Transformer fault prediction method based on multi-source data fusion and related device

Abstract

The invention provides a transformer fault prediction method based on multi-source data fusion and a related device, and belongs to the technical field of transformer fault prediction. According to the method, stationary time sequence data and non-stationary time sequence data in the transformer are obtained in real time; preprocessing the stationary time sequence data and the non-stationary time sequence data which are acquired in real time, eliminating noise, filling missing data and eliminating abnormal data to obtain preprocessed stationary time sequence data and non-stationary time sequence data; performing feature extraction on the preprocessed stationary time sequence data and non-stationary time sequence data to obtain stationary time sequence features and non-stationary time sequence features; fusing the stationary time sequence features and the non-stationary time sequence features to obtain fused features; and inputting the fused features into a classifier to obtain a transformer fault prediction result. According to the method, the problem of low accuracy of transformer fault prediction caused by dependence on a single model and limited generalization capability is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer fault prediction, and in particular relates to a transformer fault prediction method based on multi-source data fusion and related devices. Background Art

[0002] As core equipment in power systems, transformer health assessment and fault prediction play a vital role in ensuring stable operation. Traditional transformer fault prediction methods typically rely on empirical rules, manual inspection, or single-sensor data analysis. However, these methods have significant drawbacks that limit their application in complex operating conditions.

[0003] First, reliance on a single data source is a major shortcoming of traditional methods. Typically, these methods rely on a single type of sensor data, such as temperature, current, and partial discharge. This analysis approach ignores the complementarity and correlation between various transformer monitoring data. Transformer failures are often the result of multiple factors. Analysis based on a single data source cannot fully and accurately capture the health of the equipment. This is especially true when detecting subtle or hidden faults, which can lead to significant blind spots and delayed fault identification.

[0004] Secondly, deficiencies in data preprocessing and feature extraction are also major bottlenecks for traditional methods. Traditional methods often rely on simple statistics (such as mean and standard deviation) for feature extraction, failing to fully consider the temporal and multi-scale characteristics of the data. Although some methods have incorporated techniques such as wavelet transforms, many still rely on a fixed mother wavelet for signal processing, which lacks the flexibility to address the multi-scale and multi-band characteristics of the signal. Therefore, when faced with non-stationary signals and data with large instantaneous fluctuations, traditional methods are very limited in their effectiveness and are unable to effectively capture key features, which in turn affects the accuracy of transformer fault prediction.

[0005] Furthermore, during transformer operation, some signals (such as partial discharge and vibration signals) exhibit strong transient fluctuations. The non-stationary nature of these signals makes feature extraction by traditional signal processing methods challenging. Many existing methods improperly handle these non-stationary signals, making it difficult to extract effective fault features, which in turn reduces the accuracy of transformer fault prediction.

[0006] Finally, traditional temporal dependency modeling methods have certain limitations when it comes to transformer fault prediction. Although some existing time series analysis methods can process time series data, they are less capable of modeling long-term dependencies. This is especially true when dealing with large amounts of data. The models' generalization and ability to capture temporal dependencies are insufficient, resulting in poor fault prediction results under complex operating conditions.

[0007] Although wavelet transforms have been widely used in signal processing, traditional wavelet transform methods typically use a single mother wavelet for signal processing. While this method can capture certain frequency domain features, due to the limitations of fixed frequency domain resolution, it cannot adapt to the multi-scale characteristics of different signals. In particular, when processing high- and low-frequency features, some key signal characteristics may be lost. Therefore, existing wavelet transform methods are relatively limited in their ability to extract multi-scale features.

[0008] Another major limitation lies in multimodal data fusion. While some methods attempt to fuse multi-source data, most fail to effectively combine the temporal and frequency characteristics of different data sources, neglecting the dynamic relationships between these signals. This makes it difficult for existing transformer fault prediction methods to comprehensively and accurately assess the health of the equipment, especially in the efficient fusion of multi-source data, leaving significant room for improvement.

[0009] Furthermore, while LSTM (Long Short-Term Memory) networks have achieved promising results in time series data modeling, traditional LSTM models are insufficient when dealing with the complex relationships between multi-source data and non-stationary signals. LSTMs often process features from different sources unevenly, failing to effectively distinguish the importance of each feature. This leads to information loss during feature fusion, which in turn affects model performance and the accuracy of transformer fault prediction. Summary of the Invention

[0010] The purpose of the present invention is to provide a transformer fault prediction method and related devices based on multi-source data fusion, which are used to solve the problem of low accuracy of transformer fault prediction in the prior art.

[0011] In order to achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a transformer fault prediction method based on multi-source data fusion, comprising the following steps: Acquire stable time series data and non-stationary time series data in the transformer in real time, wherein the stable time series data includes oil chromatogram dissolved gas content, temperature, current and voltage, and the non-stationary time series data includes partial discharge signals and vibration signals; Preprocess the stationary time series data and non-stationary time series data acquired in real time to remove noise, fill in missing data, and remove abnormal data to obtain preprocessed stationary time series data and non-stationary time series data; Feature extraction is performed on the pre-processed stationary time series data and non-stationary time series data to obtain stationary time series features and non-stationary time series features; Fuse the stationary time series features and the non-stationary time series features to obtain the fused features; The fused features are input into the classifier to obtain the transformer fault prediction results.

[0012] A further improvement of the present invention is that, in the steps of preprocessing the acquired stationary time series data and non-stationary time series data, removing noise, filling missing data and removing abnormal data to obtain preprocessed stationary time series data and non-stationary time series data, a filtering method or a smoothing method is specifically used to remove noise, a deletion method, a mean filling method, an interpolation method or a multiple interpolation method is specifically used to fill missing data, and a statistical method or a model-based anomaly detection method is specifically used to remove abnormal data.

[0013] A further improvement of the present invention is that, in the step of extracting features from the preprocessed stationary time series data and non-stationary time series data respectively to obtain stationary time series features and non-stationary time series features, a sliding window statistics method is specifically used to extract features from the preprocessed stationary time series data, a multi-head attention mechanism and a weight gating mechanism are specifically used to improve the multi-mother wavelet joint transform, and the multi-mother wavelet joint transform improved by the multi-head attention mechanism and the weight gating mechanism is used to extract features from the preprocessed non-stationary time series data.

[0014] A further improvement of the present invention is that the calculation formula of the multi-head attention mechanism is:

[0015] in, is the output weight matrix, For the i The feature output of each wavelet branch after multi-head attention mechanism or subsequent processing, i For branches, h is the number of attention heads, To concatenate the outputs of multiple heads; The calculation formula of the weight gating mechanism is:

[0016] in, is the final fusion output (non-stationary time series features), For the i The feature output of each wavelet branch after multi-head attention mechanism or subsequent processing, is the gating weight (attention coefficient) of the corresponding branch, N is the total number of wavelet branches, symbol “ ” is to perform element-by-element weighting on the entire feature matrix.

[0017] A further improvement of the present invention is that, in the step of fusing the stationary time series features and the non-stationary time series features to obtain the fused features, a double-layer LSTM is specifically used to fuse the stationary time series features and the non-stationary time series features to obtain the fused features.

[0018] A further improvement of the present invention is that the classifier is a fully connected layer.

[0019] A further improvement of the present invention is that after inputting the fused features into the classifier to obtain the transformer fault prediction result, the present invention further includes: generating a health index according to the transformer fault prediction result; The health status of the transformer is determined according to the health index.

[0020] In a second aspect, the present invention provides a transformer fault prediction system based on multi-source data fusion, comprising a data acquisition module, a data preprocessing module, a feature extraction module, a feature fusion module and a fault prediction module; The data acquisition module is used to acquire the steady time series data and the non-steady time series data in the transformer in real time, wherein the steady time series data includes the oil chromatogram dissolved gas content, temperature, current and voltage, and the non-steady time series data includes partial discharge signals and vibration signals; The data preprocessing module is used to preprocess the stationary time series data and non-stationary time series data acquired in real time, remove noise, fill missing data and remove abnormal data, and obtain preprocessed stationary time series data and non-stationary time series data; The feature extraction module is used to extract features from the pre-processed stationary time series data and non-stationary time series data, respectively, to obtain stationary time series features and non-stationary time series features; The feature fusion module is used to fuse the stationary time series features and the non-stationary time series features to obtain the fused features; The fault prediction module is used to input the fused features into the classifier to obtain the transformer fault prediction result.

[0021] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the transformer fault prediction method based on multi-source data fusion introduced above are implemented.

[0022] In a fourth aspect, the present invention provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the transformer fault prediction method based on multi-source data fusion introduced above.

[0023] Compared with the prior art, the present invention has the following beneficial effects: This invention is an improvement. Compared with existing transformer fault prediction methods, on the one hand, it considers not only stationary time series data (oil chromatogram dissolved gas content, temperature, current, and voltage) but also non-stationary time series data (partial discharge signals and vibration signals). Considering only stationary time series data can lead to low accuracy in later transformer fault prediction. Comprehensive consideration of both stationary and non-stationary time series data can improve the accuracy of later transformer fault prediction. On the other hand, this invention fuses stationary and non-stationary time series features to obtain fused features, which can reduce the risk of overfitting in the later prediction model (classifier) and thus improve the accuracy of transformer fault prediction.

[0024] Furthermore, the present invention discloses the steps of extracting features from pre-processed stationary time series data and non-stationary time series data respectively to obtain stationary time series features and non-stationary time series features, specifically using a multi-head attention mechanism and a weight gating mechanism to improve the multi-mother wavelet joint transform, and using the multi-mother wavelet joint transform improved by the multi-head attention mechanism and the weight gating mechanism to extract features from the pre-processed non-stationary time series data. The present invention uses a multi-head attention mechanism to model the time series features extracted from each wavelet branch, which can fully explore the local and global dynamic correlations of non-stationary time series data, improve the modeling ability of non-stationary and large instantaneous fluctuation data, and the mechanism can enhance the sensitivity and robustness of the model to complex signal features, thereby ensuring higher transformer fault prediction accuracy in the later stage. The present invention also uses a weight gating mechanism to learn a dynamically adjusted importance weight for each wavelet branch, ensuring that key non-stationary time series features can be better selected during the feature fusion process, thereby improving the flexibility and accuracy of feature extraction.

[0025] Furthermore, the present invention discloses a step of fusing stationary and non-stationary time series features to obtain fused features, specifically employing a two-layer LSTM to fuse the stationary and non-stationary time series features to obtain the fused features. The two-layer LSTM can simultaneously capture the temporal dependencies between stationary and non-stationary time series features, significantly improving the comprehensiveness and robustness of feature representation. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of a transformer fault prediction method based on multi-source data fusion according to the present invention; Figure 2 Schematic diagram of a transformer fault prediction system based on multi-source data fusion according to the present invention; Figure 3 Schematic diagram of the structure of the electronic device of the present invention. DETAILED DESCRIPTION

[0027] In order to further understand the content of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the embodiments are only for explaining the present invention and are not intended to limit it.

[0028] The transformer fault prediction method based on multi-source data fusion proposed in the present invention performs feature extraction on pre-processed stationary time series data and non-stationary time series data respectively to obtain stationary time series features and non-stationary time series features; fuses the stationary time series features and the non-stationary time series features to obtain fused features; and inputs the fused features into a classifier to obtain a transformer fault prediction result. Compared with the existing technology, the present invention effectively solves the problem in the existing technology of relying on a single model and having limited generalization ability, resulting in low accuracy in transformer fault prediction.

[0029] Example 1: The flowchart of the transformer fault prediction method based on multi-source data fusion of the present invention is as follows: Figure 1 As shown, the transformer fault prediction method based on multi-source data fusion of the present invention includes the following steps: S1. Acquire the steady time series data and non-steady time series data in the transformer in real time, wherein the steady time series data includes the oil chromatogram dissolved gas content, temperature, current and voltage, and the non-steady time series data includes partial discharge signals and vibration signals.

[0030] S2. Preprocess the stationary time series data and non-stationary time series data acquired in real time to remove noise, fill in missing data, and remove abnormal data to obtain preprocessed stationary time series data and non-stationary time series data.

[0031] S3. Perform feature extraction on the preprocessed stationary time series data and non-stationary time series data respectively to obtain stationary time series features and non-stationary time series features.

[0032] S4. Fuse the stationary time series features and the non-stationary time series features to obtain fused features.

[0033] S5. Input the fused features into the classifier to obtain the transformer fault prediction result.

[0034] Example 2: The schematic diagram of the transformer fault prediction system based on multi-source data fusion of the present invention is shown in FIG. Figure 2 As shown, the transformer fault prediction system based on multi-source data fusion of the present invention includes a data acquisition module, a data preprocessing module, a feature extraction module, a feature fusion module and a fault prediction module.

[0035] The data acquisition module is used to acquire the steady time series data and non-steady time series data in the transformer in real time. The steady time series data includes the oil chromatogram dissolved gas content, temperature, current and voltage, and the non-steady time series data includes partial discharge signals and vibration signals.

[0036] The data preprocessing module is used to preprocess the stationary time series data and non-stationary time series data acquired in real time, remove noise, fill missing data and remove abnormal data, and obtain preprocessed stationary time series data and non-stationary time series data.

[0037] The feature extraction module is used to extract features from the preprocessed stationary time series data and non-stationary time series data respectively to obtain stationary time series features and non-stationary time series features.

[0038] The feature fusion module is used to fuse the stationary time series features and the non-stationary time series features to obtain the fused features.

[0039] The fault prediction module is used to input the fused features into the classifier to obtain the transformer fault prediction results.

[0040] Example 3: The transformer fault prediction method based on multi-source data fusion of the present invention comprises the following steps: S1. Acquire the stationary time series data and non-stationary time series data of the transformer in real time.

[0041] First, the stationary and non-stationary time series data of the transformer are acquired in real time. The stationary time series data includes the oil chromatogram dissolved gas content, temperature, current and voltage, while the non-stationary time series data includes partial discharge signals and vibration signals.

[0042] S2. Preprocess the stationary time series data and non-stationary time series data acquired in real time to remove noise, fill in missing data, and remove abnormal data to obtain preprocessed stationary time series data and non-stationary time series data.

[0043] In this step, filtering or smoothing methods are specifically used to remove noise, deletion, mean filling, interpolation or multiple interpolation methods are specifically used to fill missing data, and statistical methods or model-based anomaly detection methods are specifically used to remove abnormal data.

[0044] S3. Perform feature extraction on the preprocessed stationary time series data and non-stationary time series data respectively to obtain stationary time series features and non-stationary time series features.

[0045] This step specifically uses a sliding window statistical method to extract features from the preprocessed stationary time series data (oil chromatogram dissolved gas content, temperature, current, and voltage). Specifically, a multi-head attention mechanism and a weighted gating mechanism are used to improve the multi-mother wavelet joint transform. This improved multi-mother wavelet joint transform is then used to extract features from the preprocessed non-stationary time series data (partial discharge signals and vibration signals). In this embodiment, the improved multi-mother wavelet joint transform using the multi-head attention mechanism and the weighted gating mechanism is called the Gated Multi-Wavelet Attention (GMWA) module.

[0046] The following is a detailed description of the sliding window statistics method: This embodiment takes the oil chromatogram dissolved gas content (gas content is also called gas concentration) as an example, assuming that the length of the oil chromatogram dissolved gas content is N, using { 1, 2,... , N} indicates that, 1, 2,... , N represents the gas concentration at time t. Assume that the window size of the sliding window is W, that is, each time window contains W data points. The sliding window statistics method extracts features by sliding the window on the time series. The step size of each sliding is S, usually S=1 (i.e., step-by-step sliding). Each window W i The position in the time series is indexed by i Determine that the oil chromatogram dissolved gas concentration data in the window is W i ={ i , i+1 ,... , i+w-1}, the number of data points in the window is W , the starting position of the window i Satisfies 1≤i≤N−W+1.

[0047] For each sliding window W i , extract the following statistical features:

[0048]

[0049]

[0050]

[0051]

[0052]

[0053] in, For the i The average value of the dissolved gas concentration in the oil chromatogram within the window, For the i The standard deviation of the dissolved gas concentration in the oil chromatogram within a window, and They are window W i The maximum and minimum values within is the range of data within the window, For the i The rate of change of dissolved gas concentration in oil chromatogram within a window, For the i The linear trend of the change of dissolved gas concentration in oil chromatogram within a window, i is the number of windows, t For the moment, x is a variable, is the number of data points in the window.

[0054] The feature data extracted in multiple time windows are organized into a non-time series feature matrix. Each row represents a feature sample in a time window, and each column represents a feature. The calculation formula is as follows:

[0055]

[0056] Among them, all the extracted features (such as mean, standard deviation, maximum value, minimum value, rate of change and slope, etc.) form a feature vector , all windows W i The extracted eigenvectors are merged into a feature matrix F, that is, the original stationary time series data is converted into a non-time series feature matrix (also called stationary time series features) through the sliding window statistics method.

[0057] The non-time series feature matrix samples extracted by the sliding window statistics method in this embodiment can be further used for subsequent prediction model training, avoiding excessive dependence on time points, while also retaining the potential effective information in the original stationary time series data.

[0058] The following is a detailed description of the improvement of the multi-mother wavelet joint transform (multi-mother wavelet fusion module) using the multi-head attention mechanism and weight gating mechanism: This embodiment introduces a multi-mother wavelet parallel decomposition strategy (also called a multi-mother wavelet joint transform) based on traditional wavelet transforms. This strategy constructs multiple independent wavelet branches, each using a different mother wavelet function to adapt to the signal's response characteristics in different frequency bands. Different wavelet functions have varying time and frequency domain resolutions, specifically responding to certain types of fault characteristics (such as high-frequency spikes and low-frequency slow variations). By combining the transform results of multiple mother wavelet functions, the ability to capture the features of complex non-stationary signals (also known as non-stationary time series data) can be effectively improved.

[0059] Specifically, assuming that the original signal is x(t), select N mother wavelet functions {ψ1(t), ψ2(t),…, ψ N (t)}, input x(t) into each wavelet transform branch respectively to obtain the corresponding wavelet coefficient feature sequence {F1, F2, ..., F N}. The mathematical expression of each wavelet transform is as follows:

[0060] in, For the i mother wavelet function, is the scale factor, controlling the frequency resolution, is the translation factor, controlling the time positioning, For the i The time-frequency characteristics obtained by the wavelet branches are Indicates the use of i The wavelet function of the input signal x(t) is scaled by and translation factors The wavelet coefficients (time-frequency features) obtained under t For the moment, is a variable.

[0061] After completing the multi-mother wavelet decomposition and obtaining multiple wavelet feature sequences Finally, this embodiment introduces the Multi-Head Attention (MHA) mechanism to model the temporal characteristics of each wavelet branch to fully explore the dynamic correlation between local and global. i The feature sequence extracted from the wavelet branches is , where T is the time step length and d is the feature dimension of each step. Then each MHA module first projects the input features into query, key, and value vectors:

[0062] in, are sequences of vectors for query, key, and value, respectively, It is for i The learnable weight matrix of each branch is used to map the original features to the attention space.

[0063] Calculate the weighted relationship separately on each attention head:

[0064] in, For the i The first wavelet branch j The output features of the attention head, They are the query, key, and value vectors obtained by learning linear transformation of the input features in the attention head, and their dimensions are , where T is the time step length, is the feature dimension of each attention head. is the activation function, for The transposed matrix of .

[0065] The outputs of multiple heads are concatenated and linearly transformed to obtain the final attention output of the wavelet branch:

[0066] in, is the output weight matrix, For the i The feature output of each wavelet branch after multi-head attention mechanism or subsequent processing, i For branches, h is the number of attention heads, To concatenate the output of multiple heads.

[0067] After completing the time series modeling of wavelet features, this embodiment introduces a weight gating mechanism to further achieve adaptive fusion and optimal selection of the outputs of different wavelet branches. The core idea of this mechanism is to learn a dynamically adjustable importance weight for each wavelet branch to control its contribution to the final fused features, thereby enhancing the multi-mother wavelet fusion module's ability to focus on key frequency bands or effective features, and improving the discernibility and effectiveness of feature fusion.

[0068] The weight gating mechanism learns a scalar weight for it, which is used to represent the effectiveness of the branch. The calculation formula of the weight is:

[0069] in, is the gating weight (attention coefficient) of the corresponding branch (decomposed into multiple wavelet branches with different frequency characteristics through multi-mother wavelet joint transform), To perform global average pooling or maximum pooling on sequence features (obtain vector), , is the linear transformation parameter, For the Sigmoid activation function, ensure .

[0070] All gate weights Acts on the corresponding branch output , the final fusion output is:

[0071] in, is the final fusion output (non-stationary time series features), For the i The feature output of each wavelet branch after multi-head attention mechanism or subsequent processing, is the gating weight (attention coefficient) of the corresponding branch, This reflects the importance of this branch in the overall feature fusion. N is the total number of wavelet branches, symbol “ ” is to perform element-by-element weighting on the entire feature matrix.

[0072] In this embodiment, all gate weights It is usually generated dynamically through a learnable attention module or gating network. The important features of all wavelet branches are comprehensively considered to achieve adaptive fusion of multi-scale information.

[0073] S4. Fuse the stationary time series features and the non-stationary time series features to obtain fused features.

[0074] In this step, a double-layer LSTM is used to fuse the stationary time series features and the non-stationary time series features to obtain the fused features.

[0075] Specifically, an input layer is constructed for stationary time series features and non-stationary time series features, respectively. One input layer is used to receive stationary time series features, and the other input layer is used to receive non-stationary time series features. Independent LSTM layers are designed for these two input channels (stationary time series features and non-stationary time series features). Each LSTM layer will learn the temporal dependencies of the corresponding features. For each input X1 and X2 (representing non-stationary time series features and stationary time series features, respectively), after LSTM processing, two outputs h1 and h2 are obtained. The calculation formulas for h1 and h2 are as follows:

[0076]

[0077] in, and Represent the LSTM layers for processing non-stationary time series features and stationary time series features respectively.

[0078] The outputs of two LSTM layers (also called double-layer LSTM) are fused into a feature vector through concatenation or features from different sources are weightedly combined, so that the information of non-stationary time series features and stationary time series features can be utilized simultaneously.

[0079] S5. Input the fused features into the classifier to obtain the transformer fault prediction result.

[0080] The fused features are input into the classifier (the classifier in this embodiment is a fully connected layer) to obtain the transformer fault prediction result. After the fused features are input into the classifier to obtain the transformer fault prediction result, the following steps are also included: Generate a health index based on transformer fault prediction results; The health status of the transformer is determined based on the health index.

[0081] The health index h is generated according to the transformer fault prediction results, and the value range of the health index h is between [0, 1].

[0082] When the health index h is 0, the transformer is completely healthy. When the health index h is 1, the transformer is completely faulty. When the health index h is greater than 0.6, the transformer is in normal condition and in good health. When the health index h is greater than or equal to 0.4 and less than or equal to 0.6, the transformer is at potential failure risk and requires regular monitoring and maintenance. When the health index h is less than 0.4, the transformer has experienced a serious failure or is about to fail, and operations and maintenance personnel must take immediate measures to repair it.

[0083] Example 4: See also Figure 3 As shown, the present invention also provides an electronic device 100 for a transformer fault prediction method based on multi-source data fusion; the electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on the at least one processor 102, and at least one communication bus 104.

[0084] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the transformer fault prediction method based on multi-source data fusion described in Example 1 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101. The memory 101 can mainly include a program storage area and a data storage area. The program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data (such as audio data) created based on the use of the electronic device 100. In addition, the memory 101 can include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device.

[0085] The at least one processor 102 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 102 may be a microprocessor or any conventional processor, etc. The processor 102 is the control center of the electronic device 100 and connects various parts of the entire electronic device 100 using various interfaces and lines.

[0086] The memory 101 in the electronic device 100 stores a plurality of instructions to implement a transformer fault prediction method based on multi-source data fusion, and the processor 102 can execute the plurality of instructions to implement: Acquire stable time series data and non-stationary time series data in the transformer in real time, wherein the stable time series data includes oil chromatogram dissolved gas content, temperature, current and voltage, and the non-stationary time series data includes partial discharge signals and vibration signals; Preprocess the stationary time series data and non-stationary time series data acquired in real time to remove noise, fill in missing data, and remove abnormal data to obtain preprocessed stationary time series data and non-stationary time series data; Feature extraction is performed on the pre-processed stationary time series data and non-stationary time series data to obtain stationary time series features and non-stationary time series features; Fuse the stationary time series features and the non-stationary time series features to obtain the fused features; The fused features are input into the classifier to obtain the transformer fault prediction results.

[0087] Example 5: If the module / unit integrated in the electronic device 100 is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device that can carry the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory and read-only memory (ROM, Read-Only Memory).

[0088] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts 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, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0090] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0091] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A transformer fault prediction method based on multi-source data fusion, characterized in that: The following steps are involved: Acquire stable time series data and non-stationary time series data in the transformer in real time, wherein the stable time series data includes oil chromatogram dissolved gas content, temperature, current and voltage, and the non-stationary time series data includes partial discharge signals and vibration signals; Preprocess the stationary time series data and non-stationary time series data acquired in real time to remove noise, fill in missing data, and remove abnormal data to obtain preprocessed stationary time series data and non-stationary time series data; Feature extraction is performed on the pre-processed stationary time series data and non-stationary time series data to obtain stationary time series features and non-stationary time series features; Fuse the stationary time series features and the non-stationary time series features to obtain the fused features; The fused features are input into the classifier to obtain the transformer fault prediction results.

2. The transformer fault prediction method based on multi-source data fusion according to claim 1 is characterized in that: In the step of preprocessing the acquired stationary time series data and non-stationary time series data, removing noise, filling missing data and removing abnormal data to obtain preprocessed stationary time series data and non-stationary time series data, a filtering method or a smoothing method is specifically used to remove noise, a deletion method, a mean filling method, an interpolation method or a multiple interpolation method is specifically used to fill missing data, and a statistical method or a model-based anomaly detection method is specifically used to remove abnormal data.

3. The transformer fault prediction method based on multi-source data fusion according to claim 1 is characterized in that: In the step of extracting features from the preprocessed stationary time series data and non-stationary time series data respectively to obtain stationary time series features and non-stationary time series features, a sliding window statistics method is specifically used to extract features from the preprocessed stationary time series data, a multi-head attention mechanism and a weight gating mechanism are specifically used to improve the multi-mother wavelet joint transform, and the multi-mother wavelet joint transform improved by the multi-head attention mechanism and the weight gating mechanism is used to extract features from the preprocessed non-stationary time series data.

4. The transformer fault prediction method based on multi-source data fusion according to claim 3 is characterized in that: The calculation formula of the multi-head attention mechanism is: in, is the output weight matrix, For the i The feature output of each wavelet branch after multi-head attention mechanism or subsequent processing, i For branches, h is the number of attention heads, To concatenate the outputs of multiple heads; The calculation formula of the weight gating mechanism is: in, is the final fusion output (non-stationary time series features), For the i The feature output of each wavelet branch after multi-head attention mechanism or subsequent processing, is the gating weight (attention coefficient) of the corresponding branch, N is the total number of wavelet branches, symbol" ” is to perform element-by-element weighting on the entire feature matrix.

5. The transformer fault prediction method based on multi-source data fusion according to claim 1 is characterized in that: In the step of fusing the stationary time series features and the non-stationary time series features to obtain the fused features, a double-layer LSTM is specifically used to fuse the stationary time series features and the non-stationary time series features to obtain the fused features.

6. The transformer fault prediction method based on multi-source data fusion according to claim 1 is characterized in that: The classifier is a fully connected layer.

7. The transformer fault prediction method based on multi-source data fusion according to claim 1 is characterized in that: After the fused features are input into the classifier to obtain the transformer fault prediction results, the following steps are also included: generating a health index according to the transformer fault prediction result; The health status of the transformer is determined according to the health index.

8. A transformer fault prediction system based on multi-source data fusion, characterized in that: It includes data acquisition module, data preprocessing module, feature extraction module, feature fusion module and fault prediction module; The data acquisition module is used to acquire the steady time series data and the non-steady time series data in the transformer in real time, wherein the steady time series data includes the oil chromatogram dissolved gas content, temperature, current and voltage, and the non-steady time series data includes partial discharge signals and vibration signals; The data preprocessing module is used to preprocess the stationary time series data and non-stationary time series data acquired in real time, remove noise, fill missing data and remove abnormal data, and obtain preprocessed stationary time series data and non-stationary time series data; The feature extraction module is used to extract features from the pre-processed stationary time series data and non-stationary time series data, respectively, to obtain stationary time series features and non-stationary time series features; The feature fusion module is used to fuse the stationary time series features and the non-stationary time series features to obtain the fused features; The fault prediction module is used to input the fused features into the classifier to obtain the transformer fault prediction result.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the transformer fault prediction method based on multi-source data fusion according to any one of claims 1 to 7 are implemented.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the transformer fault prediction method based on multi-source data fusion according to any one of claims 1 to 7 are implemented.

Images