Classification and recognition method and device based on self-adaptive ascending learning strategy, computing equipment and storage medium

Abstract

The invention discloses a classification recognition method and device based on an adaptive ascending learning strategy, computing equipment and a storage medium, and the method comprises the steps: obtaining a plurality of data sets of different subjects in different cognitive states, and carrying out the preprocessing and feature extraction of each data set, obtaining a graph structure containing differential entropy and a Pearson coefficient of each tested individual in each data set; obtaining an optimal sub-graph of each tested individual in each data set by using a dynamic cluster center strategy and the graph structure containing differential entropy and Pearson coefficient of each tested individual in each data set; obtaining implicit representation of each order graph by using a double-graph-driven interactive adaptive order-lifting network and the optimal sub-graph of each tested individual in each data set; and constructing a decoder, and inputting the implicit representation of each order of graph into the decoder to obtain a classification recognition result.

Description

Technical Field

[0001] The present application relates to the technical field of implicit brain drive representation, and in particular to a classification and recognition method, apparatus, computing device and storage medium based on an adaptive step-up learning strategy. Background Art

[0002] EEG (Electroencephalogram) signals have become one of the important tools for assessing people's brain cognitive state due to their high temporal resolution and difficulty in faking. It is worth noting that accurate EEG decoding technology can promote the practical application value of BCI (Brain-Computer Interface), such as clinical medicine, behavioral monitoring and mind reading. However, accurately decoding brain activity patterns from unstable and complex EEG signals has always been a challenge for researchers.

[0003] Most of the current related research basically designs appropriate methods to decode the cognitive state under a specific cognitive task. Since EEG signals can extract time domain, frequency domain and spatial domain features, a large number of works extract discriminative cognitive features from these three perspectives. For example, a new multi-dimensional feature fusion network based on Gaussian time domain network and pure convolutional space-frequency domain network is proposed for fatigue detection. Considering the differences in different emotional states, a convolutional attention network is proposed to capture spatiotemporal emotional information. In addition, there is also work that combines deep learning and geometric modeling to convert EEG data into high-dimensional information to fully represent the brain's spatiotemporal information. In addition, some studies design effective transfer learning methods to narrow the differences between subjects. The essence of these methods is to capture the common features of each subject while distinguishing the difference information of each subject.

[0004] In addition, some recent work has introduced prior knowledge from neuroscience to construct a series of brain-inspired networks. The functions of multiple brain regions are combined to identify global and local fatigue representations. Neuroscience shows that there are significant differences in the regional driven interaction patterns of individuals' brains, and they drive advanced cognitive states. There are three main defects in current EEG-based research: 1) Efficient construction of brain region subgraph models: Current research lacks consideration of the importance of edge information in brain graph models to node information updates, and how to update node information to edge information to enhance graph representation; 2) Optimal division of brain region subgraphs: Most studies rely too much on prior knowledge from neuroscience and lack research on the brain region driven interaction patterns unique to each individual, so as to better divide the graph model subspace; 3) The interaction patterns between subgraphs cannot adaptively update feature information: These studies do not dynamically and adaptively interact with the information between brain region subgraphs to locally upgrade to globally to capture high-order implicit representations. Summary of the invention

[0005] The present application provides a classification and recognition method, apparatus, computing device and storage medium based on an adaptive step-up learning strategy, which solves the technical problem existing in the prior art of how to decode the implicit driving representation of the brain of different individuals in different cognitive states.

[0006] The present application provides a classification and recognition method based on an adaptive ascending learning strategy, comprising:

[0007] Obtain multiple data sets of different subjects in different cognitive states, and obtain a graph structure containing differential entropy and Pearson coefficient for each subject in each data set by preprocessing and feature extraction for each data set;

[0008] Using a dynamic cluster center strategy and a graph structure including differential entropy and Pearson coefficient of each subject in each data set, an optimal subgraph of each subject in each data set is obtained;

[0009] Obtaining implicit representations of each order graph using a dual-graph driven interactive adaptive order-raising network and the optimal subgraph of each subject in each data set;

[0010] A decoder is constructed, and a classification recognition result is obtained by inputting the implicit representations of each order graph into the decoder.

[0011] Preferably, using the dynamic cluster center strategy and the graph structure including differential entropy and Pearson coefficient of each subject in each data set, the optimal subgraph of each subject in each data set is obtained, including:

[0012] For each individual subject's Pearson coefficient in the graph structure, a cluster center node of each individual subject is generated using a dynamic cluster center strategy;

[0013] According to the cluster center node of each individual subject, the intra-cluster correlation degree between each channel and other channels in the cluster and the inter-cluster correlation degree between each channel and other clusters are calculated respectively;

[0014] Calculating a silhouette coefficient using the intra-cluster association degree and the inter-cluster association degree;

[0015] The silhouette coefficient is used to obtain the optimal subgraph for each subject in each data set.

[0016] Preferably, the dual-graph driven interactive adaptive up-order network includes a sub-graph pattern extraction module, a strong edge guided dual-graph attention module and an adaptive up-order learning module.

[0017] Preferably, the step of using the dual-graph driven interactive adaptive ascending-order network and the optimal subgraph of each subject in each data set to obtain the implicit representation of each order graph includes:

[0018] Using the subgraph pattern extraction module in the dual-graph driven interactive adaptive ascending-order network and the optimal subgraph of each subject in each data set, a strongly coupled edge feature is obtained;

[0019] The strong edge-guided dual-graph attention module and the strongly coupled edge features in the dual-graph driven interactive adaptive ascending network are used to obtain the correlation scores between the sub-graphs;

[0020] By utilizing the adaptive up-order learning module in the dual-graph driven interactive adaptive up-order network and the correlation scores between the sub-graphs, an implicit representation of each-order graph is obtained.

[0021] Preferably, the strongly coupled edge features include:

[0022]

[0023] in, represents the strongly coupled edge feature of the sth subject; Represents the characteristic information of the sth subject calculated by the unidirectional driving mechanism; [.] T represents the transpose of the matrix; W represents the adaptively generated weight matrix; Represents the differential entropy characteristics of the sth subject.

[0024] Preferably, the correlation scores between the subgraphs include:

[0025]

[0026] Among them, RS k Represents the correlation score between each subgraph of order k; Represents the strongest correlation score between each subgraph of order k; Represents the correlation score between the i-th subgraph and the j-th subgraph of order k.

[0027] Preferably, the implicit representations of the various order graphs include:

[0028] PG (k) ={E (k) ,H (k)}

[0029] Among them, PG (k) represents the implicit representation of the k-order graph; E (k) represents k-order edge interaction information; H (k) Represents the k-order implicit representation after graph convolution.

[0030] The present application also provides a classification and recognition device based on an adaptive ascending learning strategy, comprising:

[0031] An acquisition module is used to acquire multiple data sets of different subjects in different cognitive states, and obtain a graph structure containing differential entropy and Pearson coefficient of each subject in each data set by preprocessing and feature extraction of each data set; and obtain an optimal subgraph of each subject in each data set by using a dynamic cluster center strategy and the graph structure containing differential entropy and Pearson coefficient of each subject in each data set;

[0032] Obtaining implicit representation modules of each order graph, for obtaining implicit representations of each order graph by using a dual-graph driven interactive adaptive order-raising network and an optimal subgraph of each subject in each data set;

[0033] The classification recognition module is used to construct a decoder, and obtains a classification recognition result by inputting the implicit representations of each order graph into the decoder.

[0034] The present application also provides a computing device, including:

[0035] Memory and processor;

[0036] The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions. The computer executable instructions are executed by the processor to implement the steps of the classification and recognition method based on the adaptive step-up learning strategy.

[0037] The present application also provides a computer-readable storage medium storing computer-executable instructions, which are executed by a processor to implement the steps of the classification and recognition method based on an adaptive step-up learning strategy.

[0038] The beneficial effect of the present application is that the graph representation of local areas is enhanced by implementing regional sub-graph patterns for each individual's unique regional driven interaction pattern, and an adaptive graph ascending framework is constructed to dynamically capture the regional driven interaction patterns of the brain under different cognitive states. In addition, the present application is the first to gradually and adaptively ascend the learning of complex EEG information based on graph selective interaction to reveal the high-order implicit representation of the brain of an individual in different cognitive states. This strategy was verified on multiple EEG datasets, and the results showed that its decoding effect is better than the most advanced methods. Experimental results show that GOI-AAL (Graph Optional Interaction-Adaptive Ascending Learning Strategy, an adaptive ascending learning strategy based on graph selective interaction) can effectively decode the high-order implicit representations of the brain of each individual under different cognitive states. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flow chart of a classification and recognition method based on an adaptive ascending learning strategy provided by the present application;

[0040] Figure 2 It is a schematic diagram of a classification and recognition device based on an adaptive ascending learning strategy provided by the present application;

[0041] Figure 3 This is a structural diagram of the selective interactive adaptive up-grading learning strategy GOI-AAL based on the brain graph model provided in this application. DETAILED DESCRIPTION

[0042] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application. In the subsequent description, the use of suffixes such as "module", "component" or "unit" used to represent elements is only to facilitate the description of the present application and has no specific meaning in itself. Therefore, "module", "component" or "unit" can be used in a mixed manner.

[0043] This application proposes a new method called GOI-AAL, an adaptive up-grading learning strategy based on graph selective interaction, to decode the implicitly driven representations of the brain of different individuals in different cognitive states. Starting from the unique regional driven interaction pattern of each individual, this strategy enhances the graph representation of local areas and constructs an adaptive graph up-grading framework to dynamically capture the high-order implicit representations of the brain in different cognitive states. The decoding effect of this strategy on three EEG datasets confirms its advantages of robustness and efficiency, and can effectively decode the high-order implicit representations of the brain in different cognitive states. At the same time, ablation studies have confirmed that the proposed strategy is more in line with the filtering of complex EEG information, so as to better understand the cognitive level of the model in processing information of different cognitive states.

[0044] Figure 1 This is a flow chart of a classification and recognition method based on an adaptive step-up learning strategy provided by the present application. Figure 1 As shown, including:

[0045] Step S101: obtaining multiple data sets of different subjects in different cognitive states, and obtaining a graph structure including differential entropy and Pearson coefficient of each subject in each data set by preprocessing and feature extraction for each data set;

[0046] Step S102: using the dynamic cluster center strategy and the graph structure including the differential entropy and the Pearson coefficient of each subject in each data set, obtaining the optimal subgraph of each subject in each data set;

[0047] Step S103: using a dual-graph driven interactive adaptive order-raising network and the optimal subgraph of each subject in each data set to obtain implicit representations of each order graph;

[0048] Step S104: construct a decoder, and obtain a classification recognition result by inputting the implicit representations of each order graph into the decoder.

[0049] In this application, the optimal subgraph of each subject in each data set is obtained by using the dynamic cluster center strategy and the graph structure containing differential entropy and Pearson coefficient of each subject in each data set, including:

[0050] For each individual subject's Pearson coefficient in the graph structure, a cluster center node of each individual subject is generated using a dynamic cluster center strategy;

[0051] According to the cluster center node of each individual subject, the intra-cluster correlation degree between each channel and other channels in the cluster and the inter-cluster correlation degree between each channel and other clusters are calculated respectively;

[0052] Calculating a silhouette coefficient using the intra-cluster association degree and the inter-cluster association degree;

[0053] The silhouette coefficient is used to obtain the optimal subgraph for each subject in each data set.

[0054] In the present application, the dual-graph driven interactive adaptive up-order network includes a sub-graph pattern extraction module, a strong edge guided dual-graph attention module and an adaptive up-order learning module.

[0055] In the present application, the use of a dual-graph driven interactive adaptive ascending-order network and the optimal subgraph of each subject in each data set to obtain an implicit representation of each order graph includes:

[0056] Using the subgraph pattern extraction module in the dual-graph driven interactive adaptive ascending-order network and the optimal subgraph of each subject in each data set, a strongly coupled edge feature is obtained;

[0057] The strong edge-guided dual-graph attention module and the strongly coupled edge features in the dual-graph driven interactive adaptive ascending network are used to obtain the correlation scores between the sub-graphs;

[0058] By utilizing the adaptive up-order learning module in the dual-graph driven interactive adaptive up-order network and the correlation scores between the sub-graphs, an implicit representation of each-order graph is obtained.

[0059] In this application, the strongly coupled edge features include:

[0060]

[0061] in, represents the strongly coupled edge feature of the sth subject; Represents the characteristic information of the sth subject calculated by the unidirectional driving mechanism; [.] T represents the transpose of the matrix; W represents the adaptively generated weight matrix; Represents the differential entropy characteristics of the sth subject.

[0062] In this application, the correlation scores between the subgraphs include:

[0063]

[0064] Among them, RS k Represents the correlation score between each subgraph of order k; Represents the strongest correlation score between each subgraph of order k; Represents the correlation score between the i-th subgraph and the j-th subgraph of order k.

[0065] In this application, the implicit representations of the various order graphs include:

[0066] PG (k) ={E (k) ,H (k)}

[0067] Among them, PG (k) represents the implicit representation of the k-order graph; E (k) represents k-order edge interaction information; H (k) Represents the k-order implicit representation after graph convolution.

[0068] The following is a detailed introduction to the specific process of this application

[0069] like Figure 3 As shown, this application is mainly divided into a dynamic cluster center strategy, a dual-graph driven interactive adaptive up-order network and a decoder. The three data sets used in this application are D = {D d}, d = {1, 2, 3}. For each data set, after data preprocessing and feature extraction, the differential entropy and Pearson combination graph is obtained as in, is the differential entropy characteristic, is the Pearson coefficient, N and S represent the sample size and the number of subjects respectively, and G is a graph structure consisting of differential entropy as the feature and the Pearson coefficient as the adjacency matrix.

[0070] This application evaluates the proposed GOI-AAL based on three public datasets of different tasks: the first dataset is used for fatigue classification, the second dataset is used for emotion recognition, and the third dataset is used for driving behavior classification.

[0071] The SEED-VIG dataset contains EEG data from 17 brain channels of 21 subjects. The experiment established a virtual driving scene and collected EEG signals and frontal EEG signals. The first dataset divides EEG data into awake and tired according to the PERCLOS (Percentage of Eye Closure) index.

[0072] The DREAMER dataset collects multimodal physiological data from 23 subjects, including 14 channels of EEG signals and 2 channels of ECG signals; each subject watched 18 emotion-inducing videos with a duration ranging from 67 to 393 seconds, and the last 60 seconds of data were selected, as the emotion induction at this time was considered optimal. Arousal, valence, and dominance were used as labels for EEG data. If the value is greater than 3, it indicates high emotion; otherwise, it indicates low emotion.

[0073] The MPDB dataset records multimodal physiological data generated by different driving behaviors of 35 subjects. The data is divided into five labels: smooth driving, acceleration, deceleration, lane change, and turning. The EEG data of this dataset includes 59 channels, the sampling frequency is 1kHz, and the sampling length of each behavior is 2s, so each frame contains 2000*59 sampling points.

[0074] Part I: Dynamic Cluster Center Strategy

[0075] The existing method is to divide the brain into multiple fixed areas based on prior knowledge of neuroscience, and then perform modeling. However, this division method may not be applicable to all cognitive tasks. More importantly, the distribution of EEG data of different subjects varies significantly, and it is not applicable to a unified standard for measurement. Therefore, in order to divide relatively stable subgraphs, the subgraph construction process needs to consider the uniformity within the cluster and the non-uniformity between clusters. The K-means clustering method can meet the above requirements. In this application, K-means is used as the benchmark method to construct a dynamic cluster center strategy, and adaptive subgraph division is performed according to individuals.

[0076] For the structure of the sth subject The dynamic cluster center strategy proposed in this application first randomly selects K initial cluster center points μ1,μ2,...,μ K It should be emphasized that since it involves graph operations and subgraph order-raising operations, C is the number of channels. Then, the other electrode channels are assigned to the nearest initial cluster center, and the cluster center is recalculated according to the Euclidean distance, and then iterated continuously until the cluster center and number remain unchanged. In general, the objective function that needs to be achieved here is as follows:

[0077]

[0078] Among them, μ′1,μ′2,...,μ′ K is the dynamically generated cluster center point, is the Pearson coefficient of the sth subject, and N is the sample size of the subject.

[0079] It is worth noting that in order to obtain the optimal subgraph division for each subject in the data set, SC (Silhouette Coefficient) is introduced here to assist in the judgment. This method takes into account the compactness within the cluster and the separation between clusters, and verifies whether it is the optimal division by calculating the correlation between each channel and other channels in the cluster (r) and the correlation with other clusters (R):

[0080]

[0081] where r c and R c In this application, the optimal subgraph division for each subject is obtained based on SC. Among them, SC k represents the optimal contour system.

[0082]

[0083] Part II: Dual-graph driven interactive adaptive up-order network

[0084] 2.1. Sub-graph pattern extraction

[0085] Since each channel in the subgraph has strong and weak connectivity, we first set up a screening mechanism for these substructures and redefine them as Thus, a strongly connected structure is obtained after sparseness. Where F(.) represents a screening function with a learnable parameter φ. Next, a unidirectional driving mechanism is developed to aggregate the structure of frequency band information:

[0086]

[0087] Where f(.) represents the driving mechanism with weight w and bias vector b, σ soft represents the Softmax function, and tanh represents the hyperbolic tangent activation function. Represents the feature information calculated by the unidirectional driving mechanism; It represents the strong edge after the sth subject is screened, that is, the Pearson coefficient with a threshold greater than γ = 0.8. In order to enhance the representation of edge information, it is proposed to update the node information into the edge information and adaptively assign the weight matrix to obtain the edge features.

[0088]

[0089] in[.] T represents the transpose of the matrix, and W represents the adaptively generated weight matrix. represents the differential entropy feature of the sth subject, and N represents the sample size. Represents the enhanced edge features.

[0090] 2.2 Adaptive Up-scaling Learning

[0091] First, we describe the learning process of the first-order features. Given a subject, for the i-th and j-th subgraphs, define PG 1 Pair the first-order subgraphs.

[0092]

[0093] Next, we describe the designed SEG-DGA (Strong Edge Guided Dual Graph Attention), the calculation results of which will serve as the direct basis for the final pairing of the first-order subgraphs. and (C i and C j Respectively represent the number of channels of the i-th and j-th sub-images; represents the enhanced edge feature of the sth subject. 1i represents the first-order ith subgraph. Ci represents the number of channels of the ith subgraph. Represents the enhanced edge feature of the sth subject. 1j represents the jth subgraph of the first order. Cj represents the number of channels of the jth subgraph. N represents the sample size). In order to obtain the edge feature information of both at the same time, a spatiotemporal convolutional layer is constructed to learn point by point and reduce the dimension of the feature information of both.

[0094]

[0095] where w i , w j , b i , b j Represent the weights and bias vectors of the two convolutional layers respectively. A i and A j Respectively represent the mapping information of the edge features of subgraph i and subgraph j after dimensionality reduction; then A i and A j The images are fused, then feature compressed through the pooling layer, and finally the correlation scores between sub-graphs are obtained through the Sigmoid function.

[0096]

[0097] here Represents the correlation score between the first-order subgraphs and serves as the basis for pairing the first-order subgraphs with the strongest correlation. i ) T represents the average pooling layer. sig Represents the sigmoid function.

[0098]

[0099] RS1 represents the correlation score between the first-order subgraphs, and 1 represents the first order.

[0100] Next, this pair of first-order subgraphs with the strongest correlation is fed to the module AHAL (Adaptive Higher-Ascent Learning) to capture and extract the implicit features of the first-order graph. and Due to C i ≠C j , this application considers the dimensional information of both while performing interactive calculations. As shown below, we first get:

[0101]

[0102] Where N represents the sample size, Ci represents the number of channels of sub-image i, and w i ,b i are the weight and bias parameters of the linear transformation of subgraph i, Q i ,K i ,V i represent the query vector, key vector and value vector of subgraph i respectively. i and Q j Respectively with [K i ,V i ] and [K j ,V j ] to perform cross calculation to obtain first-order edge interaction information, as shown below:

[0103]

[0104] in, Represents the first-order edge interaction information of subgraph i to subgraph j. Similarly, Represents the first-order edge interaction information of subgraph j to subgraph i. soft Represents the softmax activation function.

[0105] It is worth noting that the model may have gradient explosion during the training process, so and To recalibrate:

[0106]

[0107] in, represents the calibrated first-order edge interaction information. i,j and b i,j represents the weight and bias parameters of convolution training, and ReLu is the activation function. Then, the joint Constructing new graph parties with channel information And input into the graph convolution module:

[0108]

[0109] Where θ represents the Fourier domain parameter, T m (.) represents the m-th order Chebyshev polynomial, W m represents the weighting parameter, Represents the implicit representation features after graph convolution. Then the first-order graph implicit representation PG can be obtained (1) :

[0110]

[0111] In the subsequent second-order learning, the second-order subgraph pairing is further learned with other unconnected regions based on the implicit representation of the first-order graph. Similarly, the second-order subgraph pairing can be defined as:

[0112]

[0113] Then, repeat the calculation process of formulas (7)-(14) to obtain the second-order graph implicit representation PG (2) By analogy, the network finally obtains the K-order graph implicit representation PG through adaptive order learning. (K) The entire upgrade process can be summarized as follows:

[0114]

[0115] Part 3: Decoder

[0116] In obtaining the K-th order graph implicitly representing PG (K) Finally, a decoder is constructed to obtain the final classification result.

[0117] Y pred =FC 1*class [ReLU(FC 1*C (PG (K) ;θ′));θ″] (17)

[0118] Where class represents the number of categories in the dataset, FC represents the linear layer, and θ′ and θ″ represent the parameters of the two linear layers respectively. In summary, the calculation process of GOI-AAL is shown in Algorithm 1.

[0119]

[0120]

[0121] Figure 2 is a schematic diagram of a classification and recognition device based on an adaptive ascending learning strategy provided by the present application, such as Figure 2 As shown, including:

[0122] An acquisition module is used to acquire multiple data sets of different subjects in different cognitive states, and obtain a graph structure containing differential entropy and Pearson coefficient of each subject in each data set by preprocessing and feature extraction of each data set; and obtain an optimal subgraph of each subject in each data set by using a dynamic cluster center strategy and the graph structure containing differential entropy and Pearson coefficient of each subject in each data set;

[0123] Obtaining implicit representation modules of each order graph, for obtaining implicit representations of each order graph by using a dual-graph driven interactive adaptive order-raising network and an optimal subgraph of each subject in each data set;

[0124] The classification recognition module is used to construct a decoder, and obtains a classification recognition result by inputting the implicit representations of each order graph into the decoder.

[0125] The present application also provides a computing device, including:

[0126] Memory and processor;

[0127] The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions. The computer executable instructions are executed by the processor to implement the steps of the classification and recognition method based on the adaptive step-up learning strategy.

[0128] The present application also provides a computer-readable storage medium storing computer-executable instructions, which are executed by a processor to implement the steps of the classification and recognition method based on an adaptive step-up learning strategy.

[0129] In summary, this application proposes a new method called adaptive up-order learning strategy based on graph selective interaction (GOI-AAL) to decode the implicit driving representation of the brain of different individuals in different cognitive states. First, in order to overcome the limitation of neuroscience prior knowledge on model learning performance, the brain graph model of different individuals is constructed on the dynamic cluster center strategy (DCCS) in Euclidean space, and adaptively aggregated into K subgraph patterns according to the characteristics of the graph structure. Then, the strongly connected edge information of each subgraph pattern is screened, and the channel node information is aggregated to the edge to obtain the strongly coupled edge feature. In order to explore the correlation between each subgraph pattern, a strong edge-guided dual-graph attention module (SEG-DGA) is designed to obtain the correlation score between graphs, and a dual-graph driven interactive adaptive up-order network (GDIA) is developed based on this. The graph pairs screened by the correlation score are gradually incrementally learned to finally obtain the high-order implicit representation of the graph. The proposed strategy is verified on the datasets of multiple cognitive tasks, and the results show that its decoding performance is better than the existing state-of-the-art methods. More importantly, GOI-AAL can effectively decode the high-level implicit representations of each individual's brain under different cognitive states.

[0130] The beneficial effects of the present application are as follows: (1) The present application constructs a dynamic cluster center strategy (DCCS) in Euclidean space for brain graph models of different individuals, and adaptively aggregates them into K subgraph patterns according to the characteristics of the graph structure. (2) The present application designs a SEG-DGA to obtain the correlation scores between graphs, and develops a GDIA based on this, which gradually incrementally learns the graph pairs selected by the correlation scores to finally obtain the high-order implicit representation of the graph. (3) The verification of the proposed strategy on multiple datasets shows the superiority of its decoding performance and reveals the high-order implicit representation of the brain under different cognitive states.

[0131] The preferred embodiments of the present application are described above with reference to the accompanying drawings, but the scope of the present application is not limited thereby. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present application shall be within the scope of the present application.

Claims

1. A classification and recognition method based on an adaptive ascending learning strategy, characterized in that: include: Obtain multiple data sets of different subjects in different cognitive states, and obtain a graph structure containing differential entropy and Pearson coefficient for each subject in each data set by preprocessing and feature extraction for each data set; Using a dynamic cluster center strategy and a graph structure including differential entropy and Pearson coefficient of each subject in each data set, an optimal subgraph of each subject in each data set is obtained; Obtaining implicit representations of each order graph using a dual-graph driven interactive adaptive order-raising network and the optimal subgraph of each subject in each data set; A decoder is constructed, and a classification recognition result is obtained by inputting the implicit representations of each order graph into the decoder.

2. The method according to claim 1, characterized in that Using the dynamic cluster center strategy and the graph structure including differential entropy and Pearson coefficient of each subject in each data set, the optimal subgraph of each subject in each data set is obtained, including: For each individual subject's Pearson coefficient in the graph structure, a cluster center node of each individual subject is generated using a dynamic cluster center strategy; According to the cluster center node of each individual subject, the intra-cluster correlation degree between each channel and other channels in the cluster and the inter-cluster correlation degree between each channel and other clusters are calculated respectively; Calculating a silhouette coefficient using the intra-cluster association degree and the inter-cluster association degree; The silhouette coefficient is used to obtain the optimal subgraph for each subject in each data set.

3. The method according to claim 1, characterized in that The dual-graph driven interactive adaptive up-order network includes a sub-graph pattern extraction module, a strong edge guided dual-graph attention module and an adaptive up-order learning module.

4. The method according to claim 3, characterized in that The method of using the dual-graph driven interactive adaptive order-raising network and the optimal subgraph of each subject in each data set to obtain the implicit representation of each order graph includes: Using the subgraph pattern extraction module in the dual-graph driven interactive adaptive ascending-order network and the optimal subgraph of each subject in each data set, a strongly coupled edge feature is obtained; The strong edge-guided dual-graph attention module and the strongly coupled edge features in the dual-graph driven interactive adaptive ascending network are used to obtain the correlation scores between the sub-graphs; By utilizing the adaptive up-order learning module in the dual-graph driven interactive adaptive up-order network and the correlation scores between the sub-graphs, an implicit representation of each-order graph is obtained.

5. The method according to claim 4, characterized in that The strongly coupled edge features include: in, represents the strongly coupled edge feature of the sth subject; Represents the characteristic information of the sth subject calculated by the unidirectional driving mechanism; [.] T represents the transpose of the matrix; W represents the adaptively generated weight matrix; Represents the differential entropy characteristics of the sth subject.

6. The method according to claim 5, characterized in that The correlation scores between the subgraphs include: Among them, RS k Represents the correlation score between each subgraph of order k; Represents the strongest correlation score between each subgraph of order k; Represents the correlation score between the i-th subgraph and the j-th subgraph of order k.

7. The method according to claim 6, characterized in that The implicit representations of the various order graphs include: PG (k) ={E (k) ,H (k) } Among them, PG (k) represents the implicit representation of the k-order graph; E (k) represents k-order edge interaction information; H (k) Represents the k-order implicit representation after graph convolution.

8. A classification and recognition device based on an adaptive ascending learning strategy, characterized in that: include: An acquisition module is used to acquire multiple data sets of different subjects in different cognitive states, and obtain a graph structure including differential entropy and Pearson coefficient of each subject in each data set by preprocessing and feature extraction of each data set; Using a dynamic cluster center strategy and a graph structure including differential entropy and Pearson coefficient of each subject in each data set, an optimal subgraph of each subject in each data set is obtained; Obtaining implicit representation modules of each order graph, for obtaining implicit representations of each order graph by using a dual-graph driven interactive adaptive order-raising network and an optimal subgraph of each subject in each data set; The classification recognition module is used to construct a decoder, and obtains a classification recognition result by inputting the implicit representations of each order graph into the decoder.

9. A computing device comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of a classification and recognition method based on an adaptive step-up learning strategy as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of a classification and recognition method based on an adaptive step-up learning strategy as described in any one of claims 1 to 7.

Images