A lower limb motion recognition method based on s-transform energy concentration surface electromyography decoding

Abstract

A method for lower limb movement recognition based on surface electromyography (EMG) decoding using S-transform energy concentration is proposed. First, the raw EMG signal is preprocessed. The preprocessed signal is then truncated to a useful time segment using endpoint detection. Next, the signal within this time segment undergoes segmented S-transform and energy concentration calculation. Segmented operations are used to extract signal features of a specified dimension. Movement pattern classification is performed using SVM, and multi-channel signal features are fused and analyzed for lower limb movement recognition. This invention performs segmented S-transform on the signal to preserve its temporal information, while optimizing the energy concentration calculation process to improve movement classification accuracy. An SVM multi-classifier is built, and the accuracy of lower limb movement pattern recognition is further improved through S-transform energy concentration and feature fusion.

Description

Technical Field

[0001] This invention relates to the field of pattern recognition technology in human-computer interaction, and in particular to a method for recognizing lower limb movements based on S-transform energy concentration surface electromyography decoding. Background Technology

[0002] With the continuous development of human-computer interaction technology, the needs of amputees have shifted from bulky passive prostheses to intelligent prostheses, which can recognize human movement intentions and actively adjust output torque or joint angles to achieve a gait similar to that of a healthy person, making them more comfortable to use.

[0003] Existing lower limb prostheses use mechanical sensing signals (Huang Pinggao. Research on key technologies for sensing and recognizing the movement intention of intelligent lower limb prostheses [D]. University of Chinese Academy of Sciences (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), 2020.), but the inherent time delay of mechanical signals causes their control process to lag behind human gait. Some prostheses use electromyography (EMG) signals for control, but current lower limb EMG decoding extracts simple time-frequency domain features of the signals (Sánchez-Velasco LE, Arias-Montiel M, Guzmán-Ramírez E, Lugo-González E. A low-cost EMG-controlled anthropomorphic robotic hand for power and precision grasp. Biocybern BiomedEng 2020;40:221–37.), losing phase information that is indispensable in motion pattern analysis. This results in poor recognition of lower limb movement patterns and fails to achieve natural interaction between patient intention and prosthetic control. Therefore, it is necessary to explore methods for recognizing human movement intentions by decoding electromyographic signals on the lower limb surface, so as to further improve the accuracy of lower limb movement intention recognition and enrich the movement control modes of lower limb prostheses. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, the present invention aims to provide a lower limb movement recognition method based on surface electromyography decoding using S-transform energy concentration. The method performs segmented S-transform on the signal to preserve its temporal information, while optimizing the calculation process of energy concentration to improve the accuracy of movement classification. An SVM multi-classifier is built, and the accuracy of lower limb movement pattern recognition is further improved by using S-transform energy concentration and feature fusion.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for lower limb movement recognition based on surface electromyography decoding using S-transform energy concentration is proposed. First, the raw electromyography signal is preprocessed. The preprocessed signal is then truncated to a useful time period through endpoint detection. Then, the signal within this time period is subjected to segmented S-transform and energy concentration calculation. Segmented operation is used to extract signal features of a specified dimension. Movement pattern classification is performed using SVM, and multi-channel signal features are fused and analyzed to perform lower limb movement recognition.

[0007] A method for lower limb movement recognition based on surface electromyography decoding with S-transform energy concentration includes the following steps:

[0008] 1) Signal preprocessing and active segment detection;

[0009] 2) Segmented S-transform of signal and concentrated energy calculation;

[0010] 3) Construction of a support vector machine multi-classifier and correlation analysis of multi-channel electromyography signals;

[0011] 4) Support vector machine motion classification based on multi-channel signal feature fusion analysis.

[0012] The signal preprocessing in step 1) includes: 1.1) 50Hz notch filtering; 1.2) 30Hz zero-phase-shift high-pass filtering; a zero-phase-shift filter is used in the preprocessing process to preserve the phase information of the original data sequence.

[0013] In step 1), the active segment detection adopts a dual threshold detection method based on short-time energy and short-time variance. First, the initial minimum threshold of short-time energy and the variance and low threshold representing the fluctuation are set. Both parameters are dynamically adjusted according to the actual situation. Then, the window length and window shift are defined to divide the signal into frames.

[0014] Assume the electromyography signal of the nth frame is represented as Given a frame length of N, calculate its short-time energy. As shown in equation (1), calculate its variance and sum. As shown in equation (2),

[0015] (1)

[0016] (2)

[0017] In the formula, It is the sample index of the signal, representing the sample point in each frame. From 0 to An integer of 1 represents the position of each sampling point in this frame of signal. This represents the mean of the signal.

[0018] In step 2), the signal segmentation S-transform and energy concentration calculation are as follows: the data length of one action cycle is denoted as x, a sliding window with a window length and step size is set, the S-transform is performed on each segment of data, the frequency interval is set, and a time-frequency matrix is ​​obtained; as shown in equation (3), the 1 / q root of the absolute value of the signal after the S-transform is superimposed and summed in the time-frequency domain, and then the summation result is subjected to q-square operation; by optimizing the width of the window function in the S-transform, a higher energy concentration is obtained for signals with rapidly changing frequency components, and its energy concentration calculation is shown in equation (4); the energy concentration measurement method is used as shown in equation (5), the entire time-frequency concentration is decomposed, and the window width at the discrete point is used to improve its time-frequency concentration;

[0019] (3)

[0020] (4)

[0021] (5)

[0022] In the formula: — Signal energy concentration; -- time; — Frequency; — Signal S-transformation matrix, F —Discrete frequency range; CM 1—Time-frequency concentration, q—a nonlinearity parameter that controls the degree of focusing; the larger the q, the more emphasis is placed on the high-value region. —Normalized time-frequency concentration —— The normalized form, —Generalized time-frequency concentration; —Low-power parameters typically emphasize global distribution. High-order parameters typically emphasize local concentration. --yes The normalized form ensures the scale consistency of time-frequency representation.

[0023] Step 3) involves building a multi-classifier using a support vector machine. Specifically, a multi-classifier is designed. For a given set of m classes, a binary classifier is trained for every two classes. The total number of binary classifiers is m(m-1) / 2. For a set of data that needs to be classified, it needs to be predicted by all classifiers, and a voting method is used to determine its final class attribute.

[0024] Step 3) involves the correlation analysis of multi-channel electromyographic signals. Specifically, by analyzing the correlation of electromyographic signals, the differences in classification effects between single-channel and multi-channel signal fusion are explored, and the optimal combination of lower limb muscles is found to improve the recognition accuracy of human lower limb movements.

[0025] Step 4) of the support vector machine motion classification based on multi-channel signal feature fusion analysis is as follows: The similarity between the lower limb muscles of different subjects is different, so it is necessary to calculate the similarity of the multi-channel signal features of the subjects and select the channels with better classification results from the muscle signals of the rectus femoris, vastus medialis, biceps femoris, semitendinosus, tibialis anterior, and medial gastrocnemius for fusion; by comparing the classification results of single-channel signals, the multi-channel signal features of the rectus femoris, biceps femoris and medial gastrocnemius are fused and analyzed, and the fusion coefficient is assigned by the control variable method. Then, the features with different weights are input into the classifier for motion classification.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] This invention analyzes the shortcomings of existing motion recognition methods based on input features and proposes a lower limb motion recognition method based on surface electromyography (EMG) decoding using S-transform energy concentration. The S-transform is a combination of wavelet transform and short-time Fourier transform, possessing the characteristics of Fourier transform while also providing different resolutions at different frequencies through window functions. This invention extracts simple time-frequency domain features and S-transform energy concentration features from lower limb EMG signals and uses support vector machines for classification. Furthermore, it analyzes the similarity of six lower limb muscles and improves motion recognition accuracy through multi-channel signal feature fusion. Compared to simple time-frequency domain statistical feature analysis, this invention effectively improves classification accuracy, and further enhances accuracy through feature fusion of channels such as the rectus femoris, biceps femoris, and gastrocnemius muscle. Attached Figure Description

[0028] Figure 1 This is a flowchart of lower limb movement recognition according to an embodiment of the present invention.

[0029] Figure 2 This is a flowchart of the signal active segment detection process according to an embodiment of the present invention.

[0030] Figure 3 This is the result of the signal segmentation S-transformation in an embodiment of the present invention.

[0031] Figure 4 These are the energy concentration calculation results for the four motions in the embodiments of the present invention.

[0032] Figure 5 This is a flowchart of the SVM multi-classification process in an embodiment of the present invention.

[0033] Figure 6This is a simple time-frequency domain feature recognition result of an embodiment of the present invention.

[0034] Figure 7 This is the result of energy concentration feature recognition in embodiment S of the present invention.

[0035] Figure 8 This is a comparison of the signal classification results of the rectus femoris muscle in an embodiment of the present invention.

[0036] Figure 9 This is a comparison of the results of multi-channel fusion and single-channel classification of electromyographic signals in an embodiment of the present invention. Detailed Implementation

[0037] The technical solution of the present invention will be further described and illustrated below with reference to the embodiments and accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0038] Reference Figure 1 A method for lower limb movement recognition based on surface electromyography decoding with S-transform energy concentration includes the following steps:

[0039] 1) Signal preprocessing and active segment detection;

[0040] 2) Segmented S-transform of signal and concentrated energy calculation;

[0041] 3) Construction of a support vector machine multi-classifier and correlation analysis of multi-channel electromyography signals;

[0042] 4) Support vector machine motion classification based on multi-channel signal feature fusion analysis.

[0043] The signal preprocessing in step 1) includes: 1.1) 50Hz notch filtering to remove power frequency interference; 1.2) 30Hz zero-phase-shift high-pass filtering to remove motion artifacts. In the preprocessing process, a zero-phase-shift filter is used to preserve the phase information of the original data sequence, which makes up for the phase distortion of the original filter.

[0044] Reference Figure 2 In step 1), the active segment detection adopts a dual threshold detection method based on short-time energy and short-time variance. First, the initial minimum threshold of short-time energy and the variance and low threshold representing the fluctuation are set. Both parameters are dynamically adjusted according to the actual situation. Then, the window length and window shift are defined to divide the signal into frames.

[0045] Assume the electromyography signal of the nth frame is represented as Given a frame length of N, calculate its short-time energy. As shown in equation (1), calculate its variance and sum. As shown in equation (2),

[0046] (1)

[0047] (2)

[0048] In the formula, It is the sample index of the signal, representing the sample point in each frame. From 0 to An integer of 1 represents the position of each sampling point in this frame of signal. Indicates the mean of the signal;

[0049] Reference Figure 3 In step 2), the signal segmentation S-transform and energy concentration calculation are as follows: A data length of x for one action cycle is extracted. A sliding window with a window length (x / 25) and a step size (x / 25) is set. An S-transform is performed on each data segment, with a frequency interval of 2Hz, resulting in a time-frequency matrix of size 16*x / 25. In the signal segmentation S-transform results, the darker the blue portion, the lower its time-frequency energy; as the color gradually changes to yellow, the higher its time-frequency energy. It can be seen that the energy is mainly concentrated between 100-300Hz. After the S-transform, the frequency, phase, and energy amplitude of the corresponding electromyographic signal on the action surface are displayed in a three-dimensional image.

[0050] As shown in Equation (3), the 1 / q root of the absolute value of the signal after S-transformation is superimposed and summed in the time-frequency domain, and then the summation result is subjected to q-square operation; by optimizing the width of the window function in S-transformation, a higher energy concentration can be obtained for signals with rapidly changing frequency components, and its energy concentration calculation is shown in Equation (4); the energy concentration measurement method is shown in Equation (5), the entire time-frequency concentration is decomposed, and the window width at discrete points is fully utilized to improve its time-frequency concentration;

[0051] (3)

[0052] (4)

[0053] (5)

[0054] In the formula: — Signal energy concentration; -- time; — Frequency; — Signal S-transformation matrix; F —Discrete frequency range; CM 1—Time-frequency concentration, q—a nonlinearity parameter that controls the degree of focusing; the larger the q, the more emphasis is placed on the high-value region. —Normalized time-frequency concentration —— The normalized form, —Generalized time-frequency concentration; —Low-power parameters typically emphasize global distribution. High-order parameters typically emphasize local concentration. --yes Normalization ensures scale consistency in time-frequency representation;

[0055] Reference Figure 4 In this embodiment, the segmented S-transform energy concentration calculation results of four actions, namely walking on flat ground, going up stairs, going down stairs, and crossing obstacles, retain 25 feature data of the original electromyographic signal, better preserve the phase information of the signal, and the energy amplitude difference is also more obvious. Figure 4 The image shows the energy concentration of the rectus femoris muscle under four actions (walking on flat ground, climbing stairs, descending stairs, and crossing obstacles). Each sub-image corresponds to one action, with the horizontal axis representing the number of features (from 1 to 25) and the vertical axis representing the energy value. The distribution characteristics of the four actions are as follows: Walking on flat ground: The energy value is relatively low, and the fluctuation amplitude is relatively stable, indicating a relatively uniform signal energy distribution. Climbing stairs: A significant energy peak appears near feature 15, reflecting a higher energy concentration for this specific feature. Descending stairs: Similar to climbing stairs, the overall energy fluctuation is larger, but not as high as the energy peak at feature 15. Crossing obstacles: The energy value fluctuation amplitude is larger, and the overall energy distribution is more dispersed compared to other actions, with higher energy values ​​for some features, especially in the range of features 20-25. Features and phase information: The data retains 25 features, indicating that the feature extraction through piecewise S-transform is sufficient to decompose the original signal, not only preserving phase information but also enhancing the differentiation of energy amplitude in specific actions, which is beneficial to improving the accuracy of motion recognition.

[0056] Reference Figure 5 In step 3), the support vector machine multi-classifier construction is specifically as follows: design a multi-classifier. For a given set of m classes, train a binary classifier for every two classes. The total number of binary classifiers is m(m-1) / 2. To achieve four classifications, six classifiers are needed, specifically for: class 1 and class 2, class 1 and class 3, class 1 and class 4, class 2 and class 3, class 2 and class 4, and class 3 and class 4. For a piece of data that needs to be classified, it needs to be predicted by all classifiers, and the final class attribute is determined by voting.

[0057] Step 3) involves the correlation analysis of multi-channel electromyographic signals. Specifically, by analyzing the correlation of electromyographic signals, the differences in classification effects between single-channel and multi-channel signal fusion are explored, and the optimal combination of lower limb muscles is found to improve the recognition accuracy of human lower limb movements.

[0058] Reference Figure 6 , Figure 7 , Figure 8 In this embodiment, multi-channel signal fusion involves processing the raw electromyography (EMG) data to obtain a simple time-frequency domain feature matrix and an S-transform energy concentration feature matrix of the lower limb surface EMG signal. Support vector machines are then used to classify these two types of features. A comprehensive analysis of the mean, minimum, and concentration of accuracy in the motion classification results reveals that the rectus femoris and medial gastrocnemius muscles perform exceptionally well. Considering the experimental results of both methods, single-channel muscle signals are chosen for practical applications, with the rectus femoris signal being the most suitable for motion classification and recognition. Based on the assessment of the pattern recognition performance, preparations are made for selecting appropriate channels for fusion.

[0059] Figure 6 The mean classification accuracy of six-channel electromyography (EMG) signals based on simple time-frequency domain features for the 10 participants was as follows: 80.28% (rectus femoris), 82.87% (vastus medialis), 71.71% (biceps femoris), 83.06% (semitendinosus), 81.41% (tibialis anterior), and 82.59% (gastrocnemius medialis). The maximum classification accuracy of the six-channel muscle signals was 91.89% (rectus femoris) and 100%. 97.30% (vastus medialis), 94.60% (semitendinosus), 91.67% (tibialis anterior), 90.91% (gastrocnemius medialis); the minimum classification accuracy of the six-channel muscle signals were 67.86% (rectus femoris), 52.94% (vastus medialis), 47.06% (biceps femoris), 64% (semitendinosus), 50% (tibialis anterior), and 58.82% (gastrocnemius medialis).

[0060] Five participants achieved the highest classification accuracy in the semitendinosus muscle signal classification, three participants achieved the highest classification accuracy in the rectus femoris muscle signal classification, and two participants achieved the highest classification accuracy in the vastus medialis muscle signal classification. The highest classification accuracy was 100%, which was the classification result of the vastus medialis muscle for participant nine. However, by observing the overall classification results of the vastus medialis muscle, it was found that both the mean and minimum classification accuracy were lower than those of the semitendinosus muscle, indicating that there were significant differences in muscle signal classification among different participants. The mean maximum accuracy was 83.06% (semitendinosus muscle). In summary, the muscle channels with better classification performance based on time-frequency domain features were the semitendinosus, rectus femoris, and vastus medialis muscle. Among them, the classification accuracy and stability of the first two were higher than those of other channels. The classification effect of the vastus medialis muscle varied from person to person, and its stability was slightly less than that of the semitendinosus and rectus femoris muscle classification results.

[0061] Figure 7The mean accuracy of six-channel electromyographic signal feature recognition based on S-transform energy concentration characteristics for 10 subjects was as follows: 94.20% (rectus femoris), 93.20% (vastus medialis), 95.58% (biceps femoris), 92.26% (semitendinosus), 91.93% (tibialis anterior), and 93.73% (gastrocnemius medialis). The maximum classification accuracy of the six-channel muscle signals was 100% (rectus femoris) and 97.20% (vastus medialis). 30% (vastus medialis), 100% (biceps femoris), 100% (semitendinosus), 100% (tibialis anterior), 100% (gastrocnemius medialis); the minimum classification accuracy values ​​for the six-channel muscle signals were: 85.71% (rectus femoris), 70.59% (vastus medialis), 78.57% (biceps femoris), 77.42% (semitendinosus), 64.29% (tibialis anterior), and 83.33% (gastrocnemius medialis).

[0062] Except for the vastus medialis, the maximum classification accuracy of the other five muscle signals reached 100%. A comprehensive analysis of the mean, minimum, and concentration of accuracy revealed that the rectus femoris and gastrocnemius muscles performed exceptionally well. Considering the results of both methods, in practical applications using single-channel muscle signals, the rectus femoris signal is the most suitable for action classification and recognition.

[0063] Reference Figure 8 The above analysis revealed that the semitendinosus and rectus femoris muscles showed better classification performance among simple time-frequency domain features, while the medial gastrocnemius and rectus femoris muscles showed better classification performance among S-transform energy concentration features. Taking the rectus femoris signal as an example, the classification results of ten subjects under both methods were statistically analyzed. It is evident that for each subject, the S-transform energy concentration method outperformed the simple time-frequency domain feature method in terms of classification performance. The average accuracy of the former was 93.70%, while the average accuracy of the latter was 80.71%. Further analysis of the classification performance of the six-channel signals of the ten subjects showed that the S-transform energy concentration method was superior in 91.67% of the results. Therefore, the S-transform energy concentration method does indeed have better pattern recognition performance than extracting simple time-frequency domain features.

[0064] Reference Figure 9In step 4), the support vector machine motion classification based on multi-channel signal feature fusion analysis is as follows: the similarity between lower limb muscles of different subjects varies, so similarity calculation needs to be performed on the multi-channel signal features of the subjects. Channels with better classification performance among muscle signals such as rectus femoris, vastus medialis, biceps femoris, semitendinosus, tibialis anterior, and medial gastrocnemius are selected for fusion. By comparing the classification results of single-channel signals, in order to further improve accuracy, the multi-channel signal features of rectus femoris, biceps femoris, and medial gastrocnemius are fused and analyzed. The fusion coefficient is assigned by the control variable method, and then the features with different weights are input into the classifier for motion classification. Through multi-channel muscle signal feature fusion, the accuracy of subject action recognition can be effectively improved.

[0065] The preceding analysis showed that the rectus femoris muscle classification results were superior. To further improve accuracy, multi-channel signal features of the rectus femoris, biceps femoris, and medial gastrocnemius muscles were fused and analyzed. The multi-channel fusion classification results of the four subjects were compared with the single-channel classification results. Figure 9 In the electromyographic signal channel 1, the rectus femoris is represented; in channel 2, the vastus medialis is represented; in channel 3, the biceps femoris is represented; in channel 4, the semitendinosus is represented; in channel 5, the tibialis anterior is represented; in channel 6, the gastrocnemius is represented; and in channel 7, the fusion analysis results of the rectus femoris, biceps femoris, and gastrocnemius are represented.

[0066] Assuming the fusion coefficients for the three signals are a, b, and c, in the experiment with subject 1, it was found that when a=1, b=0.2, and c=0.5, the classification accuracy reached 100%, and the subject achieved a classification accuracy of 85.7140% using the single-channel signal from the rectus femoris muscle. Similarly, in the experiments with subjects 2 and 4, using a multi-channel fusion method for classification yielded even higher classification accuracy. In the experiment with subject 3, it was found that when a=1, b=0.2~0.7, and c=0.3, the classification accuracy reached 100%, and the subject achieved a classification accuracy of 96.9697% using the single-channel signal from the rectus femoris muscle. This demonstrates that multi-channel muscle signal feature fusion can effectively improve the accuracy of motion recognition for subjects.

[0067] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the present invention, and the patent protection scope of the present invention should be defined by the claims.

Claims

1. A method for lower limb movement recognition based on S-transform energy concentration surface electromyography decoding, characterized in that, Includes the following steps: 1) Signal preprocessing and active segment detection: First, the raw electromyographic signal is preprocessed, and the preprocessed signal is then processed by endpoint detection to extract the useful time segment; 2) Segmented S-transform and concentrated energy calculation: Then, the signal within the time period is segmented and S-transformed and concentrated energy is calculated. The segmented operation is used to extract signal features of a specified dimension. The specific steps of signal segmentation S-transformation and energy concentration calculation are as follows: extract the data length of one action cycle as x, set a sliding window with a window length and step size, perform S-transformation on each segment of data, set the frequency interval, and obtain a time-frequency matrix; as shown in equation (1), the 1 / q root of the absolute value of the signal after S-transformation is superimposed and summed in the time-frequency domain, and then the summation result is subjected to q-square operation; By optimizing the width of the window function in the S-transform, a higher energy concentration can be obtained for signals with rapidly changing frequency components. The energy concentration calculation is shown in Equation (2). The energy concentration measurement method is as shown in Equation (3), which decomposes the entire time-frequency concentration and uses the window width at discrete points to improve its time-frequency concentration. （1） （2） （3） In the formula: — Signal energy concentration; -- time; — Frequency; — Signal S-transformation matrix; F —Discrete frequency range; CM 1—Time-frequency concentration, q—a non-linearity parameter that controls the degree of focusing; the larger the q, the more emphasis is placed on the high-value region. —Normalized time-frequency concentration —— The normalized form, —Generalized time-frequency concentration; —Low-power parameters, emphasizing global distribution. —Higher power parameters emphasize local concentration. --yes Normalization ensures scale consistency in time-frequency representation; 3) Construction of a support vector machine multi-classifier and correlation analysis of multi-channel electromyography signals; The multi-channel electromyography (EMG) signal correlation analysis specifically involves: analyzing the correlation of EMG signals to explore the differences in classification effects between single-channel and multi-channel signal fusion, finding the optimal lower limb muscle combination. Since the similarity between lower limb muscles varies among different subjects, the correlation between multi-channel signal features of the rectus femoris, vastus medialis, biceps femoris, semitendinosus, tibialis anterior, and gastrocnemius medialis needs to be calculated. Based on the correlation analysis results, channels with good classification effects are selected for fusion. By comparing the classification results of single-channel signals, the S-transform energy concentration features of the rectus femoris, biceps femoris, and gastrocnemius medialis are fused and analyzed. Fusion coefficients are assigned using the controlled variable method, and then features with different weights are input into the classifier for motion classification. 4) Support Vector Machine Motion Classification Based on Multi-channel Signal Feature Fusion Analysis: Motion pattern classification is performed using SVM, and multi-channel signal features are fused and analyzed to identify lower limb movements.

2. The lower limb movement recognition method according to claim 1, characterized in that, The signal preprocessing in step 1) includes: 1.1) 50Hz notch filtering to remove power frequency interference; 1.2) 30Hz zero-phase-shift high-pass filtering to remove motion artifacts; a zero-phase-shift filter is used in the preprocessing process to preserve the phase information of the original data sequence.

3. The lower limb movement recognition method according to claim 1, characterized in that, In step 1), the active segment detection adopts a dual threshold detection method based on short-time energy and short-time variance. First, the initial minimum threshold of short-time energy and the variance and low threshold representing the fluctuation are set. Both parameters are dynamically adjusted according to the actual situation. Then, the window length and window shift are defined to divide the signal into frames. Assume the electromyography signal of the nth frame is represented as Given a frame length of N, calculate its short-time energy. As shown in equation (4), calculate its variance and sum. As shown in equation (5), （4） （5） In the formula, It is the sample index of the signal, representing the sample point in each frame. From 0 to An integer of 1 represents the position of each sampling point in this frame of signal. This represents the mean of the signal.

4. The lower limb movement recognition method according to claim 1, characterized in that, Step 3) involves building a multi-classifier using a support vector machine. Specifically, a multi-classifier is designed. For a given set of m classes, a binary classifier is trained for every two classes. The total number of binary classifiers is m(m-1) / 2. For a set of data that needs to be classified, it needs to be predicted by all classifiers, and a voting method is used to determine its final class attribute.

Images