Power equipment fault risk analysis method based on multi-dimensional feature fusion

Abstract

The invention provides a power equipment fault risk analysis method based on multi-dimensional feature fusion. The method comprises the following steps: collecting time sequence monitoring data of a plurality of nodes in a power system; for the time sequence monitoring data of each node, the sliding window intercepts T continuous time steps each time to form a to-be-processed window segment, then each time step in the to-be-processed window segment, the time sequence monitoring data and a timestamp are fused to generate an enhanced feature vector, and a time sequence feature matrix formed by T enhanced feature vectors is obtained; constructing a node feature matrix based on the time sequence feature matrix of each node; on the basis of the relationship among a plurality of nodes in the power system, constructing a power utilization information transmission link topological graph, and further constructing a graph adjacency matrix; inputting the node feature matrix and the graph adjacency matrix into a preset fault risk analysis model, and determining a fault node; and outputting fault risk influence spreading range nodes by adopting an improved ant colony algorithm.

Description

Technical Field

[0001] This application relates to the field of smart grid technology, specifically to a method, system, electronic device, and storage medium for power equipment fault risk analysis based on multi-dimensional feature fusion. Background Technology

[0002] With the rapid advancement of new power system construction, the number of remote data acquisition devices deployed in power networks, such as smart meters, terminals, and concentrators, has exploded. These devices are widely distributed, covering complex physical environments from urban to rural areas, and are constantly exposed to diverse climates and electromagnetic interference, posing unprecedented challenges to their operational status monitoring and maintenance. Traditional maintenance models relying on manual inspections and post-incident troubleshooting are no longer adequate for the new situation of a large number of devices, hidden faults, and high real-time requirements. This has resulted in not only low maintenance efficiency but also a sharp increase in labor and time costs.

[0003] Meanwhile, the system gathers operational data from massive amounts of equipment daily, including multi-dimensional information such as voltage, current, communication status, and event logs. While the data volume is enormous, its value density is relatively low. Due to the data's multi-source, heterogeneous, and time-series-dependent characteristics, relying solely on threshold alarms or traditional statistical analysis methods is insufficient to deeply uncover hidden trends in equipment health degradation and early signs of failure. This lack of analytical capability leads to a delay in perceiving potential equipment risks, hindering accurate fault location and early warning. Consequently, fault detection often depends on existing business impacts, such as failed metering data collection or abnormal power consumption, directly affecting data collection success rate and data quality. Summary of the Invention

[0004] This application provides a method for analyzing the fault risk of power equipment based on multi-dimensional feature fusion, which enables the identification of the spread of the impact of power equipment fault risks, supports real-time monitoring, anomaly identification and efficient source tracing of remote power equipment faults, and promotes the intelligent operation and maintenance of new power systems.

[0005] To achieve the above objectives, this application adopts the following technical solution:

[0006] Firstly, this application provides a method for power equipment fault risk analysis based on multi-dimensional feature fusion, the method comprising:

[0007] Collect time-series monitoring data from multiple nodes in the power system, where multiple nodes represent different electrical devices in the power system;

[0008] For the time-series monitoring data of each node, the sliding window extracts T consecutive time steps each time to form a window segment to be processed. Then, for each time step within the window segment to be processed, the time-series monitoring data and timestamps are fused to generate an enhanced feature vector. The window segment to be processed is then transformed into a time-series feature matrix composed of T enhanced feature vectors.

[0009] Based on the time-series feature matrix of each node, the node feature vector of each node is extracted, and the node feature vectors of all nodes are combined to construct the node feature matrix; then, based on the physical connection and communication relationship between multiple nodes in the power system, the power consumption information transmission link topology graph is constructed, and based on the topology graph, the graph adjacency matrix is ​​constructed.

[0010] Input the node feature matrix and graph adjacency matrix into the preset fault risk analysis model to determine the faulty nodes;

[0011] Starting from the faulty node, based on the topology graph, an improved ant colony algorithm is used to simulate the propagation path of the fault risk, identify and output the nodes within which the fault risk affects the spread.

[0012] Secondly, this application provides a power equipment fault risk analysis system based on multi-dimensional feature fusion, the system comprising:

[0013] The data acquisition and processing module is used to collect time-series monitoring data from multiple nodes in the power system, where multiple nodes represent different electrical devices in the power system;

[0014] The time-series feature construction module is used to extract T consecutive time steps from the time-series monitoring data of each node in a sliding window to form a window segment to be processed. For each time step in the window segment to be processed, the time-series monitoring data and timestamp are fused to generate an enhanced feature vector, and the window segment to be processed is transformed into a time-series feature matrix composed of T enhanced feature vectors.

[0015] The node feature and graph structure modeling module is used to extract the node feature vector of each node based on the temporal feature matrix of each node, combine the node feature vectors of all nodes to construct the node feature matrix; then, based on the physical connection and communication relationship between multiple nodes in the power system, it constructs the power information transmission link topology graph, and based on the topology graph, it constructs the graph adjacency matrix.

[0016] The fault risk analysis model module is used to input the node feature matrix and graph adjacency matrix into a preset fault risk analysis model to identify faulty nodes.

[0017] The risk propagation analysis module is used to simulate the propagation path of fault risk starting from the faulty node, based on the topology graph, and using an improved ant colony algorithm to identify and output the nodes within which the fault risk affects the spread.

[0018] Thirdly, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the program, implements the power equipment fault risk analysis method based on multi-dimensional feature fusion of the first aspect.

[0019] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium including storage of a computer program or instructions, which, when executed, cause the power equipment fault risk analysis method of the first aspect based on multi-dimensional feature fusion to be executed.

[0020] In this embodiment, by constructing a power information transmission link topology diagram reflecting the relationship between physical connections and information flow, and using a graph adjacency matrix as a benchmark for information propagation and impact analysis, the transmission path of fault risks in complex systems can be effectively modeled. This solves the problems of inaccurate fault location and delayed risk perception caused by neglecting equipment correlation in traditional point-based monitoring methods. On this basis, long-term and short-term behavioral features and spatial correlation features of nodes are extracted by temporal convolutional networks and graph neural networks, respectively. An attention mechanism and a hybrid expert network are used for dynamic deep fusion, achieving high-dimensional representation and accurate measurement of abnormal node states. Finally, an improved ant colony algorithm is used to simulate the risk propagation process starting from a fault node on the topology diagram. The scope of impact spread is identified based on pheromone concentration, thus completing the leap from single fault diagnosis to systemic risk warning, significantly improving the comprehensiveness and predictability of power equipment fault risk analysis.

[0021] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the power equipment fault risk analysis method based on multi-dimensional feature fusion provided in this application embodiment. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. At the same time, in the description of the embodiments of this application, the terms "first," "second," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0025] Figure 1 This is a flowchart illustrating the power equipment fault risk analysis method based on multi-dimensional feature fusion provided in this application embodiment.

[0026] The process of the power equipment fault risk analysis method based on multi-dimensional feature fusion is as follows:

[0027] Step S1. Collect time-series monitoring data from multiple nodes in the power system, where multiple nodes represent different electrical devices in the power system.

[0028] Among them, multiple nodes represent different electrical devices in the power system, specifically: terminal equipment (smart energy meters, distribution transformer terminals, substation terminals, low-voltage concentrators, low-voltage data collectors, etc.); channels (230MHz wireless private network, wireless public network, power line carrier communication, wired network, public switched telephone network).

[0029] In addition, the time-series monitoring data includes multiple monitoring indicators, such as daily electricity consumption, load extremes, voltage, current, metering abnormal event data, transparent metering event data, communication log data between the terminal and the main station of the acquisition system, temperature, time, user address, and the area to which the user belongs, etc., without specific restrictions.

[0030] In this embodiment of the application, after collecting time-series monitoring data from multiple nodes in the power system, the method further includes: standardizing the time-series monitoring data to obtain standardized time-series monitoring data for each node, specifically:

[0031] The time-series monitoring data for each node includes multiple monitoring indicator values ​​collected at multiple time points;

[0032] For any one of the multiple monitoring indicators, the values ​​at all time points are extracted to form the original indicator sequence. Then, the Z-score standardization method is used to standardize the original indicator sequence to obtain the standardized original indicator sequence.

[0033] Based on the standardized original indicator sequence of all monitoring indicators in each node, standardized time-series monitoring data for each node is obtained.

[0034] Specifically, the standardized processing calculation formula is as follows: Where, x t The mean(x) represents the original monitoring data of the node at time t. t ) and std(x t The values ​​are the mean and standard deviation of the original monitoring data, respectively. This is the standardized data.

[0035] The main reason for standardization is to align the spatial mapping and representation of multiple monitoring indicators. That is, to transform various data from different places (such as terminal devices, channels, and environment), whether they are text (such as the "normal" status of the device), numerical values ​​in different units (such as voltage 220V, current 5A), or mixed information, into a unified format (such as converting them all to numerical values), and then adjust their size and meaning (alignment), so that the data that originally "said their own words" can "communicate in the same language", and can be analyzed together or input into the model for further processing.

[0036] Step S2. For the time series monitoring data of each node, the sliding window extracts T consecutive time steps each time to form a window segment to be processed. Then, for each time step in the window segment to be processed, the time series monitoring data and timestamp are fused to generate an enhanced feature vector. The window segment to be processed is then transformed into a time series feature matrix composed of T enhanced feature vectors.

[0037] Corresponding to the above standardization process, step S2 can also be as follows: For the standardized time-series monitoring data of each node, a sliding window extracts T consecutive time steps each time to form a window segment to be processed. Then, for each time step within the window segment to be processed, the standardized time-series monitoring data and timestamps are fused to generate an enhanced feature vector, transforming the window segment to be processed into a time-series feature matrix composed of T enhanced feature vectors. In this embodiment, fusing standardized time-series monitoring data and timestamps to generate enhanced feature vectors includes:

[0038] For each time step within the window segment, the standardized time-series monitoring data corresponding to each time step, the start timestamp and the end timestamp of data collection are concatenated to form the enhanced feature vector corresponding to each time step.

[0039] That is, for the standardized time-series monitoring data stream of each node, a sliding window of fixed length T (number of time steps) is defined and slides forward on the time axis with a preset step size S. After each slide, the data of the T consecutive time steps covered by the window constitute a window segment to be processed. For each time step in the window segment, a feature fusion operation is performed: the standardized time-series monitoring data vector of the time step, the operation event type encoding vector extracted from the log corresponding to the time step, and the start and end timestamps of the data collection period are concatenated to generate an enhanced feature vector. All T time steps in the window segment are arranged in chronological order, and the enhanced feature vectors generated by each time step are stacked as a row of a matrix to construct a two-dimensional matrix with dimensions of T rows × M columns, which is called the time-series feature matrix of the node in the current window segment. Here, M is the total dimension of the enhanced feature vectors, and T is the window length.

[0040] For example, a sliding window of size T and step size S is selected. A node detects data X within one window, where x... t ∈X, t∈T, that is, x t The data is monitored at the t-th time node; x t timestamp s at time t-1 t The timestamp at that moment e t and various types of sequence op t The concatenation is performed to obtain the enhanced feature vector H at that moment. t H t ={x t ,op t ,s t ,e t}, where s t -e t This indicates the time interval between two data monitoring sessions.

[0041] Step S3. Based on the time-series feature matrix of each node, extract the node feature vector of each node, combine the node feature vectors of all nodes to construct the node feature matrix; then, based on the physical connection and communication relationship between multiple nodes in the power system, construct the power information transmission link topology graph, and based on the topology graph, construct the graph adjacency matrix.

[0042] In this embodiment of the application, based on the temporal feature matrix of each node, the node feature vector of each node is extracted, including: performing mean pooling or max pooling on the enhanced feature vectors of all time steps in the temporal feature matrix of each node, and using the pooling result as the corresponding node feature vector.

[0043] In this embodiment of the application, a graph adjacency matrix is ​​constructed based on the topological graph, including:

[0044] Generate a basic adjacency matrix based on the connection relationships between nodes in the topological graph;

[0045] By adding an identity matrix to the basic adjacency matrix, we obtain the graph adjacency matrix.

[0046] Specifically, the basic adjacency matrix is ​​a symmetric matrix A. If there is an edge between node i and node j, then the element a in the adjacency matrix is... ij and a ji Set all values ​​to 1; otherwise, set them to 0. For example, for an undirected graph with 3 nodes, if node 1 is connected to nodes 2 and 3, and node 2 is connected to node 3, then the adjacency matrix is:

[0047] To preserve self-features in subsequent graph convolutions, a self-loop is added to the adjacency matrix A, which means adding the adjacency matrix A to the identity matrix E of the same size.

[0048] Step S4. Input the node feature matrix and graph adjacency matrix into the preset fault risk analysis model to determine the faulty nodes.

[0049] In this embodiment, the preset fault risk analysis model includes a temporal convolutional network, a multi-layer graph neural network, an attention mechanism module, a hybrid expert network, and a fault node determination module.

[0050] Temporal convolutional networks are used to extract the long-term and short-term temporal variation patterns of each node based on the node feature matrix and output the temporal feature representation of each node.

[0051] Multilayer graph neural networks are used to combine the features of each node's neighboring nodes with the node feature matrix and graph adjacency matrix to output a multilayer spatial feature representation of each node.

[0052] The attention mechanism module is used to dynamically fuse the temporal feature representation and multi-layer spatial feature representation of each node using the attention mechanism to generate the fused feature representation of each node.

[0053] A multi-layer hybrid expert network is used to compute the multi-layer anomaly probability of each node based on the fused feature representation of each node.

[0054] The fault node determination module is used to output the anomaly probability value of each node based on the multi-level anomaly probability of each node, combined with dynamic routing weighted summation, and to identify the node with the highest anomaly probability value as the fault node.

[0055] Specifically, step S4 can be understood as:

[0056] Step 1: Based on the node feature matrix, extract the long-term and short-term time series change patterns of each node and output the time series feature representation of each node.

[0057] That is, the node feature matrix input data is used as input, and TCN (Temporal Convolutional Network) is used to extract the long and short term feature representations of the node temporal features, and output the long and short term feature representation data.

[0058] For each node v i Its original features in the current sliding window c The compression includes latent vectors containing both temporal and spatial information, where T is the length of the sliding window and D is the value of x. t ,op t ,s t ,e t The length after concatenation. Temporal features are obtained using TCN convolution, and causal convolutional layers are defined to capture the local change trends of nodes:

[0059]

[0060] in, Representing the local temporal features of a node, LayerNorm(·) represents the normalization operation across all feature dimensions, k represents the kernel size, and d represents the dilation coefficient, used to gradually expand the receptive field to extract longer-term trends. This represents the weight matrix.

[0061] Temporal features Z are obtained by stacking multiple convolutional layers with different dilation coefficients. time :

[0062]

[0063] Among them, Num TCN This indicates the number of convolutional layers.

[0064] Step 2: Based on the node feature matrix and the graph adjacency matrix, collect the features of each node's neighboring nodes and output a multi-layer spatial feature representation of each node.

[0065] That is, the node feature matrix and graph adjacency matrix are used as input data, and a multi-layer GNN network is used to aggregate the adjacency association data of each node in the whole graph, and output the spatial feature data of each layer of GNN network nodes.

[0066] Define an L-layer GNN network, where each layer aggregates information about each node's neighbors. To distinguish devices at different layers in the network, the temporal features Z extracted by the TCN are... time The data is concatenated with device information and used as input. For the l-th layer GNN, attention aggregation is used to combine node v... iThe features of all neighboring nodes are concatenated, and the formula is:

[0067]

[0068] in, This represents the aggregation of features of all neighbors of node i after the computation of the first layer of the GNN, where N(i) represents the features of node v. i The set of all neighboring nodes, Indicates v i neighbor node v j The features of the GNN at layer l-1, a ij These are the attention weights. The Softmax function is used to calculate the attention weights of each neighbor node in the interval [0,1], i.e., to perform normalization. LeakyReLU(·) is a non-linear activation function that allows weak propagation of negative values. This activation function is defined as:

[0069]

[0070] The parameter α is set to 0.2. T It is a learnable attention parameter vector; This indicates that node v i and node v j The features in the (l-1)th layer are spliced ​​together.

[0071] Aggregator node v i After obtaining the neighbor node information, the neighbor information of this layer is aggregated. The target node is injected, while retaining some feature information of the target node from the previous GNN layer. A gating network is used to determine the ratio of new to old information fusion, and the formula is as follows:

[0072]

[0073] in, This is the update gate, which generates a gated vector between [0,1] that determines the amount of historical information retained and the amount of new information absorbed. σ(·) is the sigmoid function, which compresses the elements to the range [0,1]. The weight matrix W... z ∈R d×2d and bias b z ∈R d These are used to calculate parameters and enhance the flexibility of the model. ⊙ represents element-wise multiplication; This indicates the concatenation of vectors. Indicating a candidate state involves aggregating messages from the current layer. Features of the previous layer The concatenation process, where W and b are the weight matrix and bias parameter respectively, generates new features through linear and nonlinear activations. The ReLU activation function is defined as ReLU(x) = max(0,x). Through the aggregation of neighbor node information and the gating unit described above, the target node v can be obtained. i Spatial characteristics.

[0074] Step 3: Use the attention mechanism to dynamically fuse the temporal feature representation and multi-layer spatial feature representation of each node to generate the fused feature representation of each node.

[0075] That is, based on the spatial feature representation of each GNN network node, the attention mechanism is used to dynamically fuse the temporal feature representation and the spatial feature representation to output fused feature data;

[0076] Define the attention matrix β:

[0077]

[0078] Among them, W q and W k It is a learnable query and key matrix, where d is the dimension of the two features.

[0079] The feature-weighted fusion formula is defined as follows:

[0080]

[0081] This represents the node features of the current time slice and the current level. That is, if there are L layers of GNN, there are corresponding L nodes.

[0082] Step 4: Calculate the multi-level anomaly probability of each node based on the fusion feature representation of each node.

[0083] Based on fused feature representations, a multilayer perceptron is used to calculate independent anomaly prediction values ​​for each node in each layer of the GNN network, outputting the anomaly probability values ​​for each node at different levels. Traditional graph neural networks only focus on and use the results of the deepest layers, leading to the loss of effective shallow network feature information as the layers progress. In fact, graph neural networks progressively cover the information of each node's neighboring nodes, nodes with a one-hop interval, and so on, until they encompass the entire graph. Therefore, to obtain information at different levels, the model designs a MoE hybrid expert module, creating a unique expert network for each layer of the graph neural network to obtain information at different levels. Then, the outputs of all expert networks are weighted and summed to obtain the probability of anomalies at each node.

[0084] Based on the latent features of each node in layer l, a lightweight two-layer fully connected multilayer perceptron is designed. The anomaly probability of each node in the current graph neural network layer is: The calculation formula is:

[0085]

[0086] Among them, the independent trainable parameters of each layer These are their biases.

[0087] This forces each hybrid expert network to provide reasonable predictions, preventing the model from relying solely on a subset of layers. Define the loss function:

[0088]

[0089] Where γ is the attenuation coefficient, Φ is the single-layer cross-entropy loss, and N is the total number of nodes in the link.

[0090] Dynamic weight calculation is performed using an attention mechanism, mapping the output of the last layer of the graph neural network to a query vector. i Map the output of the current layer to a key vector. Formulate this process:

[0091]

[0092] Wherein, the query matrix Q∈R d×d The output of the final layer is converted into a query vector, where the key matrix K∈R. d×d The output of the current layer is converted into a key vector, where the attention vector q∈R. d Used to calculate attention scores.

[0093] Attention score Normalization is performed:

[0094]

[0095] Define an entropy loss function to distribute weights evenly, and add a variance term to prevent gated routing weights in MoE. Too concentrated, the formula is:

[0096]

[0097] in, Represents the entropy term. This represents the variance of the weights of the i-th node across all layers (a total of L layers), where L is the total number of hybrid expert layers, and l is the index of the expert layer.

[0098] Step 5: Based on the multi-level anomaly probability of each node, combined with dynamic routing weighted summation, output the anomaly probability value of each node, and take the node with the highest anomaly probability value as the fault node.

[0099] That is, the final anomaly probability is obtained by weighting and summing the node anomaly probabilities calculated by the hybrid expert network at the corresponding level based on the weights of each level.

[0100]

[0101] If the probability of an anomaly of a node is greater than P, then the prediction is that the node is faulty or abnormal.

[0102] Based on the prediction results weighted by the MoE hybrid expert network and the labeled data y in the sample i Design the main loss function L main Considering the limited number of outliers in the sample and the presence of some critical devices or nodes in the electricity information transmission link topology, a binary cross-entropy method is introduced in the main loss function, which combines category-weighted and device-criticality-weighted cross-entropy:

[0103]

[0104] in:

[0105]

[0106] Where, ω pos and ω critical These are two hyperparameters, representing the weighted average when the node actually exhibits an anomaly and when the node is a critical node, respectively. Ultimately, the loss function is:

[0107]

[0108] To prevent overfitting and improve the model's generality, L2 regularization is added. α and β are hyperparameters.

[0109] Step S5. Starting from the faulty node, based on the topology graph, use the improved ant colony algorithm to simulate the propagation path of the fault risk, identify and output the nodes within the scope of the fault risk's impact.

[0110] In this embodiment, starting from the faulty node, an improved ant colony algorithm is used based on the topology graph to simulate the propagation path of the fault risk, identify and output the nodes within the scope of the fault risk's spread, including:

[0111] S51. Initially release artificial ants at the faulty node, assign the same initial pheromone concentration value to each edge in the topology graph, and restrict the artificial ants from moving along the predetermined data transmission path direction in the topology graph;

[0112] S52. In each iteration, the artificial ant moves from the current node to an unvisited, reachable neighbor node based on the path selection probability. The path selection probability is calculated by the pheromone concentration value on the corresponding edge and the heuristic information. The heuristic information is determined based on the anomaly probability of the current node and its neighboring nodes and is used to guide the artificial ant to move to the region with a high anomaly probability.

[0113] S53. In each iteration, after the artificial ant completes the path selection, it globally updates the pheromone concentration of each edge in the topology graph. The update process includes proportionally evaporating the pheromone concentration on all edges according to a preset evaporation coefficient to simulate the natural decay of information. Based on the path selection of the artificial ant in this iteration, the corresponding amount of pheromone is added to the edges that the artificial ant actually traverses.

[0114] S54. Repeat steps S52 to S53 until the preset number of iterations is reached. Nodes connected to edges whose final pheromone concentration exceeds the preset threshold are identified as nodes within the scope of the fault risk spread and are output.

[0115] The specific process is as follows:

[0116] Step 1: Initially release artificial ants at the faulty node and restrict the artificial ants to move only along the data transmission path of the power information transmission link topology map constructed in Step C;

[0117] Define the probability of an ant choosing a node:

[0118]

[0119] in, τ represents the probability that the artificial ant k moves from the current node i to the node j; ij (t) represents the pheromone concentration along the path from i to j at time t. Initially, the pheromone concentration along each path is a small constant and equal. As the artificial ants are updated, the pheromone concentration changes over time. α is the information factor; β is the heuristic factor; J allowqd Let η represent the set of all nodes that ant k has not yet visited, and which are reachable from the current node i along the actual data transmission path; ij (t) is the path visibility from i to j, which is usually represented by the reciprocal of the weight of the edge from i to j.

[0120] Step 2: Define ant pheromone concentration and visibility, that is, calculate the average probability of anomalies of neighboring nodes, as a measure of the degree of anomalous clustering in the neighborhood of a node, and induce artificial ants to move towards nodes with high anomalous association.

[0121] Pheromone concentration is usually represented as a numerical matrix τ ij (t), where i and j represent two states or nodes in the solution space; t represents the number of algorithm iterations (time steps); τ ij (t) represents the pheromone concentration along the path from node i to node j at iteration t, specifically defined as follows: τ ij (t+1)=(1-ρ)·τ ij (t)+Δτ ij .

[0122] Where, τ ij It is the pheromone concentration along path i→j, ρ is the volatility coefficient, and Δτ ij This is a new pheromone added in this iteration.

[0123] Visibility definition:

[0124]

[0125] Where λ1 and λ2 are weighting coefficients, respectively measuring the proportion of initial anomaly prediction results and anomaly local consistency; J neighbor Let Q represent the set of all neighboring nodes of node i, and let Q represent the number of all neighbors of node i, i.e., Q = |J|. neighbor |

[0126] Step 3: By analyzing the changes in pheromone concentration and the coverage statistics of ant pathways, extract the coverage nodes whose cumulative pheromone concentration exceeds the set pheromone concentration threshold, and output the risk impact spread range nodes.

[0127] The pheromone concentration is updated using an ant-week strategy. That is, after each ant k completes its iteration, the pheromone concentration is updated for all paths it has traversed. The formula is:

[0128]

[0129] Where ρ represents the volatile factor, which is the simulated natural evaporation of pheromones, and (1-ρ) represents the concentration of residual pheromones; This represents the increase in pheromone concentration caused by ant k choosing the path from node i to node j, calculated as follows:

[0130]

[0131] Where E represents the path constant, L kThis represents the sum of the weights of the paths traversed by ant k.

[0132] When ant k's J allowed When the set is empty, that is, when there is no reachable abnormal node that ant k has not yet visited, it is considered that one round trip of ant k has ended, and all the paths that k has walked are an abnormal propagation path.

[0133] Finally, after all the ants have completed the iteration, the nodes covered by the artificial ants released from the abnormal nodes are used as the risk impact spread range nodes, and the risk impact spread range nodes are output.

[0134] In summary, in this embodiment, by constructing a power information transmission link topology map reflecting the relationship between physical connections and information flow, and using a graph adjacency matrix as the benchmark for information propagation and impact analysis, the transmission path of fault risks in complex systems can be effectively modeled. This solves the problems of inaccurate fault location and delayed risk perception caused by neglecting equipment correlation in traditional point-based monitoring methods. Furthermore, by extracting the long-term and short-term behavioral features and spatial correlation features of nodes using temporal convolutional networks and graph neural networks respectively, and by employing an attention mechanism and a hybrid expert network for dynamic deep fusion, high-dimensional representation and accurate measurement of abnormal node states are achieved. Finally, by simulating the risk propagation process starting from a fault node on the topology map using an improved ant colony algorithm, and identifying the scope of impact spread based on pheromone concentration, a leap from single fault diagnosis to systemic risk warning is achieved, significantly improving the comprehensiveness and predictability of power equipment fault risk analysis.

[0135] The above combination Figure 1 This paper describes the power equipment fault risk analysis method based on multi-dimensional feature fusion provided in the embodiments of this application. The following describes the power equipment fault risk analysis system based on multi-dimensional feature fusion provided in the embodiments of this application.

[0136] The system includes: a data acquisition and processing module, a time series feature construction module, a node feature and graph structure modeling module, a fault risk analysis model module, and a risk propagation analysis module, as detailed below.

[0137] The data acquisition and processing module is used to collect time-series monitoring data from multiple nodes in the power system, where multiple nodes represent different electrical devices in the power system;

[0138] The time-series feature construction module is used to extract T consecutive time steps from the time-series monitoring data of each node in a sliding window to form a window segment to be processed. Then, for each time step in the window segment to be processed, the time-series monitoring data and timestamps are fused to generate an enhanced feature vector, and the window segment to be processed is transformed into a time-series feature matrix composed of T enhanced feature vectors.

[0139] The node feature and graph structure modeling module is used to extract the node feature vector of each node based on the temporal feature matrix of each node, combine the node feature vectors of all nodes to construct the node feature matrix; then, based on the physical connection and communication relationship between multiple nodes in the power system, it constructs the power information transmission link topology graph, and based on the topology graph, it constructs the graph adjacency matrix.

[0140] The fault risk analysis model module is used to input the node feature matrix and graph adjacency matrix into a preset fault risk analysis model to identify faulty nodes.

[0141] The risk propagation analysis module is used to simulate the propagation path of fault risk starting from the faulty node, based on the topology graph, and using an improved ant colony algorithm to identify and output the nodes within which the fault risk affects the spread.

[0142] Furthermore, the specific implementation of the above system is basically similar to the method implementation, so the description is relatively simple. For relevant details, please refer to the description of the method implementation. Moreover, it should be noted that in the various modules of the system of this application, the components are logically divided according to the functions they are to perform. However, this application is not limited to this and can re-divide or combine the components as needed.

[0143] In another aspect, the present invention provides an electronic device for implementing the above-described method for power equipment fault risk analysis based on multi-dimensional feature fusion. This electronic device is not limited to a terminal device or server in a system. The electronic device includes, but is not limited to, a memory and a processor. The memory stores a computer program, and the processor is configured to execute the steps in any of the above-described method embodiments via the computer program.

[0144] In another aspect, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in various optional embodiments of the power equipment fault risk analysis method based on multi-dimensional feature fusion described above. The computer program is configured to execute the steps in any of the above method embodiments during runtime.

[0145] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than those shown in the embodiments and still achieve the desired result. Furthermore, the specific order or sequential order shown in the drawings is not necessarily required to achieve the desired result; in some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0146] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for power equipment fault risk analysis based on multi-dimensional feature fusion, characterized in that, include: Collect time-series monitoring data from multiple nodes in a power system, wherein the multiple nodes represent different electrical devices in the power system; For the time-series monitoring data of each node, the sliding window extracts T consecutive time steps each time to form a window segment to be processed. For each time step in the window segment to be processed, the time-series monitoring data and timestamp are fused to generate an enhanced feature vector. The window segment to be processed is then transformed into a time-series feature matrix composed of T enhanced feature vectors. Based on the time-series feature matrix of each node, the node feature vector of each node is extracted, and the node feature vectors of all nodes are combined to construct a node feature matrix; then, based on the physical connection and communication relationship between the multiple nodes in the power system, a power consumption information transmission link topology graph is constructed, and based on the topology graph, a graph adjacency matrix is ​​constructed. Input the node feature matrix and the graph adjacency matrix into a preset fault risk analysis model to determine the faulty nodes; Starting from the faulty node, based on the topology graph, an improved ant colony algorithm is used to simulate the propagation path of the fault risk, identify and output the nodes within the scope of the fault risk's spread.

2. The power equipment fault risk analysis method based on multi-dimensional feature fusion according to claim 1, characterized in that, After collecting time-series monitoring data from multiple nodes in the power system, the method further includes: The time-series monitoring data for each node includes multiple monitoring indicator values ​​collected at multiple time points; For any one of the multiple monitoring indicators, the values ​​at all time points are extracted to form the original indicator sequence. Then, the Z-score standardization method is used to standardize the original indicator sequence to obtain the standardized original indicator sequence. Based on the standardized original indicator sequence of all monitoring indicators in each node, the standardized time-series monitoring data of each node is obtained. Correspondingly, For each node's standardized time-series monitoring data, the sliding window extracts T consecutive time steps each time to form a window segment to be processed. Then, for each time step within the window segment to be processed, the standardized time-series monitoring data and timestamps are fused to generate an enhanced feature vector. The window segment to be processed is then transformed into a time-series feature matrix composed of T enhanced feature vectors.

3. The power equipment fault risk analysis method based on multi-dimensional feature fusion according to claim 2, characterized in that, The process of fusing standardized time-series monitoring data and timestamps to generate enhanced feature vectors includes: For each time step within the window segment, the standardized time-series monitoring data corresponding to each time step, the start timestamp and the end timestamp of data collection are concatenated to form the enhanced feature vector corresponding to each time step.

4. The power equipment fault risk analysis method based on multi-dimensional feature fusion according to claim 1, characterized in that, The extraction of node feature vectors for each node based on the temporal feature matrix of each node includes: For each node, the enhanced feature vectors of all time steps in the temporal feature matrix are subjected to mean pooling or max pooling, and the pooling results are used as the corresponding node feature vectors.

5. The power equipment fault risk analysis method based on multi-dimensional feature fusion according to claim 1, characterized in that, The construction of the graph adjacency matrix based on the topological graph includes: Based on the connection relationships between nodes in the topology graph, a basic adjacency matrix is ​​generated; By adding an identity matrix to the basic adjacency matrix, a graph adjacency matrix is ​​obtained.

6. The power equipment fault risk analysis method based on multi-dimensional feature fusion according to claim 1, characterized in that, The preset fault risk analysis model includes a temporal convolutional network, a multi-layer graph neural network, an attention mechanism module, a hybrid expert network, and a fault node determination module; The temporal convolutional network is used to extract the long and short-term temporal change patterns of each node based on the node feature matrix, and output the temporal feature representation of each node. The multi-layer graph neural network is used to combine the features of each node's neighboring nodes based on the node feature matrix and the graph adjacency matrix, and output a multi-layer spatial feature representation of each node. The attention mechanism module is used to dynamically fuse the temporal feature representation and multi-layer spatial feature representation of each node using the attention mechanism to generate a fused feature representation of each node. The multi-layer hybrid expert network is used to calculate the multi-layer anomaly probability of each node based on the fused feature representation of each node. The fault node determination module is used to output the anomaly probability value of each node based on the multi-level anomaly probability of each node, combined with dynamic routing weighted summation, and to identify the node with the highest anomaly probability value as the fault node.

7. The power equipment fault risk analysis method based on multi-dimensional feature fusion according to claim 1, characterized in that, Starting from the faulty node, and based on the topology graph, an improved ant colony algorithm is used to simulate the propagation path of the fault risk, identify and output the nodes within the scope of the fault risk's spread, including: S51. Initially release artificial ants at the faulty node, assign the same initial pheromone concentration value to each edge of the topology graph, and restrict the artificial ants from moving along a predetermined data transmission path on the topology graph; S52. In each iteration, the artificial ant moves from the current node to an unvisited, reachable neighbor node based on the path selection probability. The path selection probability is calculated by the pheromone concentration value on the corresponding edge and heuristic information. The heuristic information is determined based on the anomaly probability of the current node and its neighboring nodes and is used to guide the artificial ant to move to the region with a high anomaly probability. S53. In each iteration, after the artificial ant completes the path selection, it globally updates the pheromone concentration of each edge in the topology graph; wherein, the update process includes the pheromone concentration on all edges being proportionally evaporated according to a preset evaporation coefficient to simulate the natural decay of information; and then, based on the path selection of the artificial ant in this iteration, the corresponding amount of pheromone is added to the edges actually traversed by the artificial ant. S54. Repeat steps S52 to S53 until the preset number of iterations is reached. Nodes connected to edges whose final pheromone concentration exceeds the preset threshold are identified as nodes within the scope of the fault risk spread and are output.

8. A power equipment fault risk analysis system based on multi-dimensional feature fusion, characterized in that, The system includes: The data acquisition and processing module is used to acquire time-series monitoring data from multiple nodes in the power system, wherein the multiple nodes represent different electrical devices in the power system; The time-series feature construction module is used to extract T consecutive time steps each time from the time-series monitoring data of each node using a sliding window to form a window segment to be processed. For each time step in the window segment to be processed, the time-series monitoring data and timestamp are fused to generate an enhanced feature vector, and the window segment to be processed is transformed into a time-series feature matrix composed of T enhanced feature vectors. The node feature and graph structure modeling module is used to extract the node feature vector of each node based on the temporal feature matrix of each node, combine the node feature vectors of all nodes to construct the node feature matrix; then, based on the physical connection and communication relationship between the multiple nodes in the power system, it constructs the power consumption information transmission link topology graph, and based on the topology graph, it constructs the graph adjacency matrix. The fault risk analysis model module is used to input the node feature matrix and the graph adjacency matrix into a preset fault risk analysis model to determine faulty nodes; The risk propagation analysis module is used to simulate the propagation path of the fault risk based on the topology graph, starting from the fault node, and to identify and output the nodes within the scope of the fault risk's impact.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the power equipment fault risk analysis method based on multi-dimensional feature fusion as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a computer program or instructions that, when executed, cause the method as described in any one of claims 1-7 to be performed.

Images