SSVEP (Steady-State Visual Evoked Potential) electroencephalogram signal feature extraction and classification method, system and device and medium
By performing multiple signal decomposition and time-frequency feature extraction on SSVEP EEG signals, combined with attention mechanism and convolutional neural network processing, the problem of low classification accuracy of SSVEP EEG signals across subjects is solved, achieving higher classification accuracy and robustness.
Patent Information
- Application Number
- CN202510176029.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-18
- Publication Date
- 2025-06-06
AI Technical Summary
The performance of SSVEP EEG signal classification across subjects is affected by individual differences, resulting in low classification accuracy. How to improve the classification accuracy of SSVEP EEG signal classification across subjects is a major challenge.
By performing several signal decomposition processing, time-frequency feature extraction and wavelet inverse transformation processing on SSVEP EEG signals, the capture ability of multi-scale time-frequency features is enhanced, and through channel attention mechanism, spatial attention mechanism, deep convolution and separable convolution processing, signal feature weights are adaptively allocated and all features are fused to improve signal separability.
It effectively improves the multi-scale time-frequency characteristic characterization ability of SSVEP EEG signals, alleviates the impact of individual differences on BCIs performance, and significantly improves the accuracy and robustness of SSVEP EEG signals classification across subjects.
Smart Images
Figure BDA0005275528850000031 
Figure BDA0005275528850000051 
Figure BDA0005275528850000091
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of electroencephalogram (ESVEP) signal processing, and in particular to a method, system, device and medium for extracting and classifying SSVEP electroencephalogram (ESVEP) signal features. Background Art
[0002] Brain-Computer Interfaces (BCIs) are an innovative technology that enables information exchange between the brain and external devices or computers by directly reading and decoding brain neural activity. As a cutting-edge technology for human-computer interaction, BCIs have broad application potential in medical rehabilitation, assistive devices, entertainment games, and neuroscience research. In particular, BCIs have demonstrated great social value in providing communication or control means for patients who have lost their language or motor skills.
[0003] In practical applications, the performance of cross-subject SSVEP EEG signal classification has a decisive impact on the practicality of BCIs. SSVEP (Steady-state Visual Evoked Potential) is a phenomenon measured by the brain's electrical response to repetitive visual stimulation. When the human eye looks at a light source or pattern that flickers or changes at a specific frequency, the cerebral cortex will produce an electrical activity pattern synchronized with the stimulation frequency. This electrical activity can be recorded through EEG signals and used in BCIs systems.
[0004] However, due to significant differences between individuals in neuroanatomy, visual sensitivity, and cognitive status, the stability and consistency of SSVEP EEG signals are often affected, making cross-subject SSVEP EEG signal classification still a major challenge. How to improve the accuracy of cross-subject SSVEP EEG signal classification is an issue that requires continuous attention in the field of EEG signal processing technology. Summary of the invention
[0005] In order to solve the problems raised in the above background technology, the present invention provides a method, system, device and medium for extracting and classifying SSVEP electroencephalogram (EEG) signal features.
[0006] The technical solution of the present invention is as follows:
[0007] A SSVEP electroencephalogram signal feature extraction and classification method comprises the following steps:
[0008] S1, obtain SSVEP EEG signal, extract spatial features through convolution processing, obtain spatial features, extract time-frequency features of SSVEP EEG signal, and obtain reconstructed low-frequency features;
[0009] The time-frequency feature extraction process is as follows: the SSVEP EEG signal is subjected to several signal decomposition processes in sequence, wherein the signal decomposition process decomposes the SSVEP EEG signal to output a low-frequency component, and the low-frequency component is used as the input of the next signal decomposition process, and the last signal decomposition process outputs a low-frequency component and several high-frequency components.
[0010] All low-frequency components and several high-frequency components obtained by the above-mentioned several signal decomposition processes are sequentially subjected to channel number expansion, time-frequency feature extraction, feature scaling, and reshaping processing to obtain several low-frequency features and several high-frequency features.
[0011] Based on the low-frequency features obtained from the last signal decomposition process and several high-frequency features, and all low-frequency features obtained from several previous signal decomposition processes, a reconstructed low-frequency feature is obtained;
[0012] S2, adding the spatial features and the reconstructed low-frequency features to obtain integrated features, and performing batch normalization, activation function processing and random inactivation processing on the integrated features in turn to obtain multi-scale time-frequency features;
[0013] S3, after two-dimensional convolution processing of multi-scale time-frequency features to extract local features and batch normalization processing, obtain local time domain change features,
[0014] The local time domain change features are processed by channel attention mechanism and spatial attention mechanism in turn to obtain spatial enhanced features;
[0015] S4, performing deep convolution processing on the spatial enhancement features to obtain the local spatial features of each channel, and performing separable convolution processing on the local spatial features of all channels to obtain spatial domain fusion features;
[0016] S5. Flatten the spatial domain fusion features, and then map the spatial domain fusion features to the classification space containing different visual stimulation frequency categories through linear transformation. Use the classification function to obtain the probability distribution of different visual stimulation frequencies to obtain the SSVEP EEG signal classification results.
[0017] The specific process of S1 based on the low-frequency features obtained from the last signal decomposition process and several high-frequency features, and all low-frequency features obtained from the previous signal decomposition processes is as follows:
[0018] Starting from the last signal decomposition process, a low-frequency feature and all high-frequency features are connected along the channel dimension, and then the first reconstructed low-frequency feature is obtained by inverse wavelet transform. The first reconstructed low-frequency feature is connected with the low-frequency feature obtained by the previous signal decomposition process, and then the inverse wavelet transform is performed to obtain the second reconstructed low-frequency feature. The above process is repeated for the second reconstructed low-frequency feature until the nth reconstructed low-frequency feature is connected with the low-frequency feature obtained by the first signal decomposition process, and then the inverse wavelet transform is performed to obtain the reconstructed low-frequency feature.
[0019] The signal decomposition process in S1 decomposes the SSVEP EEG signal and outputs a low-frequency component, specifically:
[0020] Signal decomposition processing uses four filters to decompose the SSVEP EEG signal into a low-frequency component and three high-frequency components. The four filters are expressed as follows:
[0021]
[0022] Among them, f LL is the low-pass filter used, f LH 、f HL 、f HH These are the different high-pass filters used.
[0023] The channel attention mechanism processing and the spatial attention mechanism processing in S3 are specifically as follows: the local time domain change features are processed in parallel by global average pooling and maximum pooling, and then the outputs of the global pooling and the maximum pooling are processed by a multi-layer perceptron respectively, and then spliced to obtain a first spliced feature, and the channel attention weight is generated by the activation function for the first spliced feature, and the channel attention weight is multiplied element by element by the local time domain change feature to obtain a channel enhancement feature;
[0024] The channel enhancement features are processed by global average pooling and maximum pooling in turn, and the outputs of the global average pooling and maximum pooling are spliced to obtain the second splicing feature. The second splicing feature is convolved and the spatial attention weight is generated by the activation function. The spatial attention weight is multiplied element by element with the channel enhancement feature to obtain the spatial enhancement feature.
[0025] In S4, the spatial enhancement feature is subjected to deep convolution processing, specifically, the spatial enhancement feature is subjected to deep convolution processing using an N×1×1 convolution kernel, where N is the number of channels.
[0026] The local spatial features of all channels are subjected to separable convolution processing, specifically, a 1×16 convolution kernel is used to perform separable convolution processing on the local spatial features of all channels.
[0027] In S3, local features are extracted from the multi-scale time-frequency features by two-dimensional convolution processing, specifically, two-dimensional convolution processing is performed on the multi-scale time-frequency features using a 1×16 two-dimensional convolution kernel.
[0028] In the S1, the SSVEP EEG signal is obtained and spatial features are extracted through convolution processing, specifically, the SSVEP EEG signal is convoluted using a 5×5 convolution kernel.
[0029] The present invention also provides a SSVEP electroencephalogram signal feature extraction and classification system, comprising:
[0030] EEG signal feature extraction module: used to obtain SSVEP EEG signals, perform spatial feature extraction through convolution processing, obtain spatial features, perform time-frequency feature extraction on SSVEP EEG signals, and obtain reconstructed low-frequency features;
[0031] The time-frequency feature extraction process is as follows: performing signal decomposition processing on the SSVEP EEG signal for several times in sequence, the signal decomposition processing decomposes the SSVEP EEG signal into a low-frequency component, and the low-frequency component is used as the input of the next signal decomposition processing. The last signal decomposition processing outputs a low-frequency component and several high-frequency components. All low-frequency components and several high-frequency components obtained by the above-mentioned several signal decomposition processings are sequentially subjected to channel number expansion, time-frequency feature extraction, feature scaling, and reshaping processing to obtain several low-frequency features and several high-frequency features. Based on the low-frequency features and several high-frequency features obtained by the last signal decomposition processing and all low-frequency features obtained by the previous several signal decomposition processings, a reconstructed low-frequency feature is obtained;
[0032] Multi-scale time-frequency feature acquisition module: used to add spatial features and reconstructed low-frequency features to obtain integrated features, and perform batch normalization, activation function processing and random inactivation processing on the integrated features in turn to obtain multi-scale time-frequency features;
[0033] Feature enhancement module: It is used to extract local features from multi-scale time-frequency features through two-dimensional convolution and batch normalization to obtain local time domain change features. The local time domain change features are processed by channel attention mechanism and spatial attention mechanism in turn to obtain spatial enhancement features.
[0034] Feature fusion module: used to perform deep convolution processing on the spatial enhancement features to obtain the local spatial features of each channel, and perform separable convolution processing on the local spatial features of all channels to obtain spatial domain fusion features;
[0035] EEG signal classification module: used to flatten the spatial domain fusion features, and then map the spatial domain fusion features to the classification space containing different visual stimulation frequency categories through linear transformation processing, and use the classification function to obtain the probability distribution of different visual stimulation frequencies to obtain the SSVEP EEG signal classification results.
[0036] In addition, the present invention provides an SSVEP EEG signal feature extraction and classification device, comprising a processor and a memory, wherein the processor implements the SSVEP EEG signal feature extraction and classification method as described above when executing a computer program stored in the memory.
[0037] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the steps of the SSVEP electroencephalogram signal feature extraction and classification method as described above.
[0038] The beneficial effects of the present invention are:
[0039] 1. The present invention provides a SSVEP EEG signal feature extraction and classification method, which enhances the ability to capture the multi-scale time-frequency features of SSVEP EEG signals by performing several signal decomposition processes, time-frequency feature extraction and wavelet inverse transform processes on SSVEP EEG signals in sequence, and more comprehensively characterizes the EEG response characteristics of different subjects.
[0040] 2. The present invention performs local feature extraction on multi-scale time-frequency features through two-dimensional convolution processing and expands the number of channels, providing more channel information for subsequent processing. Through channel attention mechanism processing and spatial attention mechanism processing, the weights of significant features in the signal are adaptively allocated, noise signals are effectively filtered, and the separability of the signal is improved.
[0041] 3. The present invention extracts the local spatial features of each channel through deep convolution processing, fuses all features through separable convolution processing to obtain spatial domain fusion features, and after flattening and linear transformation processing, uses a classification function to obtain the EEG signal classification result, which effectively alleviates the impact of individual differences on BCIs performance. DETAILED DESCRIPTION
[0042] Exemplary embodiments of the present disclosure are described in more detail below.
[0043] Example
[0044] This embodiment provides a method for extracting and classifying SSVEP electroencephalogram signal features, comprising the following steps:
[0045] S1, obtain SSVEP EEG signal, extract spatial features through convolution processing, obtain spatial features, extract time-frequency features of SSVEP EEG signal, and obtain reconstructed low-frequency features;
[0046] The time-frequency feature extraction process is as follows: the SSVEP EEG signal is subjected to several signal decomposition processes in sequence, wherein the signal decomposition process decomposes the SSVEP EEG signal to output a low-frequency component, and the low-frequency component is used as the input of the next signal decomposition process, and the last signal decomposition process outputs a low-frequency component and several high-frequency components.
[0047] All low-frequency components and several high-frequency components obtained by the above-mentioned several signal decomposition processes are sequentially subjected to channel number expansion, time-frequency feature extraction, feature scaling, and reshaping processing to obtain several low-frequency features and several high-frequency features.
[0048] Based on the low-frequency features obtained from the last signal decomposition process and several high-frequency features, and all low-frequency features obtained from several previous signal decomposition processes, a reconstructed low-frequency feature is obtained.
[0049] Specifically, in step S1, the acquired SSVEP EEG signal is subjected to convolution processing to extract spatial features, specifically, a 5×5 convolution kernel is used to perform convolution processing on the SSVEP EEG signal to extract and obtain spatial features.
[0050] At the same time, the time-frequency features of the acquired SSVEP EEG signals are extracted, and the SSVEP EEG signals are subjected to several signal decomposition processes in sequence. The signal decomposition process decomposes the SSVEP EEG signals into a low-frequency component, which is specifically:
[0051] Signal decomposition processing uses four filters to decompose the SSVEP EEG signal into a low-frequency component and three high-frequency components. The four filters are expressed as follows:
[0052]
[0053] Among them, f LL is the low-pass filter used, f LH 、f HL 、f HH These are the different high-pass filters used.
[0054] The obtained low-frequency component is further subjected to the next signal decomposition processing, and the final signal decomposition processing outputs a low-frequency component and several high-frequency components.
[0055] Based on the low-frequency features obtained from the last signal decomposition processing and several high-frequency features, and all the low-frequency features obtained from the previous signal decomposition processing, the specific process is: starting from the last signal decomposition processing, a low-frequency feature and all the high-frequency features obtained are connected along the channel dimension, and then processed by inverse wavelet transform to obtain a first reconstructed low-frequency feature, and the first reconstructed low-frequency feature is connected with the low-frequency feature obtained from the previous signal decomposition processing, and then processed by inverse wavelet transform to obtain a second reconstructed low-frequency feature, and the above process is repeated for the second reconstructed low-frequency feature until the nth reconstructed low-frequency feature is connected with the low-frequency feature obtained from the first signal decomposition processing, and then processed by inverse wavelet transform to obtain the reconstructed low-frequency feature.
[0056] In this embodiment, when performing time-frequency feature extraction, the SSVEP EEG signal is set to be subjected to two signal decomposition processes in sequence. The first signal decomposition process uses four filters to decompose the input SSVEP EEG signal into a low-frequency (LL) component and three high-frequency (LH, HL, HH) components, and the spatial dimension (height and width) of each component is half of the input SSVEP EEG signal. The second signal decomposition process repeats the above process for the low-frequency (LL) component output by the first signal decomposition process, and reduces the spatial dimension of each component obtained to one-fourth of the size of the input SSVEP EEG signal. A total of two signal decomposition processes were performed.
[0057] Starting from the second signal decomposition process, the inverse wavelet transform process is performed in sequence. The specific process is: a low-frequency feature and all high-frequency features output after the second signal decomposition process are connected along the channel dimension, and then the connected features are processed by inverse wavelet transform to obtain the first reconstructed low-frequency feature. The first reconstructed low-frequency feature is connected with the low-frequency feature obtained by the first signal decomposition process, and the connected features are also processed by inverse wavelet transform to obtain the reconstructed low-frequency feature.
[0058] S2. Add the spatial features and the reconstructed low-frequency features to obtain integrated features. The integrated features are subjected to batch normalization, activation function processing and random inactivation processing in sequence to obtain multi-scale time-frequency features.
[0059] In this step, the spatial features and reconstructed low-frequency features obtained in step S1 are added to obtain integrated features to integrate multi-level information, and the integrated features are further subjected to batch normalization, activation function processing, and random dropout processing (Dropout) in sequence to extract multi-scale time-frequency features. The activation function used in step S2 is specifically the ELU activation function.
[0060] S3. After two-dimensional convolution processing and batch normalization processing, the multi-scale time-frequency features are subjected to local feature extraction and batch normalization processing to obtain local time domain change features. The local time domain change features are processed by channel attention mechanism and spatial attention mechanism in turn to obtain spatial enhanced features.
[0061] Specifically, two-dimensional convolution is performed on the multi-scale time-frequency features to extract local features, specifically, two-dimensional convolution is performed on the multi-scale time-frequency features using a 1×16 two-dimensional convolution kernel.
[0062] The obtained local time domain change features are further processed by channel attention mechanism and spatial attention mechanism in sequence. Specifically, the local time domain change features are processed in parallel by global average pooling and maximum pooling respectively, and then the outputs of global pooling and maximum pooling are processed by multi-layer perceptron respectively, and then spliced to obtain the first spliced feature, and the channel attention weight is generated for the first spliced feature through the activation function, and the channel attention weight is multiplied element by element with the local time domain change feature to obtain the channel enhanced feature.
[0063] The channel enhancement features are processed by global average pooling and maximum pooling in turn, and the outputs of the global average pooling and maximum pooling are spliced to obtain the second splicing feature. The second splicing feature is convolved and the spatial attention weight is generated by the activation function. The spatial attention weight is multiplied element by element with the channel enhancement feature to obtain the spatial enhancement feature.
[0064] The specific process formula is as follows:
[0065] M C (F) = σ(MLP[G Avg (F)]+MLP[G Max (F)]),
[0066] M S (F) = σ((f 7×7 (Concat([G Avg (F)],[G Max (F)]))),
[0067] Among them, M C (F) is the channel enhancement feature, M S (F) is the spatial enhancement feature, F is the local temporal variation feature of the input, σ is the activation function, G Avg , G Max They are global average pooling and maximum pooling, MLP is a multi-layer perceptron, Concat is concatenation, and f 7×7 It is a 7×7 convolution.
[0068] In this step, the activation function used is specifically the sigmoid activation function.
[0069] By sequentially processing the local time domain change features with the channel attention mechanism and the spatial attention mechanism, the spatial enhanced features are obtained, which can adaptively highlight the key time-frequency features related to the SSVEP EEG signals, enhance the feature expression ability, and improve the classification performance and generalization ability.
[0070] S4. Perform deep convolution processing on the spatial enhancement features to obtain the local spatial features of each channel, and perform separable convolution processing on the local spatial features of all channels to obtain spatial domain fusion features.
[0071] In step S4, deep convolution processing is performed on the spatial enhancement features, specifically, deep convolution processing is performed on the spatial enhancement features using an N×1×1 convolution kernel, where N is the number of channels.
[0072] The local spatial features of all channels are subjected to separable convolution processing, specifically, a 1×16 convolution kernel is used to perform separable convolution processing on the local spatial features of all channels.
[0073] By extracting data spatial features through deep convolution processing and then fusing all feature information through separable convolution processing, the feature representation capability can be improved and the comprehensiveness and accuracy of spatial domain fusion features can be ensured.
[0074] S5. Flatten the spatial domain fusion features, and then map the spatial domain fusion features to the classification space containing different visual stimulation frequency categories through linear transformation. Use the classification function to obtain the probability distribution of different visual stimulation frequencies to obtain the SSVEP EEG signal classification results.
[0075] The classification function used in step S5 is a Softmax function, which can generate probability distributions of different visual stimulus frequencies.
[0076] The present invention also provides a SSVEP electroencephalogram signal feature extraction and classification system, comprising:
[0077] EEG signal feature extraction module: used to obtain SSVEP EEG signals, perform spatial feature extraction through convolution processing, obtain spatial features, perform time-frequency feature extraction on SSVEP EEG signals, and obtain reconstructed low-frequency features;
[0078] The time-frequency feature extraction process is as follows: performing signal decomposition processing on the SSVEP EEG signal for several times in sequence, the signal decomposition processing decomposes the SSVEP EEG signal into a low-frequency component, and the low-frequency component is used as the input of the next signal decomposition processing. The last signal decomposition processing outputs a low-frequency component and several high-frequency components. All low-frequency components and several high-frequency components obtained by the above-mentioned several signal decomposition processings are sequentially subjected to channel number expansion, time-frequency feature extraction, feature scaling, and reshaping processing to obtain several low-frequency features and several high-frequency features. Based on the low-frequency features and several high-frequency features obtained by the last signal decomposition processing and all low-frequency features obtained by the previous several signal decomposition processings, a reconstructed low-frequency feature is obtained;
[0079] Multi-scale time-frequency feature acquisition module: used to add spatial features and reconstructed low-frequency features to obtain integrated features, and perform batch normalization, activation function processing and random inactivation processing on the integrated features in turn to obtain multi-scale time-frequency features;
[0080] Feature enhancement module: It is used to extract local features from multi-scale time-frequency features through two-dimensional convolution and batch normalization to obtain local time domain change features. The local time domain change features are processed by channel attention mechanism and spatial attention mechanism in turn to obtain spatial enhancement features.
[0081] Feature fusion module: used to perform deep convolution processing on the spatial enhancement features to obtain the local spatial features of each channel, and perform separable convolution processing on the local spatial features of all channels to obtain spatial domain fusion features;
[0082] EEG signal classification module: used to flatten the spatial domain fusion features, and then map the spatial domain fusion features to the classification space containing different visual stimulation frequency categories through linear transformation processing, and use the classification function to obtain the probability distribution of different visual stimulation frequencies to obtain the SSVEP EEG signal classification results.
[0083] In addition, the present invention also provides an SSVEP EEG signal feature extraction and classification device, comprising a processor and a memory, wherein the processor implements the SSVEP EEG signal feature extraction and classification method as described above when executing a computer program stored in the memory.
[0084] Finally, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the SSVEP EEG signal feature extraction and classification method as described above.
[0085] Experimental content
[0086] In order to verify the effectiveness and versatility of the present invention, a cross-subject comparative experiment was conducted on a public dataset benchmark dataset (BenchmarkDataset).
[0087] The specific data sets and experimental results used are described as follows:
[0088] The Benchmark Dataset dataset collects SSVEP-BCI recordings of 35 healthy subjects (17 females, aged 17-34 years, average age: 22 years), focusing on 40 characters flashing at different frequencies (8-15.8Hz, interval 0.2Hz).
[0089] The experimental results are shown in Table 1, which specifically compares the method of the present invention with canonical correlation analysis (CCA), filter bank-based canonical correlation analysis (FBCCA), convolutional neural network (CNN), and electroencephalogram signal network (EEGNet) on the BenchmarkDataset dataset. The average and variance of the classification accuracy of all subjects and the average kappa coefficient (kappa) value of all subjects are presented in Table 1. The results show that the method of the present invention outperforms other methods and has good accuracy and robustness in the classification of SSVEP cross-subject EEG signals.
[0090] Table 1 Experimental results
[0091]
[0092]
[0093] The performance of the method of the present invention is significantly improved compared with other advanced algorithms, with the highest average accuracy and kappa value, and the smallest variance. These results show that the method of the present invention can achieve high classification performance while having better robustness.
[0094] The present invention enhances the ability to capture multi-scale time-frequency features of SSVEP electroencephalogram signals through the combined effects of two discrete wavelet transforms, inverse wavelet transforms and convolution, more comprehensively characterizes the electroencephalogram response characteristics of different subjects, extracts local features through two-dimensional convolution and expands the number of channels to provide more channel information for the follow-up, adaptively allocates significant feature weights in the signal through channel attention mechanism processing and spatial attention mechanism processing, extracts spatial domain features through deep convolution, and fuses all features through separable convolution, outputs classification results through flattening processing and classification function, and effectively alleviates the impact of individual differences on brain-computer interface performance.
Claims
1. A SSVEP electroencephalogram signal feature extraction and classification method, characterized in that: The following steps are involved: S1, obtain SSVEP EEG signal, extract spatial features through convolution processing, obtain spatial features, extract time-frequency features of SSVEP EEG signal, and obtain reconstructed low-frequency features; The time-frequency feature extraction process is as follows: the SSVEP EEG signal is subjected to a plurality of signal decomposition processes in sequence, wherein the signal decomposition process decomposes the SSVEP EEG signal to output a low-frequency component, and the low-frequency component is used as the input of the next signal decomposition process, and the last signal decomposition process outputs a low-frequency component and a plurality of high-frequency components; All low-frequency components and several high-frequency components obtained by the above-mentioned several signal decomposition processes are sequentially subjected to channel number expansion, time-frequency feature extraction, feature scaling, and reshaping processing to obtain several low-frequency features and several high-frequency features. Based on the low-frequency features obtained from the last signal decomposition process and several high-frequency features, and all low-frequency features obtained from several previous signal decomposition processes, a reconstructed low-frequency feature is obtained; S2, adding the spatial features and the reconstructed low-frequency features to obtain integrated features, and performing batch normalization, activation function processing and random inactivation processing on the integrated features in turn to obtain multi-scale time-frequency features; S3, after two-dimensional convolution processing of multi-scale time-frequency features to extract local features and batch normalization processing, obtain local time domain change features, The local time domain change features are processed by channel attention mechanism and spatial attention mechanism in turn to obtain spatial enhanced features; S4, performing deep convolution processing on the spatial enhancement features to obtain the local spatial features of each channel, and performing separable convolution processing on the local spatial features of all channels to obtain spatial domain fusion features; S5. Flatten the spatial domain fusion features, and then map the spatial domain fusion features to the classification space containing different visual stimulation frequency categories through linear transformation. Use the classification function to obtain the probability distribution of different visual stimulation frequencies to obtain the SSVEP EEG signal classification results.
2. A SSVEP electroencephalogram signal feature extraction and classification method according to claim 1, characterized in that: The specific process of S1 based on the low-frequency features obtained from the last signal decomposition process and several high-frequency features, and all low-frequency features obtained from the previous signal decomposition processes is as follows: Starting from the last signal decomposition process, a low-frequency feature and all high-frequency features are connected along the channel dimension, and then the first reconstructed low-frequency feature is obtained by inverse wavelet transform. The first reconstructed low-frequency feature is connected with the low-frequency feature obtained by the previous signal decomposition process, and then the inverse wavelet transform is performed to obtain the second reconstructed low-frequency feature. The above process is repeated for the second reconstructed low-frequency feature until the nth reconstructed low-frequency feature is connected with the low-frequency feature obtained by the first signal decomposition process, and then the inverse wavelet transform is performed to obtain the reconstructed low-frequency feature.
3. A SSVEP electroencephalogram signal feature extraction and classification method according to claim 1, characterized in that: The signal decomposition process in S1 decomposes the SSVEP EEG signal and outputs a low-frequency component, specifically: Signal decomposition processing uses four filters to decompose the SSVEP EEG signal into a low-frequency component and three high-frequency components. The four filters are expressed as follows: Among them, f LL is the low-pass filter used, f LH 、f HL 、f HH These are the different high-pass filters used.
4. A SSVEP electroencephalogram signal feature extraction and classification method according to claim 1, characterized in that: The channel attention mechanism processing and the spatial attention mechanism processing in S3 are specifically as follows: the local time domain change features are processed in parallel by global average pooling and maximum pooling, and then the outputs of the global pooling and the maximum pooling are processed by a multi-layer perceptron respectively, and then spliced to obtain a first spliced feature, and the channel attention weight is generated by the activation function for the first spliced feature, and the channel attention weight is multiplied element by element by the local time domain change feature to obtain a channel enhancement feature; The channel enhancement features are processed by global average pooling and maximum pooling in turn, and the outputs of the global average pooling and maximum pooling are spliced to obtain the second splicing feature. The second splicing feature is convolved and the spatial attention weight is generated by the activation function. The spatial attention weight is multiplied element by element with the channel enhancement feature to obtain the spatial enhancement feature.
5. The SSVEP electroencephalogram signal feature extraction and classification method according to claim 1, characterized in that: In S4, the spatial enhancement feature is subjected to deep convolution processing, specifically, the spatial enhancement feature is subjected to deep convolution processing using an N×1×1 convolution kernel, where N is the number of channels. The local spatial features of all channels are subjected to separable convolution processing, specifically, a 1×16 convolution kernel is used to perform separable convolution processing on the local spatial features of all channels.
6. A SSVEP electroencephalogram signal feature extraction and classification method according to claim 1, characterized in that: In S3, local features are extracted from the multi-scale time-frequency features by two-dimensional convolution processing, specifically, two-dimensional convolution processing is performed on the multi-scale time-frequency features using a 1×16 two-dimensional convolution kernel.
7. A SSVEP electroencephalogram signal feature extraction and classification method according to claim 1, characterized in that: In the S1, the SSVEP EEG signal is obtained and spatial features are extracted through convolution processing, specifically, the SSVEP EEG signal is convoluted using a 5×5 convolution kernel.
8. A SSVEP electroencephalogram signal feature extraction and classification system, characterized in that: include: EEG signal feature extraction module: used to obtain SSVEP EEG signals, perform spatial feature extraction through convolution processing, obtain spatial features, perform time-frequency feature extraction on SSVEP EEG signals, and obtain reconstructed low-frequency features; The time-frequency feature extraction process is as follows: performing signal decomposition processing on the SSVEP EEG signal for several times in sequence, the signal decomposition processing decomposes the SSVEP EEG signal into a low-frequency component, and the low-frequency component is used as the input of the next signal decomposition processing. The last signal decomposition processing outputs a low-frequency component and several high-frequency components. All low-frequency components and several high-frequency components obtained by the above-mentioned several signal decomposition processings are sequentially subjected to channel number expansion, time-frequency feature extraction, feature scaling, and reshaping processing to obtain several low-frequency features and several high-frequency features. Based on the low-frequency features and several high-frequency features obtained by the last signal decomposition processing and all low-frequency features obtained by the previous several signal decomposition processings, a reconstructed low-frequency feature is obtained; Multi-scale time-frequency feature acquisition module: used to add spatial features and reconstructed low-frequency features to obtain integrated features, and perform batch normalization, activation function processing and random inactivation processing on the integrated features in turn to obtain multi-scale time-frequency features; Feature enhancement module: It is used to extract local features from multi-scale time-frequency features through two-dimensional convolution and batch normalization to obtain local time domain change features. The local time domain change features are processed by channel attention mechanism and spatial attention mechanism in turn to obtain spatial enhancement features. Feature fusion module: used to perform deep convolution processing on the spatial enhancement features to obtain the local spatial features of each channel, and perform separable convolution processing on the local spatial features of all channels to obtain spatial domain fusion features; EEG signal classification module: used to flatten the spatial domain fusion features, and then map the spatial domain fusion features to the classification space containing different visual stimulation frequency categories through linear transformation processing, and use the classification function to obtain the probability distribution of different visual stimulation frequencies to obtain the SSVEP EEG signal classification results.
9. A SSVEP electroencephalogram signal feature extraction and classification device, characterized in that: It comprises a processor and a memory, wherein when the processor executes the computer program stored in the memory, it implements the SSVEP electroencephalogram signal feature extraction and classification method as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that: Used to store a computer program, wherein when the computer program is executed by a processor, the SSVEP electroencephalogram signal feature extraction and classification method as described in any one of claims 1 to 7 is implemented.