Cross-user myoelectric pattern recognition method integrating domain adaptation and domain generalization

Abstract

The application discloses a cross-user electromyogram pattern recognition method integrating field adaptation and field generalization, first acquires surface electromyogram signals of a source domain user and forms a field generalization sample pair dataset; then a deep learning network model based on class separation and field alignment (CSDAN) is constructed; next, the CSDAN model is trained based on the field generalization sample pair dataset and a field adaptation sample pair dataset; finally, a dual-mode new user action pattern classifier is constructed, and new user action patterns are recognized. The application not only has robustness in cross-user electromyogram pattern recognition in a field adaptation scene requiring only one calibration data, but also has leading cross-user electromyogram pattern recognition accuracy in a field generalization scene without calibration, thereby promoting the wide application of human-computer interface devices in human-computer interaction, clinical rehabilitation and other fields.

Description

Technical Field

[0001] This invention belongs to the field of biomedical electromyography technology, specifically relating to a cross-user electromyography pattern recognition method that integrates domain adaptation and domain generalization. Background Technology

[0002] Surface electromyography (sEMG) is a bioelectrical signal generated during muscle activity. Because sEMG signals are generated before human movements and reflect rich information about muscle activity, decoding this information can reveal the body's movement intentions. Therefore, sEMG signals are often used as control commands for prostheses and exoskeletons, offering hope for motor function reconstruction. With advancements in myoelectric pattern recognition (MPR) technology, more subtle movement intentions can be decoded from sEMG signals. This makes sEMG signals not only useful in clinical rehabilitation but also promising for application in human-computer interaction (HCI) as control commands for HCI devices. This demonstrates the immense potential of EEMG-driven HCI devices in consumer electronics fields such as virtual reality and interactive games.

[0003] Unlike myoelectric prosthetic devices, which are typically used by a single user, human-machine interface (HMI) devices usually need to be able to be used by multiple users simultaneously. Currently, many researchers have developed myoelectric-driven HMI devices, which perform well in laboratory testing environments. However, in real-world applications, due to the non-stationary and random nature of electrophysiological signals, the accuracy of classifiers trained on the original user data significantly decreases or even fails when used by new users. This prevents the device from meeting the practical requirement of being used by multiple users simultaneously. New users typically need to re-collect myoelectric data for different movements and retrain the classifier to take over the same myoelectric interface. This training burden leads to the abandonment of myoelectric interface devices and hinders the development of myoelectric interaction methods.

[0004] Many studies have proposed robust EMG pattern recognition methods for new users, aiming to improve the performance of EMG interfaces across user scenarios. These methods primarily employ domain adaptation and domain generalization techniques from deep learning. Domain adaptation typically involves two approaches: supervised adaptation, which retrains new users on a small amount of labeled data to help them adapt to the classifier. However, this supervised learning method still faces a significant training burden. Therefore, some studies favor unsupervised learning as an alternative approach. Cote et al. achieved user adaptation in a completely unsupervised manner by automatically collecting unlabeled data on different movement patterns from new users during use, thereby expanding the training set and improving the recognition rate. While this method eliminates the need for deliberate collection of calibration data from end users (which is usually done in real-time during use), the user experience is not significantly improved. This is because the amount of data from new users is limited in the initial stages of EMG device integration, leading to poor recognition results and requiring a lengthy data collection process. Therefore, some studies suggest that domain generalization can eliminate the need for calibration, truly reducing the user burden. Researchers hope that models trained on the first n users can be directly applied to the (n+1)th user. However, these methods typically require limited knowledge from end-users, and there is little connection between original and new users, making it difficult to achieve high recognition rates. In summary, while domain adaptation-based methods can improve the recognition rate of EMG interfaces on new users, they still incur a high training burden. Domain generalization-based methods, while having no calibration burden for new users, generally have lower recognition rates. How to ensure high recognition rates while minimizing the calibration burden across different users is a scientific problem that remains to be solved, and its application potential and commercial value are enormous. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a cross-user electromyography (EMG) pattern recognition method integrating domain adaptation and domain generalization. First, it collects surface EMG signals from users in the source domain and constructs a domain generalization sample pair dataset. Then, it constructs a deep learning network model based on Class Separation and Domain Alignment Network (CSDAN). Next, it trains the CSDAN model based on the domain generalization sample pair dataset and the domain adaptation sample pair dataset. Finally, it constructs a bimodal new user action pattern classifier and performs pattern recognition on the new user action patterns. This invention not only demonstrates robustness for cross-user EMG pattern recognition in domain adaptation scenarios requiring only one calibration data point, but also achieves leading cross-user EMG pattern recognition accuracy in completely uncalibrated domain generalization scenarios, promoting the widespread application of human-computer interface devices in fields such as human-computer interaction and clinical rehabilitation.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows:

[0007] Step 1: Using the same high-density surface electromyography (SEMG) electrode, collect surface electromyography (EMG) signals from n users performing m pre-specified actions. After preprocessing the collected signals by data segmentation, activation segment signal extraction, and feature extraction, obtain EMG feature data with m action labels, thus forming the source domain user dataset. Randomly select p data points from the data of each action in the source domain user dataset to obtain n×m×p EMG data points as representative data.

[0008] From all data and representative data in the source domain user dataset, one electromyography data point is randomly selected to form a data pair. If the two data points in the data pair have the same movement pattern, the data pair is marked as "same"; otherwise, the data pair is marked as "different". This forms a domain generalization sample pair dataset.

[0009] Step 2: Construct a deep learning network model based on category separation and domain alignment, namely the CSDAN model;

[0010] Step 2-1: The CSDAN model has a dual-branch structure, including: a shared dual-branch feature extraction module B, a classification module C, and a category separation and domain alignment module D;

[0011] The shared dual-branch feature extraction module B consists of two feature extraction networks with the same parameters; each feature extraction network contains a network units, and each network unit consists of a convolutional layer, an activation layer, and a batch normalization layer; the first a-1 network units have a max pooling layer added after the batch normalization layer; and the a-th network unit has a flattening layer added at the end.

[0012] Step 2-2: Connect the two feature extraction networks to the classification module C and the category separation and neighborhood alignment module D, respectively;

[0013] The classification module C includes: a fully connected layer, a softmax activation layer, and a classification cross-entropy loss function;

[0014] Construct the cross-entropy loss function using equation (1):

[0015]

[0016] The category separation and domain alignment module D includes: a category separation and domain alignment loss function;

[0017] Construct the category separation and domain alignment loss function using equation (2):

[0018]

[0019] In equations (1) and (2), N represents the number of electromyographic samples in the source domain user dataset, c1 is the number of action categories, and y ic p is the one-hot encoded vector of the actual label of the i-th sample. ic It is the predicted probability of the action category obtained from the output of the softmax activation layer, that is, the predicted probability that the i-th sample belongs to the c-th class; This represents the electromyographic sample of the i-th source domain user. This represents the electromyographic sample of the i-th target domain user. Let represent the Euclidean distance of the i-th sample pair. The label represents the i-th sample pair that is "different" or "same"; m represents the boundary value; λ represents the weight of the loss function's contribution to the model;

[0020] Step 3: Train the CSDAN model based on the domain generalization sample pair dataset and the domain adaptation sample pair dataset;

[0021] Step 3-1: If a domain generalization model is used, train the domain generalization electromyography pattern recognition classifier: input the domain generalization sample pair dataset into the two input terminals of the CSDAN model in batches, use the class separation and domain alignment module D to perform gradient descent training on the CSDAN model and update the model parameters, train until the loss function converges, and obtain the domain generalization cross-user electromyography pattern recognition model.

[0022] Step 3-2: If a domain adaptation model is used, construct a domain adaptation sample pair dataset and train a domain adaptation electromyography pattern recognition model;

[0023] The calibration electromyographic signals of m actions of the new user are collected using electromyographic electrodes. After preprocessing the electromyographic signals of the new user according to step 1, q data points with action category labels are randomly selected from each action to obtain a new user calibration dataset consisting of m×q electromyographic data. The source domain user dataset and the new user calibration dataset are combined to form a domain adaptation sample pair dataset according to step 1.

[0024] The domain-adapted sample pair dataset is input into the two input terminals of CSDAN in batches, and the domain-adapted cross-user electromyography pattern recognition model is trained according to step 3-1.

[0025] Step 4: Construct a bimodal new user action pattern classifier to perform pattern recognition on the new user's action patterns; the new user is the (n+1)th user connected to the electromyography electrode, who randomly performs one of m actions;

[0026] Step 4-1: When using the domain generalization mode, the signal to be predicted is preprocessed according to step 1 to obtain electromyographic feature data. The electromyographic feature data is then input into the two input terminals of the CSDAN network. The softmax layer of the classification module C obtains the probability distribution of the largest corresponding class and then makes a decision.

[0027] Step 4-2: When using the domain-adaptive modality, update the data and model according to the process in Step 3-2, and then obtain the prediction results of the action pattern according to Step 4-1.

[0028] Preferably, the source domain user dataset contains N = 2400 electromyographic samples and C = 6 action categories.

[0029] Preferably, the boundary value m = 1; the loss function contributes a weight λ = 0.25 to the model.

[0030] The beneficial effects of this invention are as follows:

[0031] Compared with existing technologies, this invention not only demonstrates robustness for cross-user EMG pattern recognition in domain adaptation scenarios requiring only a minimum of one calibration data point, but also achieves leading cross-user EMG pattern recognition accuracy in completely uncalibrated domain generalization scenarios. Furthermore, it eliminates the need to modify the model structure when switching between domain adaptation and domain generalization scenarios.

[0032] 1. In step 2 of this invention, the dual-branch shared feature extraction module, under the condition of no new user data, improves the model's generalization ability by learning the feature relationships between electromyographic signals of different users and different action patterns in the source domain, and then directly applies it to new user pattern recognition; when calibration conditions are available, it can learn the feature relationships between new users and source domain users to improve the model's adaptability to new users, thus realizing the integration of two cross-user pattern recognition methods, domain generalization and domain adaptation, without changing the model structure.

[0033] 2. Steps 2 and 3 of this invention use category separation and domain alignment loss functions to align the data distribution of different users with the same action category during gradient descent, while separating the data distribution between different categories. This facilitates the improvement of the model's accuracy in cross-user electromyography pattern classification, thereby enhancing the robustness of the electromyography interaction system when used by different users.

[0034] 3. In step 4 of this invention, the model is trained on source domain users to achieve a high generalization ability, and can be used directly after new users join without any new user data. When new user label data is available, the model can be updated quickly. Each action category only requires a minimum of one electromyography (EMG) data, which can achieve leading recognition accuracy. At the same time, it greatly reduces the amount of data required for new users and reduces the burden of re-collecting data when recognizing EMG patterns of new users. Attached Figure Description

[0035] Figure 1 This is a flowchart of the method of the present invention;

[0036] Figure 2 This is a schematic diagram of the electromyography signal acquisition device and electrode configuration of the present invention;

[0037] Figure 3 This is a structural diagram of the deep learning network model for category separation and domain alignment in this invention;

[0038] Figure 4 This is a diagram of the electromyography signal feature extraction network structure of the present invention. Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] This invention provides a cross-user electromyography pattern recognition method that integrates domain adaptation and domain generalization. This method enables the electromyography interface to complete domain adaptation calibration under small sample size conditions, ensuring cross-user recognition rate, and reduces the calibration burden on users through domain generalization without switching model parameters, thereby improving robustness to new user data.

[0041] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0042] The present invention provides a cross-user electromyography pattern recognition method integrating domain adaptation and domain generalization, characterized by comprising the following steps:

[0043] Step 1: Using the same high-density surface electromyography (SEMG) electrode, collect surface electromyography (EMG) signals from n users performing m pre-specified actions. After preprocessing the collected signals by data segmentation, activation segment signal extraction, and feature extraction, obtain EMG feature data with m action labels, thus forming the source domain user dataset. Randomly select p data points from the data of each action in the source domain user dataset to obtain n×m×p EMG data points as representative data; n≥1, m≥1, p≥1.

[0044] From all data and representative data in the source domain user dataset, one electromyographic data point is randomly selected to form a data pair. If the two data points in the data pair have the same movement pattern, the data pair is marked as "same"; otherwise, the data pair is marked as "different". This forms a domain generalization sample pair dataset.

[0045] Step 2: Construct a deep learning network model based on Class Separation and Domain Alignment Network (CSDAN);

[0046] Step 2.1: The CSDAN model has a dual-branch structure, including: a shared dual-branch feature extraction module B, a classification module C, and a category separation and domain alignment module D;

[0047] The shared dual-branch feature extraction module B is composed of two feature extraction networks with the same parameters;

[0048] Each feature extraction network contains a network units, each consisting of a convolutional layer, an activation layer, and a batch normalization layer; the first a-1 network units have a max pooling layer added after the batch normalization layer; and the a-th network unit has a flattening layer added at the end.

[0049] Step 2.2: The two feature extraction networks are respectively connected to the classification module C and the category separation and neighborhood alignment module D;

[0050] The category separation module C includes: a fully connected layer, a softmax activation layer, and a classification cross-entropy loss function;

[0051] The cross-entropy loss function is constructed using equation (1):

[0052]

[0053] The category separation and domain alignment module D includes: category separation loss and domain alignment loss functions;

[0054] The category separation and domain alignment loss function is constructed using equation (2):

[0055]

[0056] In equations (1) and (2), N represents the number of electromyography samples y in the source domain user dataset. ic p is the one-hot encoded vector of the actual label of the i-th sample. ic It is the predicted probability of the action category obtained from the output of the softmax activation layer, that is, the predicted probability that the i-th sample belongs to the c-th class; This represents the electromyographic sample of the i-th source domain user. This represents the electromyographic sample of the i-th target domain user. Let represent the Euclidean distance of the i-th sample pair. The label represents the i-th sample pair that is "different" or "same"; m represents the boundary value; λ represents the weight of the loss function's contribution to the model;

[0057] Step 3: Train the CSDAN model based on the domain generalization sample pair dataset and the domain adaptation sample pair dataset;

[0058] Step 3.1: If a domain generalization model is used, train the domain generalization electromyography pattern recognition classifier: input the domain generalization sample pair dataset into the two input terminals of CSDAN in batches, use the class separation and domain alignment module D to perform gradient descent training on the CSDAN model and update the model parameters, train until the loss function converges, and obtain the domain generalization cross-user electromyography pattern recognition model.

[0059] Step 3.2: If a domain adaptation model is used, construct a domain adaptation sample pair dataset and train a domain adaptation electromyography pattern recognition model.

[0060] The electromyographic electrodes are used to collect calibration electromyographic signals of m actions of a new user. After preprocessing the electromyographic signals of the new user according to step one, q data points with action category labels are randomly selected from each action to obtain a new user calibration dataset consisting of m×q electromyographic data points. The source domain user dataset and the new user calibration dataset are combined to form a domain adaptation sample pair dataset according to step one.

[0061] The domain-adapted sample pair dataset is input into the two input terminals of CSDAN in batches, and the domain-adapted cross-user electromyography pattern recognition model is trained according to step 3.1.

[0062] Step 4: Construct a bimodal new user action pattern classifier to perform pattern recognition on the new user's action patterns; The new user (the (n+1)th user) is connected to electromyographic electrodes and randomly performs one of m actions;

[0063] Step 4.1: When using the domain generalization mode, the signal to be predicted is preprocessed according to step one to obtain electromyographic feature data. The electromyographic feature data is then input into the two input terminals of the CSDAN network. The softmax layer of the classification module C obtains the probability distribution of the maximum corresponding class and then makes a decision.

[0064] Step 4.2: When using the domain-adaptive modality, update the data and model according to the process in Step 3.2, and then obtain the prediction results of the action pattern according to Step 4.1.

[0065] Example:

[0066] This embodiment proposes a deep model and loss function that can be both domain-adaptive and domain-generalizable, and also proposes a dual-modal training method for this model, such as... Figure 1 As shown, it includes the following steps:

[0067] Step 1: Using the same high-density surface electromyography (SEMG) electrode, collect surface electromyography (EMG) signals from n users performing m pre-specified actions. After preprocessing the collected signals by data segmentation, activation segment signal extraction, and feature extraction, obtain EMG feature data with m action labels, thus forming the source domain user dataset. Randomly select p data points from the data of each action in the source domain user dataset to obtain n×m×p EMG data points as representative data; n≥1, m≥1, p≥1.

[0068] From all data and representative data in the source domain user dataset, one electromyography data point is randomly selected to form a data pair. If the two data points in the data pair have the same movement pattern, the data pair is marked as "same"; otherwise, the data pair is marked as "different". This forms a domain generalization sample pair dataset.

[0069] Step 1.1, For example, a 10×10 channel high-density surface electromyography (EMG) electrode configuration was used to collect surface EMG signals from the forearm finger extensor muscles of 7 subjects. During the collection process, the subjects were asked to perform 6 hand gestures (m1-m6). The collection process is as follows: Figure 2 As shown.

[0070] Step 1.2: Perform data segmentation on the acquired signal, using a sliding segmentation with a data window length of 128ms and a step size of 64ms.

[0071] Step 1.3: Extract the active segment signal from the data window. That is, the area under the signal in three consecutive data windows is greater than the area during the resting period. This is considered to be the active segment signal.

[0072] Step 1.4: Extract three electromyographic features from the activation segment signal: wavelength (time domain feature) and median frequency and average frequency (frequency domain features), and generate electromyographic feature data.

[0073] Step 1.5: Label each feature data with a representative action pattern to form the source domain user dataset; randomly select one data point from each action data in the source domain user dataset to obtain 54 (i.e., 9×6×1) electromyographic data points as representative data.

[0074] Step 1.6: Select one electromyographic data point from all data and one representative data point from the source domain user dataset to form a data pair. If the two data points in the data pair have the same action pattern, mark the data pair as "same"; otherwise, mark the data pair as "different". This forms the domain generalization sample pair dataset.

[0075] Step 2: Construct a deep learning network model based on Class Separation and Domain Alignment Network (CSDAN);

[0076] Step 2.1: The CSDAN model has a dual-branch structure, as follows: Figure 3 As shown, it includes: a shared dual-branch feature extraction module B, a classification module C, and a category separation and domain alignment module D;

[0077] The shared dual-branch feature extraction module B consists of two feature extraction networks with the same parameters;

[0078] In this embodiment, each feature extraction network contains two network units, such as... Figure 4 As shown, each network unit consists of a convolutional layer, an activation layer, and a batch normalization layer; the first network unit adds a max pooling layer after the batch normalization layer; the second network unit adds a flattening layer at the end.

[0079] Step 2.2: Connect the two feature extraction networks to the classification module C and the category separation and neighborhood alignment module D, respectively;

[0080] The category separation module C includes: a fully connected layer, a softmax activation layer, and a classification cross-entropy loss function;

[0081] Use equation (1) to construct the classification cross-entropy loss function:

[0082]

[0083] The category separation and domain alignment module D includes: category separation loss and domain alignment loss functions;

[0084] Use equation (2) to construct the category separation and domain alignment loss function:

[0085]

[0086] In this embodiment, in equations (1) and (2), the number of electromyographic samples in the source domain user dataset is N = 2400, the number of action categories is C = 6, and y ic It is the one-hot encoded vector of the actual label of the i-th sample, and the action class prediction probability p is obtained from the output of the softmax activation layer. ic The value is a decimal between 0 and 1, which is the predicted probability that the i-th sample belongs to class C. This represents the electromyographic sample of the i-th source domain user. This represents the electromyographic sample of the i-th target domain user. The Euclidean distance of the i-th sample pair is calculated using the following formula:

[0087]

[0088] The label representing the i-th sample pair that is "different" or "same"; boundary value m = 1; the weight of the loss function's contribution to the model λ = 0.25;

[0089] Step 3: Train the CSDAN model based on the domain generalization sample pair dataset and the domain adaptation sample pair dataset;

[0090] Step 3.1: If a domain generalization model is used, train the domain generalization electromyography pattern recognition classifier: input the domain generalization sample pair dataset into the two input terminals of CSDAN in batches, use the class separation and domain alignment module D to perform gradient descent training on the CSDAN model and update the model parameters, train until the loss function converges, and obtain the domain generalization cross-user electromyography pattern recognition model.

[0091] In this embodiment, the model uses the Adam optimizer during training, with a learning rate of 0.001, a batch size of 256, and 80 iterations.

[0092] Step 3.2: If a domain adaptation model is used, construct a domain adaptation sample pair dataset and train a domain adaptation electromyography pattern recognition model.

[0093] For example, electromyographic electrodes are used to collect calibration electromyographic signals of six actions of a new user. After preprocessing the electromyographic signals of the new user according to step one, one data point with an action category label is randomly selected from each action to obtain a new user calibration dataset consisting of six electromyographic data points. The source domain user dataset and the new user calibration dataset are combined to form a domain adaptation sample pair dataset according to step one.

[0094] The domain-adapted sample pair dataset is input into the two input terminals of CSDAN in batches, and the domain-adapted cross-user electromyography pattern recognition model is trained according to step 3.1.

[0095] Step 4: Construct a bimodal new user action pattern classifier and perform pattern recognition on the new user action patterns; The new user (the 9th user in this embodiment) is connected to electromyography electrodes and randomly performs one of m actions, where m = 6 in this embodiment;

[0096] Step 4.1: When using the domain generalization mode, the signal to be predicted is preprocessed according to step one to obtain electromyographic feature data. The electromyographic feature data is then input into the two input terminals of the CSDAN network. The softmax layer of the classification module C obtains the probability distribution of the maximum corresponding class and then makes a decision.

[0097] Step 4.2: When using the domain-adaptive modality, update the data and model according to the process in Step 3.2, and then measure the prediction results of the action pattern according to Step 4.1.

[0098] To evaluate the performance of the above-mentioned scheme in this invention, a performance comparison evaluation was conducted using non-deep learning methods, deep learning methods, and the method of this invention. The evaluation included cross-user electromyography pattern recognition rate and the amount of calibration data required.

[0099] In the comparison methods, TCA achieves distribution distance alignment by analyzing the transfer components between different domains. BDA and WBDA further improve the robustness of the model to class imbalance by aligning the marginal distribution and conditional distribution while analyzing the domain components. CNN directly transfers the user features of the source domain to the user features of the target domain. This method serves as the baseline level in the comparison methods. DANN distinguishes domains by training the discriminator and deceives the discriminator by training the generator to learn domain-invariant feature representations.

[0100] Table 1 Comparison of recognition rates (%) of cross-user electromyography pattern methods

[0101]

[0102] For example, in Table 1, users 1-8 are used as target domain users for testing model performance, while the electromyography (EMG) data of the other 7 users constitute the source domain user dataset for model training. From the comparison of EMG pattern recognition rates in Table 1, it can be seen that not only is the recognition rate of this invention significantly better than other methods, but the amount of data used for new user calibration is also significantly reduced. In particular, the domain generalization method completely eliminates the need for calibration processes and calibration data, greatly reducing the calibration burden on users.

Claims

1. A cross-user electromyography pattern recognition method integrating domain adaptation and domain generalization, characterized in that, Includes the following steps: Step 1: Using the same high-density surface electromyography (SEMG) electrode, collect surface electromyography (EMG) signals from n users when they perform m pre-specified actions. After preprocessing the collected signals by data segmentation, activation segment signal extraction, and feature extraction, obtain EMG feature data with m action labels, thus forming the source domain user dataset. From the data of each action in the source domain user dataset, p data points are randomly selected to obtain n×m×p electromyographic data points as representative data; From all data and representative data in the source domain user dataset, one electromyography data point is randomly selected to form a data pair. If the two data points in the data pair have the same movement pattern, the data pair is marked as "same"; otherwise, the data pair is marked as "different". This forms a domain generalization sample pair dataset. Step 2: Construct a deep learning network model based on category separation and domain alignment, namely the CSDAN model; Step 2-1: The CSDAN model has a dual-branch structure, including: a shared dual-branch feature extraction module B, a classification module C, and a category separation and domain alignment module D; The shared dual-branch feature extraction module B consists of two feature extraction networks with the same parameters; each feature extraction network contains a network units, and each network unit consists of a convolutional layer, an activation layer, and a batch normalization layer; the first a-1 network units have a max pooling layer added after the batch normalization layer; and the a-th network unit has a flattening layer added at the end. Step 2-2: Connect the two feature extraction networks to the classification module C and the category separation and neighborhood alignment module D, respectively; The classification module C includes: a fully connected layer, a softmax activation layer, and a classification cross-entropy loss function; Construct the cross-entropy loss function using equation (1): The category separation and domain alignment module D includes: a category separation and domain alignment loss function; Construct the category separation and domain alignment loss function using equation (2): In equations (1) and (2), N represents the number of electromyographic samples in the source domain user dataset, c1 is the number of action categories, and y ic p is the one-hot encoded vector of the actual label of the i-th sample. ic It is the predicted probability of the action category obtained from the output of the softmax activation layer, that is, the predicted probability that the i-th sample belongs to the c-th class; This represents the electromyographic sample of the i-th source domain user. This represents the electromyographic sample of the i-th target domain user. Let represent the Euclidean distance of the i-th sample pair. The label represents the i-th sample pair that is "different" or "same"; m represents the boundary value; λ represents the weight of the loss function's contribution to the model; Step 3: Train the CSDAN model based on the domain generalization sample pair dataset and the domain adaptation sample pair dataset; Step 3-1: If a domain generalization model is used, train the domain generalization electromyography pattern recognition classifier: input the domain generalization sample pair dataset into the two input terminals of the CSDAN model in batches, use the class separation and domain alignment module D to perform gradient descent training on the CSDAN model and update the model parameters, train until the loss function converges, and obtain the domain generalization cross-user electromyography pattern recognition model. Step 3-2: If a domain adaptation model is used, construct a domain adaptation sample pair dataset and train a domain adaptation electromyography pattern recognition model; The calibration electromyographic signals of m actions of the new user are collected using electromyographic electrodes. After preprocessing the electromyographic signals of the new user according to step 1, q data points with action category labels are randomly selected from each action to obtain a new user calibration dataset consisting of m×q electromyographic data. The source domain user dataset and the new user calibration dataset are combined to form a domain adaptation sample pair dataset according to step 1. The domain-adapted sample pair dataset is input into the two input terminals of CSDAN in batches, and the domain-adapted cross-user electromyography pattern recognition model is trained according to step 3-1. Step 4: Construct a bimodal new user action pattern classifier to perform pattern recognition on the new user's action patterns; the new user is the (n+1)th user connected to the electromyography electrode, who randomly performs one of m actions; Step 4-1: When using the domain generalization mode, the signal to be predicted is preprocessed according to step 1 to obtain electromyographic feature data. The electromyographic feature data is then input into the two input terminals of the CSDAN network. The softmax layer of the classification module C obtains the probability distribution of the largest corresponding class and then makes a decision. Step 4-2: When using the domain-adaptive modality, update the data and model according to the process in Step 3-2, and then obtain the prediction results of the action pattern according to Step 4-1.

2. The cross-user electromyography pattern recognition method integrating domain adaptation and domain generalization according to claim 1, characterized in that, The source domain user dataset contains N = 2400 electromyographic samples and C = 6 action categories.

3. The cross-user electromyography pattern recognition method integrating domain adaptation and domain generalization according to claim 1, characterized in that, The boundary value m = 1; the weight of the loss function's contribution to the model λ = 0.25.

Images