UPS (Uninterrupted Power Supply) data complementation method for fusing bidirectional space-time diagram convolution with MLP (Multiplayer Library

Abstract

The invention relates to a UPS data completion method based on bidirectional space-time diagram convolution fusion MLP anomaly perception. The method comprises the steps of obtaining UPS data and constructing a time sequence data set through a sliding window; marking a missing position; constructing a UPS data completion network model, wherein the UPS data completion network model comprises an MLP anomaly perception adjacency matrix, and two time-space-two-way graph convolution modules, a self-adaptive pooling layer and a missing value filling module which are arranged in sequence; the MLP anomaly perception adjacency matrix is used for generating a dynamic adjacency matrix and inputting the dynamic adjacency matrix into the two space-time-bidirectional graph convolution modules; the space-time-bidirectional graph convolution module comprises two bidirectional graph convolution networks and a time sequence causal convolution network so as to extract space features and time features in the UPS data; carrying out data complementation on missing positions after outputting feature reduction data dimensions; training the UPS data completion network model through the training data set; and carrying out UPS data completion by using the trained UPS data completion network model. According to the method, the real state data of UPS operation can be accurately restored.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data processing, and particularly relates to a UPS data completion method based on bidirectional space-time graph convolution fusion MLP abnormal perception. BACKGROUND

[0002] UPS equipment is the guarantee power supply for key infrastructures (such as hospitals, data centers, factories), and once a failure occurs, it may lead to serious consequences. UPS online health management and health management system are crucial for improving the reliability of the whole machine and maintaining the safety of the data room. Due to equipment failure, communication obstacles and human factors, the collected space-time data inevitably has missing values, which seriously destroys the space-time correlation of the data and reduces the application value of the data. However, the UPS health management system usually relies on complete time data, and the data missing may mask the key abnormal signals, cause the feature distribution to deviate, and directly affect the accuracy of the fault prediction model, especially the prediction model based on machine learning / deep learning, thereby affecting the accurate assessment of the system health state and increasing the risk of misjudgment or missed judgment.

[0003] Therefore, it is necessary to study the UPS data completion method. A self-supervised spatio-temporal data completion method and system based on time sequence semantic attention are disclosed in Chinese Patent No. CN202510841249.3. The spatio-temporal data is obtained, and the incomplete time sequence is divided into semantic segment units through time sequence segmentation modeling; the cross attention mechanism is used to mine the inter-segment potential time sequence dependency pattern to obtain the segmented attention calculation result; the time sequence comparison model is constructed based on the segmented attention calculation result, and the strong generalization embedding of the time sequence modeling is obtained through the self-supervised method; the embedding obtained through the time sequence modeling is used as the time sequence feature of each node, and the spatial domain modeling is performed through the graph convolution network to obtain the spatial domain modeling result. This method breaks through the strong dependence of traditional models on data integrity, and can effectively capture nonlinear spatio-temporal correlations in data missing scenarios, providing a high-robustness time sequence embedding basis for subsequent spatio-temporal node-level tasks. However, this method does not study outliers such as fault data and abnormal data, which can easily lead to misjudgment and omission of abnormal data and fault data, and can cause error completion of fault data under normal working conditions. The error completion data seriously affects the value and mineability of the data. This method performs a large amount of calculation in the time dimension, fully mines the time dimension signal features, and has a large amount of calculation. In the relatively simple time sequence signal scene, this algorithm is prone to overfitting in the time dimension, and cannot provide accurate completion effect. The graph convolution network is used in the spatial feature mining and analysis of this method, and the graph convolution network will greatly increase the calculation amount with the increase of the graph nodes. In the scene with complex spatial features, the calculation efficiency is low and the completion is slow. This method focuses on urban traffic monitoring data, and proposes an algorithm mainly based on time dimension features and supplemented by spatial dimension features. In the UPS scene, the spatial dimension feature information is complex, the time sequence dimension feature is simple, and the spatial feature usually contains important abnormal signals, so the signal mining of the spatial dimension should be paid attention to. SUMMARY

[0004] The UPS data completion method provided by the application is beneficial to accurately restoring the real state data of the UPS operation.

[0005] In order to achieve the above-mentioned purpose, the technical scheme adopted by the application is as follows: a UPS data completion method based on bidirectional spatio-temporal graph convolution fusion MLP abnormal perception, comprising the following steps:

[0006] (1) acquiring UPS data for model training, constructing a time sequence data set through a sliding window; marking the missing positions in the time sequence data set to form a training data set;

[0007] (2) constructing a UPS data completion network model, the UPS data completion network model comprising an MLP abnormal perception adjacency matrix and two spatio-temporal bidirectional graph convolution modules, an adaptive pooling layer and a missing value filling module arranged in sequence;

[0008] The input data is fed into the MLP anomaly-aware adjacency matrix and the first spatiotemporal-bidirectional graph convolutional module, respectively. The output of the first spatiotemporal-bidirectional graph convolutional module serves as the input to the second spatiotemporal-bidirectional graph convolutional module. The MLP anomaly-aware adjacency matrix processes the input data to generate a dynamic adjacency matrix. This dynamic adjacency matrix adaptively adjusts edge weights to capture anomaly signals by fusing the node dependencies learned by the MLP with the feature propagation matrix weighted by anomaly scores. The generated dynamic adjacency matrix is ​​simultaneously input into both spatiotemporal-bidirectional graph convolutional modules.

[0009] The spatiotemporal-bidirectional graph convolution module includes a first bidirectional graph convolutional network, a temporal causal convolutional network, and a second bidirectional graph convolutional network arranged sequentially. The dynamic adjacency matrix of the spatiotemporal-bidirectional graph convolution module is simultaneously input to both bidirectional graph convolutional networks. The bidirectional graph convolutional network, based on the dynamic adjacency matrix, mines the lateral spatial features of the input data through forward graph convolution and backward graph convolution combined with Chebyshev multinomials, thereby reducing computational complexity. The temporal causal convolutional network models the longitudinal temporal features of the input data, ensuring that the output depends only on the current and historical input data, avoiding interference from future information. The spatiotemporal-bidirectional graph convolution module integrates the spatial and temporal features extracted by the first bidirectional graph convolutional network, the temporal causal convolutional network, and the second bidirectional graph convolutional network as the output.

[0010] The output of the second spatiotemporal-bidirectional graph convolution module is input into an adaptive pooling layer to restore the data dimensions; then, the missing value imputation module fills in the missing data at the marked locations.

[0011] (3) The constructed UPS data completion network model is trained using the training dataset. During the training process, the loss function value is calculated and the model parameters are iteratively updated to minimize the completion error.

[0012] (4) Use the trained UPS data completion network model to complete UPS data.

[0013] Furthermore, in step (1), for the acquired UPS data, an hourly time window is constructed to cover short- and medium-term trends. Then, the UPS data is sampled through the time window to construct a time series dataset D = (x t-w ,x t-w+1 ,...,x t );

[0014] Missing locations in the time-series dataset are marked using a mask matrix, where 0 represents normal data and 1 represents missing data; thus forming the training dataset (x t-w ,x t-w+1 ,...,xt (mask).

[0015] Furthermore, in step (2), the method for constructing the dynamic adjacency matrix is ​​as follows:

[0016] The node feature pairs are modeled using an MLP model, and the edge weights f() are dynamically learned to generate an adjacency matrix.

[0017] e ij =MLP([X i ||X j ])

[0018] A attn [i,j]=softmax(e ij )

[0019] Among them, [X i ||X j ] indicates passing through node X i Node X j The node feature pairs formed by splicing; MLP represents a multilayer perceptron; e ij Represent the edge weight from node i to j; represent all edge weights e ij The organization forms an adjacency matrix A atten [i,j];

[0020] Calculate anomaly scores for feature dimensions and construct a feature propagation matrix based on these anomaly scores:

[0021]

[0022] Among them, score f The anomaly score represents the feature dimension, x. f For all samples of the f-th feature, Var(x) f ) represents the variance of the f-th feature over all time periods, mask b,t,f This indicates whether the node has missing data;

[0023]

[0024] in, A represents the normalized outlier score. anom [i,j] represents the feature propagation matrix based on anomaly scores;

[0025] By fusing the adjacency matrix and the feature propagation matrix based on anomaly scores, a dynamic adjacency matrix is ​​generated.

[0026] A[i,j]=α·A attn [i,j]+(1-α)·A anom [i,j]

[0027] Where α is the fusion weight.

[0028] Further, in step (2), the expressions for the forward propagation convolution (TD-GCN) and backward propagation convolution (BU-GCN) of the bidirectional graph convolutional network are:

[0029]

[0030] in, To add a self-loop to the positive adjacency matrix, Let A be the reverse adjacency matrix, A be the dynamic adjacency matrix, and I be the identity matrix.

[0031]

[0032] in, These represent the activation matrices of the l-th layer during forward and backward propagation, respectively. Let represent the trainable weights of the l-th layer in forward and backward propagation, respectively, and σ(·) represent the activation function.

[0033] Furthermore, in step (2), the convolution kernel is replaced by a Chebyshev polynomial, the expression of which is:

[0034]

[0035] in, These represent the forward convolution weights and the backward convolution weights, respectively. The normalized Laplace matrix, for The transpose of T k () represents the Chebyshev polynomial, and x is the input signal.

[0036] Furthermore, in step (2), the temporal causal convolutional network models the temporal dependency of the input data through causal convolution, and uses only historical and current time data to fill the input data forward, forcing the convolution to depend only on the current and past time.

[0037] y t =f(x) t ,x t-1 ,…,x1)

[0038] Among them, y t Let f() represent the output features of a temporal causal convolutional network, and x represent the convolution operation. t This represents the data at time step t.

[0039] Further, in step (2), the spatiotemporal-bidirectional graph convolution module performs secondary spatial mining of spatiotemporal features through the hierarchical structure of the first bidirectional graph convolutional network, the temporal causal convolutional network, and the second bidirectional graph convolutional network: First, the spatial features of the original input data are mined through the first bidirectional graph convolutional network; then, the temporal dependencies are learned through the temporal causal convolutional network; finally, the spatial features after fusing temporal features are deeply integrated again through the second bidirectional graph convolutional network.

[0040] Furthermore, in step (3), the loss function uses the mean squared error:

[0041]

[0042] Among them, X MsE The mean squared error is represented by y, where n is the sample size and y is the mean squared error. i Let i be the true value of the i-th sample. Let be the predicted value for the i-th sample.

[0043] The present invention also provides a computer device, including a memory and a processor, wherein the processor implements the above-described method when executing a computer program stored in the memory.

[0044] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0045] Compared with existing technologies, this invention has the following advantages: To address the complex spatiotemporal dependencies and nonlinearities in UPS data completion, and to tackle the problems of scarce fault data and severe data imbalance, this invention proposes a bidirectional spatiotemporal graph convolutional fusion MLP anomaly perception UPS data completion method. This method dynamically adjusts the adjacency matrix through MLP anomaly perception. Simultaneously, based on the statistical characteristics of abnormal data, it constructs an anomaly perception propagation matrix to guide the graph structure to focus on the connections between "potential anomaly" feature dimensions. Furthermore, it addresses the complex lateral spatial dimension features of UPS systems, while the temporal dimension features are relatively simple and exhibit small fluctuations. Based on the characteristics of this approach, a spatiotemporal-bidirectional graph convolution module was constructed, which prioritizes spatial structure and supplements it with temporal features. This module deeply mines lateral spatial features through a bidirectional graph convolutional network, ensuring that the graph structure and anomaly correlations within the spatial features are fully extracted. It also mines temporal features through temporal causal convolution, ensuring that the output features focus only on historical and current data, avoiding interference from future data. To reduce computational load, Chebyshev multinomials are used to replace the convolution kernels of the bidirectional spatiotemporal graph convolution, forming a bidirectional Chebyshev convolution. This effectively avoids the massive computational burden of bidirectional spatiotemporal graph convolution and improves the algorithm's completion efficiency. This method can recreate the true state trajectory of system operation as closely as possible, providing a more continuous and complete input sequence for online monitoring and health management systems. This assists in early warning and risk intervention, improving the effectiveness of fault location and remaining life prediction. Attached Figure Description

[0046] Figure 1 This is a schematic diagram illustrating the implementation principle of the UPS data completion method for bidirectional spatiotemporal graph convolutional fusion MLP anomaly perception provided in this embodiment of the invention. Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0048] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0049] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0050] like Figure 1As shown in the figure, this embodiment provides a UPS data completion method for bidirectional spatiotemporal graph convolutional fusion MLP anomaly perception, and its specific implementation steps are as follows.

[0051] (1) Obtain UPS data for model training and construct a time series dataset by sliding window; mark the missing positions in the time series dataset to form a training dataset.

[0052] (2) Construct a UPS data completion network model, which includes an MLP anomaly perception adjacency matrix and two spatiotemporal bidirectional graph convolutional modules, an adaptive pooling layer and a missing value imputation module set in sequence.

[0053] The input data is fed into the MLP anomaly-aware adjacency matrix and the first spatiotemporal-bidirectional graph convolutional module, respectively. The output of the first spatiotemporal-bidirectional graph convolutional module is used as the input of the second spatiotemporal-bidirectional graph convolutional module. The MLP anomaly-aware adjacency matrix processes the input data to generate a dynamic adjacency matrix. The dynamic adjacency matrix adaptively adjusts the edge weights to capture anomaly signals by fusing the node dependencies learned by MLP with the feature propagation matrix weighted by anomaly scores. The generated dynamic adjacency matrix is ​​simultaneously fed into both spatiotemporal-bidirectional graph convolutional modules.

[0054] The spatiotemporal-bidirectional graph convolution module includes a first bidirectional graph convolutional network, a temporal causal convolutional network, and a second bidirectional graph convolutional network arranged sequentially. The dynamic adjacency matrix of the spatiotemporal-bidirectional graph convolution module is simultaneously input to both bidirectional graph convolutional networks. The bidirectional graph convolutional network, based on the dynamic adjacency matrix, mines the lateral spatial features of the input data through forward graph convolution (TD-GCN) and backward graph convolution (BU-GCN), and uses Chebyshev multinomials to replace traditional convolution kernels to reduce computational complexity. The temporal causal convolutional network models the longitudinal temporal features of the input data, ensuring that the output depends only on the current and historical input data, avoiding interference from future information. The spatiotemporal-bidirectional graph convolution module integrates the spatial and temporal features extracted by the first bidirectional graph convolutional network, the temporal causal convolutional network, and the second bidirectional graph convolutional network as the output.

[0055] The UPS data completion network model processes the input dataset through two spatiotemporal-bidirectional graph convolutional modules. These modules simultaneously receive the dynamic adjacency matrix A calculated from the MLP anomaly-aware adjacency matrix. The spatiotemporal-bidirectional graph convolutional modules perform secondary spatial mining of spatiotemporal features through a hierarchical structure of a first bidirectional graph convolutional network, a temporal causal convolutional network, and a second bidirectional graph convolutional network. Specifically, this includes: first, mining the spatial features of the original input data through a bidirectional graph convolutional network; then, mining temporal features and learning temporal dependencies through a temporal causal convolutional layer; and finally, performing lateral feature mining again through a bidirectional graph convolutional network, propagating and integrating features at each time step to achieve deep integration of spatial features after fusing temporal features. This module helps alleviate the problem of spatial structural information loss caused by temporal noise or missing data.

[0056] After fully exploring the spatiotemporal features, the output of the second spatiotemporal-bidirectional graph convolution module is input into an adaptive pooling layer to restore the data dimensionality. Then, a missing value imputation module is used to complete the data at the missing labeled locations.

[0057] (3) The constructed UPS data completion network model is trained using the training dataset. During the training process, the loss function value is calculated and the model parameters are iteratively updated to minimize the completion error.

[0058] (4) Use the trained UPS data completion network model to complete UPS data.

[0059] 1. Dataset Construction

[0060] For the acquired UPS data, hourly time windows are constructed to cover short- to medium-term trends. Then, UPS data is sampled using these time windows to construct a time-series dataset D = (x... t-w ,x t-w+1 ,...,x t ).

[0061] Missing locations in the time-series dataset are marked using a mask matrix, where 0 represents normal data and 1 represents missing data; thus forming the training dataset (x t-w ,x t-w+1 ,...,x t (mask).

[0062] The core idea of ​​the sliding window is to transform time series data into a supervised learning format. For time series data, in this embodiment, the total time step W is 6, and the 6-step window, or 120 minutes of data, can cover short- to medium-term trends (such as hourly periods), avoiding noise introduced by excessively long windows or pattern loss by excessively short windows. Missing positions in the time series dataset are marked using a mask matrix, where 0 represents normal data and 1 represents missing data. The time series dataset (x) constructed using the mask matrix and the sliding window is then... t-w ,x t-w+1 ,...,x t Together they form the training dataset (x) t-w ,x t-w+1 ,...,x t (mask).

[0063] 2. MLP Anomaly-Aware Adjacency Matrix

[0064] An adjacency matrix is ​​a common way to represent graph structures in graph theory. In machine learning and deep learning, it is widely used to model the relationships (connections / dependencies) between nodes. Traditional graph convolution relies on a predefined adjacency matrix A, but in practice, graph structures are often unknown or not fixed. Using static or other dynamic adjacency matrices can easily overlook the relationships between graph structures.

[0065] This invention uses an MLP (MLP-based Adjacency Matrix Learning) model to model node feature pairs, dynamically learn edge weights f(), and generate an adjacency matrix to improve the accuracy of graph convolutional networks.

[0066] Let the feature of each node be X. i ∈R B*T Edge weights are generated using a shared MLP model, as shown in the following formula:

[0067] e ij =MLP([X i ||X j ])

[0068]

[0069] Among them, [X i ||X j ] indicates passing through node X i Node X j The node feature pairs formed by splicing; MLP stands for Multilayer Perceptron, which has nonlinear transformation capabilities; e ij Represent the edge weight from node i to j; represent all edge weights e ij The organization forms an adjacency matrix A atten[i,j]∈R N×N .

[0070] This invention constructs an anomaly-aware matrix, which is a feature propagation matrix based on anomaly scores. By combining the statistical characteristics of the data, it guides the graph structure to focus on the relationship between the feature dimensions of "potential anomalies".

[0071] First, anomaly scores are calculated for the feature dimensions, and then a feature propagation matrix based on the anomaly scores is constructed.

[0072]

[0073] Among them, score f The anomaly score represents the feature dimension, x. f For all samples of the f-th feature, Var(x) f ) represents the variance of the f-th feature over all time periods, mask b,t,f This indicates whether the node has missing data (1 indicates it exists). A higher anomaly score means that the feature is more likely to be anomaly or missing.

[0074]

[0075] in, A represents the normalized outlier score. anom [i,j] represents the feature propagation matrix based on anomaly scores, where features with higher scores have stronger connections.

[0076] By fusing the adjacency matrix and the feature propagation matrix based on anomaly scores, a dynamic adjacency matrix is ​​generated.

[0077] A[i,j]=α·A attn [i,j]+(1-α)·A anom [i,j]

[0078] Where α is the fusion weight.

[0079] 3. Bidirectional Graph Convolutional Networks

[0080] A graph consists of a set of nodes and edges connecting the nodes. There can be multiple connections between nodes, representing complex relationships. Define an undirected graph G = (V, E), where V = {v1, v2, ..., v...} n} represents the node set, and n represents the number of nodes; E = {e1, e2, ..., e m Let} represent the edge set, and m represent the number of edges. The nodes in the node set represent actual individuals, and the edges in the edge set represent the relationships between individuals.

[0081] The graph Laplacian matrix is ​​used to describe the diffusion process of nodes in a graph. The Laplacian matrix is ​​defined as follows:

[0082] L=DA

[0083] Introducing the symmetric normalized Laplace matrix L sym :

[0084]

[0085] The Bi-GCN (Bidirectional Graph Convolutional Network) consists of two independent GCN branches: forward propagation convolution (TD-GCN) and backward propagation convolution (BU-GCN). TD-GCN represents propagation along the original edge direction (parent node → child node), modeling a deep causal chain. BU-GCN represents propagation along the reverse edge direction (child node → parent node), capturing a breadth-diffusion structure.

[0086]

[0087] in, To add a self-loop to the positive adjacency matrix, Let A be the reverse adjacency matrix, A be the dynamic adjacency matrix, and I be the identity matrix.

[0088]

[0089] in, These represent the activation matrices of the l-th layer during forward and backward propagation, respectively. Let represent the trainable weights of the l-th layer in forward and backward propagation, respectively, and σ(·) represent the activation function.

[0090] 4. Chebyshev polynomials

[0091] The universal decomposition of the graph Laplacian matrix L is:

[0092] L=UΛU T

[0093] Where U represents the eigenvector matrix (the Fourier basis of the graph), and Λ represents the eigenvalue diagonal matrix (the frequency components of the graph). The spectral domain convolution of the graph signal x is defined as:

[0094] g θ ·x=Ug θ (Λ)U T x

[0095] Among them, g θ (Λ) is a frequency domain filter. To reduce computation, the graph convolution approximates the spectral graph using Chebyshev polynomials, and uses a K-order polynomial approximation. The convolution operation is expressed as:

[0096]

[0097] in, This is the scaled graph Laplace matrix; It is the k-th order Chebyshev polynomial, and its recurrence relation is:

[0098]

[0099]

[0100] Where, θ k The weight parameters corresponding to the k-th order polynomial (to be learned) enable this convolution to efficiently propagate the signal to its k-th order neighbors in the graph. The principle of the bidirectional graph convolutional network is to introduce the Laplacian matrix and its transpose onto the original ChebNet to construct forward and backward graph convolutions, corresponding to forward and backward dependency propagation, respectively. This allows the Chebyshev polynomial to replace the traditional convolution kernel, with the expression:

[0101]

[0102] in, These represent the forward convolution weights and the backward convolution weights, respectively. The normalized Laplace matrix, for The transpose of T k () represents the Chebyshev polynomial, and x is the input signal.

[0103] 5. Temporal Causal Convolutional Networks

[0104] In UPS time-series data, because the operation of the data center requires high grid stability, there are usually no significant signal fluctuations, and the time-series characteristics are relatively simple. Causal convolution is a convolutional structure specifically designed for time-series modeling tasks. Its core idea is to ensure that the output depends only on the current and historical inputs, prohibiting the use of future information, so as to strictly maintain the causal relationship of the time series.

[0105] During the online fault diagnosis data filling phase of the UPS, only historical data is received, and future data is not involved. Therefore, the temporal causal convolutional network models the temporal dependency of the input data through causal convolution, and only uses historical and current data to pad the input data, forcing the convolution to depend only on the current and past moments.

[0106] y t =f(x) t ,x t-1 ,…,x1)

[0107] Among them, y t Let f() represent the output features of a temporal causal convolutional network, and x represent the convolution operation. t This represents the data at time step t.

[0108] 6. Loss Function

[0109] In this embodiment, the loss function used for model training is the mean squared error (MSE):

[0110]

[0111] Among them, X MSE The mean squared error is represented by y, where n is the sample size and y is the mean squared error. i Let i be the true value of the i-th sample. Let be the predicted value for the i-th sample.

[0112] This embodiment also provides a computer device, including a memory and a processor, wherein the processor executes a computer program stored in the memory to implement the above-described method.

[0113] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0114] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0115] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0116] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0117] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0118] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A UPS data completion method based on bidirectional spatiotemporal graph convolutional fusion MLP anomaly detection, characterized in that, Includes the following steps: (1) Obtain UPS data for model training and construct a time series dataset using a sliding window; Label the missing locations in the time-series dataset to form the training dataset; (2) Construct a UPS data completion network model, which includes an MLP anomaly perception adjacency matrix and two spatiotemporal bidirectional graph convolutional modules, an adaptive pooling layer and a missing value imputation module set sequentially. The input data is fed into the MLP anomaly-aware adjacency matrix and the first spatiotemporal-bidirectional graph convolutional module, respectively. The output of the first spatiotemporal-bidirectional graph convolutional module serves as the input to the second spatiotemporal-bidirectional graph convolutional module. The MLP anomaly-aware adjacency matrix processes the input data to generate a dynamic adjacency matrix. This dynamic adjacency matrix adaptively adjusts edge weights to capture anomaly signals by fusing the node dependencies learned by the MLP with the feature propagation matrix weighted by anomaly scores. The generated dynamic adjacency matrix is ​​simultaneously input into both spatiotemporal-bidirectional graph convolutional modules. The spatiotemporal-bidirectional graph convolution module includes a first bidirectional graph convolutional network, a temporal causal convolutional network, and a second bidirectional graph convolutional network arranged sequentially. The dynamic adjacency matrix of the spatiotemporal-bidirectional graph convolution module is simultaneously input to both bidirectional graph convolutional networks. The bidirectional graph convolutional network, based on the dynamic adjacency matrix, mines the lateral spatial features of the input data through forward graph convolution and backward graph convolution combined with Chebyshev multinomials, thereby reducing computational complexity. The temporal causal convolutional network models the longitudinal temporal features of the input data, ensuring that the output depends only on the current and historical input data, avoiding interference from future information. The spatiotemporal-bidirectional graph convolution module integrates the spatial and temporal features extracted by the first bidirectional graph convolutional network, the temporal causal convolutional network, and the second bidirectional graph convolutional network as the output. The output of the second spatiotemporal-bidirectional graph convolution module is input into an adaptive pooling layer to restore the data dimensions; then, the missing value imputation module fills in the missing data at the marked locations. (3) The constructed UPS data completion network model is trained using the training dataset. During the training process, the loss function value is calculated and the model parameters are iteratively updated to minimize the completion error. (4) Use the trained UPS data completion network model to complete UPS data.

2. The UPS data completion method for bidirectional spatiotemporal graph convolutional fusion MLP anomaly perception according to claim 1, characterized in that, In step (1), for the acquired UPS data, an hourly time window is constructed to cover short- and medium-term trends. Then, the UPS data is sampled through the time window to construct a time series dataset D = (x t-w ,x t-w+1 ,...,x t ); Missing locations in the time-series dataset are marked using a mask matrix, where 0 represents normal data and 1 represents missing data; thus forming the training dataset (x t-w ,x t-w+1 ,...,x t (mask).

3. The UPS data completion method for bidirectional spatiotemporal graph convolutional fusion MLP anomaly perception according to claim 1, characterized in that, In step (2), the method for constructing the dynamic adjacency matrix is ​​as follows: The node feature pairs are modeled using an MLP model, and the edge weights f() are dynamically learned to generate an adjacency matrix. e ij =MLP([X i ||X j ]) A attn [i,j]=softmax(e ij ) Among them, [X i ||X j ] indicates passing through node X i Node X j The node feature pairs formed by splicing; MLP represents a multilayer perceptron; e ij Represent the edge weight from node i to j; represent all edge weights e ij The organization forms an adjacency matrix A atten [i,j]; Calculate anomaly scores for feature dimensions and construct a feature propagation matrix based on these anomaly scores: Among them, score f The anomaly score represents the feature dimension, x. f For all samples of the f-th feature, Var(x) f ) represents the variance of the f-th feature over all time periods, mask b,t,f This indicates whether the node has missing data; in, A represents the normalized outlier score. anom [i,j] represents the feature propagation matrix based on anomaly scores; By fusing the adjacency matrix and the feature propagation matrix based on anomaly scores, a dynamic adjacency matrix is ​​generated. A[i,j]=α·A attn [i,j]+(1-α)·A anom [i,j] Where α is the fusion weight.

4. The UPS data completion method for bidirectional spatiotemporal graph convolutional fusion MLP anomaly perception according to claim 1, characterized in that, In step (2), the expressions for the forward propagation convolution (TD-GCN) and backward propagation convolution (BU-GCN) of the bidirectional graph convolutional network are as follows: in, To add a positive adjacency matrix with self-loops, Let A be the reverse adjacency matrix, A be the dynamic adjacency matrix, and I be the identity matrix. in, These represent the activation matrices of the l-th layer during forward and backward propagation, respectively. Let represent the trainable weights of the l-th layer in forward and backward propagation, respectively, and σ(·) represent the activation function.

5. The UPS data completion method for bidirectional spatiotemporal graph convolutional fusion MLP anomaly perception according to claim 1, characterized in that, In step (2), the convolution kernel is replaced by a Chebyshev polynomial, the expression of which is: in, These represent the forward convolution weights and the backward convolution weights, respectively. The normalized Laplace matrix, for The transpose of T k () represents the Chebyshev polynomial, and x is the input signal.

6. The UPS data completion method for bidirectional spatiotemporal graph convolutional fusion MLP anomaly perception according to claim 1, characterized in that, In step (2), the temporal causal convolutional network models the temporal dependency of the input data through causal convolution, and uses only historical and current data to fill the input data forward, forcing the convolution to depend only on the current and past times; y t <f(x t ,x t-1 ,…,x1) Among them, y t Let f() represent the output features of a temporal causal convolutional network, and x represent the convolution operation. t This represents the data at time step t.

7. The UPS data completion method for bidirectional spatiotemporal graph convolutional fusion MLP anomaly perception according to claim 1, characterized in that, In step (2), the spatiotemporal-bidirectional graph convolution module performs secondary spatial mining of spatiotemporal features through the hierarchical structure of the first bidirectional graph convolutional network, the temporal causal convolutional network, and the second bidirectional graph convolutional network: First, the spatial features of the original input data are mined through the first bidirectional graph convolutional network; then, the temporal dependencies are learned through the temporal causal convolutional network; finally, the spatial features after fusing temporal features are deeply integrated again through the second bidirectional graph convolutional network.

8. The UPS data completion method for bidirectional spatiotemporal graph convolutional fusion MLP anomaly perception according to claim 1, characterized in that, In step (3), the loss function uses mean squared error: Among them, X MSE The mean squared error is represented by y, where n is the sample size. i Let i be the true value of the i-th sample. Let be the predicted value for the i-th sample.

9. A computer device, comprising a memory and a processor, characterized in that, When the processor executes the computer program stored in the memory, it implements the method according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method described in any one of claims 1-8.

Images