A biosignal pattern recognition method and system for rehabilitation training

Abstract

The application provides a biological signal pattern recognition method and system for rehabilitation training, and relates to the technical field of rehabilitation medicine. The technical solution points are as follows: features are extracted from the collected biological signals according to continuous time windows, the closeness of each time window feature to a target action reference and a compensatory action reference is calculated respectively, the boundary compression degree is determined by combining the category change rate of the continuous time windows and the closeness difference, when the boundary compression degree exceeds a preset threshold, only local rejudgment is performed between the target action, rest and compensatory interference, and a trigger instruction or a temporary trigger instruction is output according to the rejudgment result, so that fine processing is performed when the boundary is fuzzy, and misjudgment is avoided; the target action reference and the compensatory action reference are updated according to the effective target features and the compensatory interference features confirmed in the training, the adaptive update of the action reference is realized, and the differentiation degree of the target action and the compensatory action is prevented from being reduced due to the update.

Description

Technical Field

[0001] This application relates to the field of rehabilitation medical technology, and more specifically, to a biosignal pattern recognition method and system for rehabilitation training. Background Technology

[0002] Upper limb function recovery training for stroke patients is a crucial aspect of rehabilitation medicine. In the training area of ​​a hospital's rehabilitation department, patients typically recover function by repeatedly performing lifting, flexion, extension, and grasping movements under the guidance of therapists. To support this process, rehabilitation training stations are generally equipped with biosignal pattern recognition systems. These systems use surface electromyography (sEMG) signals as the primary input. Through acquisition, processing, recognition, control, and recording, they convert the patient's muscle activation patterns into recognizable movement intentions, thereby triggering rehabilitation equipment to provide assistance or feedback. Specifically, the system collects weak bioelectrical signals by attaching electrodes to the patient's target muscle groups (such as the deltoid and biceps brachii), amplifies them, and sends them to the host computer for filtering, feature extraction, and pattern recognition. The recognition results indicate the current movement intention or muscle activation pattern. Based on this, the control program sends instructions to the display terminal or training equipment to trigger animated feedback or drive rehabilitation equipment such as exoskeletons to perform corresponding movements, while simultaneously recording training data.

[0003] Existing biosignal pattern recognition systems typically follow a set procedure before each training session. First, the therapist verifies the patient's identity and retrieves historical training data, including the previously used electrode channel location, target action category, and recognition parameters. Then, the patient takes their place, undergoes skin cleaning, electrode attachment, and lead connection. After system startup, baseline signal readings and quality checks are performed on each channel. Before formal training, the system requires calibration, where the therapist guides the patient through pre-defined target actions, collecting biosignal samples. These raw signals are processed through filtering and feature extraction (such as root mean square value and mean absolute value) before being fed into a pre-defined template matching program or classification model to establish a correspondence between the patient's current action category and signal features. If the patient's physiological state changes significantly, the therapist must intervene manually, requesting the system to resample and generate recognition parameters for the day. After calibration, the therapist sets the training task, including action category, number of repetitions, trigger threshold, and assistance intensity. The trigger threshold is usually set based on the calibration results to determine the valid action intent.

[0004] Once training officially begins, the biosignal acquisition device continuously collects data. The host computer processes the signals continuously according to a preset time window, extracts features, performs pattern recognition, and outputs the action category and reliability. If the recognition result is the target action and the activation intensity reaches a preset threshold, the control program triggers the rehabilitation device to execute the corresponding action and provides feedback; if it is a relaxation or incorrect action, the device remains inactive. However, the patient's physiological state is not constant within the same training day. Signal waveforms and patterns may change due to slow initiation at the beginning of training, clear activation patterns in the middle, and accumulated fatigue or increased sweating in the later stages. In this situation, existing systems typically rely on therapists to manually monitor recognition performance during rest periods between sets. If false triggers increase or signal amplitude changes, training must be interrupted, signals re-acquired, and recognition thresholds or parameters adjusted. After training, the system archives the training information for future use.

[0005] However, existing biosignal pattern recognition systems face significant challenges during long-term or multi-day training. The physiological state of stroke patients is dynamic; as training progresses, target muscle groups may experience fatigue or undergo physiological adaptations to complete tasks. Initially sporadic compensatory muscle activities (such as shoulder shrugs, trunk synergistic contractions, and excessive involvement of non-target muscle groups) may gradually evolve into normalized force patterns, leading to a continuous expansion of the range of biosignal features generated by non-target muscle groups. In this situation, these evolved compensatory features begin to overlap with the feature templates of the target movement, blurring or even squeezing the originally clear boundaries between movement categories. Existing systems primarily rely on matching the signal features of the current time window with preset static movement templates. Once the signal simultaneously contains both target movement components and significant compensatory components, the recognition results frequently jump between multiple unrelated movement categories, or misclassify compensatory movements as valid target movements and trigger the device. This instability and mis-triggering not only prevent the system from accurately matching the patient's changing physiological state but also force therapists to frequently interrupt training for manual recalibration or threshold adjustments. This not only increases the number of steps and processing time for manual intervention, but also makes it easy to make judgment errors during the adjustment intervals, which seriously affects the continuity of the training process, the timeliness of response, and the patient's training experience.

[0006] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0007] The purpose of this application is to provide a biosignal pattern recognition method and system for rehabilitation training, which has the advantages of being able to dynamically adapt to changes in the patient's physiological state, effectively distinguish between target actions and compensatory interference, reduce misjudgments, improve the accuracy and continuity of rehabilitation training, thereby enhancing the patient's training experience and rehabilitation effect.

[0008] This application provides a biosignal pattern recognition method for rehabilitation training, the technical solution of which is as follows: Features are extracted from the collected biological signals according to continuous time windows. The proximity between the features of each time window and the target action reference and the compensatory action reference is calculated respectively. The degree of boundary compression is determined by combining the category change rate of the continuous time window and the proximity difference. When the degree of boundary compression exceeds the preset threshold, local re-judgment is performed only among the target action, rest and compensatory interference, and a trigger command or a delay trigger command is output according to the re-judgment result. The target action reference and the compensation action reference are updated according to the effective target features and compensation interference features confirmed during training. If the distance between the two after the update is less than the lower limit of the safe distance, the update of the target action reference is cancelled.

[0009] Furthermore, this application also proposes that the calculation of the proximity between the features of each time window and the target action reference and the compensatory action reference includes: Based on the dynamic weights of each biosignal acquisition channel, the weighted distances between the current time window features and the target action reference and the compensatory action reference are calculated respectively. Extract the target channel proportion, compensatory channel proportion, and common contraction index representing the contraction intensity of the antagonistic muscle corresponding to the current time window features; Based on the weighted distance, the target channel proportion, the compensating channel proportion, and the common contraction index, the proximity of the current time window features to the target action reference and the compensating action reference is determined.

[0010] Furthermore, this application also proposes that, before calculating the weighted distance between the current time window features and the target action reference and the compensatory action reference based on the dynamic weights of each biosignal acquisition channel, the method further includes: Acquire the historical stability, signal-to-noise ratio of the day, contact quality, and resting baseline offset of each acquisition channel; The initial weights of each acquisition channel are calculated based on the historical stability, the signal-to-noise ratio of the current day, the contact quality, and the degree of resting baseline offset. During the feature extraction process, when it is detected that the signal variance of an acquisition channel increases or the noise rises beyond a set value, the weight of that acquisition channel is reduced to form the dynamic weight of each biological signal acquisition channel.

[0011] Furthermore, this application also proposes that the step of performing local re-judgment only among the target action, rest, and compensatory interference, and outputting a trigger command or a delayed trigger command based on the re-judgment result, includes: Reduce the dynamic weights of acquisition channels that have shown compensatory increase characteristics, and extract multi-window continuity scores; The target score is calculated based on the proximity of the current time window to the target action reference, the target channel ratio, and the multi-window continuity score. The compensation score is calculated based on the proximity of the current time window to the compensation action reference, the proportion of the compensation channel, and the common contraction index. Compare the target score with the compensation score, and when the target score is higher than the compensation score by a preset difference, output the corresponding trigger command; When the compensation score is higher than the target score, a delay trigger command for the corresponding compensation interference is output and the system enters a pending confirmation state.

[0012] Furthermore, this application also proposes that, after the output of the delayed triggering instruction corresponding to the compensatory interference, it further includes: Prohibit sending normal mechanical assistive movement commands and control the rehabilitation equipment to perform one of the following actions: hold the image while waiting, reduce the level of assistance, or output movement correction prompts; When multiple consecutive time windows are in the pending confirmation state during the same action attempt process, the corresponding attempt rounds will be recorded as compensation enhancement rounds; The time window features corresponding to the compensation enhancement rounds are included in the update set of the compensation action reference.

[0013] Furthermore, this application also proposes that the method further includes: Continuously monitor system status across multiple update cycles; When any of the following conditions are met: the minimum number of valid target features are not obtained in a consecutive preset update cycle, the proportion of the same training group entering the pending confirmation state continues to exceed the upper limit, or the distance between the updated target action reference and the compensatory action reference is continuously lower than the lower limit of the safe distance, a manual intervention prompt carrying a restricted action identifier is generated. Output the aforementioned manual intervention prompt to instruct the acquisition of a short-term calibration signal on a specified acquisition channel or action.

[0014] Furthermore, this application also proposes that the method further includes: The current fatigue score is calculated by combining the median frequency and envelope amplitude of the biological signal; When the current fatigue score increases, and the proportion of the compensatory channel and the common contraction index do not exceed the preset level, the lower limit threshold for judging the target action reference is relaxed. When the current fatigue score increases, and the proportion of the compensation channel or the common contraction index increases simultaneously beyond a preset level, the upper limit threshold for judging the compensation action reference is tightened.

[0015] Furthermore, this application also proposes that updating the target action reference and the compensatory action reference based on the effective target features and compensatory interference features confirmed during training includes: When updating the action references separately, the update rate for the target action reference is set to be less than the update rate for the compensatory action reference. Extract and save the distribution drift vector between the mean of the trusted target window in the first half and the mean of the trusted target window in the second half of the current training round; During cross-group or cross-day training handover, if it is detected that the new round of initial signal distribution is close to the compensating action reference according to the distribution drift vector, or if it is detected that the distance between the target action reference and the compensating action reference is reduced, then the restricted output mode is entered. In the restricted output mode, only the specified target category and the compensatory interference category are retained for identification until a sufficient number of valid target features are reacquired before full multi-class identification is restored.

[0016] Furthermore, this application also proposes that, in the restricted output mode, only the specified target category and the compensatory interference category are retained for identification until a sufficient number of valid target features are reacquired before restoring complete multi-class identification, including: Obtain the number of times each category not involved in the identification process and the currently specified target category jointly enter the local re-judgment before entering the restricted output mode; The recovery order for each category not involved in the identification is determined according to the number of times the identification was performed, from least to most. After reacquiring the required number of valid target features, if the degree of boundary compression and the number of local re-judgments corresponding to the current identification range are both lower than the corresponding thresholds, a category that did not participate in the identification is added to the current identification range according to the recovery order.

[0017] Furthermore, this application also proposes, and includes: After each time a non-identified category is added to the current identification range, the degree of boundary compression, the category change rate of the continuous time window, and the number of local re-judgments are obtained within the preset observation window; When at least one of the boundary compression degree, the category change rate of the continuous time window, and the number of local re-judgments exceeds the corresponding threshold, the non-identified category is removed from the current identification range, and the identification range before its inclusion is maintained. When the degree of boundary compression, the rate of change of category in the continuous time window, and the number of local re-judgments do not exceed the corresponding thresholds, one category that was not identified is retained, and the next category that was not identified is added according to the recovery order.

[0018] Furthermore, this application also proposes, and includes: After each time a non-identified category is added to the current identification range, the degree of boundary compression, the category change rate of the continuous time window, and the number of local re-judgments are obtained within the preset observation window; When at least one of the boundary compression degree, the category change rate of the continuous time window, and the number of local re-judgments exceeds the corresponding threshold, the non-identified category is removed from the current identification range, and the identification range before its inclusion is maintained. When the degree of boundary compression, the rate of change of category in the continuous time window, and the number of local re-judgments do not exceed the corresponding thresholds, one category that was not identified is retained, and the next category that was not identified is added according to the recovery order.

[0019] Furthermore, this application also proposes a biosignal pattern recognition system for rehabilitation training, used to perform the above-described method, comprising: The first module is used to extract features from the collected biological signals according to continuous time windows, calculate the proximity of the features of each time window to the target action reference and the compensatory action reference, and determine the degree of boundary compression by combining the category change rate and proximity difference of the continuous time window. The second module is used to perform local re-judgment only among target action, rest and compensatory interference when the degree of boundary compression exceeds the preset threshold, and output trigger command or postpone trigger command according to the re-judgment result. The third module is used to update the target action reference and the compensation action reference according to the valid target features and compensation interference features confirmed in the training. If the distance between the two after the update is less than the lower limit of the safe distance, the update of the target action reference is cancelled.

[0020] The above solution provides a system capable of implementing the aforementioned methods, offering hardware support for rehabilitation training and improving the level of intelligence in rehabilitation training.

[0021] As can be seen from the above, the biosignal pattern recognition method and system provided in this application for rehabilitation training extracts features from the collected biosignals according to continuous time windows, calculates the proximity of each time window feature to the target action reference and the compensatory action reference, and determines the degree of boundary compression by combining the category change rate and proximity difference of the continuous time window, which can dynamically evaluate the clarity of the recognition boundary; when the degree of boundary compression exceeds a preset threshold, local re-judgment is performed only between the target action, resting state and compensatory interference, and a trigger command or a delayed trigger command is output according to the re-judgment result, thereby performing refined processing when the recognition boundary is blurred and avoiding misjudgment; based on the effective results confirmed during training... The target feature and compensatory interference feature update the target action reference and compensatory action reference respectively. If the distance between the two after the update is less than the lower limit of the safe distance, the update of the target action reference is cancelled, thus realizing the adaptive update of the action reference. At the same time, it prevents the update from reducing the distinction between the target and the compensatory action. It effectively solves the problems of unstable recognition, frequent false triggers, and frequent manual intervention required by the existing biosignal pattern recognition system when the patient's physiological state changes dynamically. It has the advantages of being able to dynamically adapt to changes in the patient's physiological state, effectively distinguish between the target action and the compensatory interference, reduce misjudgment, improve the accuracy and coherence of rehabilitation training, and thus improve the patient's training experience and rehabilitation effect. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a biosignal pattern recognition method for rehabilitation training provided in this application.

[0023] Figure 2 This is a schematic diagram of the structure of a biosignal pattern recognition system for rehabilitation training provided in this application.

[0024] In the diagram: 1. First module; 2. Second module; 3. Third module. Detailed Implementation

[0025] The technical solutions of this application will be clearly and completely described below with reference to the embodiments thereof. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0026] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0027] In the context of upper limb function recovery training after stroke in the rehabilitation department of a hospital, patients, under guidance, use rehabilitation equipment equipped with biosignal acquisition and recognition systems to perform repetitive functional training. These systems typically use surface electromyography (EMG) signals as input and trigger equipment assistance or provide feedback by recognizing the patient's intention to move. In the early stages of training, the patient's physiological state is relatively stable, and the system can effectively distinguish between target movements and background noise based on the movement templates and judgment thresholds established during initial calibration. However, as training continues, especially in the middle and later stages of a single training session or during training over several consecutive days, the patient's physiological state will change.

[0028] A common problem is the evolution of compensatory behavior. Initially, patients may only exhibit sporadic compensatory muscle activity to complete a movement, such as slightly shrugging their shoulders when attempting to raise their arms. These weak signals from non-target muscle groups are still within the system's preset discrimination limits and are considered tolerable interference. However, as the target muscle groups become more fatigued, or as patients develop new physiological adaptations to achieve training tasks, the originally sporadic compensatory behavior gradually solidifies into a normalized force exertion pattern. At this point, the range of biosignal characteristics generated by non-target muscle groups continues to expand, and they begin to overlap with the characteristics of the target movement in the feature space. This phenomenon can be understood as the originally clear classification boundary being continuously squeezed by compensatory signals, leading to a decrease in discrimination.

[0029] This boundary compression directly leads to the instability of recognition performance. When a patient attempts to perform a single target action, the recognition result may frequently jump between multiple unrelated action categories because the signal contains both target and compensatory components. Or, due to interference from compensatory signals, there may be significant recognition lag and false triggering. This makes it impossible for the recognition to accurately match the patient's current changed physiological state. The therapist has to frequently interrupt the training to manually recalibrate or adjust the threshold. This not only increases the operational burden and prolongs the processing time, but also disrupts the continuity of the training process and the timeliness of the response.

[0030] To address the challenges of avoiding inconsistent or false triggers in recognition results, reducing frequent manual recalibration, and maintaining the continuity and timeliness of the training process in situations where patients' compensatory behaviors become increasingly normalized, signal features overlap with multiple action templates, and classification boundaries are squeezed, this application provides a biosignal pattern recognition method for rehabilitation training. The core of this method lies in adding a dynamic discrimination and adaptive update mechanism oriented towards compensatory evolution between traditional feature extraction and final category output. This method first assesses the relative intensity of the target action component and the compensatory action component in the current signal, and based on this, determines the stable state of the classification boundary. Then, it decides whether to output directly, enter local re-judgment, postpone triggering, or perform a separate update of the judgment reference.

[0031] Specifically, refer to Figure 1 This application provides a biosignal pattern recognition method for rehabilitation training, comprising the following steps: S1. Extract features from the collected biological signals according to continuous time windows, calculate the proximity of each time window feature to the target action reference and the compensatory action reference, and determine the degree of boundary compression by combining the category change rate and proximity difference of the continuous time window. S2. When the degree of boundary compression exceeds the preset threshold, local re-judgment is performed only among the target action, rest and compensatory interference, and a trigger command or a delay trigger command is output according to the re-judgment result. S3. Update the target action reference and the compensation action reference according to the effective target features and compensation interference features confirmed in the training. If the distance between the two after the update is less than the lower limit of the safe distance, then cancel the update of the target action reference.

[0032] Specifically, this method can be deployed in the host computer of the upper limb function recovery station in the hospital's rehabilitation department. The signal acquisition part is responsible for synchronously acquiring surface electromyographic signals from the target muscle groups and common compensatory muscle groups of the patient's upper limb. For example, when performing elbow flexion training, target channel electrodes can be placed on the main elbow flexor muscle groups such as the biceps brachii and brachioradialis, while compensatory monitoring channel electrodes can be placed on the upper trapezius, anterior deltoid, and the triceps brachii as an antagonist muscle. This multi-channel configuration allows for the simultaneous observation of the main driving source of the target movement and the common manifestations of compensatory behavior.

[0033] After signal acquisition, a continuous stream of biological signal data is received at a fixed sampling frequency, such as 1,000 times per second. The received raw signal is first preprocessed, including bandpass filtering to remove motion artifacts and high-frequency noise, power frequency notch filtering to suppress power line interference, and full-wave rectification and envelope extraction. In order to achieve a balance between real-time performance and stability, the processing adopts a sliding time window approach. For example, the time window length is set to 200 milliseconds, and it slides once every 50 milliseconds to form a continuous analysis unit. For each time window, a set of feature vectors is extracted, including the root mean square value and the average absolute value reflecting the signal strength, as well as waveform length and zero-crossing count that can characterize muscle activity patterns.

[0034] Next, it is necessary to establish benchmarks for comparison, namely the target action reference and the compensatory action reference. These two references can be dynamically updated. Specifically, they can be the center point of a feature space, i.e., the mean vector. The target action reference represents the signal feature distribution of the patient at the current stage, which has been identified as having less compensation and can truly reflect the patient's action intention. The compensatory action reference represents the signal feature distribution of non-standard force patterns such as shoulder shrugging and abnormal co-contraction that have been identified in previous training.

[0035] For each new time window feature, a bidirectional comparison is performed, that is, the proximity of the time window feature to the target action reference and the compensatory action reference corresponding to the current task are calculated separately. The proximity calculation can be based on some distance metric between feature vectors; the smaller the distance, the higher the proximity. Through this bidirectional comparison, the degree to which the current signal simultaneously contains target action components and compensatory action components can be quantified.

[0036] Subsequently, the degree of boundary compression was determined to assess the stability of the current recognition boundary. This indicator was calculated by combining two aspects of information: the rate of change of categories within a continuous time window and the proximity difference. The rate of change of categories refers to the frequency with which the output of the original basic classification jumps between different action categories within a preset short time period. The output of the original basic classification refers to the initial judgment label of a relatively traditional multi-class action recognizer for each time window. A high rate of change indicates that the recognition result is extremely unstable. The proximity difference reflects the relative distance between the current time window feature and the target action reference and the compensatory action reference. When a time window feature is close to both the target reference and the compensatory reference, its proximity difference will be very small, indicating that the signal is in a blurry and difficult-to-distinguish region. Combining the rate of change of categories and the proximity difference can effectively quantify the compression of the classification boundary by compensatory behavior, thereby distinguishing between instantaneous signal fluctuations and continuous changes in patient status.

[0037] When the calculated boundary compression exceeds a preset threshold, it indicates that the current recognition environment is unstable, and traditional full-class classification is prone to errors. At this time, a local reclassification mechanism is triggered. Local reclassification narrows the candidate set to the most relevant items in the current context, namely the target action, resting state, and compensatory interference. When the boundary compression exceeds the preset threshold, the action category outputs that are irrelevant to the current training task in the basic multi-class discrimination will be blocked. For example, when training a unidirectional lift, the candidate set is strictly limited to three categories: lifting action, resting state, and shoulder shrug compensation interference.

[0038] If the reassessment result shows that the target action component in the current signal still dominates, a trigger command will be output to drive the rehabilitation device to perform the corresponding auxiliary action or provide positive feedback. If the reassessment result shows that the compensatory interference component is too strong, a delayed trigger command will be output. This is a buffer state, in which the rehabilitation device may enter a waiting state, reduce the level of assistance, or provide correction prompts, giving the patient an opportunity to adjust and correct their movements, while avoiding the reinforcement of incorrect movement patterns.

[0039] Finally, during the training process, the time window features are classified according to the actual interaction results. Those time window features that successfully trigger the device, receive positive feedback, and are not rejected by the therapist are identified as valid target features, while those time window features that are blocked by the local re-judgment mechanism or identified as having excessive compensatory components are identified as compensatory interference features. The system will use the set of valid target features to update the target action reference and the set of compensatory interference features to update the compensatory action reference.

[0040] To prevent compensatory feature contamination—that is, to prevent the system from learning compensatory actions as correct actions as it trains—a risk control mechanism is implemented. After each update, the distance between the new target action reference and the compensatory action reference is calculated. If this distance is less than a preset safe distance lower limit, it means that the two are too close in the feature space, and the target model may have been severely eroded by compensatory components. In this case, the update to the target action reference is revoked, and only the update to the compensatory action reference is retained.

[0041] In some specific implementations, regarding the mechanism for revoking target action reference updates, after each update of the target action reference using valid target features, the Euclidean distance between the updated target action reference center and the current compensatory action reference center in the multidimensional feature space is calculated. The lower limit of the safe distance is based on the initial distance between the target action and the compensatory action during the patient's initial calibration, typically set to 70% of the initial distance. If the calculated current distance is less than this lower limit of the safe distance, it indicates that the valid target features included this time contain too many compensatory components, causing the target model to shift towards the compensatory model. In this case, the newly calculated target action reference is discarded, and the parameters of the target action reference are rolled back to their state before the update, while the updates to the compensatory action reference are retained to ensure the purity of the target model.

[0042] Through the above scheme, this method can intelligently address the normalization of compensatory behaviors caused by factors such as fatigue in patients. When ambiguity and uncertainty occur, it avoids misjudgment and misoperation by local re-judgment and deferred triggering. Through a separate reference update mechanism, it continuously learns the changes in the patient's state while adhering to the correct training direction. This greatly reduces the reliance on manual intervention and improves the coherence, safety and effectiveness of the training process.

[0043] In one specific implementation, the process of calculating the proximity between each time window feature and the target action reference and the compensatory action reference includes: calculating the weighted distance between the current time window feature and the target action reference and the compensatory action reference according to the dynamic weight of each biosignal acquisition channel; extracting the target channel proportion, the compensatory channel proportion, and the common contraction index characterizing the antagonistic muscle contraction intensity corresponding to the current time window feature; and determining the proximity between the current time window feature and the target action reference and the compensatory action reference based on the weighted distance, the target channel proportion, the compensatory channel proportion, and the common contraction index.

[0044] The concept of dynamic weighting is introduced when calculating distance, meaning that not all acquisition channels have equal importance in distance calculation. Different weights are assigned to each channel based on its real-time reliability. A channel with high reliability will have a greater weight in determining the final distance, and vice versa. The weighted distance can be calculated in the form of weighted Euclidean distance. For example, for a target reference, the weighted distance can be expressed as the difference between the feature vector of the current time window and the mean vector of the target reference. The norm is transformed by a diagonal weight matrix composed of the weights of each channel. This weighting method can effectively suppress the distance calculation deviation caused by noise, poor electrode contact, or signal drift in a single or a few channels, thus improving the robustness of distance assessment.

[0045] In addition to the overall feature distance, three key relative features were extracted, which are crucial for distinguishing between target and compensatory movements. The first is the target channel proportion, which quantifies the proportion of signal energy from the preset target muscle group channel relative to the total energy of all channels within the current time window. A high target channel proportion indicates that the current muscle activity is primarily contributed by the target muscle group. The second is the compensatory channel proportion, corresponding to the target channel proportion. This indicator quantifies the proportion of signal energy from the preset compensatory muscle group channel. A high compensatory channel proportion indicates a high level of involvement of compensatory behavior. The third is the co-contraction index, specifically used to assess the degree of synchronous activation of antagonistic muscle groups. In normal motor control, when the agonist muscle contracts, the antagonist muscle usually relaxes. However, in cases of impaired motor control, abnormal synchronous contraction of the antagonist muscles often occurs, i.e., co-contraction. The co-contraction index can be obtained by calculating the ratio of the overlapping portion of the signal envelope of the antagonist muscle to the total active portion. A high co-contraction index indicates an abnormal movement pattern.

[0046] The final proximity score is determined by integrating all the above information. This score can be a comprehensive rating or a vector containing multiple components. For example, proximity to a target action reference could be a function that combines a lower weighted distance, a higher target channel proportion, a lower compensating channel proportion, and a lower common contraction index. This multi-dimensional comprehensive evaluation allows for a more accurate characterization of the current signal's attributes. Even with a high total signal amplitude, if the compensating channel proportion or common contraction index is also high, it can be determined that this is not a high-quality target action, thus avoiding misjudgments that rely solely on signal strength for triggering decisions.

[0047] In one specific implementation, before calculating the weighted distance between the current time window features and the target action reference and the compensatory action reference based on the dynamic weights of each biosignal acquisition channel, the method further includes: acquiring the historical stability, signal-to-noise ratio of the day, contact quality, and resting baseline offset of each acquisition channel; calculating the initial weight of each acquisition channel based on the historical stability, signal-to-noise ratio of the day, contact quality, and resting baseline offset; and during the feature extraction process, when it is detected that the signal variance of an acquisition channel increases or the noise increases beyond a set value, the weight of that acquisition channel is reduced to form the dynamic weights of each biosignal acquisition channel.

[0048] Before training begins or during the initial calibration phase, an initial weight is calculated for each acquisition channel. This calculation integrates four aspects of information: First, historical stability, extracted from the patient's historical training records, reflects the dispersion of signal characteristics of a specific channel for the same target action in previous training sessions. A channel that shows stability and low dispersion in historical data is considered more reliable and receives a higher stability score. Second, the signal-to-noise ratio (SNR) for the day, assessed after acquiring the resting signal and calibration action signal for the day; a higher SNR indicates better signal quality. Third, contact quality, assessed by measuring the impedance between the electrode and the skin; low impedance generally indicates good electrode contact quality. Fourth, the degree of resting baseline deviation, comparing the baseline of the resting signal acquired that day with the baseline level in historical records; a channel with a smaller deviation indicates greater stability. These four dimensions of assessment are weighted and combined, for example, using a linear formula, to calculate a comprehensive initial weight for each channel. Thus, at the start of formal training, the influence of each channel is pre-set based on its long-term historical performance and current immediate state.

[0049] During the actual training, the weights are not static but dynamically adjusted in real time. A common indicator of signal quality degradation is an abnormal increase in signal variance or a significant rise in noise levels. This is usually related to factors such as changes in the conductive adhesive properties of the electrodes due to patient sweating, loosening of the electrodes due to body movement, or external electromagnetic interference. At each time window of feature extraction, these indicators for each channel are monitored. Once the signal variance or noise level of a channel exceeds a preset dynamic adjustment threshold, the reliability of that channel is considered to have decreased at that moment. In response, the weight of that channel is immediately reduced. This reduction can be a fixed percentage or a dynamic adjustment based on the degree of exceeding the threshold.

[0050] This real-time, negative feedback adjustment ensures that any channel that suddenly becomes unreliable during the entire training process will have its negative impact on the final recognition result suppressed immediately. The resulting dynamic weights accurately reflect the true contribution value of each channel at each moment, thus providing the most reliable input for subsequent weighted distance calculation and pattern recognition.

[0051] In one specific implementation, the process of performing local re-judgment only among the target action, resting state, and compensatory interference, and outputting a trigger command or a delayed trigger command based on the re-judgment result, includes: reducing the dynamic weight corresponding to the acquisition channel that has shown a compensatory increase characteristic, and extracting a multi-window continuity score; calculating a target score based on the proximity of the current time window to the target action reference, the proportion of the target channel, and the multi-window continuity score; calculating a compensation score based on the proximity of the current time window to the compensatory action reference, the proportion of the compensatory channel, and the common contraction index; comparing the target score with the compensation score, and outputting a corresponding trigger command when the target score is higher than the compensation score by a preset difference; and outputting a delayed trigger command for the corresponding compensatory interference and entering a pending confirmation state when the compensation score is higher than the target score.

[0052] For acquisition channels that already exhibit a clear upward trend in compensatory activity before entering the re-judgment state, their dynamic weight will be further reduced. This means that in the subsequent scoring calculation, the influence of these known interference source channels will be further weakened, allowing the decision-making process to focus more on those channels that may still carry the true target intent. Simultaneously, a multi-window continuity score will be extracted. This score is used to assess whether the signal pattern in the current time window is part of a coherent action process, rather than an isolated, sudden signal spike. For example, in a real action initiation process, the amplitude of the electromyographic signal typically increases smoothly and remains stable for a period of time. By analyzing the characteristic change trends of the current time window and several adjacent time windows, such a continuity score can be calculated. A high continuity score increases the likelihood that the current signal represents a meaningful action intent.

[0053] The target score is calculated by combining three factors: the proximity of the current time window to the target action reference, which reflects the similarity of overall features; the proportion of the target channel, which reflects the contribution of the target muscle group; and the aforementioned multi-window continuity score, which reflects the temporal coherence of the movement. These three factors are weighted and summed to form a target score that can comprehensively assess the intensity and quality of the target action intention.

[0054] At the same time, the calculation of the compensation score also takes into account several factors: the proximity of the current time window to the reference of the compensation action, which reflects the similarity to the known compensation pattern; the proportion of the compensation channel, which reflects the participation of the compensation muscle group; and the common contraction index. These three factors are also weighted and summed to form a compensation score that quantifies the intensity of compensation interference.

[0055] The system determines whether the target score is higher than the compensatory score by a preset difference. This difference is to ensure the reliability of the decision. A positive judgment is only made when the evidence of the target intention is significantly stronger than the evidence of compensatory interference. If this condition is met, a corresponding trigger command is output to activate the rehabilitation device. Conversely, if the compensatory score is higher than the target score, or if the target score is slightly higher but does not exceed the preset difference, the current primary activity is determined to be compensatory interference. In this case, a corresponding delay trigger command for compensatory interference is output, and the system enters a pending confirmation state, thereby avoiding responses to erroneous actions.

[0056] In one specific implementation, after outputting the corresponding delayed triggering instruction for compensatory interference, the method further includes: prohibiting the sending of normal mechanical assistive action instructions, and controlling the rehabilitation device to perform one of maintaining the screen and waiting, reducing the assistive level, or outputting action correction prompts; when multiple consecutive time windows are in a pending confirmation state during the same action attempt, recording the corresponding attempt round as a compensatory enhancement round; and incorporating the time window features corresponding to the compensatory enhancement round into the updated set of compensatory action references.

[0057] Once the system enters the pending confirmation state, sending normal mechanical assistive movement commands to the rehabilitation device is prohibited. This is to prevent physical reinforcement of detected compensatory behaviors. Next, one of three preset intervention methods is selected for execution. The first is to maintain a waiting screen, meaning the training task interface on the screen remains unchanged, providing no physical assistance and giving the patient an opportunity to adjust and retry, thus maintaining the continuity of training. The second is to reduce the level of assistance. If the rehabilitation device has multi-level assistance functions, it can be adjusted to a lower level. This means the device still provides some support or guidance, but mainly relies on the patient's active exertion, encouraging the patient to find the correct exertion pattern with reduced external assistance. The third is to output movement correction prompts, providing specific and targeted corrective suggestions on the display terminal in the form of text, images, or animations, such as prompts to relax the shoulder or pay attention to forearm posture. All three methods aim to effectively manage and guide compensatory behaviors without interrupting the training process.

[0058] If the pending state only appears briefly for one or two time windows during a certain action attempt, it will be considered an occasional event and will continue to be observed without special recording. However, if multiple consecutive time windows are in the pending state during the same action attempt, for example, lasting for more than half a second, the compensatory behavior in this attempt will be determined to be continuous and significant. In this case, the entire attempt round will be marked and recorded as a compensatory enhancement round.

[0059] Signal data recorded as rounds of compensatory enhancement are incorporated into the feature set used to update the compensatory action reference. In this way, the system can continuously learn new or more normalized compensatory patterns in patients, enabling the compensatory action reference model to evolve dynamically and continuously improve its ability to identify and distinguish various compensatory behaviors. This allows for earlier and more accurate detection and handling of these compensatory interferences in future identifications.

[0060] In one specific implementation, the method further includes: continuously monitoring the system status within multiple update cycles; generating a manual intervention prompt carrying a restricted action identifier when any of the following conditions are met: failing to acquire the minimum number of valid target features within a consecutive preset update cycle, the proportion of the same training group entering the pending confirmation state continuously exceeding the upper limit, or the distance between the updated target action reference and the compensating action reference continuously falling below the lower limit of the safe distance; and outputting the manual intervention prompt to instruct the acquisition of short-term calibration signals on a specified acquisition channel or action.

[0061] The system continuously monitors its operational status across multiple update cycles. An update cycle can be a training set or a fixed number of action repetitions. The monitored metrics are several key performance indicators. The first indicator is the efficiency of acquiring effective target features. If, over several consecutive update cycles, the minimum number of confirmed effective target features are not collected, it may mean that the patient is having difficulty correctly performing the target action, or that the current recognition parameter settings are significantly deviating from the patient's actual state. The second indicator is the frequency of compensatory interference. If, within the same training set, the proportion of time windows entering the pending confirmation state consistently exceeds a preset upper limit, such as 30%, it indicates that compensatory behavior is very prevalent and seriously affects the effectiveness of training. The third indicator is the safety of model updates. If, after multiple consecutive updates, the distance between the target action reference and the compensatory action reference is lower than the safe distance limit, it indicates that the classification boundary is being severely squeezed, and the system faces the risk of mislearning compensatory patterns as target patterns.

[0062] When any one of the above three conditions is met, it will be determined that the current situation requires manual intervention. At this time, a manual intervention prompt carrying a restricted action identifier will be automatically generated. This prompt can contain specific problem diagnosis information. For example, the prompt information will clearly indicate which action has insufficient effective feature acquisition or which action has an excessively high compensation ratio.

[0063] Finally, this prompt is output to the therapist, for example, displayed on the therapist's terminal. The prompt directly provides suggested interventions, such as instructing the therapist to collect a short calibration signal on a specific acquisition channel that may be experiencing problems, or for a specific, difficult-to-recognize movement. For example, the prompt might suggest reconfirming the upper trapezius electrode position, or instructing the patient to relax their shoulder and repeat elbow flexion twice. This type of manual intervention request, providing specific guidance, avoids unnecessary interruptions and makes the therapist's intervention more precise and efficient, significantly reducing the time spent on troubleshooting and recalibration.

[0064] In one specific implementation, the method further includes: calculating the current fatigue score by combining the median frequency and envelope amplitude of the biological signal; relaxing the lower threshold for judging the target action reference when the current fatigue score increases and the proportion of compensatory channels and the common contraction index do not exceed a preset level; and tightening the upper threshold for judging the compensatory action reference when the current fatigue score increases and the proportion of compensatory channels or the common contraction index increases simultaneously beyond a preset level.

[0065] First, a current fatigue score is calculated in real time, based on two physiological indicators of muscle fatigue: median frequency and envelope amplitude. During muscle fatigue, the median frequency of the electromyographic signal typically shifts towards lower frequencies, while the envelope amplitude decreases. These two indicators of the target channel's electromyographic signal are continuously tracked and compared with reference values ​​from the initial training phase or a non-fatigue state to calculate a quantitative fatigue score. A higher score indicates a more severe degree of muscle fatigue.

[0066] Next, based on changes in fatigue scores and the performance of compensatory indicators, a dual-path threshold adaptive adjustment is performed. The first scenario is where the current fatigue score increases, but simultaneously, the two key compensatory indicators—the proportion of compensatory channels and the co-contraction index—do not exceed preset normal levels. This indicates that although the patient's target muscle group is fatigued, leading to decreased exertion capacity and weakened signals, the patient is still striving to complete the movement with the correct movement pattern and has not exhibited significant compensatory behavior. To encourage patients to attempt correctly under these challenging conditions and avoid misjudging movements as invalid due to weakened signals, the lower threshold for judging target movements is appropriately relaxed. This means that a weaker but correctly patterned movement still has a high probability of being identified as a valid target movement.

[0067] The second scenario involves a simultaneous increase in the fatigue score and a corresponding rise in the proportion of compensatory pathways or the co-contraction index. This indicates that the patient, due to fatigue in the target muscle group, is resorting to incorrect compensatory strategies to complete the task. In this case, simply relaxing the target recognition threshold would be extremely dangerous, as it could lead to these fatigue-induced compensatory movements being misjudged as correct movements. Therefore, the opposite strategy is adopted: tightening the upper limit threshold for judging compensatory movements. This means a lower tolerance for compensatory behavior, and even signals with some similarity to the compensatory reference are more likely to be judged as compensatory interference. This differentiated adjustment strategy allows for maintaining sensitivity to correct effort while strengthening the inhibition of incorrect compensation when the patient is fatigued, thus maintaining the quality and correct orientation of training throughout the dynamically changing physiological state.

[0068] In one specific implementation, the process of updating the target action reference and the compensatory action reference according to the valid target features and compensatory interference features confirmed during training includes: setting the update rate of the target action reference to be less than the update rate of the compensatory action reference when updating the action references separately; extracting and saving the distribution drift vector between the mean of the credible target window in the first half and the mean of the credible target window in the second half of the current training round; when training is handed over across groups or days, if it is detected that the distribution of the new round of initial signal is close to the compensatory action reference according to the distribution drift vector, or if it is detected that the distance between the target action reference and the compensatory action reference has decreased, then the restricted output mode is entered; in the restricted output mode, only the specified target category and the compensatory interference category are retained for recognition until a sufficient number of valid target features are reacquired before full multi-class recognition is restored.

[0069] In the design of the update mechanism, a differentiated update rate is adopted. The update rate for the target action reference is set to a relatively small value. This means that the target reference model is more conservative and slower in absorbing new data. This slow update strategy can effectively filter out occasional noise or slight compensatory contamination, ensuring the stability and purity of the target action reference, enabling it to represent a high-quality, ideal action pattern over the long term. In contrast, the update rate for the compensatory action reference can be set slightly higher. This allows the system to learn and capture newly emerging compensatory patterns in patients more quickly, thereby timely incorporating these new interfering patterns into the monitoring scope and improving the sensitivity of compensatory behavior identification.

[0070] To quantify and record state changes during a single training session, a distribution drift vector is extracted and saved. Specifically, all target windows identified as reliable in the current training round are divided into a first half and a second half in chronological order. The mean values ​​of the features in these two parts are then calculated. The difference vector between these two means is defined as the distribution drift vector for this training session. This vector, in both direction and magnitude, indicates the direction in which the patient's stable target movement pattern has drifted during this training session. This is usually related to the accumulation of fatigue or minor adjustments in motor learning. This drift vector serves as an important state summary for this training session and is saved along with the updated reference model.

[0071] At the transition points between cross-group or cross-day training, status assessment is performed using information saved from the previous training. When a new round of training begins and initial calibration signals are collected, two key conditions are checked. First, does the new initial signal distribution, under the influence of the drift vector from the previous training, move closer to the compensatory action reference? Second, does the new initial signal distribution result in a significant reduction in the gap between the target action reference and the compensatory action reference? If either of these conditions is met, it indicates that the patient's current state may have adversely changed, and the classification boundary has been squeezed. In this case, performing full-class recognition carries a high risk, and as a preventative measure, the system automatically switches to restricted output mode.

[0072] Limited output mode is a temporary, highly secure working state. In this mode, the recognition scope is temporarily narrowed; it no longer attempts to distinguish all action categories, but only retains the core target category and compensatory interference category of the current training task for recognition. This greatly simplifies the recognition problem, reduces the risk of misjudgment due to state drift and boundary ambiguity, and ensures stability and security in the early stages of training when uncertainty is high. Training continues in this limited mode until a sufficient number of new, effective target features are accumulated during the training process. Only when the patient's condition is confirmed to be stable, or a new reference model has been successfully established, will the system gradually and in a controlled manner revert to a complete, multi-class recognition mode.

[0073] In one specific implementation, the process of retaining only the specified target category and the compensatory interference category for identification in the restricted output mode until a sufficient number of valid target features are reacquired before restoring complete multi-class identification includes: obtaining the number of times each non-identified category and the current specified target category jointly enter local re-judgment before entering the restricted output mode; determining the restoration order of each non-identified category according to the number of times from least to most; and after reacquiring a sufficient number of valid target features, when the boundary compression degree and the number of local re-judgments corresponding to the current identification range are both lower than the corresponding thresholds, adding a non-identified category to the current identification range according to the restoration order.

[0074] When entering the restricted output mode, a historical data analysis is performed to review the training records before entering the restricted mode. The number of times each action category that was not involved in the recognition at that time and the action category that is currently retained as the designated target jointly enter the local re-judgment is counted. This number is an indicator of the degree of confusion between the two categories. The more times they jointly enter the local re-judgment, the more blurred the boundary between the two categories in the feature space, and the greater the possibility of interference between them.

[0075] Based on this statistical result, a recovery order is determined, ranking all categories that did not participate in the identification process in ascending order of the number of times they jointly entered local re-judgment. This means that those categories that have historically had the least confusion with the current target category and are least likely to cause interference will be placed at the front of the recovery queue, while those categories that have historically been the most likely to cause confusion will be placed at the back.

[0076] Next, the recovery process requires strict preconditions. First, in restricted mode, a preset number of effective target features must be reacquired through training. This ensures that the current core recognition task has stabilized. On this basis, the degree of boundary compression and the number of local re-judgments are checked within the recognition range that currently only contains target categories and compensatory interference. Only when both of these stability indicators are lower than their respective preset thresholds will the current recognition environment be considered stable enough, and it will be possible to start expanding the recognition range.

[0077] Once all conditions are met, the recovery step is initiated, adding only one category that was not identified to the current identification range at a time. For example, the category that is first in the recovery order is added to the identification set. This recovery method of adding one category at a time ensures that the potential risks introduced by expanding the identification range each time are minimized and controllable.

[0078] In one specific implementation, the method further includes: after each addition of a non-identified category to the current identification range, obtaining the degree of boundary compression within a preset observation window, the category change rate of a continuous time window, and the number of local re-judgments; when at least one of the degree of boundary compression, the category change rate of a continuous time window, and the number of local re-judgments exceeds the corresponding threshold, removing a non-identified category from the current identification range and maintaining the identification range before its addition; when the degree of boundary compression, the category change rate of a continuous time window, and the number of local re-judgments do not exceed the corresponding threshold, retaining a non-identified category and continuing to add the next non-identified category according to the recovery order.

[0079] After each new, previously unrecognized category is added to the current recognition range according to the recovery order, a short observation period is initiated. Within this pre-defined observation window, such as the next thirty seconds or ten action attempts, three key stability metrics are monitored: boundary squeezing degree, reflecting whether the newly added category has caused blurring of the classification boundary; the category change rate over a continuous time window, reflecting whether the stability of the recognition results has been affected; and the number of local reclassifications, reflecting the frequency of uncertainty. These three metrics can quickly and objectively assess the immediate impact of adding a new category on the overall recognition performance.

[0080] Decisions are made based on monitoring results during the observation period. If at least one of these three indicators exceeds its corresponding safety threshold—for example, a sudden spike in boundary squeezing or frequent jumps in recognition results—it is determined that the newly added category is causing excessive interference in the current state. In this case, a rollback operation is triggered, removing the newly added category from the current recognition range and restoring it to its previous, verified stable state. This immediate rollback mechanism effectively prevents the entire system from falling back into instability due to inappropriate category restoration.

[0081] Conversely, if all three stability metrics remain below safe thresholds throughout the observation window, the category recovery is considered successful. In this case, the newly added category is retained, and then the next category not yet identified is added according to the predetermined recovery order, repeating the same observation and verification process. Through this trial-verification-confirmation / rollback cycle, the recognition range can be gradually and dynamically restored from the restricted mode to the complete state in a very safe and robust manner, achieving a smooth transition while ensuring training quality and consistency.

[0082] Secondly, referring to Figure 2 This application also proposes a biosignal pattern recognition system for rehabilitation training, used to perform any one of the steps in the above methods, including: The first module 1 is used to extract features from the collected biological signals according to continuous time windows, calculate the proximity of each time window feature to the target action reference and the compensatory action reference, and determine the degree of boundary compression by combining the category change rate and proximity difference of the continuous time window. The second module 2 is used to perform local re-judgment only among target action, rest and compensatory interference when the degree of boundary compression exceeds the preset threshold, and output trigger command or postpone trigger command according to the re-judgment result. The third module 3 is used to update the target action reference and the compensation action reference according to the valid target features and compensation interference features confirmed in the training. If the distance between the two after the update is less than the lower limit of the safe distance, the update of the target action reference is cancelled.

[0083] By extracting features from the collected biosignals according to continuous time windows, the proximity of each time window feature to the target action reference and the compensatory action reference is calculated. The degree of boundary compression is determined by combining the category change rate and proximity difference of the continuous time windows, enabling dynamic evaluation of the clarity of the recognition boundary. When the boundary compression exceeds a preset threshold, local re-judgment is performed only among the target action, resting state, and compensatory interference. Based on the re-judgment result, a trigger command or a delayed trigger command is output, thus refining the recognition when the boundary is blurred and avoiding misjudgment. The target is updated according to the effective target features and compensatory interference features confirmed during training. If the distance between the action reference and the compensatory action reference is less than the lower limit of the safe distance after the update, the update of the target action reference is cancelled. This achieves adaptive updating of the action reference and prevents the update from reducing the distinguishability between the target and the compensatory action. It effectively solves the problems of unstable recognition, frequent false triggers, and frequent manual intervention required by existing biosignal pattern recognition systems when the patient's physiological state changes dynamically. It has the advantages of being able to dynamically adapt to changes in the patient's physiological state, effectively distinguish between the target action and the compensatory interference, reduce misjudgment, improve the accuracy and continuity of rehabilitation training, and thus enhance the patient's training experience and rehabilitation effect.

[0084] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A biosignal pattern recognition method for rehabilitation training, characterized in that, include: Features are extracted from the collected biological signals according to continuous time windows. The proximity between the features of each time window and the target action reference and the compensatory action reference is calculated respectively. The degree of boundary compression is determined by combining the category change rate of the continuous time window and the proximity difference. When the degree of boundary compression exceeds the preset threshold, local re-judgment is performed only among the target action, rest and compensatory interference, and a trigger command or a delay trigger command is output according to the re-judgment result. The target action reference and the compensation action reference are updated according to the effective target features and compensation interference features confirmed during training. If the distance between the two after the update is less than the lower limit of the safe distance, the update of the target action reference is cancelled.

2. The biosignal pattern recognition method for rehabilitation training according to claim 1, characterized in that, The calculation of the proximity between the features of each time window and the target action reference and the compensatory action reference includes: Based on the dynamic weights of each biosignal acquisition channel, the weighted distances between the current time window features and the target action reference and the compensatory action reference are calculated respectively. Extract the target channel proportion, compensatory channel proportion, and common contraction index representing the contraction intensity of the antagonistic muscle corresponding to the current time window features; Based on the weighted distance, the target channel proportion, the compensating channel proportion, and the common contraction index, the proximity of the current time window features to the target action reference and the compensating action reference is determined.

3. The biosignal pattern recognition method for rehabilitation training according to claim 2, characterized in that, Before calculating the weighted distance between the current time window features and the target action reference and the compensatory action reference based on the dynamic weights of each biosignal acquisition channel, the method further includes: Acquire the historical stability, signal-to-noise ratio of the day, contact quality, and resting baseline offset of each acquisition channel; The initial weights of each acquisition channel are calculated based on the historical stability, the signal-to-noise ratio of the current day, the contact quality, and the degree of resting baseline offset. During the feature extraction process, when it is detected that the signal variance of an acquisition channel increases or the noise rises beyond a set value, the weight of that acquisition channel is reduced to form the dynamic weight of each biological signal acquisition channel.

4. The biosignal pattern recognition method for rehabilitation training according to claim 2, characterized in that, The method of performing local re-judgment only among target action, rest, and compensatory interference, and outputting a trigger command or a delayed trigger command based on the re-judgment result, includes: Reduce the dynamic weights of acquisition channels that have shown compensatory increase characteristics, and extract multi-window continuity scores; The target score is calculated based on the proximity of the current time window to the target action reference, the target channel ratio, and the multi-window continuity score. The compensation score is calculated based on the proximity of the current time window to the compensation action reference, the proportion of the compensation channel, and the common contraction index. Compare the target score with the compensation score, and when the target score is higher than the compensation score by a preset difference, output the corresponding trigger command; When the compensation score is higher than the target score, a delay trigger command for the corresponding compensation interference is output and the system enters a pending confirmation state.

5. The biosignal pattern recognition method for rehabilitation training according to claim 4, characterized in that, Following the output of the delayed triggering instruction corresponding to the compensatory interference, the following is also included: Prohibit sending normal mechanical assistive movement commands and control the rehabilitation equipment to perform one of the following actions: hold the image while waiting, reduce the level of assistance, or output movement correction prompts; When multiple consecutive time windows are in the pending confirmation state during the same action attempt process, the corresponding attempt rounds will be recorded as compensation enhancement rounds; The time window features corresponding to the compensation enhancement rounds are included in the update set of the compensation action reference.

6. The biosignal pattern recognition method for rehabilitation training according to claim 4, characterized in that, The method further includes: Continuously monitor system status across multiple update cycles; When any of the following conditions are met: the minimum number of valid target features are not obtained in a consecutive preset update cycle, the proportion of the same training group entering the pending confirmation state continues to exceed the upper limit, or the distance between the updated target action reference and the compensatory action reference is continuously lower than the lower limit of the safe distance, a manual intervention prompt carrying a restricted action identifier is generated. Output the aforementioned manual intervention prompt to instruct the acquisition of a short-term calibration signal on a specified acquisition channel or action.

7. The biosignal pattern recognition method for rehabilitation training according to claim 2, characterized in that, The method further includes: The current fatigue score is calculated by combining the median frequency and envelope amplitude of the biological signal; When the current fatigue score increases, and the proportion of the compensatory channel and the common contraction index do not exceed the preset level, the lower limit threshold for judging the target action reference is relaxed. When the current fatigue score increases, and the proportion of the compensation channel or the common contraction index increases simultaneously beyond a preset level, the upper limit threshold for judging the compensation action reference is tightened.

8. The biosignal pattern recognition method for rehabilitation training according to claim 1, characterized in that, The step of updating the target action reference and the compensatory action reference based on the effective target features and compensatory interference features confirmed during training includes: When updating the action references separately, the update rate for the target action reference is set to be less than the update rate for the compensatory action reference. Extract and save the distribution drift vector between the mean of the trusted target window in the first half and the mean of the trusted target window in the second half of the current training round; During cross-group or cross-day training handover, if it is detected that the new round of initial signal distribution is close to the compensating action reference according to the distribution drift vector, or if it is detected that the distance between the target action reference and the compensating action reference is reduced, then the restricted output mode is entered. In the restricted output mode, only the specified target category and the compensatory interference category are retained for identification until a sufficient number of valid target features are reacquired before full multi-class identification is restored.

9. The biosignal pattern recognition method for rehabilitation training according to claim 8, characterized in that, The process of retaining only the specified target category and compensatory interference category for identification in the restricted output mode, until a sufficient number of valid target features are reacquired, and then restoring complete multi-class identification, includes: Obtain the number of times each category not involved in the identification process and the currently specified target category jointly enter the local re-judgment before entering the restricted output mode; The recovery order for each category not involved in the identification is determined according to the number of times the identification was performed, from least to most. After reacquiring the required number of valid target features, if the degree of boundary compression and the number of local re-judgments corresponding to the current identification range are both lower than the corresponding thresholds, a category that did not participate in the identification is added to the current identification range according to the recovery order.

10. A biosignal pattern recognition system for rehabilitation training, used to perform the method according to any one of claims 1 to 9, characterized in that, include: The first module is used to extract features from the collected biological signals according to continuous time windows, calculate the proximity of each time window feature to the target action reference and the compensatory action reference, and determine the degree of boundary compression by combining the category change rate of the continuous time window and the proximity difference. The second module is used to perform local re-judgment only among target action, rest and compensatory interference when the degree of boundary compression exceeds a preset threshold, and output a trigger command or a delay trigger command according to the re-judgment result. The third module is used to update the target action reference and the compensation action reference according to the effective target features and compensation interference features confirmed in the training. If the distance between the two after the update is less than the lower limit of the safe distance, the update of the target action reference is cancelled.

Images