Electroencephalogram signal detection method based on separation time correlation and auditory disturbance

Abstract

An electroencephalogram signal detection method based on separation time correlation and auditory disturbance comprises the steps that unmarked electroencephalogram data is collected, pre-training is carried out through a multi-view frequency spectrum-space-time mask automatic encoder model, frequency spectrum-space-time features are obtained, and then fine adjustment is carried out through a linear classifier to obtain the specific state of the electroencephalogram data. According to the invention, time interference and hearing interference can be effectively eliminated, and high-precision meditation state classification is realized. Implementation of the technical scheme is described in detail, and innovation of experimental design and a data acquisition mode and a self-supervised learning algorithm matched with the experimental design and the data acquisition mode are emphatically elaborated.

Description

Technical Field

[0001] The present invention relates to a technology in the field of electroencephalogram (EEG) signal processing, in particular to an EEG signal detection method based on separation of time association and auditory interference. Background Art

[0002] Meditation state monitoring is an important research direction in the field of biomedical informatics in recent years. Traditional methods mainly rely on the analysis of electroencephalogram (EEG) data. Existing technologies usually use methods based on feature engineering and machine learning to distinguish between meditation and non-meditation states. These methods mainly rely on pre-labeled data sets for training and verification. However, there are two significant challenges in traditional meditation state monitoring methods: one is time interference, that is, the EEG data collected at different time points are highly correlated, which may cause the model to overfit to the time characteristics rather than the meditation state itself; the other is auditory interference. Audio-guided meditation will affect the characteristics of EEG signals, making it difficult to distinguish the difference between EEG signals with and without audio guidance. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention proposes an EEG signal detection method based on separating time association and auditory interference, which can effectively eliminate time interference and auditory interference and achieve high-precision meditation state classification. The implementation of this technical solution is described in detail below, focusing on the innovativeness of the experimental design and data collection method, as well as the self-supervised learning algorithm that is compatible with it.

[0004] The present invention is achieved through the following technical solutions:

[0005] The present invention relates to an EEG signal detection method based on separating time association and auditory interference. The method collects unlabeled EEG data, pre-trains a multi-view spectrum-space-time mask autoencoder model to obtain spectrum-space-time features, and then fine-tunes a linear classifier to obtain the specific state of the EEG data.

[0006] The present invention relates to a system for implementing the above method, comprising: a state labeling unit, an EEG acquisition unit and a meditation state analysis unit, wherein: the state labeling unit plays meditation prompts according to a preset meditation prompt program when alternating, and adds additional data to the collected EEG data to clarify the environment and conditions for collecting EEG at that moment; the EEG acquisition unit is composed of an EEG cap and a signal amplifier and is associated with the state labeling unit, and collects EEG data in real time according to the program shown in the state labeling unit, and collects a section of EEG data containing 64 channels; the meditation state analysis unit performs data analysis based on the collected EEG data, performs preprocessing such as noise removal, and then adopts a deep learning method including a self-supervised learning mechanism to obtain the result of real-time detection of meditation state of EEG state. Technical Effects

[0007] The present invention introduces multiple data collections, and performs training and testing between different collections; by including meditation experiments with audio guidance and without audio guidance, the differences in EEG signals under different conditions are analyzed, and the model can better distinguish different meditation states, thereby improving the robustness of the model under various experimental conditions. Through self-supervised learning technology, the model can be pre-trained on a large amount of unlabeled EEG data, so as to learn more extensive and more general features. Compared with the prior art, the present invention can effectively reduce the impact of time correlation on EEG signal classification, thereby improving the accuracy of meditation state classification. It reduces the dependence on labeled data and enhances the adaptability of the model under different experimental conditions and between different individuals. The emotional EEG dataset SEED, which is unrelated to the meditation classification task, is used for pre-training to improve the performance on the meditation detection task. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a data collection flow chart of the present invention;

[0009] Figure 2 This is a flow chart of EEG data processing of the present invention;

[0010] Figure 3 It is a schematic diagram of the effect of the embodiment. DETAILED DESCRIPTION

[0011] like Figure 1 and Figure 2 As shown, this embodiment relates to an EEG signal detection method based on separation of time association and auditory interference, comprising:

[0012] Step 1. Multiple data collection: Each subject participated in three independent experiments, with an interval of 14 days between each experiment. In each collection, the subject performed audio-guided meditation, non-audio-guided meditation, and resting state experiments. The time interval between each collection can effectively reduce the temporal correlation of EEG signals in different states, avoid overfitting of the model to temporal features, and thus improve the accuracy of classification.

[0013] In the data collection, during each collection process, the subjects alternately perform meditation with audio guidance and meditation without audio guidance to explore the impact of auditory interference on EEG signals in the meditation state. Specifically, each collection includes 5 minutes of resting state, about 20 minutes of meditation state, and another 5 minutes of resting state. The meditation stage includes alternating audio guidance with prompts and audio guidance without prompts. The subjects enter the meditation state through audio instructions, and in the stage without audio guidance, the subjects are required to continue meditating in a quiet environment. Through this alternating design, EEG signals under different conditions can be collected separately, so as to more accurately analyze the impact of auditory interference on the meditation state.

[0014] Step 2: Pre-train the unlabeled EEG data set obtained in step 1 through the multi-view spectrum-space-time masked autoencoder model (MV-SSTMA) to obtain general spectrum-space-time features. Specifically, the multi-view spectrum-space-time masked autoencoder (MV-SSTMA) algorithm adopts a CNN-Transformer-based architecture. First, the unlabeled EEG data is randomly masked through the self-supervised pre-training stage, and then the spectrum, space and time features are extracted using multi-scale causal convolutional layers and multi-head self-attention mechanisms. In the pre-training stage, the algorithm uses unlabeled EEG data for general feature learning. Through random masking, the EEG data is divided into visible subsets and masked subsets. The model captures spectrum, space and time information by reconstructing the masked data, and the extracted features have stronger generalization ability. The key operations include projecting the input into the D-dimensional spectrum embedding space and dividing the EEG channels as "patches" to retain the spatial position information. Among them, the multi-scale causal convolutional layer is used to process the temporal features of the EEG signal. Causal convolution ensures that the output at time t depends only on the input at time t and before, preventing future information leakage. The multi-scale design (different kernel sizes) enables the model to capture different ranges of time steps to adapt to EEG signals. The multi-head self-attention layer is used to capture the spatial dependencies between EEG channels, which helps to extract the whole brain information of multi-channel EEG signals. The convolutional layer and self-attention outputs of each scale (small, medium, and large) are aggregated to obtain complete spatial embedding information.

[0015] The multi-view spectral-spatial-temporal masked autoencoder model (MV-SSTMA) is implemented using but not limited to the technology described in "A Multi-view Spectral-Spatial-Temporal Masked Autoencoder for Decoding Emotions with Self-supervised Learning" by Rui Li et al. (https: / / doi.org / 10.1145 / 3503161.3548243).

[0016] The parameters of the multi-view spectrum-space-time mask autoencoder model include:

[0017] Spectral embedding dimension D: used to map input data to a specific spectral feature space.

[0018] Convolution kernel size k s ,k m ,k l : Corresponding to small, medium and large scale convolution layers, used for feature extraction in different time ranges.

[0019] The number of self-attention heads: determines the number of heads in the spatial attention layer, thus affecting the capture of dependencies between EEG channels.

[0020] Batch Normalization: Used to stabilize the model training process and prevent gradient explosion or disappearance.

[0021] Position encoding: implemented in the spatial position encoding layer to preserve the spatial position information of different EEG channels.

[0022] Step 3: Fine-tune the spectrum-space-time features obtained in step 2, specifically: input the extracted spectrum-space-time features into the linear classifier, and finally output whether the subject is in a meditative state or a resting state. After pre-training is completed, fine-tune on a labeled target dataset to adapt the model to specific tasks, such as the classification of meditation and resting states. Through training with a small learning rate and selectively freezing some layers or unfreezing them all, the model is gradually optimized on a labeled dataset so that it learns the features required for specific tasks. The model achieves high-precision meditation and resting state classification by fine-tuning the labeled meditation dataset, combining feature extractors and linear classifiers. Through these designs, MV-SSTMA has significant advantages in denoising, improving EEG classification accuracy, and enhancing model generalization capabilities.

[0023] After specific practical experiments, in the experimental environment settings of cross-subjects and the same subject, the self-supervised learning MV-SSTMA model is the core, the pre-training data set is the unlabeled SEED emotional EEG data set, and the above method is fine-tuned on the meditation and resting state labeled data. Compared with the existing technology, this method reduces the interference of time correlation on EEG signals by introducing multiple acquisitions in data acquisition and setting a 14-day interval. At the same time, the meditation experiment design with and without audio guidance is alternately carried out, which effectively eliminates the influence of auditory interference on meditation state monitoring. In terms of algorithm, the multi-view spectral-spatial-temporal masked autoencoder (MV-SSTMA) model is used. After self-supervised learning pre-training and fine-tuning on labeled data, the model can more accurately distinguish between meditation and resting states. The specific experimental results show that in the experiment of the same subject, the F1 score reached 0.7776 and the accuracy was 0.7863; in the experiment across subjects, the F1 score reached 0.7239. Compared with the traditional method, the classification performance has been significantly improved in terms of F1 score, accuracy, precision and recall rate, verifying the superiority of this method in eliminating interference factors and improving classification accuracy.

[0024] In terms of method improvement, this method improves the data collection design: introducing an alternating design with audio guidance and without audio guidance in multiple collections to avoid the influence of time and auditory interference on the classification of meditation states. The principle is to reduce the influence of time correlation on the classification results by introducing an interval period through multiple collections. The alternating use of experimental designs with audio guidance and without audio guidance can more accurately learn the characteristics of EEG signals under different meditation conditions, thereby avoiding auditory interference. The specific operation is that each subject participates in three independent experiments, each experiment is 14 days apart, and each collection includes experiments in the resting state and the meditation state. The meditation state experiment further alternates between audio-guided and non-audio-guided meditations to collect corresponding EEG signal data.

[0025] This method improves on the meditation EEG recognition algorithm, introduces a self-supervised learning algorithm, and uses a multi-view spectrum-space-time masked autoencoder (MV-SSTMA) to improve the generalization ability of the model through self-supervised pre-training. The principle is to perform self-supervised pre-training on unlabeled data, so that the model can autonomously learn more general spectrum-space-time features. This method reduces the dependence on labeled data, making the model widely adaptable to different experimental conditions and individuals. The specific operation is to use the MV-SSTMA model to pre-train on an unlabeled emotional EEG dataset (such as the SEED dataset). This process randomly masks part of the EEG signal and reconstructs it to learn the universal features of spectrum, space and time. The pre-trained model is fine-tuned and the classification performance of the model is optimized using labeled data so that it can adapt to the classification tasks of meditation and resting state.

[0026] Based on the above improvements, this method achieves: 1) Improved classification accuracy: Through multiple acquisitions and interval design, the impact of time and auditory interference is significantly reduced, allowing the model to more accurately distinguish between meditation and resting states. 2) Enhanced generalization ability: Self-supervised pre-training enables the model to acquire general feature extraction capabilities, improving its adaptability under different experimental conditions and individual differences. 3) Enhanced robustness: The system's evaluation indicators such as accuracy and F1 score in meditation state monitoring are significantly better than traditional technical methods, and it shows higher robustness across individuals and under different meditation conditions. The results are as follows: Figure 3 shown.

[0027] In summary, the present invention has achieved innovation in EEG monitoring of meditation state. First, by improving the data collection method, multiple collections are introduced and set at a 14-day interval, which reduces the interference of time correlation on EEG signals; at the same time, the alternating audio-guided and non-audio-guided meditation experiment design is adopted to eliminate the influence of auditory interference. In addition, in terms of algorithm, a multi-view spectral-spatial-temporal masked autoencoder (MV-SSTMA) model is proposed. The spectral-spatial-temporal features are extracted from unlabeled EEG data through self-supervised learning, and then fine-tuned on labeled meditation data, which significantly improves the classification performance and the generalization ability of the model, so that the model can more accurately distinguish between meditation and resting state.

[0028] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A method for detecting electroencephalogram signals based on separating temporal association and auditory interference, characterized in that: By collecting unlabeled EEG data, pre-training the multi-view spectrum-space-time mask autoencoder model to obtain the spectrum-space-time features, and then fine-tuning the linear classifier to obtain the specific state of the EEG data.

2. The method for detecting EEG signals based on separating temporal correlation and auditory interference according to claim 1, wherein: include: Step 1: Multiple data collection: Each subject participated in three independent experiments, with a 14-day interval between each experiment; each collection was conducted with audio-guided meditation, without audio-guided meditation, and resting state experiments; Step 2: Pre-training the unlabeled EEG dataset obtained in step 1 through a multi-view spectrum-space-time mask autoencoder model (MV-SSTMA) to obtain a general spectrum-space-time feature; Step 3: Fine-tune the spectrum-space-time features obtained in step 2. Specifically, input the extracted spectrum-space-time features into a linear classifier, and finally output whether the subject is in a meditative state or a resting state.

3. The method for detecting EEG signals based on separating time association and auditory interference according to claim 2, characterized in that: The time interval between each acquisition can effectively reduce the temporal correlation of EEG signals under different states, avoid overfitting of the model to temporal features, and thus improve the accuracy of classification.

4. The method for detecting EEG signals based on separating time association and auditory interference according to claim 2, characterized in that: In the data collection, during each collection process, the subjects alternated between meditation with audio guidance and meditation without audio guidance to explore the impact of auditory interference on the EEG signals in the meditation state. Specifically, each collection included 5 minutes of resting state, about 20 minutes of meditation state, and another 5 minutes of resting state.

5. The method for detecting EEG signals based on separating time association and auditory interference according to claim 4 is characterized in that: By alternating between audio guidance with and without prompts, the subjects were guided into a meditative state through audio instructions, whereas in the phase without audio guidance, the subjects were asked to continue meditating in a quiet environment.

6. The method for detecting EEG signals based on separating time association and auditory interference according to claim 2, characterized in that: The step 2 is specifically as follows: the multi-view spectral-spatial-temporal masked autoencoder (MV-SSTMA) algorithm adopts a CNN-Transformer-based architecture, firstly randomly masks the unlabeled EEG data through a self-supervised pre-training phase, and then uses a multi-scale causal convolutional layer and a multi-head self-attention mechanism to extract spectral, spatial and temporal features; Unlabeled EEG data are used for general feature learning in the self-supervised pre-training stage: the EEG data are divided into a visible subset and a masked subset through random masking, and the spectral, spatial and temporal information is captured by reconstructing the masked data. The extracted features have stronger generalization ability.

7. The method for detecting EEG signals based on separating time association and auditory interference according to claim 2, characterized in that: The fine-tuning means that through training with a small learning rate and selectively freezing some layers or unfreezing all layers, the model is gradually optimized on a labeled dataset so that it learns the features required for a specific task.

8. An electroencephalogram signal detection system based on separation of time correlation and auditory interference for implementing any of the methods described in claims 1-7, characterized in that: include: The state labeling unit, the EEG acquisition unit and the meditation state analysis unit, wherein: the state labeling unit plays the meditation prompts according to the preset meditation prompt program, and adds additional data to the collected EEG data to clarify the environment and conditions for collecting EEG at that moment; the EEG acquisition unit is composed of an EEG cap and a signal amplifier and is associated with the state labeling unit, and collects EEG data in real time according to the program shown in the state labeling unit, and collects a section of EEG data containing 64 channels; the meditation state analysis unit performs data analysis on the collected EEG data, performs preprocessing such as noise removal, and then adopts a deep learning method including a self-supervised learning mechanism to obtain the result of real-time detection of meditation state of EEG state.

Images