Semi-supervised motor imagery electroencephalogram classification algorithm based on double-graph convolutional network mutual supervision
Through the semi-supervised method of mutual supervision of dual-graph convolutional networks, the pseudo-label optimization model parameters are used to solve the problem of high dependence on labeled data in the existing technology, and efficient classification of motor imagination EEG signals is achieved, reducing the cost of the brain-computer interface system.
Patent Information
- Application Number
- CN202510674069.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-23
- Publication Date
- 2025-08-15
AI Technical Summary
The existing technology requires a large amount of labeling data when training deep learning models, which leads to high cost of brain-computer interface systems and is difficult to effectively improve performance. Especially in the classification of sports imagination EEG signal, there are problems such as large individual differences and high labeling costs.
The semi-supervised motion imaginary EEG classification algorithm based on mutual supervision of dual-graph convolution network is adopted. By constructing an adjacency matrix and feature matrix, two independently initialized graph convolution models are used to generate pseudo-labels on labeled and unlabeled data, optimize model parameters, and reduce dependence on labeled data.
It improves the classification accuracy and stability of the electroencephalogram signal of the motor imagination, reduces the need for large-scale annotation of data, and reduces the cost of design and performance evaluation of brain-computer interface systems.
Smart Images
Figure CN120493019A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of biomedical engineering brain-computer interface, and in particular relates to a semi-supervised motor imagery EEG classification algorithm based on mutual supervision of dual-graph convolutional networks. Background Art
[0002] A brain-computer interface (BCI) is a communication system that establishes communication and control between the human brain and devices without relying on peripheral nerves and muscles. It is widely used in medical fields such as brain-controlled robotic arms, intelligent prosthetics, and neurorehabilitation. Motor imagery (MI), a commonly used paradigm in BCI systems, utilizes motor imagery signals from the brain to control external devices, providing a new means of communication and control for patients with movement disorders. This approach holds significant research significance and promising application prospects in helping patients with movement disorders improve their quality of life and accelerate their recovery.
[0003] The EEG decoding module is a core component of the motor imagery-based brain-computer interface (MI-BCI) system. The key to decoding lies in extracting effective features from complex EEG signals and achieving accurate classification. Deep learning has become a popular and highly effective method for EEG decoding, but training deep learning models typically requires extensive labeled data. In practice, acquiring high-quality labeled EEG data often faces challenges such as high time costs, complex operations, and significant individual variability. These challenges limit the performance of MI-BCI systems and significantly increase their cost.
[0004] To address the above issues, this study proposed a semi-supervised motor imagery EEG classification algorithm based on mutual supervision of dual graph convolutional networks to reduce the labeling cost of EEG data. This method uses two independently initialized graph convolutional models to predict the input data respectively, and generates pseudo-labels for each other on unlabeled data, thereby constructing a supervised loss function and a pseudo-supervised loss function. By minimizing the total loss function, the model parameters are optimized to improve classification accuracy. In this process, the dynamic topological structure between electrodes is constructed based on the Pearson correlation coefficient, and the spatiotemporal information of EEG is extracted in combination with specific EEG features, enabling the model to more effectively learn the potential feature distribution of EEG signals. This method reduces the dependence on large-scale labeled data, improves the stability and robustness of motor imagery EEG signal classification, and provides a new technical solution for low-cost and efficient BCI applications. Summary of the Invention
[0005] In view of this, the present invention aims to propose a semi-supervised motor imagery EEG classification algorithm based on mutual supervision of dual-graph convolutional networks, which can reduce the demand for labeled EEG data and lower the application cost of motor imagery brain-computer interface.
[0006] To achieve the above object, the technical solution of the present invention is achieved as follows: S1, preprocessing the original EEG signal, performing bandpass filtering, time window truncation and baseline correction in sequence; S2. Construct an adjacency matrix based on the correlation coefficient between electrodes, extract specific EEG features to construct a feature matrix, and build a graph convolution model with two independently initialized parameters; S3. Both models use labeled data for supervised learning and use cross entropy loss as the supervision loss; S4. The two models use unlabeled data to generate pseudo labels as the supervision signal of the other model, calculate the bidirectional cross entropy loss as the pseudo supervision loss, and implement model training by combining the supervision loss and the pseudo supervision loss.
[0007] Furthermore, in step S1, the original EEG signal is preprocessed in sequence, including Butterworth band-pass filtering, trial segmentation, time window truncation and baseline correction.
[0008] Furthermore, in step S2, an adjacency matrix is constructed based on the correlation coefficient between electrodes, a feature matrix is constructed by extracting specific EEG features, and a graph convolution model with two independently initialized parameters is constructed, including the following steps: S201, for each trial data after preprocessing, calculate the Pearson correlation coefficient between channels based on the sampling points; S202, set a threshold, if the absolute value of the correlation coefficient is greater than the threshold, it is recorded as 1, otherwise it is recorded as 0, and generate and output the channel adjacency matrix; S203, extracting specific EEG features of the EEG signal of each channel as node features, and constructing a feature matrix; S204. Construct a graph convolutional network model with two independently initialized parameters based on the adjacency matrix and the feature matrix.
[0009] Furthermore, in step S3, the two models are supervised using labeled data and cross entropy loss is used as the supervision loss, including the following steps: S301: Input the pre-processed labeled EEG data into two independently initialized graph convolutional networks, and output the corresponding predicted classification confidence vectors; S302: After outputting the classification confidence vector, a loss function is calculated based on the label of the corresponding EEG data for supervision, and cross entropy loss is selected as the loss function; S303. Add the cross entropy losses of the two models to obtain the supervised loss, and set the weight to control the proportion of the pseudo supervised loss to the overall loss.
[0010] Furthermore, the classification confidence vector in step S301 is the network output after softmax normalization, representing the probability distribution of different categories.
[0011] Furthermore, the supervision loss function formula in step S303 is as follows: in, Represents the classification confidence vector of the two network outputs, Represents the corresponding true labels of the two network input data, is the cross entropy loss function, are the loss function weight parameters of the two networks.
[0012] Furthermore, in step S4, the two models use unlabeled data to generate pseudo labels as the supervision signal of the other model, calculate the bidirectional cross entropy loss as the pseudo supervision loss, and integrate the supervision loss and pseudo supervision loss to implement model training, including the following steps: S401, the unlabeled EEG data is passed through two graph convolutional networks to obtain two confidence vectors; S402. Use the argmax function to obtain the class with the highest confidence and generate the unique hot vectors of the two networks.
[0013] S403. Define pseudo-supervision loss, which is bidirectional and consists of the cross entropy loss between the one-hot label vector obtained by one network and the classification confidence vector of another network; S404. The total loss function consists of supervised loss and pseudo-supervised loss. Through the back-propagation algorithm, the model will gradually update the weights and biases of the network to minimize the total loss function and converge to optimize the model and achieve its best performance.
[0014] Furthermore, in step S403, the formula of the pseudo-supervision loss function is as follows: in, is the confidence vector of the classification The calculated one-hot label vector.
[0015] Furthermore, in step S404, the total loss function of the model is in is the weight, that is, the proportion of pseudo-supervision loss to the overall loss, and the Sigmoid function is used for real-time adjustment.
[0016] Furthermore, through the back-propagation algorithm, the model will gradually update the weights and biases of the network so that the loss function Minimize and converge to optimize the model and reach its best performance.
[0017] A computer-readable storage medium is characterized by storing a computer program, which, when executed by a processor, implements a semi-supervised motor imagery EEG classification algorithm based on mutual supervision of a dual-graph convolutional network.
[0018] Compared with the existing technology, the semi-supervised motor imagery EEG classification algorithm based on dual-graph convolutional network mutual supervision described in the present invention has the following advantages: (1) The classification method described in the present invention is based on a graph convolutional neural network model. By constructing an adjacency matrix and a feature matrix, it can effectively capture the spatial features between electrodes in the EEG signal and the temporal features of each channel, thereby improving the efficiency of applying different types of features. (2) The classification method described in the present invention promotes mutual learning between the two models through a semi-supervised method based on pseudo-label consistency supervision, and can obtain higher model classification performance with less labeled EEG data and a large amount of unlabeled EEG signals, thereby reducing the demand for the amount of labeled motor imagery EEG data and lowering the cost of brain-computer interface system design and performance evaluation. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] The accompanying drawings, which constitute part of the present invention, are provided to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are provided to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings: Figure 1 This is a flowchart of a semi-supervised motor imagery EEG classification algorithm based on mutual supervision of dual-graph convolutional networks according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a semi-supervised motor imagery EEG classification algorithm based on mutual supervision of dual-graph convolutional networks according to an embodiment of the present invention; DETAILED DESCRIPTION
[0020] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.
[0021] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.
[0022] like Figures 1 to 2 As shown in FIG, a semi-supervised motor imagery EEG classification algorithm based on mutual supervision of dual-graph convolutional networks includes the following steps: S1, preprocessing the original EEG signal, performing bandpass filtering, time window truncation and baseline correction in sequence; S2. Construct an adjacency matrix based on the correlation coefficient between electrodes, extract specific EEG features to construct a feature matrix, and build a graph convolution model with two independently initialized parameters; S3. Both models use labeled data for supervised learning and use cross entropy loss as the supervision loss; S4. The two models use unlabeled data to generate pseudo labels as the supervision signal of the other model, calculate the bidirectional cross entropy loss as the pseudo supervision loss, and implement model training by combining the supervision loss and the pseudo supervision loss.
[0023] In step S1, the original EEG signal is preprocessed in sequence, including Butterworth bandpass filtering, trial segmentation, time window truncation and baseline correction.
[0024] In step S2, an adjacency matrix is constructed based on the correlation coefficient between electrodes, EEG features are extracted to construct a feature matrix, and a graph convolution model with two independently initialized parameters is constructed, including the following steps: S201, for each trial data after preprocessing, calculate the Pearson correlation coefficient between channels based on the sampling points; S202, set a threshold, if the absolute value of the correlation coefficient is greater than the threshold, it is recorded as 1, otherwise it is recorded as 0, and generate and output the channel adjacency matrix; S203, extracting specific EEG features of the EEG signal of each channel as node features, and constructing a feature matrix; S204. Construct a graph convolutional network model with two independently initialized parameters based on the adjacency matrix and the feature matrix.
[0025] In step S3, the two models use labeled data for supervised learning, using cross entropy loss as the supervised loss, including the following steps: S301: Input the pre-processed labeled EEG data into two independently initialized graph convolutional networks, and output the corresponding predicted classification confidence vectors; S302: After outputting the classification confidence vector, a loss function is calculated based on the label of the corresponding EEG data for supervision, and cross entropy loss is selected as the loss function; S303. Add the cross entropy losses of the two models to obtain the supervised loss, and set the weight to control the proportion of the pseudo supervised loss to the overall loss.
[0026] In step S301, the classification confidence vector is the network output after softmax normalization, representing the probability distribution of different categories.
[0027] In step S303, the supervision loss function formula is as follows: in, Represents the classification confidence vector of the two network outputs, Represents the corresponding true labels of the two network input data, is the cross entropy loss function, are the loss function weight parameters of the two networks.
[0028] In step S4, the two models use unlabeled data to generate pseudo labels as the supervision signal of the other model, calculate the bidirectional cross entropy loss as the pseudo supervision loss, and integrate the supervision loss and pseudo supervision loss to implement model training, including the following steps: S401, the unlabeled EEG data is passed through two graph convolutional networks to obtain two confidence vectors; S402. Use the argmax function to obtain the class with the highest confidence and generate the unique hot vectors of the two networks.
[0029] S403. Define pseudo-supervision loss, which is bidirectional and consists of the cross entropy loss between the one-hot label vector obtained by one network and the classification confidence vector of another network; S404. The total loss function consists of supervised loss and pseudo-supervised loss. Through the back-propagation algorithm, the model will gradually update the weights and biases of the network to minimize the total loss function and converge to optimize the model and achieve its best performance.
[0030] In step S403, the formula of the pseudo-supervision loss function is as follows: in, is the confidence vector of the classification The calculated one-hot label vector.
[0031] In step S404, the total loss function of the model is in is the weight, that is, the proportion of pseudo-supervision loss to the overall loss, and the Sigmoid function is used for real-time adjustment.
[0032] Those skilled in the art will appreciate that the units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.
[0033] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.
[0034] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A semi-supervised motor imagery EEG classification algorithm based on mutual supervision of dual-graph convolutional networks, comprising the following steps: S1, preprocessing the original EEG signal, performing bandpass filtering, time window truncation and baseline correction in sequence; S2. Construct an adjacency matrix based on the correlation coefficient between electrodes, extract specific EEG features to construct a feature matrix, and build a graph convolution model with two independently initialized parameters; S3. Both models use labeled data for supervised learning and use cross entropy loss as the supervision loss; S4. The two models use unlabeled data to generate pseudo labels as the supervision signal of the other model, calculate the bidirectional cross entropy loss as the pseudo supervision loss, and implement model training by combining the supervision loss and the pseudo supervision loss.
2. The semi-supervised motor imagery EEG classification algorithm based on dual-graph convolutional network mutual supervision according to claim 1 is characterized by: In step S1, the raw EEG signal is preprocessed in sequence, including Butterworth bandpass filtering, trial segmentation, time window truncation and baseline correction.
3. The semi-supervised motor imagery EEG classification algorithm based on dual-graph convolutional network mutual supervision according to claim 2 is characterized by: In step S2, an adjacency matrix is constructed based on the inter-electrode correlation coefficient, specific EEG features are extracted to construct a feature matrix, and a graph convolution model with two independently initialized parameters is constructed, including the following steps: S201, for each trial data after preprocessing, calculate the Pearson correlation coefficient between channels based on the sampling points; S202, set a threshold, if the absolute value of the correlation coefficient is greater than the threshold, it is recorded as 1, otherwise it is recorded as 0, and generate and output the channel adjacency matrix; S203, extracting specific EEG features of the EEG signal of each channel as node features, and constructing a feature matrix; S204. Construct a graph convolutional network model with two independently initialized parameters based on the adjacency matrix and the feature matrix.
4. The semi-supervised motor imagery EEG classification algorithm based on dual-graph convolutional network mutual supervision according to claim 3, characterized in that: In step S3, the two models use the labeled data for supervised learning, using cross entropy loss as the supervised loss, including the following steps: S301: Input the pre-processed labeled EEG data into two independently initialized graph convolutional networks, and output the corresponding predicted classification confidence vectors; S302: After outputting the classification confidence vector, a loss function is calculated based on the label of the corresponding EEG data for supervision, and cross entropy loss is selected as the loss function; S303. Add the cross entropy losses of the two models to obtain the supervised loss, and set the weight to control the proportion of the pseudo-supervised loss to the overall loss.
5. The semi-supervised motor imagery EEG classification algorithm based on mutual supervision of dual-graph convolutional networks according to claim 4, characterized in that: In step S4, the two models use unlabeled data to generate pseudo labels as the supervision signal of the other model, calculate the bidirectional cross entropy loss as the pseudo supervision loss, and integrate the supervision loss and pseudo supervision loss to implement model training, including the following steps: S401, the unlabeled EEG data is passed through two graph convolutional networks to obtain two confidence vectors; S402. Use the argmax function to obtain the class with the highest confidence and generate the unique hot vectors of the two networks.
6. S403. Define pseudo-supervision loss. This loss is bidirectional and consists of the cross entropy loss between the one-hot label vector obtained by one network and the classification confidence vector of the other network. S404. The total loss function consists of supervised loss and pseudo-supervised loss. Through the back-propagation algorithm, the model will gradually update the weights and biases of the network to minimize the total loss function and converge to optimize the model and achieve its best performance.
7. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the semi-supervised motor imagery EEG classification algorithm based on mutual supervision of a dual-graph convolutional network according to any one of claims 1 to 6 is implemented.