A surface myoelectric robust intelligent recognition method for robots

Abstract

The application discloses a surface myoelectric robust intelligent identification method for a robot, and comprises the following steps: firstly, collecting a plurality of groups of original human myoelectric signals by using a surface myoelectric sensor; then, performing sliding window slicing processing on the collected sensor data, and dividing training data sets and test data sets; taking the training data sets as the input of a time domain processing subnetwork and a frequency domain processing subnetwork to perform network training, extracting time domain features and frequency domain features respectively after the network training, taking the time domain features and the frequency domain features as the input of an identification classifier, and training the identification classifier; finally, inputting a test sample set into a trained signal processing identification model to perform surface myoelectric signal processing identification; the application effectively combines time domain and frequency domain information, and compared with a traditional complete combination network, the application improves the identification accuracy and stability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent identification, and particularly relates to a surface electromyography robust intelligent identification method for a robot. BACKGROUND

[0002] With the rapid development of computer technology, artificial intelligence, signal processing, human-computer interaction, brain-like chip and other technologies, as the signal generated by the nervous system when the muscle contracts, surface electromyography (sEMG) controls and reflects the physiological and anatomical characteristics of the muscle, and its signal analysis capability has been significantly improved. Researchers have adopted various signal preprocessing, feature extraction and classification techniques to achieve more efficient processing of electromyography signals. As a physiological signal closely related to human motion, sEMG signals have been widely used in human-computer interaction systems. Analysis of sEMG signals can monitor, evaluate and control muscle activity, thereby promoting the development of human-computer interaction.

[0003] However, there are still some challenges in the process of sEMG signal identification. Deep learning algorithms are widely used in sEMG-based signal identification due to their strong classification ability. Deep learning-based methods usually assume that the training and test data are consistent. However, in real-world applications such as sEMG identification systems, training and test data are often disturbed by various factors. Users need to perform specific gestures when obtaining supervised samples, and the classifier trained from these samples performs significantly worse when applied to other users or the same user wearing the device. From an algorithmic perspective, the generalization ability of current sEMG-based classification and identification systems is limited. In addition, as more and more complex features are extracted, the computational burden also increases. SUMMARY

[0004] The present application is a surface electromyography robust intelligent identification method for a robot, which is based on a signal processing identification model to achieve more accurate surface electromyography signal identification, and can effectively solve the problem of significant decline in generalization effect and high computational complexity of current sEMG classification and identification algorithms in processing multi-subject analysis.

[0005] The technical solution of the present application is as follows: a surface electromyography robust intelligent identification method for a robot, comprising the following steps:

[0006] Step 1: Use surface electromyography sensors to collect several groups of original electromyography signals of the human body, slice the original electromyography signals through a sliding window, and divide the training data set and the test data set.

[0007] Step 2: Construct a signal processing identification model, including a time domain processing subnetwork, a frequency domain processing subnetwork and a DBN network.

[0008] Step 3: The training data set is taken as the input data of the time domain processing subnetwork, and the time domain processing subnetwork is trained to obtain time domain features:

[0009] The time domain processing subnetwork performs SSA processing on the input data to obtain an SSA result, connects the SSA result with the original electromyographic signal corresponding to the SSA result to obtain a time domain training sample, extracts features through an Xception module, and then obtains time domain features through a global adaptive average pooling layer.

[0010] Step 4: The training data set is taken as the input data of the frequency domain processing subnetwork, and the frequency domain processing subnetwork is trained to obtain frequency domain features:

[0011] The frequency domain processing subnetwork performs FFT processing on the input data to obtain an FFT result, connects the FFT result with the original electromyographic signal corresponding to the FFT result to obtain a frequency domain training sample, extracts features through an OmniScale module, and then obtains frequency domain features through a global adaptive average pooling layer.

[0012] Step 5: A DBN network is constructed, the time domain features and the frequency domain features are taken as the input of the DBN network, the DBN network is taken as a recognition classifier, the DBN network is trained, and a trained signal processing recognition model is obtained:

[0013] The DBN network performs feature splicing on the time domain features and the frequency domain features to obtain time-frequency features, then takes the time-frequency features as the input of the RBM layer of the DBN network for unsupervised training, finally adjusts the network parameters through supervised training of the DBN network to obtain the trained signal processing recognition model.

[0014] Step 6: The test data set is input into the trained signal processing recognition model to obtain a recognition result, and the surface electromyographic signal processing recognition for a robot is performed.

[0015] Compared with the prior art, the present application has the following advantages:

[0016] (1) The present application first proposes a surface electromyographic robust intelligent recognition method for a robot, which is based on a signal processing recognition model and applied to the intelligent recognition of surface electromyographic signals.

[0017] (2) The designed surface electromyographic robust intelligent recognition method for a robot effectively combines time domain and frequency domain information, so that the model can comprehensively understand and process surface electromyographic signals. Compared with the traditional complete combination network, the signal processing recognition model network algorithm designed in the present application has better stability, stronger generalization, and higher fault diagnosis accuracy in the human upper limb surface electromyographic signal recognition task. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a surface electromyography robust intelligent recognition method flow chart for a robot.

[0019] Figure 2 is a confusion matrix diagram of the recognition result of the signal processing recognition model.

[0020] Figure 3 is a bar chart of the diagnostic accuracy of different recognition methods under different tasks. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be described in detail below with reference to the drawings and specific embodiments.

[0022] sEMG contains accurate and rich human information, and the technical research of sEMG fusion human robot has attracted great attention of many scholars. At present, the generalization ability of the sEMG-based classification recognition system is limited, and with the extraction of more and complex features, the calculation burden also increases. Aiming at the problem that the generalization ability of the sEMG-based classification recognition network is limited, the effect is significantly reduced when processing multi-agent analysis and the high computational complexity problem, a signal processing recognition model is designed. In order to fully consider the time domain and frequency domain signals of the original electromyography signal, the time domain processing sub-network and the frequency domain processing sub-network are introduced to extract the features of the time domain and frequency domain signals respectively, and finally the deep belief network DBN is used to further fuse and process the extracted features of the two networks to obtain the classification result of sEMG. The network can obtain excellent recognition effect without complex parameter adjustment in upper limb movement classification, and has good generalization effect in multi-agent analysis.

[0023] In combination with Figure 1 , the surface electromyography robust intelligent recognition method for a robot provided by the present application first uses the MA400-18 acquisition system and the MA-420 electromyography signal amplifier to collect experimental data; the experiment is set to 12 groups of specified upper limb movements, and the experimental personnel collects the surface electromyography signal according to the specified movement within the specified time; then, the collected sensor data is subjected to sliding window slicing processing, and the training data set and the test data set are divided; the training data set is used as the input of the time domain processing sub-network, the time domain processing sub-network is trained, the time domain features are obtained, the training data set is used as the input of the frequency domain processing sub-network, the frequency domain processing sub-network is trained, the frequency domain features are obtained, and then the two extracted features are combined as the input of the restricted Boltzmann machine layer in the deep belief network for unsupervised training; finally, the network parameters are fine-tuned through the supervised training of the deep belief network, and the training of the whole signal processing and recognition model is completed; in combination with Figure 2 、 Figure 3It can be seen that this process effectively combines time domain and frequency domain information, enabling the model to comprehensively understand and process surface electromyography signals, significantly improving recognition accuracy and stability compared to traditional fully combined networks. The specific steps are as follows:

[0024] Step 1: Collect several groups of human original electromyography signals using surface electromyography sensors; use MA400-18 collection system and MA-420 electromyography signal amplifier to collect experimental data, set 12 groups of specified upper limb actions, and collect surface electromyography signals according to the specified actions within the specified time. Slice the collected multi-sensor related signals and divide them into training data set and test data set.

[0025] The experiment collected relevant data from the right hands of four healthy men. During data collection, all participants completed 12 basic actions according to the instructions of the observers, guided by a metronome. Each action was repeated five times, each lasting 5 seconds. After each basic action, the participants rested for 2 minutes to prevent muscle fatigue. This task mode continued until all experimental actions were completed. The joint actions mainly included shoulder flexion, shoulder extension, shoulder abduction, internal rotation, external rotation, elbow flexion, forearm elevation, forearm pronation, wrist flexion, wrist extension, ulnar deviation, and radial deviation.

[0026] Step 2: Construct a signal processing recognition model, including a time domain processing subnetwork, a frequency domain processing subnetwork, and a DBN network.

[0027] Step 3: Use the training data set as the input data of the time domain processing subnetwork to train the time domain processing subnetwork and obtain the time domain features.

[0028] The time domain processing subnetwork processes the input data using SSA to obtain SSA results, connects the SSA results with their corresponding original electromyography signals to obtain time domain training samples, extracts features through the Xception module, and then obtains time domain features through the global adaptive average pooling layer. The specific steps are as follows:

[0029] Step 3.1: The training data set is first processed using SSA to obtain SSA results.

[0030] Let represent the original surface electromyography signal, where c represents the number of channels, and the signal of channel c is represented as A c = (a0, a1, a2...a N ), where N is the length of the time series, a N represents the Nth point of the time series; T represents transposition.

[0031] SSA is based on multivariate statistical principle, which reveals the potential structure and pattern in the original EMG signal by extracting the main components of time series; each channel is processed separately, define Len as the length of sliding window, X w represents the sequence segment generated by sliding window, X w = (f w-1 ,f w ,...f w+Len-2 ) T , where f w represents the value of time series at position w, (f w-1 ,f w ,...f w+Len-2 ) T represents a column vector containing time series data from w-1 to w+Len-2; then it is constructed into trajectory matrix X = (X1, X2...X m ), which represents the result of the entire time series signal segmented by sliding window, where m represents the number of sequence segments generated by sliding window.

[0032] Then it is converted into the sum of d rank-1 basic matrices, where d is the number of non-zero eigenvalues in matrix S = X·X T , arranged in descending order:

[0033] X = E1 + E2 +…+E i …+E d

[0034] Basic matrix E i is represented as:

[0035]

[0036] where U i is the corresponding eigenvector, λ i is the eigenvalue, and the right eigenvector The contribution of basic matrix E i to the norm of X gradually decreases as the index i increases; the matrix with i < 3 is reconstructed by the reverse process of singular value decomposition, and then the denoised signal S i is obtained by matrix inverse operation, the SSA result is saved as

[0037] Step 3.2: splice the SSA result with its corresponding original EMG signal to obtain the time domain training sample.

[0038] Step 3.3: the time domain training sample first passes through the Xception module and the global adaptive average pooling layer, and the output of the global adaptive average pooling layer GAP is taken as the time domain feature.

[0039] S3.3.1, construct an Xception module, the Xception module adopts a nonlinear normalization method, and the nonlinear normalization formula is as follows:

[0040]

[0041] wherein x b represents an input scalar to be normalized, a nonlinear normalization parameter μ = 256 is used, and sign(x) represents a sign function; F(x b ) represents a nonlinear normalization function;

[0042] The Xception module uses a depth separable convolution block to capture the time and spatial information content of the sparse multi-channel surface electromechanical signal; the depth separable convolution block includes two parallel paths, the first path first passes through a one-dimensional convolution module, and then passes through three depth separable convolutions in parallel; the second path first passes through a maximum pooling, and then passes through a one-dimensional convolution; four depth separable convolution blocks are stacked together as the main part of the Xception module, then a residual connection is introduced to avoid overfitting, and finally a global adaptive average pooling layer is used to output the time domain features; the specific process is as follows:

[0043] One-dimensional convolution Conv1D: for a one-dimensional data sequence input x, the convolution operation Conv1D(x) is represented as:

[0044] Conv1D(x) = A*x + B

[0045] where * is a convolution operation, A is a weight of a convolution kernel, and B is a bias.

[0046] Depth separable convolution is an efficient convolution method, mainly composed of two parts: depth convolution and point-by-point convolution; this method can significantly reduce the parameter quantity and computational complexity of the model while maintaining or improving performance; one-dimensional depth separable convolution is composed of two parts: depth convolution and point-by-point convolution.

[0047] For input wherein L in is the length, C in is the number of input channels, and the operation definition of depth convolution is:

[0048] Y c = x' c * φ c

[0049] wherein x' c is the cth channel of the input x', φ c is a convolution kernel dedicated to the cth channel, * represents a convolution operation, and Y c is the result of the cth channel of the output.

[0050] Point-wise convolution receives the output of depth-wise convolution where L' is the length after convolution operation, and C in maps the information of the channel to C out channel, the operation is represented as:

[0051] ξ = Y * φ

[0052] where Y is the output of depth-wise convolution, φ is a convolution kernel, which is actually a weight matrix with size C in × C out , and ξ is the final output with dimension L' × C out .

[0053] Max pooling: for one-dimensional data sequence input x, the element at position w' of the output signal sequence generated by the max pooling operation is represented as:

[0054]

[0055] x (w'×s+i',c) is the value of the input signal at the w' × s + i'th position on the channel c, a is the size of the pooling window, s is the stride, i.e. the length of the window moving each time, and Maxpooling() represents the max pooling function. Max pooling can effectively reduce the dimension of the signal, reduce the amount of data, and extract significant features in the signal.

[0056] S3.3.2, the global adaptive average pooling operation is as follows:

[0057]

[0058] where x1 is the output result of the Xception module, is the length of the sequence, x l is the value of the sequence at position l, and GAP() represents the global adaptive average pooling function. Global adaptive average pooling can simplify the model structure, adapt to different length inputs, reduce training and inference time, and extract more robust features.

[0059] Step 4: Take the training data set as the input data of the frequency domain processing subnetwork, train the frequency domain processing subnetwork, and obtain the frequency domain feature:

[0060] The frequency domain processing subnetwork performs FFT processing on the input data to obtain the FFT result, connects the FFT result with its corresponding original electromyogram, obtains the frequency domain training sample, extracts features through the OmniScale module, and obtains the frequency domain feature through the global adaptive average pooling layer.

[0061] Step 4.1: The training data set is first processed by FFT to obtain the FFT result;

[0062] For the training data set, the formula of Fourier transform FFT is as follows:

[0063]

[0064] where e -jωt is a complex exponential function, j is an imaginary unit, ω represents an angular frequency, and t represents time; x(t) represents the time domain signal corresponding to the training data set.

[0065] The FFT result is saved as Here, represents the set of Fourier transform results of all channels, and each F c is the Fourier transform result of channel c.

[0066] Step 4.2: The FFT result is spliced with its corresponding original electromyographic signal to obtain the frequency domain training sample.

[0067] Step 4.3: The frequency domain training sample is first processed by the OmniScale module for feature extraction, then the residual connection is introduced to enhance the feature representation capability, combined with the multi-head self-attention mechanism for dynamic weight distribution, realizing the fusion of multi-scale features, and finally passing through the global adaptive average pooling layer GAP, taking the output of the global adaptive average pooling layer GAP as the frequency domain feature.

[0068] S4.3.1, build the OmniScale module, use the OmniScale module to capture important scale information, as follows:

[0069] OmniScale is a three-layer multi-core structure, each core performs the same padding convolution on the input, using to represent the kernel size set of the qth layer;

[0070]

[0071] q∈{1,2} represents the first two layers, at this time the prime number size convolution kernel is used, i.e. {1,2,3,5,...,p max}, where p max represents the largest prime number. q=3 represents the third layer, using convolution kernels with sizes 1 and 2, i.e. {1,2}. OmniScale can obtain more comprehensive feature extraction, increase the robustness of the model to input data changes, and improve the generalization ability of the model.

[0072] S4.3.2, multi-head self-attention mechanism: for an input with shape (L,d model ), where L is the sequence length and dmodel is the feature dimension of each time step, the input is first passed through three different linear transformations, generating query Query, key Key and value Value matrices; for each head δ, the transformation is as follows:

[0073]

[0074] where are the weight matrices for query, key and value, respectively, with shape (d model ,d k ), d k is the dimension of query and key for each head.

[0075] For each head, the dot product between query and key is calculated to obtain the attention score:

[0076]

[0077] where is the dot product of query and key, representing the correlation between each time step, is a scaling factor to prevent the numerical value of the dot product from being too large, and Softmax is used for normalization to obtain the attention weight; Attention δ represents the output of the δth head.

[0078] For the input vector z = [z1, z2,..., z z ..., z Δ ], z z represents the zth value in the input vector; Δ represents the length of the input vector.

[0079] The output of the Softmax function is a probability distribution vector p = [p1, p2,..., p p ..., p Δ ], where p p represents the probability of the class corresponding to the pth component of the input vector.

[0080] The definition of the Softmax function is:

[0081]

[0082] where exp(z z ) is the exponential function, representing the exponential of z z , is the sum of the exponents of all inputs, which serves as a normalization factor to ensure that the sum of all probabilities p p in the output is 1; ε represents the index.

[0083] After calculating the self-attention for each head independently, the outputs of all heads are concatenated:

[0084] MultiHead=Concat(Attention1,...,Attention δ )W O

[0085] where W O is the linear transformation weight matrix of the output, Attention δ represents the output of the δth head, MultiHead represents the output result of multi-head attention, and Concat() represents the concatenation operation.

[0086] The multi-head self-attention mechanism can help the model better understand the input data and extract more rich and valuable features, thereby significantly improving the performance of the model.

[0087] S4.3.3, after the multi-head self-attention mechanism, global adaptive average pooling is performed to obtain the frequency domain feature:

[0088]

[0089] where x2 represents the result obtained after the multi-head self-attention mechanism.

[0090] Step 5: Constructing a DBN network, taking the time domain feature and the frequency domain feature as the input of the DBN network, taking the DBN network as a recognition classifier, training the DBN network to obtain a trained signal processing recognition model. The DBN network is used to fuse the time domain and frequency domain features, which can fully utilize different domain information and improve the model generalization ability by virtue of the powerful feature learning ability.

[0091] The DBN network performs feature concatenation on the time domain feature and the frequency domain feature to obtain a time-frequency feature, which is then taken as the input of the RBM layer of the DBN network for unsupervised training. Finally, the network parameters are fine-tuned through supervised training of the DBN network to obtain a trained signal processing recognition model.

[0092] The DBN network is constructed as follows:

[0093] The DBN network is constructed by stacking several RBMs, and its training process starts from an unsupervised pre-training phase. The weights and biases of each RBM in the network are set through pre-training to establish a good initial weight for the subsequent fine-tuning phase. The training of each layer is independent and is performed sequentially. After pre-training, the network is further fine-tuned through a supervised learning phase. The entire network is fine-tuned through the standard backpropagation algorithm to minimize the error of the output layer. The backpropagation update rule of the error depends on the gradient of the loss function.

[0094] The energy function of the RBM is defined as:

[0095]

[0096] where v α is the state of a visible unit, h β is the state of a hidden unit, a α is the bias of a visible unit a, b β is the bias of a hidden unit b, and w αβ is the weight between a visible unit a and a hidden unit b.

[0097] The joint probability distribution P(v, h) of the visible and hidden layers in an RBM:

[0098]

[0099] where v is a visible layer unit representing the input data, h is a hidden layer unit representing the latent features captured in the data, and E(v, h) is the energy function of the RBM.

[0100] where Z is the normalization factor:

[0101]

[0102] The marginalized probability distribution P(v) of the visible layer is:

[0103]

[0104] The activation probability P(h β =1 | v) of a hidden unit is:

[0105]

[0106] The activation probability P(v α =1 | h) of a visible unit given the condition of a hidden unit is:

[0107]

[0108] where σ(x) is the sigmoid function:

[0109]

[0110] The output process of a deep belief network (DBN) involves multiple steps. During DBN training, starting from the input data, the output of the topmost hidden layer is passed to the Softmax layer, which converts the output into a probability distribution P(y θ | z) for each class:

[0111]

[0112] where y θw and b θ and b θ are the weight and bias corresponding to the θ-th class respectively; η l and b l are the weight and bias corresponding to the l-th class respectively;

[0113] The model gives the classification recognition result according to the class corresponding to the maximum probability That is:

[0114]

[0115] is an operation of finding a parameter that makes the function maximum, in this formula, it is to find the class that makes the conditional probability P(y θ |z) maximum among all different classes k, finally represents the predicted class obtained after the input data is classified by the DBN.

[0116] Step 6: input the test data set into the trained signal processing recognition model, that is, input into the time domain processing subnetwork and the frequency domain processing subnetwork and pass through the DBN network to obtain the recognition result, and perform surface electromyogram signal processing recognition for the robot.

[0117] Example 1: sEMG experimental data verification

[0118] In this experiment, the MA400 EMG system series product of Motion Lab Systems Company was used to collect sEMG data. Relevant data were collected from the right hands of four healthy men. The ages of these participants were between 22 and 30 years old, and they all had no muscle disease. Before data collection, each participant was familiar with all the upper limb movements. During data collection, all participants completed 12 basic movements according to the instructions of the observer, guided by the metronome. Each movement was repeated five times, each lasting 5 seconds. After each basic movement, the participant rested for 2 minutes to prevent muscle fatigue. This task mode continued until all experimental movements were completed. The joint movements mainly included shoulder flexion (S1), shoulder extension (S2), shoulder abduction (S3), internal rotation (S4), external rotation (S5), elbow flexion (E1), forearm elevation (E2), forearm pronation (E3), wrist flexion (W1), wrist extension (W2), ulnar deviation (W3), and radial deviation (W4). After completing data collection, the data were divided into 4 (number of participants) x 12 (total number of movements) groups.

[0119] Task A is set to use single surface electromyogram signal for model training and testing, and task B is set to use multiple surface electromyogram signals for model training and testing.

[0120] Figure 2 The confusion matrix of the signal processing recognition model under task A and task B is shown. Figure 3 The bar chart of the diagnostic accuracy of different methods under different tasks is shown. The average accuracy of the signal processing recognition model under task A is 97.2%, and the average accuracy under task B is 85.9%.

[0121] In summary, in view of the problem that the current sEMG-based classification recognition network has limited generalization ability, the effect significantly decreases when processing multi-subject analysis, and the calculation complexity is high, the signal processing recognition model is designed. The network model can achieve excellent recognition effect without complex parameter adjustment in upper limb movement classification, and has good generalization effect in multi-subject analysis. In order to fully consider the time domain and frequency domain signals of the original electromyographic signal, the time domain processing sub-network and the frequency domain processing sub-network are introduced to extract the features of the time domain and frequency domain signals respectively, and then the extracted features of the two networks are further fused by using the deep belief network DBN for processing, and finally the classification result of the sEMG is obtained. The experiment shows that the model has high accuracy.

[0122] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A surface myoelectric robust intelligent recognition method for a robot, characterized in that, The method comprises the following steps: Step 1: acquiring a plurality of groups of original surface electromyography signals of a human body by using a surface electromyography sensor, slicing the original surface electromyography signals by using a sliding window, and dividing the original surface electromyography signals into a training data set and a test data set; Step 2: constructing a signal processing and recognition model, including a time domain processing subnetwork, a frequency domain processing subnetwork and a DBN network; Step 3: taking the training data set as input data of the time domain processing subnetwork, training the time domain processing subnetwork, and obtaining time domain features: The time domain processing subnetwork performs SSA processing on the input data to obtain SSA results, connects the SSA results with corresponding original surface electromyography signals, obtains time domain training samples, extracts features through an Xception module, and then obtains time domain features through a global adaptive average pooling layer; The Xception module adopts a nonlinear normalization method, uses a depth separable convolution block to capture time and space information content of sparse multi-channel surface electromyography signals, the depth separable convolution block includes two parallel paths, the first path first passes through a one-dimensional convolution module and then passes through three depth separable convolutions in parallel; the second path first passes through a maximum pooling and then passes through a one-dimensional convolution; four depth separable convolution blocks are stacked together as a main part of the Xception module, then a residual connection is introduced to avoid overfitting, and finally a global adaptive average pooling layer is used for outputting the time domain features; Step 4: taking the training data set as input data of the frequency domain processing subnetwork, training the frequency domain processing subnetwork, and obtaining frequency domain features: The frequency domain processing subnetwork performs FFT processing on the input data to obtain FFT results, connects the FFT results with corresponding original surface electromyography signals, obtains frequency domain training samples, extracts features through an OmniScale module, and then obtains frequency domain features through a global adaptive average pooling layer; The OmniScale module is constructed, and the OmniScale module is used to capture important scale information, and the specific process is as follows: OmniScale is a three-layer multi-kernel structure, each kernel performs the same padded convolution on the input, using represents a set of kernel sizes for the qth layer; q e {1,2} means the first two layers, in which prime number size of convolution kernel is used, i.e. p max represents the largest prime number; q = 3 means the third layer, using convolution kernels of size 1 and 2, i.e. Step 5: constructing a DBN network, taking the time domain features and the frequency domain features as input of the DBN network, taking the DBN network as a recognition classifier, training the DBN network, and obtaining a trained signal processing and recognition model: The DBN network performs feature splicing on the time domain features and the frequency domain features to obtain time-frequency features, takes the time-frequency features as input of an RBM layer in the DBN network for unsupervised training, and finally fine-tunes network parameters through supervised training of the DBN network to obtain the trained signal processing and recognition model; Step 6: inputting the test data set into the trained signal processing and recognition model to obtain a recognition result, and performing surface electromyography signal processing and recognition for a robot.

2. The sEMG robust intelligent recognition method for robots according to claim 1, characterized in that: In step 3, the training data set is taken as input data of the time domain processing subnetwork, the time domain processing subnetwork is trained, and time domain features are obtained, and the specific process is as follows: Step 3.1: performing SSA processing on the training data set to obtain SSA results: With representing the original surface electromyography signal, where c represents the number of channels, the signal of channel c is represented as A c = (a0, a1, a2...a N ), where N is the length of the time series, a N represents the Nth point of the time series; T represents the transpose; SSA is based on multivariate statistical principles, which reveals the underlying structure and patterns in the original EMG signal by extracting the principal components of the time series; each channel is processed separately, define Len as the length of the sliding window, use X w to represent the sequence fragments generated by the sliding window, X w = (f w-1 ,f w ,...f w+Len-2 ) T , where f w represents the value of the time series at position w, (f w-1 ,f w ,...f w+Len-2 ) T represents a column vector containing time series data from w-1 to w+Len-2; then it is constructed as a trajectory matrix X = (X1, X2...X m ), which represents the results of the entire time series signal segmented by the sliding window, where m represents the number of sequence fragments generated by the sliding window; It is then converted into the sum of d rank-1 elementary matrices, where d is the number of non-zero eigenvalues in the matrix S = X · X T in descending order: X = E1 + E2 +... + E i ... + E d Fundamental matrix E i is expressed as: where U i is the corresponding eigenvector, λ i is the eigenvalue, the right eigenvector With the increase of the index i, the contribution of the basic matrix E i to the norm of X gradually decreases; the matrix with i < 3 is reconstructed by the reverse process of singular value decomposition, and then the denoising signal S i is obtained by matrix inverse operation, and the SSA result is saved as Step 3.2: splicing the SSA results with corresponding original surface electromyography signals to obtain time domain training samples; Step 3.3: An Xception module is constructed, and the time domain training sample sequentially passes through the Xception module and the global adaptive average pooling layer, and the output of the global adaptive average pooling layer is taken as the time domain feature; The global adaptive average pooling operation is as follows: where x1 is the output result of the Xception module, is the length of the sequence, is the value of the sequence at position GAP() represents the global adaptive average pooling function.

3. The surface myoelectric robust intelligent recognition method for robots according to claim 2, characterized in that: In step 3.3, the Xception module is constructed as follows: The nonlinear normalization formula is as follows: where x b represents the input scalar to be normalized, sign(x) represents the sign function, and F(x b ) represents the nonlinear normalization function. One-dimensional convolution Conv1D: for one-dimensional data sequence input x, the convolution operation Conv1D(x) is represented as: Conv1D(x) = A*x + B Where * is the convolution operation, A is the weight of the convolution kernel, and B is the bias; The depth separable convolution is an efficient convolution method, mainly composed of two parts: depth convolution and point-by-point convolution; one-dimensional depth separable convolution is composed of two parts: depth convolution and point-by-point convolution; For input where L in is the length, C in is the number of input channels, and the operation of depthwise convolution is defined as: Y c = x' c * φ c where x' is the input c is the c-th channel of the input x', φ c is a convolution kernel specific to the c-th channel, * denotes a convolution operation, Y c is the result of the c-th channel of the output; Point-wise convolutions then receive the output of the depthwise convolutions where L' is the length after the convolution operation and maps the information of the input channels to the output channels, the operation is represented as: ξ = Y*φ where Y is the output of the deep convolution, φ is a convolution kernel, which is actually a weight matrix of size C in × C out ; ξ is the final output, with dimensions L' x C out ; C out denotes the number of output channels; Max pooling: for one-dimensional data sequence input x, the output signal sequence generated by the max pooling operation is represented as: x (w'×s+i',c) is the value of the input signal at the w'xs+i' position on channel c, a is the size of the pooling window, s is the stride, i.e. the length of the window movement at each time, and Maxpooling() denotes the max pooling function.

4. The surface myoelectric robust intelligent recognition method for robots according to claim 3, characterized in that: In step 4, the training data set is taken as the input of the frequency domain processing subnetwork, and the frequency domain processing subnetwork is trained to obtain the frequency domain feature, as follows: Step 4.1: The training data set is subjected to FFT processing to obtain the FFT result; For the training data set, the Fourier transform FFT formula is as follows: where e -jωt is the complex exponential function, j is the imaginary unit, ω denotes the angular frequency, t denotes time; x(t) denotes the time-domain signal corresponding to the training data set; Save the FFT result as The above A set of Fourier transform results representing all channels, each F c is the Fourier transform result of channel c; Step 4.2: The FFT result is spliced with the original electromyogram corresponding thereto to obtain a frequency domain training sample; Step 4.3: An OmniScale module is constructed, and the frequency domain training sample is subjected to preliminary feature extraction through the OmniScale module, combined with a multi-head self-attention mechanism for dynamic weight allocation, to realize fusion of multi-scale features, and finally subjected to a global adaptive average pooling layer, and the output of the global adaptive average pooling layer is taken as the frequency domain feature; Multi-head self-attention mechanism: for an input of shape (L, d model ), where L is the sequence length, d model is the feature dimension at each time step, the input is passed through three different linear transformations to generate query Query, key Key, and value Value matrices; for each head δ, the transformation is as follows: wherein are weight matrices for queries, keys, and values, respectively, of shape (d model ,d k ), d k is the dimension of queries and keys for each head. For each head, the dot product between the query and the key is calculated to obtain the attention score: where is the dot product of query and key, representing the relevance between each time step, is a scaling factor to prevent the numerical value of the dot product from being too large, and Softmax is used to normalize the attention weight; Attention δ represents the output of the δth head. For an input vector z = [z1, z2,..., z z ...,z Δ ], z z denotes the z-th value in the vector; Δ denotes the length of the input vector; The output of the softmax function is a probability distribution vector p = [p1, p2,..., p p ] where p Δ represents the probability of the class corresponding to the pth component of the input vector. p ​ The definition of the Softmax function is as follows: where exp(z z ) is the exponential function, representing the exponent of z z , is the sum of the exponents of all inputs, and as a normalization factor, ensures that the sum of all probabilities p p output is 1; ε denotes the index; After each head independently calculates the self-attention, the outputs of all heads are spliced together: MultiHead = Concat(Attention1,..., Attention δ ) W O where W O is the linear transformation weight matrix of the output, Attention δ denotes the output of the δth head, MultiHead denotes the output result of multi-head attention, and Concat() denotes a concatenation operation. After the multi-head self-attention mechanism and the global adaptive average pooling, the frequency domain feature is obtained: Where x2 represents the result obtained after the multi-head self-attention mechanism.

5. The sEMG robust intelligent recognition method for robots according to claim 4, characterized in that: In step 5, a DBN network is constructed as follows: A DBN network is constructed by stacking several RBMs, and its training process starts from an unsupervised pre-training phase; the weights and biases of each RBM in the network are set layer by layer through pre-training to establish a good initial weight for the subsequent fine-tuning stage; The training of each layer is independent and is performed sequentially; after pre-training, the network is further fine-tuned through a supervised learning phase; the entire network is fine-tuned through a standard backpropagation algorithm to minimize the error of the output layer; the error backpropagation update rule depends on the gradient of the loss function; The energy function E(v, h) of the RBM is defined as: where v α is the state of the visible unit, h β is the state of the hidden unit, a α is the bias of the visible unit α, b β is the bias of the hidden unit β, and ω αβ is the weight between the visible unit α and the hidden unit β. The joint probability distribution P(v, h) of the visible layer and the hidden layer in the RBM is: Where v is the visible layer unit, representing the input data, h is the hidden layer unit, representing the hidden features captured in the data; E(v, h) represents the energy function of the RBM; Where Z is the normalization factor: The marginalization probability distribution P(v) of the visible layer is: The activation probability P(h β =1 | v) of a hidden unit is: Given the condition of the hidden unit, the activation probability P(v α =1 | h) is: where σ(x) is a sigmoid function: During DBN training, starting from the input data, the output of the topmost hidden layer is passed to the Softmax layer, which converts the output into a probability distribution for each class wherein, represents the weight and bias of the th class, and are the weight and bias of the th class, respectively; η l and b l are the weight and bias of the lth class, respectively. The model gives a classification recognition result according to a category corresponding to a maximum probability That is, is an operation that finds a parameter that maximizes a function, and in this formula, it is to find a class k that maximizes the conditional probability P (C = k | X) among all different classes k, and finally represents a predicted class obtained by classifying input data through the DBN.

Images