A network architecture search method and system for electroencephalogram signal processing

Abstract

The application provides a network architecture search method and system for electroencephalogram signal processing, and realizes effective migration from high-channel EEG data to low-channel EEG data. In particular, for children with autism, it is difficult for them to cooperate for a long time to collect EEG signals, which may result in incomplete data. The application searches for the optimal neural network architecture by using the multi-agent deep deterministic policy gradient algorithm to search for the neural network architecture, so that a high classification accuracy is maintained in the process of migrating from high-channel data to low-channel data.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram (EEG) signal processing, and in particular to a network architecture search method and system for EEG signal processing. Background Art

[0002] Autism Spectrum Disorder (ASD) is a broad neurodevelopmental disorder characterized by core symptoms such as social communication impairment, language communication impairment, and repetitive behaviors. Due to the large number of individuals with ASD (hereafter referred to as autism), early screening and accurate diagnosis are crucial for patient rehabilitation and intervention.

[0003] However, the current assessment and diagnosis of autism mainly relies on expert interviews and behavioral scales. These methods are usually highly subjective and inefficient, making it difficult to meet the needs of large-scale screening. Electroencephalography (EEG), as a means of recording the electrophysiological activity of the cerebral cortex, has the advantages of being non-invasive, easy to collect, and having high temporal resolution. Through feature engineering and traditional machine learning techniques, automatic identification of people with autism can be achieved. Its technical path includes: (1) Preprocessing. Filtering is used to remove artifacts and retain specific frequency band signals to obtain more effective features; (2) Feature extraction. Information is extracted from the signal through analysis methods such as time, frequency, time-frequency, and entropy to form a feature set; (3) Feature selection. Filtering or wrapping feature selection methods are used to select effective feature subsets; (4) Autism identification. By training a classifier or a combination of classifiers, the classifier parameters are optimized using labeled data, and the optimized model is used to identify unlabeled samples. The advantage of this solution is that the training complexity is low and it does not rely on large-scale data and high-performance computing; but its disadvantage is that it requires high prior knowledge and professional skills.

[0004] With the advent of big data and increased computing power, deep learning technologies such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have achieved remarkable success in computer vision and natural language processing, even surpassing human performance in certain tasks. Consequently, applying deep learning techniques to EEG signal processing to achieve high-precision autism diagnosis has become a research hotspot in the field of brain health. However, high-density EEG acquisition equipment is expensive and places an increased burden on subjects, especially children with autism, significantly increasing the cost and difficulty of EEG signal acquisition. Furthermore, some EEG channels have no significant effect on autism diagnosis and can even reduce the efficiency and accuracy of model training.

[0005] Therefore, for the application of deep learning technology in the field of EEG signal processing and classification, it is necessary to propose a method with wider adaptability and higher accuracy. Summary of the Invention

[0006] The present invention provides a network architecture search method and system for EEG signal processing, which are used to solve the defects of the existing technology that the network architecture used when deep learning technology is applied in the field of EEG signal processing classification lacks universality and high accuracy.

[0007] In a first aspect, the present invention provides a network architecture search method for EEG signal processing, comprising:

[0008] Collecting EEG signal data of the target subject;

[0009] Preprocessing the EEG signal data to obtain preprocessed EEG signal data;

[0010] Using the preprocessed EEG signal data, a multi-agent deep deterministic policy gradient algorithm is used to perform architecture search training on a neural network model to obtain an original EEG signal acquisition model;

[0011] The original EEG signal acquisition model is evaluated and optimized to determine the optimal EEG signal acquisition model.

[0012] According to a network architecture search method for EEG signal processing provided by the present invention, EEG signal data of a target subject is collected, comprising:

[0013] Collect high-density 62-channel 1000Hz resting-state EEG signal data.

[0014] According to a network architecture search method for EEG signal processing provided by the present invention, the EEG signal data is preprocessed to obtain preprocessed EEG signal data, including:

[0015] Downsampling, filtering, and artifact removal are sequentially performed on the EEG signal data to obtain filtered EEG signal data;

[0016] Segmenting the filtered EEG signal data into a plurality of EEG segments using a fixed time window;

[0017] Performing discrete Fourier transform and bandpass filtering on the multiple EEG segments to obtain EEG frequency domain signals and retaining EEG time domain signals;

[0018] The EEG frequency domain signals and the EEG time domain signals are randomly mixed according to the subjects and divided into a training set, a validation set and a test set.

[0019] According to a network architecture search method for EEG signal processing provided by the present invention, the preprocessed EEG signal data is used to perform architecture search training on a neural network model using a multi-agent deep deterministic policy gradient algorithm to obtain an original EEG signal acquisition model, including:

[0020] Determining that the neural network model includes an input layer, multiple hidden layers, and an output layer to form a multi-agent decision-making architecture;

[0021] Searching for the neural network model based on the multi-agent deep deterministic policy gradient algorithm, and searching for the optimal network architecture for high- and low-channel EEG data migration through collaboration and policy optimization between agents;

[0022] Initial training is performed on preset high-channel EEG data to obtain training data, and the training data is transferred to preset low-channel EEG data for learning to obtain the original EEG signal acquisition model.

[0023] According to a network architecture search method for EEG signal processing provided by the present invention, it is determined that the neural network model includes an input layer, multiple hidden layers and an output layer, forming a multi-agent decision architecture, including:

[0024] The input layer receives EEG data features;

[0025] The multiple hidden layers use convolution layers and pooling layers of different sizes to extract time-frequency features from the EEG data features;

[0026] The output layer generates classification results using the Softmax function.

[0027] According to a network architecture search method for EEG signal processing provided by the present invention, the neural network model is searched based on the multi-agent deep deterministic policy gradient algorithm. Through collaboration and policy optimization between agents, the optimal network architecture for high- and low-channel EEG data migration is searched, including:

[0028] Determine that the intelligent agent includes a policy network and an evaluation network, wherein the input of the policy network is the intelligent agent observation data, and the output is the action vector adapted to the action space; the input of the evaluation network is the intelligent agent observation and action, and the output is the action quality of the intelligent agent;

[0029] The evaluation network evaluates the agent's action quality Q value by minimizing the loss function and updates the evaluation network;

[0030] The policy network generates a Q-value-maximizing action by calculating the policy gradient and updates the policy network;

[0031] The target network is added to the strategy network and the evaluation network respectively, and soft updates are performed to obtain the optimal network architecture for high- and low-channel EEG data migration.

[0032] According to a network architecture search method for EEG signal processing provided by the present invention, the original EEG signal acquisition model is evaluated and optimized to determine the optimal EEG signal acquisition model, including:

[0033] Calculate the classification accuracy of the original EEG signal acquisition model on the preset low-channel EEG data after migration on the test set, compare the EEG data under different channels, and verify the effectiveness of the multi-agent deep deterministic policy gradient algorithm;

[0034] Multiple rounds of cross-validation and segmented experiments were used to test the stability of the original EEG signal acquisition model under different input conditions;

[0035] Perform hyperparameter optimization based on classification accuracy and model stability evaluation results;

[0036] The selected optimal network architecture is transplanted to a low sampling rate EEG device, and the optimal network architecture is subjected to model lightweight processing to obtain the optimal EEG signal acquisition model.

[0037] In a second aspect, the present invention further provides a network architecture search system for EEG signal processing, comprising:

[0038] An acquisition module, used to acquire EEG signal data of a target subject;

[0039] A preprocessing module, used to preprocess the EEG signal data to obtain preprocessed EEG signal data;

[0040] A training module is used to use the pre-processed EEG signal data to perform architecture search training on the neural network model using a multi-agent deep deterministic policy gradient algorithm to obtain an original EEG signal acquisition model;

[0041] The optimization module is used to evaluate and optimize the original EEG signal acquisition model to determine the optimal EEG signal acquisition model.

[0042] In a third aspect, the present invention also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the network architecture search method for EEG signal processing as described above is implemented.

[0043] In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a network architecture search method for EEG signal processing as described in any one of the above.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This invention effectively improves the efficiency of network architecture search through collaborative search and strategy optimization between intelligent agents, making the model more adaptable to the complex distribution of EEG data features. By optimizing model parameters through transfer learning, the model can still maintain high classification accuracy on low-channel data, improving the adaptability and versatility of the algorithm.

[0046] After the model is transplanted to a low-sampling-rate device, the present invention can still maintain efficient classification performance and real-time response capabilities, ensuring that EEG signals can be quickly processed and classified in practical applications to meet the needs of real-time analysis.

[0047] This model boasts high device compatibility and is suitable for EEG devices with varying sampling rates and channel counts. The provided technical support and deployment documentation simplify the model's integration and application, providing a technical foundation for the large-scale deployment of low-cost EEG devices. Through lightweight and optimized design, the model's hardware dependency is effectively reduced, making it suitable for low-cost EEG devices and providing a cost-effective solution for resource-constrained devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 1 is a flow chart of a network architecture search method for EEG signal processing provided by the present invention;

[0050] Figure 2 It is the overall structural principle diagram provided by the present invention;

[0051] Figure 3 It is a structural diagram of the network architecture search system for EEG signal processing provided by the present invention. DETAILED DESCRIPTION

[0052] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0053] To address the challenges of existing technologies, this paper proposes an adaptive neural network architecture search method. This method aims to reduce the patient cooperation required for EEG acquisition by effectively migrating high-channel EEG data to low-channel EEG data. This method is particularly suitable for the collection and analysis of EEG data from children with autism. This method utilizes the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm to search for neural network architectures, identify the optimal architecture, and achieve migration between high- and low-channel data, ensuring high classification accuracy and reliability.

[0054] Figure 1 FIG is a flow chart of a network architecture search method for EEG signal processing provided by an embodiment of the present invention, such as Figure 1 Shown, including:

[0055] Step 100: Collecting EEG signal data of a target subject;

[0056] Step 200: preprocessing the EEG signal data to obtain preprocessed EEG signal data;

[0057] Step 300: Using the pre-processed EEG signal data, a multi-agent deep deterministic policy gradient algorithm is used to perform architecture search training on a neural network model to obtain an original EEG signal acquisition model;

[0058] Step 400: Evaluate and optimize the original EEG signal acquisition model to determine the optimal EEG signal acquisition model.

[0059] Specifically, the embodiment of the present invention proposes an adaptive neural network architecture search method for migrating high-channel EEG data to low-channel EEG data. The core of this method is to automatically search for neural network architectures using a multi-agent deep deterministic policy gradient algorithm to effectively migrate between different channel data and maintain a high classification accuracy during the migration process. The specific implementation steps include:

[0060] Step 1: Data acquisition. This work uses high-density 62-channel 1000Hz resting-state EEG signal data as the data source for training the model.

[0061] Step 2: Data preprocessing. The data collected by the EEG acquisition device is a 3-5 minute raw signal. The processing required in this step is divided into the following parts: (1) Using downsampling, bandpass filtering and artifact removal algorithms to reduce the impact of noise such as electrooculogram and electromyography on subsequent analysis; (2) Using a fixed time window to divide the raw signal into multiple time segments as the minimum unit for autism discrimination, individual discrimination results can then be generated by voting; (3) Discrete Fourier transform processing is performed on each EEG segment to make the frequency domain features more prominent; (4) The data is divided into training set, validation set and test set by randomly selecting different people.

[0062] Step 3: Model construction and training. (1) Network architecture initialization: Based on the multi-layer structure of the neural network, the initial model includes an input layer, multiple hidden layers, and an output layer. (2) MADDPG algorithm training: The architecture of the model is searched using a multi-agent deep deterministic policy gradient algorithm. This algorithm searches for the optimal network architecture suitable for the migration of high- and low-channel EEG data through collaboration and policy optimization between agents. (3) Transfer learning: Transfer training is performed using high-channel data as input to adapt to the feature distribution of low-channel data and optimize model parameters.

[0063] Step 4: Evaluate and optimize the classification accuracy of the model. Specific evaluation indicators include:

[0064] Classification accuracy: Evaluate the classification accuracy after migration of high- and low-channel EEG data.

[0065] Model stability: Ensure the stability and adaptability of the model during the migration between different data channels.

[0066] Step 5: Porting and Deployment. The optimal neural network architecture ultimately selected is suitable for classification tasks involving low-channel EEG data. This model can be deployed on low-sampling-rate EEG devices, further reducing equipment acquisition costs.

[0067] The present invention uses a multi-agent deep deterministic policy gradient algorithm to search for a neural network architecture and find the optimal neural network architecture, so as to maintain a high classification accuracy in the process of migrating from high-channel data to low-channel data.

[0068] Based on the above embodiment, step 1 includes:

[0069] 1.1 EEG signal acquisition process

[0070] EEG signal acquisition is usually carried out in a professional EEG acquisition room. Sensors are attached to the surface of the subject's scalp to record the brain's electrophysiological activity data in resting or task states, and the real-time collected signals are transmitted to a computer for storage.

[0071] 1.2 Preliminary Preparation and Environment Setup

[0072] To ensure the quality of signal acquisition, subjects were asked to remove items that might affect signal acquisition, such as glasses and earrings. Furthermore, during the acquisition process, subjects were given a headgear made of special materials to ensure that the electrodes fit tightly against the scalp and reduce external interference.

[0073] 1.3 Electrode arrangement and conductive treatment

[0074] Electrodes corresponding to the 62 channels, a reference electrode, and a ground wire were threaded through the headgear, ensuring they adhered tightly to the subject's scalp. To further reduce resistance and improve signal acquisition stability, conductive paste was applied to the contact points.

[0075] 1.4 Signal Acquisition and Recording

[0076] The EEG data collection process requires the subject to maintain steady breathing, sit quietly, and remain still with their eyes open for 3-5 minutes. The collected EEG signals will be stored along with the subject's basic information for subsequent analysis.

[0077] Based on the above embodiment, step 2 includes:

[0078] 2.1 Data Downsampling and Filtering

[0079] First, the collected EEG data is downsampled to 256Hz to enhance its noise immunity and reduce storage and computing resource consumption. Next, bandpass filtering (0.5-45Hz) is used to remove interference signals outside the EEG frequency band, effectively preserving the main components of EEG activity. Furthermore, an artifact removal algorithm is used to remove artifacts such as electrooculography and electromyography from the data to ensure data purity.

[0080] 2.2 Data fragmentation processing

[0081] Each subject's complete EEG recording was segmented into 4-second segments, which served as the basic unit for subsequent processing. During the discrimination process, the voting results of multiple segments were used to generate the final classification judgment for the subject, thereby improving the robustness of individual discrimination.

[0082] 2.3 Time Domain and Frequency Domain Feature Extraction

[0083] Each EEG segment is subjected to a discrete Fourier transform (DFT) to convert the data from the time domain to the frequency domain, making the frequency domain features more prominent while preserving the time domain features. After DFT processing, a bandpass filter is applied again to retain only the 0.5-45Hz frequency band.

[0084] 2.4 Dataset Division

[0085] All samples were randomly mixed by person and divided into training, validation, and test sets (with a ratio of 3:1:1). The training set was used for model training, the validation set was used for parameter optimization, and the test set was used to evaluate the final model performance.

[0086] Based on the above embodiment, step 3 includes:

[0087] 3.1 Network Architecture Initialization

[0088] Based on the multi-layered structure of a neural network, the initial model design includes an input layer, multiple hidden layers, and an output layer, forming a multi-agent decision-making architecture. This model architecture is suitable for migrating high- and low-channel EEG data. The input layer receives EEG data features, the hidden layers use convolutional layers and pooling layers of varying sizes to extract time-frequency features from the data, and the output layer uses a softmax function to generate classification results, completing multi-classification tasks.

[0089] 3.2 Multi-agent Deep Deterministic Policy Gradient Algorithm Training

[0090] The MADDPG algorithm is used to search for model architectures and conduct collaborative training across multiple agents to achieve the optimal strategy for the EEG data migration task. The MADDPG algorithm specifically includes an Actor network and a Critic network, which are used for action generation and value assessment, respectively.

[0091] 3.2.1 Agent Design

[0092] Policy Network (Actor Network): The MADDPG algorithm assigns a separate Actor Network to each agent. The Actor Network takes input from the agent's observation space and outputs actions within its action space. Each Actor Network takes input from the agent's observation data and outputs an action vector adapted to the action space.

[0093] Evaluation Network (Critic Network): The Critic Network is responsible for evaluating the quality of each agent's actions. In a multi-agent environment, the input to each Critic Network includes the observations and actions of all agents, allowing the strategies of other agents to be considered during the evaluation process, improving the effectiveness of collaborative learning.

[0094] 3.2.2 Evaluation Network Update

[0095] The evaluation network evaluates the quality of the agent's actions by minimizing the loss function. The loss function of the Critic network is:

[0096]

[0097] in, Represents the Q value output by the Critic network, and the parameter is ; For TD target, For the current reward, is the discount factor, is the next state, and The actions of all agents in the next state generated by the Actor network. The Critic network optimizes the parameters by minimizing the loss function. .

[0098] 3.3 Policy Network Update

[0099] The goal of the policy network is to generate the optimal policy so that the agent's actions can maximize the Q value. The policy gradient calculation formula of the actor network is:

[0100]

[0101] in, Represents the gradient of the Actor network, and the parameters are ,and is the Q value output by the Critic network, Represents the policy function of the Actor network, which calculates the agent's action under the current observation. Through backpropagation, the Actor network is updated to generate actions that maximize the Q value.

[0102] 3.4 Target Network Soft Update

[0103] In order to maintain the stability of the network, a target network is added to the Actor and Critic networks, and soft updates are performed. The soft update formula is:

[0104]

[0105]

[0106] Among them, τ is the soft update coefficient, which is usually small (such as 0.001) and is used to gradually update the parameters of the target network to keep the training stable.

[0107] 3.5 Transfer Learning

[0108] Initial training is performed on high-channel EEG data, and transfer learning is used to apply the knowledge gained from this training to low-channel EEG data, reducing training time. This step retains the data characteristics of high-channel training. Retraining on low-channel data adjusts model parameters, allowing the model to adapt to the characteristic distribution of low-channel data, thereby improving classification accuracy and stability.

[0109] Based on the above embodiment, step 4 includes:

[0110] 4.1 Classification Accuracy Evaluation

[0111] The model was evaluated by calculating the classification accuracy of the model on low-channel EEG data after transfer on the test set, and EEG data under different channels were used for comparison to verify the effectiveness of the MADDPG algorithm in transfer learning.

[0112] 4.2 Model stability evaluation

[0113] Multiple rounds of cross-validation and segmented experiments are used to test the stability of the model under different input conditions, including the training convergence speed and performance differences under different channel numbers, to ensure the adaptability and high-performance output of the migration process.

[0114] 4.3 Optimization and Adjustment

[0115] Based on the classification accuracy and model stability evaluation results, hyperparameter optimization (such as learning rate, batch size, layer selection, and dropout rate adjustment) is performed to improve the model's ability to handle complex features while ensuring model adaptability and stability.

[0116] Based on the above embodiment, step 5 includes:

[0117] The optimal network architecture was ultimately selected and applied to the classification task of low-channel EEG data. To further reduce equipment acquisition costs, the model was ported to EEG devices with lower sampling rates and lightweighted, such as reducing the number of convolutional and pooling layers and the number of model parameters. During deployment, the model was verified to maintain efficient and accurate classification performance on low-channel EEG data, meeting real-time requirements and device computational limitations, providing technical support for the application of low-cost EEG devices.

[0118] The embodiments of the present invention are described below with specific examples. Figure 2 Shown, including:

[0119] The first step is data collection. EEG signal collection is typically performed in a dedicated EEG acquisition room. To ensure signal quality, the environment should be quiet and the power supply voltage stable. Avoid selecting subjects with thick or coarse hair, and ask them to clean their heads and remove any headwear before donning the headgear. Electrodes are accurately applied to the subject's scalp at the corresponding locations, and a special conductive paste is applied to the contact points to reduce resistivity, ensuring that each resistance is less than 100 kΩ (optimally less than 30 kΩ). During the acquisition process, subjects must keep their eyes open, remain still, and remain silent. They should maintain steady breathing and a relaxed state, and use air conditioning to prevent sweating that can cause signal drift. Finally, the 3-5 minutes of collected data are stored along with the subject's detailed information.

[0120] In this example, the default EEG data frequency collected is 1000 Hz, a resolution far exceeding the basic data accuracy requirements of deep neural networks. To enhance noise immunity, downsampling is used to reduce the frequency to 256 Hz. Subsequently, a bandpass filter is used to remove noise interference between 0-0.5 Hz and above 45 Hz. Each subject's EEG recording is divided into 4-second segments, which serve as processing units for subsequent analysis; any tail segments shorter than 4 seconds are discarded. For classification, the overall classification result is based on the classification results for at least half of all segments of the individual. Segments with heavy noise or significant signal drift are manually screened and removed. Each EEG segment is then subjected to a discrete Fourier transform, converting the signal from the time domain to the frequency domain to enhance feature clarity. The effective frequency band of 0.5-45 Hz is retained, while also preserving the original time domain features. To facilitate subsequent analysis, each sample is resized to a three-dimensional structure with dimensions of [1 × 62 × 196], where 62 represents the number of channels and 196 is the feature length of the 0.5-45 Hz frequency band. Finally, samples were randomly mixed by individual (not fragment) and divided into training set, validation set and test set in a ratio of 3:1:1. The first two were used for model training and optimization, and the latter was used for model testing. The label format of each sample was [0, 1] or [1,0], corresponding to autistic individuals and normally developing individuals, respectively.

[0121] The model in this embodiment is designed based on a multi-layer neural network structure, which mainly includes an input layer, multiple hidden layers and an output layer. The input layer is responsible for receiving EEG data features and extracting the time and frequency features of the EEG signal through different convolutional layers and pooling layers. The hidden layer applies batch normalization and activation functions to enhance the generalization ability of the model and avoid overfitting. At the same time, a dropout layer is set in the hidden layer to randomly discard neurons with a certain probability to further improve the stability of the model. Finally, the output layer uses a softmax activation function to convert the output into a probability for high and low awareness states or other multi-classification tasks. The architecture of the model supports migration between EEG data with different numbers of channels, which helps to cope with different data input formats. In order to search for the optimal network structure suitable for EEG data migration, the multi-agent deep deterministic policy gradient (MADDPG) algorithm is used for collaborative training. The MADDPG algorithm contains multiple agents, each of which consists of an independent actor network and a critic network. (1) Actor network: An independent actor network is assigned to each agent. The actor network obtains input from the observation space of the agent and outputs an action vector that adapts to the action space. The input of the policy network is the observation data of each agent, and the output is the action, which indicates the strategy of each agent. (2) Critic network: used to evaluate the action quality of each agent. The input of each critic network includes the observations and actions of all agents, and the strategies of other agents are taken into account during the evaluation, thereby achieving collaborative optimization. The critic network adjusts the parameters to optimize the Q value of the action by minimizing the loss function. The update goal of the critic network is to minimize the Q value deviation of the action, that is, to make the policy output in the current state close to the optimal. In order to ensure that the model obtains better feature representation on complex EEG data, a dynamic reward function is used to reward high-precision classification output and encourage the actor and critic networks to optimize the adaptability. Through the actor-critic mechanism in the MADDPG algorithm, the model obtains better migration ability between EEG data with different numbers of channels.

[0122] To evaluate and optimize the model's transfer learning performance for low-channel EEG data, we first calculated the model's classification accuracy on a test set and compared the classification results under different channel numbers to verify the effectiveness of the MADDPG algorithm in transfer learning. Subsequently, we used multiple rounds of cross-validation and segmented experiments to test the model's stability under different input conditions, including the speed of training convergence and performance differences under varying channel numbers, thereby ensuring the model's adaptability and output performance during the transfer process. Based on the classification accuracy and stability evaluation results, we optimized the model's hyperparameters (such as the learning rate, batch size, number of hidden layers, and dropout rate), ultimately improving the model's ability to handle complex features while maintaining its adaptability and stability.

[0123] The optimal neural network architecture finally selected is suitable for the classification task of low-channel EEG data. To reduce the equipment acquisition and computing costs, the model was lightweighted and then ported to EEG devices with low sampling rates. The specific processing steps include reducing the number of convolution and pooling layers and optimizing the number of model parameters to reduce storage and computing requirements. In addition, based on the characteristics of low-sampling rate devices, the input data format of the model was readjusted to ensure that the model can adapt to the reduction in sampling rate while maintaining classification accuracy. After the model's performance was verified on low-channel EEG data, real-time classification prediction can be achieved to meet the actual application needs of low-cost EEG devices, providing technical support for deployment on mobile devices or in resource-constrained scenarios.

[0124] The network architecture search system for EEG signal processing provided by the present invention is described below. The network architecture search system for EEG signal processing described below and the network architecture search method for EEG signal processing described above can be referenced to each other.

[0125] Figure 3 FIG. 1 is a schematic diagram of a network architecture search system for EEG signal processing provided by an embodiment of the present invention. Figure 3 As shown, it includes: an acquisition module 31, a pre-processing module 32, a training module 33 and an optimization module 34, wherein:

[0126] The acquisition module 31 is used to collect the EEG signal data of the target subject; the preprocessing module 32 is used to preprocess the EEG signal data to obtain preprocessed EEG signal data; the training module 33 is used to use the preprocessed EEG signal data and utilize the multi-agent deep deterministic policy gradient algorithm to perform architecture search training on the neural network model to obtain the original EEG signal acquisition model; the optimization module 34 is used to evaluate and optimize the original EEG signal acquisition model to determine the optimal EEG signal acquisition model.

[0127] 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 aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A network architecture search method for EEG signal processing, characterized in that: include: Collecting EEG signal data of the target subject; Preprocessing the EEG signal data to obtain preprocessed EEG signal data; Using the preprocessed EEG signal data, a multi-agent deep deterministic policy gradient algorithm is used to perform architecture search training on a neural network model to obtain an original EEG signal acquisition model; Evaluating and optimizing the original EEG signal acquisition model to determine the optimal EEG signal acquisition model; The pre-processed EEG signal data is used to perform architecture search training on the neural network model using a multi-agent deep deterministic policy gradient algorithm to obtain the original EEG signal acquisition model, including: Determining that the neural network model includes an input layer, multiple hidden layers, and an output layer to form a multi-agent decision-making architecture; Searching for the neural network model based on the multi-agent deep deterministic policy gradient algorithm, and searching for the optimal network architecture for high- and low-channel EEG data migration through collaboration and policy optimization between agents; Performing initial training on preset high-channel EEG data to obtain training data, migrating the training data to preset low-channel EEG data for learning, and obtaining the original EEG signal acquisition model; The neural network model is searched based on the multi-agent deep deterministic policy gradient algorithm. Through collaboration and policy optimization between agents, the optimal network architecture for high- and low-channel EEG data migration is searched, including: Determine that the intelligent agent includes a policy network and an evaluation network, wherein the input of the policy network is the intelligent agent observation data, and the output is the action vector adapted to the action space; the input of the evaluation network is the intelligent agent observation and action, and the output is the action quality of the intelligent agent; The evaluation network evaluates the agent's action quality Q value by minimizing the loss function and updates the evaluation network; The policy network generates a Q-value-maximizing action by calculating the policy gradient and updates the policy network; The target network is added to the strategy network and the evaluation network respectively, and soft updates are performed to obtain the optimal network architecture for high- and low-channel EEG data migration.

2. The network architecture search method for EEG signal processing according to claim 1, characterized in that: Collect EEG signal data from the target subject, including: Collect high-density 62-channel 1000Hz resting-state EEG signal data.

3. The network architecture search method for EEG signal processing according to claim 1, characterized in that: Preprocessing the EEG signal data to obtain preprocessed EEG signal data includes: Downsampling, filtering, and artifact removal are sequentially performed on the EEG signal data to obtain filtered EEG signal data; Segmenting the filtered EEG signal data into a plurality of EEG segments using a fixed time window; Performing discrete Fourier transform and bandpass filtering on the multiple EEG segments to obtain EEG frequency domain signals and retaining EEG time domain signals; The EEG frequency domain signals and the EEG time domain signals are randomly mixed according to the subjects and divided into a training set, a validation set and a test set.

4. The network architecture search method for EEG signal processing according to claim 1, characterized in that: Determining that the neural network model includes an input layer, multiple hidden layers, and an output layer, forming a multi-agent decision-making architecture, including: The input layer receives EEG data features; The multiple hidden layers use convolution layers and pooling layers of different sizes to extract time-frequency features from the EEG data features; The output layer generates classification results using the Softmax function.

5. The network architecture search method for EEG signal processing according to claim 1, characterized in that: Evaluating and optimizing the original EEG signal acquisition model to determine the optimal EEG signal acquisition model includes: Calculate the classification accuracy of the original EEG signal acquisition model on the preset low-channel EEG data after migration on the test set, compare the EEG data under different channels, and verify the effectiveness of the multi-agent deep deterministic policy gradient algorithm; Multiple rounds of cross-validation and segmented experiments were used to test the stability of the original EEG signal acquisition model under different input conditions; Perform hyperparameter optimization based on classification accuracy and model stability evaluation results; The selected optimal network architecture is transplanted to a low sampling rate EEG device, and the optimal network architecture is subjected to model lightweight processing to obtain the optimal EEG signal acquisition model.

6. A network architecture search system for EEG signal processing, based on the network architecture search method for EEG signal processing according to any one of claims 1 to 5, characterized in that: include: An acquisition module, used to acquire EEG signal data of a target subject; A preprocessing module, used to preprocess the EEG signal data to obtain preprocessed EEG signal data; A training module is used to use the pre-processed EEG signal data to perform architecture search training on the neural network model using a multi-agent deep deterministic policy gradient algorithm to obtain an original EEG signal acquisition model; The optimization module is used to evaluate and optimize the original EEG signal acquisition model to determine the optimal EEG signal acquisition model.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the network architecture search method for EEG signal processing as described in any one of claims 1 to 5 is implemented.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the network architecture search method for EEG signal processing according to any one of claims 1 to 5 is implemented.

Images