A rehabilitation movement evaluation method

Abstract

The application discloses a rehabilitation action evaluation method, comprising the following steps: setting a rehabilitation action, selecting a plurality of training actions capable of evaluating the rehabilitation action and a plurality of significant motion features corresponding to the plurality of training actions respectively, establishing a demonstration action template, and dividing demonstration action template data into feature sections; collecting user rehabilitation action data divided into feature sections by using a sensing device, and calculating the likelihood values of each feature section to the template; completing fuzzy comprehensive evaluation of the test action by using a fuzzy comprehensive evaluation model, and calculating the comprehensive score value thereof; taking the likelihood values of each feature section in combination with a scaling degree index as characteristic values of a tree model method, taking the comprehensive score value of the fuzzy comprehensive evaluation as a corresponding label, and quantifying the score by using a gradient boosting tree model algorithm to evaluate the current execution rehabilitation action; and realizing a relatively accurate evaluation result by using less action data and lower calculation amount, and the algorithm is simple, the cost is low, the model is easy to deploy, and the applicability is strong.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation movement assessment, and more specifically to a method for evaluating rehabilitation movements. Background Technology

[0002] When people experience motor impairments due to physical dysfunction, rehabilitation training is usually necessary to eliminate or alleviate the impairments and restore fine motor skills. During rehabilitation training, people typically need to complete several different rehabilitation exercises designed by professional rehabilitation medical experts. Guidance for rehabilitation training is often provided through direct observation and assistance from clinicians. However, due to objective limitations, over 90% of rehabilitation training in my country is currently completed at home. This results in many users lacking professional and objective rehabilitation training assessments and feedback, directly impacting the effectiveness and sustainability of the training.

[0003] Intuitive and timely assessment feedback can improve user motivation and training effectiveness. Therefore, sensor-based rehabilitation movement evaluation methods have emerged, and currently, existing rehabilitation movement evaluation methods can be mainly divided into the following three types:

[0004] (1) Threshold method: By setting a data threshold, when the received signal exceeds (or falls below) the preset threshold, it is judged that the action is not demonstrated.

[0005] (2) Distance measurement: The distance and template matching positioning method is used to determine whether the rehabilitation action is demonstrated, and the distance between the user's action sequence information and the characteristics of the demonstrated action sequence is calculated.

[0006] (3) Probabilistic model: Represent rehabilitation actions as a state sequence, and calculate and extract various evaluation indicators or features related to the actions based on the output state sequence and evaluation content of the model.

[0007] Current rehabilitation exercise evaluation methods, taking lower limb rehabilitation exercises as an example, typically break down lower limb rehabilitation exercises into several training movements, such as knee flexion and adduction exercises, knee flexion and abduction exercises, and straight leg raise exercises. These training movements are then further broken down into individual movements that need to be repeated multiple times; for example, straight leg raise exercises are broken down into ten repetitions of the straight leg raise. Finally, an overall evaluation is made for a specific single movement based on the overall average index of the individual movement, such as giving an overall score for the fifth repetition of the straight leg raise. However, the overall average index is insufficient to fully understand the characteristics and progress of the rehabilitation exercise. Specific single movements usually consist of different sub-parts; for example, the straight leg raise exercise can be broken down according to the time sequence into sub-parts such as raising the leg, knee flexion, maintaining extension, leg retraction, and lowering. Based on the characteristics of rehabilitation movements, there will be differences in the numerical values ​​and trends of characteristic indicators in different time segments of the training movement. Taking straight leg raise as an example, more attention is paid to the intensity and pattern of electromyographic activity when raising the leg, while stability is more important when lowering the leg. Moreover, these data may be affected by factors such as rehabilitation progress, physical condition, and individual differences among patients. Therefore, each segment actually has different evaluation focuses. However, existing evaluation methods all score a single movement as a whole, ignoring the differences between each segment. Therefore, the current technology is not accurate enough for the overall evaluation results of a single repetition of a single movement, because the existing methods are not good at grasping the subtle differences of specific rehabilitation movements. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the present invention aims to provide a rehabilitation movement evaluation method that achieves the entire process from movement segmentation to evaluation through a single model. The evaluation results are relatively accurate, the algorithm is simple, the cost is low, the model is easy to deploy, and it has strong applicability.

[0009] To achieve the above objectives, the present invention employs the following technical solution:

[0010] A method for evaluating rehabilitation movements includes the following steps:

[0011] S1. Set up rehabilitation movements, select multiple training movements that can evaluate rehabilitation movements and their corresponding multiple significant motion features; collect significant motion feature data of repeated standard movements of multiple demonstrators through sensing devices, preprocess the data, and use the Gaussian mixture model algorithm to cluster and fit the preprocessed data on the time series to obtain the Gaussian mixture model parameters and divide the feature segments, and establish a demonstration movement template for a single significant motion feature under a single repetition.

[0012] S2. Use sensor devices to collect user rehabilitation test action data, and scale the single action cycle time to match the demonstration action template through interpolation. Use the action template constructed in S1 using the Gaussian mixture model algorithm to cluster and divide the integrated data into feature segments, and calculate the likelihood value of each feature segment.

[0013] S3. Perform a fuzzy comprehensive evaluation on the user's rehabilitation test action data using a fuzzy comprehensive evaluation model, and calculate its comprehensive score.

[0014] S4. The likelihood values ​​of each feature segment calculated in S2, combined with the scaling index, are used as the feature values ​​of the tree model method, and the comprehensive score calculated in S3 is used as the label of the tree model method. The score is quantified by the gradient boosting tree model algorithm to evaluate the current rehabilitation action.

[0015] The present invention also has the following technical features:

[0016] Preferably, the sensing devices described in S1 and S2 include a dual-channel electromyography sensor and an inertial sensor.

[0017] Furthermore, the data processing method described in S1 includes filtering of data acquired by the dual-lead electromyography sensor, regularization of negative data, Gaussian smoothing, and normalization.

[0018] Preferably, the calculation method for clustering using the Gaussian mixture model algorithm described in S1 and S2 includes:

[0019] 1) Calculate the probability density of a single Gaussian component fitted to the salient motion feature data, which is expressed as:

[0020]

[0021] In the formula, μ k It is the expectation of the k-th component, with a dimension of 2; Σ k It is the covariance of the k-th component; y j These are the example data points in the action loop, j = 1, 2, ..., N, where N is determined by the sensor sampling frequency;

[0022] 2) Fit the Gaussian components of multiple significant motion feature data. The mixture function of the multiple significant motion feature data components is expressed as:

[0023]

[0024] 3) The likelihood value is calculated using the likelihood function of the Gaussian mixture model, and its functional expression is as follows:

[0025]

[0026] 4) The EM algorithm is used to solve for the final parameters of the model.

[0027] Furthermore, the E-step aims to complete the information related to the latent variables through the initialized parameters, and its function is expressed as:

[0028]

[0029] In the formula, Given the parameter θ and the example data y, the calculated γ ik The expected value represents the probability that the demonstration data point j comes from component k under the current model parameters, which is the responsiveness of component k to the demonstration data point j: θ (i) For the model parameters of the previous iteration, θ was used in the first iteration. (i) Parameter θ for initialization (0) ;

[0030] The M-step calculates the parameters for round i+1 by maximizing the likelihood, and its function is expressed as:

[0031]

[0032] The model parameters for the new iteration are obtained through the E-step and M-step, and the function is as follows:

[0033]

[0034]

[0035]

[0036] Repeat the E-step and M-step until convergence, i.e., the parameters tend to stabilize ||θ. i+1 -θ i ||<ε, thus obtaining the determined parameters of the Gaussian mixture model.

[0037] Preferably, the calculation method for scaling the single action loop time by interpolation as described in S2 is as follows:

[0038]

[0039] Among them, X z For scaling features, tr represents the number of original interpolation data points, and ts represents the number of pre-determined template data points.

[0040] Preferably, the method for completing fuzzy comprehensive evaluation using a fuzzy comprehensive evaluation model as described in S2 includes:

[0041] A fuzzy relation matrix is ​​established based on significant motion feature factors. The weights of the significant motion feature factors are calculated, and the results are used as the evaluation index weight vector of the fuzzy evaluation method. The weighted average operator is used to perform fuzzy comprehensive evaluation on each sample data in turn, and its comprehensive score is calculated.

[0042] Furthermore, the method for completing fuzzy comprehensive evaluation using a fuzzy comprehensive evaluation model as described in S2 includes:

[0043] 1) Determine the factor domain of the evaluation object. If there are m indicators in total:

[0044] U = {u1, u2, ..., u} m}

[0045] 2) Define the domain of the rating scale. The rating set is the collection of all possible overall ratings that the evaluator may make about the evaluated object:

[0046] V = {ν1,ν2,…,ν} n}

[0047] 3) Establish the fuzzy relation matrix R:

[0048]

[0049] Where, r mn Indicates that the evaluated object is determined from factor u m Let's look at the v n Membership degree of hierarchical fuzzy subsets;

[0050] 4) By constructing the fuzzy weight vector A of the evaluation factors using the entropy weight method combined with expert weights, the fuzzy comprehensive evaluation model is expressed as:

[0051]

[0052] Among them, b n This indicates the evaluation object's performance on v under all evaluation factors. j The membership degree of a fuzzy subset of ratings, assuming the score of the comment set is defined as S(4, 3, 2, 1), is calculated as follows:

[0053] B=b1×4+b2×3+b3×2+b4×1.

[0054] Preferably, the gradient boosting tree boosting method described in S4 employs an additive model and a forward distribution algorithm to complete the learning process by gradually reducing the residuals.

[0055] Furthermore, the gradient boosting tree algorithm used is Adaboost, XGBoost, LightGBM, or GBDT.

[0056] Compared with the prior art, the present invention has the following technical effects:

[0057] This invention proposes a method for differential feature segments, which decomposes a single cycle of motion into a series of posture segments, meeting the needs of more detailed motion evaluation. It fits the motion into several feature segments using a Geometric Model (GMM) and uses the likelihood value calculated by the GMM as the feature of the tree model. At the same time, it proposes a new feature of scaling degree to characterize the difference in duration between motions, thereby generating an evaluation that is more in line with reality. It also reveals the influence of differential motion components on the performance of complete motion. Dynamic weights are assigned to different feature segments based on visual features. Each motion is comprehensively evaluated from aspects such as standardization, speed, and completeness, and finally, a quantitative evaluation index of the motion is obtained, providing a fast, objective assessment and real-time feedback for the user's rehabilitation motion.

[0058] This invention uses only dual-lead electromyography sensors and inertial sensors to bring the accuracy as close as possible to the physician's evaluation results without increasing system complexity. It achieves an accuracy rate of 84% in the physician's fuzzy comprehensive evaluation to complete the quantitative scoring of the action. It is low-cost, has a simple algorithm, and has strong promotional value. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the rehabilitation evaluation system of the present invention;

[0060] Figure 2 The diagram illustrates the preprocessing of salient motion feature data of a demonstrator performing repeated training movements in an example embodiment.

[0061] Figure 3 This is an example diagram of sample fitting for EMG and ROM feature segmentation based on standard samples, as shown in the embodiment.

[0062] Figure 4 The images show a portion of the raw EMG data (a) and the interpolated and scaled data (b) from the embodiment.

[0063] Figure 5 Contribution maps of different feature sections in the embodiment;

[0064] Figure 6 This is an example diagram illustrating the use of EMG and IMU sensors to acquire electromyographic and angle data in this embodiment. Detailed Implementation

[0065] The specific content of the present invention will be further explained in detail below with reference to the embodiments.

[0066] This embodiment provides a method for evaluating rehabilitation movements, including the following steps:

[0067] S1. Set up rehabilitation exercises, select multiple training exercises that can evaluate the rehabilitation exercises and their corresponding significant motion features; collect significant motion feature data of repeated training exercises from multiple demonstrators using dual-channel electromyography (EMG) sensors and inertial sensors; perform filtering, negative data regularization, Gaussian smoothing, and normalization preprocessing on the EMG data; and use the Gaussian Mixture Model (GMM) algorithm to cluster the preprocessed data. The calculation method of GMM clustering includes:

[0068] 1) Calculate the probability density of a single Gaussian component fitted to the salient motion feature data, which is expressed as:

[0069]

[0070] In the formula, μ k It is the expectation of the k-th component, with a dimension of 2; Σ k It is the covariance of the k-th component; y j These are the example data points in the action loop, j = 1, 2, ..., N, where N is determined by the sensor sampling frequency;

[0071] 2) Fit the Gaussian components of multiple significant motion feature data. The mixture function of the multiple significant motion feature data components is expressed as:

[0072]

[0073] 3) The likelihood value is calculated using the likelihood function of the Gaussian mixture model, and its functional expression is as follows:

[0074]

[0075] 4) The EM algorithm is used to solve for the final parameters of the model.

[0076] Furthermore, the E-step aims to complete the information related to the latent variables through the initialized parameters, and its function is expressed as:

[0077]

[0078] In the formula, Given the parameter θ and the example data y, the calculated γ ik The expected value represents the probability that the demonstration data point j comes from component k under the current model parameters, which is the responsiveness of component k to the demonstration data point j: θ (i) For the model parameters of the previous iteration, θ was used in the first iteration. (i) Parameter θ for initialization (0) ;

[0079] The M-step calculates the parameters for round i+1 by maximizing the likelihood, and its function is expressed as:

[0080]

[0081] The model parameters for the new iteration are obtained through the E-step and M-step, and the function is as follows:

[0082]

[0083]

[0084]

[0085] Repeat the E-step and M-step until convergence, i.e., the parameters tend to stabilize ||θ. i+1 -θ i Given ||<ε, the parameters of the Gaussian mixture model are determined. Each demonstration action is fitted to this template. After calculating the probability of all components corresponding to each data point in the demonstration action, the component with the highest probability is selected as the component k to which the demonstration data point belongs. The calculation of all demonstration action data points is completed in sequence, and each demonstration action data point is assigned to its corresponding component with the highest probability. The feature segment division of the demonstration action is completed, and a demonstration action template for a single motion feature under a single repetition is established.

[0086] S2. Use sensor devices to collect user rehabilitation movement data, and use interpolation to scale the cycle time of a single movement to match the demonstration movement template.

[0087] Because the duration of a patient's single action varies, there are action loops with significant differences in time dimension, possibly due to being too fast or too slow. Unlike algorithms such as DTW, which can automatically handle time series data of unequal length, data of unequal length can lead to errors in the probability model fitting. To ensure the accuracy of action stages distinguished by feature segments, the time of a single action loop is scaled to match the template through interpolation. The error caused by this scaling is incorporated into the features of that single loop. The calculation method for the scaled features is as follows:

[0088]

[0089] Where Xz represents the scaling feature, tr represents the number of original interpolated data points, and ts represents the number of pre-determined template data points. The proposed scaling feature Xz represents the ratio of the number of data points added by the interpolation operation to the number of template data points. If it is close to 0, it means that the interpolation operation has almost no effect on the size of the dataset. If it is close to 1, it means that the interpolation operation has significantly increased the number of data points.

[0090] The GMM algorithm is used to cluster the integrated data into feature segments, and the likelihood value of each feature segment is calculated.

[0091] S3. Establish a fuzzy relation matrix for user rehabilitation movement data based on significant movement feature factors, calculate the weights of the significant movement feature factors, and use the results as the evaluation index weight vector for the fuzzy evaluation method. Use a weighted average operator to perform fuzzy comprehensive evaluation on each sample data in turn, and calculate its comprehensive score. The specific method includes:

[0092] 1) Determine the factor domain of the evaluation object. If there are m indicators in total:

[0093] U = {u1, u2, ..., u} m}

[0094] 2) Define the domain of the rating scale. The rating set is the collection of all possible overall ratings that the evaluator may make about the evaluated object:

[0095] V = {ν1,ν2,…,ν} n}

[0096] 3) Establish the fuzzy relation matrix R:

[0097]

[0098] Where, r mn Indicates that the evaluated object is determined from factor u m Let's look at the v n Membership degree of hierarchical fuzzy subsets;

[0099] 4) By constructing the fuzzy weight vector A of the evaluation factors using the entropy weight method combined with expert weights, the fuzzy comprehensive evaluation model is expressed as:

[0100]

[0101] Among them, b n This indicates the evaluation object's performance on v under all evaluation factors. j The membership degree of a fuzzy subset of ratings, assuming the score of the comment set is defined as S(4, 3, 2, 1), is calculated as follows:

[0102] B=b1×4+b2×3+b3×2+b4×1.

[0103] S4. The likelihood values ​​of each feature segment calculated in S2, combined with the scaling index and the comprehensive score calculated in S3, are used as feature values ​​for the tree model method. The gradient boosting tree model algorithm quantifies the score to evaluate the currently performed rehabilitation action. The gradient boosting tree boosting method uses an additive model and a forward distribution algorithm to learn by progressively reducing the residuals. The gradient boosting tree algorithm uses Adaboost, XGBoost, LightGBM, or GBDT. When constructing the dataset, the input is the likelihood value of each feature segment. One complete cycle of the action is considered as a sample. The likelihood values ​​of the feature segments of feature A are Xa_1, Xa_2, Xa_3…, and the likelihood values ​​of the feature segments of feature B are Xb_1, Xb_2, Xb_3…. At the same time, a scaling factor Xz, which reflects the error in the time of a single action, is added to the feature. The label is the fuzzy comprehensive evaluation score Y.

[0104] This embodiment focuses on the evaluation method for rehabilitation exercises involving seated straight leg raises. EMG and IMU sensors were used to acquire electromyographic and angular data. Ten volunteers were recruited, including five healthy individuals and five individuals with a history of knee injury or mobility impairment. A single-lead GER electromyographic sensor was selected and attached to the middle deltoid of the quadriceps femoris muscle. The MPU6050 inertial sensor (15g) was securely fixed to the lower tibia of the subject using a strap above the ankle. Figure 6 As shown;

[0105] The MPU6050 is a low-cost inertial sensor that incorporates an accelerometer, gyroscope, and other sensors. It can calculate joint angles using only a single sensor's acceleration values: because the straight leg raise is a single-degree-of-freedom movement, meaning it moves around a single axis in a plane, it can be approximated as a change in angle in the sagittal plane—that is, a change in angle on a two-dimensional plane that divides the body into left and right parts. The angle of the joint in the sagittal plane is calculated using the acceleration collected by the inertial sensor. The calculation formula is:

[0106]

[0107] Where θ(t) is the knee joint extension / flexion angle, acc.x is the acceleration on the x-axis collected by the sensor, and acc.z is the acceleration on the z-axis.

[0108] At the start of the experiment, participants calibrated their sensors, acquired EMG baseline data, and obtained IMU baseline data on a horizontal plane to reset the sensor settings. Each participant repeated the seated straight leg raise exercise approximately 30 times, completing the repetitions as best they could manage. Missed or incorrectly collected samples were removed, resulting in a total of 292 data sets.

[0109] Electromyography (EMG) data were processed using high-pass and low-pass filters in the 10-500Hz range and a band-stop filter at 50Hz; negative data were regularized; to preserve features as much as possible, smoothing parameters with minimal feature loss but clear trends were selected, and Gaussian smoothing with a window of 400ms was used. To eliminate individual differences, the maximum EMG values ​​of each subject were collected simultaneously after smoothing, and the data were normalized from max to min, as shown below. Figure 2 As shown;

[0110] Because the characteristics of inertial sensors are relatively clear, no preprocessing was performed on the inertial data in order to preserve the original data features to the greatest extent.

[0111] After acquiring sensor-based time-series data, 20 sets of standard motion data were selected and fitted using a Gaussian Mixture Model (GMM) to construct a standard motion template. Feature segment division is as follows: Figure 3 As shown. Figure 3 In this study, an electromyographic template for the standard movement was fitted using seven components, as was an angular template for the standard movement. During the fitting process, the EMG and angular data tended to stabilize after 110 and 50 iterations, respectively, and the GMM parameters at these points were used as the standard model parameters.

[0112] To ensure accurate differentiation of movement stages by feature segments, rehabilitation training data was collected, and the cycle time of a single movement was scaled to match the template through interpolation. Figure 4 As shown, the error caused by this scaling will be incorporated into the features of this single loop, and the likelihood values ​​of 14 feature segments, including EMG and angle, will be calculated. The scaling degree index will be used as the feature label x.

[0113] Constructing label y using a fuzzy method:

[0114] Based on the studied movements, a fuzzy relation matrix R is established using four factors (muscle exertion u1, movement limit angle u2, joint stability u3, speed control u4) and four indicators (excellent, qualified, average, poor).

[0115] The factor weights are calculated by weighting the results of the entropy weight method and the expert factor scores by w=0.5. The resulting factor weight calculation results are used as the evaluation index weight vector of the fuzzy evaluation method.

[0116] The weighted average type M(*,+) operator is used to perform fuzzy comprehensive evaluation on each sample data in turn, and its comprehensive score is calculated as the label Y of the dataset.

[0117] The tree-based Boosting algorithm was chosen because it is suitable for the established dataset and allows for intuitive interpretation of the features. GBDT, XGBoost, AdaBoost, and LightGBM models were trained on the dataset using Boosting methods. Optimal parameters were tuned for each algorithm. The test and training sets were split in a 7:3 ratio, and 10-fold cross-validation was used. The boosting model performed well on the dataset, achieving approximately 80% evaluation accuracy after 10-fold cross-validation. The contribution of different feature segments in the LightGBM model is shown below. Figure 5 As shown in the feature contribution graph of the LightGBM results, the angular features Xb_5 and Xb_2, and the electromyographic features Xa_3 and Xa_1, contribute the most to the results. In addition, the Xz_b angular index also contributes significantly to the results.

[0118] A single cyclic movement is decomposed into a series of posture segments. Each posture segment is then comprehensively evaluated based on factors such as standardization, speed, and completeness, ultimately yielding a quantitative evaluation index for the movement. Based on this quantitative movement evaluation method, a multimodal sensor-based movement evaluation model is proposed and validated using a straight leg raise movement. Results show that, for the selected movement, the proposed system uses a Gaussian Model (GMM) to subdivide the time-series movement into feature segments, and then trains a tree model to obtain a quantitative score. Requiring only one IMU and one EMG, it can achieve nearly 84% accuracy in quantitatively scoring movements, comparable to physician fuzzy comprehensive evaluation, providing fast, objective assessment and real-time feedback. The entire solution is low-cost, simple in algorithm, and has strong potential for widespread application.

Claims

1. A method for evaluating rehabilitation movements, characterized in that, Includes the following steps: S1. Set up rehabilitation movements, and select multiple training movements that can evaluate rehabilitation movements and their corresponding multiple significant motor features. The significant motion feature data of repeated standard actions of multiple demonstrators were collected by sensing devices. The data was preprocessed and then clustered and fitted on the time series using the Gaussian mixture model algorithm to obtain the Gaussian mixture model parameters and divide the feature segments to establish a demonstration action template with a single significant motion feature in a single repetition. S2. Use sensor devices to collect user rehabilitation test action data, and scale the single action cycle time to match the demonstration action template through interpolation. Use the action template constructed in S1 using the Gaussian mixture model algorithm to cluster and divide the integrated data into feature segments, and calculate the likelihood value of each feature segment. The computational methods for clustering using the Gaussian mixture model algorithm described in S1 and S2 include: 1) Calculate the probability density of a single Gaussian component fitted to the salient motion feature data, which is expressed as: In the formula, It is the expectation of the k-th component, with a dimension of 2; It is the covariance of the k-th component; These are sample data points in the action loop. N is determined by the sensor sampling frequency; 2) Fit the Gaussian components of multiple significant motion feature data. The mixture function of the multiple significant motion feature data components is expressed as: 3) The likelihood value is calculated using the likelihood function of the Gaussian mixture model, which is expressed as: 4) The EM algorithm is used to solve for the final parameters of the model; The E-step aims to complete the information related to latent variables by initializing the parameters, and its function is expressed as: In the formula, Given the parameters θ and the example data y, the result is... The expected value represents the probability that the demonstration data point j comes from component k under the current model parameters, which is the responsiveness of component k to the demonstration data point j. These are the model parameters from the previous iteration. In the first iteration, Parameters for initialization ; The M-step calculates the parameters for round i+1 by maximizing the likelihood, and its function is expressed as: The model parameters for the new iteration are obtained through the E-step and M-step, and the function is as follows: Repeat the E-step and M-step until convergence, i.e., the parameters tend to stabilize. The parameters of the Gaussian mixture model are thus determined. S3. Perform a fuzzy comprehensive evaluation on the user's rehabilitation test action data using a fuzzy comprehensive evaluation model, and calculate its comprehensive score. S4. The likelihood values ​​of each feature segment calculated in S2, combined with the scaling index, are used as the feature values ​​of the tree model method, and the comprehensive score calculated in S3 is used as the label of the tree model method. The score is quantified by the gradient boosting tree model algorithm to evaluate the current rehabilitation action.

2. The rehabilitation movement evaluation method as described in claim 1, characterized in that, The sensing devices described in S1 and S2 include a dual-channel electromyography sensor and an inertial sensor.

3. The rehabilitation movement evaluation method as described in claim 2, characterized in that, The data processing methods described in S1 include filtering of data acquired by the dual-lead electromyography sensor, regularization of negative data, Gaussian smoothing, and normalization.

4. The rehabilitation movement evaluation method as described in claim 1, characterized in that, The calculation method for scaling the single action loop time by interpolation as described in S2 is as follows: Among them, X z For scaling characteristics, tr represents the number of original interpolation data points, and ts represents the number of pre-determined template data points.

5. The rehabilitation movement evaluation method as described in claim 1, characterized in that, The method for completing fuzzy comprehensive evaluation using a fuzzy comprehensive evaluation model, as described in S2, includes: A fuzzy relation matrix is ​​established based on significant motion feature factors. The weights of the significant motion feature factors are calculated, and the results are used as the evaluation index weight vector of the fuzzy evaluation method. The weighted average operator is used to perform fuzzy comprehensive evaluation on each sample data in turn, and its comprehensive score is calculated.

6. The rehabilitation movement evaluation method as described in claim 5, characterized in that, The method for completing fuzzy comprehensive evaluation using a fuzzy comprehensive evaluation model, as described in S2, includes: 1) Determine the factor domain of the evaluation object. If there are m indicators in total: 2) Define the domain of evaluation ratings. The evaluation set is the collection of all possible overall evaluation results that an evaluator might make about the evaluated object: 3) Establish the fuzzy relation matrix R: Where, r mn Indicates that the evaluated object is determined from factor u m Let's look at the v n Membership degree of hierarchical fuzzy subsets; 4) By constructing the fuzzy weight vector A of the evaluation factors using the entropy weight method combined with expert weights, the fuzzy comprehensive evaluation model is expressed as: Among them, b n This indicates the evaluation object's performance on v under all evaluation factors. j The membership degree of a fuzzy subset of ratings, assuming the score of the comment set is defined as S(4, 3, 2, 1), is calculated as follows: B = b1×4 + b2×3 + b3×2 + b4×1.

7. The rehabilitation movement evaluation method as described in claim 1, characterized in that, The gradient boosting tree boosting method described in S4 uses an additive model and a forward distribution algorithm to complete the learning process by gradually reducing the residuals.

8. The rehabilitation movement evaluation method as described in claim 7, characterized in that, The gradient boosting tree algorithm used is Adaboost, XGBoost, LightGBM, or GBDT.

Images