Physiological signal channel aggregation method and system for graph neural network

Abstract

The invention provides a physiological signal channel aggregation method and system for a graph neural network. The method comprises the steps of obtaining channel node features and an adjacent matrix of an upstream task; calculating the similarity between channel nodes, initializing a sorting matrix, and carrying out row and column normalization on the sorting matrix; sorting the channel node features and the adjacent matrixes by using a sorting matrix; the method comprises the following steps: grouping channels, constructing mapping, and converting a sequenced adjacent matrix into a mapping matrix; aggregating the sorted channel node features and the mapping matrix; and constructing a regional aggregation loss function to constrain the adjacent matrix. The technical problems that universality is poor, functional similarity changes between channels are difficult to dynamically capture, knowledge in the preset field is relied on, and capture features are single are solved.

Description

Technical Field

[0001] The present invention relates to the field of physiological signal processing, and in particular to a method and system for aggregating physiological signal channels in a graph neural network. Background Art

[0002] Graph neural networks, due to their ability to process non-Euclidean data, have become a powerful tool for extracting physiological signal features. Existing methods predefine static partitioning methods for the output of graph neural networks. This static partitioning method ignores the dynamic interactions between physiological signal channels, thereby weakening the ability to identify individual differences between physiological signals. Moreover, the regional boundaries defined by the static partitioning method are fixed, ignoring the functional overlap between regions formed during the dynamic interaction of physiological signals, thereby limiting the model's ability to analyze the coordination mechanism between regions.

[0003] The existing invention patent application document, "A Brain Network Representation Learning Method Based on a Self-Attention Dynamic Graph Neural Network," with publication number CN115146668A, includes the following steps: first, performing a short-time Fourier transform on the EEG signal to obtain its time-frequency characteristics; then, modeling the multi-channel EEG features and using the topological structure of the EEG signal to learn a more discriminative EEG feature representation; and finally, using a long short-term memory network to extract the temporal dynamic evolution characteristics of the brain network spatial feature representation. However, the aforementioned existing solution, based on a self-attention dynamic graph neural network for brain network representation learning, relies on a predefined spatial distance threshold to divide adjacency relationships and cannot dynamically adapt to changes in functional similarity between physiological signal channels.

[0004] The existing invention patent application document "An epilepsy localization method, terminal device and storage medium" with publication number CN118750007A, the existing method includes: inputting the patient's electroencephalogram (EEG) data pair; dividing the EEG data into fixed time window segments and sequentially performing node feature extraction, adjacency matrix calculation, graph construction, graph neural network node embedding extraction and subgraph comparative learning; inputting the EEG sample to be diagnosed into the parameterized network model for data training and testing to achieve epilepsy classification and localization. However, the aforementioned existing solution is for epilepsy localization, and the adjacency matrix integrates semantics and spatial distances, but does not provide a general channel aggregation method and cannot be migrated to other physiological signals, such as electrocardiogram and electromyography. At the same time, the existing solution also needs to rely on preset domain knowledge, such as epilepsy-specific channel associations.

[0005] The existing invention patent application document with publication number CN113269239A is "A method for classifying relational network nodes based on a multi-channel convolutional neural network". The existing method includes: obtaining initial graph network information containing nodes to be classified, the initial graph network information including initial graph network structure information and node feature information of each node in the initial graph network; constructing a feature graph based on the feature similarity between nodes in the initial graph network, and obtaining the structural information of the feature graph; establishing a node classification model, and pre-training the node classification model, inputting the structural information of the initial graph network, the feature information of each node in the initial graph network, and the structural information of the feature graph into the pre-trained node classification model to determine the type of each node; determining the label type to which each node belongs based on the label type in the initial graph network. The aforementioned existing scheme classifies relational network nodes based on a multi-channel convolutional neural network. However, this method focuses on node classification and does not involve channels, such as aggregation at the edge and adjacency matrix levels, and cannot handle the synergistic effects between channels in physiological signals.

[0006] In summary, existing technologies have technical problems such as poor universality, difficulty in dynamically capturing changes in functional similarity between channels, reliance on preset domain knowledge, and capturing single features. Summary of the Invention

[0007] The technical problem to be solved by the present invention is: how to solve the technical problems in the existing technology of poor universality, difficulty in dynamically capturing changes in functional similarity between channels, reliance on preset domain knowledge, and single feature capture.

[0008] The present invention solves the above technical problems by adopting the following technical solutions: A physiological signal channel aggregation method for graph neural networks includes:

[0009] S1. For the upstream graph neural network, obtain the channel node output and adjacency matrix of the feature extraction task;

[0010] S2. Calculate the similarity between channel nodes, construct a channel node similarity matrix, initialize the sorting matrix, and iterate the row and column normalization of the sorting matrix;

[0011] S3. Use the sorting matrix to sort the channel node features and the adjacency matrix respectively to obtain sorted channel node features and sorted adjacency matrix;

[0012] S4, defining the number of channels in each region, constructing a mapping, and transforming the sorted adjacency matrix into a mapping matrix;

[0013] S5. Perform aggregation operations on the sorting channel node features and mapping matrix;

[0014] S6. Construct a regional aggregation loss function and perform constraint operations on the sorted adjacency matrix.

[0015] This method aggregates physiological signal channels within graph neural networks. By performing row-column synchronized matrix sorting on the adjacency matrix, channels with similar functional expressions are aggregated within the adjacency matrix, effectively reflecting the local structural information of physiological signals. This method also possesses strong universality and can be directly embedded into physiological signal processing models based on graph neural networks without adjusting the upstream network structure. It is widely applicable to methods whose upstream task is to extract physiological signal features based on graph neural networks.

[0016] In a more specific technical solution, in S1, the channel node output in the upstream task is obtained Adjacency Matrix

[0017] In a more specific technical solution, in S2, for the channel node feature X, the channel node similarity matrix is calculated Using the following logic, calculate the element s in the i-th row and j-th column of the channel node similarity matrix S: i,j :

[0018]

[0019] Where, δ i,j is the Kronecker function;

[0020] Constructing a sorting matrix Using the following logic, we can get the initial sorting matrix:

[0021] P (0) =S

[0022] Where,

[0023] Initialize the sorting matrix P (0) Perform T iterations of row and column normalization;

[0024] For the sorting matrix of the Tth iteration Let P = P (T) to obtain the sorting matrix.

[0025] In a more specific technical solution, during the row and column normalization of T iterations, for the t-th iteration, the row normalization operation is performed using the following logic:

[0026]

[0027] The column normalization operation is performed using the following logic:

[0028]

[0029] Where N represents the number of channel nodes.

[0030] The present invention constructs a sorting matrix to sort the adjacency matrix and channel node features, and constructs a mapping and designs a regional aggregation loss function to constrain the adjacency matrix to achieve dynamic aggregation of channel nodes, which can dynamically capture the functional similarity changes between physiological signal channels.

[0031] In a more specific technical solution, in S3, the following logic is used to use the sorting matrix P to synchronously sort the rows and columns of the adjacency matrix A to obtain the sorted adjacency matrix

[0032]

[0033] Using the following logic, the sorting matrix P is used to sort the channel node features X to obtain the sorted channel node features

[0034] X′=PX.

[0035] The present invention performs row-column synchronized matrix sorting on the adjacency matrix, so that channels with similar functional expressions are aggregated on the adjacency matrix, effectively reflecting the local structural information of physiological signals and overcoming the problem that static partitioning methods are difficult to adapt to dynamic interactions between channels.

[0036] The present invention directly sorts and groups channel node features and adjacency matrices to capture dynamic functional associations between channels without being limited to node features.

[0037] In a more specific technical solution, in S4, N channel node features are aggregated into R regions, and the number of channel nodes in each region is preset:

[0038] C={c1,c2,…,c R};

[0039] Use the following logic to calculate the number of channel nodes c in the jth region: j :

[0040]

[0041] Where, r is the corresponding remainder;

[0042] Defines a mapping of channel nodes to regions Using the following logic, calculate the element m in the jth row and ith column of the mapping j,i :

[0043]

[0044] Use the following logic to construct a mapping matrix between channel nodes and aggregation areas

[0045] Z=A′M.

[0046] The present invention provides a universal channel aggregation framework that adapts to the channel characteristics of different physiological signals through sorting matrix and region division without the need for pre-set domain knowledge.

[0047] The construction of the sorting and mapping matrix in the present invention only relies on matrix multiplication, can be processed in parallel, and is applicable to large-scale physiological signal data.

[0048] In a more specific technical solution, in S5, the sorting channel node features Mapping Matrix Using the following logic, calculate the aggregate features

[0049] Y = ReLU(ZX′W+b)

[0050] Where W is the trainable weight matrix, b is the trainable bias, and ReLU is the activation function.

[0051] In a more specific technical solution, in S6, the sorted adjacency matrix Calculate matrix elements and weight matrix W A′ :

[0052]

[0053] Building a block diagonal mask matrix

[0054] Using the following logic, define the regional aggregation loss function L:

[0055]

[0056] Where ε is a small positive number and ⊙ represents element-by-element multiplication.

[0057] The present invention dynamically aggregates the channel node features of physiological signals through a sorting process, and designs a regional aggregation loss function to constrain the adjacency matrix to a block diagonal matrix form, thereby avoiding the limitations of predefined static partitioning methods, thereby better describing the dynamic interaction process between physiological signal channels and the functional overlap between regions formed in the dynamic interaction process.

[0058] In a more specific technical solution, the following logic is used to express the block diagonal mask matrix middle:

[0059]

[0060] Where, Indicates size c i ×c i0 represents a matrix of all 1s of the corresponding size.

[0061] In a more specific technical solution, the physiological signal channel aggregation system for graph neural networks includes:

[0062] The node output and adjacency matrix acquisition module is used to obtain the channel node output and adjacency matrix of the feature extraction task from the upstream graph neural network;

[0063] The row and column normalization iteration module is used to calculate the similarity between channel nodes, construct the channel node similarity matrix, initialize the sorting matrix, and perform row and column normalization iteration on the sorting matrix. The row and column normalization iteration module is connected to the node output and adjacency matrix acquisition module;

[0064] A sorting module is used to use the sorting matrix to sort the channel node features and the adjacency matrix respectively, and obtain the sorted channel node features and the sorted adjacency matrix. The sorting module is connected with the row and column normalization iteration module and the node output and adjacency matrix acquisition module;

[0065] A mapping construction module is used to define the number of channels in each region, construct a mapping, and transform the sorted adjacency matrix into a mapping matrix. The mapping construction module is connected to the sorting module;

[0066] Aggregation module, used to perform aggregation operations on sorting channel node features and mapping matrices. The aggregation module is connected to the mapping construction module and the sorting module.

[0067] The adjacency matrix constraint module is used to construct the regional aggregation loss function and perform constraint operations on the sorted adjacency matrix. The adjacency matrix constraint module is connected to the aggregation module.

[0068] Compared with the prior art, the present invention has the following advantages:

[0069] This method aggregates physiological signal channels within graph neural networks. By performing row-column synchronized matrix sorting on the adjacency matrix, channels with similar functional expressions are aggregated within the adjacency matrix, effectively reflecting the local structural information of physiological signals. This method also possesses strong universality and can be directly embedded into physiological signal processing models based on graph neural networks without adjusting the upstream network structure. It is widely applicable to methods whose upstream task is to extract physiological signal features based on graph neural networks.

[0070] The present invention constructs a sorting matrix to sort the adjacency matrix and channel node features, and constructs a mapping and designs a regional aggregation loss function to constrain the adjacency matrix to achieve dynamic aggregation of channel nodes, which can dynamically capture the functional similarity changes between physiological signal channels.

[0071] The present invention performs row-column synchronized matrix sorting on the adjacency matrix, so that channels with similar functional expressions are aggregated on the adjacency matrix, effectively reflecting the local structural information of physiological signals and overcoming the problem that static partitioning methods are difficult to adapt to dynamic interactions between channels.

[0072] The present invention directly sorts and groups channel node features and adjacency matrices to capture dynamic functional associations between channels without being limited to node features.

[0073] The present invention provides a universal channel aggregation framework that adapts to the channel characteristics of different physiological signals through sorting matrix and region division without the need for pre-set domain knowledge.

[0074] The construction of the sorting and mapping matrix in the present invention only relies on matrix multiplication, can be processed in parallel, and is applicable to large-scale physiological signal data.

[0075] The present invention dynamically aggregates the channel node features of physiological signals through a sorting process, and designs a regional aggregation loss function to constrain the adjacency matrix to a block diagonal matrix form, thereby avoiding the limitations of predefined static partitioning methods, thereby better describing the dynamic interaction process between physiological signal channels and the functional overlap between regions formed in the dynamic interaction process.

[0076] The present invention solves the technical problems existing in the prior art, such as poor universality, difficulty in dynamically capturing functional similarity changes between channels, reliance on preset domain knowledge, and single feature capture. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a schematic diagram of the basic steps of the physiological signal channel aggregation method for graph neural network according to Example 1 of the present invention;

[0078] Figure 2 This is a schematic diagram of the structure of a physiological signal channel aggregation model for a graph neural network according to Example 1 of the present invention;

[0079] Figure 3 Schematic diagram of the upstream graph neural network of Example 1 of the present invention;

[0080] Figure 4 Schematic diagram of the sorting process of Example 1 of the present invention;

[0081] Figure 5 This is a dynamic brain area distribution map of the subject in Example 2 of the present invention under a negative emotional state. DETAILED DESCRIPTION

[0082] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0083] Example 1

[0084] like Figure 1 As shown, the physiological signal channel aggregation method for graph neural network provided by the present invention includes the following basic steps:

[0085] S1. Obtain channel node features and adjacency matrix of upstream tasks;

[0086] like Figure 2 As shown, in this embodiment, the TFDN (Temporal Frequency Diffusion Network) model is used as the graph neural network used in the upstream task. The TFDN model is an upstream model for physiological signal feature extraction. Specifically, it is an EEG signal emotion recognition model. The specific structure of the TFDN model can be found in Figure 3 , get the channel node output of TFDN and the adjacency matrix

[0087] S2. Calculate the similarity between channel nodes, initialize the sorting matrix, and normalize the rows and columns of the sorting matrix;

[0088] In this embodiment, the similarity between channel nodes is calculated, a channel node similarity matrix is constructed, a sorting matrix is initialized, and then the sorting matrix is iterated by row and column normalization;

[0089] Specifically, for the channel node feature X, calculate the channel node similarity matrix For the element s in row i and column j of S i,j , and its calculation formula is:

[0090]

[0091] Among them, ||x i -x j ||2 represents the Euclidean distance between the feature vectors of channel node i and channel node j, which is used to measure the similarity between the two in the feature space; σ represents the bandwidth parameter, which is used to control the similarity decay rate; δ i,j is the Kronecker function, which takes the value 1 when i = j and 0 otherwise, and is used to ensure the similarity of the channel nodes themselves;

[0092] Constructing a sorting matrix Use the channel node similarity matrix S to initialize the sorting matrix P, that is, P (0) =S, where

[0093] For the initialized sorting matrix P (0) Perform row and column normalization for T iterations. For the tth iteration, the row normalization calculation formula is:

[0094]

[0095] The column normalization calculation formula is:

[0096]

[0097] Where N represents the number of channel nodes, p i,j (t-1) Represents the element in row i and column j in the t-1 iteration of the sorting matrix;

[0098] For the sorting matrix of the Tth iteration Let P = P (T) to obtain the sorting matrix.

[0099] S3, sorting the channel node features and adjacency matrix using the sorting matrix;

[0100] like Figure 4 As shown, in this embodiment, the channel node features and the adjacency matrix are sorted using a sorting matrix. The sorted channel node features and the adjacency matrix are obtained;

[0101] Specifically, use the sorting matrix P to sort the adjacency matrix A in row and column synchronization to obtain the sorted adjacency matrix The calculation formula is:

[0102]

[0103] Among them, the purpose of P multiplying the adjacency matrix A on the left is to sort the rows of A. The purpose of right-multiplying the adjacency matrix A is to sort the columns of A;

[0104] Use the sorting matrix P to sort the channel node features X to obtain the sorted channel node features The calculation formula is:

[0105] X′=PX

[0106] Sorting the channel node features X ensures that in the subsequent steps, the adjacency matrix A′ corresponds to the channel of the channel node features X′;

[0107] S4, constructing a mapping by grouping the channels, and transforming the sorted adjacency matrix into a mapping matrix;

[0108] In this embodiment, the number of channels contained in each region is defined, and a mapping is constructed to transform the sorted adjacency matrix into a mapping matrix;

[0109] Specifically, for N channel node features, they are aggregated into R regions, and the number of channel nodes contained in each region is specified as C = {c1, c2, ..., c R}, for the number of channel nodes c in the jth region j , and its calculation formula is:

[0110]

[0111] in, r is the corresponding remainder, so that the number of channel nodes contained in each area tends to be balanced;

[0112] Define the mapping of channel nodes to regions The element m in the jth row and ith column j,i The calculation formula is:

[0113]

[0114] Construct a mapping matrix between channel nodes and aggregation areas The calculation formula is:

[0115]

[0116] S5, aggregating the sorted channel node features and mapping matrix;

[0117] In this embodiment, for the sorted channel node features With the mapping matrix Calculating aggregate features The calculation method is:

[0118] Y = ReLU(ZX′W+b)

[0119] Among them, W and b are the trainable weight matrix and bias respectively, and ReLU is the activation function;

[0120] S6. Construct a regional aggregation loss function to constrain the adjacency matrix.

[0121] In this embodiment, a regional aggregation loss function is constructed to constrain the sorted adjacency matrix to conform to a block diagonal matrix form during the back-propagation process. Specific operations include but are not limited to:

[0122] For the sorted adjacency matrix Calculate the matrix elements and W A′ , which is calculated as follows:

[0123]

[0124] Building a block diagonal mask matrix Specifically:

[0125]

[0126] in, Indicates size c i ×c i The matrix of all 1s, c i ∈{c1,c2,…,c R}, 0 represents a matrix of all 0s of the corresponding size;

[0127] Define the regional aggregation loss function L, which is calculated as follows:

[0128]

[0129] Here, ε is a small positive number, and ⊙ represents element-by-element multiplication. The regional aggregation loss function L is used during the model backpropagation process to guide the model to enhance the intra-region connection density while effectively reducing the interference between regions, thereby obtaining an adjacency matrix with a block structure.

[0130] Example 2

[0131] In this embodiment, a specific individual EEG signal emotion recognition task based on the SEED dataset is described. The SEED dataset includes but is not limited to: 62 EEG channels, corresponding to the international 10-20 lead layout, which records the EEG signals of subjects when watching emotion-inducing videos, and is used to identify the three emotional states of pleasure, neutrality, and sadness. The method of this application is embedded in the upstream TFDN model. For subject-dependent experiments, a five-fold cross-validation is used, that is, each subject's data is divided into 5 subsets separately to avoid cross-subject generalization interference. Only the first session data is used to eliminate multi-session data differences, focusing on the dynamic aggregation of channels in a single scenario.

[0132] In this embodiment, the channel node features of the TFDN model in the upstream task are obtained With the adjacency matrix

[0133] Specifically, calculate the channel node similarity matrix The calculation formula is: Where σ=1; construct the sorting matrix And initialize P (0) =S; for the initialized sorting matrix P (0)Perform row and column normalization for T = 10 iterations. For the tth iteration, the row normalization calculation formula is: The column normalization calculation formula is: For the T=10th iteration, the sorting matrix P=P (10) Let to obtain the sorting matrix.

[0134] Specifically, use the sorting matrix P to sort the adjacency matrix A in row and column synchronization to obtain the sorted adjacency matrix The calculation formula is: Use the sorting matrix P to sort the channel node features X to obtain the sorted channel node features The calculation formula is: X′=PX;

[0135] In this embodiment, N=62 channel node features are aggregated into R=16 regions, and the number of channel nodes contained in each region is specified as C={c1, c2, ..., c R}, for the number of channel nodes c in the jth region j , and its calculation formula is: in q=3, r is the corresponding remainder, r=2; define the mapping from channel nodes to regions The element m in the jth row and ith column j,i The calculation formula is: Construct a mapping matrix between channel nodes and aggregation areas The calculation formula is:

[0136] In this embodiment, for the sorted channel node features With the mapping matrix Calculating aggregate features The calculation method is: Y = ReLU (ZX′W + b), where W and b are the trainable weight matrix and bias respectively, and ReLU is the activation function; for the sorted adjacency matrix Calculate the matrix elements and W A′ , which is calculated as follows: Building a block diagonal mask matrix Specifically: in Indicates size c i ×c i The full 1 matrix, 0 represents the full 0 matrix of the corresponding size; define the regional aggregation loss function L, which is calculated as follows: Where ε is a small positive number and ⊙ represents element-by-element multiplication. The present invention constructs a sorting matrix to sort the channel node features and the adjacency matrix, and designs a regional aggregation loss function so that the model sorts channels with similar functional expressions in the adjacent order during the learning process.

[0137] Table 1 Comparison of the method of the present invention and other channel polymerization methods under subject-dependent conditions

[0138] Method Name Dataset Accuracy (%) / Standard deviation (%) TFDN+static 7-area channel aggregation SEED 93.89 / 05.54 TFDN+static 10-area channel aggregation SEED 93.12 / 06.91 TFDN+static 11-area channel aggregation SEED 92.97 / 07.48 TFDN+static 17-area channel aggregation SEED 94.53 / 05.72 TFDN+ local channel aggregation method SEED 94.69 / 03.93

[0139] Table 1 above shows the comparison between the method of the present invention and other static channel aggregation methods under a subject-dependent experiment. The comparison results show that the method of the present invention is superior to other methods in terms of both accuracy and standard deviation.

[0140] like Figure 5 As shown, the dynamic brain area distribution of subject No. 15 in the SEED dataset under a negative emotional state is displayed. The present invention successfully divides 62 brain electrodes into 16 functionally related dynamic brain areas, demonstrating the effectiveness of the present invention.

[0141] The present invention can be used for physiological signal channel aggregation of graph neural networks and for emotion recognition tasks using EEG signals (a type of physiological signal). EEG signals contain multiple channels, and different channels correspond to different functional areas of the brain. Emotion recognition needs to capture the dynamic functional coordination between channels.

[0142] In summary, this invention aggregates physiological signal channels using graph neural networks. By performing row-column synchronized matrix sorting on the adjacency matrix, channels with similar functional expressions are aggregated on the adjacency matrix, effectively reflecting the local structural information of physiological signals. This invention also has strong universality and can be directly embedded in physiological signal processing models based on graph neural networks without adjusting the upstream network structure. It can be widely applied to methods whose upstream task is to extract physiological signal features based on graph neural networks.

[0143] The present invention constructs a sorting matrix to sort the adjacency matrix and channel node features, and constructs a mapping and designs a regional aggregation loss function to constrain the adjacency matrix to achieve dynamic aggregation of channel nodes, which can dynamically capture the functional similarity changes between physiological signal channels.

[0144] The present invention performs row-column synchronized matrix sorting on the adjacency matrix, so that channels with similar functional expressions are aggregated on the adjacency matrix, effectively reflecting the local structural information of physiological signals and overcoming the problem that static partitioning methods are difficult to adapt to dynamic interactions between channels.

[0145] The present invention directly sorts and groups channel node features and adjacency matrices to capture dynamic functional associations between channels without being limited to node features.

[0146] The present invention provides a universal channel aggregation framework that adapts to the channel characteristics of different physiological signals through sorting matrix and region division without the need for pre-set domain knowledge.

[0147] The construction of the sorting and mapping matrix in the present invention only relies on matrix multiplication, can be processed in parallel, and is applicable to large-scale physiological signal data.

[0148] The present invention dynamically aggregates the channel node features of physiological signals through a sorting process, and designs a regional aggregation loss function to constrain the adjacency matrix to a block diagonal matrix form, thereby avoiding the limitations of predefined static partitioning methods, thereby better describing the dynamic interaction process between physiological signal channels and the functional overlap between regions formed in the dynamic interaction process.

[0149] The present invention solves the technical problems existing in the prior art, such as poor universality, difficulty in dynamically capturing functional similarity changes between channels, reliance on preset domain knowledge, and single feature capture.

[0150] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A physiological signal channel aggregation method for graph neural networks, characterized by: The method comprises: S1. For the upstream graph neural network, obtain the channel node output and adjacency matrix of the feature extraction task; S2. Calculate the similarity between channel nodes, construct a channel node similarity matrix, initialize a sorting matrix, and iterate row and column normalization on the sorting matrix; S3. Use the sorting matrix to sort the channel node features and the adjacency matrix respectively to obtain sorted channel node features and a sorted adjacency matrix; S4, defining the number of channels in each region, constructing a mapping, and transforming the sorted adjacency matrix into a mapping matrix; S5. performing an aggregation operation on the sorting channel node features and the mapping matrix; S6. Construct a regional aggregation loss function and perform a constraint operation on the sorted adjacency matrix.

2. The physiological signal channel aggregation method for graph neural network according to claim 1 is characterized in that: In S1, the channel node output in the upstream task is obtained The adjacency matrix 3. The physiological signal channel aggregation method for graph neural network according to claim 1 is characterized in that: In S2, for the channel node feature X, calculate the channel node similarity matrix Using the following logic, calculate the element s in the i-th row and j-th column of the channel node similarity matrix S i,j : Where, δ i,j is the Kronecker function; Construct the sorting matrix Using the following logic, we can get the initial sorting matrix: P (0) =S Where, Initialize the sorting matrix P (0) Perform T iterations of row and column normalization; For the sorting matrix of the Tth iteration Let P = P (T) To obtain the sorting matrix.

4. The physiological signal channel aggregation method for graph neural network according to claim 3 is characterized in that: In the row and column normalization of T iterations, for the t-th iteration, the row normalization operation is performed using the following logic: The column normalization operation is performed using the following logic: Where N represents the number of channel nodes.

5. The physiological signal channel aggregation method for graph neural network according to claim 1 is characterized in that: In S3, the following logic is used to use the sorting matrix P to synchronously sort the rows and columns of the adjacency matrix A to obtain the sorted adjacency matrix Using the following logic, the channel node feature X is sorted using the sorting matrix P to obtain the sorted channel node feature X′=PX.

6. The physiological signal channel aggregation method for graph neural network according to claim 1, characterized in that: In S4, the N channel node features are aggregated into R regions, and the number of channel nodes in each region is preset: C={c1,c2,…,c R }; Use the following logic to calculate the number of channel nodes c in the jth region: j : Where, r is the corresponding remainder; Defines a mapping of channel nodes to regions Using the following logic, calculate the element m in the jth row and ith column of the mapping j,i : The mapping matrix between the channel nodes and the aggregation areas is constructed using the following logic: Z=A′M.

7. The physiological signal channel aggregation method for graph neural network according to claim 1 is characterized in that: In said S5, the characteristics of the sorting channel nodes are The mapping matrix Using the following logic, calculate the aggregate features Y=ReLU(ZX ′ W+b) Where W is the trainable weight matrix, b is the trainable bias, and ReLU is the activation function.

8. The physiological signal channel aggregation method for graph neural network according to claim 1 is characterized in that: In S6, the sorted adjacency matrix Calculate matrix elements and weight matrix W A′ : Building a block diagonal mask matrix The regional aggregation loss function L is defined using the following logic: Where ε is a small positive number and ⊙ represents element-by-element multiplication.

9. The physiological signal channel aggregation method for graph neural network according to claim 8, characterized in that: The block diagonal mask matrix is expressed using the following logic: middle: Where, Indicates size c i ×c i 0 represents a matrix of all 1s of the corresponding size.

10. A physiological signal channel aggregation system for graph neural networks, characterized by: The system comprises: The node output and adjacency matrix acquisition module is used to obtain the channel node output and adjacency matrix of the feature extraction task from the upstream graph neural network; A row and column normalization iteration module is used to calculate the similarity between channel nodes, construct a channel node similarity matrix, initialize a sorting matrix, and perform row and column normalization iteration on the sorting matrix. The row and column normalization iteration module is connected to the node output and adjacency matrix acquisition module; A sorting module, configured to use the sorting matrix to sort the channel node features and the adjacency matrix respectively, and obtain sorted channel node features and a sorted adjacency matrix, wherein the sorting module is connected to the row and column normalization iteration module and the node output and adjacency matrix acquisition module; a mapping construction module, used to define the number of channels in each region, construct a mapping, and transform the sorting adjacency matrix into a mapping matrix, the mapping construction module being connected to the sorting module; An aggregation module, configured to perform aggregation operations on the sorting channel node features and the mapping matrix, the aggregation module being connected to the mapping construction module and the sorting module; An adjacency matrix constraint module is used to construct a regional aggregation loss function and perform a constraint operation on the sorted adjacency matrix. The adjacency matrix constraint module is connected to the aggregation module.

Images