A hand action recognition method based on a multimodal deep forest framework

Abstract

The present invention discloses a hand motion recognition method based on a multimodal deep forest framework, comprising: S1. obtaining electromyographic signals and acceleration signals corresponding to hand motions; S2. extracting the time domain features corresponding to the electromyographic signals and acceleration signals respectively to obtain electromyographic signal feature vectors and acceleration signal feature vectors; S3. fusing the electromyographic signal feature vectors and acceleration signal feature vectors to obtain a fused feature vector set; S4. optimizing the deep forest with determined hyperparameters; S5. inputting the feature vector set into the optimized deep forest to obtain an output recognition result of the hand motion. The present invention aims to explore the complementary effect of acceleration signals on electromyographic signals, prove that the fusion of acceleration signals can improve the accuracy of motion recognition, and promote the research and development of theories related to human motion, especially hand motion, as well as the application of electromyographic signal decoding in rehabilitation medicine, human-computer interaction and other inventions.

Description

Technical Field

[0001] The present invention relates to the field of sports rehabilitation, and in particular to a hand motion recognition method based on a multimodal deep forest framework. Background Art

[0002] Generating human gestures is a complex process. In recent years, with the advancement of human-computer interaction and the rapid development of computer technology, gesture recognition technology has been widely used in prosthetic medicine, rehabilitation training, and sports training. Despite these technological advances, existing prosthetics still cannot fully meet the actual dexterity needs of amputees. Therefore, how to naturally implant a multifunctional and dexterous prosthesis has become a challenge. A fundamental challenge for ensuring the accurate operation of these robotic devices is how to distinguish the user's intention from the surface electromyography signals obtained. Although significant progress has been made in gesture recognition accuracy in recent years, there is still significant room for improvement in classification accuracy for a large number of gestures. Summary of the Invention

[0003] The purpose of the present invention is to address the shortcomings of the existing technology and provide a hand motion recognition method based on a multimodal deep forest framework.

[0004] In order to achieve the above objectives, the present invention adopts the following technical solutions:

[0005] A hand action recognition method based on a multimodal deep forest framework, comprising:

[0006] S1. Acquire electromyographic signals and acceleration signals corresponding to hand movements;

[0007] S2. Extract the time domain features corresponding to the EMG signal and the acceleration signal, respectively, to obtain the EMG signal feature vector and the acceleration signal feature vector;

[0008] S3. Fusing the EMG signal feature vector and the acceleration signal feature vector to obtain a fused feature vector set;

[0009] S4. Optimize the deep forest with the determined hyperparameters.

[0010] S5. Input the feature vector set into the optimized deep forest to obtain the output recognition result of the hand movement.

[0011] Furthermore, the time domain features of the electromyographic signal in step S2 include the absolute average value of the electromyographic signal, the waveform length of the electromyographic signal, the autoregressive coefficient of the electromyographic signal, the number of zero crossings of the electromyographic signal, and the number of times the slope sign of the electromyographic signal changes.

[0012] Furthermore, the absolute average value of the electromyographic signal is expressed as:

[0013]

[0014] Wherein, MAV represents the absolute average value of the electromyographic signal; N represents the sliding window size; x i represents the instantaneous sEMG value at time point i.

[0015] Furthermore, the waveform length of the electromyographic signal is expressed as:

[0016]

[0017] Among them, N represents the sliding window size; x i represents the instantaneous sEMG value at time point i.

[0018] Furthermore, the autoregressive coefficient of the electromyographic signal is expressed as:

[0019]

[0020] Among them, x i,k represents the autoregressive coefficient of the electromyographic signal; p represents the model order of the autoregressive model; p j represents the jth coefficient of the autoregressive model; ε t represents the residual white noise; k represents the kth sampling point in the electromyographic signal sequence; x i represents the instantaneous sEMG value at time point i.

[0021] Furthermore, the zero crossing number of the electromyographic signal is expressed as:

[0022]

[0023] in,

[0024]

[0025] Where ε represents the threshold to avoid low-level noise; N represents the sliding window size; x i represents the instantaneous sEMG value at time point i.

[0026] Furthermore, the number of times the slope sign of the electromyographic signal changes is expressed as:

[0027]

[0028] in,

[0029]

[0030] Where ε represents the threshold to avoid low-level noise; N represents the sliding window size; x i represents the instantaneous sEMG value at time point i.

[0031] Further, the time-domain feature of the acceleration signal in step S2 includes the application average value of the acceleration signal.

[0032] Further, in step S3, dimensionality reduction analysis is also included for the acceleration signal feature vector, where the dimensionality reduction analysis is specifically as follows:

[0033] Set the acceleration matrix as X, expressed as:

[0034]

[0035] where n represents the number of samples and m represents the dimension of each sample;

[0036] Perform principal component analysis on the acceleration matrix, expressed as:

[0037] where

[0038] where, x1, x2,..., x p represents the original feature vector; z1, z2,..., z p (m < p) represents the new feature vector; l ij represents the coefficient.

[0039] ]Further, step S5 specifically includes:

[0040] S51. Send the obtained m-dimensional feature vector to the random forest of the deep forest algorithm, and each random forest obtains an n-dimensional data; where, each layer structure in the deep forest has 4 random forests, two of which are completely random forests and two are ordinary random forests, so 4 n-dimensional probability data are obtained;

[0041] S52. Concatenate the 4 n-dimensional probability data obtained in step S51 with the original acceleration matrix to obtain m + 4n-dimensional data, and use the obtained m + 4n-dimensional data as the input of the next layer, then 4 n-dimensional data are obtained;

[0042] S53. Concatenate the 4 n-dimensional probability data obtained in step S52 with the original acceleration matrix to obtain m + 4n-dimensional data, and use the obtained m + 4n-dimensional data as the input of the next layer, then 4 n-dimensional data are obtained;

[0043] S54. Continue to execute step S53 until the last layer is completed, and finally output 4 n-dimensional data; <]

[0044] S55. Take the average of the finally output 4 n-dimensional data to obtain an n-dimensional data;

[0045] S56. Take the largest value in the obtained n-dimensional data as the prediction result.

[0046] Compared with the existing technology, this invention proposes a multimodal deep forest framework based on the deep forest algorithm to identify hand movements, aiming to explore the complementary role of acceleration signals on electromyographic signals, and prove that the fusion of acceleration signals can improve the accuracy of motion recognition, and is conducive to promoting the research and development of theories related to human body movement, especially hand movement, and promoting the application of electromyographic signal decoding in rehabilitation medicine, human-computer interaction and other inventions. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a hand motion recognition method based on a multimodal deep forest framework provided in Example 1;

[0048] Figure 2 This is a schematic diagram of the sliding window method used in signal processing provided in Example 1;

[0049] Figure 3 This is a process diagram of feature extraction provided in Example 1;

[0050] Figure 4 This is a schematic diagram of the cascade forest principle provided in Example 1;

[0051] Figure 5 This is a hyperparameter confirmation diagram provided in Example 1;

[0052] Figure 6 is a confusion matrix diagram of a healthy subject and an amputee subject provided in Example 1;

[0053] Figure 7 is a schematic diagram of classification accuracy results for different inputs provided in Example 1;

[0054] Figure 8 This is a schematic diagram of the comparison results of the classification accuracy of the first embodiment with other traditional classifiers. DETAILED DESCRIPTION

[0055] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0056] The purpose of the present invention is to address the shortcomings of the existing technology and provide a hand motion recognition method based on a multimodal deep forest framework.

[0057] Example 1

[0058] This embodiment provides a hand action recognition method based on a multimodal deep forest framework. Figure 1 Shown, including:

[0059] S1. Acquire electromyographic signals and acceleration signals corresponding to hand movements;

[0060] S2. Extract the time domain features corresponding to the EMG signal and the acceleration signal, respectively, to obtain the EMG signal feature vector and the acceleration signal feature vector;

[0061] S3. Fusing the EMG signal feature vector and the acceleration signal feature vector to obtain a fused feature vector set;

[0062] S4. Optimize the deep forest with the determined hyperparameters.

[0063] S5. Input the feature vector set into the optimized deep forest to obtain the output recognition result of the hand movement.

[0064] In step S1 , myoelectric signals and acceleration signals corresponding to hand movements are acquired.

[0065] The human hand movements are preprocessed to extract the signals corresponding to the hand movements, and the extracted signals are divided into electromyographic signals and acceleration signals by sliding window analysis method. Figure 2 As shown, this embodiment sets a sliding window of 265ms and an increment of 50ms.

[0066] In step S2, the time domain features corresponding to the electromyographic signal and the acceleration signal are extracted respectively to obtain the electromyographic signal feature vector and the acceleration signal feature vector.

[0067] The five time domain features corresponding to the electromyographic signal are extracted, namely, the absolute average MAV, the waveform length WL, the autoregressive coefficient AR4, the number of zero crossings ZC, and the number of slope sign changes SSC; Figure 3 As shown in the figure, the methods for extracting the five time domain features of the electromyographic signal are as follows:

[0068] Mean Absolute Value (MAV): The MAV is a popular time-domain feature value in surface electromyography (sEMG) signal processing. It not only characterizes the magnitude of myoelectric power but can also be used to detect the starting position of the active segment of the sEMG signal. The MAV is expressed as:

[0069]

[0070] Wherein, MAV represents the absolute average value of the electromyographic signal; N is the sliding window size, x i is the instantaneous sEMG value at time point i.

[0071] Waveform Length (WL): Waveform length can be used to measure the complexity of surface EMG signals. Waveform length is expressed as:

[0072]

[0073] Where N is the sliding window size, x i is the instantaneous sEMG value at time point i.

[0074] Autoregressive coefficients (AR): The autoregressive model is a prediction model that treats the surface electromyographic signal as a pattern and then constructs different models to represent different actions. p , can be used as the eigenvalue matrix autoregressive coefficient, expressed as:

[0075]

[0076] Among them, x i,k represents the autoregressive coefficient of the electromyographic signal; p represents the model order of the autoregressive model; p j represents the jth coefficient of the autoregressive model; ε t represents the residual white noise; k represents the kth sampling point in the EMG signal sequence; N is the sliding window size, x i is the instantaneous sEMG value at time point i.

[0077] Zero Crossing: Counts the frequency of the sEMG signal passing through zero, while setting ε ≥ 0 to reduce the influence of noise in the signal. The zero crossing number is expressed as:

[0078]

[0079] in,

[0080]

[0081] ε is the threshold to avoid low-level noise, N is the sliding window size, x i is the instantaneous sEMG value at time point i. In this patent, ε=2 is set.

[0082] Slope Sign Change Count (SSC): This feature measures the frequency of slope changes. Given three consecutive samples, if the samples are as follows, then the SSC increases by 1. The slope sign change count is expressed as:

[0083]

[0084] Among them,

[0085]

[0086] ε ≥ 0, which is a preset threshold for reducing the influence of noise on features. N is the size of the sliding window, and x i is the instantaneous sEMG value at time point i.

[0087] For the acceleration signal, in this embodiment, the mean value (MEAN) is selected to extract features. It is expressed as:

[0088]

[0089] Among them, N is the number of sampling points used to calculate the MEAN value.

[0090] In step S3, the feature vectors of the EMG signal and the acceleration signal are fused to obtain a fused feature vector set.

[0091] After fusing the features of the surface EMG signal and the acceleration signal and conducting a hand movement pattern experiment, good experimental results can be obtained. However, the dimensionality of the feature values input to the classifier is too high, which not only reduces the classification ability and generalization ability of the classifier, but is also more unfavorable for embedding the algorithm into hardware in the future. Therefore, this embodiment considers using the principal component analysis method (PCA) to reduce the dimensionality of the feature values. However, according to multiple experiments, the dimensionality reduction effect of the feature values of the surface EMG signal is very poor. Therefore, this embodiment only performs dimensionality reduction analysis on the feature values of the acceleration signal. The dimensionality reduction analysis is specifically as follows:

[0092] Set the acceleration matrix as X, which is expressed as: [[ID=三十一]]

[0093] [[ID=3十四]] [[ID=三十五]]

[0094] Among them, n represents the number of samples, and m represents the dimension of each sample;

[0095] Perform principal component analysis on the acceleration matrix, which is expressed as:

[0096] Among them

[0097] Among them, x1, x2,..., x p represents the original feature vector; z1, z2,..., z p (m < p) represents the new feature vector; l ij represents the coefficient. The determination rule is as follows: z i and z jMutually unrelated, i, j∈[1,m], z1 is x1, x2,…, x p The maximum variance is the same as z m is related to z1, z2, ..., z m-1 All x1, x2, ..., x are unrelated p The maximum variance of z1, z2, ..., z is finally calculated. m That is, the original feature vector is x1, x2, ..., x p The main component of .

[0098] The measurement device in this embodiment has 12 electrodes, and the acceleration signal is used as a three-dimensional inertial measurement unit, which has 36 dimensions of raw data, that is, m is 36. After various data display, the 36 dimensions will be reduced to 18 dimensions to achieve the maximum accuracy. Subsequently, the acceleration features after dimensionality reduction are spliced ​​with the electromyographic signal features to form a raw feature vector T, which can be expressed as follows:

[0099] After feature extraction, the feature vector of the EMG sample can be defined as:

[0100] S EMG =[f 1, f 2, …,f 12 ],

[0101] in,

[0102] f i =[MAV i ,WL i ,AR i ,SSC i ,ZC i ].

[0103] Where i ranges from 1 to 12, indicating the channel number. Therefore, the length of the EMG feature vector is 96. For the ACC feature vector, the original length is 36, including 12 channels and 3 coordinate directions for each channel. In order to reduce the redundant information in the ACC signal, PCA-based dimensionality reduction is applied to obtain the first 18 main components, which can be defined as

[0104] S ACC =[MEAN1, MEAN2,…MEAN 18 ],

[0105] Among them, MEAN i represents the i-th principal component after PCA.

[0106] Therefore, the eigenvector T can be expressed as:

[0107] T=[S EMG ,S ACC,y],

[0108] Among them, y represents the label of 40 gestures.

[0109] In step S4, the deep forest with determined hyperparameters is optimized.

[0110] like Figure 4 The cascade forest principle diagram shown in the figure shows that in the cascade forest (deep forest), each level receives the feature information processed by the previous level to achieve layer-by-layer processing. Each layer contains four random forests. Two random forests set the number of split features of the tree node with the square root of the total number of features, and the other two forests set it with the logarithm of the total number of features. Each completely random forest is generated by randomly assigning a feature to split at each node and growing the tree to a pure leaf, that is, each leaf node contains only instances of the same category. Similarly, each random forest selects features as candidates (d is the number of input features) and select the feature with the best Gini value for splitting. The number of trees in each forest is a hyperparameter. The deep forest model initially assumed that the number of trees contained in it was 500 trees, whether it was a completely random forest or an ordinary random forest. In general, increasing the number of trees in each random forest will improve the accuracy of the experiment, and can also make the deviation constant and variance smaller. But this improvement is not endless. After the accuracy converges, the accuracy of the random forest will not increase with the increase in number. In order to balance the computational burden and experimental accuracy, the most appropriate number of trees was explored. This example used 40 categories of gestures from 20 complete subjects of NinaPro DB7 as experimental data. In order to determine the number of trees in the forest, the surface electromyography data of the first subject was used to evaluate how the number of trees affects the classification accuracy. As Figure 5 It shows how the accuracy changes as the number of trees increases. Figure 5 As can be seen in the figure, the accuracy improves significantly with the increase in the number of trees. When the number of trees increases to 125, the accuracy improvement becomes stable. To balance computing resources and classification accuracy, this example sets the number of trees in each random forest to 125.

[0111] The training process of deep forests can be formalized in terms of the supervised learning problem of learning a mapping from feature space X to label space Y, where Y = {1,2,3…,C}; let Z = [0,1] c and training set S=((x1,y1),…,(x m ,y m )) is drawn from the distribution D. The deep forest model is defined by the triple (h,f,l), where

[0112] h=(h1,…,h T ), where h tis the forest set of t layers, h t It is the hypothesis class H t members.

[0113] ·f=(f 1, …,f T ), where f t is a cascade forest up to level t.

[0114] l=(l1,…l T ), where l t is the validation error at layer t.

[0115] Number of layers t∈{1,…T}, f t :X→Z is defined as follows:

[0116]

[0117] In each layer t, h t (·) and f t (·) will produce a set of probability vectors where p i is the predicted probability of category i. When the number of layers is t=1, h t The input of the layer is x, and the input of other layers is [x,f t-1 (x)].

[0118] At layer t, if the input verification error l at layer t-1 t-1 Verification error l with layer number t t The difference between t-1 -l t ) is less than or equal to the threshold η, then it needs to pass the next level. Until it reaches the Tth layer, the difference between the verification error and the T-1 layer is greater than the threshold η, and the number of layers of the deep forest is finally determined. Each triple (h, f, l) defines a deep forest model g:X→Y as follows:

[0119]

[0120] where t′ = arg t∈{1,…,T} (l t-1 -l t )≤η.

[0121] Algorithm 1 summarizes the training algorithm of Deep Forest.

[0122] In the current study, each forest in a layer can generate 40 probability values ​​for each category (i.e., gesture). Therefore, each layer will generate 160 probability values ​​as the enhanced features produced by the four random forests. Except for the first layer, the input of all subsequent layers is the sum of the 160 enhanced features and the original 114 feature vectors from EMG and ACC.

[0123] The implementation of the deep forest training algorithm is shown in Table 1:

[0124] Table 1

[0125]

[0126]

[0127] In step S5, the feature vector set is input into the optimized deep forest to obtain the output recognition result of the hand movement.

[0128] S51. For n-class data, the m-dimensional feature vector T is fed into a random forest. Each random forest generates an n-dimensional data set. Each layer of the cascade forest has four random forests, two of which are completely random forests and two are ordinary random forests. This gives us four n-dimensional probability data sets.

[0129] S52. The original X is concatenated to form m+4n-dimensional data. This data is used as the input for the next layer, and four n-dimensional data are obtained.

[0130] S53. Then concatenate it with the original X to obtain m+4n-dimensional input data, which serves as the input for the next layer.

[0131] S54. When it is passed to the last layer, there is no need to splice the original data together again, because it is the last layer and does not need to be spliced ​​together as the input of the next layer. Finally, 4 n-dimensional data are output;

[0132] S55. Average the four n-dimensional data to obtain an n-dimensional data;

[0133] S56. Take the largest value in this n-dimensional data as the prediction result.

[0134] The number of layers in the multi-granularity cascade forest is adaptively adjusted. During the construction phase of the cascade forest, as long as the cross-validated verification accuracy of the current layer does not improve compared to the previous layer, the construction of the cascade forest stops and the entire structure is completed.

[0135] Each gesture class in the database of this example has six replicates. In each experiment, we used five replicates to train the Deep Forest, following the same protocol as the database, and the remaining replicate was used to evaluate the Deep Forest. This process was then repeated using a different evaluation fold in each iteration, resulting in six-fold cross-validation. The classification results from these six replicates were then averaged.

[0136] like Figure 6The confusion matrices for intact subject S1 and amputee S22, obtained using the present invention, are shown. As can be seen, the overall accuracy for S1 and S22 is approximately 90% and 80%, respectively. Both matrices demonstrate that most classes are correctly predicted, and that the present invention outperforms intact subjects compared to amputee subjects. Furthermore, it can be seen that most misclassifications occur between adjacent classes due to similarity.

[0137] at the same time, Figure 7 The classification accuracy of intact subjects and amputee subjects for three different data inputs, EMG+ACC, EMG, and ACC, is shown respectively. (A) is the experimental result before dimensionality reduction, and (B) is the result of dimensionality reduction of ACC features using PCA. Figure 7 The results also demonstrate that dimensionality reduction of the ACC signal can enhance the performance of deep forests. Specifically, the accuracy of EMG+ACC dimensionality reduction in intact subjects was 91.40±2.02%, 2.07% higher than the accuracy without dimensionality reduction. Furthermore, in amputee subjects, the accuracy of EMG+ACC dimensionality reduction was 77.80±9.61%, 2.57% higher than the accuracy without dimensionality reduction.

[0138] In this example, SVM, KNN, and RF were also used to test the data. During the SVM tuning process, the penalty coefficient was set to 0.8, and the linear kernel function was selected as the classification parameter. In KNN, the parameter was set to k=5. In the random forest, the parameter of the decision tree was set to 110, and the impurity was based on information entropy. Figure 8 As shown, by comparing with deep forests using other classic classifiers, the gap can be clearly seen, which can prove the superiority of the present invention.

[0139] in, Figure 8 is with Figure 6 Another corresponding form of data visualization.

[0140] Compared with the existing technology, this invention proposes a multimodal deep forest framework based on the deep forest algorithm to identify hand movements, aiming to explore the complementary role of acceleration signals on electromyographic signals, and prove that the fusion of acceleration signals can improve the accuracy of motion recognition, and is conducive to promoting the research and development of theories related to human body movement, especially hand movement, and promoting the application of electromyographic signal decoding in rehabilitation medicine, human-computer interaction and other inventions.

[0141] Note that the above are only preferred embodiments of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments, and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments and may include many other equivalent embodiments without departing from the concept of the present invention. The scope of the present invention is determined by the scope of the appended claims.

Claims

1. A hand action recognition method based on a multimodal deep forest framework, characterized in that: include: S1. Acquire electromyographic signals and acceleration signals corresponding to hand movements; S2. Extract the time domain features corresponding to the EMG signal and the acceleration signal respectively to obtain the EMG signal feature vector and the acceleration signal feature vector; the time domain features of the EMG signal include the absolute average value of the EMG signal, the waveform length of the EMG signal, the autoregressive coefficient of the EMG signal, the number of zero crossings of the EMG signal, and the number of changes in the slope sign of the EMG signal; the time domain features of the acceleration signal include the application average value of the acceleration signal; S3. Fusion of the EMG signal feature vector and the acceleration signal feature vector to obtain a fused feature vector set; further comprising dimensionality reduction analysis of the acceleration signal feature vector; S4. Optimize the deep forest with the determined hyperparameters. S5. Input the feature vector set into the optimized deep forest to obtain the output recognition result of the hand movement; Step S5 specifically includes: S51. The obtained m-dimensional feature vector is sent to the random forest of the Deep Forest algorithm. Each random forest generates an n-dimensional data. Each layer of the Deep Forest algorithm has four random forests, two black ones are completely random forests, and two are ordinary random forests, thus generating four n-dimensional probability data. S52. Concatenate the four n-dimensional probability data obtained in step S51 with the original acceleration matrix to obtain m + 4n-dimensional data. The obtained m + 4n-dimensional data is used as the input of the next layer, thereby obtaining four n-dimensional data. S53. Concatenate the four n-dimensional probability data obtained in step S52 with the original acceleration matrix to obtain m + 4n-dimensional data. The obtained m + 4n-dimensional data is used as the input of the next layer, thus obtaining four n-dimensional data. S54. Continue to execute step S53 until the last layer is executed, and finally output 4 n-dimensional data; S55. Average the last four n-dimensional data output to obtain an n-dimensional data; S56. Take the largest value in the obtained n-dimensional data as the prediction result.

2. The hand motion recognition method based on a multimodal deep forest framework according to claim 1, characterized in that: The absolute average value of the electromyographic signal is expressed as: ， Wherein, MAV represents the absolute average value of the electromyographic signal; N represents the sliding window size; represents the instantaneous sEMG value at time point i.

3. The hand motion recognition method based on a multimodal deep forest framework according to claim 2 is characterized in that: The waveform length of the electromyographic signal is expressed as: ， Where N represents the sliding window size; represents the instantaneous sEMG value at time point i.

4. The hand motion recognition method based on a multimodal deep forest framework according to claim 2, characterized in that: The autoregressive coefficient of the electromyographic signal is expressed as: ， in, represents the autoregressive coefficient of the electromyographic signal; p represents the model order of the autoregressive model; represents the j-th coefficient of the autoregressive model; represents the residual white noise; k represents the kth sampling point in the electromyographic signal sequence; represents the instantaneous sEMG value at time point i.

5. The hand action recognition method based on a multimodal deep forest framework according to claim 2, characterized in that: The zero crossing number of the electromyographic signal is expressed as: , in, , in, represents the threshold to avoid low-level noise; N represents the sliding window size; represents the instantaneous sEMG value at time point i.

6. The hand motion recognition method based on a multimodal deep forest framework according to claim 2, characterized in that: The number of times the slope sign of the electromyographic signal changes is expressed as: ， in, ， in, represents the threshold to avoid low-level noise; N represents the sliding window size; represents the instantaneous sEMG value at time point i.

7. The hand motion recognition method based on a multimodal deep forest framework according to claim 6, characterized in that: The dimensionality reduction analysis in step S3 is specifically as follows: Set the acceleration matrix to X, expressed as: ， Among them, n represents the number of samples, and m represents the dimension of each sample; Perform principal component analysis on the acceleration matrix, which is expressed as: ,in , in, represents the original feature vector; represents the new feature vector; Represents the coefficient, m <p。

Images