Power equipment fault prediction method and system

Abstract

The invention provides a power equipment fault prediction method and system. The method comprises the steps of performing multi-modal fusion on acquired historical data of power equipment to obtain feature vectors of all modals; fusing the feature vectors by adopting principal component analysis to obtain fusion vectors; constructing the fusion vector into a graph to obtain a space-time diagram; learning spatial features of the spatial information extraction data by using a space-time diagram in a graph convolution mode to obtain a spatial feature representation vector; fusing the spatial feature representation vector in combination with the attention weight to generate an attention fusion vector; learning time sequence information by using a gating loop unit to obtain a time characteristic representation vector; performing weighted feature fusion on the spatial feature representation vector and the time feature representation vector to obtain a space-time fusion representation vector; and processing the space-time fusion representation vector by using a multi-layer perceptron, and outputting power equipment fault prediction. Based on the method, the invention further provides a power equipment fault prediction system. According to the invention, the accuracy and robustness of power equipment fault prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment fault prediction, and in particular relates to a power equipment fault prediction method and system. Background Art

[0002] With the continuous development of the power system, the types of equipment and operating environment of power grid companies have become more and more complex, and equipment failure prediction has become one of the key tasks to ensure the stable operation of the power grid. During the operation of power equipment, a large amount of multimodal data will be generated, including sensor data, operating status data, etc. These data often have the characteristics of diverse structures, large data volume, and strong real-time performance.

[0003] Traditional equipment fault prediction methods mainly rely on statistical analysis and classical machine learning techniques. These methods usually require a lot of expert knowledge and historical data, and have limited performance when dealing with complex spatiotemporal data. With the increase in data volume and complexity, and the rapid development of big data and artificial intelligence technologies, data-driven fault prediction methods have gradually attracted attention and application. Graph Neural Networks (GNNs), as a deep learning model that can process graph-structured data, show great potential in the field of equipment fault prediction. GNNs can capture complex spatial dependencies by modeling the relationships between nodes and edges. However, when simply relying on graph neural networks to process time series data, the dynamic change characteristics in the time dimension are often ignored. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a method and system for predicting power equipment faults, so as to improve the accuracy and robustness of power equipment fault prediction.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for predicting faults of electric power equipment comprises the following steps:

[0007] Acquire historical data of power equipment, perform multimodal data fusion on the historical data of power equipment to obtain feature vectors of each mode; fuse the feature vectors of each module by principal component analysis to obtain a fusion vector; construct the fusion vector into a graph by cosine similarity to obtain a feature adjacency matrix, and generate a spatiotemporal graph;

[0008] Using the obtained spatiotemporal graph, spatial information is learned by graph convolution, and the spatial features of the data are extracted to obtain a spatial feature representation vector; the spatial feature representation vector is fused with the attention weight by using a temporal attention network to generate a time-series enhanced attention fusion vector; the temporal information is learned by using a gated recurrent unit, and the temporal features of the attention fusion vector are extracted to obtain a temporal feature representation vector; the spatial feature representation vector and the temporal feature representation vector are weightedly fused to obtain a spatiotemporal fusion representation vector;

[0009] The result of spatiotemporal fusion representation vector processing using a multilayer perceptron is input into the activation function, and the power equipment fault prediction is completed based on the output of the activation function.

[0010] A power equipment fault prediction system, comprising: a multimodal data fusion module, a network construction learning module and a fault prediction module;

[0011] The multimodal data fusion module is used to obtain historical data of power equipment, perform multimodal data fusion on the historical data of power equipment to obtain feature vectors of each mode; fuse the feature vectors of each module by principal component analysis to obtain a fusion vector; construct the fusion vector into a graph by cosine similarity to obtain a feature adjacency matrix, and generate a spatiotemporal graph;

[0012] The network construction learning module is used to use the obtained spatiotemporal graph to learn spatial information by graph convolution, extract the spatial features of the data to obtain a spatial feature representation vector; use the temporal attention network to fuse the spatial feature representation vector with the attention weight to generate a time-series enhanced attention fusion vector; then use the gated recurrent unit to learn the time sequence information, extract the time features of the attention fusion vector to obtain a time feature representation vector; perform weighted feature fusion on the spatial feature representation vector and the time feature representation vector to obtain a spatiotemporal fusion representation vector;

[0013] The fault prediction module is used to use a multi-layer perceptron to process the result of the spatiotemporal fusion representation vector and input it into an activation function, and complete the power equipment fault prediction according to the output of the activation function.

[0014] The effects provided in the content of the invention are only the effects of the embodiments, not all the effects of the invention. One of the above technical solutions has the following advantages or beneficial effects:

[0015] The present invention proposes a method and system for predicting faults of electric equipment, which includes the following steps: obtaining historical data of electric equipment, fusing the historical data of electric equipment with multimodal data to obtain feature vectors of each mode; fusing the feature vectors of each module by principal component analysis to obtain a fusion vector; constructing the fusion vector into a graph by cosine similarity to obtain a feature adjacency matrix and generate a spatiotemporal graph; using the obtained spatiotemporal graph, learning spatial information by graph convolution, extracting the spatial features of the data to obtain a spatial feature representation vector; fusing the spatial feature representation vector with attention weights by using a temporal attention network to generate a time-enhanced attention fusion vector; learning the time-series information by using a gated recurrent unit, extracting the time features of the attention fusion vector to obtain a time feature representation vector; performing weighted feature fusion on the spatial feature representation vector and the time feature representation vector to obtain a spatiotemporal fusion representation vector; using a multilayer perceptron to process the spatiotemporal fusion representation vector, the result is input into an activation function, and the prediction of faults of electric equipment is completed according to the output of the activation function. Based on a method for predicting faults of electric equipment, a system for predicting faults of electric equipment is also proposed. The present invention uses graph neural networks to capture spatial dependencies and extracts spatiotemporal features in combination with changes in the time dimension. The introduction of an attention mechanism in the time dimension can dynamically focus on the impact of different moments on the prediction results and improve the accuracy of time feature extraction.

[0016] By combining the spatiotemporal graph neural network and the temporal attention mechanism, the present invention can simultaneously capture the spatial dependencies and temporal dynamic characteristics in the equipment operation data, thereby improving the accuracy and robustness of power equipment fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flow chart of a method for predicting faults of electric power equipment proposed in Embodiment 1 of the present invention;

[0018] Figure 2 A data flow processing diagram of a method for predicting power equipment faults based on a spatiotemporal graph neural network and time attention provided in Example 1 of the present disclosure;

[0019] Figure 3 This is a schematic diagram of a power equipment fault prediction system proposed in Example 2 of the present invention. DETAILED DESCRIPTION

[0020] In order to clearly illustrate the technical features of the present solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings. The disclosure below provides many different embodiments or examples for realizing different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below. In addition, the present invention may repeat reference numbers and / or letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed. It should be noted that the components illustrated in the accompanying drawings are not necessarily drawn to scale. The present invention omits the description of known components and processing techniques and processes to avoid unnecessary limitations on the present invention.

[0021] Example 1

[0022] Embodiment 1 of the present invention proposes a method for predicting power equipment faults, which is used to solve the technical problems existing in the prediction of power equipment faults in the prior art. The present invention first uses a multimodal data fusion method to fuse data from multiple sources, such as data from sensors, equipment maintenance information, equipment attribute information, and environmental data, into a comprehensive vector; then uses a cosine similarity algorithm to calculate the similarity between fused vectors, and constructs a graph based on the similarity; uses a graph structure to learn the spatial features of the data through graph convolution; uses a temporal attention network to weight the time series data to facilitate the learning of time series features; uses a gated recurrent unit (GRU) to learn the time series features of the data; uses a weighted average method to fuse spatiotemporal features to obtain a fused spatiotemporal feature representation vector; and then after a multi-layer perceptron (MLP) and a fully connected layer, performs a power equipment fault prediction task.

[0023] Figure 1 This is a flow chart of a method for predicting faults of electric power equipment proposed in Example 1 of the present invention.

[0024] In step S1, the massive amount of power equipment data is collected, specifically, the data of 363 power equipment in a certain time period of a certain power equipment is obtained, and the system collects one piece of data every hour, 10214 pieces of data for each device, a total of 4065718 pieces of data. The data contains information such as equipment number, time, voltage, current, etc.

[0025] In the field of power equipment, it is often necessary to process multiple data types, which cover information from different sources, including power parameters (such as current, voltage, power, etc.), operating status data (recording changes in the operating status of equipment, such as switch status, load changes and fault occurrence, etc.), temperature, humidity, energy consumption and environmental data.

[0026] In step S2, a large amount of power equipment data is preprocessed, including historical power equipment data, and the acquired equipment data is preprocessed, including data cleaning, missing data completion, data definition and storage.

[0027] In step S3, the historical data of the power equipment is fused by multimodal data to obtain the feature vectors of each mode. The multimodal data fusion method is used to integrate these data from different modes:

[0028] F fusion =f(X (1) ,X (2) ,…,X (M) ); (1)

[0029] Among them, F fusion represents the fused data; M represents M different data modalities, specifically represented by X (1) ,X (2) ,…,X (M0 ; Each modality’s data contains N samples, expressed as Where m=1, 2,…,M.

[0030] If there are n samples and each sample has m features, the original data can be represented as an n×m matrix X. To calculate the mean of X as μ, the data is first centered, that is, each feature is subtracted from its mean to obtain the centered data matrix The goal of PCA is to find a k-dimensional linear subspace so that the projection on this subspace can preserve the variance of the original data to the greatest extent. This linear subspace consists of a set of orthogonal bases, called principal components. Specifically, the PCA process can be divided into the following steps:

[0031] calculate The covariance matrix Σ of is:

[0032]

[0033] Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors, sort the eigenvalues ​​by size, and select the eigenvectors corresponding to the first k largest eigenvalues ​​as principal components; project the original data onto the selected principal components to obtain the reduced-dimensional data. The reduced-dimensional matrix of the covariance matrix Σ is:

[0034]

[0035] Where Z is the fusion vector; V k A matrix consisting of the first k eigenvectors.

[0036] In step S4, the fused vectors are constructed into a graph based on cosine similarity. For the fused vector set, the cosine similarity between each pair of vectors needs to be calculated. The vector set is represented as {U1,U2,…,U N}, where each vector has dimension D. Two vectors U p and U q The cosine similarity sim(U p , U q ) can be calculated by the following formula:

[0037]

[0038] Among them, U p is the pth subvector in the fusion vector Z; U q is the qth subvector in the fusion vector Z; p,q∈[1,N] and p≠q; then the similarity matrix calculated by cosine similarity is used to construct the adjacency matrix of the graph. The inventor sets the similarity threshold to 0.5, and defines it according to the similarity spatiotemporal graph as:

[0039]

[0040] Among them, A ij represents a space-time graph; D is a similarity threshold, which is set to 0.5 in the present application, and those skilled in the art can make reasonable settings based on actual conditions.

[0041] In step S5, the obtained spatiotemporal graph is used to learn spatial information by graph convolution, and the spatial features of the data are extracted to obtain a spatial feature representation vector;

[0042] Figure 2 A data flow processing diagram of a method for predicting power equipment faults based on a spatiotemporal graph neural network and time attention provided in Example 1 of the present disclosure;

[0043] The main purpose of graph convolutional neural networks is to obtain the representation of each node in the graph while considering the influence of adjacent nodes. The adjacency matrix A is used to represent the connection relationship of the graph G, and the graph convolution operation of GCN is a feature aggregation process based on the adjacency matrix.

[0044] The space-time graph A ij Middle node v i Updated to The update formula is expressed as:

[0045]

[0046] in, is node v j In the feature representation of layer l, v j Yesi Neighbor nodes of W (l) is the weight matrix, N(i) is the node v i The set of neighbor nodes, c ij is the normalization coefficient used to scale the feature representation of neighbor nodes; σ is the activation function; v i A ij The i-th node in v j A ij The jth node in ;

[0047] The activation function used in this application is Relu, for each node v i , perform weighted summation on the feature representations of its neighboring nodes, and then obtain a new feature representation through a nonlinear activation function This process can be seen as a process of obtaining information from neighboring nodes and updating their own feature representations. Here, multiple graph convolutional layers are stacked to build a deep GCN model. The feature representation of each layer is updated and propagated based on the feature representation of the previous layer, and finally a high-level graph feature representation is obtained. Next, through the global maximum pooling operation, the feature vectors of all nodes are merged into a vector of fixed dimension. Specifically, for each feature dimension, the maximum value of all nodes in that dimension is selected. In other words, for each feature dimension, the maximum value of the corresponding feature values ​​of all nodes in the graph is found, and these maximum values ​​are combined in order into a new vector, denoted as f GCN .

[0048] In step S6, the spatial feature representation vector is combined with the attention weight using the temporal attention network to generate a temporally enhanced attention fusion vector;

[0049] In order to more effectively capture the temporal information of the data, the inventors use a temporal attention network to learn the importance of data at different times.

[0050] f GCN As the input of the temporal attention network at the current time t, denoted as h t , and the node feature representation {h t-1 ,h t-2 ,…,h t-k}, where k is the time window size; h t-1 h is the input of the temporal attention network at time t-1; t-2 h is the input of the temporal attention network at time t-2; t-k Input to the temporal attention network tk at a time;

[0051] The attention weights of the feature representations of historical moments are calculated and used to weighted sum the feature representations of each historical moment, so that the attention aggregation feature representation of the current moment is obtained as follows:

[0052] α t-s =softmax(w T tanh(Uh t-s +b)); (7)

[0053] Among them, w is the first learning parameter, U is the second learning parameter and b is the third learning parameter; tanh is the hyperbolic tangent activation function, α t-s is the attention weight of historical moment ts; h t-s Input to the temporal attention network at time ts;

[0054] The calculated attention weights are used to perform weighted summation on the node feature representations at the historical moments, and the attention aggregation representation at the current moment is obtained as:

[0055] Among them, h att Aggregate attention for the present moment;

[0056] The node feature at the current moment is represented by h t With the attention aggregation representation h att After fusion, we get the time-series enhanced attention fusion vector feature, which is specifically expressed as:

[0057] h att+gcn =concat(h t ,h att ); (9)

[0058] Among them, h att+gcn Temporal enhanced attention fusion vector features.

[0059] In step S7, the gated recurrent unit is used to learn the timing information, and the time feature of the attention fusion vector is extracted to obtain the time feature representation vector;

[0060] Use recurrent neural networks to capture temporal dependencies in time series. Traditional RNNs have defects such as inability to retain long-term memory and gradient explosion. Variants of RNN provide solutions to such problems. GRU replaces the forget gate and input gate with an update gate based on LSTM, which reduces training time and improves training efficiency while maintaining considerable computational accuracy. Specifically, the GRU model can be seen as a combination of a reset gate and an update gate. At each time step t, the input vector required by GRU includes the current input vector x t and the hidden state h at the previous time step t-1 . The temporal attention fusion feature representation hatt+gcn As the input vector x at this time step t In GRU, σ and tanh represent the sigmoid function and tanh activation function respectively, representing the gate signal. Taking the gate at time t as an example, r t is the reset gate, which is used to reset the state h of the previous moment t-1 With the current information x t Combined to get the candidate hidden state c t , z t is the update gate, used to determine h t-1 How much should be discarded and how much should be discarded? t What new information needs to be incorporated to get the final hidden state h t .

[0061] The update gate is represented as: t =σ(W z ·[x t ,h t-1 ]+b z ); (10)

[0062] The reset gate is represented by: r t =σ(W r ·[x t ,h t-1 ]+b r ); (11)

[0063] Among them, z t represents the update gate, r t represents the reset gate, σ represents the activation function; W z is the weight matrix of the update gate; b z is the bias term of the update gate; W r is the weight matrix of the reset gate, b r is the bias term for the reset gate; x t is the information of the current moment; for the calculation at the initial moment, GRU will use the initialized hidden state, that is, the zero vector, to calculate the values ​​of the update gate and the reset gate.

[0064] The candidate hidden state at the current moment is expressed as:

[0065]

[0066] in, is the candidate hidden state at the current moment; ⊙ represents the element-by-element multiplication operation, W h is the weight matrix, b h is the bias vector;

[0067] Update the current hidden state:

[0068]

[0069] Among them, h t is the output of the recurrent neural network, denoted by f GRU .

[0070] In step S8, the spatial feature representation vector and the temporal feature representation vector are weightedly fused to obtain a spatiotemporal fusion representation vector;

[0071] The spatial feature vector f extracted by GCN is extracted by weighted average. GCN And the temporal feature vector f extracted by GRU GRU The fusion is:

[0072] f fusion =β·f GCN +(1-β)·f GRU ; (14)

[0073] Among them, β is the fusion weight parameter; f fusion is the space-time fusion representation vector.

[0074] In step S9, the result of processing the spatiotemporal fusion representation vector using a multilayer perceptron is input into an activation function, and the power equipment fault prediction is completed according to the output of the activation function.

[0075] Input layer to first hidden layer: f (1) =σ(W1·f fusion +b1); (15)

[0076] Among them, f (1) is the output of the first hidden layer of the multilayer perceptron; b1 is the bias vector of the first hidden layer; W1 is the weight matrix of the first hidden layer; σ represents the activation function;

[0077] From the first hidden layer to the second hidden layer: f (2) =σ(W2·f (1) +b2); (16)

[0078] Among them, f (2) is the output of the second hidden layer of the multilayer perceptron; b2 is the bias vector of the second hidden layer; W2 is the weight matrix of the second hidden layer; and so on, until the output f of the last hidden layer is obtained. (L) ;

[0079] f (L) As the input of the fully connected layer, the fully connected layer undergoes linear and nonlinear transformations to obtain the output of the fully connected layer; (L+1) =σ(W 全连接 ·f (L) +b 全连接); (17)

[0080] Among them, f (L+1) is the output of the fully connected layer; W 全连接 is the weight matrix of the fully connected layer; b 全连接 is the bias vector of the fully connected layer; where L is the number of layers of the multilayer perceptron;

[0081] f (L+1) As the input of the softmax function, the output vector is converted into a probability distribution through the softmax function to obtain the final classification result: y = softmax (h (L+1) );(18)

[0082] Among them, y is the final output vector, and the softmax function converts the output vector into the probability distribution of power equipment fault prediction.

[0083] In step S10, the power equipment fault prediction is performed on the test sample, and the results are compared with the actual situation. This embodiment uses mean square error (MSE) and mean absolute error (MAE) as evaluation indicators, regards the data per hour as a timestamp, and sets the prediction window length to 192 timestamps. Table 1 shows the comparison of power equipment fault prediction methods.

[0084] Table 1: Comparison of power equipment failure prediction methods

[0085] Model MSE MAE Linear-regression 0.188291 0.294101 MLP 0.148177 0.189586 RNN 0.079508 0.157658 LSTM 0.023895 0.166168 GRU 0.031139 0.134277 Conv-GRU 0.010938 0.098412 This embodiment 0.008384 0.079234

[0086] It can be seen from Table 1 that the performance of the power equipment fault prediction method proposed in the embodiment of the present invention is better than other methods.

[0087] In a method for predicting faults of electric power equipment proposed in Example 1 of the present invention, the method further includes adjusting the parameters in the multilayer perceptron based on the output fault prediction results using a gradient back propagation algorithm to continuously optimize the fault prediction accuracy. After the model training is completed, the model is tested using data that did not participate in the training process, the model test results are compared with the actual situation, and the model is optimized and adjusted, thereby continuously optimizing the weight data in the model and continuously improving the fault prediction accuracy.

[0088] A method for predicting faults of power equipment proposed in Example 1 of the present invention uses a graph neural network to capture spatial dependencies and extracts spatiotemporal features in combination with changes in the time dimension. The introduction of an attention mechanism in the time dimension can dynamically focus on the impact of different moments on the prediction results and improve the accuracy of time feature extraction.

[0089] A method for predicting power equipment faults proposed in Example 1 of the present invention, by combining a spatiotemporal graph neural network and a temporal attention mechanism, can simultaneously capture the spatial dependencies and temporal dynamic characteristics in equipment operation data, thereby improving the accuracy and robustness of power equipment fault prediction.

[0090] Example 2

[0091] Based on the power equipment fault prediction method proposed in embodiment 1 of the present invention, embodiment 2 of the present invention further proposes a power equipment fault prediction system, which includes: a multimodal data fusion module, a network construction learning module and a fault prediction module;

[0092] The multimodal data fusion module is used to obtain the historical data of the power equipment, perform multimodal data fusion on the historical data of the power equipment to obtain the feature vectors of each mode; fuse the feature vectors of each module by principal component analysis to obtain a fusion vector; construct the fusion vector into a graph by cosine similarity to obtain a feature adjacency matrix, and generate a space-time graph;

[0093] The network construction learning module is used to use the obtained spatiotemporal graph to learn spatial information by graph convolution, extract the spatial features of the data to obtain a spatial feature representation vector; use the temporal attention network to fuse the spatial feature representation vector with the attention weight to generate a time-series enhanced attention fusion vector; then use the gated recurrent unit to learn the temporal information, extract the time features of the attention fusion vector to obtain a time feature representation vector; perform weighted feature fusion on the spatial feature representation vector and the time feature representation vector to obtain a spatiotemporal fusion representation vector;

[0094] The fault prediction module is used to use the multi-layer perceptron to process the result of the spatiotemporal fusion representation vector and input it into the activation function, and complete the power equipment fault prediction according to the output of the activation function.

[0095] The system also includes a feedback update module, which is used to adjust the parameters in the model based on the output fault prediction results using the gradient back propagation algorithm to continuously optimize the fault prediction accuracy. After the model training is completed, the model test results are compared with the actual situation, and the model is optimized and adjusted, thereby continuously optimizing the weight data in the model and continuously improving the fault prediction accuracy.

[0096] During the execution of the multimodal data fusion module: the historical data of the power equipment is subjected to multimodal data fusion to obtain a fusion vector; the characteristic vectors of each module are fused by principal component analysis to obtain a fusion vector, and the process includes:

[0097] F fusion =f(X (1) ,X(2) ,…,X (M) ); (1)

[0098] Among them, F fusion represents the fused data; M represents M different data modalities, specifically represented by X (1) ,X (2) ,…,X (M) ; Each modality’s data contains N samples, expressed as Where m = 1, 2, ..., M;

[0099] The original power equipment historical data is represented as an n×m matrix X; the mean of X is calculated as μ, and X is centralized to obtain the centralized data matrix

[0100] calculate The covariance matrix Σ of is:

[0101]

[0102] Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. Sort the eigenvalues ​​by size and select the eigenvectors corresponding to the first k largest eigenvalues ​​as the principal components. The matrix after dimensionality reduction of the covariance matrix Σ is:

[0103]

[0104] Where Z is the fusion vector; V k A matrix consisting of the first k eigenvectors.

[0105] The fusion vector is constructed into a graph by using cosine similarity to obtain a feature adjacency matrix. The process of generating a spatiotemporal graph includes:

[0106] The fusion vector Z is represented by {U1,U2,…,U N}, where each vector has dimension D; two vectors U p and U q The cosine similarity sim(U p , U q )The calculation process is:

[0107]

[0108] Among them, U p is the pth subvector in the fusion vector Z; U q is the qth subvector in the fused vector Z; p,q∈[1,N] and p≠q;

[0109] The space-time graph is defined as:

[0110]

[0111] Among them, A ij represents the space-time graph; D is the similarity threshold.

[0112] During the execution of the network construction learning module: using the obtained spatiotemporal graph, adopting the graph convolution method to learn spatial information, extracting the spatial features of the data to obtain the spatial feature representation vector, specifically including:

[0113] The space-time graph A ij Middle node v i Updated to The update formula is expressed as:

[0114]

[0115] in, is node v j In the feature representation of layer l, v j Yes i Neighbor nodes of W (l) is the weight matrix, N(i) is the node v i The set of neighbor nodes, c ij is the normalization coefficient; σ is the activation function; v i A ij The i-th node in v j A ij The jth node in ;

[0116] Then, the feature vectors of all nodes in the spatiotemporal graph are merged into a vector of fixed dimension through the global maximum pooling operation, which is recorded as f GCN .

[0117] The process of fusing the spatial feature representation vector with the attention weight using the temporal attention network to generate a temporally enhanced attention fusion vector includes:

[0118] f GCN As the input of the temporal attention network at the current time t, denoted as h t , and the node feature representation {h t-1 ,h t-2 ,…,h t-k}, where k is the time window size; h t-1 h is the input of the temporal attention network at time t-1; t-2 h is the input of the temporal attention network at time t-2; t-k Input to the temporal attention network tk at a time;

[0119] The attention weights of the feature representations of historical moments are calculated and used to weighted sum the feature representations of each historical moment, so that the attention aggregation feature representation of the current moment is obtained as follows:

[0120] α t-s =softmax(w T tanh(Uh t-s +b)); (7)

[0121] Among them, w is the first learning parameter, U is the second learning parameter and b is the third learning parameter; tanh is the hyperbolic tangent activation function, α t-s is the attention weight of historical moment ts; h t-s Input to the temporal attention network at time ts;

[0122] The current moment of attention aggregation is expressed as:

[0123] Among them, h att Aggregate attention for the present moment;

[0124] The node feature at the current moment is represented by h t With the attention aggregation representation h att After fusion, we get the time-series enhanced attention fusion vector feature, which is specifically expressed as:

[0125] h att+gcn =concat(h t ,h att ); (9)

[0126] Among them, h att+gcn Temporal enhanced attention fusion vector features.

[0127] Then, the gated recurrent unit is used to learn the timing information and extract the temporal features of the attention fusion vector to obtain the temporal feature representation vector.

[0128] The update gate is represented as: t =σ(W z ·[x t ,h t-1 ]+b z ); (10)

[0129] The reset gate is represented by: r t =σ(W r ·[x t ,h t-1 ]+b r ); (11)

[0130] Among them, z t represents the update gate, r trepresents the reset gate, σ represents the activation function; W z is the weight matrix of the update gate; b z is the bias term of the update gate; W r is the weight matrix of the reset gate, b r is the bias term for the reset gate; x t Information for the current moment;

[0131] The candidate hidden state at the current moment is expressed as:

[0132]

[0133] in, is the candidate hidden state at the current moment; ⊙ represents the element-by-element multiplication operation, W h is the weight matrix, b h is the bias vector;

[0134] Update the current hidden state:

[0135]

[0136] Among them, h t is the output of the recurrent neural network, denoted by f GRU .

[0137] The specific process of performing weighted feature fusion on the spatial feature representation vector and the temporal feature representation vector to obtain the spatiotemporal fusion representation vector includes: using a weighted average method to perform weighted averaging on the spatial feature vector f extracted by GCN GCN And the temporal feature vector f extracted by GRU GRU The fusion is:

[0138] f fusion =β·f GCN +(1-β)·f GRU ; (14)

[0139] Among them, β is the fusion weight parameter; f fusion is the space-time fusion representation vector.

[0140] During the execution of the fault prediction module, the result of processing the spatiotemporal fusion representation vector using a multilayer perceptron is input into the activation function. The process of completing the power equipment fault prediction according to the output of the activation function includes:

[0141] Input layer to first hidden layer:

[0142] f (1) =σ(W1·f fusion +b1); (15)

[0143] Among them, f (1)is the output of the first hidden layer of the multilayer perceptron; b1 is the bias vector of the first hidden layer; W1 is the weight matrix of the first hidden layer; σ represents the activation function;

[0144] From the first hidden layer to the second hidden layer:

[0145] f (2) =σ(W2·f (1) +b2); (16)

[0146] Among them, f (2) is the output of the second hidden layer of the multilayer perceptron; b2 is the bias vector of the second hidden layer; W2 is the weight matrix of the second hidden layer; and so on, until the output f of the last hidden layer is obtained. (L) ;

[0147] f (L) As the input of the fully connected layer, the fully connected layer undergoes linear and nonlinear transformations to obtain the output of the fully connected layer; (L+1) =σ(W 全连接 ·f (L) +b 全连接 ); (17)

[0148] Among them, f (L+1) is the output of the fully connected layer; W 全连接 is the weight matrix of the fully connected layer; b 全连接 is the bias vector of the fully connected layer; where L is the number of layers of the multilayer perceptron;

[0149] f (L+1) As the input of the softmax function, the output vector is converted into a probability distribution through the softmax function to obtain the final classification result:

[0150] y=softmax(h (L+1) ); (18)

[0151] Among them, y is the final output vector, and the softmax function converts the output vector into the probability distribution of power equipment fault prediction.

[0152] A power equipment fault prediction system proposed in Example 2 of the present invention uses a graph neural network to capture spatial dependencies and extracts spatiotemporal features in combination with changes in the time dimension. The introduction of an attention mechanism in the time dimension can dynamically focus on the impact of different moments on the prediction results and improve the accuracy of time feature extraction.

[0153] An electric power equipment fault prediction system proposed in Example 2 of the present invention, by combining a spatiotemporal graph neural network and a temporal attention mechanism, can simultaneously capture the spatial dependencies and temporal dynamic characteristics in equipment operation data, thereby improving the accuracy and robustness of electric power equipment fault prediction.

[0154] The description of the relevant parts of an electric power equipment fault prediction system provided in Example 2 of the present application can be found in the detailed description of the corresponding parts of an electric power equipment fault prediction method provided in Example 1 of the present application, and will not be repeated here.

[0155] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the term "include", "comprise" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment that includes a series of elements are inherent to the elements. In the absence of more restrictions, the elements limited by the sentence "comprise one..." do not exclude the presence of other identical elements in the process, method, article or equipment that includes the elements. In addition, the above-mentioned technical solution provided in the embodiment of the present application is consistent with the corresponding technical solution in the prior art in principle, and the part is not described in detail, so as not to repeat too much.

[0156] Although the above describes the specific implementation of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. For those skilled in the art, other different forms of modifications or deformations can be made on the basis of the above description. It is not necessary and impossible to list all the implementation methods here. On the basis of the technical solution of the present invention, various modifications or deformations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. A method for predicting faults of electric power equipment, characterized in that: The following steps are involved: Acquire historical data of power equipment, perform multimodal data fusion on the historical data of power equipment to obtain feature vectors of each mode; fuse the feature vectors of each module by principal component analysis to obtain a fusion vector; construct the fusion vector into a graph by cosine similarity to obtain a feature adjacency matrix, and generate a spatiotemporal graph; Using the obtained spatiotemporal graph, we adopt graph convolution to learn spatial information and extract the spatial features of the data to obtain the spatial feature representation vector. The spatial feature representation vector is combined with the attention weight by using the temporal attention network to generate a time-series enhanced attention fusion vector; the temporal information is then learned by using the gated recurrent unit to extract the time feature of the attention fusion vector to obtain the temporal feature representation vector; the spatial feature representation vector and the temporal feature representation vector are weightedly fused to obtain a spatiotemporal fusion representation vector; The result of spatiotemporal fusion representation vector processing using a multilayer perceptron is input into the activation function, and the power equipment fault prediction is completed based on the output of the activation function.

2. A method for predicting faults of electric power equipment according to claim 1, characterized in that: The method also includes: using a gradient back propagation algorithm to adjust parameters in the multi-layer perceptron to continuously optimize the accuracy of fault prediction.

3. A method for predicting faults of electric power equipment according to claim 1, characterized in that: The process of fusing the historical data of the power equipment with multi-modal data to obtain a fusion vector; and fusing the characteristic vectors of each module by principal component analysis to obtain a fusion vector includes: F fusion =f(X (1) ,X (2) ,…,X (M) );(1) Among them, F fusion represents the fused data; M represents M different data modalities, specifically represented by X (1) ,X (2) ,…,X (M) ; Each modality’s data contains N samples, expressed as Where m = 1, 2, ..., M; The original power equipment historical data is represented as an n×m matrix X; the mean of X is calculated as μ, and X is centralized to obtain the centralized data matrix calculate The covariance matrix Σ is: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. Sort the eigenvalues ​​by size and select the eigenvectors corresponding to the first k largest eigenvalues ​​as the principal components. The matrix after dimensionality reduction of the covariance matrix Σ is: Where Z is the fusion vector; V k A matrix consisting of the first k eigenvectors.

4. A method for predicting faults of electric power equipment according to claim 3, characterized in that: The fusion vector is constructed into a graph by using cosine similarity to obtain a feature adjacency matrix. The process of generating a spatiotemporal graph includes: The fusion vector Z is represented by {U1,U2,…,U N }, where each vector has dimension D; two vectors U p and U q The cosine similarity sim(U p , U q )The calculation process is: Among them, U p is the pth subvector in the fusion vector Z; U q is the qth subvector in the fused vector Z; p,q∈[1,N] and p≠q; The space-time graph is defined as: Among them, A ij represents the space-time graph; D is the similarity threshold.

5. A method for predicting faults of electric power equipment according to claim 4, characterized in that: The method of using the obtained spatiotemporal graph to learn spatial information by graph convolution and extracting the spatial features of the data to obtain the spatial feature representation vector specifically includes: The space-time graph A ij Middle node v i Updated to The update formula is expressed as: in, is node v j In the feature representation of layer l, v j Yes i Neighbor nodes of (l) is the weight matrix, N(i) is the node v i The set of neighbor nodes, c ij is the normalization coefficient; σ is the activation function; v i A ij The i-th node in v j A ij The jth node in ; Then, the feature vectors of all nodes in the spatiotemporal graph are merged into a vector of fixed dimension through the global maximum pooling operation, which is recorded as f GCN .

6. A method for predicting faults of electric power equipment according to claim 5, characterized in that: The process of fusing the spatial feature representation vector with the attention weight using the temporal attention network to generate a temporally enhanced attention fusion vector includes: f GCN As the input of the temporal attention network at the current time t, denoted as h t , and the node feature representation {h t-1 ,h t-2 ,…,h t-k }, where k is the time window size; h t-1 h is the input of the temporal attention network at time t-1; t-2 h is the input of the temporal attention network at time t-2; t-k Input to the temporal attention network tk at a time; The attention weights of the feature representations of historical moments are calculated and used to weighted sum the feature representations of each historical moment, so that the attention aggregation feature representation of the current moment is obtained as follows: α t-s =softmax(w T tanh(Uh t-s +b));(7) Among them, w is the first learning parameter, U is the second learning parameter and b is the third learning parameter; tanh is the hyperbolic tangent activation function, α t-s is the attention weight of historical moment ts; h t-s Input to the temporal attention network at time ts; The current moment of attention aggregation is expressed as: Among them, h att Aggregate attention for the present moment; The node feature at the current moment is represented by h t With the attention aggregation representation h att After fusion, we get the time-series enhanced attention fusion vector feature, which is specifically expressed as: h att+gcn =concat(h t ,h att );(9) Among them, h att+gcn Temporal enhanced attention fusion vector features.

7. A method for predicting faults of electric power equipment according to claim 6, characterized in that: Then, the gated recurrent unit is used to learn the timing information and extract the temporal features of the attention fusion vector to obtain the temporal feature representation vector. The update gate is represented as: t =σ(W z ·[x t ,h t-1 ]+b z );(10) The reset gate is represented by: r t =σ(W r ·[x t ,h t-1 ]+b r );(11) Among them, z t represents the update gate, r t represents the reset gate, σ represents the activation function; W z is the weight matrix of the update gate; b z is the bias term of the update gate; W r is the weight matrix of the reset gate, b r is the bias term for the reset gate; x t Information for the current moment; The candidate hidden state at the current moment is expressed as: in, is the candidate hidden state at the current moment; ⊙ represents the element-by-element multiplication operation, W h is the weight matrix, b h is the bias vector; Update the current hidden state: Among them, h t is the output of the recurrent neural network, denoted by f GRU .

8. A method for predicting faults of electric power equipment according to claim 7, characterized in that: The specific process of performing weighted feature fusion on the spatial feature representation vector and the temporal feature representation vector to obtain the spatiotemporal fusion representation vector includes: using a weighted average method to perform weighted averaging on the spatial feature vector f extracted by GCN GCN And the temporal feature vector f extracted by GRU GRU The fusion is: f fusion =β·f GCN +(1-β)·f GRU ;(14) Among them, β is the fusion weight parameter; f fusion is the space-time fusion representation vector.

9. A method for predicting faults of electric power equipment according to claim 8, characterized in that: The result of processing the spatiotemporal fusion representation vector using a multilayer perceptron is input into the activation function. The process of completing the power equipment fault prediction according to the output of the activation function includes: Input layer to first hidden layer: f (1) =σ(W1 f fusion +b1);(15) Among them, f (1) is the output of the first hidden layer of the multilayer perceptron; b1 is the bias vector of the first hidden layer; W1 is the weight matrix of the first hidden layer; σ represents the activation function; From the first hidden layer to the second hidden layer: f (2) =σ(W2·f (1) +b2);(16) Among them, f (2) is the output of the second hidden layer of the multilayer perceptron; b2 is the bias vector of the second hidden layer; W2 is the weight matrix of the second hidden layer; and so on, until the output f of the last hidden layer is obtained. (L) ; f (L) As the input of the fully connected layer, the fully connected layer undergoes linear and nonlinear transformations to obtain the output of the fully connected layer; (L+1) =σ(W 全连接 ·f (L) +b 全连接 ); (17) Among them, f (L+1) is the output of the fully connected layer; W 全连接 is the weight matrix of the fully connected layer; b 全连接 is the bias vector of the fully connected layer; where L is the number of layers of the multilayer perceptron; f (L+1) As the input of the softmax function, the output vector is converted into a probability distribution through the softmax function to obtain the final classification result: and softmax(h) (L+1) );(18) Among them, y is the final output vector, and the softmax function converts the output vector into the probability distribution of power equipment fault prediction.

10. A power equipment fault prediction system, characterized in that: include: Multimodal data fusion module, network construction learning module and fault prediction module; The multimodal data fusion module is used to obtain historical data of power equipment, perform multimodal data fusion on the historical data of power equipment to obtain feature vectors of each mode; fuse the feature vectors of each module by principal component analysis to obtain a fusion vector; construct the fusion vector into a graph by cosine similarity to obtain a feature adjacency matrix, and generate a spatiotemporal graph; The network construction learning module is used to use the obtained spatiotemporal graph to learn spatial information by graph convolution, extract the spatial features of the data and obtain a spatial feature representation vector; The spatial feature representation vector is combined with the attention weight by using the temporal attention network to generate a time-series enhanced attention fusion vector; the temporal information is then learned by using the gated recurrent unit to extract the time feature of the attention fusion vector to obtain the temporal feature representation vector; the spatial feature representation vector and the temporal feature representation vector are weightedly fused to obtain a spatiotemporal fusion representation vector; The fault prediction module is used to use a multi-layer perceptron to process the result of the spatiotemporal fusion representation vector and input it into an activation function, and complete the power equipment fault prediction according to the output of the activation function.

Images