Exoskeleton intention recognition window parameter selection method based on surface electromyogram signals

Abstract

The invention discloses an exoskeleton intention recognition window parameter selection method based on a surface electromyogram signal, and relates to the technical field of biomedical signal processing. The method comprises the following steps: respectively acquiring electromyographic signals on the surfaces of an upper limb and a lower limb through a Ninapro data set and an sEMG data set, preprocessing the acquired data by adopting a filter and mu-Law normalization, dividing windows by combining different parameters of window length and overlapping degree according to the preprocessed data, performing feature extraction according to a time-domain feature value group and a frequency-domain feature value group, and performing feature extraction according to a time-domain feature value group and a frequency-domain feature value group. Calculating an optimal window length and an optimal overlapping degree by using a category divergence parameter selection algorithm; performing comparison verification according to the obtained optimal result and calculation results of a plurality of classifiers; according to the exoskeleton intention recognition window parameter selection method based on the surface electromyogram signals, a window category distribution measurement algorithm for measuring different window lengths and overlapping degrees is provided, and a large number of repeated classifier training and verification processes are avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of biomedical signal processing, and in particular to a method for selecting window parameters for exoskeleton intention recognition based on surface electromyogram signals. Background Art

[0002] With the aggravation of the problem of global population aging, the demand for exoskeleton rehabilitation robots is increasing day by day. How to accurately and quickly recognize the intention of the wearer using sEMG signals is one of the research goals in the field of exoskeleton applications; as a non-contact, non-invasive, and advanced physiological signal, sEMG is the preferred way to control exoskeletons. Surface electromyogram signals are generated by muscle movement. When neurons control muscle fibers to generate excitement, action potentials will be generated in the muscle fibers. The action potentials generated by all muscle fibers on the skin surface are superimposed on the skin surface to form surface electromyogram signals. Surface electromyogram signals contain intention information of movement and are widely used in posture and action recognition.

[0003] Based on feature extraction, related research has explored the relationship between window length and overlap degree on accuracy. The window size can be reduced to obtain higher accuracy. However, after combining the window length and overlap degree, multiple classifiers need to be trained on multiple data sets to find the optimal combination in the given set combination. This process requires a large number of experiments and takes a long time. And from specific parameters, different data sets are different. Although many studies have obtained the optimal combination in specific scenarios, they lack interpretability and general applicability. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for selecting window parameters for exoskeleton intention recognition based on surface electromyogram signals, and proposes a window class distribution metric algorithm for measuring different window lengths and overlap degrees. After sampling each combination with the minimum number of samples, the traces of the between-class matrix and the within-class matrix are used as the basis to select the combination of these two parameters, avoiding a large number of repeated classifier training and verification processes.

[0005] To achieve the above object, the present invention provides a method for selecting window parameters for exoskeleton intention recognition based on surface electromyogram signals, including the following steps:

[0006] S1. Collect electromyogram signals on the surfaces of the upper and lower limbs through the Ninapro data set and the EMG data set respectively;

[0007] S2. Preprocess the data collected in S1 by using a fourth-order Butterworth filter and a μ-Law normalization method;

[0008] S3. Divide different parameter combinations of window length and overlap degree according to the data preprocessed in S2 to form data sets;

[0009] S4. Extract features from the new dataset divided in S3 according to the time-domain eigenvalue group and the frequency-domain eigenvalue group;

[0010] S5. Calculate the optimal window length and overlap degree using the class divergence parameter selection algorithm based on the eigenvalues extracted in S4;

[0011] S6. Compare and verify the optimal result obtained in S5 with the calculation results of several classifiers.

[0012] Preferably, the calculation formula of the Butterworth filter in S2 is as follows:

[0013]

[0014] where |H(f)| 2 is the amplitude square function of the filter, f is the frequency, f c is the cut-off frequency, N is the order of the filter, and here N = 4 is taken.

[0015] Preferably, the calculation formula of the μ-law normalization in S2 is as follows:

[0016]

[0017] where x norm represents the data after normalization, μ is the compression parameter, x max is the maximum absolute value of the input signal, and sign(x) represents the sign of the original data retained.

[0018] Preferably, the window division process in S3 is as follows:

[0019] Suppose a time series is T and the data length is |T|, then the number of available samples N after window segmentation of this time series is:

[0020]

[0021] where l is the length of each sliding window and s is the data length of the overlapping part between windows;

[0022] Suppose the window length set is L and the overlap degree set is S, and the parameter sets of the two are combined together as {(s i , l j ) | s i ∈S, l j ∈L}; for different parameter combinations, the number of samples N is also different. Therefore, each time a comparison is made, the samples with the least number in the set combination are sampled according to the class ratio, and the formula is as follows:

[0023]

[0024] Among them, N sample is the number of samples after proportional sampling, and N i，j is the sample digital code under the parameter combination of (s i , l j ). i and j are the subscripts of the overlapping degree and window length.

[0025] Preferably, after the window division in S3, the feature extraction process and formula for each window are as follows:

[0026] f t = FE(w t ); (5)

[0027] Pred t = CLS(f t ); (6)

[0028] Among them, FE represents the feature extraction function, w t represents the t-th window, Pred t represents the predicted classification result, and w t ∈T, f t represents the features after passing through the feature extractor, including all features in the time domain and frequency domain here, and CLS represents the classifier.

[0029] Preferably, the time domain eigenvalue group described in S4 includes: Mean Absolute Value (MAV), Root Mean Square (RMS), and Waveform Length (WL);

[0030] The calculation formula for the Mean Absolute Value (MAV) is as follows:

[0031]

[0032] The calculation formula for the Root Mean Square (RMS) is as follows:

[0033]

[0034] The calculation formula for the Waveform Length (WL) is as follows:

[0035]

[0036] Among them, x i represents the sEMG data with a window length of L;

[0037] The frequency domain eigenvalue group includes: Median Frequency (MF), Mean Power Frequency (MPF), and Mean Frequency (MNF);

[0038] The calculation formula for the Median Frequency (MF) is as follows:

[0039]

[0040] Among them, f median is the median frequency, and S(f i ) represents the power spectral density of each frequency component;

[0041] The calculation formula for the mean power frequency MPF is as follows:

[0042]

[0043] The calculation formula for the mean frequency MNF is as follows:

[0044]

[0045] Among them, X(f i ) represents the frequency domain data after the fast Fourier transform FFT.

[0046] Preferably, the specific steps of the category divergence parameter selection algorithm in S5 are as follows:

[0047] First, perform normalization processing on each sample feature, and the formula is as follows:

[0048]

[0049] x t is the normalized feature data. For the extracted feature f t in the normalization process, here we adopt the maximum-minimum normalization, and respectively represent the minimum and maximum values of f t ;

[0050] Suppose there are C categories, then the within-class scatter matrix S w and the between-class scatter matrix S b are defined by the following formulas:

[0051]

[0052] Among them, x represents the sample data, k ∈ C, and n k represents the number of samples after sampling in the k-th category, CC k represents the mean vector of the feature space of the k-th class, and OC represents the mean vector of all samples;

[0053] The above-mentioned CC k and OC are calculated by the following formulas respectively:

[0054]

[0055] The final optimization objective is as follows:

[0056]

[0057] where, trace(S w ) represents the trace of the within-class scatter matrix, and trace(S b ) represents the trace of the between-class scatter matrix.

[0058] Preferably, the classifier described in S6 includes: K-Nearest Neighbor Algorithm (KNN), Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), Random Forest (RF), and Neural Network (NN), and the training of each classifier adopts the method of five-fold cross-validation.

[0059] Therefore, the present invention adopts a method for selecting window parameters for exoskeleton intention recognition based on surface electromyogram signals as described above. Compared with the prior art, it has the following beneficial effects:

[0060] 1. This application uses the between-class distribution matrix and the within-class distribution matrix to select the optimal parameter combination, which is an effective feature selection and optimization method. It can improve the performance of the machine learning model, and at the same time provide more robust and interpretable results. By maximizing the between-class difference and minimizing the within-class difference, different classes of data can be better separated, thereby improving the accuracy and performance of the classifier, and reducing the computational complexity of subsequent processing and analysis, and improving the efficiency of the entire system;

[0061] 2. Through window partitioning, this application can more accurately capture the key information in the signal, while the selection of class divergence parameters helps to identify which signal features are the most important for classification. Combining the two can improve the accuracy of overall signal processing. By selecting the parameters that best represent the class differences, the model can better generalize to new data and reduce the risk of overfitting. Window partitioning and parameter selection can also reduce unnecessary feature calculations and only focus on the features that contribute the most to classification, thereby reducing the computational complexity and resource consumption.

[0062] Next, through the drawings and embodiments, the technical solutions of the present invention will be further described in detail. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is the overall flowchart of a method for selecting window parameters for exoskeleton intention recognition based on surface electromyogram signals according to the present invention;

[0064] Figure 2 is the relationship diagram of window length - trace of the scatter matrix for a method for selecting window parameters for exoskeleton intention recognition based on surface electromyogram signals according to the present invention;

[0065] Figure 3 is the relationship diagram of window length - classifier accuracy (time-domain features) for a method for selecting window parameters for exoskeleton intention recognition based on surface electromyogram signals according to the present invention;

[0066] Figure 4It is a graph showing the relationship between the window length and the classifier accuracy (frequency domain features) of a method for selecting window parameters for exoskeleton intention recognition based on surface electromyogram signals in the present invention;

[0067] Figure 5 It is a graph showing the relationship between the overlapping degree and the trace of the divergence matrix of a method for selecting window parameters for exoskeleton intention recognition based on surface electromyogram signals in the present invention;

[0068] Figure 6 It is a graph showing the relationship between the overlapping degree and the classifier accuracy (time domain features) of a method for selecting window parameters for exoskeleton intention recognition based on surface electromyogram signals in the present invention;

[0069] Figure 7 It is a graph showing the relationship between the overlapping degree and the classifier accuracy (frequency domain features) of a method for selecting window parameters for exoskeleton intention recognition based on surface electromyogram signals in the present invention. Detailed implementation manners

[0070] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "upper", "lower", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the product of the present invention is usually placed during use. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention.

[0071] Embodiment

[0072] As Figure 1 shown, a method for selecting window parameters for exoskeleton intention recognition based on surface electromyogram signals of the present invention includes the following steps:

[0073] S1. Collect electromyogram signals on the surfaces of the upper and lower limbs through the Ninapro dataset and the EMG dataset respectively;

[0074] Among them, the EMG dataset in Lower LimbData was collected by the MWX8 Datalog Biometrics device, including the analysis of the relationship between muscle behavior and the knee joint of 11 healthy humans: walking, stretching the leg from a sitting position, and knee flexion while standing. During the collection process, 4 electrodes (rectus femoris, semitendinosus, biceps femoris, and vastus medialis) and a knee joint angle measurement device were used. Each proposed movement included 3 to 5 repetitions in each group. The sampling rate of the dataset was 1 kHz, and it was called the lower limb surface electromyogram intention recognition dataset; the Ninapro dataset included sEMG and motion data of 27 healthy subjects during 52 repeated hand movements (A, B, C) plus a rest posture. The experiment included finger, wrist, and hand movements. The data was obtained using 10 Otto Bock MyoBock 13E200 electrodes, that is, 10 channels. The sampling rate of the dataset was 2 kHz, and it was called the upper limb surface electromyogram intention recognition dataset;

[0075] S2. Preprocess the data collected in S1 using a fourth-order Butterworth filter and the μ-Law normalization method; the calculation formula of the Butterworth filter in S2 is as follows:

[0076]

[0077] Among them, |H(f)| 2 is the magnitude squared function of the filter, f is the frequency, f c is the cut-off frequency, N is the order of the filter, and here N = 4 is taken.

[0078] The calculation formula of the μ-law normalization is as follows:

[0079]

[0080] Among them, x norm represents the normalized data, μ is the compression parameter (default value 255), x max is the maximum absolute value of the input signal, and sign(x) represents retaining the sign of the original data;

[0081] S3. Divide the windows with different parameter combinations of window length and overlap degree according to the data preprocessed in S2 to form a dataset;

[0082] The process of dividing the windows is as follows:

[0083] Suppose a time series is T and the data length is |T|, then the number of available samples N after window slicing of this time series is:

[0084]

[0085] Among them, l is the length of each sliding window, and s is the data length of the overlapping part between windows. To ensure the real-time performance of exoskeleton control, l should be guaranteed to be within 400 ms. However, for performance comparison, the test is still carried out up to 450 ms.

[0086] Let the set of window lengths be L and the set of overlapping degrees be S. The parameter sets of the two are combined together as {(s i ,l j )|s i ∈S,l j ∈L}; for different parameter combinations, the number of samples N is also different. Therefore, each time for comparison, the samples with the least number in the set combination are sampled according to the class ratio. The formula is as follows:

[0087]

[0088] Among them, N sample is the number of samples after proportional sampling, N i，j is the sample number under the parameter combination of (s i ,l j ), and i and j respectively correspond to the subscripts of the overlapping degree and window length;

[0089] After window division, the feature extraction process for each window and the formula are as follows:

[0090] f t =FE(w t ); (5)

[0091] Pred t =CLS(f t ); (6)

[0092] Among them, FE represents the function of feature extraction, w t represents the t-th window, Pred t represents the predicted classification result, and w t ∈T, f t represents the features after passing through the feature extractor, including all features in the time domain and frequency domain here, and CLS represents the classifier;

[0093] S4. Feature extraction is performed on the new data set divided in S3 according to the time-domain eigenvalue group and frequency-domain eigenvalue group; among them, the time-domain eigenvalue group includes: mean absolute value MAV, root mean square value RMS, and waveform length WL;

[0094] The calculation formula for the mean absolute value MAV is as follows:

[0095]

[0096] The calculation formula for the root mean square value RMS is as follows:

[0097]

[0098] The calculation formula for the waveform length WL is as follows:

[0099]

[0100] where x i represents the sEMG data with a window length of L;

[0101] The frequency domain eigenvalue group includes: median frequency MF, mean power frequency MPF, and mean frequency MNF;

[0102] The calculation formula for the median frequency MF is as follows:

[0103]

[0104] where f median is the median frequency, and S(f i ) represents the power spectral density of each frequency component;

[0105] The calculation formula for the mean power frequency MPF is as follows:

[0106]

[0107] The calculation formula for the mean frequency MNF is as follows:

[0108]

[0109] where X(f i ) represents the frequency domain data after the fast Fourier transform FFT;

[0110] S5. Use the class divergence parameter selection algorithm to calculate the optimal window length and overlap degree based on the eigenvalues extracted in S4; the specific steps of the class divergence parameter selection algorithm are as follows:

[0111] First, perform normalization processing on each sample feature, and the formula is as follows:

[0112]

[0113] x t The normalized feature data, for the extracted feature f t In the normalization process, here we use the maximum-minimum normalization, and represent the minimum and maximum values of f t respectively;

[0114] Assume there are C classes, then the within-class scatter matrix S wand the between-class scatter matrix S b The defining formula is as follows:

[0115]

[0116] where x represents the sample data, k ∈ C, and n k represents the number of samples after sampling for the k-th class, CC k represents the mean vector of the feature space of the k-th class, and OC represents the mean vector of all samples;

[0117] The above CC k and the calculation formulas for OC are shown as follows respectively:

[0118]

[0119] The trace of the within-class scatter matrix measures the degree of dispersion among samples of the same class. A small trace value indicates high similarity among samples within a class. The trace of the between-class scatter matrix measures the degree of separation between different classes. A large trace value indicates better distinguishability between classes;

[0120] In a classification task, it is usually desired that the trace of the between-class scatter matrix is as large as possible, while the trace of the within-class scatter matrix is as small as possible, so that it is easier to distinguish between classes;

[0121] The final optimization objective is as follows:

[0122]

[0123] where trace(S w ) represents the trace of the within-class scatter matrix, and trace(S b ) represents the trace of the between-class scatter matrix;

[0124] That is, in a given set of common combinations, find the combination that makes the difference in traces as small as possible. This combination is the optimal window length and overlap degree;

[0125] S6. Compare and verify the optimal result obtained according to S5 with the calculation results of several classifiers; the classifiers include: K-Nearest Neighbors (KNN), Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), Random Forest (RF), Neural Network (NN), and the training of each classifier adopts the method of five-fold cross-validation.

[0126] In the specific implementation process, we performed experimental operations (cross-validation) on the sEMG signals of each subject in the two datasets. First, we divided the time-domain and frequency-domain features of the time-series data into five channels, sampled the samples after normalization, and calculated them according to various combinations. The window length set is {50ms, 100ms, 150ms, 200ms, 250ms, 300ms, 350ms, 400ms, 450ms}, and the overlap degree set is {10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%}, and we combined these two sets.

[0127] At the same time, for visual comparison and analysis, among the selected parameter combinations, the overlap degree was fixed at 50%, and the data with the window length changing and the window length fixed at 350ms and the overlap degree changing were compared. To verify the effectiveness of the algorithm, we gave the mean and variance of the accuracy of different classifiers in the lower limb dataset under different combinations to explore the optimal window accuracy found.

[0128] For each subject, the distribution of the training set and the test set is shown in Table 1:

[0129] Table 1 Distribution of the training set and the test set

[0130] Number of subjects Number of actions Training set Test set Dataset 1 11 4 1，2，4，5 3 Dataset 2 27 12 1，3，5，6 2，4

[0131] As Figures 2 - 4 shown, the changing trends of the distance metric algorithm and the accuracy results of each classifier corresponding to the fixed overlap degree of 50% and the changing window length are given; it can be seen from Figure 2 that as the window length increases, the inter-class distance increases, and at the same time the intra-class distance decreases; at the same time, the effect of time-domain feature extraction is generally better than that of frequency-domain features, which is reflected in the inter-class distance being larger than that of the frequency domain and the intra-class distance being smaller than that of the frequency domain. Combining with Figure 3 and Figure 4 it can be seen that the changing trends of the classification accuracy and the distance scale are the same, and when it is 350ms - 450ms, the local optimum of the distance scale is reached, and the accuracy of each classifier also reaches the highest at this window size; at the same time, the classifier results of time-domain features are generally better than those of frequency-domain features, which is the same as the changing trend of Difference in Figure 2 , reflecting the effectiveness of the algorithm.

[0132] Therefore, when the overlap degree is 50%, the window length of this dataset can be set to 350ms, which also meets the response speed requirements of the exoskeleton; at the same time, the experiment also proves that for some classifiers with high linear requirements, such as SVM and LDA, it proves the effectiveness of this algorithm. For KNN, as the intra-class distance decreases, its accuracy naturally increases.

[0133] As Figures 5 - 7 shown, the changing trends of the distance metric algorithms corresponding to a window length of 350 ms with a fixed overlap degree and the accuracy results of each classifier are given.

[0134] It can be seen from Figure 5 that as the overlap degree increases, the inter-class distance increases while the intra-class distance decreases. The changing trend of the classification accuracy is the same as that of the distance metric. Combining Figure 2 and Figure 5 it is concluded that the influence degree of the overlap degree is smaller than that of the window length. This is mainly because the increase in the overlap degree results in more stable extracted features and a higher similarity degree among samples, as there is more repeated data before and after. Similarly, since the Difference of the time-domain features is better than that of the frequency-domain features when the overlap degree increases, the overall accuracy of the classifier in Figure 6 is better than that in Figure 7 .

[0135] Therefore, for a window size of 350 ms, the optimal overlap degree of this dataset is 90%, which meets the response speed requirements. At the same time, the experiment also proves that the window size and overlap degree have little influence on random forests and neural networks.

[0136] Therefore, the present invention adopts a method for selecting window parameters for exoskeleton intention recognition based on surface electromyography signals as described above, explores the relationship between the window length and overlap degree and the classification results, and designs a window category distribution metric algorithm to select the optimal window size and overlap degree to achieve the intention recognition of sEMG. After sampling each parameter combination with the minimum number of samples, the traces of the inter-class matrix and intra-class matrix are used as the basis to find the optimal parameter combination.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for selecting window parameters for exoskeleton intention recognition based on surface electromyography signals, characterized in that: S1, collect electromyographic signals of the upper and lower limbs through the Ninapro dataset and EMG dataset respectively; S2, preprocess the data collected by S1 using the fourth-order Butterworth filter and μ-Law normalization method; S3, dividing the windows into data sets by combining different parameters of window length and overlap degree according to the data preprocessed by S2; S4, extracting features from the new data set divided by S3 according to the time domain eigenvalue group and the frequency domain eigenvalue group; S5, using the category divergence parameter selection algorithm based on the feature values ​​extracted in S4 to calculate the optimal window length and overlap degree; S6. Compare and verify the optimal result obtained in S5 with the calculation results of several classifiers.

2. The method for selecting parameters of an exoskeleton intention recognition window based on surface electromyography signals according to claim 1, characterized in that: The calculation formula of the Butterworth filter in S2 is as follows: Where |H(f)| 2 is the square function of the filter amplitude, f is the frequency, f c is the cut-off frequency, N is the order of the filter, here N=4.

3. The method for selecting parameters of an exoskeleton intention recognition window based on surface electromyography signals according to claim 2, characterized in that: The calculation formula for μ-law normalization described in S2 is as follows: Among them, x norm represents the normalized data, μ is the compression parameter, x max is the maximum absolute value of the input signal, and sign(x) means retaining the sign of the original data.

4. The method for selecting parameters of an exoskeleton intention recognition window based on surface electromyography signals according to claim 3, characterized in that: The window division process described in S3 is as follows: Suppose a time series is T and the data length is |T|, then the number of available samples N after this time series is split by the window is: Among them, l is the length of each sliding window, and s is the length of the data overlapping between windows; Assume that the window length set is L, the overlap degree set is S, and the parameter sets of the two are combined into {(s i ,l j )|s i ∈S,l j ∈L}; For different parameter combinations, the number of samples N is also different, so each time the comparison is performed, the least sample in the set combination is sampled according to the category ratio. The formula is as follows: Among them, N sample is the number of samples after proportional sampling, N i，j for (s i ,l j ) parameter combination, i and j correspond to the subscripts of overlap degree and window length respectively.

5. The method for selecting parameters of an exoskeleton intention recognition window based on surface electromyography signals according to claim 4, characterized in that: After the window division described in S3, the feature extraction process and formula for each window are as follows: f t =FE(w t ); (5) Pred t =CLS(f t ); (6) Among them, FE represents the function of feature extraction, w t represents the tth window, Pred t represents the predicted classification result, and w t ∈T,f t It represents the features after the feature extractor, including all the features in the time domain and frequency domain. CLS represents the classifier.

6. The method for selecting parameters of an exoskeleton intention recognition window based on surface electromyography signals according to claim 5, characterized in that: The time domain feature value group in S4 includes: mean absolute value MAV, root mean square value RMS and waveform length WL; The calculation formula of the mean absolute value MAV is as follows: The calculation formula of the root mean square value RMS is as follows: The calculation formula of the waveform length WL is as follows: Among them, x i represents sEMG data with a window length of L; The frequency domain feature value group includes: median frequency MF, mean power frequency MPF and mean frequency MNF; The calculation formula of median frequency MF is as follows: MF=f median in Among them, f median is the median frequency, S(f i ) represents the power spectral density of each frequency component; The calculation formula of average power frequency MPF is as follows: The calculation formula of mean frequency MNF is as follows: Among them, X(f i ) represents the frequency domain data after fast Fourier transform FFT.

7. The method for selecting parameters of an exoskeleton intention recognition window based on surface electromyography signals according to claim 6, characterized in that: The specific steps of the category divergence parameter selection algorithm in S5 are as follows: First, each sample feature is normalized, the formula is as follows: x t After normalization, the feature data is extracted. t Normalization process, here we use maximum and minimum value normalization, and Represents f t The minimum and maximum values ​​of Suppose there are C categories, then the intra-class scatter matrix S w and the inter-class scatter matrix S b The definition formula is as follows: Among them, x represents sample data, k∈C, n k Represents the number of samples of the kth category, CC k represents the mean vector of the feature space of class k, and OC represents the mean vector of all samples; The above CC k The calculation formulas for OC and OC are as follows: The final optimization goals are as follows: Among them, trace(S w ) represents the trace of the intra-class scatter matrix, trace(S b ) represents the trace of the between-class scatter matrix.

8. The method for selecting parameters of an exoskeleton intention recognition window based on surface electromyography signals according to claim 7, characterized in that: The classifiers described in S6 include: K-nearest neighbor algorithm KNN, linear discriminant analysis LDA, support vector machine SVM, random forest RF, neural network NN, and the training of each classifier adopts a five-fold cross-validation method.

Images