Robust intelligent recognition method for surface myoelectricity of robot

Abstract

The invention discloses a surface myoelectricity robust intelligent identification method for a robot. The method comprises the following steps: firstly, collecting a plurality of groups of human body original myoelectricity signals by using a surface myoelectricity sensor; then, performing sliding window slicing processing on the acquired sensor data, and dividing a training data set and a test data set; performing network training by taking the training data set as input of a time domain processing sub-network and a frequency domain processing sub-network, respectively extracting time domain features and frequency domain features through network training, and training an identification classifier by taking the time domain features and the frequency domain features as input of the identification classifier; and finally, inputting the test sample set into the trained signal processing and recognition model to carry out surface electromyogram signal processing and recognition. According to the method, time domain and frequency domain information is effectively combined, and compared with a traditional complete combination network, the method has the advantage that the recognition accuracy and stability are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recognition technology, and in particular to a surface electromyography robust intelligent recognition method for a robot. Background Art

[0002] With the rapid development of computer technology, artificial intelligence, signal processing, human-computer interaction, brain-like chips and other technologies, surface electromyography (sEMG), as a signal generated by the nervous system during muscle contraction, controls and reflects the physiological and anatomical characteristics of muscles, and its signal analysis capabilities have been significantly improved. Researchers have adopted various signal preprocessing, feature extraction and classification techniques to achieve more efficient processing of electromyographic signals. As a physiological signal closely related to human movement, sEMG signals have been widely used in human-computer interaction systems. The 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 recognition. Deep learning algorithms are widely used in sEMG-based signal recognition due to their powerful classification capabilities. Deep learning-based methods usually assume that the distribution of training and test data is consistent. But in real-world applications, such as sEMG recognition systems, training and test data are often interfered with by various factors. Users need to perform specific gestures when obtaining supervised samples, and the classifiers trained by these samples have significantly reduced performance when applied to other users or the same user wearing the device again. From an algorithmic perspective, the generalization ability of current sEMG-based classification and recognition systems is limited. In addition, as more and more complex features are extracted, the computational burden also increases. Summary of the invention

[0004] The present invention is a surface electromyography robust intelligent recognition method for robots. Based on a signal processing recognition model, it achieves more accurate surface electromyography signal recognition and can effectively solve the problem that the generalization effect of the current sEMG classification and recognition algorithm is significantly reduced when processing multi-subject analysis and has high computational complexity.

[0005] The technical solution to realize the present invention is: a surface electromyography robust intelligent recognition method for a robot, comprising the following steps:

[0006] Step 1: Use surface electromyography sensors to collect several groups of human raw electromyographic signals, slice the raw electromyographic signals through a sliding window, and divide them into training data sets and test data sets.

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

[0008] Step 3: Use the training data set as the input data of the time domain processing sub-network, train the time domain processing sub-network, and obtain the time domain features:

[0009] The time domain processing subnetwork performs SSA processing on the input data to obtain SSA results, connects the SSA results with their corresponding original electromyographic 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.

[0010] Step 4: Use the training data set as the input data of the frequency domain processing sub-network, train the frequency domain processing sub-network, and obtain the frequency domain features:

[0011] 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 EMG signal to obtain the frequency domain training sample, extracts features through the OmniScale module, and then obtains the frequency domain features through the global adaptive average pooling layer.

[0012] Step 5: Build a DBN network, use the time domain features and frequency domain features as the input of the DBN network, use the DBN network as a recognition classifier, train the DBN network, and obtain a trained signal processing recognition model:

[0013] The DBN network concatenates the time domain features and frequency domain features to obtain the time-frequency features, which are then used as the input of the RBM layer in 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 and recognition model.

[0014] Step 6: Input the test data set into the trained signal processing and recognition model to obtain the recognition result, and perform surface electromyography signal processing and recognition for the robot.

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

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

[0017] (2) The designed robust intelligent recognition method for surface electromyography for robots effectively combines time domain and frequency domain information, enabling the model to fully understand and process surface electromyography signals. Compared with the traditional fully integrated network, the signal processing recognition model network algorithm designed in this paper has better stability and stronger generalization, and has achieved higher fault diagnosis accuracy in the task of surface electromyography signal recognition of human upper limbs. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0020] Figure 3 It 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 solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

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

[0023] Combination Figure 1 The present invention discloses a method for robust intelligent recognition of surface electromyography for robots. First, the MA400-18 acquisition system and the MA-420 electromyography signal amplifier are used to collect experimental data. The experiment is set as 12 groups of prescribed upper limb movements, and the experimenter collects surface electromyography signals according to the prescribed movements within a prescribed time. Then, the collected sensor data are processed by sliding window slicing to divide the training data set and the test data set. The training data set is used as the input of the time domain processing subnetwork, and the time domain processing subnetwork is trained to obtain time domain features. The training data set is used as the input of the frequency domain processing subnetwork, and the frequency domain processing subnetwork is trained to obtain frequency domain features. The two extracted features are then 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 supervised training of the deep belief network to complete the training of the entire signal processing and recognition model. Figure 2 , Figure 3It can be seen that this process effectively combines time domain and frequency domain information, enabling the model to fully understand and process surface electromyography signals, significantly improving recognition accuracy and stability compared to the traditional fully combined network. The specific steps are as follows:

[0024] Step 1: Use surface electromyography sensors to collect several groups of human original electromyographic signals; use MA400-18 acquisition system and MA-420 electromyographic signal amplifier to collect experimental data. The experiment is set as 12 groups of prescribed upper limb movements. The experimenters collect surface electromyographic signals according to the prescribed movements within the prescribed time. The collected multi-sensor related signals are sliced ​​and processed to divide the training data set and the test data set.

[0025] The experiment collected relevant data from the right hands of four healthy men. During the data collection process, all participants completed 12 basic movements according to the instructions of the observer, guided by a metronome. Each movement was repeated five times, each lasting 5 seconds. After each basic movement, the participants rested for 2 minutes to prevent muscle fatigue. This task mode continued until all experimental movements were completed. Joint movements 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 the time domain processing subnetwork, the frequency domain processing subnetwork and the DBN network.

[0027] Step 3: Use the training data set as the input data of the time domain processing sub-network, train the time domain processing sub-network, and obtain the time domain features.

[0028] The time domain processing subnetwork performs SSA processing on the input data to obtain the SSA result, connects the SSA result with its corresponding original electromyographic signal to obtain the time domain training sample, extracts the feature through the Xception module, and then obtains the time domain feature through the global adaptive average pooling layer, as follows:

[0029] Step 3.1: The training data set is first processed by SSA to obtain the SSA result.

[0030] use represents the original surface electromyographic signal, Where c represents the number of channels, and the signal of channel c is represented by 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 transpose.

[0031] SSA is based on the principle of multivariate statistics. It reveals the potential structure and pattern in the original electromyographic signal by extracting the main components of the time series. Each channel is processed separately, and Len is defined as the sliding window length. w represents the sequence fragment 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 ), represents the result of the entire time series signal being segmented by the sliding window, where m represents the number of sequence segments generated by the sliding window.

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

[0033] X=E1+E2+…+E i …+E d

[0034] Basic Matrix E i It is expressed as:

[0035]

[0036] Among them, U i is the corresponding eigenvector, λ i is the eigenvalue, the right eigenvector As the index i increases, the basic matrix E i The contribution to the norm of X gradually decreases; the matrix with i < 3 is selected and reconstructed through the reverse process of singular value decomposition, and then the denoised signal S is obtained through matrix inversion operation i , the SSA results are saved as

[0037] Step 3.2: Concatenate the SSA results with their corresponding original EMG signals to obtain time domain training samples.

[0038] Step 3.3: The time domain training samples pass through the Xception module and the global adaptive average pooling layer successively, and the output of the global adaptive average pooling layer GAP is used as the time domain feature.

[0039] S3.3.1. Construct the Xception module. The Xception module uses a nonlinear normalization method. The nonlinear normalization formula is as follows:

[0040]

[0041] Among them, x b represents the input scalar to be normalized, using a nonlinear normalization parameter μ = 256, sign(x) represents the sign function; F(x b ) represents a nonlinear normalization function;

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

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

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

[0045] Among them, * is the convolution operation, A is the weight of the convolution kernel, and B is the bias.

[0046] Depthwise separable convolution is an efficient convolution method, which mainly consists of two parts: depthwise convolution and pointwise convolution. This method can significantly reduce the number of parameters and computational complexity of the model while maintaining or improving performance. One-dimensional depthwise separable convolution consists of two parts: depthwise convolution and pointwise convolution.

[0047] For input Where L in is the length, C in is the number of input channels, and the depthwise convolution operation is defined as:

[0048] Y c =x' c *φ c

[0049] where x' c is the cth channel of input x', φ c is the convolution kernel dedicated to the cth channel, * indicates the convolution operation, Y c It 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 the convolution operation, and C in The information of the channel is mapped to C out Channel,operations are expressed as:

[0051] ξ=Y*φ

[0052] Where Y is the output of the depthwise convolution, φ is a convolution kernel, which is actually a convolution kernel of size C. in ×C out The weight matrix of ξ is the final output, with a dimension of L′×C out .

[0053] Max pooling: For a one-dimensional data sequence input x, the element of the output signal sequence generated by the max pooling operation at position w' 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 channel c, a is the size of the pooling window, s is the stride, that is, the length of each window movement, and Maxpooling() represents the maximum pooling function. Maximum pooling can effectively reduce the dimension of the signal, reduce the amount of data, and extract significant features from the signal.

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

[0057]

[0058] Where x1 is the output 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 inputs of different lengths, reduce training and inference time, and make feature extraction more robust.

[0059] Step 4: Use the training data set as the input data of the frequency domain processing sub-network, train the frequency domain processing sub-network, and obtain the frequency domain features:

[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 EMG signal to obtain the frequency domain training sample, extracts features through the OmniScale module, and then obtains the frequency domain features 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 for Fourier transform FFT is as follows:

[0063]

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

[0065] Save the FFT results as Here Represents the set of Fourier transform results of all channels, each F c is the Fourier transform result of channel c.

[0066] Step 4.2: Concatenate the FFT result with the corresponding original EMG signal to obtain the frequency domain training sample.

[0067] Step 4.3: The frequency domain training samples are first subjected to feature extraction by the OmniScale module, and then the residual connection is introduced to enhance the feature representation capability. The multi-head self-attention mechanism is combined for dynamic weight allocation to achieve multi-scale feature fusion. Finally, the global adaptive average pooling layer GAP is passed through, and the output of the global adaptive average pooling layer GAP is used as the frequency domain feature.

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

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

[0070]

[0071] q∈{1,2} represents the first two layers, and a prime-sized 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 of size 1 and 2, i.e. {1,2}. OmniScale can obtain more comprehensive feature extraction, increase the model's robustness to changes in input data, and improve the model's generalization ability.

[0072] S4.3.2. Multi-head self-attention mechanism: For the shape (L,d model ), where L is the sequence length, dmodel is the feature dimension of each time step. First, the input is transformed through three different sets of linear transformations to generate query, key, and value matrices. For each head δ, the transformation is as follows:

[0073]

[0074] in are weight matrices for query, key, and value, respectively, with shapes (d model ,d k ), d k are the dimensions of the query and key for each header.

[0075] For each head, we compute the dot product between the query and the key to get the attention score:

[0076]

[0077] in is the dot product of the query and the key, indicating the correlation between each time step, is a scaling factor used to prevent 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.

[0078] For the input vector z=[z1,z2,...,z z ...,z Δ ],z z represents the zth value in the 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 category corresponding to the pth component of the input vector.

[0080] The definition of the Softmax function is:

[0081]

[0082] Among them, exp(z z ) is an exponential function, indicating that z z The index of is the exponential sum of all inputs, which acts as a normalization factor to ensure that all probabilities p of the output p The sum of is 1; ε represents the index.

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

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

[0085] Among them, W O is the output linear transformation weight matrix, 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 richer and more valuable features, thereby significantly improving the performance of the model.

[0087] S4.3.3, after the multi-head self-attention mechanism and then the global adaptive average pooling, the frequency domain features are obtained:

[0088]

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

[0090] Step 5: Build a DBN network, use the time domain features and frequency domain features as the input of the DBN network, use the DBN network as a recognition classifier, train the DBN network, and obtain a trained signal processing recognition model. Using the DBN network to fuse the time domain and frequency domain features can make full use of information in different domains and improve the generalization ability of the model with its powerful feature learning ability.

[0091] The DBN network concatenates the time domain features and frequency domain features to obtain the time-frequency features, which are then used as the input of the RBM layer in 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 and recognition model.

[0092] Construct a DBN network as follows:

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

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

[0095]

[0096] 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 hidden unit β, ω αβ is the weight between the visible unit α and the hidden unit β.

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

[0098]

[0099] Where v is the visible layer unit, representing the input data, h is the hidden layer unit, representing the capture of implicit features in the data; E(v,h) represents the energy function of 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 of the hidden unit P(h β =1|v) is:

[0105]

[0106] Given the hidden unit, the activation probability of the visible unit P(v α =1|h) is:

[0107]

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

[0109]

[0110] The output process of the deep belief network DBN involves multiple steps. When DBN is trained, it starts with the input data, which is passed layer by layer through multiple hidden layers. The output of the top hidden layer is passed to the Softmax layer, which converts the output into the probability distribution P(y θ |z):

[0111]

[0112] Among them, y θrepresents the weight and bias of the θth category, η θ and b θ are the weight and bias corresponding to the θth category respectively; η l and b l are the weight and bias corresponding to the lth category respectively;

[0113] The model gives the classification recognition result according to the category corresponding to the maximum probability Right now:

[0114]

[0115] It is an operation to find the parameters that maximize the function. In this formula, we need to find the conditional probability P(y θ |z) the largest category, and finally It represents the predicted category obtained after classifying the input data through DBN.

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

[0117] Example 1: sEMG experimental data verification

[0118] The MA400 EMG system series products of Motion Lab Systems were used to collect sEMG data in this experiment. The relevant data were collected from the right hands of four healthy males. The ages of these participants ranged from 22 to 30 years old, and none of them had any muscle diseases. Before data collection, each participant was familiarized with all upper limb movements. During data collection, all participants completed 12 basic movements according to the instructions of the observer, guided by a metronome. Each movement was repeated five times, each lasting 5 seconds. After each basic movement, the participants 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) × 12 (total number of movements) groups.

[0119] Task A is set to use the surface electromyography signals of a single person to train and test the model, and Task B is set to use the surface electromyography signals of multiple people to train and test the model.

[0120] Figure 2 The confusion matrix of the signal processing recognition model under Task A and Task B is shown. Figure 3 The diagnostic accuracy bar graphs of different methods under different tasks are 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, a signal processing recognition model was designed to solve the problem that the current classification and recognition network based on sEMG has limited generalization ability, significantly reduced effect when dealing with multi-subject analysis, and high computational complexity. This network model can achieve excellent recognition effect in upper limb movement classification without complex parameter adjustment, 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 subnetwork and the frequency domain processing subnetwork are introduced to extract features of the time domain and frequency domain signals respectively. Finally, the deep belief network DBN is used to further fuse the extracted features of the two networks, and finally the classification result of sEMG is obtained. Experiments show that the model has a high accuracy rate.

[0122] The above description is only a specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a technician familiar with the technical field within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A robust intelligent recognition method for surface electromyography of a robot, characterized in that: The following steps are involved: Step 1: Use surface electromyography sensors to collect several groups of human original electromyographic signals, slice the original electromyographic signals through a sliding window, and divide them into training data sets and test data sets; Step 2: Construct a signal processing recognition model, including a time domain processing subnetwork, a frequency domain processing subnetwork, and a DBN network; Step 3: Use the training data set as the input data of the time domain processing sub-network, train the time domain processing sub-network, and obtain the time domain features: The time domain processing subnetwork performs SSA processing on the input data to obtain the SSA result, connects the SSA result with its corresponding original electromyographic signal to obtain the time domain training sample, extracts features through the Xception module, and then obtains the time domain features through the global adaptive average pooling layer; Step 4: Use the training data set as the input data of the frequency domain processing sub-network, train the frequency domain processing sub-network, and obtain the frequency domain features: 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 electromyographic signal to obtain the frequency domain training sample, extracts features through the OmniScale module, and then obtains the frequency domain features through the global adaptive average pooling layer; Step 5: Build a DBN network, use the time domain features and frequency domain features as the input of the DBN network, use the DBN network as a recognition classifier, train the DBN network, and obtain a trained signal processing recognition model: The DBN network concatenates the time domain features and the frequency domain features to obtain the time-frequency features, which are then used as the input of the RBM layer in 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. Step 6: Input the test data set into the trained signal processing and recognition model to obtain the recognition result, and perform surface electromyography signal processing and recognition for the robot.

2. A surface electromyography robust intelligent recognition method for robots according to claim 1, characterized in that: In step 3, the training data set is used as the input data of the time domain processing sub-network, and the time domain processing sub-network is trained to obtain the time domain features, as follows: Step 3.1: Perform SSA processing on the training data set to obtain the SSA result: use represents the original surface electromyographic signal, Where c represents the number of channels, and the signal of channel c is represented by 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; SSA is based on the principle of multivariate statistics. It reveals the potential structure and pattern in the original electromyographic signal by extracting the main components of the time series. Each channel is processed separately, and Len is defined as the sliding window length. w represents the sequence fragment 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 ), represents the result after the entire time series signal is segmented by the sliding window, where m represents the number of sequence segments generated by the sliding window; Then it is converted into the sum of d rank-1 basic matrices, where d is the matrix S = X·X T The number of nonzero eigenvalues ​​in , in descending order: X=E1+E2+…+E i …+E d Basic Matrix E i It is expressed as: Among them, U i is the corresponding eigenvector, λ i is the eigenvalue, the right eigenvector As the index i increases, the basic matrix E i The contribution to the norm of X gradually decreases; the matrix with i < 3 is selected and reconstructed through the reverse process of singular value decomposition, and then the denoised signal S is obtained through matrix inversion operation i , the SSA results are saved as Step 3.2: Concatenate the SSA result with the corresponding original electromyographic signal to obtain the time domain training sample; Step 3.3: Construct the Xception module. The time domain training samples pass through the Xception module and the global adaptive average pooling layer successively. The output of the global adaptive average pooling layer GAP is used as the time domain feature. Among them, the global adaptive average pooling operation is as follows: Among them, x1 is the output result of the Xception module, is the length of the sequence, is the sequence at position The value of , GAP() represents the global adaptive average pooling function.

3. A surface electromyography robust intelligent recognition method for robots according to claim 2, characterized in that: In step 3.3, build the Xception module as follows: The Xception module uses a nonlinear normalization method and uses a depth-separable convolution block to capture the temporal and spatial information content of sparse multi-channel surface electromechanical 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 maximum pooling and then one-dimensional convolution; four depth-separable convolution blocks are stacked together as the main part of the Xception module, and then residual connections are introduced to avoid overfitting. Finally, the time domain features are output through a global adaptive average pooling layer; the details are as follows: The nonlinear normalization formula is as follows: Among them, x b represents the input scalar to be normalized, using a nonlinear normalization parameter μ = 256, sign(x) represents the sign function; F(x b ) represents a nonlinear normalization function; One-dimensional convolution Conv1D: For a one-dimensional data sequence input x, the convolution operation Conv1D(x) is expressed as: Conv1D(x)=A*x+B Among them, * is the convolution operation, A is the weight of the convolution kernel, and B is the bias; Depthwise separable convolution is an efficient convolution method, which mainly consists of two parts: depthwise convolution and pointwise convolution; one-dimensional depthwise separable convolution consists of two parts: depthwise convolution and pointwise convolution; For input Where L in is the length, C in is the number of input channels, and the depthwise convolution operation is defined as: Y c =x' c *f c where x' c is the cth channel of input x', φ c is the convolution kernel dedicated to the cth channel, * indicates the convolution operation, Y c is the result of the cth channel of the output; Point-wise convolution receives the output of depth-wise convolution Where L′ is the length after the convolution operation, and maps the information of the input channel to the output channel. The operation is expressed as: ξ=Y*φ Among them, Y is the output of the depthwise convolution, φ is a convolution kernel, which is actually a convolution kernel of size C. in ×C out The weight matrix of ξ is the final output, with a dimension of L′×C out ; C out Indicates the number of output channels; Max pooling: For a one-dimensional data sequence input x, the element of the output signal sequence generated by the max pooling operation at position w' is represented as: x (w'×s+i',c) is the value of the input signal at the w'×s+i'th position on channel c, a is the size of the pooling window, s is the stride, that is, the length of each window movement, and Maxpooling() represents the maximum pooling function.

4. The method for robust intelligent recognition of surface electromyography for robots according to claim 3, characterized in that: In step 4, the training data set is used as the input of the frequency domain processing sub-network, and the frequency domain processing sub-network is trained to obtain frequency domain features, as follows: Step 4.1: Perform FFT processing on the training data set to obtain the FFT result; For the training data set, the formula for Fourier transform FFT is as follows: Among them, e -jωt is a complex exponential function, j is an imaginary unit, ω represents the angular frequency, and t represents time; x(t) represents the time domain signal corresponding to the training data set; Save the FFT results as Above Represents the set of Fourier transform results of all channels, each F c is the Fourier transform result of channel c; Step 4.2: Concatenate the FFT result with the corresponding original electromyographic signal to obtain the frequency domain training sample; Step 4.3: Construct the OmniScale module. The frequency domain training samples are subjected to preliminary feature extraction by the OmniScale module. The multi-head self-attention mechanism is combined to perform dynamic weight allocation to achieve multi-scale feature fusion. Finally, the global adaptive average pooling layer GAP is passed through the global adaptive average pooling layer GAP, and the output of the global adaptive average pooling layer GAP is used as the frequency domain feature. Multi-head self-attention mechanism: For the shape (L,d model ), where L is the sequence length, d model is the feature dimension of each time step. The input is transformed through three different sets of linear transformations to generate query, key and value matrices. For each head δ, the transformation is as follows: in are weight matrices for query, key, and value, respectively, with shapes (d model ,d k ), d k are the dimensions of the query and key for each header; For each head, we compute the dot product between the query and the key to get the attention score: in is the dot product of the query and the key, indicating the correlation between each time step, is a scaling factor used to prevent 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 the input vector z=[z1,z2,...,z z ...,z Δ ],z z represents the zth value in the vector; Δ represents the length of the input vector; 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 category corresponding to the pth component of the input vector; The definition of the Softmax function is: Among them, exp(z z ) is an exponential function, indicating that z z The index of is the exponential sum of all inputs, which acts as a normalization factor to ensure that all probabilities p of the output p The sum of is 1; ε represents the index; After each head calculates self-attention independently, the outputs of all heads are concatenated: MultiHead=Concat(Attention1,...,Attention δ )W O Where W O is the output linear transformation weight matrix, Attention δ Represents the output of the δth head, MultiHead represents the output result of multi-head attention, and Concat() represents the concatenation operation; After the multi-head self-attention mechanism and then the global adaptive average pooling, the frequency domain features are obtained: Among them, x2 represents the result obtained after the multi-head self-attention mechanism.

5. The method for robust intelligent recognition of surface electromyography for robots according to claim 4, characterized in that: In step 4.3, build the OmniScale module and use the OmniScale module to capture important scale information as follows: OmniScale is a three-layer multi-core structure, each kernel performs the same padded convolution on the input, using represents the kernel size set of the qth layer; q∈{1,2} represents the first two layers, and a convolution kernel of prime size is used, that is, p max represents the largest prime number; q=3 represents the third layer, using convolution kernels of size 1 and 2, i.e.

6. The method for robust intelligent recognition of surface electromyography for robots according to claim 5, characterized in that: In step 5, the DBN network is constructed as follows: The DBN network is constructed by stacking several RBMs. The training process starts from the unsupervised pre-training stage. 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. Each layer is trained independently and sequentially; after pre-training, the network is further fine-tuned through a supervised learning phase; the entire network is fine-tuned using a standard back-propagation algorithm to minimize the error at the output layer; the back-propagation update rule for the error depends on the gradient of the loss function; The energy function E(v,h) of RBM is defined as: Among them, 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 hidden unit β, 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 hidden layer in RBM is: Where v is the visible layer unit, representing the input data, and h is the hidden layer unit, representing the capture of implicit features in the data; E(v,h) represents the energy function of RBM; where Z is the normalization factor: The marginalized probability distribution P(v) of the visible layer is: The activation probability of the hidden unit P(h β =1|v) is: Given the hidden unit, the activation probability of the visible unit P(v α =1|h) is: Where σ(x) is the sigmoid function: When DBN is trained, it starts with the input data, which is passed through multiple hidden layers layer by layer. The output of the top hidden layer is passed to the Softmax layer, which converts the output into the probability distribution P(y θ |z): Among them, y θ represents the weight and bias of the θth category, η θ and b θ are the weight and bias corresponding to the θth category respectively; η l and b l are the weight and bias corresponding to the lth category respectively; The model gives the classification recognition result according to the category corresponding to the maximum probability Right now: It is an operation to find the parameters that maximize the function. In this formula, we need to find the conditional probability P(y θ |z) the largest category, and finally It represents the predicted category obtained after classifying the input data through DBN.

Images