Classification method and device for multi-level feature fusion based on autism electroencephalogram signals

Abstract

The invention discloses a classification method and device for multi-level feature fusion based on autism electroencephalogram signals, relates to the field of electroencephalogram signal analysis, and solves the problem that an existing deep learning method only pays attention to high-level features and neglects low-level features in autism electroencephalogram signal classification. The classification method comprises the following steps: acquiring an electroencephalogram signal of an autism patient, and executing a nonlinear multi-dimensional preprocessing step on the electroencephalogram signal; performing time domain-frequency domain double-domain joint feature extraction on the preprocessed electroencephalogram signals; constructing a hierarchical fusion and refinement module with decomposition-fusion-refinement three-stage processing capability; and based on the extracted hierarchical fused features, constructing a multi-level feature fusion model with an adaptive feature weight distribution mechanism, and training the multi-level feature fusion model based on autism electroencephalogram signals by adopting a self-supervised training strategy to realize classification of autism and non-autism samples. The method is also suitable for the field of deep learning electroencephalogram signal feature fusion.

Description

Technical Field

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

[0002] Autism Spectrum Disorder (ASD) is a complex neurodevelopmental disorder that mainly affects an individual's social interaction, communication ability, and behavior pattern. Traditional ASD diagnosis methods mainly rely on clinical observation and behavioral assessment, lacking objective biological indicators, which may lead to diagnostic differences and delays. In recent years, electroencephalogram (EEG), as a non-invasive and high-time-resolution tool, has been widely used in the research and diagnosis of ASD. However, EEG signals have characteristics such as non-linearity, non-stationarity, and high dimensionality, and traditional feature extraction and classification methods have certain limitations in dealing with complex EEG signals.

[0003] In the classification of EEG signals of Autism Spectrum Disorder, although existing deep learning methods can automatically extract features, they often focus on the fusion of high-level features, resulting in the loss of important low-level feature information and being unable to comprehensively capture the subtle changes and patterns in EEG signals. Secondly, traditional feature fusion strategies lack a fine distinction between unique and shared information in the time domain and frequency domain, and fail to effectively decouple the complementarity and redundancy of features in each domain, reducing the discriminative ability of the model. Summary of the Invention

[0004] The present invention aims at the limitations of existing deep learning methods in the classification of autistic EEG signals, such as only focusing on high-level features and ignoring low-level features. To solve the above technical problems, the present invention is implemented through the following technical solutions: The present invention proposes a classification method for multi-level feature fusion based on autistic EEG signals, and the classification method includes the following steps: Step 1: Obtain the EEG signals of autistic patients, and perform a non-linear multi-dimensional preprocessing step on the EEG signals of the autistic patients; Step 2: Perform time-frequency dual-domain joint feature extraction on the preprocessed EEG signals of autistic patients 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 spectral analysis methods, and obtaining a multi-level hierarchical feature representation from the above extraction paths; Step 3: Construct a hierarchical fusion and refinement module HDFR with the processing ability of three stages of decomposition-fusion-refinement, including a feature decoupling unit and an orthogonal fusion unit, for hierarchically fusing the features extracted in Step 2; Step 4: Based on the hierarchically fused features extracted in Step 3, construct a multi-level feature fusion model with an adaptive feature weight allocation mechanism, and adopt 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.

[0005] Further, a preferred implementation is provided. 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 reference using the whole-brain average re-reference method, performing frequency-domain purification by combining a Butterworth band-pass filter and a notch filter, decomposing the signal components using the independent component analysis (ICA) technique, automatically identifying and removing electrooculogram (EOG) and electromyogram (EMG) physiological artifacts, and applying Z-Score normalization to the filtered signals.

[0006] Further, a preferred implementation is provided. The method for extracting time-frequency dual-domain joint features from the preprocessed autistic EEG signals in Step 2 is: realizing the collaborative extraction and representation learning of time-domain and frequency-domain features by constructing a deep neural network with adaptive parameter adjustment ability.

[0007] Further, a preferred implementation is provided. The method for realizing the collaborative extraction of time-domain and frequency-domain features by constructing a deep neural network with adaptive parameter adjustment ability is: After inputting the time series of the preprocessed autistic EEG signals, extract features of different scales through multi-scale residual blocks, and introduce an attention mechanism in the three dimensions of channels, space, and kernels to achieve dynamic capture of the key features of autistic EEG signals.

[0008] After inputting the time series of the preprocessed autistic EEG signals, convert them into the frequency spectrum domain through the short-time Fourier transform, and extract frequency-domain features through full-dimensional dynamic convolutions with different kernel sizes.

[0009] Further, a preferred implementation is provided. The hierarchically fused and refined module HDFR constructed in Step 3 includes constructing a differential fusion block module DFB and an orthogonal fusion block module OFB. The differential fusion block module DFB is used to separate the unique and shared features of the time domain and the frequency domain from the time-domain and frequency-domain features. The orthogonal fusion block module OFB is used to fuse multi-level features of the time domain and the frequency domain.

[0010] Further, a preferred implementation is provided. A Frobenius orthogonality constraint is also set in the orthogonal fusion block module OFB to ensure that different feature representations maintain maximum independence in the high-dimensional space.

[0011] Further, a preferred implementation is provided. The method for constructing a multi-level feature fusion model based on autistic EEG signals in step 4 is as follows: Time-domain feature extraction, which 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, which 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, which is used to fuse the time-domain features and frequency-domain features of each level using the hierarchical fusion and refinement module.

[0012] Global average pooling, which is used to perform global average pooling on the fused feature map and output the constructed multi-level feature fusion model based on autistic EEG signals. Solution 2: A classification device based on multi-level feature fusion of autistic EEG signals. The classification device includes: A preprocessing module, which is used to obtain the EEG signals of autistic patients and perform non-linear multi-dimensional preprocessing steps on the EEG signals of the autistic patients; A feature extraction module, which is used to perform time-frequency dual-domain joint feature extraction on the preprocessed EEG signals of autistic patients 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 spectral analysis methods, and obtaining a multi-level hierarchical feature representation from the above extraction paths; A hierarchical fusion module, which is used to construct a hierarchical fusion and refinement module HDFR with the processing capabilities of decomposition-fusion-refinement, including a feature decoupling unit and an orthogonal fusion unit, for performing hierarchical fusion on the features extracted in the feature extraction module; A classification module, which constructs a multi-level feature fusion model with an adaptive feature weight allocation mechanism based on the hierarchically fused features extracted by the hierarchical fusion module, and uses 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.

[0013] Solution 3: A computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the classification method described in any one of Solution 1.

[0014] Solution 4: A computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the classification method described in any one of Solution 1.

[0015] The advantages of the present invention are as follows: The present invention proposes a classification method based on multi-level feature fusion of autistic EEG signals. By extracting time-domain features and frequency-domain features for inter-level fusion, and combining full-dimensional dynamic convolution and hierarchical fusion and refinement HDFR modules, the present invention realizes refined feature extraction and efficient fusion of autistic EEG signals, and improves the accuracy and robustness of classification.

[0016] The present invention is also applicable to the field of deep learning EEG signal feature fusion. Brief Description of the Drawings

[0017] Figure 1 It is a schematic diagram of the multi-level feature fusion model architecture described in Embodiment 1.

[0018] Figure 2 It is a schematic diagram of the model architecture of the multi-scale residual block described in Embodiment 1.

[0019] Figure 3 It is a schematic diagram of the differential fusion block DFB architecture described in Embodiment 5.

[0020] Figure 4 It is a schematic diagram of the orthogonal fusion block OFB architecture described in Embodiment 5.

[0021] Figure 5 It is a schematic diagram of the confusion matrix of the subject-related experiment in Embodiment 11.

[0022] Figure 6 It is a schematic diagram of the confusion matrix of the subject-unrelated experiment in Embodiment 11.

[0023] Figure 7 It is a schematic diagram of the comparison experiment results of the subject-related experiment in Embodiment 11.

[0024] Figure 8 It is a schematic diagram of the comparison test results of the subject-unrelated experiment in Embodiment 11. Detailed Embodiments

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments.

[0026] Embodiment 1. This embodiment provides a classification method based on multi-level feature fusion of autistic EEG signals. The classification method includes the following steps: Step 1: Obtain the EEG signals of autistic patients, and perform a non-linear multi-dimensional preprocessing step on the EEG signals of the autistic patients; Step 2: Perform time-frequency dual-domain joint feature extraction on the preprocessed EEG signals of autistic patients, 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 spectral analysis methods, and obtaining a multi-level hierarchical feature representation from the above extraction paths; Step 3: Construct a hierarchical fusion and refinement module HDFR with the ability of three-stage processing of decomposition-fusion-refinement, including a feature decoupling unit and an orthogonal fusion unit, for hierarchically fusing the features extracted in Step 2; Step 4: Based on the hierarchically fused features extracted in Step 3, construct a multi-level feature fusion model with an adaptive feature weight allocation mechanism, and adopt 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.

[0027] Embodiment 2: This embodiment further limits the classification method of multi-level feature fusion based on autistic EEG signals described in Embodiment 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 reference using the whole-brain average re-reference method, performing frequency-domain purification by combining a Butterworth band-pass filter and a notch filter, decomposing the signal components using the independent component analysis ICA technique, automatically identifying and removing electrooculogram and electromyogram physiological artifacts, and applying Z-Score normalization processing to the filtered signals. Embodiment 3: This embodiment further limits the classification method of multi-level feature fusion based on autistic EEG signals described in Embodiment 1. The method for performing time-domain-frequency-domain dual-domain joint feature extraction on the preprocessed autistic EEG signals in Step 2 is: realizing the collaborative extraction and representation learning of time-domain and frequency-domain features by constructing a deep neural network with adaptive parameter adjustment ability.

[0028] Embodiment 4: This embodiment further limits the classification method of multi-level feature fusion based on autistic EEG signals described in Embodiment 1. The method for realizing the collaborative extraction of time-domain and frequency-domain features by constructing a deep neural network with adaptive parameter adjustment ability is: After inputting the time series of the preprocessed autistic EEG signals, extract time-domain features of different scales through a multi-scale residual block, and introduce an attention mechanism in the three dimensions of channels, space, and kernels to realize the dynamic capture of key features of autistic EEG signals; After inputting the time series of the preprocessed autistic EEG signals, convert them into the frequency spectrum domain through the short-time Fourier transform, and extract frequency-domain features through full-dimensional dynamic convolutions with different kernel sizes.

[0029] Embodiment 5. This embodiment further limits the classification method based on multi-level feature fusion of autistic EEG signals described in Embodiment 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. The differential fusion block module DFB is used to separate the unique and shared features of the time domain and frequency domain from the time domain and frequency domain features. The orthogonal fusion block module OFB is used to fuse the multi-level features of the time domain and frequency domain features.

[0030] Embodiment 6. This embodiment further limits the classification method based on multi-level feature fusion of autistic EEG signals described in Embodiment 5. A Frobenius orthogonal constraint is also set in the orthogonal fusion block module OFB to ensure that different feature representations maintain maximum independence in the high-dimensional space.

[0031] Embodiment 7. This embodiment further limits the classification method based on multi-level feature fusion of autistic EEG signals described in Embodiment 1. The method for constructing a multi-level feature fusion model based on autistic EEG signals in step 4 is as follows: 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 is 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 the constructed multi-level feature fusion model based on autistic EEG signals. Embodiment 8. This embodiment proposes a classification device based on multi-level feature fusion of autistic EEG signals. The classification device includes: A preprocessing module is used to obtain the EEG signals of autistic patients and perform non-linear multi-dimensional preprocessing steps on the EEG signals of the autistic patients. A feature extraction module is used to perform time-frequency dual-domain joint feature extraction on the preprocessed EEG signals of autistic patients 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 spectral analysis methods, and obtaining multi-level hierarchical feature representations from the above extraction paths. Hierarchical fusion module, used to 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 hierarchically fusing the features extracted in the feature extraction module; Classification module, based on the hierarchically fused features extracted by the hierarchical fusion module, constructs a multi-level feature fusion model with an adaptive feature weight allocation mechanism, and uses a self-supervised training strategy to train the multi-level feature fusion model based on autistic EEG signals to achieve classification of autistic and non-autistic samples.

[0032] Embodiment 9. This embodiment proposes a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the classification method for multi-level feature fusion based on autistic EEG signals according to any one of Embodiments 1 to 7.

[0033] Embodiment 10. This embodiment proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the classification method for multi-level feature fusion based on autistic EEG signals according to any one of Embodiments 1 to 7.

[0034] Embodiment 11. This embodiment provides the following examples for explaining Embodiments 1 to 10 above. The examples are specifically as follows: See Figure 1 As shown, the present invention proposes a classification method for multi-level feature fusion based on autistic EEG signals. The method specifically includes the following steps: Step 1: Obtain data and preprocess the data to reduce the influence of noise and artifacts.

[0035] Step 2: In order to enhance the ability to capture the time-domain and frequency-domain features of EEG signals, we construct a feature extraction network in combination with full-dimensional dynamic convolution.

[0036] Step 3: In order to achieve inter-level feature fusion, a hierarchical fusion and refinement (HDFR) module is designed. Through this improvement, while enhancing the ability to capture EEG signal features, the problem of loss of low-level features when fusing time-domain and frequency-domain features is also solved.

[0037] Step 4: Construct a multi-level feature fusion model based on autistic EEG signals for classification.

[0038] Step 5: Conduct a comparative experiment between the custom network and other network models.

[0039] Further, Step 1 includes the following steps: Step 1.1. Data preprocessing: Use EEGLAB in MATLAB software to preprocess the EEG signals, including: Downsampling: To reduce the computational load, the sampling rate of the EEG signals is reduced to 256 HZ.

[0040] Filtering: Use a Butterworth band-pass filter with cut-off frequencies set at 0.5 Hz and 60 Hz to remove low-frequency drift and high-frequency noise; for power-line interference (50 Hz or 60 Hz), use a notch filter to suppress it.

[0041] Re-referencing: To improve the signal quality, enhance cross-study comparability, and improve the accuracy of signal analysis, perform a whole-brain average reference on the EEG signals.

[0042] Artifact removal: Perform ICA labeling on the preprocessed data to identify and automatically label and remove possible physiological artifacts such as electrooculogram and electromyogram.

[0043] Normalization: To avoid a certain eigenvalue having an excessive impact on the model, use Z-Score normalization to normalize the preprocessed data.

[0044] Step 1.2. Perform data augmentation on the preprocessed data using a sliding window.

[0045] Step 1.3. Divide the data after data augmentation into datasets according to subject-related and subject-unrelated experimental types.

[0046] Furthermore, Step 2 includes the following steps: Step 2.1. Construct a time-domain feature extraction model FrequencyFeatureExtractor. The input is the spectrogram after SFTF processing. The specific settings are as follows: Omni-dimensional dynamic convolution (ODConv2d): The model uses four ODConv2d to perform autonomous feature extraction on the EEG signals. The kernel sizes of the four ODConv2d are 9, 7, 5, and 3 respectively, and the strides are all 1×1. The output of each convolutional layer passes through the Relu activation function.

[0047] BatchNorm2d: Use nn.BatchNorm2d(128) to normalize the convolutional feature maps to stabilize the training process of the model.

[0048] Activation='ReLU': Use the ReLU activation function to increase the non-linear representation ability of the model.

[0049] Step 2.2: Create a time-domain feature extraction model TimeFeatureExtractor. After inputting the preprocessed EEG signal time series, it first passes through a multi-scale residual block to extract features at different scales. Figure 2 It is a schematic diagram of the multi-scale residual block structure.

[0050] Multi-scale residual block: The time-domain feature extraction module is designed with a multi-scale residual structure. Three parallel ODConv2ds are configured with kernel sizes of 15, 31, and 63 respectively to capture the neural electrical activity patterns of 60ms, 120ms, and 250ms in the time series. The features extracted by each branch are fused with the original signal through residual connections and non-linearly mapped by the ReLU activation function.

[0051] Omni-dimensional dynamic convolution (ODConv2d): The model then uses three ODConv2ds to further extract features from the EEG signal. The kernel sizes of the three ODConv2ds are 7, 5, and 3 respectively, and the stride is 1×1. The output of each convolutional layer passes through the Relu activation function; BatchNorm2d: Use nn.BatchNorm2d(out_channels) to normalize the convolutional feature map.

[0052] activation='ReLU': The activation function uses ReLU (Rectified Linear Unit), which is a non-linear activation function commonly used in deep learning.

[0053] Application of the feature extraction model: The introduction of the omni-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 an attention mechanism in the three dimensions of channels, space, and kernels. Especially in the multi-scale residual structure, convolutional kernels of different sizes can simultaneously focus on microscopic and macroscopic temporal patterns, effectively dealing with the complex non-stationary characteristics in EEG signals. Through residual connections, the model retains the key information of the original signal, while avoiding the problem of gradient disappearance in the training of deep networks, improving the richness and discriminative ability of feature expression, laying a solid foundation for the subsequent time-frequency domain feature fusion, and thus significantly improving the accuracy and robustness of autism EEG signal recognition.

[0054] Furthermore, Step 3 includes the following steps: Step 3.1, as Figure 3 shown, construct a differential fusion block (DFB) module to separate the unique and shared features of the time domain and frequency domain for time domain and frequency domain features. This part specifically includes: Conv2d: First, concatenate the time-domain kernel frequency-domain features along the channel dimension, and then perform feature extraction of common features on the concatenated features through convolution operations.

[0055] Subtraction: Subtract the time-domain and frequency-domain features after ReLU for each level to amplify the unique features in the time-domain and frequency-domain features.

[0056] Pooling: Respectively pass the subtracted features through global average pooling to reduce the output dimension.

[0057] Attention mechanism: Generate global attention for the pooled features by using the SHSA attention mechanism for the two branches respectively, and suppress redundant or interfering feature information.

[0058] Adaptive weighting: Multiply the attention weights generated by the SHSA attention mechanism by the original level features to obtain weighted feature representations. Feature summation: Add the common features obtained through Conv2d to the weighted features to obtain two feature representations.

[0059] Step 3.2, as Figure 4 shown, construct an Orthogonal Fusion Block (OFB) module for fusing multi-level features of time-domain and frequency-domain features. This part specifically includes: Attention mechanism: First, concatenate the two output feature representations after DFB along the channel dimension, and then perform local attention weighting through the Acmix attention mechanism to suppress unimportant feature information.

[0060] Adaptive weighting: Weight the two features output after DFB with the attention weights generated by the Acmix attention mechanism to suppress locally unimportant features.

[0061] Orthogonal constraint: Apply orthogonal constraints to the two features after adaptive weighting to enhance the complementarity and discriminability of the two types of feature information.

[0062] Cosine similarity calculation: Concatenate the features after orthogonal constraint along the channel dimension, and then calculate the cosine similarity with the feature after the previous HDFR fusion.

[0063] Adaptive weighted fusion: Constrain the calculated cosine similarity to the range [0, 1] through Sigmoid, and then use , where is the feature obtained by concatenating the features after orthogonal constraint along the channel dimension, α is the dynamic adaptive weight coefficient, obtained through cosine similarity calculation and mapped through the Sigmoid function, and is a dynamically adjusted weight parameter, is the feature after HDFR fusion, It is the previous feature after HDFR fusion.

[0064] Conv2d: Further refine the fused features through Conv2d for feature extraction.

[0065] Application of the hierarchical fusion and refinement module: The DFB component accurately separates the shared and unique features in the time and frequency domains, avoiding information redundancy and mutual interference problems in traditional fusion methods. Secondly, the OFB component introduces the Frobenius orthogonality constraint to ensure that different feature representations maintain maximum independence in the high-dimensional space, greatly enhancing the discriminative ability of the features. In addition, HDFR adopts a multi-level fusion strategy to effectively transfer low-level feature information to the high-level network, 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 degree of each layer of features according to different data samples, improving the adaptability and generalization ability of the model on EEG data with significant individual differences.

[0066] Furthermore, step 4 includes the following steps: Step 4.1, integrating steps 1 to 3, creates the Class EEG Multi-Level Feature FusionNetwork (nn.Module) class to construct a complete multi-level feature fusion model for autistic EEG signals. This part specifically includes: Time-domain feature extraction: Pass the time-domain input time_input through the time-domain feature extraction model to obtain the time-domain features of each level.

[0067] Frequency-domain feature extraction: Pass the frequency-domain input freq_input through the frequency-domain feature extraction model to obtain the frequency-domain features of each level.

[0068] Feature fusion: Use the hierarchical fusion and refinement module to fuse the time-domain features and frequency-domain features of each level.

[0069] Global Average Pooling: Perform global average pooling on the fused feature map to reduce the dimensionality. self.fc1: Input the pooled feature vector into the fully connected layer self.fc1, and after passing through the ReLU activation function, obtain the hidden layer representation.

[0070] self.fc2: Finally, through the fully connected layer self.fc2, obtain the output result of the model.

[0071] Application of the multi-level feature fusion model: Dynamic convolution is combined in time-domain feature extraction to construct a dynamic residual network, which can adaptively adjust the convolution kernel parameters and enhance the ability to capture the time-varying features of EEG signals. At the same time, by fusing time-domain and frequency-domain features, the dynamic changes and complex patterns of EEG signals are deeply revealed. And the time-domain features and frequency-domain features are effectively jointly represented. This method provides a more accurate and robust solution for the auxiliary diagnosis of electroencephalogram (EEG) signals of autism spectrum disorder (ASD).

[0072] Further, step 5 includes the following steps: Step 5.1: The model of the present invention is tested under both subject-related and subject-independent experimental conditions. Figure 5 and Figure 6 respectively show the confusion matrix results under these two experimental conditions. Compared with the deep learning models in Figure 7 and Figure 8 where the deep learning models in Figure 7 include SVM, EEGNet, ShallowConvNet, DeepConvNet, Tawhid, Ari; Figure 8 where the deep learning models in

[0073] Those skilled in the art can understand that the features described in the various embodiments of the present disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in the present disclosure. In particular, without departing from the spirit and teachings of the present disclosure, the various embodiments of the present disclosure can be combined in various ways. All such combinations fall within the scope of the present disclosure.

[0074] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concepts. Therefore, the appended embodiments are intended to be construed as including the preferred embodiments and all changes and modifications falling within the scope of the present invention. Obviously, those skilled in the art can make various changes and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the equivalent technology of the present invention, the present invention is also intended to include these changes and variations.

Claims

1. A classification method based on multi-level feature fusion of autistic EEG signals, characterized in that The classification method includes the following steps: Step 1: Obtain the electroencephalogram (EEG) signals of autistic patients, and perform a non-linear multi-dimensional preprocessing step on the EEG signals of the autistic patients; Step 2: Perform time-frequency dual-domain joint feature extraction on the preprocessed EEG signals of autistic patients 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 spectral analysis methods, and obtaining a multi-level hierarchical feature representation from the above extraction paths; Step 3: Construct a hierarchical fusion and refinement module HDFR with the ability of decomposition-fusion-refinement three-stage processing, including a feature decoupling unit and an orthogonal fusion unit, for hierarchically fusing the features extracted in Step 2; Step 4: Based on the hierarchically fused features extracted in Step 3, construct a multi-level feature fusion model with an adaptive feature weight allocation mechanism, and use 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.

2. The classification method based on multi-level feature fusion of autistic EEG signals according to claim 1, wherein 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 reference using the whole-brain average re-reference method, performing frequency-domain purification by combining a Butterworth band-pass filter and a notch filter, decomposing the signal components using independent component analysis (ICA) technology, automatically identifying and removing electrooculogram (EOG) and electromyogram (EMG) physiological artifacts, and applying Z-Score normalization processing to the filtered signals.

3. The classification method based on multi-level feature fusion of autistic EEG signals according to claim 1, characterized in that The method for performing time-domain-frequency-domain dual-joint feature extraction on the preprocessed autistic EEG signals in Step 2 is: realizing the collaborative extraction and representation learning of time-domain and frequency-domain features by constructing a deep neural network with adaptive parameter adjustment ability.

4. The classification method based on multi-level feature fusion of autistic EEG signals according to claim 3, wherein, The method for realizing the collaborative extraction of time-domain and frequency-domain features by constructing a deep neural network with adaptive parameter adjustment ability is: After inputting the time series of the preprocessed autistic EEG signals, different-scale time-domain features are extracted through multi-scale residual blocks, and an attention mechanism is introduced in the three dimensions of channels, space, and kernels to realize the dynamic capture of key features of autistic EEG signals; After inputting the time series of the preprocessed autistic EEG signals, they are transformed into the frequency spectrum domain through the short-time Fourier transform, and frequency-domain features are extracted through fully dimensional dynamic convolutions with different kernel sizes.

5. The classification method based on multi-level feature fusion of autistic EEG signals according to claim 1, wherein 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. The differential fusion block module DFB is used to separate the unique and shared features of the time domain and the frequency domain for time-domain and frequency-domain features; The orthogonal fusion block module OFB is used to fuse multi-level features of the time domain and the frequency domain.

6. The classification method based on multi-level feature fusion of autistic EEG signals according to claim 5, wherein A Frobenius orthogonal constraint is also set in the orthogonal fusion block module OFB to ensure that different feature representations maintain maximum independence in the high-dimensional space.

7. The classification method based on multi-level feature fusion of autistic EEG signals according to claim 1, wherein, The method for constructing a multi-level feature fusion model based on autistic EEG signals in Step 4 is: Time-domain feature extraction is used to obtain the time-domain features of each level by passing the time-domain input time_input through the time-domain feature extraction model; Frequency-domain feature extraction is used to obtain the frequency-domain features of each level by passing the frequency-domain input freq_input through the frequency-domain feature extraction model; Feature Fusion is 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 the constructed multi-level feature fusion model based on autistic EEG signals.

8. A classification device for multi-level feature fusion based on autistic electroencephalogram signals, characterized in that, The classification device includes: A preprocessing module for obtaining the EEG signals of autistic patients and performing non-linear multi-dimensional preprocessing steps on the EEG signals of the autistic patients; A feature extraction module for performing time-frequency dual-domain joint feature extraction on the preprocessed EEG signals of autistic patients 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 spectral analysis methods, and obtaining a multi-level hierarchical feature representation from the above extraction paths; A hierarchical fusion module for constructing a hierarchical fusion and refinement module HDFR with the ability of three-stage processing of decomposition-fusion-refinement, including a feature decoupling unit and an orthogonal fusion unit, for performing hierarchical fusion on the features extracted in the feature extraction module; A classification module constructs a multi-level feature fusion model with an adaptive feature weight allocation mechanism based on the hierarchically fused features extracted by the hierarchical fusion module, and uses 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.

9. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1-7 are implemented.

Images