Emotion recognition method based on multi-band adaptive graph convolution

Abstract

The invention discloses an emotion recognition method based on multi-band adaptive graph convolution, and the method comprises the steps: introducing an adaptive topological graph which carries out the self-adaption according to a sample, and generating a time smooth topological graph in combination with a memory attenuation mechanism, thereby carrying out the smoothing and constraint of the adaptive topological graph, and reducing the impact on topology estimation from instantaneous noise disturbance. A frequency band topological graph is constructed based on the joint constraint of a prior topological graph and a frequency band offset topological graph, and on the basis of keeping a certain frequency band sharing structure constraint, depiction of frequency band difference connection is introduced, so that the modeling complexity and uncertainty brought by learning a complete dense topology for each frequency band are reduced. A fusion strategy for multi-band representation is provided, adaptive adjustment is carried out on information contribution of different bands, the repetitive effect of redundant bands is inhibited to a certain extent, information collaboration of complementary bands is enhanced, and therefore more stable cross-band representation support is provided for emotion recognition.

Description

Technical Field

[0001] This invention relates to the technical field of EEG signal decoding, and in particular to an emotion recognition method based on multi-band adaptive graph convolution. Background Technology

[0002] Due to its high temporal resolution and non-invasive acquisition, EEG signals have application value in fields such as emotion recognition, brain-computer interfaces, and neural state assessment. However, EEG signals have a low signal-to-noise ratio, and there are significant individual differences and distribution shifts among different subjects, which can easily lead to decreased generalization performance and insufficient stability in cross-subject emotion recognition tasks.

[0003] Existing technologies often employ convolutional neural networks (CNNs) or recurrent neural networks (RNNs) to model multi-channel signals or their frequency domain features, but they are insufficient in modeling the non-Euclidean topological dependencies between channels, making it difficult to characterize brain region functional connectivity patterns. In recent years, graph convolutional neural networks (GCNNs) have provided a new modeling method for EEG emotion recognition by constructing adjacency matrices to represent the functional connectivity relationships between channels. However, the adjacency matrices of existing graph methods are mostly statically defined based on priors such as electrode spatial distance or fixed coherence, lacking the ability to adaptively learn dynamic functional connectivity under task-driven conditions. Furthermore, the connection structure may contain many redundant edge weights, potentially introducing noise propagation and increasing the risk of overfitting. In addition, multi-band information is usually obtained through simple splicing or independent weighted fusion, making it difficult to effectively distinguish the contributions of redundant and complementary frequency bands. Summary of the Invention

[0004] The purpose of this invention is to provide an emotion recognition method based on multi-band adaptive graph convolution. By constructing a frequency-band time-varying topology graph, it achieves joint modeling of the spatial topology and temporal dynamic changes of EEG signals. The dynamic graph topology is updated through a memory decay mechanism, and a frequency band fusion strategy is introduced to suppress redundant frequency bands and enhance the synergistic contribution of complementary frequency bands, thereby achieving more effective modeling and fusion of the spatiotemporal topology of EEG signals.

[0005] To achieve the above objectives, the technical solution provided by this invention is: a multi-band adaptive graph convolution emotion recognition method, comprising the following steps:

[0006] S1. Acquire raw multi-channel EEG signals and truncate them according to a preset window length. Each truncation corresponds to one emotion recognition task, resulting in task EEG signals. Segment the task EEG signals according to a preset time step to obtain multi-channel short-time EEG signals at multiple time steps. Filter and divide the multi-channel short-time EEG signals at each time step according to preset frequency bands to obtain multi-frequency band EEG signals. Extract features from the multi-frequency band EEG signals to construct a four-dimensional EEG feature tensor with frequency band, time step, channel, and feature dimensions.

[0007] S2. For the four-dimensional EEG feature tensor, the two-dimensional EEG feature tensor corresponding to each frequency band and each time step is extracted sequentially. An adaptive topology graph is constructed based on the two-dimensional EEG feature tensor to characterize the instantaneous functional connectivity between channels of the brain network at that frequency band and time step. A memory decay mechanism is introduced, that is, the adaptive topology graphs generated at different time steps are weighted and fused to obtain a time-smoothed topology graph, thereby suppressing instantaneous noise and maintaining temporal consistency. A prior topology graph shared across frequency bands is constructed, and a frequency band bias topology graph is learned for each frequency band to characterize the differences in connectivity between frequency bands. The two are fused to obtain a frequency band topology graph. The time-smoothed topology graph and the frequency band topology graph are fused to obtain the frequency band time-varying topology graph of the corresponding frequency band at each time step. The frequency band time-varying topology graphs of each time step are stacked in chronological order to form a sequence of frequency band time-varying topology graphs for that frequency band.

[0008] S3. Using the three-dimensional EEG feature tensor corresponding to the "time step-channel-feature" of each frequency band of the four-dimensional EEG feature tensor as input, and the frequency band time-varying topology sequence of the corresponding frequency band as constraints, frequency band features with fused spatiotemporal information are extracted through spatiotemporal graph convolution to form frequency band feature vectors for each frequency band; a frequency band similarity matrix is ​​constructed, and based on the frequency band similarity matrix, a redundancy suppression matrix for suppressing information redundancy and a complementary enhancement matrix for enhancing information complementarity are constructed respectively; a frequency band coupling matrix is ​​constructed, and based on the frequency band coupling matrix combined with the redundancy suppression matrix and the complementary enhancement matrix, a frequency band weight vector is obtained; the frequency band feature vectors of each frequency band are weighted and summed based on the frequency band weight vector to obtain the final fused EEG feature vector; the fused EEG feature vector is input into a classifier constructed by a fully connected layer, and the emotion category prediction result is output.

[0009] Furthermore, in step S1, let the window length be... The acquired EEG signal for the task is represented as follows:

[0010] ;

[0011] In the formula, the symbol The symbol "belongs to" indicates that the variable preceding the symbol belongs to a set following it. Represents the set of real numbers. Indicates the number of channels. This represents the number of sampling points. If the sampling frequency is... ,but ;

[0012] The task EEG signals are processed according to a preset time step. Divide the time into segments of seconds to obtain the number of time steps. The multi-channel short-time EEG signals at each time step are filtered and divided channel by channel according to a preset frequency band to obtain multi-frequency EEG signals. :

[0013] ;

[0014] In the formula, The number of frequency bands. This represents the filtered EEG signal in the first frequency band. This represents the filtered EEG signal in the second frequency band. Indicates the first Filtered EEG signals in each frequency band, Indicates the first Filtered EEG signals in each frequency band;

[0015] Subsequently, multi-band EEG signals were analyzed. Feature extraction was performed to construct a four-dimensional EEG feature tensor with frequency band, time step, channel, and feature as dimensions. ,in Indicates the number of features.

[0016] Furthermore, step S2 includes the following steps:

[0017] S21, Targeting the four-dimensional EEG feature tensor Take out each frequency band in sequence With time step The corresponding two-dimensional EEG feature tensor of "channel-feature" and respectively through and Get the query matrix AND key matrix ,in and The weight matrix is ​​a learnable linear mapping. This indicates the transpose operation. express and Feature dimensions; ultimately generating an adaptive topology graph. To characterize the instantaneous functional connectivity between channels of the brain network at this frequency band and time step:

[0018] ;

[0019] In the formula, Represents a normalized mapping, symbol This indicates element-wise multiplication. A mask matrix constructed based on ROI or geometric rules, used to shield node pairs that do not conform to the preset connection rules;

[0020] S22. Introduce a memory decay mechanism to obtain a time-smoothed topology graph based on an adaptive topology graph. That is, for frequency bands In time step hour, In time step At that time, the current time step Adaptive topology graph Compared with the previous time step Time-smoothed topology graph Weighted fusion yields the current time step. Time-smoothed topology graph This suppresses transient noise and maintains timing consistency.

[0021] ;

[0022] In the formula, frequency band The memory coefficient, the value of which depends on the frequency band;

[0023] S23. Using EEG channels as graph nodes, and comprehensively considering brain region ROI division, channel geometric proximity, and left-right hemisphere symmetry, construct a cross-frequency band shared prior topology graph. The connection weight between channels within the same ROI is higher than the connection weight between channels in different ROIs. Based on this, to characterize the connection differences across different frequency bands, for each frequency band... Set frequency band offset topology diagram and to Sparse constraints are applied during learning to suppress redundant connections and improve topology interpretability; finally, the two are fused to obtain the frequency band. Frequency band topology diagram :

[0024] ;

[0025] S24. Smooth the time topology of each frequency band. Frequency band topology diagram By merging, frequency bands are obtained. Time step Frequency band time-varying topology diagram :

[0026] ;

[0027] Frequency band time-varying topology diagram at each time step By stacking them in chronological order, a sequence of time-varying topology maps for that frequency band can be formed. :

[0028] ;

[0029] In the formula, Indicates frequency band The frequency band time-varying topology diagram of the first time step. Indicates frequency band The frequency band time-varying topology diagram of the second time step. Indicates frequency band , No. Frequency band time-varying topology diagram for each time step. Indicates frequency band , No. Frequency band time-varying topology diagram for each time step.

[0030] Furthermore, step S3 includes the following steps:

[0031] S31. A three-dimensional EEG feature tensor based on the "time step-channel-feature" corresponding to each frequency band of the four-dimensional EEG feature tensor. As input, use the frequency band time-varying topology sequence corresponding to the frequency band. To constrain this, spatiotemporal joint convolution is performed using the spatiotemporal graph convolution model STGCN, and along... and Perform global pooling to generate frequency band feature vectors. ,in, The dimension of the frequency band feature vector;

[0032] S32, will The feature vectors of each frequency band are stacked to obtain a matrix. :

[0033] ;

[0034] In the formula, This represents the frequency band feature vector of the first frequency band. This represents the frequency band feature vector of the second frequency band. Indicates the first Frequency band feature vectors of each frequency band Indicates the first Frequency band feature vectors of each frequency band;

[0035] right Normalize by row to obtain the matrix. Construct a frequency band similarity matrix ,in This represents a column vector with all elements equal to 1; subsequently, a redundancy suppression matrix is ​​constructed based on this frequency band similarity matrix. and complementary enhancement matrix ,in, For the Sigmoid function, The suppression intensity coefficient;

[0036] S33. Constructing the frequency band coupling matrix And combined with the redundancy suppression matrix and complementary enhancement matrix Obtain the frequency band weight vector :

[0037] ;

[0038] In the formula, For the collaboration enhancement coefficient, This represents the weighting coefficient for the first frequency band. This represents the weighting coefficient for the second frequency band. Indicates the first Weighting coefficients for each frequency band Indicates the first Weighting coefficients for each frequency band;

[0039] S34. The feature vectors of each frequency band are weighted and summed to obtain the final fused EEG feature vector. :

[0040] ;

[0041] Then, the EEG feature vectors will be fused. Input a classifier constructed from fully connected layers, output a probability distribution vector of predicted sentiment categories. :

[0042] ;

[0043] In the formula, This represents the mapping of the fully connected layer, used to map the fused EEG feature vectors to the emotion category space; finally, the category with the highest probability is taken as the emotion recognition result.

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

[0045] 1. This invention introduces an adaptive topology graph based on samples and further combines a memory decay mechanism to generate a time-smoothed topology graph, thereby smoothing and constraining the adaptive topology graph, reducing the impact of instantaneous noise disturbances on topology estimation, and improving the stability of topology modeling and the feasibility of model training.

[0046] 2. Based on the joint constraints of the prior topology graph and the frequency band offset topology graph, this invention constructs a frequency band topology graph. While maintaining certain shared structural constraints of frequency bands, it introduces the characterization of frequency band difference connections, thereby reducing the modeling complexity and uncertainty caused by learning the complete dense topology for each frequency band separately, and providing a clearer structural foundation for subsequent cross-frequency band modeling and interpretation.

[0047] 3. This invention proposes a fusion strategy for multi-band representation. By adaptively adjusting the information contribution of different frequency bands, it can suppress the repetitive effect of redundant frequency bands and strengthen the information synergy of complementary frequency bands to a certain extent, thereby providing more stable cross-band representation support for emotion recognition. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the method of the present invention.

[0049] Figure 2 This is a schematic diagram of the construction of a time-varying topology map for a frequency band. Detailed Implementation

[0050] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0051] like Figure 1 and Figure 2 As shown in the figure, this embodiment discloses a multi-band adaptive graph convolution emotion recognition method, the specific details of which are as follows:

[0052] S1: Acquire the raw multi-channel EEG signal, and truncate the raw multi-channel EEG signal according to a preset window length. Each truncation corresponds to one emotion recognition task. Set the window length to 20 seconds, and obtain the task EEG signal:

[0053] ;

[0054] In the formula, the symbol The symbol "belongs to" indicates that the variable preceding the symbol belongs to a set following it. Represents the set of real numbers. Indicates the number of channels, if the sampling rate ,but Indicates the number of sampling points;

[0055] The task EEG signals are processed according to a preset time step. Divide the time into segments of seconds to obtain the number of time steps. The multi-channel short-time EEG signals are then processed, and the multi-channel short-time EEG signals at each time step are filtered and divided channel by channel according to a preset frequency band. If the frequency band is selected as... The specific range of each frequency band is set as follows: Frequency band (0.5-3Hz), Frequency band (4-7Hz) Frequency band (8-13Hz) Frequency band (14-30Hz) Multi-band EEG signals were obtained in the frequency band (31-5Hz). :

[0056] ;

[0057] In the formula, express Filtered EEG signals in the frequency band express Filtered EEG signals in the frequency band express Filtered EEG signals in the frequency band express Filtered EEG signals, express Filtered EEG signals;

[0058] Multi-band EEG signals were then extracted. Using power spectral density and differential entropy density as features, a four-dimensional EEG feature tensor is constructed with frequency band, time step, channel, and feature dimensions. ,in Indicates the number of frequency bands. Indicates the number of time steps. Indicates the number of channels. Indicates the number of features.

[0059] S2: For the four-dimensional EEG feature tensor, the two-dimensional EEG feature tensor corresponding to the "channel-feature" of each frequency band and each time step is extracted sequentially. An adaptive topology graph is constructed based on the two-dimensional EEG feature tensor to characterize the instantaneous functional connectivity between channels of the brain network at that frequency band and time step. A memory decay mechanism is introduced, that is, the adaptive topology graphs generated at different time steps are weighted and fused to obtain a time-smoothed topology graph, thereby suppressing instantaneous noise and maintaining temporal consistency. A prior topology graph shared across frequency bands is constructed, and a frequency band bias topology graph is learned for each frequency band to characterize the differences in connectivity between frequency bands. The two are fused to obtain a frequency band topology graph. The time-smoothed topology graph and the frequency band topology graph are fused to obtain the frequency band time-varying topology graph of the corresponding frequency band at each time step. The frequency band time-varying topology graphs of each time step are stacked in chronological order to form a sequence of frequency band time-varying topology graphs for that frequency band. This includes the following steps:

[0060] S21, Targeting the four-dimensional EEG feature tensor Take out each frequency band in sequence With time step The corresponding two-dimensional EEG feature tensor of "channel-feature" ,like , and If the weight matrix is ​​a learnable linear mapping, then it can be obtained by... and Get the query matrix AND key matrix ,in This represents the transpose operation; the final result is an adaptive topology graph. To characterize the instantaneous functional connectivity between channels of the brain network at this frequency band and time step:

[0061] ;

[0062] In the formula, Represents a normalized mapping, symbol This indicates element-wise multiplication. A mask matrix constructed based on ROI or geometric rules, used to shield node pairs that do not conform to the preset connection rules;

[0063] S22. Introduce a memory decay mechanism to obtain a time-smoothed topology graph based on an adaptive topology graph. That is, for Each frequency band in, such as Frequency band, in time step hour, In time step At that time, the current time step Adaptive topology graph Compared with the previous time step Time-smoothed topology graph Weighted fusion yields the current time step. Time-smoothed topology graph This suppresses transient noise and maintains timing consistency.

[0064] ;

[0065] In the formula, frequency band The memory coefficient;

[0066] S23. Using 62 EEG channels as graph nodes, these channels are divided into ROI subsets according to brain anatomy: frontal lobe, temporal lobe, parietal lobe, occipital lobe, and central region. For channels within the same ROI, a connection weight of 0.8 is set; for channels between adjacent ROIs, a connection weight of 0.3 is set; for electrode pairs symmetrical between the left and right hemispheres, their connection weight is increased by 0.2. A cross-band shared prior topology graph is constructed. Based on this, to characterize the connectivity differences between different frequency bands, each frequency band is designated as a frequency band. Configure frequency band offset topology And the initial value is a random value, then... Apply Norm sparse constraints are learned to suppress redundant connections and improve topology interpretability; finally, the two are fused to obtain the frequency band. Frequency band topology diagram :

[0067] ;

[0068] S24. Smooth the time topology of each frequency band. Frequency band topology diagram Integration, such as for frequency bands Its time step Frequency band time-varying topology diagram for:

[0069] ;

[0070] Frequency band time-varying topology diagram at each time step By stacking them in chronological order, a sequence of time-varying topology maps for that frequency band can be formed. For example, for frequency bands :

[0071] ;

[0072] In the formula, Indicates frequency band The frequency band time-varying topology diagram of the first time step. Indicates frequency band The frequency band time-varying topology diagram of the second time step. Indicates frequency band , No. Frequency band time-varying topology diagram for each time step. Indicates frequency band Frequency band time-varying topology diagram at the 40th time step.

[0073] S3: Using the three-dimensional EEG feature tensor corresponding to the "time step-channel-feature" of each frequency band of the four-dimensional EEG feature tensor as input, and the frequency band time-varying topology sequence of the corresponding frequency band as constraints, frequency band features with fused spatiotemporal information are extracted through spatiotemporal graph convolution to form frequency band feature vectors for each frequency band; constructing a frequency band similarity matrix, and based on the frequency band similarity matrix, constructing a redundancy suppression matrix for suppressing information redundancy and a complementary enhancement matrix for enhancing information complementarity; constructing a frequency band coupling matrix, and obtaining a frequency band weight vector based on the frequency band coupling matrix combined with the redundancy suppression matrix and the complementary enhancement matrix; weighted summing of the frequency band feature vectors of each frequency band based on the frequency band weight vector to obtain the final fused EEG feature vector; inputting the fused EEG feature vector into a classifier constructed by a fully connected layer, and outputting the emotion category prediction result; including the following steps:

[0074] S31. A three-dimensional EEG feature tensor based on the "time step-channel-feature" corresponding to each frequency band of the four-dimensional EEG feature tensor. As input, use the frequency band time-varying topology sequence corresponding to the frequency band. To constrain this, spatiotemporal joint convolution is performed using the spatiotemporal graph convolutional model STGCN. The graph convolutional layer employs Chebyshev multinomial graph convolution, calculating the graph Laplacian matrix based on the time-varying adjacency matrix, and then performing convolution operations on node features based on Chebyshev multinomial expansion to capture spatial channel topological relationships. The temporal convolutional layer uses 1D convolution with a kernel size of 3 and a stride of 1, performing convolution operations on the features output by the graph convolutional layer in the temporal dimension to capture dynamic temporal changes; subsequently, along... and Perform global average pooling to generate frequency band feature vectors. ,in, The dimension of the frequency band feature vector;

[0075] S32, will The feature vectors of these five frequency bands are stacked to obtain a matrix. :

[0076] ;

[0077] In the formula, express Frequency band feature vector, express Frequency band feature vector, express Frequency band feature vector, express Frequency band feature vector, express Frequency band feature vector;

[0078] right Normalize by row to obtain the matrix. Construct a frequency band similarity matrix ,in Indicates all elements are The column vectors; then, based on this frequency band similarity matrix, a redundancy suppression matrix is ​​constructed. and complementary enhancement matrix ,in, For the Sigmoid function, The suppression intensity coefficient;

[0079] S33. Constructing the frequency band coupling matrix And combined with the redundancy suppression matrix and complementary enhancement matrix Obtain the frequency band weight vector :

[0080] ;

[0081] In the formula, For the collaboration enhancement coefficient, express Frequency band weighting coefficients express Frequency band weighting coefficients express Frequency band weighting coefficients express Frequency band weighting coefficients express Frequency band weighting coefficients;

[0082] S34. The feature vectors of each frequency band are weighted and summed to obtain the final fused EEG feature vector. :

[0083] ;

[0084] Then, the EEG feature vectors will be fused. Input a classifier constructed from fully connected layers, output a probability distribution vector of predicted sentiment categories. :

[0085] ;

[0086] In the formula, The mapping of the fully connected layers is used to map the fused EEG feature vectors to the emotion category space. Two fully connected layers can be used. The first fully connected layer has an input dimension of 16 and an output dimension of 32, using the ReLU activation function. The second fully connected layer has an input dimension of 32 and an output dimension of 16, also using the ReLU activation function. Batch normalization is added after each fully connected layer. Finally, the category with the highest probability is taken as the emotion recognition result.

[0087] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A sentiment recognition method using multi-band adaptive graph convolution, characterized in that, Includes the following steps: S1. Collect raw multi-channel EEG signals and truncate them according to the preset window length. Each truncation corresponds to an emotion recognition task, and the task EEG signal is obtained. The task EEG signal is segmented according to a preset time step to obtain multi-channel short-time EEG signal with multiple time steps. The multi-channel short-time EEG signal of each time step is filtered and divided into channels according to a preset frequency band to obtain multi-frequency EEG signal. Feature extraction was performed on multi-band EEG signals to construct a four-dimensional EEG feature tensor with frequency band, time step, channel, and feature as dimensions; S2. For the four-dimensional EEG feature tensor, the two-dimensional EEG feature tensor corresponding to each frequency band and each time step is extracted sequentially. An adaptive topology graph is constructed based on the two-dimensional EEG feature tensor to characterize the instantaneous functional connectivity between channels of the brain network at that frequency band and time step. A memory decay mechanism is introduced, that is, the adaptive topology graphs generated at different time steps are weighted and fused to obtain a time-smoothed topology graph, thereby suppressing instantaneous noise and maintaining temporal consistency. A prior topology graph shared across frequency bands is constructed, and a frequency band bias topology graph is learned for each frequency band to characterize the differences in connectivity between frequency bands. The two are fused to obtain a frequency band topology graph. The time-smoothed topology graph and the frequency band topology graph are fused to obtain the frequency band time-varying topology graph of the corresponding frequency band at each time step. The frequency band time-varying topology graphs of each time step are stacked in chronological order to form a sequence of frequency band time-varying topology graphs for that frequency band. S3. Using the three-dimensional EEG feature tensor corresponding to the "time step-channel-feature" of each frequency band of the four-dimensional EEG feature tensor as input, and the frequency band time-varying topology sequence of the corresponding frequency band as constraints, frequency band features with fused spatiotemporal information are extracted through spatiotemporal graph convolution to form frequency band feature vectors for each frequency band; a frequency band similarity matrix is ​​constructed, and based on the frequency band similarity matrix, a redundancy suppression matrix for suppressing information redundancy and a complementary enhancement matrix for enhancing information complementarity are constructed respectively; a frequency band coupling matrix is ​​constructed, and based on the frequency band coupling matrix, the redundancy suppression matrix and the complementary enhancement matrix are combined to obtain the frequency band weight vector; the frequency band feature vectors of each frequency band are weighted and summed based on the frequency band weight vector to obtain the final fused EEG feature vector; the fused EEG feature vector is input into a classifier constructed by a fully connected layer, and the emotion category prediction result is output.

2. The emotion recognition method based on multi-band adaptive graph convolution according to claim 1, characterized in that, In step S1, let the window length be... The acquired EEG signal for the task is represented as follows: ； In the formula, the symbol The symbol "belongs to" indicates that the variable preceding the symbol belongs to a set following it. Represents the set of real numbers. Indicates the number of channels. This represents the number of sampling points. If the sampling frequency is... ,but ; The task EEG signals are processed according to a preset time step. Divide the time into segments of seconds to obtain the number of time steps. The multi-channel short-time EEG signals at each time step are filtered and divided channel by channel according to a preset frequency band to obtain multi-frequency EEG signals. : ； In the formula, The number of frequency bands. This represents the filtered EEG signal in the first frequency band. This represents the filtered EEG signal in the second frequency band. Indicates the first Filtered EEG signals in each frequency band, Indicates the first Filtered EEG signals in each frequency band; Subsequently, multi-band EEG signals were analyzed. Feature extraction was performed to construct a four-dimensional EEG feature tensor with frequency band, time step, channel, and feature as dimensions. ,in Indicates the number of features.

3. The emotion recognition method based on multi-band adaptive graph convolution according to claim 2, characterized in that, Step S2 includes the following steps: S21, Targeting the four-dimensional EEG feature tensor Take out each frequency band in turn With time step The corresponding "channel-feature" two-dimensional EEG feature tensor and respectively through and Get the query matrix AND key matrix ,in and The weight matrix is ​​a learnable linear mapping. This indicates the transpose operation. express and Feature dimensions; ultimately generating an adaptive topology graph. To characterize the instantaneous functional connectivity between channels of the brain network at this frequency band and time step: ； In the formula, Represents a normalized mapping, symbol This indicates element-wise multiplication. A mask matrix constructed based on ROI or geometric rules, used to shield node pairs that do not conform to the preset connection rules; S22. Introduce a memory decay mechanism to obtain a time-smoothed topology graph based on an adaptive topology graph. That is, for frequency bands In time step hour, In time step At that time, the current time step Adaptive topology graph Compared with the previous time step Time-smoothed topology graph Weighted fusion yields the current time step. Time-smoothed topology graph This suppresses transient noise and maintains timing consistency. ； In the formula, frequency band The memory coefficient, the value of which depends on the frequency band; S23. Using EEG channels as graph nodes, and comprehensively considering brain region ROI division, channel geometric proximity, and left-right hemisphere symmetry, construct a cross-frequency band shared prior topology graph. The connection weight between channels within the same ROI is higher than the connection weight between channels in different ROIs. Based on this, to characterize the connection differences across different frequency bands, for each frequency band... Set frequency band offset topology diagram and to Sparse constraints are applied during learning to suppress redundant connections and improve topology interpretability; finally, the two are fused to obtain the frequency band. Frequency band topology diagram : ； S24. Smooth the time topology of each frequency band. Frequency band topology diagram By merging, frequency bands are obtained. Time step Frequency band time-varying topology diagram : ； Frequency band time-varying topology diagram at each time step By stacking them in chronological order, a sequence of time-varying topology maps for that frequency band can be formed. : ； In the formula, Indicates frequency band The frequency band time-varying topology diagram of the first time step. Indicates frequency band The frequency band time-varying topology diagram of the second time step. Indicates frequency band , No. Frequency band time-varying topology diagram for each time step. Indicates frequency band , No. Frequency band time-varying topology diagram for each time step.

4. The emotion recognition method based on multi-band adaptive graph convolution according to claim 3, characterized in that, Step S3 includes the following steps: S31. A three-dimensional EEG feature tensor based on the "time step-channel-feature" corresponding to each frequency band of the four-dimensional EEG feature tensor. As input, use the frequency band time-varying topology sequence corresponding to the frequency band. To constrain this, spatiotemporal joint convolution is performed using the spatiotemporal graph convolution model STGCN, and along... and Perform global pooling to generate frequency band feature vectors. ,in, The dimension of the frequency band feature vector; S32, will The feature vectors of each frequency band are stacked to obtain a matrix. : ； In the formula, This represents the frequency band feature vector of the first frequency band. This represents the frequency band feature vector of the second frequency band. Indicates the first Frequency band feature vectors of each frequency band Indicates the first Frequency band feature vectors of each frequency band; right Normalize by row to obtain the matrix. Construct a frequency band similarity matrix ,in This represents a column vector with all elements equal to 1; subsequently, a redundancy suppression matrix is ​​constructed based on this frequency band similarity matrix. and complementary enhancement matrix ,in, For the Sigmoid function, The suppression intensity coefficient; S33. Constructing the frequency band coupling matrix And combined with the redundancy suppression matrix and complementary enhancement matrix Obtain the frequency band weight vector : ； In the formula, For the collaboration enhancement coefficient, This represents the weighting coefficient for the first frequency band. This represents the weighting coefficient for the second frequency band. Indicates the first Weighting coefficients for each frequency band Indicates the first Weighting coefficients for each frequency band; S34. The feature vectors of each frequency band are weighted and summed to obtain the final fused EEG feature vector. : ； Then, the EEG feature vectors will be fused. Input a classifier constructed from fully connected layers, output a probability distribution vector of predicted sentiment categories. : ； In the formula, This represents the mapping of the fully connected layer, used to map the fused EEG feature vectors to the emotion category space; finally, the category with the highest probability is taken as the emotion recognition result.

Images