Pulse graph neural network space-time signal detection method and device, medium and product

Abstract

The invention discloses a pulse graph neural network space-time signal detection method and device, a medium and a product, and belongs to the technical field of space-time signal detection. According to the method, graph structure data of a system network is processed through a trained multi-source information fusion pulse graph neural network model, and the multi-source information fusion pulse graph neural network model comprises a graph convolutional neural network and a pulse graph neural network. The pulse graph neural network comprises an input layer, a self-adaptive mixed space-time pulse neuron model layer, an attention layer, a full connection layer and an output layer which are connected in sequence, and the advantages of the graph convolutional neural network and the pulse neural network are combined, so that efficient detection of the dynamic space-time signals of the complex topology network is realized; and the real-time performance and the accuracy of signal processing are improved.

Description

Technical Field

[0001] The present application relates to the technical field of spatio-temporal signal detection, and particularly to a spatio-temporal signal detection method, device, medium and product based on a spiking graph neural network. Background Art

[0002] With the development of information technology, the processing of dynamic signals in various networks (such as communication networks, sensor networks, etc.) has become increasingly important. Traditional signal processing methods often face problems such as large computational complexity and low efficiency when dealing with large-scale and high-complexity network signals, resulting in the real-time performance and accuracy of signal processing being affected. Summary of the Invention

[0003] The purpose of the present application is to provide a spatio-temporal signal detection method, device, medium and product based on a spiking graph neural network, so as to improve the real-time performance and accuracy of signal processing.

[0004] To achieve the above purpose, the present application provides the following solutions:

[0005] In a first aspect, the present application provides a spatio-temporal signal detection method based on a spiking graph neural network, including:

[0006] Obtaining graph structure data of various system networks, where each piece of graph structure data includes static graph structure data and dynamic spatio-temporal data. The static graph structure data is the fixed topological structure information of each node and edge in the system network. The static graph structure data includes a node feature matrix and an adjacency matrix, and the adjacency matrix is used to represent the connection relationship between nodes. The dynamic spatio-temporal data is the network signal data that changes with time on each node;

[0007] Taking each piece of graph structure data as an input, and using a trained spiking graph neural network model for multi-source information fusion to detect anomalies and identify anomaly categories. The spiking graph neural network model for multi-source information fusion includes a graph convolutional neural network and a spiking graph neural network. The spiking graph neural network includes an input layer, an adaptive hybrid spatio-temporal spiking neuron model layer, an attention layer, a fully connected layer, and an output layer connected in sequence. The graph convolutional neural network is used to extract features from the static graph structure data to obtain the node representation after graph convolution. The spiking graph neural network is used to perform anomaly detection based on the node representation that fuses spatio-temporal information, and the node representation that fuses spatio-temporal information is a vector obtained based on the node representation after graph convolution and the dynamic spatio-temporal data.

[0008] In a second aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. The processor executes the computer program to implement the spatio-temporal signal detection method based on a spiking graph neural network described in the first aspect above.

[0009] In a third aspect, the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the pulse graph neural network spatio-temporal signal detection method described in the first aspect above is implemented.

[0010] In a fourth aspect, the present application provides a computer program product, including a computer program, and when the computer program is executed by a processor, the pulse graph neural network spatio-temporal signal detection method described in the first aspect above is implemented.

[0011] According to the specific embodiments provided by the present application, the present application has the following technical effects:

[0012] The present application provides a pulse graph neural network spatio-temporal signal detection method, device, medium and product. The method processes the graph structure data of the system network through a trained multi-source information fusion pulse graph neural network model. The multi-source information fusion pulse graph neural network model includes a graph convolutional neural network and a pulse graph neural network. The pulse graph neural network includes an input layer, an adaptive hybrid spatio-temporal pulse neuron model layer, an attention layer, a fully connected layer and an output layer connected in sequence. By collecting static graph structure data (node feature matrix, adjacency matrix) and dynamic spatio-temporal data (time series signal), the graph convolutional neural network (Graph Convolutional Network, GCN) is used to extract topological features, spatio-temporal information is fused through pulse coding, the fused spatio-temporal information is processed by an adaptive hybrid spatio-temporal pulse neuron model (AHSNM), and the key information capture ability is enhanced by combining the pulse-induced graph attention mechanism. Finally, the abnormal category is determined by statistically analyzing the pulse firing intensity and frequency, and an abnormal probability vector is output. The present application combines the advantages of the graph convolutional neural network and the pulse neural network, realizes the efficient detection of dynamic spatio-temporal signals of complex topological networks, and improves the real-time performance and accuracy of signal processing. The present application has high efficiency, robustness and versatility, and is applicable to fields such as power systems, communication networks, and sensor networks. Description of the Drawings

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0014] Figure 1 It is an application environment diagram of a pulse graph neural network spatio-temporal signal detection method in Embodiment 1 of the present application;

[0015] Figure 2Schematic flow chart of a spiking graph neural network spatio-temporal signal detection method provided in Embodiment 1 of this application;

[0016] Figure 3 Schematic processing flow chart of a spiking graph neural network model for multi-source information fusion in Embodiment 1 of this application;

[0017] Figure 4 Schematic structural diagram of a computer device provided in Embodiment 2 of this application. Detailed implementation manners

[0018] Next, the technical solutions in the embodiments of this application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this application.

[0019] Spiking Neural Networks (SNNs) show great potential in network signal processing due to their biologically inspired spike coding and low-power consumption characteristics. However, existing SNN models still have limitations in processing graph-structured data, especially in capturing positional information and dynamic features between nodes.

[0020] In response to this, the purpose of this application is to provide a spiking graph neural network spatio-temporal signal detection method, device, medium, and product. This method aims to efficiently process and detect dynamic signals of different complex topological network structures. By combining Graph Convolutional Network (GCN), AHSNM model, and spike-induced graph attention mechanism, it realizes efficient and accurate feature extraction of spatio-temporal dynamic signals in various networks, improving the real-time performance and accuracy of signal processing.

[0021] To make the above objects, features, and advantages of this application more obvious and understandable, the following further detailed description of this application will be given in conjunction with the accompanying drawings and specific implementation manners.

[0022] Embodiment 1

[0023] The spiking graph neural network spatio-temporal signal detection method provided in the embodiment of this application can be applied to, for example Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be set separately, integrated on the server 104, placed on the cloud or other servers. The terminal 102 can send the graph structure data of various system networks to the server 104. After receiving the graph structure data of various system networks, the server 104 uses each of the graph structure data as input, and uses the trained multi-source information fusion spiking graph neural network model to detect anomalies and identify anomaly categories. Among them, the multi-source information fusion spiking graph neural network model includes a graph convolutional neural network and a spiking graph neural network. The spiking graph neural network includes an input layer, an adaptive hybrid spatio-temporal spiking neuron model layer, an attention layer, a fully connected layer, and an output layer connected in sequence. The graph convolutional neural network is used to extract features from the static graph structure data to obtain the node representation after graph convolution. The spiking graph neural network is used to perform anomaly detection based on the node representation that fuses spatio-temporal information. The node representation that fuses spatio-temporal information is a vector obtained based on the node representation after graph convolution and the dynamic spatio-temporal data. The server 104 can feedback the detected anomalies and anomaly categories to the terminal 102. In addition, in some embodiments, the video label processing method can also be implemented separately by the server 104 or the terminal 102. For example, the terminal 102 can directly use the spiking graph neural network spatio-temporal signal detection method to process the graph structure data of various system networks to be processed, or the server 104 can obtain the graph structure data of various system networks to be processed from the data storage system and use the spiking graph neural network spatio-temporal signal detection method to process it.

[0024] Among them, the terminal 102 can be, but is not limited to, various desktop computers, laptop computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. The portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers, and can also be a cloud server.

[0025] In an exemplary embodiment, a spiking graph neural network spatio-temporal signal detection method is provided. This method is executed by a computer device, and can be specifically executed alone by a computer device such as a terminal or a server, or jointly executed by a terminal and a server. In the embodiments of the present application, taking this method applied to Figure 1 the server 104 in as an example for illustration, it includes the following steps 201 to step 202. Among them:

[0026] Step 201: Obtain the graph structure data of various system networks. Each piece of the graph structure data includes static graph structure data and dynamic spatio-temporal data. The static graph structure data is the fixed topological structure information of each node and edge in the system network. The static graph structure data includes a node feature matrix and an adjacency matrix. The adjacency matrix is used to represent the connection relationship between nodes. The dynamic spatio-temporal data is the network signal data that changes with time on each node.

[0027] Step 202: Use the trained pulsed graph neural network model for multi-source information fusion to detect anomalies and identify anomaly categories with each piece of the graph structure data as the input. The pulsed graph neural network model for multi-source information fusion includes a graph convolutional neural network and a pulsed graph neural network. The pulsed graph neural network includes an input layer, an adaptive hybrid spatio-temporal pulsed neuron model layer, an attention layer, a fully connected layer, and an output layer connected in sequence. The graph convolutional neural network is used to extract features from the static graph structure data to obtain the node representation after graph convolution. The pulsed graph neural network is used to perform anomaly detection based on the node representation that fuses spatio-temporal information. The node representation that fuses spatio-temporal information is a vector obtained based on the node representation after graph convolution and the dynamic spatio-temporal data.

[0028] To make those skilled in the art more clear about the specific execution process of the pulsed graph neural network spatio-temporal signal detection method provided in this embodiment, the following is specifically explained in combination with Figures 2 - 3 for specific interpretation.

[0029] As Figure 2 and Figure 3 shown, the pulsed graph neural network spatio-temporal signal detection method includes:

[0030] Step 1: Data preparation and preprocessing.

[0031] (1) Collect the graph information of the dynamic complex network topology:

[0032] First, adopt the time window division strategy. For a general complex network system, select a suitable time interval as the size of the time window. This time interval should be short enough to capture the rapid changes in the network state, and at the same time long enough to ensure that enough data can be collected within each time window.

[0033] Within each time window, use image recognition technology and sensors to collect network information. Taking the power system network as an example, power line diagrams, substation layout diagrams, etc. can be used to obtain the static structure of the network, and the information of nodes (such as substations, generators) and edges (such as transmission lines, distribution lines) can be extracted through image recognition technology. At the same time, dynamic signal data, such as voltage, current, power, etc., are collected through sensors or monitoring devices installed in the power grid.

[0034] (2) Convert the collected image information and sensor data into the input format of the spiking graph neural network to obtain graph structure data (static graph structure data and dynamic spatio-temporal data).

[0035] Specifically, convert the information of nodes and edges into the node feature matrix X and the adjacency matrix A. The node feature matrix X contains the feature information of each node, such as position, speed, power, etc.; the adjacency matrix A describes the connection relationship between nodes in the network. Among them, the node feature matrix X ∈ R N×F (N is the number of nodes, F is the number of features), and the adjacency matrix A ∈ R N×N , and the time series data V of the nodes t ∈ R N×T (that is, dynamic spatio-temporal data, where T is the number of time steps).

[0036] Classify the collected data nodes to determine the abnormal category. Before training, divide the data set, and split the graph data into training, validation, and test sets according to the ratio of 7:1.5:1.5.

[0037] Step 2: Input the graph structure data into the spiking graph neural network with multi-source information fusion for training.

[0038] Step 2.1: Use the graph convolutional neural network to extract features from the static graph structure data in the graph structure data, fuse the node features and the adjacency matrix information, generate the node representation after graph convolution, and capture the topological structure information of the network.

[0039] H = σ(A · S X · W);

[0040] Among them, S X is the placeholder of the node feature matrix to be encoded, W is the learnable weight matrix, σ is the activation function, and the node representation H after graph convolution ∈ R N×F , and F is the feature dimension after graph convolution).

[0041] Step 2.2: Perform spiking encoding and fusion on the node features extracted by graph convolution (that is, the generated node representation after graph convolution) and the time series data of the nodes.

[0042] Node spiking encoding: Assume that the magnitude of the encoded spiking probability is positively correlated with the importance of the node features, that is, use the Bernoulli function to encode the relationship between the probability p of each feature triggering a spike and the node features. The calculation formula for the node feature spike sequence is:

[0043] S Hij (t) ∼ Bernoulli(H(i, f));

[0044] Among them, i represents the node index, and j represents the feature dimension index. represents the impulse value (0 or 1) of the j-th feature of node i at time t. The Bernoulli function determines the probability of triggering an impulse according to the magnitude of the feature value H(i, f).

[0045] Similarly, the impulse coding of the time series data of the node (time series impulse sequence) can also be obtained:

[0046] S Vi (t) = Bernoulli(V ι (i, t));

[0047] Fusing spatio-temporal information: Concatenate the node feature impulse sequence and the time series impulse sequence in the time dimension to form a new input vector (i.e., the node representation fusing spatio-temporal information).

[0048] S i (t) = [S Hi1 (t), S Hi2 (t), S Hi3 (t), S Hi4 (t),..., S Hij (t), S Vi (t)];

[0049] Among them, j is the number of feature dimensions, and S i (t) is the fused input vector of node i at time t.

[0050] Step 2.3: Use the Adaptive Hybrid Spatio-Temporal Spiking Neuron Model (AHSNM) to process the node representation S i (t) that fuses spatio-temporal information, and introduce position encoding to capture the position information in the graph structure, intelligently concatenate the position impulse and the feature impulse, and generate the output impulse sequence of the AHSNM neuron through the adaptive threshold mechanism and the impulse generation rule to achieve efficient processing of dynamic signals.

[0051] (1) Neuron model basis and parameter description

[0052] Under the framework of the Spiking Neural Network (SNN), an Adaptive Hybrid Spatio-Temporal Spiking Neuron AHSNM model is proposed. The membrane potential update equation of this model and its parameter description are as follows:

[0053]

[0054] Among them, Denote the membrane potential of the $i$-th neuron at time $t$. $\lambda$ represents the decay factor of the membrane potential, with a value in the range $(0, 1)$. $W_{0}$ represents the connection weight from the $j$-th neuron to the $i$-th neuron. Although the position encoding and the feature vector share the same weight matrix, they are distinguished by different coefficients. Denote the feature pulse amplitude of the $j$-th neuron, which is a learnable parameter. Denote the feature pulse of the $j$-th neuron at time $t$, taking values of -1, 0, or 1. Denote the position encoding weight, which is determined according to the relative position relationship between node $k$ and node $j$, and is also a learnable parameter. Denote the position encoding of position $k$, which has the same dimension as the feature vector. Denote the dynamic threshold of the $i$-th neuron at time $t$. $\delta(x)$ represents the ternary neuron mechanism function, where $\delta(x) = 1$ when $x\geq0$; $\delta(x) = -1$ when $x < 0$.

[0055] (2) Ternary Pulse Transmission and Firing Mechanism

[0056] The pulse firing of neurons follows the following ternary mechanism:

[0057]

[0058] (3) Dynamic Threshold Mechanism

[0059] Dynamic Threshold The update equation is:

[0060]

[0061] Where, Denote the dynamic threshold of the $i$-th neuron at time $t$, Denote the dynamic threshold of the $i$-th neuron at time $t - 1$, $\eta$ is the threshold adjustment rate, $\gamma$ is the input-dependent adjustment coefficient, is the long-term average value of the membrane potential of neuron $i$, $|W tj |$ represents the absolute value of the weight, which is used to adjust the speed of threshold change.

[0062] (4) Fusion of Position Encoding

[0063] To more effectively capture the position information in the graph structure, the position encoding strategy is incorporated into the input of neurons. The position encoding can be generated by advanced methods such as Laplacian eigenvectors (LSPE) or random walks (RWPE), and is converted into pulse form through Poisson encoding technology to enhance the model's sensitivity and processing ability to position information. Let the pulse form of the position encoding be where $k$ represents the position index of the node. The position pulse and the feature pulse are intelligently concatenated to jointly form the input signal of the neuron:

[0064]

[0065] Subsequently, the pulse input integrated with position information is substituted into the membrane potential update equation, enabling the neuron to simultaneously consider feature information and position information when receiving and processing information:

[0066]

[0067] It should be noted that during this process, the trainable pulse amplitude α only acts on the feature pulse while weighted by the position encoding weight β to reflect the different contributions of different position information to the update of the neuron's membrane potential. This pulse input integrated with position information and weight adjustment is substituted into the membrane potential update equation, enabling the neuron to more precisely capture and process the position relationship in the graph structure.

[0068] (5) Parameter Optimization and Model Training

[0069] During the training phase of the model, the backpropagation algorithm is used to iteratively optimize the threshold V u,ι of the neuron, the connection weight W0, the trainable pulse amplitude α, and the position weight β kj The goal of optimization is to minimize the loss function, such as the mean squared error (MSE), to improve the prediction accuracy and generalization ability of the model. The parameter update equation can be expressed as:

[0070]

[0071] where lr represents the learning rate, which controls the step size and speed of parameter update, and θ Δ,i represents the update amount of the learnable parameters of the i-th neuron in the network, where θ generally refers to the set of all learnable parameters, and V Δ,i represents the change in the membrane potential of the i-th neuron, represents the gradient of the loss function L with respect to the learnable parameter θ of the i-th neuron. During the inference phase of the model, to further improve the computational efficiency and maintain the high energy efficiency characteristics of the spiking neural network (SNN), the trainable pulse amplitude α i can be folded into the connection weight W0 to simplify the calculation process:

[0072]

[0073] In this way, no additional multiplication operation is required during the inference phase, thus significantly improving the computational efficiency of the model.

[0074] Step 2.4: Adopt the pulse-induced graph attention mechanism. According to the output pulse sequence of the AHSNM layer and the node representation after graph convolution, calculate the attention weights, and update the node representation that fuses spatio-temporal information, enhancing the model's ability to capture important information.

[0075] Attention weight A ij can be calculated based on the pulse sequences between node i and node j. Assume S i and S j represent the output pulse sequences of node i and node j respectively, then the attention weight A ij can be defined as:

[0076]

[0077] Using the calculated attention weight A ij , the node S′ in the graph structure can be updated i =∑A ij ·S j .

[0078] Step 2.5: Further process the output pulse sequence of the AHSNM layer through a fully connected layer, then obtain the pulse sequences of different nodes according to the output of the fully connected layer, determine the values of the corresponding categories by statistically counting the pulse emission intensity and frequency, and finally the output is a vector, and the vector dimension is the total number of abnormal categories.

[0079] Assume the output pulse sequence of the AHSNM layer is converted into a feature vector h i , then the output of the fully connected layer can be expressed as:

[0080] y i =σ(W·h i +b);

[0081] where, W is the weight matrix of the fully connected layer, b is the bias term, and σ is the activation function (such as ReLU, sigmoid, etc.).

[0082] Step 3: Use the trained multi-source information fusion pulse graph neural network model to perform real-time monitoring and alarming of node dynamic information. Specifically, use the trained multi-source information fusion pulse graph neural network to detect anomalies in various complex topological network system data. The input is real-time network system data, and the real-time abnormal category data is detected. When network node anomalies are detected, the model can provide timely alarm information and output the abnormal category at the same time, so as to facilitate maintenance personnel to efficiently locate and handle faults, thereby ensuring the stability and safety of industrial production.

[0083] In Step 1, only an example of the power system is given. In fact, the spatio-temporal signal detection method of the pulsed graph neural network with multi-source information fusion proposed in this embodiment can be flexibly applied to various complex network systems, such as power system networks, communication networks, sensor networks, and electroencephalogram signal recognition. A complex network system is a large or small network system that can collect information data of different devices under different conditions in different network systems, not limited to a specific network system. Different conditions refer to normal conditions and various abnormal conditions. For different complex network systems, relevant graph structure data need to be collected according to the actual situation to train the model, and then the trained model is used to monitor and alarm the dynamic information of the nodes in real time.

[0084] Specifically, for a communication network, the static structure can be represented by a network topology diagram. This includes nodes (such as routers, switches, base stations, etc.) and edges (such as fiber optic links, wireless channels, etc.). This information can be obtained from network planning documents, device lists, or network management systems and converted into a graphical representation. The dynamic data in a communication network may include traffic data (such as packet size, transmission rate), signal quality data (such as signal-to-noise ratio, bit error rate), device status data (such as temperature, humidity, power supply voltage), etc. This data is usually collected by network monitoring devices or sensors.

[0085] For a sensor network, the static structure usually consists of the positions of sensor nodes and their interconnection relationships. This information can be obtained from deployment plans, geographic information systems (GIS), or configuration files of the sensor network. In the graphical representation, nodes represent sensors, and edges may represent communication links or spatial proximity between sensors. The types of dynamic data collected in a sensor network depend on the type of sensor and the application scenario. For example, environmental sensors may collect data such as temperature, humidity, and air pressure; motion sensors may collect data such as acceleration, speed, and position. This data is usually transmitted wirelessly or wired to a central processing unit or a data collection node.

[0086] For the field of electroencephalogram signal recognition, the static structure usually refers to the structural information of the brain, such as the connection pattern (functional connection or structural connection) between brain regions. This information can be obtained through neuroimaging techniques (such as functional magnetic resonance imaging fMRI, diffusion tensor imaging DTI) and converted into a graphical representation. Nodes may represent different brain regions, and edges may represent the connection strength or functional correlation between brain regions. The dynamic data in electroencephalogram signal recognition usually refers to electroencephalogram (EEG) data, which records the electrical activity on the surface of the brain. EEG data is usually collected through an electrode array placed on the scalp and represented in the form of a time series. This data contains the activity patterns of the brain in different states, such as rest, thinking, and movement.

[0087] A method for detecting spatio-temporal signals of a pulsed graph neural network with multi-source information fusion provided by this embodiment aims to efficiently process and detect dynamic signals of different complex topological network structures. This method can be widely applied to fields such as power system networks, communication networks, sensor networks, and electroencephalogram signal recognition. This method includes: adopting a time window division strategy to collect graph information of a dynamic complex network topology and constructing corresponding graph data vectors; inputting the graph data into a pulsed graph neural network with multi-source information fusion for training, including using a graph convolutional neural network (GCN) to complete feature extraction, using a Bernoulli function for pulse coding and spatio-temporal data fusion, adopting an adaptive hybrid spatio-temporal pulse neuron model (AHSNM), a pulse-induced graph attention mechanism, and a fully connected layer to concatenate and extract spatio-temporal data information in multiple dimensions to learn node features and train the model; using the trained model to detect dynamic signals of a complex topological network structure, and when the dynamic signals are abnormal, providing alarm information to facilitate efficient positioning and processing by maintenance personnel.

[0088] The technical solution provided by this embodiment has the following advantages:

[0089] Generalizability: This technical solution is applicable to the detection of dynamic spatio-temporal signals in various networks, including but not limited to communication networks, sensor networks, power system networks, etc., and has high flexibility and scalability. First, the input and output of the pulsed graph neural network with multi-source information fusion can be adjusted and optimized according to specific tasks and data sets to adapt to different application scenarios. Second, the similarity measurement method in the calculation of attention weights can be selected and adjusted according to actual needs, such as cosine similarity, Hamming distance, etc. This flexibility enables this solution to be more widely applied to various complex graph network signal detection and alarm problems.

[0090] Adaptability and robustness: The adaptive threshold mechanism of the AHSNM layer not only enables the model to adapt to different signal characteristics, but also enhances the robustness of the model to noise and abnormal data by combining feature spatio-temporal information and graph structure information. Even in the case of low data quality or the existence of outliers, this solution can still maintain high accuracy and stability. In addition, the AHSNM model can automatically adjust the threshold to adapt to signal changes, thereby maintaining stable performance in various network environments.

[0091] Fusion of multi-source information: This solution can efficiently process graph network signals with spatiotemporal characteristics by combining the advantages of SNN and GNN. GNN focuses on processing graph structure data and can mine the complex relationships between nodes and edges; while SNN is good at processing time series data and can capture dynamic changes between nodes. At the same time, the proposed AHSNM model introduces the position encoding information of the graph structure. Finally, the AHSNM model uses ternary position adaptive pulse neurons to more accurately capture multi-source information. Multi-source information fusion makes this solution more efficient and accurate in processing spatiotemporal data.

[0092] Good bioinspiration: The design of the SNN model is inspired by the pulse emission mechanism of biological neurons, which makes this solution have significant advantages in bioinspiration. By simulating the processing method of the biological nervous system, this solution can process spatiotemporal data more naturally and simulate the information processing process of the human brain to a certain extent.

[0093] Unique advantages of AHSNM model:

[0094] Accurately capture features: The AHSNM model can more accurately capture the feature information in the network signal and improve the accuracy of signal processing by separating the ternary position-adaptive pulse neurons.

[0095] Position awareness: The model incorporates a position encoding strategy, which enables neurons to take both feature information and position information into account when receiving and processing information, thus enhancing the model’s ability to perceive the network topology.

[0096] High configurability: The parameters of the AHSNM model (such as thresholds, connection weights, pulse amplitudes, position encoding weights, etc.) can all be optimized through training, allowing the model to be flexibly configured according to different application scenarios to meet diverse needs.

[0097] In summary, this technical solution has significant advantages in terms of efficient processing of spatiotemporal data, powerful feature extraction capabilities, high flexibility and scalability, good bioinspiration, and strong robustness. These advantages make this solution have broad application prospects and great potential in the fields of power systems, intelligent transportation, social network analysis, etc.

[0098] Example 2

[0099] This embodiment provides a computer device, which may be a server or a terminal. Its internal structure diagram may be as follows: Figure 4As shown in the figure. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data in the method for detecting spatio-temporal signals of a pulsed graph neural network. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements the method for detecting spatio-temporal signals of a pulsed graph neural network in Embodiment 1.

[0100] Those skilled in the art can understand that Figure 4 the structure shown in the figure is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0101] Embodiment 3

[0102] This embodiment provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, it implements the method for detecting spatio-temporal signals of a pulsed graph neural network in Embodiment 1.

[0103] Embodiment 4

[0104] This embodiment provides a computer program product including a computer program, and when the computer program is executed by a processor, it implements the method for detecting spatio-temporal signals of a pulsed graph neural network in Embodiment 1.

[0105] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0106] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tapes, floppy disks, flash memories, optical memories, high-density embedded non-volatile memories, resistive random access memories (ReRAM), magnetoresistive random access memories (MRAM), ferroelectric random access memories (FRAM), phase change memories (PCM), graphene memories, etc. Volatile memories can include random access memory (RAM) or external cache memories, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0107] The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logics, data processing logics based on quantum computing, etc., without limitation.

[0108] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0109] Specific examples are used in this article to elaborate on the principles and implementation manners of the present application. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A pulse graph neural network spatiotemporal signal detection method, characterized in that: The pulse graph neural network spatiotemporal signal detection method comprises: Obtain graph structure data of various system networks, wherein each of the graph structure data includes static graph structure data and dynamic spatiotemporal data, wherein the static graph structure data is the fixed topological structure information of each node and edge in the system network, the static graph structure data includes a node feature matrix and an adjacency matrix, the adjacency matrix is ​​used to characterize the connection relationship between nodes, and the dynamic spatiotemporal data is the network signal data on each node that changes over time; Each of the graph structure data is taken as input, and a trained multi-source information fusion pulse graph neural network model is used to detect anomalies and identify anomaly categories, wherein the multi-source information fusion pulse graph neural network model includes a graph convolutional neural network and a pulse graph neural network, and the pulse graph neural network includes an input layer, an adaptive hybrid spatiotemporal pulse neuron model layer, an attention layer, a fully connected layer and an output layer connected in sequence, and the graph convolutional neural network is used to extract features from the static graph structure data to obtain the node representation after graph convolution, and the pulse graph neural network is used to perform anomaly detection based on the node representation of fused spatiotemporal information, and the node representation of fused spatiotemporal information is a vector obtained based on the node representation after graph convolution and the dynamic spatiotemporal data.

2. The pulse graph neural network spatiotemporal signal detection method according to claim 1, characterized in that: The training process of the multi-source information fusion impulse graph neural network model specifically includes: Acquire training samples, wherein the training samples include graph structure data samples and abnormal categories, and the graph structure data samples include static graph structure data samples and dynamic spatiotemporal data samples; Using a graph convolutional neural network to extract features from the static graph structure data sample to obtain a node representation after convolution of the sample graph; Performing pulse coding on the node representation after convolution of the sample graph and the dynamic spatiotemporal data sample respectively to obtain a sample node feature pulse sequence and a sample time series pulse sequence; The sample node feature pulse sequence and the sample time series pulse sequence are spliced ​​in the time dimension to obtain the node representation of the sample fusion spatiotemporal information; The input signal is processed by using an adaptive hybrid spatiotemporal pulse neuron model layer based on a dynamic threshold mechanism and a pulse generation rule to obtain a sample output pulse sequence, wherein the input signal is a vector obtained by concatenating the node representation of the sample fusion spatiotemporal information and the position pulse, and the position pulse is a vector obtained by processing the static graph structure data sample by using position coding and pulse coding; Using the attention layer based on the pulse-induced graph attention mechanism to update the node representation of the sample fusion spatiotemporal information, and using the updated node representation of the sample fusion spatiotemporal information as the new node representation of the sample fusion spatiotemporal information, return to step "using the adaptive hybrid spatiotemporal spike neuron model layer to process the input signal based on the dynamic threshold mechanism and the pulse generation rule to obtain the sample output pulse sequence"; Outputting a pulse sequence to the sample using a fully connected layer to obtain pulse sequences of different nodes; Detecting abnormal categories according to category values ​​using the output layer, wherein the category values ​​are values ​​determined according to pulse emission intensities and frequencies of pulse sequences of different nodes; When the loss function is minimized, the training of the multi-source information fused impulse graph neural network model is completed.

3. The pulse graph neural network spatiotemporal signal detection method according to claim 2, characterized in that: The pulse encoding is performed on the node representation after the convolution of the sample graph and the dynamic spatiotemporal data sample to obtain the sample node feature pulse sequence and the sample time series pulse sequence, which specifically includes: The Bernoulli function is used to perform pulse coding on the node representation after the convolution of the sample graph and the dynamic spatiotemporal data sample, respectively, to obtain a sample node feature pulse sequence and a sample time series pulse sequence.

4. The pulse graph neural network spatiotemporal signal detection method according to claim 2, characterized in that: The expression of the membrane potential update equation of the adaptive mixed spatiotemporal pulse neuron model in the adaptive mixed spatiotemporal pulse neuron model layer is: in, represents the membrane potential of the ith neuron at time t; λ represents the attenuation factor of the membrane potential; represents the membrane potential of the i-th neuron at time t-1; W0 represents the connection weight from the j-th neuron to the i-th neuron; represents the characteristic pulse amplitude of the jth neuron; represents the characteristic pulse of the jth neuron at time t; represents the position encoding weight, which is determined according to the relative position relationship between node k and node j; represents the position code of position k; represents the dynamic threshold of the i-th neuron at time t; δ(x) represents the ternary neuron mechanism function.

5. The pulse graph neural network spatiotemporal signal detection method according to claim 4, characterized in that: The expression of the dynamic threshold mechanism is: in, represents the dynamic threshold of the i-th neuron at time t; represents the dynamic threshold of the i-th neuron at time t-1; η is the threshold adjustment rate; μ Vi is the long-term average of the membrane potential of neuron i; γ is the input-dependent adjustment coefficient; |W tj | represents the absolute value of the weight.

6. The method for detecting spatiotemporal signals using a pulse graph neural network according to claim 4, characterized in that: The expression of the parameter update equation of the adaptive hybrid spatiotemporal pulse neuron model is: Among them, lr represents the learning rate; θ Δ,i V represents the update amount of the learnable parameters of the i-th neuron in the network; Δ,i Represents the change in membrane potential of the i-th neuron; represents the gradient of the loss function L with respect to the learnable parameter θ of the i-th neuron.

7. The pulse graph neural network spatiotemporal signal detection method according to claim 2, characterized in that: The attention layer is used to update the node representation of sample fusion spatiotemporal information based on the pulse-induced graph attention mechanism, including: Attention weights based on the sample output pulse sequence of the i-th node and the sample output pulse sequence of the j-th node; Update the node representation of the i-th node’s sample fusion spatiotemporal information according to the attention weight.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the pulse graph neural network spatiotemporal signal detection method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the pulse graph neural network spatiotemporal signal detection method described in any one of claims 1 to 7 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the pulse graph neural network spatiotemporal signal detection method described in any one of claims 1 to 7 is implemented.

Images