Fault diagnosis method and device, electronic equipment, storage medium and program product

Abstract

The application provides a fault diagnosis method and device, electronic equipment, a storage medium and a program product, and relates to the technical field of power grids. The fault diagnosis method generates evidence weighted fusion data corresponding to the initial multi-source time series data set by performing weighted evidence fusion processing on the initial multi-source time series data set, simplifies the multi-source data fusion process without complex additional signal preprocessing, improves the fusion efficiency of multi-source data, solves the problem of poor accuracy of fixed evidence weight or subjective experience-based allocation in traditional methods, makes the data with high volatility and high correlation have higher priority after weighted evidence fusion, further inputs the evidence weighted fusion data into a deep belief network model for fault diagnosis, improves the accuracy of fault feature extraction, realizes high-precision fault classification, and improves the accuracy of fault diagnosis classification results.

Description

Technical Field

[0001] This application relates to the field of power grid technology, and in particular to a fault diagnosis method, device, electronic equipment, storage medium, and program product. Background Technology

[0002] As the core equipment of the power system, power transformers undertake the critical functions of voltage transformation, energy transmission, and stable operation of the power grid. With the rapid development of the economy and society, the power grid load continues to rise. Transformers (such as oil-immersed self-cooled transformers) operate under high temperature, high voltage, high load fluctuation, and complex electromagnetic environment for a long time. Various types of faults, such as low temperature overheating, medium and low temperature overheating, high temperature overheating, low energy discharge, high energy discharge, and partial discharge, occur frequently. If early latent faults are not diagnosed in time, it may lead to equipment shutdown or even large-scale power outages, causing huge economic losses and social impact.

[0003] Traditional transformer fault diagnosis techniques often rely on manual experience or a single data source (such as oil chromatography data or partial discharge signals), but these suffer from low diagnostic efficiency, high false positive rates, and difficulty in adapting to complex scenarios with multiple types of faults coexisting. Therefore, there is an urgent need for an intelligent fault diagnosis method that requires no complex signal processing, effectively integrates multi-source data, automatically extracts fault features, and achieves rapid and accurate classification. This would improve the diagnostic efficiency and accuracy of various transformer faults and ensure the reliability of power grid supply. Summary of the Invention

[0004] This application provides a fault diagnosis method, device, electronic equipment, storage medium, and program product, which is an intelligent fault diagnosis method that can integrate multi-source data, automatically extract fault features, and achieve rapid and accurate fault classification. It can improve the diagnostic efficiency and accuracy of various types of transformer faults and ensure the reliability of power grid supply.

[0005] In a first aspect, this application provides a fault diagnosis method, comprising: acquiring an initial multi-source time series dataset of a target transformer; performing weighted evidence fusion processing on the initial multi-source time series dataset to generate evidence-weighted fusion data corresponding to the initial multi-source time series dataset; and inputting the evidence-weighted fusion data into a deep belief network model for fault diagnosis to obtain the fault diagnosis result output by the deep belief network model.

[0006] In one possible implementation, weighted evidence fusion processing is performed on the initial multi-source time-series dataset to generate weighted evidence fusion data corresponding to the initial multi-source time-series dataset. This includes: determining the criterion layer weights and scheme layer weights corresponding to the initial multi-source time-series dataset based on the improved analytic hierarchy process (AHP); combining the criterion layer weights and scheme layer weights to obtain the evidence weights of each evidence body in the initial multi-source time-series dataset; weighting and summing the evidence weights of each evidence body and the basic probability allocation of each evidence body under a preset identification framework according to each proposition in the preset identification framework to obtain the weighted evidence of each proposition, where the preset identification framework is a set of transformer fault and fault-free states; fusing the weighted evidence of each proposition based on preset evidence synthesis rules to obtain the comprehensive trust level of each proposition; and concatenating the comprehensive trust levels of each proposition into a fixed-dimensional feature vector according to the order of each proposition in the preset identification framework to generate weighted evidence fusion data.

[0007] In one possible implementation, based on the improved analytic hierarchy process (AHP), the criterion-level weights and scheme-level weights corresponding to the initial multi-source time-series dataset are determined. This includes: determining the quantitative values ​​of each piece of evidence in the initial multi-source time-series dataset under each preset criterion based on the quantitative values ​​of each piece of evidence under each preset criterion; calculating the covariance matrix between each preset criterion and the variance between each piece of evidence under each preset criterion based on the quantitative values ​​of each piece of evidence under each preset criterion; converting the covariance matrix into a relative covariance matrix and calculating the criterion-level weights based on the relative covariance matrix; constructing the fuzzy preference relation matrix corresponding to each preset criterion based on the variance between each piece of evidence under each preset criterion; converting the fuzzy preference relation matrix corresponding to each preset criterion into a consistency matrix and calculating the scheme-level weights based on the consistency matrix.

[0008] In one possible implementation, the deep belief network model is trained as follows: a sample set of data is acquired, including training and test data; weighted evidence fusion is performed on the sample set data to generate weighted evidence fusion sample data, which includes weighted evidence fusion training and test data; the weighted evidence fusion training data is input into the deep belief network model to be trained to train the model parameters, resulting in a pre-trained deep belief network model; and the weighted evidence fusion test data is input into the pre-trained deep belief network model to validate the model parameters, resulting in the final deep belief network model.

[0009] In one possible implementation, the evidence-weighted fusion training set data is input into the deep belief network model to be trained to train the model parameters, resulting in a pre-trained deep belief network model. This includes: inputting the evidence-weighted fusion training set data into the deep belief network model to be trained; performing layer-by-layer unsupervised training and reconstruction optimization on the evidence-weighted fusion training set data based on a Restricted Boltzmann Machine (RBM) to extract high-level abstract features; inputting the high-level abstract features into an attention mechanism module for feature enhancement to obtain globally enhanced features; and inputting the globally enhanced features into a top-level classifier for backpropagation global fine-tuning based on a gradient algorithm to optimize all model parameters of the deep belief network model to be trained, thus obtaining the pre-trained deep belief network model.

[0010] In one possible implementation, before performing weighted evidence fusion processing on the initial multi-source time series dataset, the fault diagnosis method further includes: normalizing the initial multi-source time series dataset to obtain a normalized initial multi-source time series dataset.

[0011] Secondly, this application provides a fault diagnosis device, comprising:

[0012] The acquisition module is used to acquire the initial multi-source time-series dataset of the target transformer;

[0013] The weighted fusion module is used to perform weighted evidence fusion processing on the initial multi-source time series dataset to generate weighted fused evidence data corresponding to the initial multi-source time series dataset.

[0014] The fault diagnosis module is used to input evidence-weighted fusion data into the deep belief network model for fault diagnosis, and obtain the fault diagnosis results output by the deep belief network model.

[0015] In one possible implementation, the weighted fusion module is specifically used for: determining the criterion layer weights and scheme layer weights corresponding to the initial multi-source time series dataset based on the improved analytic hierarchy process (AHP); combining the criterion layer weights and scheme layer weights to obtain the evidence weights of each evidence body in the initial multi-source time series dataset; weighting and summing the evidence weights of each evidence body and the basic probability allocation of each evidence body under the preset identification framework according to each proposition in the preset identification framework to obtain the weighted evidence of each proposition, where the preset identification framework is a set of transformer fault and fault-free states; fusing the weighted evidence of each proposition based on the preset evidence synthesis rules to obtain the comprehensive trust level of each proposition; and concatenating the comprehensive trust levels of each proposition into a fixed-dimensional feature vector according to the order of each proposition in the preset identification framework to generate weighted fusion evidence data.

[0016] In one possible implementation, the weighted fusion module is further configured to: determine the quantitative values ​​of each piece of evidence in the initial multi-source time-series dataset under each preset criterion based on the quantitative values ​​of each piece of evidence under each preset criterion; calculate the covariance matrix between each preset criterion and the variance between each piece of evidence under each preset criterion based on the quantitative values ​​of each piece of evidence under each preset criterion; convert the covariance matrix into a relative covariance matrix and calculate the criterion layer weights based on the relative covariance matrix; construct the fuzzy preference relation matrix corresponding to each preset criterion based on the variance between each piece of evidence under each preset criterion; convert the fuzzy preference relation matrix corresponding to each preset criterion into a consistency matrix and calculate the scheme layer weights based on the consistency matrix.

[0017] In one possible implementation, the deep belief network model is trained as follows: a sample set of data is acquired, including training and test data; weighted evidence fusion is performed on the sample set data to generate weighted evidence fusion sample data, which includes weighted evidence fusion training and test data; the weighted evidence fusion training data is input into the deep belief network model to be trained to train the model parameters, resulting in a pre-trained deep belief network model; and the weighted evidence fusion test data is input into the pre-trained deep belief network model to validate the model parameters, resulting in the final deep belief network model.

[0018] In one possible implementation, the evidence-weighted fusion training set data is input into the deep belief network model to be trained to train the model parameters, resulting in a pre-trained deep belief network model. This includes: inputting the evidence-weighted fusion training set data into the deep belief network model to be trained; performing layer-by-layer unsupervised training and reconstruction optimization on the evidence-weighted fusion training set data based on a Restricted Boltzmann Machine (RBM) to extract high-level abstract features; inputting the high-level abstract features into an attention mechanism module for feature enhancement to obtain globally enhanced features; and inputting the globally enhanced features into a top-level classifier for backpropagation global fine-tuning based on a gradient algorithm to optimize all model parameters of the deep belief network model to be trained, thus obtaining the pre-trained deep belief network model.

[0019] In one possible implementation, before performing weighted evidence fusion processing on the initial multi-source time series dataset, the fault diagnosis device further includes a processing module (not shown) for: normalizing the initial multi-source time series dataset to obtain a normalized initial multi-source time series dataset.

[0020] Thirdly, this application provides an electronic device, including: a processor and a memory communicatively connected to the processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the fault diagnosis method provided in the first aspect above.

[0021] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the fault diagnosis method provided in the first aspect above.

[0022] Fifthly, this application provides a computer program product, comprising: a computer program that, when executed by a processor, implements the fault diagnosis method provided in the first aspect above.

[0023] This application provides a fault diagnosis method, apparatus, electronic device, storage medium, and program product. The fault diagnosis method acquires an initial multi-source time-series dataset of a target transformer, performs weighted evidence fusion processing on the initial multi-source time-series dataset to generate weighted evidence fusion data corresponding to the initial multi-source time-series dataset, and further inputs the weighted evidence fusion data into a deep belief network model for fault diagnosis, obtaining the fault diagnosis result output by the deep belief network model. This application, by performing weighted evidence fusion processing on the initial multi-source time-series dataset to generate weighted evidence fusion data corresponding to the initial multi-source time-series dataset, eliminates the need for complex additional signal preprocessing, simplifies the multi-source data fusion process, and improves the fusion efficiency of multi-source data. Simultaneously, it solves the problem of fixed evidence weights or poor accuracy based on subjective experience in traditional methods, giving higher priority to highly volatile and highly correlated data after weighted evidence fusion. Furthermore, inputting the weighted evidence fusion data into a deep belief network model for fault diagnosis improves the accuracy of fault feature extraction, achieves high-precision fault classification, and enhances the accuracy of fault diagnosis classification results. Attached Figure Description

[0024] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0025] Figure 1 A flowchart illustrating the fault diagnosis method provided in the embodiments of this application. Figure 1 ;

[0026] Figure 2 A schematic diagram of the hierarchical structure model corresponding to the improved AHP provided in the embodiments of this application;

[0027] Figure 3 A flowchart illustrating the fault diagnosis method provided in the embodiments of this application. Figure 2 ;

[0028] Figure 4 A flowchart illustrating the training method of the deep belief network model provided in the embodiments of this application;

[0029] Figure 5 A schematic diagram illustrating the process of layer-by-layer unsupervised training and reconstruction optimization based on RBM provided in this application embodiment;

[0030] Figure 6 This is a schematic diagram of the structure of the confusion matrix provided in an embodiment of this application;

[0031] Figure 7 A schematic diagram of the model loss curve and accuracy curve of the deep belief network model provided in the embodiments of this application;

[0032] Figure 8 This is a schematic diagram of the structure of the fault diagnosis device provided in the embodiments of this application;

[0033] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0034] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0035] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0036] When a transformer (such as an oil-immersed self-cooled transformer) is operating, the energy loss in the core and windings is converted into heat. This heat is dissipated to the surrounding area via a radiator through the temperature difference between the inside and outside of the transformer and the convection of the oil, maintaining the transformer's normal operation. However, because self-cooling is limited by ambient temperature and system load, in summer when system load is high and external temperatures are consistently high, without external auxiliary cooling measures, the heat generated by the oil-immersed self-cooled transformer cannot be effectively dissipated. This will accelerate the aging of the transformer's insulation, shorten its service life, and seriously threaten the reliability of the power grid. Therefore, establishing an intelligent decision-making system for transformer fault diagnosis and repair can promptly and accurately identify early latent faults in the transformer, improve equipment utilization, reduce power equipment operating costs, and provide reliable technical and management guarantees for ensuring the long-term safe and stable operation of transformers.

[0037] In related technologies, transformer fault diagnosis techniques mostly rely on a single data source (such as oil chromatography data or partial discharge signals) or traditional algorithm models, but both have significant drawbacks:

[0038] 1) Diagnostic methods based on transformer expert systems are limited by the bottleneck of acquiring a complete knowledge base, have weak reasoning ability and are difficult to maintain knowledge, making them difficult to adapt to complex scenarios where multiple types of faults coexist.

[0039] 2) Neural network ensembles, support vector machines and other algorithms have problems such as difficulty in determining network structure and parameters, large amount of computation, and slow convergence speed. In addition, support vector machines need to achieve multi-classification through multiple transformations, which can easily lead to error accumulation and classification overlap.

[0040] 3) The rule-making of the fuzzy comprehensive evaluation method relies too much on human experience and lacks standardized processes, making it difficult to balance diagnostic accuracy and decision-making efficiency when dealing with complex systems.

[0041] 4) Emerging algorithms such as artificial immune systems are still at a low level in terms of mechanism understanding, algorithm construction and engineering application, and their diagnostic reliability is insufficient.

[0042] 5) Many related technologies require complex signal preprocessing and feature extraction steps, have strict requirements on data acquisition cycle, have poor versatility, and do not fully integrate complementary information from multi-source sensor data, resulting in limited reliability of diagnostic results.

[0043] Therefore, there is an urgent need for an intelligent fault diagnosis method that does not require complex signal processing, can effectively integrate multi-source data, automatically extract fault features, and achieve rapid and accurate classification, so as to improve the diagnostic efficiency and accuracy of various types of transformer faults and ensure the reliability of power grid supply.

[0044] Based on the problems existing in related technologies, the embodiments of this application generate evidence-weighted fused data corresponding to the initial multi-source time series dataset by performing weighted evidence fusion processing on the initial multi-source time series dataset. This eliminates the need for complex additional signal preprocessing, simplifies the multi-source data fusion process, and improves the fusion efficiency of multi-source data. At the same time, it solves the problem of fixed evidence weights or poor accuracy based on subjective experience in traditional methods, so that highly volatile and highly correlated data have higher priority after weighted evidence fusion. Furthermore, the evidence-weighted fused data is input into a deep belief network model for fault diagnosis, improving the accuracy of fault feature extraction, achieving high-precision fault classification, and improving the accuracy of fault diagnosis classification results.

[0045] The application scenarios of the embodiments of this application will be described below first.

[0046] The fault diagnosis method provided in this application is applicable to real-time fault diagnosis scenarios for power transformers, and is especially applicable to the diagnosis of various types of faults (such as partial discharge, overheating, high-energy discharge, etc.) in oil-immersed self-cooled transformers under high load and high temperature environments.

[0047] The specific implementation of the fault diagnosis method provided in this application will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0048] Figure 1 A flowchart illustrating the fault diagnosis method provided in the embodiments of this application. Figure 1 .like Figure 1 As shown, a specific implementation of this fault diagnosis method may include the following steps:

[0049] S101, Obtain the initial multi-source time-series dataset of the target transformer.

[0050] For example, the target transformer can be an oil-immersed self-cooled transformer.

[0051] For example, the initial multi-source time-series dataset can be high-dimensional time-series data, such as partial discharge time-series datasets, vibration time-series datasets, and oil chromatography time-series datasets. Specifically, the partial discharge time-series dataset can include multiple partial discharge values ​​arranged in chronological order, the vibration time-series dataset can include multiple vibration amplitude values ​​arranged in chronological order, and the oil chromatography time-series dataset can include multiple concentration values ​​of various characteristic gases in the oil arranged in chronological order.

[0052] For example, partial discharge time series data can be acquired by a partial discharge detector with a dedicated sensor, vibration time series data can be acquired by a vibration sensor (such as an accelerometer) with a data acquisition device, and oil chromatography time series data can be acquired by an oil chromatography monitoring device.

[0053] S102, perform weighted evidence fusion processing on the initial multi-source time series dataset to generate weighted evidence fusion data corresponding to the initial multi-source time series dataset.

[0054] For example, one possible implementation can employ a synergistic mechanism combining the improved Analytic Hierarchy Process (AHP) with Dempster's evidence theory. First, the improved AHP assigns weights to the partial discharge time-series data, vibration time-series data, and oil chromatography time-series data in the initial multi-source time-series dataset. Second, the partial discharge time-series data, vibration time-series data, and oil chromatography time-series data are converted into Basic Probability Assignment (BPA) within the Dempster evidence theory framework, completing the mapping from data to evidence information. Finally, based on the weight coefficients determined by the improved AHP, each piece of evidence is weighted and corrected, and then the conflict resolution and information fusion of multi-source evidence are completed through Dempster's synthesis rules, ultimately outputting weighted fused evidence data that comprehensively reflects the core information of each data source.

[0055] For example, the improved AHP provided in this application embodiment is a method based on data statistical characteristics.

[0056] Figure 2 This is a schematic diagram of the hierarchical structure model corresponding to the improved AHP provided in the embodiments of this application. For example... Figure 2As shown, the evaluation purpose of the top-level target layer in the hierarchical model corresponding to this improved AHP is the credibility of the evidence. That is, the target layer is used to determine the reliability of different evidence bodies (such as partial discharge time series data, vibration time series data, and oil chromatography time series data). The criterion layer includes three core indicators for evaluating the credibility of evidence: the difference criterion, the conflict criterion, and the uncertainty criterion. These are used to measure the quality of the evidence. The difference criterion is used to evaluate the degree of information difference between different evidence bodies (the greater the difference, the stronger the uniqueness of each piece of evidence and the higher the complementary value). The conflict criterion is used to evaluate the degree of contradiction between different evidence bodies (the smaller the conflict, the better the consistency of the evidence and the higher the credibility). The uncertainty criterion is used to evaluate the degree of ambiguity or unknownness of a single piece of evidence (the lower the uncertainty, the stronger the clarity of the evidence and the higher the credibility). The scheme layer includes three specific evidence bodies to be evaluated: evidence body 1, evidence body 2, and evidence body 3 (which can correspond to the evidence transformed from the three data sources: partial discharge time series data, vibration time series data, and oil chromatography time series data, respectively).

[0057] from Figure 2 As can be seen from this, the hierarchical structure model corresponding to the improved AHP first assigns criterion-level weights to the three criteria of difference, conflict, and uncertainty at the criterion level. Then, at the alternative level, the alternative weights of each piece of evidence under these three criteria are evaluated. Finally, based on the criterion-level weights and alternative-level weights, the evidence weight (evidence credibility score) of each piece of evidence is calculated, thereby determining which piece of evidence is more reliable.

[0058] It should be noted that Figure 2 is only an example: on the one hand, the number of evidence bodies in the scheme layer can be adjusted according to actual monitoring needs; on the other hand, the indicators in the criterion layer can also be supplemented based on data characteristics. This application embodiment does not limit the specific number of dimensions in the hierarchical model, only requiring that the logical closed loop of the target layer - criterion layer - scheme layer be satisfied.

[0059] S103, input the evidence-weighted fusion data into the deep belief network model for fault diagnosis, and obtain the fault diagnosis results output by the deep belief network model.

[0060] For example, the fault diagnosis results may include one fault type or multiple different fault types.

[0061] For example, fault types may include low temperature overheating fault, medium and low temperature overheating fault, high temperature overheating fault, low energy discharge fault, high energy discharge fault, partial discharge fault, vibration fault, and no fault.

[0062] For example, a Deep Belief Network (DBN) model can be a DBN model containing multiple hidden layers, wherein the number of input layer nodes, the number of output layer nodes, the model depth, and the number of nodes in each hidden layer are determined based on the dimension of the fault samples and the number of fault types.

[0063] For example, the model parameters of a DBN model include, but are not limited to, momentum, learning rate, initial error, and number of iterations.

[0064] In this embodiment, weighted evidence fusion processing is performed on the initial multi-source time-series dataset to generate weighted evidence fusion data corresponding to the initial multi-source time-series dataset. This eliminates the need for additional signal preprocessing, simplifies the multi-source data fusion process, and improves the fusion efficiency of multi-source data. At the same time, it solves the problem of poor accuracy in traditional methods where evidence weights are fixed or based on subjective experience. This gives higher priority to highly volatile and highly correlated data after weighted evidence fusion. Furthermore, the weighted evidence fusion data is input into a deep belief network model for fault diagnosis, improving the accuracy of fault feature extraction, achieving high-precision fault classification, and improving the accuracy of fault diagnosis classification results.

[0065] The following is combined with Figure 3 The specific implementation method of step S102, which performs weighted evidence fusion processing on the initial multi-source time series dataset to generate the evidence weighted fusion data corresponding to the initial multi-source time series dataset, is described in detail.

[0066] Figure 3 A flowchart illustrating the fault diagnosis method provided in the embodiments of this application. Figure 2 .like Figure 3 As shown, in this fault diagnosis method, a specific implementation of weighted evidence fusion processing on the initial multi-source time series dataset to generate weighted evidence fusion data corresponding to the initial multi-source time series dataset may include the following steps:

[0067] S301, based on the improved analytic hierarchy process, determines the criterion layer weights and scheme layer weights corresponding to the initial multi-source time series dataset based on the initial multi-source time series dataset.

[0068] The improved analytic hierarchy process is similar to the one described above, and will not be elaborated upon here.

[0069] Alternatively, one possible implementation of this step may include the following steps:

[0070] S3011, Based on the initial multi-source time series dataset, determine the criterion quantitative value of each piece of evidence in the initial multi-source time series dataset under each preset criterion.

[0071] For example, the evidence bodies in the initial multi-source time series dataset can be partial discharge time series data, vibration time series data, and oil chromatography time series data, respectively.

[0072] For example, each preset criterion can be as described above. Figure 2 The difference criterion, conflict criterion, and uncertainty criterion are shown in the figure.

[0073] For example, for the difference criterion, the data difference degree between each pair of evidence is calculated based on partial discharge time series data, vibration time series data, and oil chromatography time series data, respectively, and the average value is used as the criterion quantitative value for each evidence under the difference criterion; for the conflict criterion, the data conflict degree between each pair of evidence is calculated based on partial discharge time series data, vibration time series data, and oil chromatography time series data, respectively, and the average value is used as the criterion quantitative value for each evidence under the conflict criterion; for the uncertainty criterion, the uncertainty index (such as standard deviation) of the time series data of each evidence is calculated based on partial discharge time series data, vibration time series data, and oil chromatography time series data, respectively, and the standard deviation is used as the criterion quantitative value for each evidence under the uncertainty criterion.

[0074] Table 1 shows the quantitative values ​​of each piece of evidence in the initial multi-source time-series dataset provided in this application embodiment under each preset criterion.

[0075] Table 1

[0076]

[0077] S3012, based on the quantitative values ​​of each piece of evidence under each preset criterion, calculate the covariance matrix between each preset criterion and the variance between each piece of evidence under each preset criterion.

[0078] For example, the covariance matrix is ​​used to reflect the degree of linear correlation among the three pre-defined criteria: the difference criterion, the conflict criterion, and the uncertainty criterion.

[0079] For example, each element in the covariance matrix can be represented by the following formula:

[0080]

[0081] in, This represents the covariance values ​​corresponding to preset criteria i and preset criteria j, and n represents the number of evidence bodies (e.g., n=3). This represents the quantitative value of the k-th piece of evidence under the preset criterion i. This represents the average of the quantitative values ​​of the n evidence criteria under the preset criterion i. This represents the quantitative value of the k-th piece of evidence under the pre-defined criterion j. This represents the average value of the quantitative values ​​of the n evidence criteria under the preset criterion j.

[0082] For example, when preset criterion i is a difference criterion and preset criterion j is a conflict criterion,

[0083]

[0084]

[0085] For example, the covariance matrix can be a matrix.

[0086] For example, the variance can be the variance between each piece of evidence under each preset criterion, which is used to reflect the degree of dispersion of each piece of evidence under any preset criterion.

[0087] For example, the variance of evidence body m under preset criterion i can be expressed by the following formula:

[0088]

[0089] in, Let represent the variance of evidence body m under the preset criterion i.

[0090] S3013 converts the covariance matrix into a relative covariance matrix and calculates the criterion layer weights based on the relative covariance matrix.

[0091] For example, in one possible implementation, the covariance matrix is ​​normalized to a relative covariance matrix with elements of 1 on the main diagonal.

[0092] For example, each element in the relative covariance matrix can be represented by the following formula:

[0093]

[0094] For example, one possible way to calculate the criterion layer weights based on the relative covariance matrix is ​​to: standardize the relative covariance matrix into pairwise comparison matrices, solve for the largest eigenvalue of each pairwise comparison matrix and the eigenvector corresponding to the largest eigenvalue, and further normalize the eigenvectors to obtain the criterion layer weights.

[0095] For example, each element in a pairwise comparison matrix can be represented by the following formula:

[0096]

[0097] For example, the criterion layer weights can be expressed as .in, This indicates the weight corresponding to the difference criterion. This indicates the weights corresponding to conflict criteria. This represents the weight corresponding to the uncertainty criterion.

[0098] Understandably, the higher the weight of the criterion, the greater its impact on the credibility of the evidence.

[0099] S3014, Based on the variance, construct the fuzzy preference relation matrix corresponding to each preset criterion.

[0100] For example, each element in the fuzzy preference relation matrix corresponding to the difference criterion can be represented by the following formula:

[0101]

[0102] in, This represents the maximum variance among all evidence bodies under this difference criterion. This represents the minimum variance among all evidence bodies under this difference criterion. This represents the variance corresponding to the body of evidence mi under this difference criterion. This indicates that under the difference criterion, the body of evidence... The corresponding variance.

[0103] S3015 converts the fuzzy preference relation matrix corresponding to each preset criterion into a consistency matrix, and calculates the scheme layer weights based on the consistency matrix.

[0104] For example, in one possible implementation, the fuzzy preference relation matrix corresponding to each preset criterion is subjected to row summation and normalization transformation to obtain the consistency matrix corresponding to each preset criterion.

[0105] For example, each element in the consistency matrix can be represented as:

[0106]

[0107] For example, one possible implementation of calculating the scheme layer weights based on the consistency matrix is ​​as follows: based on the consistency matrix, calculate the priority index of each evidence body under each preset criterion, and then calculate the average of the priority indices under all preset criterions to obtain the scheme layer weights.

[0108] For example, the priority index of each piece of evidence under each preset criterion can be expressed by the following formula:

[0109]

[0110] in, Indicating evidence Superior to the body of evidence Priority indicators.

[0111] For example, body of evidence The priority index in the entire set of evidence can be expressed by the following formula:

[0112]

[0113] For example, body of evidence The scheme layer weights can be expressed as:

[0114]

[0115] S302, combine the criterion layer weights and scheme layer weights to obtain the evidence weights of each evidence body in the initial multi-source time series dataset.

[0116] For example, body of evidence The weight of evidence can be expressed by the following formula:

[0117]

[0118] in, Represents the weights of the criterion layer. Indicating evidence The weight of evidence.

[0119] S303, the evidence weight of each piece of evidence and the basic probability allocation of each piece of evidence under the preset identification framework are weighted and summed according to each proposition in the preset identification framework to obtain the weighted evidence of each proposition.

[0120] The preset identification framework is a set of transformer fault and fault-free states.

[0121] For example, the preset recognition framework can be represented as:

[0122]

[0123] Accordingly, each proposition in the preset identification framework can be a partial discharge fault, a vibration fault, or no fault.

[0124] For example, the basic probability allocation of each piece of evidence under the preset identification framework can be calculated based on the time-series data corresponding to each piece of evidence.

[0125] Understandably, in DS evidence theory, the basic probability allocation of each piece of evidence within a pre-defined identification framework is used to represent the degree to which the evidence supports the propositions within the pre-defined identification framework.

[0126] For example, the weighted evidence for each proposition in the predefined identification framework can be represented by the following formula:

[0127]

[0128] in, Indicating evidence For propositions in the pre-defined recognition framework The basic allocation probability.

[0129] For example, the weighted evidence for each proposition in the pre-defined identification framework satisfies the normalization requirement.

[0130] S304, based on preset evidence synthesis rules, integrates weighted evidence from each proposition to obtain the overall trust level of each proposition.

[0131] For example, the preset evidence synthesis rule can be the DS evidence theory synthesis rule (including the conflict factor K judgment).

[0132] For example, the conflict factor K is used to reflect the degree of conflict between two pieces of evidence. The larger K is, the higher the degree of conflict between the two pieces of evidence.

[0133] For example, the conflict factor K can be expressed by the following formula:

[0134]

[0135] For example, the overall confidence level of each proposition can be expressed by the following formula:

[0136]

[0137] S305, the comprehensive trust level of each proposition is concatenated into a fixed-dimensional feature vector according to the arrangement order of each proposition in the preset recognition framework, and evidence-weighted fusion data is generated.

[0138] For example, evidence-weighted fusion data can be represented as: {W (partial discharge fault), W (vibration fault), W (no fault)}.

[0139] In this embodiment, based on an improved analytic hierarchy process (AHP), the criterion-level weights and scheme-level weights corresponding to the initial multi-source time-series dataset are determined. These weights are then combined to obtain the evidence weights of each piece of evidence in the initial multi-source time-series dataset. Furthermore, the evidence weights of each piece of evidence and their basic probability allocation within a preset identification framework are weighted and summed according to each proposition within the preset framework to obtain weighted evidence for each proposition. Based on preset evidence synthesis rules, the weighted evidence for each proposition is fused to obtain the overall trust level of each proposition. The system concatenates each proposition into a fixed-dimensional feature vector according to its arrangement in the preset recognition framework, generating evidence-weighted fusion data. This eliminates the need for complex additional signal preprocessing, simplifies the multi-source data fusion process, and improves the fusion efficiency of multi-source data. At the same time, it solves the problem of poor accuracy in traditional methods where evidence weights are fixed or based on subjective experience. This gives higher priority to highly volatile and highly correlated data after weighted evidence fusion. Furthermore, the evidence-weighted fusion data is input into a deep belief network model for fault diagnosis, improving the accuracy of fault feature extraction, achieving high-precision fault classification, and improving the accuracy of fault diagnosis and classification results.

[0140] Optionally, the following is combined with Figure 4 The training method of the deep belief network model provided in the embodiments of this application will be described in detail.

[0141] Figure 4 This is a schematic diagram illustrating the training method of the deep belief network model provided in the embodiments of this application. Figure 4 As shown, a specific implementation of the training method for this deep belief network model may include the following steps:

[0142] S401, Obtain sample set data, which includes training set data and test set data.

[0143] For example, the sample set data may include multi-source time-series data of the target transformer under different fault modes and during normal operation (such as partial discharge time-series data, vibration time-series data, and oil chromatography time-series data).

[0144] For example, the number of samples in the sample set data can be 260, and the number of samples corresponding to different fault types and normal states in the sample set data can be the same.

[0145] For example, dividing the sample set data according to a certain ratio can yield training set data and test set data. For instance, a 7:3 ratio would allocate 70% of the sample set data to the training set and 30% to the test set data.

[0146] S402, perform weighted evidence fusion processing on the sample set data to generate weighted fused evidence sample data corresponding to the sample set data.

[0147] The evidence-weighted fusion sample data includes evidence-weighted fusion training set data and evidence-weighted fusion test set data.

[0148] In this step, the sample set data undergoes weighted evidence fusion processing to generate the corresponding weighted fused evidence sample data. The specific implementation method is similar to that described above and will not be repeated here.

[0149] One possible implementation may include normalizing the sample set data before performing weighted evidence fusion processing on the sample set data to eliminate the problem of inconsistent orders of magnitude in the sample set data.

[0150] S403, input the evidence-weighted fusion training set data into the deep belief network model to be trained to train the model parameters and obtain the pre-trained deep belief network model.

[0151] For example, the deep belief network model to be trained can be a DBN model containing multiple hidden layers. The number of input layer nodes, output layer nodes, model depth, and the number of nodes in each hidden layer of the DBN model to be trained are determined based on the dimension of the fault samples and the number of fault types.

[0152] For example, before inputting the evidence-weighted fusion training set data into the deep belief network model to be trained for model parameter training, it also includes initializing the model parameters of the DBN model to be trained (such as momentum, learning rate, initial error, number of iterations, etc.).

[0153] Optionally, this step may include the following steps:

[0154] S4031: Input the evidence-weighted fusion training set data into the deep belief network model to be trained, and perform unsupervised training and reconstruction optimization on the evidence-weighted fusion training set data layer by layer based on RBM to extract high-level abstract features.

[0155] Figure 5 This is a schematic diagram illustrating the process of layer-by-layer unsupervised training and reconstruction optimization based on RBM, as provided in an embodiment of this application. Figure 5 As shown, from the first layer of the deep belief network model to be trained, RMB (corresponding to) Figure 5 The feature pre-training begins with the data input layer in the deep belief network model. This involves inputting the evidence-weighted fusion training set data into the first visible layer of the RBM (Rich Base Model) to be trained, and performing feature pre-training in a bottom-up order, first training the bottom layer RBM, and then training the upper layers RBM in sequence.

[0156] like Figure 5 As shown, iterative training is performed for each layer of the RBM. The RBM performs layer-by-layer forward unsupervised training and internal reconstruction to optimize parameters on the input evidence-weighted fusion training set data. When the current iteration number corresponding to the current layer is greater than or equal to the preset maximum iteration number, the training of the current layer is completed. The hidden layer output of the previous layer of the RBM is used as the visible layer input of the next layer of the RBM, and the above feature pre-training is continued until the current training layer number is greater than or equal to the preset total number of layers, at which point the high-level abstract features are output.

[0157] S4032 inputs high-level abstract features into the attention mechanism module for feature enhancement, resulting in globally enhanced features.

[0158] For example, the attention mechanism module includes a global pooling unit and a fully connected network unit.

[0159] For example, one possible implementation of this step is to input high-level abstract features into the attention mechanism module for dynamic weighted feature enhancement, and dynamically learn the weights of different channels through global pooling and fully connected networks to enhance features important to the classification task, suppress redundant or noisy features, and obtain globally enhanced features.

[0160] Understandably, by employing an attention mechanism module to enhance the high-level abstract feature input, the model's sensitivity to key features can be improved, thereby enhancing the model's ability to distinguish fault types.

[0161] S4033 inputs global enhancement features into the top-level classifier, performs reverse global fine-tuning based on the gradient algorithm, optimizes all model parameters of the deep belief network model to be trained, and obtains the pre-trained deep belief network model.

[0162] For example, the top-level classifier includes a fully connected classification unit used to calculate the cross-entropy loss.

[0163] For example, in one possible implementation, global enhancement features are input into the top-level classifier, cross-entropy loss is backpropagated based on the gradient algorithm and global fine-tuning is performed to continuously optimize all model parameters of the deep belief network model to be trained until the cross-entropy loss calculated by the fully connected classification unit converges, thus obtaining the pre-trained deep belief network model.

[0164] S404. Input the evidence-weighted fusion test set data into the pre-trained deep belief network model to verify the model parameters and obtain the deep belief network model.

[0165] For example, in one possible implementation, evidence-weighted fusion test set data is input into a pre-trained deep belief network model for fault prediction, resulting in fault prediction results output by the pre-trained deep belief network model. A confusion matrix is ​​constructed using the predicted fault category in the fault prediction results as columns and the actual fault categories corresponding to the evidence-weighted fusion test set data as rows, with each element representing the number of samples that are both true to the category in that row and predicted to be the category in that column. The fault classification accuracy is then calculated based on this confusion matrix to evaluate the fault classification accuracy of the pre-trained deep belief network model. Random samples from the evidence-weighted fusion test set data are taken, and the fault prediction results output by the pre-trained deep belief network model are reconstructed in reverse to form a visualization consistent with the input data structure to evaluate the reconstruction accuracy. When the fault classification accuracy and reconstruction accuracy meet the application requirements, the pre-trained deep belief network model is determined as a deep belief network model.

[0166] Figure 6 This is a schematic diagram of the structure of the confusion matrix provided in an embodiment of this application. Figure 6 As shown, in this confusion matrix, the predicted fault category corresponding to the fault prediction result is the column, and the true fault category corresponding to the evidence-weighted fusion test set data is the row. Each element represents the number of samples that are both true to the category in that row and predicted to be the category in that column. Figure 6 As can be seen, the recognition accuracy for medium and low temperature overheating, high temperature overheating, low energy discharge, high energy discharge and normal state is 100%, the recognition accuracy for partial discharge is approximately 93%, and the overall fault classification accuracy of the pre-trained deep belief network model can reach 98.72%.

[0167] based on Figure 6 As can be seen from the confusion matrix shown, the deep belief network model provided in this application embodiment has high fault classification accuracy and strong fault diagnosis capability when performing different types of fault diagnosis.

[0168] Figure 7 This is a schematic diagram of the model loss curve and accuracy curve of the deep belief network model provided in the embodiments of this application. Figure 7 As shown, Figure 7 In Figure 7a, the model loss curve is shown, where the horizontal axis represents the number of training epochs and the vertical axis represents the loss. Curve 71 represents the validation loss curve, and curve 72 represents the training loss curve. Figure 7 In Figure 7b, the accuracy curve is represented by the horizontal axis representing the number of training rounds and the vertical axis representing the accuracy. Curve 73 represents the validation accuracy curve and curve 74 represents the training accuracy curve.

[0169] from Figure 7As can be seen from the results, the deep belief network model provided in this application embodiment has good model convergence performance, with a final verification loss of 0.0694, indicating good model generalization.

[0170] In this embodiment, sample set data including training set data and test set data is obtained, and weighted evidence fusion processing is performed on the sample set data to generate evidence-weighted fused sample data corresponding to the sample set data. The evidence-weighted fused training set data is then input into the deep belief network model to be trained to train the model parameters, resulting in a pre-trained deep belief network model. Then, the evidence-weighted fused test set data is input into the pre-trained deep belief network model to verify the model parameters, resulting in a deep belief network model. This improves the fault diagnosis accuracy and generalization ability of the trained deep belief network model.

[0171] Optionally, the fault diagnosis method provided in this application embodiment further includes, before performing weighted evidence fusion processing on the initial multi-source time series dataset, normalizing the initial multi-source time series dataset to obtain a normalized initial multi-source time series dataset.

[0172] The specific implementation method is similar to that described above, and will not be repeated here.

[0173] In summary, the fault diagnosis method provided in this application has the following beneficial effects:

[0174] 1) By performing weighted evidence fusion processing on the initial multi-source time series dataset, weighted evidence fusion data corresponding to the initial multi-source time series dataset is generated. This eliminates the need for complex additional signal preprocessing, simplifies the multi-source data fusion process, and improves the fusion efficiency of multi-source data. At the same time, it solves the problem of poor accuracy in traditional methods where evidence weights are fixed or based on subjective experience. This gives higher priority to highly volatile and highly correlated data after weighted evidence fusion. Furthermore, the weighted evidence fusion data is input into a deep belief network model for fault diagnosis, improving the accuracy of fault feature extraction, achieving high-precision fault classification, and improving the accuracy of fault diagnosis classification results.

[0175] 2) Compared with related technologies, which require signal preprocessing of the original fault data when performing fault diagnosis, the embodiments of this application do not rely on other signal processing technologies, do not have complex signal processing steps, and do not have strict periodic requirements for signal acquisition when performing fault diagnosis on the initial multi-source time series dataset based on the target transformer, thus improving the efficiency and practicality of fault diagnosis.

[0176] 3) Compared with related technologies that perform fault diagnosis based on a single data type, the embodiments of this application utilize the initial multi-source time series dataset of the target transformer for fault diagnosis, realize comprehensive analysis and decision-making of multi-dimensional parameters, improve the accuracy of fault diagnosis results, and can simultaneously diagnose multiple different fault types, thereby improving fault diagnosis efficiency.

[0177] 4) The embodiments of this application, through the powerful automatic feature extraction and nonlinear data processing capabilities of the deep belief network model, can effectively characterize the relationship between transformer parameters and system health status, diagnose various transformer faults, and ensure the stable and efficient operation of the transformer.

[0178] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0179] Figure 8 This is a schematic diagram of the fault diagnosis device provided in an embodiment of this application. Figure 8 As shown, the fault diagnosis device 80 includes an acquisition module 810, a weighted fusion module 820, and a fault diagnosis module 830.

[0180] Among them, the acquisition module 810 is used to acquire the initial multi-source time series dataset of the target transformer;

[0181] The weighted fusion module 820 is used to perform weighted evidence fusion processing on the initial multi-source time series dataset to generate weighted evidence fusion data corresponding to the initial multi-source time series dataset.

[0182] The fault diagnosis module 830 is used to input evidence-weighted fusion data into the deep belief network model for fault diagnosis and obtain the fault diagnosis results output by the deep belief network model.

[0183] In one possible implementation, the weighted fusion module 820 is specifically used for: determining the criterion layer weights and scheme layer weights corresponding to the initial multi-source time series dataset based on the improved analytic hierarchy process and the initial multi-source time series dataset; combining the criterion layer weights and scheme layer weights to obtain the evidence weights of each evidence body in the initial multi-source time series dataset; weighting and summing the evidence weights of each evidence body and the basic probability allocation of each evidence body under the preset identification framework according to each proposition in the preset identification framework to obtain the weighted evidence of each proposition, where the preset identification framework is a set of transformer fault and fault-free states; fusing the weighted evidence of each proposition based on the preset evidence synthesis rules to obtain the comprehensive trust level of each proposition; and concatenating the comprehensive trust levels of each proposition into a fixed-dimensional feature vector according to the order of each proposition in the preset identification framework to generate evidence weighted fusion data.

[0184] In one possible implementation, the weighted fusion module 820 is further configured to: determine the quantitative values ​​of each piece of evidence in the initial multi-source time-series dataset under each preset criterion based on the quantitative values ​​of each piece of evidence under each preset criterion; calculate the covariance matrix between each preset criterion and the variance between each piece of evidence under each preset criterion based on the quantitative values ​​of each piece of evidence under each preset criterion; convert the covariance matrix into a relative covariance matrix and calculate the criterion layer weights based on the relative covariance matrix; construct the fuzzy preference relation matrix corresponding to each preset criterion based on the variance between each piece of evidence under each preset criterion; convert the fuzzy preference relation matrix corresponding to each preset criterion into a consistency matrix and calculate the scheme layer weights based on the consistency matrix.

[0185] In one possible implementation, the deep belief network model is trained as follows: a sample set of data is acquired, including training and test data; weighted evidence fusion is performed on the sample set data to generate weighted evidence fusion sample data, which includes weighted evidence fusion training and test data; the weighted evidence fusion training data is input into the deep belief network model to be trained to train the model parameters, resulting in a pre-trained deep belief network model; and the weighted evidence fusion test data is input into the pre-trained deep belief network model to validate the model parameters, resulting in the final deep belief network model.

[0186] In one possible implementation, the evidence-weighted fusion training set data is input into the deep belief network model to be trained to train the model parameters, resulting in a pre-trained deep belief network model. This includes: inputting the evidence-weighted fusion training set data into the deep belief network model to be trained; performing layer-by-layer unsupervised training and reconstruction optimization on the evidence-weighted fusion training set data based on RBM to extract high-level abstract features; inputting the high-level abstract features into the attention mechanism module for feature enhancement to obtain globally enhanced features; and inputting the globally enhanced features into the top-level classifier, performing reverse global fine-tuning based on the gradient algorithm to optimize all model parameters of the deep belief network model to be trained, thus obtaining the pre-trained deep belief network model.

[0187] In one possible implementation, before performing weighted evidence fusion processing on the initial multi-source time series dataset, the fault diagnosis device further includes a processing module (not shown) for: normalizing the initial multi-source time series dataset to obtain a normalized initial multi-source time series dataset.

[0188] The fault diagnosis device provided in this application embodiment can be used to execute the method steps of the above method embodiment. The specific implementation and technical effects are similar, and will not be repeated here.

[0189] Figure 9This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 9 As shown, the electronic device 90 provided in this embodiment includes at least one processor 901 and a memory 902. Optionally, the electronic device 90 further includes a communication component 903. The processor 901, memory 902, and communication component 903 are connected via a bus 904.

[0190] In a specific implementation, at least one processor 901 executes computer execution instructions stored in memory 902, causing at least one processor 901 to perform the above-described method.

[0191] The specific implementation process of processor 901 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0192] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0193] The memory may include random access memory (RAM) and non-volatile memory (NVM), such as at least one disk storage device.

[0194] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0195] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0196] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0197] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0198] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0199] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0200] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0201] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0202] If a function is implemented as 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 technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0203] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0204] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A fault diagnosis method, characterized in that, include: Obtain the initial multi-source time-series dataset of the target transformer; The initial multi-source time series dataset is subjected to weighted evidence fusion processing to generate weighted evidence fusion data corresponding to the initial multi-source time series dataset. The evidence-weighted fusion data is input into a deep belief network model for fault diagnosis, and the fault diagnosis results output by the deep belief network model are obtained.

2. The fault diagnosis method according to claim 1, characterized in that, The step of performing weighted evidence fusion processing on the initial multi-source time-series dataset to generate weighted evidence fusion data corresponding to the initial multi-source time-series dataset includes: Based on the improved analytic hierarchy process, the criterion layer weights and scheme layer weights corresponding to the initial multi-source time series dataset are determined according to the initial multi-source time series dataset. The criterion layer weights and the scheme layer weights are combined to obtain the evidence weights of each evidence body in the initial multi-source time series dataset. The evidence weights of each piece of evidence and the basic probability allocation of each piece of evidence under the preset identification framework are weighted and summed according to each proposition in the preset identification framework to obtain the weighted evidence of each proposition. The preset identification framework is a set of transformer fault and fault-free states. Based on the preset evidence synthesis rules, the weighted evidence of each proposition is integrated to obtain the comprehensive trust level of each proposition; The overall trust level of each proposition is concatenated into a fixed-dimensional feature vector according to the arrangement order of each proposition in the preset recognition framework, thereby generating the evidence weighted fusion data.

3. The fault diagnosis method according to claim 2, characterized in that, The method based on the improved analytic hierarchy process (AHP) determines the criterion layer weights and scheme layer weights corresponding to the initial multi-source time series dataset, including: Based on the initial multi-source time series dataset, determine the criterion quantitative value of each piece of evidence in the initial multi-source time series dataset under each preset criterion; Based on the quantitative values ​​of each piece of evidence under each preset criterion, calculate the covariance matrix between each preset criterion and the variance between each piece of evidence under each preset criterion. The covariance matrix is ​​converted into a relative covariance matrix, and the criterion layer weights are calculated based on the relative covariance matrix. Based on the variance between each piece of evidence under each preset criterion, construct the fuzzy preference relation matrix corresponding to each preset criterion; The fuzzy preference relation matrix corresponding to each preset criterion is converted into a consistency matrix, and the scheme layer weight is calculated based on the consistency matrix.

4. The fault diagnosis method according to any one of claims 1 to 3, characterized in that, The deep belief network model was trained in the following way: Obtain sample set data, which includes training set data and test set data; The sample set data is subjected to weighted evidence fusion processing to generate evidence weighted fusion sample data corresponding to the sample set data. The evidence weighted fusion sample data includes evidence weighted fusion training set data and evidence weighted fusion test set data. The evidence-weighted fusion training set data is input into the deep belief network model to be trained to train the model parameters and obtain the pre-trained deep belief network model. The evidence-weighted fusion test set data is input into the pre-trained deep belief network model to verify the model parameters, thereby obtaining the deep belief network model.

5. The fault diagnosis method according to claim 4, characterized in that, The step of inputting the evidence-weighted fusion training set data into the deep belief network model to be trained for model parameter training, to obtain a pre-trained deep belief network model, includes: The evidence-weighted fusion training set data is input into the deep belief network model to be trained. The evidence-weighted fusion training set data is then subjected to unsupervised training and reconstruction optimization layer by layer based on the Restricted Boltzmann Machine (RBM) to extract high-level abstract features. The high-level abstract features are input into the attention mechanism module for feature enhancement to obtain globally enhanced features; The global enhancement features are input into the top-level classifier, and reverse global fine-tuning is performed based on the gradient algorithm to optimize all model parameters of the deep belief network model to be trained, thereby obtaining the pre-trained deep belief network model.

6. The fault diagnosis method according to any one of claims 1 to 3, characterized in that, Before performing weighted evidence fusion processing on the initial multi-source time-series dataset, the following steps are also included: The initial multi-source time series dataset is normalized to obtain a normalized initial multi-source time series dataset.

7. A fault diagnosis device, characterized in that, include: The acquisition module is used to acquire the initial multi-source time-series dataset of the target transformer; The weighted fusion module is used to perform weighted evidence fusion processing on the initial multi-source time series dataset to generate weighted evidence fusion data corresponding to the initial multi-source time series dataset. The fault diagnosis module is used to input the evidence-weighted fusion data into the deep belief network model for fault diagnosis, and obtain the fault diagnosis results output by the deep belief network model.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the fault diagnosis method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the fault diagnosis method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, include: A computer program, which, when executed by a processor, implements the fault diagnosis method as described in any one of claims 1 to 6.

Images