Threshold self-adaptive noise reduction gesture recognition method and system based on surface myoelectricity
By introducing threshold adaptive noise reduction technology into gesture recognition system based on surface electromyography signals, the gesture recognition accuracy problems under the influence of multiple noises are solved, and higher recognition accuracy and adaptability are achieved.
Patent Information
- Application Number
- CN202411429892.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-14
- Publication Date
- 2025-03-18
AI Technical Summary
Existing gesture recognition methods based on surface electromyography signals cannot effectively remove the influence of multiple types of noise, resulting in a decrease in gesture recognition accuracy.
Adaptive noise reduction based on threshold is adopted, and pre-processed and feature extraction is performed by obtaining historical surface electromyography signals, and initial features are extracted using one-dimensional convolution and residual connections, combined with Dropout and batch normalization, and then adaptive noise reduction is performed through soft threshold functions, and finally gesture classification is performed through the full connection layer.
It effectively reduces the impact of noise on the recognition model, improves the accuracy and generalization ability of gesture recognition, and can perform well in different noise environments.
Smart Images
Figure CN119646389A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of gesture recognition, and in particular to a threshold adaptive noise reduction gesture recognition method and system based on surface electromyography. Background Art
[0002] Surface electromyography (sEMG) is a biomedical signal that measures the current generated by muscles during contraction and represents neuromuscular activity. Surface electromyography contains rich movement information and reflects the subject's intention. Gesture recognition based on surface electromyography has broad application prospects in predicting various muscle activities through human-computer interaction due to its non-invasiveness. It can be used to control rehabilitation equipment, prostheses and hand exoskeleton robots based on surface electromyography to facilitate the movement of special groups such as the disabled and the elderly.
[0003] Academia and industry have conducted extensive research in the field of gesture recognition based on surface electromyography signals: (1) Atzori et al. proposed a method using a one-dimensional convolutional neural network (CNN) to decode surface electromyography signals that have been filtered out of 50Hz power line interference. Compared with existing machine learning methods, the performance is improved by 2%-5%; (2) Fajardo et al. used a 150Hz low-pass filter to filter the surface electromyography signals, merged the features in CNN and TD to complete the classification of 8 gesture classes, and the average accuracy reached 81.54%; (3) Rahimian et al. proposed an attention-based transformer architecture that captures the global information in the surface electromyography signals through a low-pass Butterworth filter, achieving an accuracy of 82.93% in gesture recognition; (4) Li et al. fused ResNet50 with the attention mechanism, and the performance was significantly improved compared with the baseline model.
[0004] Although the above methods have made some progress in gesture recognition based on surface electromyography signals, they ignore the fact that traditional filters cannot completely remove the influence of noise. In addition, these filters can only eliminate the influence of single noise, while surface electromyography signals are affected by multiple types of noise. The noise remaining in the surface electromyography signals will still affect the performance of the recognition model, thereby affecting the accuracy of gesture recognition.
[0005] Therefore, how to provide a threshold adaptive noise reduction gesture recognition method and system based on surface electromyography to improve gesture recognition accuracy has become a technical problem that needs to be solved urgently. Summary of the invention
[0006] The technical problem to be solved by the present invention is to provide a threshold adaptive noise reduction gesture recognition method and system based on surface electromyography to improve the gesture recognition accuracy.
[0007] In a first aspect, the present invention provides a threshold adaptive noise reduction gesture recognition method based on surface electromyography, comprising the following steps:
[0008] Step S1, obtaining a large number of historical surface electromyographic signals for gesture recognition, and constructing a data set after preprocessing each of the historical surface electromyographic signals;
[0009] Step S2, creating a gesture recognition model based on a feature extraction module, an adaptive noise reduction module and a gesture classification module; the feature extraction module, the adaptive noise reduction module and the gesture classification module are connected in sequence;
[0010] Step S3, training the gesture recognition model using the data set, and continuously optimizing the hyperparameters of the gesture recognition model during the training process;
[0011] Step S4: deploying the trained gesture recognition model, and performing gesture recognition using the deployed gesture recognition model.
[0012] Furthermore, the step S1 is specifically as follows:
[0013] A large number of historical surface electromyographic signals for identifying gestures from different subjects are obtained, and each of the historical surface electromyographic signals is overlapped and segmented through a sliding window. A 2 Hz high-pass filtering operation is performed on each of the overlapping and segmented historical surface electromyographic signals, and then the gesture category is labeled for each of the historical surface electromyographic signals after the high-pass filtering operation to complete the preprocessing of each of the historical surface electromyographic signals, and a data set is constructed based on each of the preprocessed historical surface electromyographic signals.
[0014] Furthermore, in step S2, the feature extraction module is used to extract initial features from historical surface electromyographic signals through one-dimensional convolution, and in the process of extracting the initial features, a preset number of neurons are randomly deleted through Dropout to suppress overfitting, and then the initial features are output through a residual connection for retaining effective features and alleviating gradient disappearance, and the initial features are data fitted to obtain fitting results through a nonlinear activation function, and finally the fitting results are batch normalized to obtain gesture features, and the gesture features are output to the adaptive noise reduction module;
[0015] The adaptive noise reduction module is used to adaptively reduce the noise of the gesture features through a soft threshold function to obtain noise reduction features, and output the noise reduction features to the gesture classification module;
[0016] The gesture classification module is used to perform a pooling operation on the noise reduction features and then perform a dimensionality reduction process through a fully connected layer to perform gesture classification and further perform gesture recognition.
[0017] Furthermore, the step S3 is specifically as follows:
[0018] Dividing the data set into a training set and a validation set based on a preset ratio, training the gesture recognition model with the training set, and continuously optimizing hyperparameters of the gesture recognition model including at least a learning rate, a learning decay rate, a constraint weight, a batch size, an optimizer, and a pooling window size during the training process;
[0019] The trained gesture recognition model is verified using the verification set.
[0020] Furthermore, the step S4 is specifically as follows:
[0021] The trained gesture recognition model is deployed to the cloud server to obtain the real-time surface electromyography signal of the gesture to be recognized, and the real-time surface electromyography signal is filtered and denoised before inputting into the gesture recognition model deployed on the cloud server to obtain the gesture recognition result. The gesture recognition model is continuously optimized based on the gesture recognition result, and the recognition log is recorded in real time.
[0022] In a second aspect, the present invention provides a threshold adaptive noise reduction gesture recognition system based on surface electromyography, comprising the following modules:
[0023] A data set construction module, used to obtain a large number of historical surface electromyographic signals for gesture recognition, and to construct a data set after preprocessing each of the historical surface electromyographic signals;
[0024] A gesture recognition model creation module, used to create a gesture recognition model based on a feature extraction module, an adaptive noise reduction module and a gesture classification module; the feature extraction module, the adaptive noise reduction module and the gesture classification module are connected in sequence;
[0025] A gesture recognition model training module, used to train the gesture recognition model using the data set, and continuously optimize the hyperparameters of the gesture recognition model during the training process;
[0026] The gesture recognition module is used to deploy the trained gesture recognition model and perform gesture recognition using the deployed gesture recognition model.
[0027] Furthermore, the data set construction module is specifically used for:
[0028] A large number of historical surface electromyographic signals for identifying gestures from different subjects are obtained, and each of the historical surface electromyographic signals is overlapped and segmented through a sliding window. A 2 Hz high-pass filtering operation is performed on each of the overlapping and segmented historical surface electromyographic signals, and then the gesture category is labeled for each of the historical surface electromyographic signals after the high-pass filtering operation to complete the preprocessing of each of the historical surface electromyographic signals, and a data set is constructed based on each of the preprocessed historical surface electromyographic signals.
[0029] Furthermore, in the gesture recognition model creation module, the feature extraction module is used to extract initial features from historical surface electromyography signals through one-dimensional convolution. In the process of extracting the initial features, a preset number of neurons are randomly deleted through Dropout to suppress overfitting. Then, the initial features are output through a residual connection for retaining effective features and alleviating gradient disappearance. The initial features are fitted with data through a nonlinear activation function to obtain fitting results. Finally, the fitting results are batch normalized to obtain gesture features, and the gesture features are output to the adaptive noise reduction module.
[0030] The adaptive noise reduction module is used to adaptively reduce the noise of the gesture features through a soft threshold function to obtain noise reduction features, and output the noise reduction features to the gesture classification module;
[0031] The gesture classification module is used to perform a pooling operation on the noise reduction features and then perform a dimensionality reduction process through a fully connected layer to perform gesture classification and further perform gesture recognition.
[0032] Furthermore, the gesture recognition model training module is specifically used for:
[0033] Dividing the data set into a training set and a validation set based on a preset ratio, training the gesture recognition model with the training set, and continuously optimizing hyperparameters of the gesture recognition model including at least a learning rate, a learning decay rate, a constraint weight, a batch size, an optimizer, and a pooling window size during the training process;
[0034] The trained gesture recognition model is verified using the verification set.
[0035] Furthermore, the gesture recognition module is specifically used for:
[0036] The trained gesture recognition model is deployed to the cloud server to obtain the real-time surface electromyography signal of the gesture to be recognized, and the real-time surface electromyography signal is filtered and denoised before inputting into the gesture recognition model deployed on the cloud server to obtain the gesture recognition result. The gesture recognition model is continuously optimized based on the gesture recognition result, and the recognition log is recorded in real time.
[0037] The advantages of the present invention are:
[0038] A data set is constructed by acquiring a large amount of historical surface electromyographic signals for identifying gestures and preprocessing them; then a gesture recognition model is created based on the feature extraction module, the adaptive noise reduction module and the gesture classification module, the loss function of the gesture recognition model is set, the gesture recognition model is trained through the data set, the hyperparameters of the gesture recognition model are continuously optimized during the training process, and finally the trained gesture recognition model is deployed, and gesture recognition is performed using the deployed gesture recognition model; that is, gesture recognition is performed through a gesture recognition model including a feature extraction module, an adaptive noise reduction module and a gesture classification module, and before gesture recognition, the input surface electromyographic signal is preprocessed for noise reduction through sliding window segmentation and high-pass filtering to reduce the influence of line interference and thermal noise, and to achieve real-time gesture recognition; the feature extraction module is used to extract the historical surface electromyographic signals through one-dimensional convolution. Initial features are extracted from electromyographic signals. During the initial feature extraction process, a preset number of neurons are randomly deleted through Dropout to suppress overfitting and improve generalization ability. Then, the initial features are output through residual connections for retaining effective features and alleviating gradient disappearance. The initial features are fitted with data through nonlinear activation functions to obtain fitting results, which have the ability to fit complex nonlinear relationships. Finally, the fitting results are batch normalized to obtain gesture features to reduce data deviation. The adaptive denoising module is used to perform adaptive denoising on gesture features through a soft threshold function to overcome the problem that traditional methods can only eliminate single noise. The gesture classification module is used to perform dimensionality reduction processing on the denoised features after pooling operations, so as to classify gestures through a fully connected layer. It can learn the associations and differences between gestures, and ultimately greatly improve the accuracy of gesture recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] The present invention will be further described below in conjunction with embodiments with reference to the accompanying drawings.
[0040] Figure 1 The present invention is a flow chart of a threshold adaptive noise reduction gesture recognition method based on surface electromyography.
[0041] Figure 2 It is a structural schematic diagram of a threshold adaptive noise reduction gesture recognition system based on surface electromyography of the present invention.
[0042] Figure 3 It is a schematic diagram of the sliding window segmentation of the present invention.
[0043] Figure 4 It is a flow chart of gesture recognition of the present invention.
[0044] Figure 5 Schematic diagram of the gesture recognition model of the present invention. DETAILED DESCRIPTION
[0045] The technical solution in the embodiment of the present application has the following general idea: gesture recognition is performed through a gesture recognition model including a feature extraction module, an adaptive noise reduction module and a gesture classification module. Before gesture recognition, the input surface electromyography signal is pre-processed by sliding window segmentation and high-pass filtering to reduce the influence of line interference and thermal noise; the feature extraction module is used to extract initial features from historical surface electromyography signals through one-dimensional convolution. During the initial feature extraction process, a preset number of neurons are randomly deleted by Dropout to suppress overfitting and improve generalization ability. Then, the initial features are output through residual connections for retaining effective features and alleviating gradient disappearance. The initial features are fitted to data through a nonlinear activation function to obtain fitting results, which have the ability to fit complex nonlinear relationships. Finally, the fitting results are batch normalized to obtain gesture features to reduce data deviation; the adaptive noise reduction module is used to adaptively reduce the noise of gesture features through a soft threshold function to overcome the problem that only single noise can be eliminated in the traditional method; the gesture classification module is used to perform dimensionality reduction processing through a fully connected layer after pooling the noise reduction features to perform gesture classification, and can learn the associations and differences between gestures to improve gesture recognition accuracy.
[0046] Please refer to Figures 1 to 5 As shown, a preferred embodiment of a threshold adaptive noise reduction gesture recognition method based on surface electromyography of the present invention comprises the following steps:
[0047] Step S1, obtaining a large number of historical surface electromyographic signals for gesture recognition, and constructing a data set after preprocessing each of the historical surface electromyographic signals;
[0048] Step S2, creating a gesture recognition model (TADsEMG) based on a feature extraction module, an adaptive noise reduction module and a gesture classification module; the feature extraction module, the adaptive noise reduction module and the gesture classification module are connected in sequence;
[0049] Step S3, training the gesture recognition model using the data set, and continuously optimizing the hyperparameters of the gesture recognition model during the training process;
[0050] Step S4: deploying the trained gesture recognition model, and performing gesture recognition using the deployed gesture recognition model.
[0051] The present invention utilizes the learning ability of deep neural networks (gesture recognition models) to automatically learn the gesture features of surface electromyographic signals related to gestures, and automatically sets the denoising threshold through a soft threshold function to reduce the impact of noise, thereby improving the feature extraction ability of deep neural networks and further improving the accuracy of gesture recognition.
[0052] In the model ablation of the present invention, the adaptive noise reduction module can eliminate the influence of noise and improve the model performance; in different noise environments, the performance of the model is better than the existing model; under the condition of no noise, the average recognition accuracy of four data sets is better than the existing model; in terms of scalability, the adaptive noise reduction module can be transplanted into the residual structure based on one-dimensional convolution, and the portability is strong. That is, the present invention effectively solves the influence of residual noise on data quality by organically integrating the feature extraction module and the adaptive noise reduction module, improves the feature extraction ability of the model, and has achieved remarkable results in noisy conditions and average recognition performance, providing a new exploration path for human-computer interaction based on surface electromyography signals.
[0053] The step S1 is specifically as follows:
[0054] A large number of historical surface electromyographic signals for identifying gestures from different subjects are obtained, and each of the historical surface electromyographic signals is overlapped and segmented through a sliding window. A 2 Hz high-pass filtering operation is performed on each of the overlapping and segmented historical surface electromyographic signals, and then the gesture category is labeled for each of the historical surface electromyographic signals after the high-pass filtering operation to complete the preprocessing of each of the historical surface electromyographic signals, and a data set is constructed based on each of the preprocessed historical surface electromyographic signals.
[0055] In order to realize real-time gesture recognition based on surface electromyography signals, overlapping segmentation is selected, and a sliding window is used to simulate the real-time processing and recognition of surface electromyography signals to realize real-time gesture recognition.
[0056] In the step S2, the feature extraction module is used to extract initial features from historical surface electromyographic signals through one-dimensional convolution, and in the process of extracting the initial features, a preset number of neurons are randomly deleted through Dropout to suppress overfitting, and then the initial features are output through a residual connection for retaining effective features and alleviating gradient disappearance, and the initial features are fitted to the data through a nonlinear activation function to obtain a fitting result, and finally the fitting result is batch normalized to obtain a gesture feature, and the gesture feature is output to the adaptive noise reduction module;
[0057] The feature extraction module allows the direct extraction of gesture features in the time domain without any data conversion, making the gesture feature extraction process more convenient and faster. In one-dimensional convolution, the convolution kernel moves along the time step to extract gesture features. As the network deepens, richer features are learned and have a larger receptive field. After the one-dimensional convolution, a Dropout of 0.2 is used to randomly delete neurons from the neural network during training to reduce the complexity of the network and improve the generalization ability of the model. Batch normalization (BatchNorm) is used to reduce data deviation, standardize output, and accelerate model training. The ReLU function is used as a nonlinear activation function for feature extraction, so that the neural network has the ability to fit complex nonlinear relationships.
[0058] The adaptive noise reduction module is used to adaptively reduce the noise of gesture features through a soft threshold function to obtain noise reduction features, and output the noise reduction features to the gesture classification module; that is, thresholds are set for different samples through a neural network, and the learning ability of the neural network is used to reduce the noise in the surface electromyography signal to improve the feature extraction ability;
[0059] The adaptive denoising module uses global average pooling and absolute value to calculate the average value of each feature channel, which is used to compress features and extract main features. Then, a two-layer FC network is used to obtain scaling parameters. Subsequently, the output features are constrained by the ReLU function, and the negative part of the output is set to zero. This process makes some neurons in the neural network inactive, thereby enhancing the sparsity of the model. The threshold is dynamically set according to the feature distribution of each sample, and the soft threshold algorithm uses the threshold to complete signal filtering.
[0060] The formula of the soft threshold function is:
[0061]
[0062] τ=y c α;
[0063]
[0064] α=ReLU(s)=max(s,0);
[0065] Among them, y represents the noise reduction feature; s represents the gesture feature; τ represents the threshold learned by the neural network; y c Represents the output of the Absolute layer (absolute value function) and the GAP layer (global average pooling); i, j, c are the width, height, and channel index of the gesture feature; α represents the scaling parameter obtained by the two FC layers and the ReLU function;
[0066] The derivative of the denoising feature relative to the gesture feature is 1 or 0, which effectively alleviates the problem of gradient disappearance and explosion:
[0067]
[0068] The gesture classification module is used to perform a pooling operation on the noise reduction features, and then perform a dimensionality reduction process through a fully connected layer to perform gesture classification, and then perform gesture recognition, that is, to map the number of network dimensions to the number of gesture categories.
[0069] The gesture classification module undergoes a pooling operation and sends highly abstract denoising features to the fully connected layer for gesture recognition. These denoising features will be further processed and transformed so that the model can learn the associations and differences between gestures and ultimately achieve accurate classification. In the fully connected layer, the input dimension is 512 and the output dimension is the number of gestures.
[0070] The step S3 is specifically as follows:
[0071] Dividing the data set into a training set and a validation set based on a preset ratio, training the gesture recognition model with the training set, and continuously optimizing hyperparameters of the gesture recognition model including at least a learning rate, a learning decay rate, a constraint weight, a batch size, an optimizer, and a pooling window size during the training process;
[0072] The trained gesture recognition model is verified using the verification set.
[0073] The step S4 is specifically as follows:
[0074] The trained gesture recognition model is deployed to the cloud server to obtain the real-time surface electromyography signal of the gesture to be recognized, and the real-time surface electromyography signal is filtered and denoised before inputting into the gesture recognition model deployed on the cloud server to obtain the gesture recognition result. The gesture recognition model is continuously optimized based on the gesture recognition result, and the recognition log is recorded in real time.
[0075] A preferred embodiment of a threshold adaptive noise reduction gesture recognition system based on surface electromyography of the present invention includes the following modules:
[0076] A data set construction module, used to obtain a large number of historical surface electromyographic signals for gesture recognition, and to construct a data set after preprocessing each of the historical surface electromyographic signals;
[0077] A gesture recognition model creation module, used to create a gesture recognition model (TADsEMG) based on a feature extraction module, an adaptive noise reduction module and a gesture classification module; the feature extraction module, the adaptive noise reduction module and the gesture classification module are connected in sequence;
[0078] A gesture recognition model training module, used to train the gesture recognition model using the data set, and continuously optimize the hyperparameters of the gesture recognition model during the training process;
[0079] The gesture recognition module is used to deploy the trained gesture recognition model and perform gesture recognition using the deployed gesture recognition model.
[0080] The present invention utilizes the learning ability of deep neural networks (gesture recognition models) to automatically learn the gesture features of surface electromyographic signals related to gestures, and automatically sets the denoising threshold through a soft threshold function to reduce the impact of noise, thereby improving the feature extraction ability of deep neural networks and further improving the accuracy of gesture recognition.
[0081] In the model ablation of the present invention, the adaptive noise reduction module can eliminate the influence of noise and improve the model performance; in different noise environments, the performance of the model is better than the existing model; under the condition of no noise, the average recognition accuracy of four data sets is better than the existing model; in terms of scalability, the adaptive noise reduction module can be transplanted into the residual structure based on one-dimensional convolution, and the portability is strong. That is, the present invention effectively solves the influence of residual noise on data quality by organically integrating the feature extraction module and the adaptive noise reduction module, improves the feature extraction ability of the model, and has achieved remarkable results in noisy conditions and average recognition performance, providing a new exploration path for human-computer interaction based on surface electromyography signals.
[0082] The data set construction module is specifically used for:
[0083] A large number of historical surface electromyographic signals for identifying gestures from different subjects are obtained, and each of the historical surface electromyographic signals is overlapped and segmented through a sliding window. A 2 Hz high-pass filtering operation is performed on each of the overlapping and segmented historical surface electromyographic signals, and then the gesture category is labeled for each of the historical surface electromyographic signals after the high-pass filtering operation to complete the preprocessing of each of the historical surface electromyographic signals, and a data set is constructed based on each of the preprocessed historical surface electromyographic signals.
[0084] In order to realize real-time gesture recognition based on surface electromyography signals, overlapping segmentation is selected, and a sliding window is used to simulate the real-time processing and recognition of surface electromyography signals to realize real-time gesture recognition.
[0085] In the gesture recognition model creation module, the feature extraction module is used to extract initial features from historical surface electromyography signals through one-dimensional convolution. In the process of extracting the initial features, a preset number of neurons are randomly deleted through Dropout to suppress overfitting. Then, the initial features are output through residual connections for retaining effective features and alleviating gradient disappearance. The initial features are fitted with data through a nonlinear activation function to obtain fitting results. Finally, the fitting results are batch normalized to obtain gesture features, and the gesture features are output to the adaptive noise reduction module.
[0086] The feature extraction module allows the direct extraction of gesture features in the time domain without any data conversion, making the gesture feature extraction process more convenient and faster. In one-dimensional convolution, the convolution kernel moves along the time step to extract gesture features. As the network deepens, richer features are learned and have a larger receptive field. After the one-dimensional convolution, a Dropout of 0.2 is used to randomly delete neurons from the neural network during training to reduce the complexity of the network and improve the generalization ability of the model. Batch normalization (BatchNorm) is used to reduce data deviation, standardize output, and accelerate model training. The ReLU function is used as a nonlinear activation function for feature extraction, so that the neural network has the ability to fit complex nonlinear relationships.
[0087] The adaptive noise reduction module is used to adaptively reduce the noise of gesture features through a soft threshold function to obtain noise reduction features, and output the noise reduction features to the gesture classification module; that is, thresholds are set for different samples through a neural network, and the learning ability of the neural network is used to reduce the noise in the surface electromyography signal to improve the feature extraction ability;
[0088] The adaptive denoising module uses global average pooling and absolute value to calculate the average value of each feature channel, which is used to compress features and extract main features. Then, a two-layer FC network is used to obtain scaling parameters. Subsequently, the output features are constrained by the ReLU function, and the negative part of the output is set to zero. This process makes some neurons in the neural network inactive, thereby enhancing the sparsity of the model. The threshold is dynamically set according to the feature distribution of each sample, and the soft threshold algorithm uses the threshold to complete signal filtering.
[0089] The formula of the soft threshold function is:
[0090]
[0091] τ=y c α;
[0092]
[0093] α=ReLU(s)=max(s,0);
[0094] Among them, y represents the noise reduction feature; s represents the gesture feature; τ represents the threshold learned by the neural network; y c Represents the output of the Absolute layer (absolute value function) and the GAP layer (global average pooling); i, j, c are the width, height, and channel index of the gesture feature; α represents the scaling parameter obtained by the two FC layers and the ReLU function;
[0095] The derivative of the denoising feature relative to the gesture feature is 1 or 0, which effectively alleviates the problem of gradient disappearance and explosion:
[0096]
[0097] The gesture classification module is used to perform a pooling operation on the noise reduction features, and then perform a dimensionality reduction process through a fully connected layer to perform gesture classification, and then perform gesture recognition, that is, to map the number of network dimensions to the number of gesture categories.
[0098] The gesture classification module undergoes a pooling operation and sends highly abstract denoising features to the fully connected layer for gesture recognition. These denoising features will be further processed and transformed so that the model can learn the associations and differences between gestures and ultimately achieve accurate classification. In the fully connected layer, the input dimension is 512 and the output dimension is the number of gestures.
[0099] The gesture recognition model training module is specifically used for:
[0100] Dividing the data set into a training set and a validation set based on a preset ratio, training the gesture recognition model with the training set, and continuously optimizing hyperparameters of the gesture recognition model including at least a learning rate, a learning decay rate, a constraint weight, a batch size, an optimizer, and a pooling window size during the training process;
[0101] The trained gesture recognition model is verified using the verification set.
[0102] The gesture recognition module is specifically used for:
[0103] The trained gesture recognition model is deployed to the cloud server to obtain the real-time surface electromyography signal of the gesture to be recognized, and the real-time surface electromyography signal is filtered and denoised before inputting into the gesture recognition model deployed on the cloud server to obtain the gesture recognition result. The gesture recognition model is continuously optimized based on the gesture recognition result, and the recognition log is recorded in real time.
[0104] In summary, the advantages of the present invention are:
[0105] A data set is constructed by acquiring a large amount of historical surface electromyographic signals for identifying gestures and preprocessing them; then a gesture recognition model is created based on the feature extraction module, the adaptive noise reduction module and the gesture classification module, the loss function of the gesture recognition model is set, the gesture recognition model is trained through the data set, the hyperparameters of the gesture recognition model are continuously optimized during the training process, and finally the trained gesture recognition model is deployed, and gesture recognition is performed using the deployed gesture recognition model; that is, gesture recognition is performed through a gesture recognition model including a feature extraction module, an adaptive noise reduction module and a gesture classification module, and before gesture recognition, the input surface electromyographic signal is preprocessed for noise reduction through sliding window segmentation and high-pass filtering to reduce the influence of line interference and thermal noise, and to achieve real-time gesture recognition; the feature extraction module is used to extract the historical surface electromyographic signals through one-dimensional convolution. Initial features are extracted from electromyographic signals. During the initial feature extraction process, a preset number of neurons are randomly deleted through Dropout to suppress overfitting and improve generalization ability. Then, the initial features are output through residual connections for retaining effective features and alleviating gradient disappearance. The initial features are fitted with data through nonlinear activation functions to obtain fitting results, which have the ability to fit complex nonlinear relationships. Finally, the fitting results are batch normalized to obtain gesture features to reduce data deviation. The adaptive denoising module is used to perform adaptive denoising on gesture features through a soft threshold function to overcome the problem that traditional methods can only eliminate single noise. The gesture classification module is used to perform dimensionality reduction processing on the denoised features after pooling operations, so as to classify gestures through a fully connected layer. It can learn the associations and differences between gestures, and ultimately greatly improve the accuracy of gesture recognition.
[0106] Although the specific implementation modes of the present invention are described above, those skilled in the art should understand that the specific implementation modes described are only illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included in the scope of protection of the claims of the present invention.
Claims
1. A threshold adaptive noise reduction gesture recognition method based on surface electromyography, characterized in that: The steps include: Step S1, obtaining a large number of historical surface electromyographic signals for gesture recognition, and constructing a data set after preprocessing each of the historical surface electromyographic signals; Step S2, creating a gesture recognition model based on a feature extraction module, an adaptive noise reduction module and a gesture classification module; the feature extraction module, the adaptive noise reduction module and the gesture classification module are connected in sequence; Step S3, training the gesture recognition model using the data set, and continuously optimizing the hyperparameters of the gesture recognition model during the training process; Step S4: deploying the trained gesture recognition model, and performing gesture recognition using the deployed gesture recognition model.
2. The method for gesture recognition based on surface electromyography with threshold adaptive noise reduction as claimed in claim 1, characterized in that: The step S1 is specifically as follows: A large number of historical surface electromyographic signals for identifying gestures from different subjects are obtained, and each of the historical surface electromyographic signals is overlapped and segmented through a sliding window. A 2 Hz high-pass filtering operation is performed on each of the overlapping and segmented historical surface electromyographic signals, and then the gesture category is labeled for each of the historical surface electromyographic signals after the high-pass filtering operation to complete the preprocessing of each of the historical surface electromyographic signals, and a data set is constructed based on each of the preprocessed historical surface electromyographic signals.
3. The method for gesture recognition based on surface electromyography with threshold adaptive noise reduction as claimed in claim 1, characterized in that: In the step S2, the feature extraction module is used to extract initial features from historical surface electromyographic signals through one-dimensional convolution, and in the process of extracting the initial features, a preset number of neurons are randomly deleted through Dropout to suppress overfitting, and then the initial features are output through a residual connection for retaining effective features and alleviating gradient disappearance, and the initial features are fitted to the data through a nonlinear activation function to obtain a fitting result, and finally the fitting result is batch normalized to obtain a gesture feature, and the gesture feature is output to the adaptive noise reduction module; The adaptive noise reduction module is used to adaptively reduce the noise of the gesture features through a soft threshold function to obtain noise reduction features, and output the noise reduction features to the gesture classification module; The gesture classification module is used to perform a pooling operation on the noise reduction features and then perform a dimensionality reduction process through a fully connected layer to perform gesture classification and further perform gesture recognition.
4. The method for gesture recognition based on surface electromyography with threshold adaptive noise reduction as claimed in claim 1, characterized in that: The step S3 is specifically as follows: Dividing the data set into a training set and a validation set based on a preset ratio, training the gesture recognition model with the training set, and continuously optimizing hyperparameters of the gesture recognition model including at least a learning rate, a learning decay rate, a constraint weight, a batch size, an optimizer, and a pooling window size during the training process; The trained gesture recognition model is verified using the verification set.
5. The method for gesture recognition based on surface electromyography with threshold adaptive noise reduction as claimed in claim 1, characterized in that: The step S4 is specifically as follows: The trained gesture recognition model is deployed to the cloud server to obtain the real-time surface electromyography signal of the gesture to be recognized, and the real-time surface electromyography signal is filtered and denoised before inputting into the gesture recognition model deployed on the cloud server to obtain the gesture recognition result. The gesture recognition model is continuously optimized based on the gesture recognition result, and the recognition log is recorded in real time.
6. A threshold adaptive noise reduction gesture recognition system based on surface electromyography, characterized in that: Includes the following modules: A data set construction module, used to obtain a large number of historical surface electromyographic signals for gesture recognition, and to construct a data set after preprocessing each of the historical surface electromyographic signals; A gesture recognition model creation module, used to create a gesture recognition model based on a feature extraction module, an adaptive noise reduction module and a gesture classification module; the feature extraction module, the adaptive noise reduction module and the gesture classification module are connected in sequence; A gesture recognition model training module, used to train the gesture recognition model using the data set, and continuously optimize the hyperparameters of the gesture recognition model during the training process; The gesture recognition module is used to deploy the trained gesture recognition model and perform gesture recognition using the deployed gesture recognition model.
7. A threshold adaptive noise reduction gesture recognition system based on surface electromyography as claimed in claim 6, characterized in that: The data set construction module is specifically used for: A large number of historical surface electromyographic signals for identifying gestures from different subjects are obtained, and each of the historical surface electromyographic signals is overlapped and segmented through a sliding window. A 2 Hz high-pass filtering operation is performed on each of the overlapping and segmented historical surface electromyographic signals, and then the gesture category is labeled for each of the historical surface electromyographic signals after the high-pass filtering operation to complete the preprocessing of each of the historical surface electromyographic signals, and a data set is constructed based on each of the preprocessed historical surface electromyographic signals.
8. The threshold adaptive noise reduction gesture recognition system based on surface electromyography according to claim 6, characterized in that: In the gesture recognition model creation module, the feature extraction module is used to extract initial features from historical surface electromyography signals through one-dimensional convolution. In the process of extracting the initial features, a preset number of neurons are randomly deleted through Dropout to suppress overfitting. Then, the initial features are output through residual connections for retaining effective features and alleviating gradient disappearance. The initial features are fitted with data through a nonlinear activation function to obtain fitting results. Finally, the fitting results are batch normalized to obtain gesture features, and the gesture features are output to the adaptive noise reduction module. The adaptive noise reduction module is used to adaptively reduce the noise of the gesture features through a soft threshold function to obtain noise reduction features, and output the noise reduction features to the gesture classification module; The gesture classification module is used to perform a pooling operation on the noise reduction features and then perform a dimensionality reduction process through a fully connected layer to perform gesture classification and further perform gesture recognition.
9. The threshold adaptive noise reduction gesture recognition system based on surface electromyography according to claim 6, characterized in that: The gesture recognition model training module is specifically used for: Dividing the data set into a training set and a validation set based on a preset ratio, training the gesture recognition model with the training set, and continuously optimizing hyperparameters of the gesture recognition model including at least a learning rate, a learning decay rate, a constraint weight, a batch size, an optimizer, and a pooling window size during the training process; The trained gesture recognition model is verified using the verification set.
10. The threshold adaptive noise reduction gesture recognition system based on surface electromyography according to claim 6, characterized in that: The gesture recognition module is specifically used for: The trained gesture recognition model is deployed to the cloud server to obtain the real-time surface electromyography signal of the gesture to be recognized, and the real-time surface electromyography signal is filtered and denoised before inputting into the gesture recognition model deployed on the cloud server to obtain the gesture recognition result. The gesture recognition model is continuously optimized based on the gesture recognition result, and the recognition log is recorded in real time.