rate of the preset rehabilitation goal. When the first deviation value is less than or equal to the preset accuracy deviation threshold and the second deviation value is less than or equal to the preset target achievement deviation threshold, the model and path update is deemed effective. The updated initial EEG decoding model, adaptation parameters and rehabilitation path are directly associated and stored in the storage unit, overwriting the original old data. If the first deviation value is greater than the preset accuracy deviation threshold or the second deviation value is greater than the preset pass rate deviation threshold, the update effect is determined to be unsatisfactory, and the parameter callback mechanism is activated. If only the first deviation value is greater than the preset accuracy deviation threshold, the updated feature template parameters are retained, the network layer parameters of the initial EEG decoding model are called back to the most recent valid update version, the decoding accuracy is recalculated, and the accuracy is stored after it meets the threshold. If the second deviation value is greater than the preset target achievement rate deviation threshold, the updated model parameters are retained, and the core parameters such as the intensity increment gradient of the rehabilitation path and the training frequency are called back to the previous version. After fine-tuning based on the patient's rehabilitation effect feedback data, the model is stored. If the two deviation values ​​are greater than the preset threshold, the system will completely revert to the previous version of the model, adaptation parameters, and recovery path, record the reason for the update failure, and not perform the storage overwrite operation.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By integrating EEG signals of upper limb motor imagery from multiple populations and constructing a high-quality dataset through triple screening based on signal-to-noise ratio, artifact persistence rate, and feature integrity, and combining transfer learning algorithms to fuse EEG features of healthy individuals and patients, while optimizing the initial EEG decoding model and feature template through similarity-weighted fusion, the impact of individual differences and signal noise on the decoding results is effectively reduced, significantly improving the accuracy and stability of upper limb rehabilitation intention recognition.

[0017] 2. Input clinical data such as patient injury type and muscle strength level, construct a rehabilitation assessment matrix including core indicators such as intention execution completion and movement correction frequency, divide the training into three levels of basic, intensive and consolidation training and dynamically adjust the path parameters to achieve precise matching between the rehabilitation plan and the patient's individual foundation and rehabilitation progress, and avoid the problem of insufficient adaptability caused by fixed templates.

[0018] 3. Equipped with visual, auditory, and tactile multimodal neurofeedback units, the feedback intensity is dynamically linked to the degree of intention execution, intensity stability, and patient attention concentration. Combined with the precise assistance control of exoskeleton rehabilitation devices, a complete closed loop of "intention-action-feedback-adjustment" is formed, which guides patients to optimize their training state in a timely manner and improves training participation and effectiveness.

[0019] 4. An incremental transfer learning algorithm is adopted. Based on the EEG signal segments and feedback response data that meet the quality inspection standards and are intended to be executed in accordance with the requirements, the EEG decoding model and feature template parameters are continuously updated. The adaptation threshold and evaluation criteria are corrected in sync to ensure that the model can be dynamically optimized as the patient's neurological function recovers and EEG characteristics change, thus ensuring the adaptability and stability of long-term rehabilitation training. Attached Figure Description

[0020] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of an upper limb rehabilitation intention recognition feedback training method based on EEG signal analysis provided in an embodiment of the present invention. Detailed Implementation

[0021] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] Reference Figure 1 In some embodiments of this application, a method for upper limb rehabilitation intention recognition feedback training based on electroencephalogram (EEG) signal analysis includes the following steps: Step S100: Integrate upper limb motor imagery EEG signals from different populations and perform signal quality screening to construct a dataset. The dataset contains effective EEG features corresponding to various upper limb movements. Construct a feature extraction module and an intent classification unit. Use transfer learning algorithm to train and generate an initial EEG decoding model and feature template. Input patient clinical data and set the parameter adjustment threshold for the initial EEG decoding model. Associate and store the determined initial EEG decoding model, feature template, and clinical information in the storage unit. Step S200: Deploy the signal acquisition unit to cover the corresponding scalp area of ​​the brain. After acquiring EEG signals, perform signal quality screening simultaneously. Input the effective signal segments that are synchronous with the motor imagination period and pass the quality inspection into the initial EEG decoding model to extract features and generate action intention vectors. Match feature templates to decode information and perform information quality inspection. Associate time series data, clinical information and information quality inspection results to form a decoding report. Step S300: Configure the multimodal neurofeedback unit to output feedback information, deploy the exoskeleton rehabilitation device to adjust the assistance force according to the decoding report to guide the execution of the movement, record the patient's feedback response behavior, extract the qualified decoding data and response data and combine them with the initial clinical information to construct a rehabilitation assessment matrix, set the training period division criteria, generate a personalized rehabilitation path and dynamically adjust the parameters; Step S400: Select EEG signal segments that meet the quality inspection standards and are intended to be executed in accordance with the requirements. Combine the feedback response data and the execution status of the rehabilitation path, use the incremental transfer learning algorithm to update the parameters of the initial EEG decoding model and feature template. The personalized adaptation unit simultaneously corrects the adaptation threshold and evaluation matrix judgment criteria of the initial EEG decoding model. The updated initial EEG decoding model, adaptation parameters and rehabilitation path are stored in the storage unit.

[0023] The above embodiments integrate upper limb motor imagery EEG signals from multiple populations and construct a high-quality dataset through triple screening based on signal-to-noise ratio, artifact persistence rate, and feature integrity. A transfer learning algorithm is then used to fuse EEG features from healthy individuals and patients. Simultaneously, similarity-weighted fusion optimizes the initial EEG decoding model and feature templates, effectively reducing the impact of individual differences and signal noise on the decoding results, significantly improving the accuracy and stability of upper limb rehabilitation intention recognition. Clinical data such as patient injury type and muscle strength level are input to construct a rehabilitation assessment matrix containing core indicators such as intention execution completion and movement correction frequency. This matrix is ​​divided into three training phases: basic, intensive, and consolidation, with dynamic adjustments to path parameters to achieve precise matching between the rehabilitation plan and the patient's individual foundation and rehabilitation progress, avoiding the inadequacy of adaptability caused by fixed templates. Multimodal neurofeedback units (visual, auditory, and tactile) are configured, with feedback intensity dynamically linked to intention execution completion, intensity stability, and patient attention concentration. Combined with precise assistive control from exoskeleton rehabilitation devices, a complete closed loop of "intention-movement-feedback-adjustment" is formed, guiding patients to optimize their training state in a timely manner, improving training participation and effectiveness. An incremental transfer learning algorithm is used to continuously update the EEG decoding model and feature template parameters based on EEG signal segments and feedback response data that meet the quality inspection standards and are intended to be executed in accordance with the requirements. The adaptation threshold and evaluation criteria are adjusted in sync to ensure that the model can be dynamically optimized as the patient's neurological function recovers and EEG characteristics change, thus ensuring the adaptability and stability of long-term rehabilitation training.

[0024] Specifically, a transfer learning algorithm is used to train and generate an initial EEG decoding model and feature template, including: The upper limb movement type, EEG signal acquisition scenario, signal quality level, and EEG characteristic intensity level of different populations were determined as the model training feature group, and the injury type, muscle strength level, and joint range of motion in the patient's clinical data were determined as the model adaptation feature group. Calculate the similarity coefficients between the model training feature group, the model adaptation feature group, and the historical training feature group and historical adaptation feature group in the historical model library in the dataset, respectively. A preset similarity coefficient threshold is set to filter out all historical training feature groups and historical adaptation feature groups whose similarity coefficients are greater than the preset similarity coefficient threshold, and training selection feature groups and adaptation selection feature groups are constructed respectively. Based on the training and screening feature groups, the effective EEG feature subset of the dataset is determined. Then, the EEG features of healthy individuals are transferred to the patient data training process using the transfer learning algorithm to train and generate the initial EEG decoding model. The core feature dimensions and matching thresholds of the feature template are determined based on the adaptive filtering feature groups, thus forming the initial feature template.

[0025] Specifically, the effective subset of EEG features in the dataset is determined based on the training feature set, and an initial EEG decoding model is generated through training. Furthermore, the core feature dimensions and matching thresholds of the feature template are determined based on the adaptation feature set, including: Obtain the set of training feature groups with the maximum similarity coefficient and the set of fitting feature groups with the maximum similarity coefficient from the training feature group and the fitting feature group respectively; If the training feature set with the maximum similarity coefficient contains only a single historical feature set, and the matching feature set with the maximum similarity coefficient contains only a single historical feature set, then the EEG feature subset corresponding to the historical training feature set is directly used as the effective feature subset of the dataset, and the initial EEG decoding model is generated by combining it with the transfer learning algorithm; the feature dimension and threshold corresponding to the historical matching feature set are directly used as the core feature dimension and matching threshold of the initial feature template. If there are multiple historical feature groups in the training feature group set and the fitting feature group set with the maximum similarity coefficient, the EEG feature dimensions of each historical training feature group are aligned first. Then, the weighted fusion value of the corresponding EEG feature is calculated using the similarity coefficient of each historical training feature group as the weight, forming an effective subset of EEG features in the dataset. The initial EEG decoding model is then generated by combining the transfer learning algorithm. Simultaneously, using the similarity coefficient of each historical adaptation feature group as the weight, the weighted average of the corresponding feature dimension parameters and matching threshold is calculated as the core feature dimension and matching threshold of the initial feature template.

[0026] Specifically, the upper limb movement types, EEG signal acquisition scenarios, signal quality levels, and EEG characteristic intensity levels based on EEG signal energy normalization for different populations, including healthy individuals and patients with different types of upper limb injuries, are determined as the model training feature group. The clearly defined injury type (e.g., upper limb hemiplegia after stroke, upper limb motor dysfunction due to spinal cord injury), muscle strength level (classified according to medical muscle strength grading standards), and joint range of motion (recorded according to joint range of motion measurement standards) from the patients' clinical data are determined as the model adaptation feature group. The cosine similarity algorithm is used to calculate the similarity coefficients between the model training feature group, the model adaptation feature group, and the historical training feature group and historical adaptation feature group in the centralized historical model library. The similarity coefficients are used to quantify the similarity between the two groups. The degree of correlation between group features; based on the distribution of historical model library data, a preset similarity coefficient threshold is set, and all historical training feature groups and historical adaptation feature groups with similarity coefficients greater than the preset threshold are selected, and training selection feature groups and adaptation selection feature groups are constructed respectively; from the training selection feature groups and adaptation selection feature groups, the set of training feature groups with the highest similarity coefficient and the set of adaptation feature groups with the highest similarity coefficient are obtained respectively (the highest similarity coefficient refers to the highest similarity value among feature groups with a value higher than the preset threshold). If the set of training feature groups with the highest similarity coefficient contains only a single historical feature group and the set of adaptation feature groups with the highest similarity coefficient contains only a single historical feature group, and the single historical feature group matches the current patient's model training feature group and adaptation feature group with a high degree of similarity, then the following criteria are considered: Optimal approach: The subset of EEG features corresponding to the historical training feature group is directly used as the effective feature subset of the dataset. A pre-training-fine-tuning transfer learning algorithm is employed, using the basic model parameters trained on the EEG features of healthy individuals as initial values. The model is then fine-tuned based on this effective feature subset to generate an initial EEG decoding model. Simultaneously, the feature dimensions and thresholds corresponding to the historical fitting feature group are directly used as the core feature dimensions and matching thresholds of the initial feature template. If both the maximum similarity coefficient training feature group set and the maximum similarity coefficient fitting feature group set contain multiple historical feature groups, the EEG feature dimensions of each historical training feature group are first aligned (by retaining the common core feature dimensions of each feature group and using zero-padding or feature dimensionality reduction for unique dimensions). To achieve dimension unification, the similarity coefficients of each historical training feature group are used as weights. The weighted fusion value of the corresponding EEG feature is calculated according to "weighted fusion value = Σ(similarity coefficients of each historical training feature group × corresponding EEG feature value) / Σsimilarity coefficients of each historical training feature group". This forms an effective subset of EEG features in the dataset. The initial EEG decoding model is trained using a pre-training-fine-tuning transfer learning algorithm. At the same time, the weighted average of the corresponding feature dimension parameters and matching thresholds is calculated using the similarity coefficients of each historical adaptation feature group as weights. This feature dimension needs to cover key EEG features related to upper limb motor imagery (such as frequency features and temporal amplitude features corresponding to μ waves and β waves). Finally, these become the core feature dimensions and matching thresholds of the initial feature template.

[0027] Understandably, by constructing a dual-dimensional feature system—a model training feature set and a model adaptation feature set—combining similarity coefficients to screen historical feature sets and determining effective EEG feature subsets and feature template parameters for different scenarios, and simultaneously introducing transfer learning algorithms to achieve efficient transfer of EEG features from healthy individuals to patient data, multiple beneficial effects are achieved: It solves the problem of insufficient model adaptability caused by single-dimensional feature selection; and by accurately matching historical data with dual feature sets, it enables the initial EEG decoding model and feature templates to be specifically adapted to individual clinical characteristics of patients, such as injury type and muscle strength level, thereby improving the initial decoding accuracy and feature matching of the model. Accuracy; and by handling the feature set with the maximum similarity coefficient in different cases (direct reuse of a single feature set and weighted fusion of multiple feature sets), the efficiency of model training and the comprehensiveness of feature representation are taken into account, avoiding data redundancy or missing key features; at the same time, the transfer learning effectively makes up for the shortage of patient EEG training data, enabling the model to have a high generalization ability in the initial stage, reducing the iteration cost of subsequent incremental training, and providing reliable model and template support for the rapid start and accurate implementation of personalized rehabilitation training, ultimately improving the real-time performance and accuracy of upper limb rehabilitation intention recognition and the personalized adaptation effect of rehabilitation training.

[0028] Specifically, the data used for signal quality screening include the signal-to-noise ratio, artifact persistence rate, and feature integrity of the EEG signal. Feature completeness refers to the coverage ratio of effective EEG feature dimensions; Preset quality standard data to determine whether the EEG signal passes quality inspection, including: When comparing the signal quality assessment data with the preset quality standard data, if the signal-to-noise ratio is greater than or equal to the preset signal-to-noise ratio threshold, the artifact residue rate is less than or equal to the preset residue rate threshold, and the feature integrity is greater than or equal to the preset integrity threshold, then the EEG signal is deemed to be qualified for quality inspection and retained as a valid signal segment for synchronization during the motor imagery period. When the signal-to-noise ratio is less than the preset signal-to-noise ratio threshold, the artifact persistence rate is greater than the preset persistence rate threshold, or the feature integrity is less than the preset integrity threshold, the EEG signal is determined to be unqualified for quality inspection. The signal segment is removed, the reason for the abnormality is recorded, and the corresponding acquisition time sequence information is marked synchronously.

[0029] Specifically, the data used for signal quality screening include the signal-to-noise ratio (SNR) of the EEG signal, artifact persistence rate, and feature integrity. The SNR is the ratio of the average amplitude of the effective signal in the 8-30Hz target frequency band related to motor imagery to the average amplitude of the noise signal from non-target frequency bands and environmental interference. The artifact persistence rate is the proportion of artifact signals (including residual eye movement, electromyography, and power frequency interference) that were not removed after independent component analysis (ICA) to remove OMG artifacts, EMG artifacts, and 50Hz power frequency notch filtering. Feature integrity refers to the percentage of the total signal that was not removed. The ratio of the number of dimensions of preset core EEG features (including time-domain amplitude, frequency-domain energy, and phase characteristics of μ and β waves) extracted from EEG signals to the total number of preset core feature dimensions; preset quality standard data including signal-to-noise ratio threshold, artifact persistence threshold, and feature integrity threshold; and clearly defining the criteria for determining synchronization during motor imagery: the deviation between the EEG signal acquisition timestamp and the patient's motor imagery command timestamp is ≤100ms, and the signal contains motor imagery feature activation responses with β wave amplitude suppression and μ wave amplitude enhancement. Based on these criteria, the judgment is made... Whether the EEG signal passes quality inspection: The signal quality judgment data is compared with the preset quality standard data. When the signal-to-noise ratio is greater than or equal to the preset signal-to-noise ratio threshold, the artifact residue rate is less than or equal to the preset residue rate threshold, and the feature integrity is greater than or equal to the preset integrity threshold, and the synchronization requirements of the motor imagery period are met, the EEG signal is judged to pass quality inspection and is retained as a valid signal segment. When the signal-to-noise ratio is less than the preset signal-to-noise ratio threshold, the artifact residue rate is greater than the preset residue rate threshold, or the feature integrity is less than the preset integrity threshold, or the synchronization requirements of the motor imagery period are not met, the EEG signal is judged to fail quality inspection, the signal segment is immediately removed, and the reasons for the abnormality are recorded in detail (such as the signal-to-noise ratio not meeting the standard due to excessive environmental interference or poor electrode contact, the artifact residue rate exceeding the standard due to improper preprocessing parameter settings or excessive patient movement amplitude, the feature integrity not meeting the standard due to insufficient signal acquisition time or electrode detachment, and the synchronization not meeting the standard due to command transmission delay). The synchronization is marked with the corresponding acquisition device number, signal acquisition start-end time, corresponding motor imagery action type, and other time sequence and scene information to provide a basis for subsequent acquisition parameter optimization and abnormal investigation.

[0030] Understandably, constructing a three-dimensional judgment system based on signal-to-noise ratio, artifact residue rate, and feature integrity, combined with synchronous verification of motor imagery time periods and anomaly tracing mechanisms, brings multiple key benefits: Firstly, through multi-dimensional indicator collaborative verification, invalid EEG signals affected by environmental interference, artifact residue, and feature loss are comprehensively eliminated, effectively avoiding signal quality misjudgments caused by single indicator screening, ensuring that the retained effective signal segments have both high purity and complete feature representation, providing reliable data support for subsequent motor intention decoding and rehabilitation assessment matrix construction; Secondly, by clarifying the core feature coverage ratio definition of feature integrity and the synchronous judgment criteria for motor imagery time periods, the accuracy and consistency of signal screening are improved, reducing the interference of invalid data on the initial EEG decoding model training and intention recognition results, indirectly improving the model decoding accuracy and the personalized adaptability of rehabilitation path generation; At the same time, by recording the reasons for screening anomalies in detail and marking the collection time sequence information, a traceable basis is provided for subsequent optimization of signal collection parameters (such as electrode contact status, anti-interference settings), adjustment of the collection environment, or adaptation to the patient's motor state, reducing the risk of rehabilitation training interruption or effect deviation due to data quality issues, ultimately ensuring the stability of upper limb rehabilitation intention recognition and the efficiency of rehabilitation training.

[0031] Specifically, the data for judging information quality inspection includes the type matching degree, strength stability, and completion credibility of the decoded information; Preset information quality standard data to determine whether the decoded information passes quality inspection, including: When the type matching degree is greater than or equal to the preset matching degree threshold, the strength stability is greater than or equal to the preset stability threshold, and the completion reliability is greater than or equal to the preset reliability threshold, the decoded information is judged to be qualified for quality inspection, directly included in the decoding report and associated with the corresponding time series data and clinical information; If the type matching degree is less than the preset matching degree threshold, the strength stability is less than the preset stability threshold, or the completion reliability is less than the preset reliability threshold, then the decoded information is determined to be unqualified and a second decoding process is initiated. If the information quality inspection meets the standards after secondary decoding, the secondary decoding result will be included in the decoding report and marked as a secondary correction record. If the standard is still not met after secondary decoding, the decoding information is discarded, the reason for the decoding abnormality is recorded, the corresponding valid signal segment and acquisition timing information are marked simultaneously, and it is not included in the subsequent rehabilitation assessment data.

[0032] Specifically, the data for information quality inspection includes the type matching degree, intensity stability, and completion reliability of the decoded information. The decoded information specifically covers upper limb movement types (such as pre-set rehabilitation movements like raising the arm, flexing the elbow, clenching the fist, and extending the fingers), the intensity level of the movement execution, and the progress of intention completion. Type matching degree refers to the cosine similarity between the decoded intention vector and the corresponding pre-set rehabilitation movement feature vector in the feature template (calculated as the dot product of the two vectors divided by the product of their magnitudes). Intensity stability refers to the coefficient of variation (i.e., the ratio of the standard deviation to the mean) of the EEG feature intensity values ​​obtained within 50ms. Completion reliability refers to the consistency between the initial EEG decoding model output intention classification results and the target EEG intensity values. The probability value (calculated using the softmax function of the model output layer); preset information quality standard data including matching degree threshold, stability threshold, and credibility threshold; and specifying the association rule between decoded information and effective signal segments as follows: the deviation between the decoding timestamp and the effective signal segment acquisition timestamp is ≤50ms. Based on this, it is determined whether the decoded information passes quality inspection: when the type matching degree is greater than or equal to the preset matching degree threshold, the intensity stability is greater than or equal to the preset stability threshold, and the completion credibility is greater than or equal to the preset credibility threshold, the decoded information is determined to pass quality inspection, is directly included in the decoding report, and associated with the corresponding time series data (including decoding start-end time, effective signal segment acquisition number, motion...). The decoded information is analyzed using the imagined action command timestamp and clinical information (including the patient's current muscle strength level, injury type, and joint range of motion). If the type matching degree is less than a preset matching degree threshold, the strength stability is less than a preset stability threshold, or the completion reliability is less than a preset reliability threshold, the decoded information is deemed unqualified for quality control. A second decoding process is immediately initiated. Specifically, the bandpass filter range of the feature extraction module is adjusted (from the original 8-30Hz to 5-35Hz), the matching weights between the action intent vector and the feature template are recalculated (increasing the weight ratio of μ-wave and β-wave features), and then the decoding operation is performed again. If the information quality control meets the standards after the second decoding, the second decoding result is included in the decoding report and marked. Record the secondary correction (including the adjusted values ​​of filter parameters and the optimized ratio of matching weights); if the standard is still not met after the secondary decoding, remove the decoding information and record in detail the reasons for the decoding anomaly (such as the failure of type matching due to the feature template matching threshold being set too strictly or the EEG feature extraction dimension being missing, the poor intensity stability due to the patient's lack of concentration or signal acquisition fluctuations, and the low reliability of completion due to insufficient adaptation between the model parameters and the patient's EEG features). Simultaneously mark the acquisition device number, the start-end time of the signal segment, and the type of motor imagery action for the corresponding effective signal segment. This decoding information will not be included in the data scope of subsequent rehabilitation assessment matrix construction and rehabilitation path adjustment to avoid invalid information interfering with rehabilitation training decisions.

[0033] Understandably, by constructing a three-dimensional judgment system encompassing type matching degree, intensity stability, and completion credibility, coupled with a secondary decoding error correction mechanism and full-link anomaly tracing markers, multiple key beneficial effects are achieved: Firstly, through multi-dimensional indicator collaborative verification, the accuracy, stability, and reliability of decoded information are accurately identified, avoiding mis-screening of qualified information or omission of invalid information due to single-indicator judgment. This ensures that the data included in the decoding report possesses high matching degree, low volatility, and strong credibility, providing accurate decision-making basis for adjusting the force of exoskeleton rehabilitation devices and generating personalized rehabilitation pathways. Secondly, through the secondary decoding process (optimizing filtering parameters and adjusting feature matching weights), This approach corrects and remedies for non-compliant information, significantly improving the utilization rate of effective decoded data, reducing data waste caused by single decoding deviations, and lowering the risk of rehabilitation training interruption. Simultaneously, by meticulously recording the causes of decoding anomalies and associating them with effective signal segments and acquisition timing information, it provides a traceable, targeted optimization basis for subsequent optimization of initial EEG decoding model parameters, adjustment of feature template matching thresholds, and improvement of signal acquisition quality, continuously enhancing the accuracy and stability of intent recognition. Finally, by rigorously eliminating invalid decoded information at each stage, it avoids interference with the construction of the rehabilitation assessment matrix and the dynamic adjustment of the rehabilitation pathway, ensuring the personalization, efficiency, and safety of upper limb rehabilitation training.

[0034] Specifically, the feedback types of the multimodal neural feedback unit include visual feedback, auditory feedback, and tactile feedback, and the feedback intensity is positively correlated with the degree of intention execution completion and intensity stability in the decoding report; The dynamic adjustment process for output feedback information includes: Two preset feedback intensity thresholds, high and medium, are used to compare the intent execution completion rate and intensity stability in the decoding report with the preset thresholds: When the completion rate and intensity stability of the intention execution are both greater than or equal to the advanced feedback threshold, the three-level collaborative feedback mode is activated: visual feedback displays the complete motion trajectory animation, auditory feedback plays high-frequency excitation sound effects, and tactile feedback outputs strong vibration prompts. When the completion rate and intensity stability of the intent execution are both between the medium and high feedback thresholds, the second-level collaborative feedback mode is activated, retaining the combination of visual and auditory feedback, adjusting the sound effect frequency and animation clarity to a medium level, and turning off haptic feedback. When the completion rate and intensity stability of the intention are both less than or equal to the intermediate feedback threshold, the basic feedback mode is activated, retaining only visual feedback, displaying simplified action guidance icons, and simultaneously reducing the frequency of feedback triggering to avoid overstimulation. The system collects real-time data on the patient's attention concentration during feedback responses. If the attention concentration is less than the preset focus threshold, the system automatically increases the current feedback intensity level by one gradient, and then returns to the original level after three action cycles.

[0035] Specifically, the multimodal neurofeedback unit provides feedback types including visual, auditory, and tactile feedback. Visual feedback is achieved through AR glasses or a rehabilitation training screen; auditory feedback is achieved through built-in speakers or headphones; and tactile feedback is achieved through the built-in vibration modules of the exoskeleton rehabilitation device (distributed at the wrist and elbow joints). The feedback intensity is positively correlated with the intention execution completion rate (the overlap rate between the actual movement trajectory and the preset rehabilitation movement trajectory, calculated using a coordinate point matching algorithm) and intensity stability (the fluctuation coefficient of EEG characteristic intensity within 100ms, i.e., the ratio of standard deviation to mean) in the decoding report. Furthermore, the feedback intensity grading is correlated with the synergistic achievement of both indicators. The degree is directly linked; the dynamic adjustment process of output feedback information specifically includes: preset high and medium-level feedback intensity thresholds (the high threshold corresponds to an intention execution completion rate ≥85% and intensity stability ≤0.2, and the medium threshold corresponds to an intention execution completion rate ≥60% and intensity stability ≤0.4), comparing the intention execution completion rate and intensity stability in the decoding report with the preset thresholds, and specifying that the feedback triggering time is within 300ms after the decoding report is generated and synchronized with the patient's action execution process in real time; when the intention execution completion rate and intensity stability are both greater than or equal to the high-level feedback threshold, the three-level collaborative feedback mode is activated, and the visual feedback displays a complete motion trajectory animation at a frame rate of 60fps (including... The system includes: real-time motion overlay and comparison with standard trajectories; auditory feedback playing high-frequency excitation sound effects of 1800-2200Hz (duration 1s, volume 60-70dB); haptic feedback outputting a strong vibration cue of 50Hz frequency and 80%-100% intensity (duration 500ms); when the intention execution completion rate and intensity stability are both between the medium and high feedback thresholds, a secondary collaborative feedback mode is activated, retaining the combination of visual and auditory feedback. The visual feedback is adjusted to a medium-definition animation with a frame rate of 30fps, the auditory feedback plays a mid-frequency excitation sound effect of 1000-1500Hz (duration 0.8s, volume 50-60dB), and haptic feedback is turned off. When the completion rate and intensity stability of the intention execution are both less than or equal to the intermediate feedback threshold, the basic feedback mode is activated, retaining only visual feedback and displaying simplified line-style action guidance indicators (highlighting the movement path of key joints). Simultaneously, the feedback trigger frequency is reduced from once every 500ms to once every 1000ms to avoid overstimulating the patient's senses. At the same time, the patient's attention concentration data in the feedback response behavior is collected in real time through the energy ratio of theta waves (4-7Hz) to beta waves (13-30Hz) in the EEG signal (attention concentration = beta wave energy / (theta wave energy + beta wave energy), the higher the ratio, the more focused the attention). If the attention concentration is less than the preset focus threshold (0.5) Automatically increase the current feedback intensity level by one gradient (specifically, increase the visual animation frame rate by 20fps, the auditory sound effect frequency by 500Hz or the volume by 10dB, or activate tactile feedback from a disabled state or increase the vibration intensity by 30%), continuing for 3 action cycles (a single action cycle is defined as the standard duration from the start to the completion of the action, such as 3 seconds), and then return to the original level, ensuring that the patient can still effectively receive feedback guidance when their attention is distracted.

[0036] Understandably, by constructing a three-dimensional feedback system integrating visual, auditory, and tactile senses, and dynamically adjusting the feedback mode based on the completion rate and intensity stability of the intention execution in the decoding report, coupled with an attention-linked enhancement mechanism, multiple key beneficial effects are achieved: Firstly, by dividing the three feedback modes into two levels of thresholds, the feedback intensity is precisely matched with the patient's training performance—when the training effect is excellent, the three-level collaborative feedback is activated to strengthen the incentive; when the performance is moderate, the core feedback combination is retained to balance guidance and efficiency; and when the performance is poor, the feedback is simplified to avoid overstimulation, fully adapting to the patient's state at different training stages and improving sensory acceptance and feedback compliance; Secondly, through multimodal collaborative feedback... This approach leverages the complementary advantages of visual and auditory feedback. Visual feedback provides intuitive movement guidance, auditory feedback enhances immediate motivation, and tactile feedback strengthens somatosensory perception, delivering training information in multiple dimensions to help patients quickly correct movement deviations and improve the accuracy and standardization of intention execution. Simultaneously, a focus-based adjustment mechanism automatically increases feedback intensity when the patient's attention is diverted, ensuring uninterrupted guidance and effectively maintaining training focus. Ultimately, through personalized and dynamic feedback design, it strengthens the memory and consolidation of correct movements while positively incentivizing patients to train, reducing resistance to training caused by inappropriate feedback, and significantly improving the efficiency and effectiveness of upper limb rehabilitation training.

[0037] Specifically, the adjustment of the assist level of exoskeleton rehabilitation devices is based on the intensity level in the decoding report, the muscle strength level in the patient's clinical data, and the real-time movement execution deviation. Precise control of the assist level, including: A preset muscle strength matching coefficient table and deviation compensation coefficient are used. The basic assistance coefficient is matched from the matching coefficient table according to the patient's muscle strength level, and the initial assistance force is determined by combining the intensity level in the decoding report. Real-time acquisition of position and speed deviation data during the execution of the action, and calculation of the comprehensive deviation value: when the comprehensive deviation value is less than or equal to the preset deviation threshold, the initial assist force remains unchanged; When the overall deviation value is greater than the preset deviation threshold but less than or equal to twice the preset deviation threshold, the assist force is increased by 10%-30% according to the deviation compensation coefficient, with a focus on compensating for joints with greater resistance to the action. When the overall deviation value exceeds twice the preset deviation threshold, the current motion guidance is paused, deviation prompts are sent through the multimodal neurofeedback unit, and the assistance intensity is reduced to 50% of the base value. Motion guidance is restarted after the patient adjusts their posture, and abnormal deviation data is recorded simultaneously for subsequent rehabilitation pathway optimization.

[0038] Specifically, the adjustment of the assist level of the exoskeleton rehabilitation device is based on the intensity level in the decoding report (levels 1-5 based on EEG characteristic energy normalization, with level 1 corresponding to the lowest motor intention intensity and level 5 corresponding to the highest motor intention intensity), the muscle strength level (0-5, where level 0 represents complete paralysis and level 5 represents normal muscle strength) in the patient's clinical data according to the MMT (Medical Muscle Strength Grading Standard), and real-time movement execution deviations (including positional and speed deviations). Precise control of the assist level adjustment includes: a preset muscle strength adaptation coefficient table (muscle strength level 0 corresponds to coefficient 1.0, level 1 to 0.9, level 2 to 0.7, level 3 to 0.5, level 4 to 0.3, and level 5 to 0.1; the lower the muscle strength level, the larger the adaptation coefficient, ensuring that patients with weak muscle strength receive sufficient assistance) and a deviation compensation coefficient (range 0.2-0). 0.5, used to quantify the correlation between deviation and the magnitude of assist adjustment), the initial assist force is calculated according to the formula "Initial Assist Force = Basic Assist Coefficient × Intensity Level Coefficient × Rated Assist Value of the Instrument" (where the intensity level coefficients 1-5 correspond to 0.4, 0.6, 0.7, 0.8, and 1.0, respectively). That is, the basic assist coefficient is matched from the adaptation coefficient table according to the patient's muscle strength level, and the intensity level coefficient and rated assist value of the instrument corresponding to the intensity level in the decoding report are combined to determine the initial assist force; through the encoder and inertial measurement unit (IMU) deployed at the joints of the exoskeleton rehabilitation device, the position deviation (Euclidean distance between the actual position and the preset standard trajectory position) and speed deviation (difference between the actual speed and the preset standard speed) data of each joint during the execution of the movement are collected in real time, according to " Deviation Comprehensive Value = 0.6 × Position Deviation + 0.The deviation comprehensive value is calculated using a 4× velocity deviation (position deviation has a higher weight than velocity deviation, prioritizing the accuracy of the movement trajectory): When the deviation comprehensive value is less than or equal to the preset deviation threshold, the initial assist force remains unchanged; when the deviation comprehensive value is greater than the preset deviation threshold but less than or equal to twice the preset deviation threshold, the adjustment ratio is determined according to "adjustment range = deviation compensation coefficient × (deviation comprehensive value - preset deviation threshold) / preset deviation threshold × 100%" (limited to a range of 10%-30%), focusing on compensating joints with high resistance to movement (determined by position deviation distribution, velocity deviation peak, and the patient's clinical joint movement limitation information, such as stroke patients often needing to focus on compensating the elbow and wrist joints); when the deviation comprehensive value is greater than twice the preset deviation threshold, the initial assist force remains unchanged. When a deviation threshold is set, the current movement guidance is immediately paused. A deviation warning message is sent via the multimodal neurofeedback unit (visual feedback displays the text "Movement deviation too large, please adjust posture," auditory feedback plays a 1500Hz warning sound, and tactile feedback outputs intermittent vibrations). Simultaneously, the assistance intensity is reduced to 50% of the baseline value to lower the risk of movement execution. Once the exoskeleton sensor detects that the patient's posture has been adjusted to the preset safe range (overall deviation value ≤ preset deviation threshold), movement guidance is restarted. Abnormal deviation data (including overall deviation value, time of occurrence of the abnormality, affected joints, and the patient's muscle strength at the time) is recorded synchronously for subsequent adjustments to movement difficulty, optimization of assistance parameters, and correction of model adaptation thresholds in the rehabilitation pathway.

[0039] Understandably, personalized initial assistance is achieved through a muscle strength matching coefficient table, combined with dynamic adjustments to the magnitude of the assistance based on the overall deviation value—stable assistance for small deviations, targeted resistance compensation for medium deviations, and pausing with a warning when exceeding the threshold, ensuring training safety. Abnormal data is recorded simultaneously to optimize the rehabilitation pathway, achieving full-link management of assistance from "personalized adaptation to dynamic correction to safe execution to iterative optimization," improving the standardization, safety, and efficiency of upper limb rehabilitation training, and accelerating the recovery of motor function.

[0040] Specifically, the core indicators of the rehabilitation assessment matrix include the degree of intention execution completion and the rate of achievement of intensity level in the decoded data, the frequency of movement correction and coordination consistency in the feedback response data, and the muscle strength level and range of motion in the initial clinical data. Establish training period segmentation criteria and generate personalized rehabilitation pathways, including: The system presets threshold values ​​for the basic training period, intensive training period, and consolidation training period, including a first preset threshold, a second preset threshold, a correction threshold, and a synergistic threshold. The first preset threshold is lower than the second preset threshold. The core indicators of the rehabilitation assessment matrix are then compared with these preset thresholds. When the completion rate of intention execution and the rate of achievement of intensity level are less than or equal to the first preset threshold, and the frequency of action correction is greater than the preset correction threshold, it is determined that the basic training period has been entered. The path parameters are set as basic action training combination, low frequency, fewer number of times per set of actions, gentle intensity increase gradient and long training cycle. When the completion rate of intent execution and the rate of achievement of intensity level are between the first preset threshold and the second preset threshold, and the coordination consistency is greater than the preset coordination threshold, it is determined that the intensive training period is entered, and the path parameters are adjusted to compound action training combination, medium frequency, moderate number of actions per set, medium intensity incremental gradient and regular training cycle. When the completion rate of intention execution and the rate of achievement of intensity level are greater than or equal to the second preset threshold, and the frequency of action correction is less than or equal to the preset correction threshold, it is determined that the training period is to be entered. The path parameters are set as comprehensive action training combination, high frequency, more number of times per set of actions, steep gradient of intensity and short training cycle. Each week, the updated rehabilitation assessment matrix of intention execution completion rate and intensity level achievement rate is extracted and compared with the corresponding training period threshold. If both key upgrade conditions are met at the same time, the path parameters are adjusted upward according to the gradient. If either intention execution completion rate or intensity level achievement rate falls back to the threshold range of the previous training period, the parameters of the previous training period are reverted. Simultaneously, the training plan and movement demonstration for the next day are updated through the AR device.

[0041] Specifically, the decoded data includes the degree of intention execution completion (overlap rate between actual action and preset trajectory) and the intensity level achievement rate (percentage of signal segments that meet the standard); the feedback response data includes the action correction frequency (number of corrections per unit time) and coordination consistency (time sequence matching degree between patient and exoskeleton); and the initial clinical data includes the muscle strength grade (MMT grade) and the range of motion of the joints (recorded according to medical measurement standards). The training period division criteria and personalized rehabilitation pathways are set as follows: Preset thresholds for basic / intensive / consolidation training periods, including a first preset threshold, a second preset threshold (first < second), a correction threshold, and a synergy threshold. Core indicators are compared with these thresholds: If the completion rate of intent execution and the rate of achieving the intensity level are ≤ the first threshold, and the movement correction frequency is > the correction threshold, the basic training period begins. Parameters are set as basic movement combinations, low frequency, fewer repetitions per set, a gradual increasing gradient, and a long cycle. If both indicators are between the two thresholds and the synergy consistency is > the synergy threshold, the intensive training period begins. Parameters are adjusted to compound movements, medium frequency and number of repetitions per set, a medium increasing gradient, and a regular cycle. If both indicators are ≥ the second threshold and the movement correction frequency is ≤ the correction threshold, the consolidation training period begins. Parameters are set as comprehensive movements, high frequency, more repetitions per set, a steep increasing gradient, and a short cycle. The rehabilitation assessment matrix is ​​updated weekly. If both key upgrade conditions (indicators meeting the standard and remaining stable for 2 weeks) are met simultaneously, parameters are increased; if any indicator declines, parameters are reverted. The training plan and movement demonstration for the next day are updated via an AR device.

[0042] Understandably, a rehabilitation assessment matrix is ​​constructed using multi-dimensional core indicators to accurately quantify patients' training performance and basic clinical status. Personalized pathway parameters are dynamically matched according to a three-tiered training phase: a foundational phase to solidify the basics, an intensive phase to achieve breakthroughs and improvements, and a consolidation phase to consolidate the results, adapting to the needs of different rehabilitation stages. Pathway parameters are adjusted bi-directionally based on weekly indicator updates, and the training plan is synchronized in real-time using an AR device, achieving precision, dynamism, and visualization of the rehabilitation program. This effectively improves training targeting and patient compliance, accelerating the recovery of upper limb motor function.

[0043] Specifically, the selection criteria for EEG signal segments that meet the quality inspection standards and whose intention execution meets the requirements include the signal quality inspection pass rate, the matching degree of intention execution, and the accuracy of action completion. The parameters of the initial EEG decoding model and feature template are updated using an incremental transfer learning algorithm, including: The system presets thresholds for the number of valid signal segments, a threshold for full compliance, and a threshold for partial compliance. It then compares the filtered signal segment judgment data with these preset thresholds. When the number of signal segments that pass the quality inspection is greater than or equal to the preset threshold, and the intention execution matching degree and action completion accuracy are both greater than or equal to the full compliance threshold, the EEG features of a single batch of signal segments are extracted as incremental training data and integrated into the original training dataset according to the first weight ratio. The network parameters of the initial EEG decoding model and the core feature dimension matching threshold of the feature template are updated through the incremental transfer learning algorithm. When the number of signal segments that pass the quality inspection is between 50% and 100% of the preset threshold, and the intention execution matching degree and action completion accuracy are greater than or equal to the partial pass threshold, the sub-signal segments in the batch of signal segments that fully meet the intention execution are extracted as incremental training data and integrated into the original training dataset according to the second weight ratio. Only the key layer parameters and feature template matching thresholds of the initial EEG decoding model are updated, and the core feature dimensions are retained. When the number of qualified signal segments is less than or equal to 50% of the preset threshold, or when the intention execution matching degree or action completion accuracy is less than the partial qualification threshold, the model parameters will not be updated on a large scale. Instead, the personalized adaptation unit will fine-tune the adaptation threshold and the fault tolerance range of the initial EEG decoding model and the evaluation matrix judgment standard based on the deviation data of the rehabilitation path, without changing the core parameters and feature template structure.

[0044] Specifically, the selection criteria for EEG signal segments that meet quality inspection standards and whose intended execution meets requirements include the signal quality inspection pass rate (the ratio of the number of qualified signal segments in a single batch to the total number of signal segments, where a single batch refers to the signal segments collected within a single 30-minute rehabilitation training cycle), the intention execution matching degree (the cosine similarity between the decoded action intention and the patient's actual motor imagination intention, calculated through the feature vectors of both), and the action completion accuracy (the mean positional error between the actual action trajectory and the preset rehabilitation action trajectory, in mm). An incremental transfer learning algorithm is used to update the parameters of the initial EEG decoding model and feature template, specifically including: a preset threshold for the number of valid signal segments (50 segments), a complete compliance threshold (0.9), and a partial compliance threshold (0.7). The selected signal segment judgment data is compared with the preset thresholds. When the number of qualified signal segments is ≥50, and the intention execution matching degree and action completion accuracy are both ≥0.9, the EEG features of that single batch of signal segments (including the temporal amplitude and frequency of μ waves and β waves) are extracted. Domain energy features) are used as incremental training data and integrated into the original training dataset at the first weight ratio (30%). Through the pre-training-fine-tuning mode of incremental transfer learning, the network full parameters and core feature dimension matching thresholds of the initial EEG decoding model are updated. When the number of qualified signal segments is between 25 and 49, and the intention execution matching degree and action completion accuracy are ≥0.7, the sub-signal segments in this batch that fully meet the intention execution are extracted as incremental data and integrated into the dataset at the second weight ratio (15%). Only the key layer parameters (fully connected layer, attention layer) and feature template matching thresholds of the model are updated, while the core feature dimensions related to μ waves and β waves are retained. When the number of qualified signal segments is ≤25, or any indicator is <0.7, large-scale updates are not initiated. Instead, the personalized adaptation unit integrated in the storage unit fine-tunes the model adaptation threshold and the fault tolerance range (±5%) of the evaluation matrix judgment standard based on the rehabilitation path execution deviation data through the gradient descent algorithm, without changing the core parameters and feature template structure.

[0045] Understandably, high-quality EEG signal segments are screened from multiple dimensions, including signal quality inspection pass rate, intent execution matching degree, and action completion accuracy, to ensure the high reliability of incremental training data and avoid invalid data interfering with model optimization. A scenario-based incremental transfer learning strategy is adopted: when data is sufficient, the model and template parameters are fully updated; when data is moderate, key layers are updated selectively; and when data is insufficient, only the adaptation threshold is fine-tuned. This balances model iteration efficiency and core structural stability while maximizing the use of effective data. Simultaneously, dynamic fine-tuning through personalized adaptation units ensures the model's real-time adaptability to individual patient rehabilitation progress. Ultimately, this achieves continuous optimization of the initial EEG decoding model and feature templates, significantly improving the accuracy of intent recognition and personalized adaptation capabilities, providing more reliable technical support for rehabilitation training, and contributing to continuous improvement in training effectiveness.

[0046] Specifically, before storing the updated initial EEG decoding model, adaptation parameters, and rehabilitation pathway into the storage unit, the effectiveness of the update needs to be verified by comparing historical data. The comparison criteria include model decoding accuracy, pathway execution compliance rate, and patient rehabilitation effect feedback data. The verification and final storage process includes: Preset accuracy deviation threshold and target achievement deviation threshold, calculate the first deviation value between the decoding accuracy of the updated model and the historical training best decoding accuracy, and the second deviation value between the execution achievement rate of the updated rehabilitation path and the target achievement rate of the preset rehabilitation goal. When the first deviation value is less than or equal to the preset accuracy deviation threshold and the second deviation value is less than or equal to the preset target achievement deviation threshold, the model and path update is deemed effective. The updated initial EEG decoding model, adaptation parameters and rehabilitation path are directly associated and stored in the storage unit, overwriting the original old data. If the first deviation value is greater than the preset accuracy deviation threshold or the second deviation value is greater than the preset pass rate deviation threshold, the update effect is determined to be unsatisfactory, and the parameter callback mechanism is activated. If only the first deviation value is greater than the preset accuracy deviation threshold, the updated feature template parameters are retained, the network layer parameters of the initial EEG decoding model are called back to the most recent valid update version, the decoding accuracy is recalculated, and the accuracy is stored after it meets the threshold. If the second deviation value is greater than the preset target achievement rate deviation threshold, the updated model parameters are retained, and the core parameters such as the intensity increment gradient of the rehabilitation path and the training frequency are called back to the previous version. After fine-tuning based on the patient's rehabilitation effect feedback data, the model is stored. If the two deviation values ​​are greater than the preset threshold, the system will completely revert to the previous version of the model, adaptation parameters, and recovery path, record the reason for the update failure, and not perform the storage overwrite operation.

[0047] Specifically, before storing the updated initial EEG decoding model, adaptation parameters, and rehabilitation pathway into the storage unit, the effectiveness of the update needs to be verified through historical data comparison. The comparison criteria include: model decoding accuracy (the percentage of correct decodings of the updated model on the EEG test set of 500 groups of patients), pathway execution target achievement rate (the percentage of preset rehabilitation actions completed and achieved by the patient when training according to the updated pathway), and patient rehabilitation effect feedback data (including clinical muscle strength grade improvement value, joint range of motion increase, and patient subjective training comfort score (1-5 points)). The verification and final storage process specifically includes: preset accuracy deviation threshold (≤5%), target achievement rate deviation threshold (≤10%), the optimal decoding accuracy in historical training is defined as the highest decoding accuracy achieved by the model in the patient's last 3 effective training cycles, and the preset rehabilitation target achievement rate is 85%. The first deviation value is calculated as |updated model decoding accuracy - optimal historical decoding accuracy|, and the second deviation value is calculated as |updated pathway execution target achievement rate - 85%|. When the first deviation value is ≤5% and the second deviation value is ≤10%, the update is deemed valid. The updated model, adaptation parameters, and rehabilitation path are associated and stored in the storage unit (associated with the update timestamp and verification result report), overwriting the old data. When either deviation value exceeds the limit, a parameter callback mechanism is activated: If only the first deviation value is >5%, the updated feature template parameters are retained, and the model network layer parameters (convolutional layer, fully connected layer) are callback to the most recent valid update version. The decoding accuracy must be ≥ the historical best value - 3% before storage. If only the second deviation value is >10%, the updated model parameters are retained, and the core parameters of the path are callback (intensity increment gradient is callback to 70%-80% of the original gradient, training frequency is callback to 80% of the original frequency). Based on feedback data (e.g., if muscle strength does not improve, the increment gradient is further reduced by 5%), and then fine-tuned (with an amplitude of ±10%) before storage. If both deviation values ​​exceed the limit, the system completely reverts to the version before the update, records the reason for failure (e.g., poor incremental data quality, excessive parameter optimization, etc.) and associated abnormal data identifiers, and does not perform the overwrite operation.

[0048] Understandably, the effectiveness of updates is validated across multiple dimensions, including model decoding accuracy, path execution compliance rate, and patient rehabilitation feedback, ensuring that updates do not deviate from the core rehabilitation goals. A deviation threshold is used as a benchmark to precisely control update quality, coupled with a scenario-based callback mechanism—retaining valid parameters when only the model or path fails to meet the standard, and completely reverting when both fail, thus avoiding ineffective updates that impact training results and minimizing the waste of effective optimization gains. Simultaneously, the reasons for update failures are recorded to provide a basis for subsequent targeted optimization. Ultimately, this ensures the reliability and adaptability of stored data, maintains the stability, continuity, and accuracy of rehabilitation training, and promotes the continuous and efficient iteration of the personalized rehabilitation system.

[0049] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0050] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0051] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0052] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for upper limb rehabilitation intention recognition and feedback training based on electroencephalogram (EEG) signal analysis, characterized in that, include: A dataset is constructed by integrating upper limb motor imagery EEG signals from different populations and screening the signal quality. The dataset contains effective EEG features corresponding to various upper limb movements. A feature extraction module and an intention classification unit are constructed. An initial EEG decoding model and feature template are trained and generated using a transfer learning algorithm. Patient clinical data are entered and the parameter adjustment threshold of the initial EEG decoding model is set. The determined initial EEG decoding model, feature template and clinical information are associated and stored in the storage unit. The signal acquisition unit is deployed to cover the corresponding scalp area of ​​the brain. After acquiring EEG signals, the signal quality is screened simultaneously. Valid signal segments that are synchronized with the motor imagination period and pass the quality inspection are input into the initial EEG decoding model to extract features and generate action intention vectors. The feature template is matched to decode information and information quality is checked. The time series data, clinical information and information quality inspection results are associated to form a decoding report. Configure a multimodal neurofeedback unit to output feedback information, deploy an exoskeleton rehabilitation device to adjust the assist force according to the decoding report to guide the execution of the movement, record the patient's feedback response behavior, extract qualified decoding data and response data and combine them with initial clinical information to construct a rehabilitation assessment matrix, set training period division standards, generate personalized rehabilitation paths and dynamically adjust parameters; Select EEG signal segments that meet the quality inspection standards and are intended to be executed in accordance with the requirements. Combine feedback response data and rehabilitation pathway execution status, use incremental transfer learning algorithm to update the parameters of the initial EEG decoding model and feature template. The personalized adaptation unit simultaneously corrects the adaptation threshold and evaluation matrix judgment criteria of the initial EEG decoding model. The updated initial EEG decoding model, adaptation parameters and rehabilitation pathway are stored in the storage unit.

2. The upper limb rehabilitation intention recognition feedback training method based on EEG signal analysis according to claim 1, characterized in that, An initial EEG decoding model and feature template were generated using a transfer learning algorithm, including: The upper limb movement type, EEG signal acquisition scenario, signal quality level, and EEG characteristic intensity level of different populations were determined as the model training feature group, and the injury type, muscle strength level, and joint range of motion in the patient's clinical data were determined as the model adaptation feature group. Calculate the similarity coefficients between the model training feature group, the model adaptation feature group, and the historical training feature group and historical adaptation feature group in the historical model library in the dataset; A preset similarity coefficient threshold is set to filter out all historical training feature groups and historical adaptation feature groups whose similarity coefficients are greater than the preset similarity coefficient threshold, and training selection feature groups and adaptation selection feature groups are constructed respectively. Based on the training and screening feature group, an effective subset of EEG features in the dataset is determined. Then, the EEG features of healthy individuals are transferred to the patient data training process using a transfer learning algorithm to train and generate an initial EEG decoding model. Based on the adaptive filtering feature group, the core feature dimensions and matching thresholds of the feature template are determined to form an initial feature template.

3. The upper limb rehabilitation intention recognition feedback training method based on EEG signal analysis according to claim 2, characterized in that, The process of determining an effective subset of EEG features from the dataset based on the training feature group and training an initial EEG decoding model, and determining the core feature dimensions and matching thresholds of the feature template based on the adaptation feature group, includes: The training feature set and the fitting feature set with the maximum similarity coefficient are obtained from the training feature set and the fitting feature set, respectively. If the maximum similarity coefficient training feature set contains only a single historical feature set, and the maximum similarity coefficient fitting feature set contains only a single historical feature set, then the EEG feature subset corresponding to the historical training feature set is directly used as the effective feature subset of the dataset, and the initial EEG decoding model is generated by combining the transfer learning algorithm; the feature dimension and threshold corresponding to the historical fitting feature set are directly used as the core feature dimension and matching threshold of the initial feature template. If there are multiple historical feature groups in the maximum similarity coefficient training feature group set and the maximum similarity coefficient fitting feature group set, then the EEG feature dimensions of each historical training feature group are aligned first, and then the weighted fusion value of the corresponding EEG feature is calculated using the similarity coefficient of each historical training feature group as the weight, forming an effective EEG feature subset of the dataset, and then the initial EEG decoding model is generated by combining the transfer learning algorithm. Simultaneously, using the similarity coefficient of each historical adaptation feature group as the weight, the weighted average of the corresponding feature dimension parameters and matching threshold is calculated as the core feature dimension and matching threshold of the initial feature template.

4. The upper limb rehabilitation intention recognition feedback training method based on EEG signal analysis according to claim 3, characterized in that, The data used for signal quality screening include the signal-to-noise ratio, artifact persistence rate, and feature integrity of the EEG signal. The feature completeness refers to the coverage ratio of effective EEG feature dimensions; Preset quality standard data to determine whether the EEG signal passes quality inspection, including: When comparing the signal quality judgment data with the preset quality standard data, if the signal-to-noise ratio is greater than or equal to the preset signal-to-noise ratio threshold, the artifact residue rate is less than or equal to the preset residue rate threshold, and the feature integrity is greater than or equal to the preset integrity threshold, then the EEG signal is judged to be qualified for quality inspection and retained as an effective signal segment for synchronization during the motor imagery period. When the signal-to-noise ratio is less than a preset signal-to-noise ratio threshold, the artifact persistence rate is greater than a preset persistence rate threshold, or the feature integrity is less than a preset integrity threshold, the EEG signal is determined to be unqualified for quality inspection. The signal segment is removed, the reason for the abnormality is recorded, and the corresponding acquisition time sequence information is marked synchronously.

5. The upper limb rehabilitation intention recognition feedback training method based on EEG signal analysis according to claim 1, characterized in that, The information quality inspection criteria include the type matching degree, strength stability, and completion reliability of the decoded information. Preset information quality standard data to determine whether the decoded information passes quality inspection, including: When the type matching degree is greater than or equal to the preset matching degree threshold, the strength stability is greater than or equal to the preset stability threshold, and the completion reliability is greater than or equal to the preset reliability threshold, the decoded information is determined to be qualified for quality inspection, and is directly included in the decoding report and associated with the corresponding time series data and clinical information; If the type matching degree is less than a preset matching degree threshold, the strength stability is less than a preset stability threshold, or the completion reliability is less than a preset reliability threshold, then the decoded information is determined to be unqualified, and a second decoding process is initiated. If the information quality inspection meets the standards after secondary decoding, the secondary decoding result will be included in the decoding report and a secondary correction record will be marked. If the standard is still not met after secondary decoding, the decoding information is discarded, the reason for the decoding abnormality is recorded, the corresponding valid signal segment and acquisition timing information are marked simultaneously, and it is not included in the subsequent rehabilitation assessment data.

6. The upper limb rehabilitation intention recognition feedback training method based on EEG signal analysis according to claim 1, characterized in that, The feedback types of the multimodal neural feedback unit include visual feedback, auditory feedback and tactile feedback, and the feedback intensity is positively correlated with the intention execution completion degree and intensity stability in the decoding report; The dynamic adjustment process for the output feedback information includes: Two preset feedback intensity thresholds, high and medium, are used to compare the intent execution completion rate and intensity stability in the decoding report with the preset thresholds: When the completion rate and intensity stability of the intention execution are both greater than or equal to the advanced feedback threshold, the three-level collaborative feedback mode is activated: visual feedback displays the complete motion trajectory animation, auditory feedback plays high-frequency excitation sound effects, and tactile feedback outputs strong vibration prompts. When the completion rate and intensity stability of the intent execution are both between the medium and high feedback thresholds, the second-level collaborative feedback mode is activated, retaining the combination of visual and auditory feedback, adjusting the sound effect frequency and animation clarity to a medium level, and turning off haptic feedback. When the completion rate and intensity stability of the intention are both less than or equal to the intermediate feedback threshold, the basic feedback mode is activated, retaining only visual feedback, displaying simplified action guidance icons, and simultaneously reducing the frequency of feedback triggering to avoid overstimulation. The system collects real-time data on the patient's attention concentration during feedback responses. If the attention concentration is less than the preset focus threshold, the system automatically increases the current feedback intensity level by one gradient, and then returns to the original level after three action cycles.

7. The upper limb rehabilitation intention recognition feedback training method based on EEG signal analysis according to claim 2, characterized in that, The adjustment of the assist force of the exoskeleton rehabilitation device is based on the intensity level in the decoding report, the muscle strength level in the patient's clinical data, and the real-time movement execution deviation. The precise control of the adjustment of the assist force includes: A preset muscle strength matching coefficient table and deviation compensation coefficient are used. The basic assistance coefficient is matched from the matching coefficient table according to the patient's muscle strength level, and the initial assistance force is determined by combining the intensity level in the decoding report. Real-time acquisition of position and speed deviation data during the execution of the action, and calculation of the comprehensive deviation value: when the comprehensive deviation value is less than or equal to the preset deviation threshold, the initial assist force remains unchanged; When the overall deviation value is greater than the preset deviation threshold but less than or equal to twice the preset deviation threshold, the assist force is increased by 10%-30% according to the deviation compensation coefficient, with a focus on compensating for joints with greater resistance to the action. When the overall deviation value exceeds twice the preset deviation threshold, the current motion guidance is paused, deviation prompts are sent through the multimodal neurofeedback unit, and the assistance intensity is reduced to 50% of the base value. Motion guidance is restarted after the patient adjusts their posture, and abnormal deviation data is recorded simultaneously for subsequent rehabilitation pathway optimization.

8. The upper limb rehabilitation intention recognition feedback training method based on EEG signal analysis according to claim 1, characterized in that, The core indicators of the rehabilitation assessment matrix include the intention execution completion rate and intensity level achievement rate in the decoded data, the movement correction frequency and coordination consistency in the feedback response data, and the muscle strength level and joint range of motion in the initial clinical data. The process of setting training period division criteria and generating personalized rehabilitation pathways includes: The system presets threshold values ​​for the basic training period, intensive training period, and consolidation training period, including a first preset threshold, a second preset threshold, a correction threshold, and a synergistic threshold. The first preset threshold is lower than the second preset threshold. The core indicators of the rehabilitation assessment matrix are then compared with these preset thresholds. When the completion rate of intention execution and the rate of achievement of intensity level are less than or equal to the first preset threshold, and the frequency of action correction is greater than the preset correction threshold, it is determined that the basic training period has been entered. The path parameters are set as basic action training combination, low frequency, fewer number of times per set of actions, gentle intensity increase gradient and long training cycle. When the completion rate of intent execution and the rate of achievement of intensity level are between the first preset threshold and the second preset threshold, and the coordination consistency is greater than the preset coordination threshold, it is determined that the intensive training period is entered, and the path parameters are adjusted to compound action training combination, medium frequency, moderate number of actions per set, medium intensity incremental gradient and regular training cycle. When the completion rate of intention execution and the rate of achievement of intensity level are greater than or equal to the second preset threshold, and the frequency of action correction is less than or equal to the preset correction threshold, it is determined that the training period is to be entered. The path parameters are set as comprehensive action training combination, high frequency, more number of times per set of actions, steep gradient of intensity and short training cycle. Each week, the updated rehabilitation assessment matrix of intention execution completion rate and intensity level achievement rate is extracted and compared with the corresponding training period threshold. If both key upgrade conditions are met at the same time, the path parameters are adjusted upward according to the gradient. If either intention execution completion rate or intensity level achievement rate falls back to the threshold range of the previous training period, the parameters of the previous training period are reverted. Simultaneously, the training plan and movement demonstration for the next day are updated through the AR device.

9. The upper limb rehabilitation intention recognition feedback training method based on EEG signal analysis according to claim 1, characterized in that, The criteria for selecting EEG signal segments that meet the quality inspection standards and whose intended execution is in line with the requirements include the signal quality inspection pass rate, the degree of matching of intention execution, and the accuracy of action completion. The method of updating the parameters of the initial EEG decoding model and feature template using incremental transfer learning algorithm includes: The system presets thresholds for the number of valid signal segments, a threshold for full compliance, and a threshold for partial compliance. It then compares the filtered signal segment judgment data with these preset thresholds. When the number of signal segments that pass the quality inspection is greater than or equal to the preset threshold, and the intention execution matching degree and action completion accuracy are both greater than or equal to the full compliance threshold, the EEG features of a single batch of signal segments are extracted as incremental training data and integrated into the original training dataset according to the first weight ratio. The network parameters of the initial EEG decoding model and the core feature dimension matching threshold of the feature template are updated through the incremental transfer learning algorithm. When the number of signal segments that pass the quality inspection is between 50% and 100% of the preset number threshold, and the intention execution matching degree and action completion accuracy are greater than or equal to the partial pass threshold, the sub-signal segments in the batch signal segments that fully meet the intention execution are extracted as incremental training data and integrated into the original training dataset according to the second weight ratio. Only the key layer parameters and feature template matching thresholds of the initial EEG decoding model are updated, and the core feature dimensions are retained. When the number of qualified signal segments is less than or equal to 50% of the preset threshold, or when the intention execution matching degree or action completion accuracy is less than the partial qualification threshold, the model parameters will not be updated on a large scale. Instead, the personalized adaptation unit will fine-tune the adaptation threshold and the fault tolerance range of the initial EEG decoding model and the evaluation matrix judgment standard based on the deviation data of the rehabilitation path, without changing the core parameters and feature template structure.

10. The upper limb rehabilitation intention recognition feedback training method based on EEG signal analysis according to claim 9, characterized in that, Before storing the updated initial EEG decoding model, adaptation parameters and rehabilitation path into the storage unit, the effectiveness of the update needs to be verified by comparing with historical data. The comparison criteria include model decoding accuracy, path execution compliance rate and patient rehabilitation effect feedback data. The verification and final storage process includes: Preset accuracy deviation threshold and target achievement deviation threshold, calculate the first deviation value between the decoding accuracy of the updated model and the historical training best decoding accuracy, and the second deviation value between the execution achievement rate of the updated rehabilitation path and the target achievement rate of the preset rehabilitation goal. When the first deviation value is less than or equal to the preset accuracy deviation threshold and the second deviation value is less than or equal to the preset target achievement deviation threshold, the model and path update is deemed effective. The updated initial EEG decoding model, adaptation parameters and rehabilitation path are directly associated and stored in the storage unit, overwriting the original old data. If the first deviation value is greater than the preset accuracy deviation threshold or the second deviation value is greater than the preset pass rate deviation threshold, the update effect is determined to be unsatisfactory, and the parameter callback mechanism is activated. If only the first deviation value is greater than the preset accuracy deviation threshold, the updated feature template parameters are retained, the network layer parameters of the initial EEG decoding model are called back to the most recent valid update version, the decoding accuracy is recalculated, and the accuracy is stored after it meets the threshold. If the second deviation value is greater than the preset target achievement rate deviation threshold, the updated model parameters are retained, and the core parameters such as the intensity increment gradient of the rehabilitation path and the training frequency are called back to the previous version. After fine-tuning based on the patient's rehabilitation effect feedback data, the model is stored. If the two deviation values ​​are greater than the preset threshold, the system will completely revert to the previous version of the model, adaptation parameters, and recovery path, record the reason for the update failure, and not perform the storage overwrite operation.

Images