Classification device based on multi-level feature fusion of autism EEG signals

Abstract

A classification device based on multi-level feature fusion of autistic EEG signals relates to the field of EEG signal analysis and solves the problem that existing deep learning methods only focus on high-level features while ignoring low-level features in the classification of autistic EEG signals. The classification method includes the following steps: obtaining EEG signals of autistic patients and performing nonlinear multidimensional preprocessing steps on the EEG signals; performing time-domain-frequency-domain dual-domain joint feature extraction on the preprocessed EEG signals; constructing a hierarchical fusion and refinement module with a three-stage processing capability of decomposition-fusion-refinement; based on the extracted hierarchical fusion features, constructing a multi-level feature fusion model with an adaptive feature weight allocation mechanism, and using a self-supervised training strategy to train the multi-level feature fusion model based on autistic EEG signals to achieve the classification of autistic and non-autistic samples. The present invention is also applicable to the field of deep learning EEG signal feature fusion.

Description

Technical Field

[0001] The present invention relates to the field of electroencephalogram (EEG) signal analysis, and in particular to a classification device based on multi-level feature fusion of autism EEG signals. Background Art

[0002] Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder that primarily affects an individual's social interaction, communication skills, and behavioral patterns. Traditional ASD diagnostic methods rely primarily on clinical observation and behavioral assessments, lacking objective biological indicators, which can lead to diagnostic variability and delays. In recent years, electroencephalography (EEG), as a non-invasive, high-temporal-resolution tool, has been widely used in the research and diagnosis of ASD. However, EEG signals are characterized by nonlinearity, non-stationarity, and high dimensionality. Traditional feature extraction and classification methods have certain limitations when processing complex EEG signals.

[0003] While existing deep learning methods for classifying EEG signals for autism spectrum disorder can automatically extract features, they often focus on fusing high-level features, resulting in the loss of important low-level feature information and an inability to fully capture subtle changes and patterns in EEG signals. Furthermore, traditional feature fusion strategies lack the ability to distinguish between unique and shared information in the time and frequency domains, failing to effectively decouple the complementarity and redundancy of features in each domain, reducing the model's discriminative ability. Summary of the Invention

[0004] The present invention addresses the limitations of existing deep learning methods in autism EEG signal classification, which only focus on high-level features while ignoring low-level features. To solve the above technical problems, the present invention is implemented through the following technical solutions:

[0005] The present invention proposes a classification device based on multi-level feature fusion of autism EEG signals, the classification device comprising:

[0006] a preprocessing module, configured to obtain an EEG signal of an autistic patient and perform a nonlinear multidimensional preprocessing step on the EEG signal of the autistic patient;

[0007] A feature extraction module is used to perform time-frequency dual-domain joint feature extraction on the EEG signals of autistic patients preprocessed in the preprocessing module, including constructing a time-domain feature extraction model based on a neural network with dynamic parameter adjustment, constructing a frequency-domain feature extraction model based on signal transformation and spectrum analysis methods, and obtaining a multi-level hierarchical feature representation from the above extraction paths;

[0008] The hierarchical fusion module is used to construct a hierarchical fusion and refinement module HDFR with a three-stage processing capability of decomposition-fusion-refinement. It includes a feature decoupling unit and an orthogonal fusion unit, which are used to perform hierarchical fusion of the features extracted in the feature extraction module.

[0009] The classification module is used to construct a multi-level feature fusion model with an adaptive feature weight allocation mechanism based on the hierarchical fusion features extracted by the hierarchical fusion module. The self-supervised training strategy is used to train the multi-level feature fusion model based on autism EEG signals to achieve the classification of autistic and non-autistic samples.

[0010] The feature extraction module uses a method to extract dual features in the time and frequency domains from the pre-processed autism EEG signals: a deep neural network with adaptive parameter adjustment capabilities is constructed to achieve collaborative extraction and representation learning of time and frequency domain features;

[0011] The method of collaboratively extracting time domain and frequency domain features by constructing a deep neural network with adaptive parameter adjustment capabilities is as follows:

[0012] After inputting the preprocessed autism EEG signal time series, the multi-scale residual block is used to extract time domain features at different scales. An attention mechanism is introduced in the three dimensions of channel, space, and kernel to dynamically capture the key features of autism EEG signals.

[0013] After inputting the preprocessed autism EEG signal time series, it is converted into the spectral domain through short-time Fourier transform, and the frequency domain features are extracted through full-dimensional dynamic convolution with different kernel sizes;

[0014] The hierarchical fusion and refinement module HDFR constructed in the hierarchical fusion module includes the differential fusion block module DFB and the orthogonal fusion block module OFB.

[0015] Differential fusion block module DFB is used to separate the unique and shared features in the time domain and frequency domain;

[0016] The orthogonal fusion block module OFB is used to fuse multi-level features of time domain and frequency domain features.

[0017] Furthermore, a preferred embodiment is provided, in which the preprocessing module preprocesses the EEG signals of autistic patients, including adjusting the signal frequency to a predetermined sampling rate through an adaptive downsampling method, recalibrating the signal baseline using a whole-brain average re-reference method, combining a Butterworth bandpass filter and a notch filter for frequency domain purification, decomposing the signal components using independent component analysis (ICA) technology, automatically identifying and removing electrooculogram and electromyography physiological artifacts, and applying Z-Score normalization processing to the filtered signal.

[0018] Furthermore, a preferred embodiment is provided, in which a Frobenius orthogonal constraint is further provided in the orthogonal fusion block module OFB, so as to ensure that different feature representations maintain maximum independence in a high-dimensional space.

[0019] Furthermore, a preferred embodiment is provided, in which a method for constructing a multi-level feature fusion model based on autism EEG signals in a classification module is as follows:

[0020] Time domain feature extraction is used to pass the time domain input time_input through the time domain feature extraction model to obtain the time domain features of each level;

[0021] Frequency domain feature extraction is used to pass the frequency domain input freq_input through the frequency domain feature extraction model to obtain the frequency domain features of each level;

[0022] Feature fusion is used to fuse the time domain features and frequency domain features of each level using the hierarchical fusion and refinement module.

[0023] Global average pooling is used to perform global average pooling on the fused feature map and output a multi-level feature fusion model based on autism EEG signals.

[0024] The present invention is beneficial in that:

[0025] This paper proposes a classification method based on multi-level feature fusion of autism EEG signals. By extracting time-domain and frequency-domain features and fusing them across levels, combined with full-dimensional dynamic convolution and a hierarchical fusion and refinement HDFR module, this method achieves refined feature extraction and efficient fusion of autism EEG signals, improving classification accuracy and robustness.

[0026] The present invention is also applicable to the field of deep learning EEG signal feature fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic diagram of the multi-level feature fusion model architecture described in Implementation Method 1.

[0028] Figure 2 Schematic diagram of the model architecture of the multi-scale residual block described in Implementation Method 1.

[0029] Figure 3 This is a schematic diagram of the differential fusion block DFB architecture described in the fifth implementation mode.

[0030] Figure 4 This is a schematic diagram of the orthogonal fusion block OFB architecture described in the fifth implementation mode.

[0031] Figure 5 Schematic diagram of the confusion matrix of the subject-related experiment in Implementation Method 11.

[0032] Figure 6 Schematic diagram of the confusion matrix for the subject-irrelevant experiment in Implementation 11.

[0033] Figure 7 Schematic diagram of comparative experimental results of subject-related experiments in Implementation Method 11.

[0034] Figure 8 This is a schematic diagram of the comparative test results of the subject-irrelevant experiment in Implementation Method 11. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solutions and advantages of the implementation methods of this application clearer, the technical solutions in the implementation methods of this application will be clearly and completely described below in combination with the drawings in the implementation methods of this application. Obviously, the described implementation methods are only part of the implementation methods of this application, not all of the implementation methods.

[0036] Embodiment 1: This embodiment provides a classification method based on multi-level feature fusion of autism EEG signals, the classification method comprising the following steps:

[0037] Step 1: obtaining an EEG signal of an autistic patient, and performing a nonlinear multidimensional preprocessing step on the EEG signal of the autistic patient;

[0038] Step 2: performing time-frequency dual-domain joint feature extraction on the EEG signals of the autistic patients preprocessed in step 1, including constructing a time-domain feature extraction model based on a neural network with dynamic parameter adjustment, constructing a frequency-domain feature extraction model based on signal transformation and spectrum analysis methods, and obtaining a multi-level hierarchical feature representation from the above extraction paths;

[0039] Step 3: Construct a hierarchical fusion and refinement module HDFR with three-stage processing capabilities of decomposition-fusion-refinement, including a feature decoupling unit and an orthogonal fusion unit, for hierarchical fusion of the features extracted in step 2;

[0040] Step 4: Based on the hierarchical fusion features extracted in step 3, a multi-level feature fusion model with an adaptive feature weight allocation mechanism is constructed, and a self-supervised training strategy is used to train the multi-level feature fusion model based on autism EEG signals to achieve the classification of autism and non-autism samples.

[0041] Implementation method 2. This implementation method further limits the classification method based on multi-level feature fusion of autism EEG signals described in implementation method 1. The preprocessing of the EEG signals of autistic patients in step 1 includes adjusting the signal frequency to a predetermined sampling rate through an adaptive downsampling method, recalibrating the signal baseline using a whole-brain average re-reference method, combining Butterworth bandpass filter and notch filter for frequency domain purification, using independent component analysis ICA technology to decompose signal components, automatically identifying and removing electrooculogram and electromyography physiological artifacts, and applying Z-Score standardization processing to the filtered signal.

[0042] Implementation method three. This implementation method further limits the classification method based on multi-level feature fusion of autism EEG signals described in implementation method one. The method for performing time domain-frequency domain dual-domain joint feature extraction on the preprocessed autism EEG signals in step 2 is: by constructing a deep neural network with adaptive parameter adjustment capability to realize the collaborative extraction and representation learning of time domain and frequency domain features.

[0043] Implementation 4: This implementation further limits the classification method based on multi-level feature fusion of autism EEG signals described in Implementation 1. The method for collaboratively extracting time domain and frequency domain features by constructing a deep neural network with adaptive parameter adjustment capability is as follows:

[0044] After inputting the preprocessed autism EEG signal time series, the multi-scale residual block is used to extract time domain features at different scales. An attention mechanism is introduced in the three dimensions of channel, space, and kernel to dynamically capture the key features of autism EEG signals.

[0045] After inputting the preprocessed autism EEG signal time series, it is converted into the spectral domain through short-time Fourier transform, and the frequency domain features are extracted through full-dimensional dynamic convolution with different kernel sizes.

[0046] Implementation 5: This implementation further defines the classification method based on multi-level feature fusion of autism EEG signals described in Implementation 1. The hierarchical fusion and refinement module HDFR constructed in step 3 includes constructing a differential fusion block module DFB and an orthogonal fusion block module OFB.

[0047] Differential fusion block module DFB is used to separate the unique and shared features in the time domain and frequency domain;

[0048] The orthogonal fusion block module OFB is used to fuse multi-level features of time domain and frequency domain features.

[0049] Implementation method six: This implementation method further limits the classification method based on multi-level feature fusion of autism EEG signals described in implementation method five. The orthogonal fusion block module OFB is also provided with a Frobenius orthogonal constraint to ensure that different feature representations maintain maximum independence in high-dimensional space.

[0050] Implementation 7: This implementation further limits the classification method based on multi-level feature fusion of autism EEG signals described in Implementation 1. The method for constructing a multi-level feature fusion model based on autism EEG signals in step 4 is:

[0051] Time domain feature extraction is used to pass the time domain input time_input through the time domain feature extraction model to obtain the time domain features of each level;

[0052] Frequency domain feature extraction is used to pass the frequency domain input freq_input through the frequency domain feature extraction model to obtain the frequency domain features of each level;

[0053] Feature fusion, which is used to fuse the time domain features and frequency domain features of each level using the hierarchical fusion and refinement module;

[0054] Global average pooling is used to perform global average pooling on the fused feature map and output a multi-level feature fusion model based on autism EEG signals.

[0055] Embodiment 8: This embodiment proposes a classification device based on multi-level feature fusion of autism EEG signals, the classification device comprising:

[0056] a preprocessing module, configured to obtain an EEG signal of an autistic patient and perform a nonlinear multidimensional preprocessing step on the EEG signal of the autistic patient;

[0057] A feature extraction module is used to perform time-frequency dual-domain joint feature extraction on the EEG signals of autistic patients preprocessed in the preprocessing module, including constructing a time-domain feature extraction model based on a neural network with dynamic parameter adjustment, constructing a frequency-domain feature extraction model based on signal transformation and spectrum analysis methods, and obtaining a multi-level hierarchical feature representation from the above extraction paths;

[0058] The hierarchical fusion module is used to construct a hierarchical fusion and refinement module HDFR with a three-stage processing capability of decomposition-fusion-refinement. It includes a feature decoupling unit and an orthogonal fusion unit, which are used to perform hierarchical fusion of the features extracted in the feature extraction module.

[0059] The classification module and the hierarchical fusion features extracted by the hierarchical fusion module are used to construct a multi-level feature fusion model with an adaptive feature weight allocation mechanism. A self-supervised training strategy is used to train the multi-level feature fusion model based on autism EEG signals to realize the classification of autism and non-autism samples.

[0060] Implementation method nine. This implementation method proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the classification method based on multi-level feature fusion of autism EEG signals as described in any one of implementation methods one to seven is implemented.

[0061] Implementation method 10. This implementation method proposes a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the classification method based on multi-level feature fusion of autism EEG signals as described in any one of implementation methods 1 to 7 is implemented.

[0062] Embodiment 11. This embodiment provides the following examples to explain the above-mentioned embodiments 1 to 10. The specific examples are as follows:

[0063] See also Figure 1 As shown, the present invention proposes a classification method based on multi-level feature fusion of autism EEG signals, which specifically includes the following steps:

[0064] Step 1: Acquire data and preprocess the data to reduce the influence of noise and artifacts.

[0065] Step 2: In order to enhance the ability to capture the time domain and frequency domain features of EEG signals, we combine full-dimensional dynamic convolution to construct a feature extraction network.

[0066] Step 3: To achieve inter-level feature fusion, a Hierarchical Fusion and Refinement (HDFR) module was designed. This improvement not only enhances the ability to capture EEG signal features but also addresses the issue of low-level feature loss when fusing time-domain and frequency-domain features.

[0067] Step 4: Construct a multi-level feature fusion model based on autism EEG signals for classification.

[0068] Step 5: Compare the custom network with other network models.

[0069] Furthermore, step 1 includes the following steps:

[0070] Step 1.1. Data preprocessing: Use EEGLAB in MATLAB to preprocess the EEG signals, including:

[0071] Downsampling: To reduce the amount of computation, the sampling rate of the EEG signal is reduced to 256 Hz.

[0072] Filtering: A Butterworth bandpass filter is used with cutoff frequencies set to 0.5 Hz and 60 Hz to remove low-frequency drift and high-frequency noise. A notch filter is used to suppress power frequency interference (50 Hz or 60 Hz).

[0073] Re-reference: To improve signal quality, enhance comparability across studies, and improve the accuracy of signal analysis, EEG signals were referenced to a whole-brain average.

[0074] Artifact removal: After pre-processing, ICAlable is performed to sort out possible physiological artifacts such as electrooculography and electromyography and automatically mark and remove them.

[0075] Standardization: To avoid a certain eigenvalue from having too great an impact on the model, Z-Score standardization is used to standardize the preprocessed data.

[0076] Step 1.2: Perform data enhancement on the preprocessed data by sliding the window.

[0077] Step 1.3: Divide the data after data augmentation into data sets according to subject-related and subject-independent experimental types.

[0078] Furthermore, step 2 includes the following steps:

[0079] Step 2.1: Build the FrequencyFeatureExtractor model for time domain feature extraction. The input is the spectrum after SFTF processing. The specific settings are as follows:

[0080] Full-Dimensional Dynamic Convolution (ODConv2d): This model uses four ODConv2d layers to extract features from EEG signals. The kernel sizes of the four ODConv2d layers are 9, 7, 5, and 3, respectively, with a stride of 1×1. The output of each convolutional layer is activated using the ReLU function.

[0081] BatchNorm2d: Use nn.BatchNorm2d(128) to normalize the feature maps after convolution to stabilize the model training process.

[0082] Activation='ReLU': Use the ReLU activation function to increase the nonlinear representation capability of the model.

[0083] Step 2.2: Create a time domain feature extraction model TimeFeatureExtractor, in which the preprocessed EEG signal time series is input and then passed through a multi-scale residual block to extract features of different scales. Figure 2 It is a schematic diagram of the multi-scale residual block structure.

[0084] Multi-scale residual block: The time-domain feature extraction module uses a multi-scale residual structure. Three parallel ODConv2d blocks are configured with kernel sizes of 15, 31, and 63, respectively, to capture neural activity patterns at 60ms, 120ms, and 250ms in the time series. The features extracted by each branch are fused with the original signal through a residual connection and then nonlinearly mapped using the ReLU activation function.

[0085] Full-Dimensional Dynamic Convolution (ODConv2d): The model uses three ODConv2d layers to further extract features from the EEG signal. The kernel sizes of the three ODConv2d layers are 7, 5, and 3, respectively, with a stride of 1×1. The output of each convolutional layer is activated by the ReLU function.

[0086] BatchNorm2d: Use nn.BatchNorm2d(out_channels) to normalize the feature map after convolution.

[0087] activation='ReLU': The activation function uses ReLU (rectified linear unit), which is a nonlinear activation function commonly used in deep learning.

[0088] Application of feature extraction model: The introduction of full-dimensional dynamic convolution ODConv significantly enhances the adaptability and flexibility of feature extraction. This module realizes the dynamic capture of key features of EEG signals by introducing attention mechanisms in three dimensions: channel, space, and kernel. In particular, in the multi-scale residual structure, convolution kernels of different sizes can focus on both microscopic and macroscopic temporal patterns at the same time, effectively processing the complex non-stationary characteristics of EEG signals. Through residual connections, the model retains the key information of the original signal while avoiding the gradient vanishing problem in deep network training, improving the richness and discrimination ability of feature expression, and laying a solid foundation for the subsequent time-domain-frequency domain feature fusion, thereby significantly improving the accuracy and robustness of autism EEG signal recognition.

[0089] Furthermore, step 3 includes the following steps:

[0090] Step 3.1, such as Figure 3 As shown in the figure, a differential fusion block (DFB) module is constructed to separate the unique and shared features of the time domain and frequency domain features. This part specifically includes:

[0091] Conv2d: First, the time domain and frequency domain features are spliced along the channel dimension, and the common features of the spliced features are extracted through convolution operation.

[0092] Subtraction: Subtract the time domain and frequency domain features after ReLU at each level to amplify the unique features in the time domain and frequency domain features.

[0093] Pooling: The subtracted features are subjected to global average pooling to reduce the output dimension.

[0094] Attention mechanism: By using the SHSA attention mechanism on the two branches respectively, global attention is generated on the pooled features to suppress redundant or interfering feature information.

[0095] Adaptive weighting: Multiply the attention weights generated by the SHSA attention mechanism with the original level features to obtain the weighted feature representation.

[0096] Feature summation: Add the common features obtained by Conv2d and the weighted features to obtain two feature representations.

[0097] Step 3.2, such as Figure 4 As shown in the figure, an Orthogonal Fusion Block (OFB) module is constructed to fuse multi-level features of time domain and frequency domain features. This part specifically includes:

[0098] Attention mechanism: The two output feature representations after DFB are first spliced along the channel dimension, and then the Acmix attention mechanism is used to perform local attention weighting to suppress unimportant feature information.

[0099] Adaptive weighting: The attention weights generated by the Acmix attention mechanism are weighted with the two features output by DFB to suppress local unimportant features.

[0100] Orthogonal constraint: The two adaptively weighted features are orthogonally constrained to improve the complementarity and discriminability of the two feature information.

[0101] Cosine similarity calculation: The features after orthogonal constraints are spliced along the channel dimension, and then the cosine similarity is calculated with the previous feature after HDFR fusion.

[0102] Adaptive weighted fusion: The calculated cosine similarity is constrained to the interval [0,1] by Sigmoid, and then ,in It is the feature after the orthogonal constraint is spliced along the channel dimension. α is the dynamic adaptive weight coefficient, which is obtained by calculating the cosine similarity and mapping it through the Sigmoid function. It is also a dynamically adjusted weight parameter. It is the feature after HDFR fusion. It is the previous feature after HDFR fusion.

[0103] Conv2d: The fused features are further refined through Conv2d to extract the fused features.

[0104] Application of hierarchical fusion and refinement modules: The DFB component accurately separates the shared and unique features in the time domain and frequency domain, avoiding the problems of information redundancy and mutual interference in traditional fusion methods. Secondly, the OFB component introduces Frobenius orthogonal constraints to ensure that different feature representations maintain maximum independence in high-dimensional space, greatly enhancing the discriminative ability of features. In addition, HDFR adopts a multi-level fusion strategy to effectively transfer low-level feature information to high-level networks, overcoming the common problem of low-level feature loss in deep learning. Through the inter-layer adaptive weighting mechanism, the model can dynamically adjust the contribution of each layer's features according to different data samples, improving the model's adaptability and generalization ability on EEG data with significant individual differences.

[0105] Furthermore, step 4 includes the following steps:

[0106] Step 4.1: Integrate steps 1 to 3 to create the Class EEG Multi-Level Feature FusionNetwork (nn.Module) class to build a complete multi-level feature fusion model for autism EEG signals. This section specifically includes:

[0107] Time domain feature extraction: The time domain input time_input is used through the time domain feature extraction model to obtain the time domain features of each level.

[0108] Frequency domain feature extraction: The frequency domain input freq_input is passed through the frequency domain feature extraction model to obtain the frequency domain features of each level.

[0109] Feature fusion: Use the hierarchical fusion and refinement module to fuse the time domain features and frequency domain features of each level.

[0110] Global Average Pooling: Perform global average pooling on the fused feature map to reduce the dimensionality.

[0111] self.fc1: The pooled feature vector is input into the fully connected layer self.fc1 and activated by the ReLU function to obtain the hidden layer representation.

[0112] self.fc2: Finally, the output of the model is obtained through the fully connected layer self.fc2.

[0113] Application of a multi-level feature fusion model: This model incorporates dynamic convolution in time-domain feature extraction to construct a dynamic residual network. This model can adaptively adjust convolution kernel parameters, enhancing the ability to capture temporal variations in EEG signals. Furthermore, by fusing time-domain and frequency-domain features, it can deeply reveal the dynamic changes and complex patterns of EEG signals. Furthermore, it effectively combines time-domain and frequency-domain features. This method provides a more accurate and robust solution for assisting the diagnosis of autism spectrum disorder (ASD) electroencephalogram (EEG) signals.

[0114] Furthermore, step 5 includes the following steps:

[0115] Step 5.1: The model of the present invention was tested under both subject-related and subject-independent experimental conditions. Figure 5 and Figure 6 The confusion matrix results under these two experimental conditions are shown respectively. Figure 7 and Figure 8 Compared with deep learning models, Figure 7 Medium-depth learning models include SVM, EEGNet, ShallowConvNet, DeepConvNet, Tawhid, and Ari; Figure 8 Among the deep learning models, SVM, EEGNet, ShallowConvNet, DeepConvNet, and Ullah, the model of the present invention performs excellently in performance evaluation.

[0116] Those skilled in the art will appreciate that the features described in the various embodiments of this disclosure may be combined or coupled in various ways, even if such combinations or couplings are not explicitly described in this disclosure. In particular, the various embodiments of this disclosure may be combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations are intended to fall within the scope of this disclosure.

[0117] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic inventive concepts. Therefore, the accompanying detailed description is intended to be interpreted as including the preferred embodiment and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, the present invention is intended to include such changes and modifications as fall within the scope of the technical equivalents of the present invention.

Claims

1. A classification device based on multi-level feature fusion of autism EEG signals, characterized by: The classification device comprises: a preprocessing module, configured to obtain an EEG signal of an autistic patient and perform a nonlinear multidimensional preprocessing step on the EEG signal of the autistic patient; A feature extraction module is used to perform time-frequency dual-domain joint feature extraction on the EEG signals of autistic patients preprocessed in the preprocessing module, including constructing a time-domain feature extraction model based on a neural network with dynamic parameter adjustment, constructing a frequency-domain feature extraction model based on signal transformation and spectrum analysis methods, and obtaining a multi-level hierarchical feature representation from the paths extracted by the time-domain feature extraction model and the frequency-domain feature extraction model; The hierarchical fusion module is used to construct a hierarchical fusion and refinement module HDFR with a three-stage processing capability of decomposition-fusion-refinement. It includes a feature decoupling unit and an orthogonal fusion unit, which are used to perform hierarchical fusion of the features extracted in the feature extraction module. The classification module is used to construct a multi-level feature fusion model with an adaptive feature weight allocation mechanism based on the hierarchical fusion features extracted by the hierarchical fusion module. The self-supervised training strategy is used to train the multi-level feature fusion model based on autism EEG signals to achieve the classification of autistic and non-autistic samples. The feature extraction module uses a method to extract dual features in the time and frequency domains from the pre-processed autism EEG signals: a deep neural network with adaptive parameter adjustment capabilities is constructed to achieve collaborative extraction and representation learning of time and frequency domain features; The method of collaboratively extracting time domain and frequency domain features by constructing a deep neural network with adaptive parameter adjustment capabilities is as follows: After inputting the preprocessed autism EEG signal time series, the multi-scale residual block is used to extract time domain features at different scales. An attention mechanism is introduced in the three dimensions of channel, space, and kernel to dynamically capture the key features of autism EEG signals. After inputting the pre-processed autism EEG signal time series, it is converted into the spectral domain through short-time Fourier transform, and the frequency domain features are extracted through full-dimensional dynamic convolution with different kernel sizes; The hierarchical fusion and refinement module HDFR constructed in the hierarchical fusion module includes the differential fusion block module DFB and the orthogonal fusion block module OFB. Differential fusion block module DFB is used to separate the unique and shared features in the time domain and frequency domain; The orthogonal fusion block module OFB is used to fuse multi-level features of time domain and frequency domain features.

2. The classification device based on multi-level feature fusion of autism EEG signals according to claim 1, characterized in that: The preprocessing module preprocesses the EEG signals of autistic patients, including adjusting the signal frequency to a predetermined sampling rate through an adaptive downsampling method, recalibrating the signal baseline using a whole-brain average re-reference method, combining Butterworth bandpass filters and notch filters for frequency domain purification, decomposing signal components using independent component analysis (ICA) technology, automatically identifying and removing electrooculogram and electromyography physiological artifacts, and applying Z-Score standardization processing to the filtered signal.

3. The classification device based on multi-level feature fusion of autism EEG signals according to claim 1, characterized in that: The orthogonal fusion block module OFB is also provided with a Frobenius orthogonal constraint to ensure that different feature representations maintain maximum independence in high-dimensional space.

4. The classification device based on multi-level feature fusion of autism EEG signals according to claim 1, characterized in that: The method for constructing a multi-level feature fusion model based on autism EEG signals in the classification module is: Time domain feature extraction is used to pass the time domain input time_input through the time domain feature extraction model to obtain the time domain features of each level; Frequency domain feature extraction is used to pass the frequency domain input freq_input through the frequency domain feature extraction model to obtain the frequency domain features of each level; feature Fusion, used to fuse the time domain features and frequency domain features of each level using the hierarchical fusion and refinement module; Global average pooling is used to perform global average pooling on the fused feature map and output a multi-level feature fusion model based on autism EEG signals.

Images