Transformer oil chromatography online fault monitoring method, device and equipment based on fuzzy neural network algorithm and medium

Abstract

The invention discloses a transformer oil chromatography online fault monitoring method, device and equipment based on a fuzzy neural network algorithm and a medium. The method comprises the following steps: constructing a monitoring point cluster; performing wavelet decomposition on the concentration time sequence data of each monitoring point in the monitoring point cluster to obtain cleaning concentration time sequence data of each monitoring point; judging whether the monitoring points in the monitoring point cluster are outlier or not; constructing an enhancement vector of the monitoring point according to the cleaning concentration time sequence data of the monitoring point and the time domain feature, the frequency domain feature and the time-frequency domain feature of the cleaning concentration time sequence data; and inputting the enhancement vector of each monitoring point in the monitoring point cluster into a preset fuzzy neural network fault diagnosis model to obtain the fault type, the fault level and the fault probability confidence coefficient of the transformer oil chromatography. The invention belongs to the field of transformer detection. The fault diagnosis efficiency of the transformer can be improved.

Description

Technical Field

[0001] The present invention relates to the field of transformer detection, and in particular to a transformer oil chromatogram online fault monitoring method, device, equipment and medium based on a fuzzy neural network algorithm. Background Art

[0002] A transformer is an electrical device used to change the voltage of alternating current (AC). Composed primarily of an iron core and windings, it transmits and converts electrical energy through the principle of electromagnetic induction. It plays a key role in power systems, performing voltage adjustments, circuit isolation, and power regulation. It is widely used in power transmission and distribution systems to improve energy transmission efficiency and ensure the safe operation of electrical equipment.

[0003] Traditional transformer fault diagnosis methods often rely on a single criterion or simple threshold judgment, lacking a comprehensive analysis of multiple factors. Simple threshold judgment methods fail to fully consider factors such as gas production rate and correlations between gas components, making them prone to misdiagnosis or omissions. Furthermore, when faced with complex faults, a single criterion cannot accurately describe the fault's full picture, making it difficult to accurately determine the fault type, severity, and development trend. This makes it impossible to provide comprehensive and accurate fault information to maintenance personnel, hindering the timeliness and effectiveness of fault handling. Summary of the Invention

[0004] The present invention solves the technical problem of low transformer fault diagnosis accuracy in the prior art by providing a transformer oil chromatogram online fault monitoring method, device, equipment and medium based on a fuzzy neural network algorithm, and achieves the technical effect of improving the transformer fault diagnosis accuracy.

[0005] In a first aspect, the present invention provides a transformer oil chromatogram online fault monitoring method based on a fuzzy neural network algorithm, comprising: Acquire the concentration data of multiple gases at each monitoring point in the transformer oil chromatogram, and build a monitoring point cluster centered on the monitoring point when the gas concentration data is greater than the preset concentration threshold corresponding to the gas; Perform wavelet decomposition on the concentration time series data of each monitoring point in the monitoring point cluster to obtain the cleaning concentration time series data of each monitoring point. The time length of each cleaning concentration time series data is the same. Obtain the template concentration time series data of the monitoring point cluster, and determine whether the monitoring point in the monitoring point cluster is out of the cluster based on the template concentration time series data and the cleaning concentration time series data, and repair the outlier monitoring point; Extract the time domain features, frequency domain features and time-frequency domain features of the cleaning concentration time series data of each monitoring point in the monitoring point cluster, and construct the enhancement vector of the monitoring point based on the cleaning concentration time series data of the monitoring point and the time domain features, frequency domain features and time-frequency domain features of the cleaning concentration time series data; The enhanced vector of each monitoring point in the monitoring point cluster is input into the preset fuzzy neural network fault diagnosis model to obtain the fault type, fault level and fault probability confidence of the transformer oil chromatogram.

[0006] Furthermore, the concentration time series data of each monitoring point in the monitoring point cluster is subjected to wavelet decomposition to obtain the cleaning concentration time series data of each monitoring point, including: Decompose the concentration time series data into high-frequency detail components and low-frequency approximate components, including:

[0007] in, for The concentration time series data at each moment, for Moment The low-frequency approximate component of the layer, for Moment The high-frequency detail components of the layer; Denoise each high-frequency detail component, including:

[0008] in, for Moment The high-frequency detail component after layer denoising, is the preset vector; Each low-frequency approximation component Input the preset LSTM model to approximate the low-frequency components Error correction is performed, where the input formula of the preset LSTM model is:

[0009] in, for The current input gate state at time t, For the The weight matrix of the layer, is the hidden state at the previous moment, For the The bias of the layer, is the Sigmoid activation function; The corrected low-frequency approximate component It is spliced ​​with the corresponding denoised high-frequency detail components to obtain the cleaning concentration time series data.

[0010] Furthermore, the template concentration time series data of the monitoring point cluster is obtained, including: The historical concentration data of multiple gases at each monitoring point in the monitoring point cluster are intercepted to obtain several historical concentration time series data of each monitoring point, where the duration of the historical concentration time series data is the same as the duration of the cleaning concentration time series data; The mean of the historical concentration time series data of the monitoring points in the monitoring point cluster is used as the template concentration time series data of the monitoring point cluster.

[0011] Furthermore, determining whether a monitoring point in the monitoring point cluster is outliers includes: Determine the cumulative distance between the cleaning concentration time series data of the monitoring point and the template concentration time series data of the monitoring point cluster, including:

[0012] in, Cleaning concentration time series data Time series data with template concentration The cumulative distance between Cleaning concentration time series data Middle data and template concentration time series data The data, is an alignment path, is the set of all alignment paths;

[0013] in, For monitoring points The local reachable density of For monitoring points The neighborhood size between the nearest monitoring points, For monitoring points The nearest monitoring point, From the monitoring point The accessible distance to the monitoring point;

[0014] in, For monitoring points The local outlier factor, For monitoring points The local reachable density of If the local outlier factor of a monitoring point is greater than a preset value, the monitoring point is considered an outlier monitoring point.

[0015] Furthermore, repair outlier monitoring points, including: Eliminate abnormal data from the cleaning concentration time series data of outlier monitoring points; The outlier monitoring points are filled using cubic spline interpolation.

[0016] Furthermore, the time domain features, frequency domain features, and time-frequency domain features of the cleaning concentration time series data of each monitoring point in the monitoring point cluster are extracted, including:

[0017] in, For the The gas concentration at each monitoring point, is the length of time for cleaning concentration time series data, is the central trend of gas concentration;

[0018] in, is the variance;

[0019] in, is the wavelet packet energy entropy, For the frequency bands, is the number of frequency bands.

[0020] Furthermore, the enhanced vector of the monitoring point is constructed based on the cleaning concentration time series data of the monitoring point, the time domain features, the frequency domain features and the time-frequency domain features of the cleaning concentration time series data, including: The cleaning concentration time series data of the monitoring point, the time domain features, the frequency domain features and the time-frequency domain features of the cleaning concentration time series data are spliced ​​to obtain a high-dimensional feature vector; The principal components are obtained by decomposing the high-dimensional eigenvectors through the eigenvalues ​​of the preset covariance matrix; The preset sorted features in the principal component are fused to obtain the enhanced vector of the monitoring point.

[0021] In a second aspect, the present invention provides a transformer oil chromatogram online fault monitoring device based on a fuzzy neural network algorithm, comprising: An acquisition module is used to obtain the concentration data of multiple gases at each monitoring point in the transformer oil chromatogram, and when the concentration data of a gas is greater than a preset concentration threshold corresponding to the gas, a monitoring point cluster is constructed with the monitoring point as the center; A decomposition module is used to perform wavelet decomposition on the concentration time series data of each monitoring point in the monitoring point cluster to obtain the cleaning concentration time series data of each monitoring point, and the time length of each cleaning concentration time series data is the same; The repair module is used to obtain the template concentration time series data of the monitoring point cluster, and determine whether the monitoring point in the monitoring point cluster is out of the cluster based on the template concentration time series data and the cleaning concentration time series data, and repair the outlier monitoring point; The enhancement module is used to extract the time domain features, frequency domain features and time-frequency domain features of the cleaning concentration time series data of each monitoring point in the monitoring point cluster, and construct the enhancement vector of the monitoring point based on the cleaning concentration time series data of the monitoring point and the time domain features, frequency domain features and time-frequency domain features of the cleaning concentration time series data; The output module is used to input the enhanced vector of each monitoring point in the monitoring point cluster into the preset fuzzy neural network fault diagnosis model to obtain the fault type, fault level and fault probability confidence of the transformer oil chromatogram.

[0022] In a third aspect, the present invention provides an electronic device, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to execute to implement the transformer oil chromatogram online fault monitoring method based on the fuzzy neural network algorithm as provided in the first aspect.

[0023] In a fourth aspect, the present invention provides a non-temporary computer-readable storage medium. When the instructions in the non-temporary computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to implement the transformer oil chromatography online fault monitoring method based on the fuzzy neural network algorithm provided in the first aspect.

[0024] One or more technical solutions provided in the present invention have at least the following technical effects or advantages: This invention effectively solves the problem of processing non-stationary fluctuation data through multi-dimensional data preprocessing, improves data accuracy and integrity, and provides a reliable data foundation for subsequent fault diagnosis. Fuzzy neural networks combine multiple criteria for hierarchical fault diagnosis, fully considering the impact of various factors on fault diagnosis and improving the accuracy of fault diagnosis.

[0025] The technical solution provided by the present invention can adapt to various complex working conditions during the operation of the transformer. Whether it is a normal operating state or a sudden load change or a sudden change in ambient temperature, it can stably and accurately monitor and diagnose. In particular, in terms of processing non-stationary fluctuating data, the adaptability is improved by cleaning the data processing method, effectively improving the reliability and stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0027] Figure 1A schematic flow chart of the transformer oil chromatogram online fault monitoring method based on the fuzzy neural network algorithm provided by the present invention; Figure 2 This is a structural schematic diagram of the transformer oil chromatogram online fault monitoring device based on the fuzzy neural network algorithm provided by the present invention. DETAILED DESCRIPTION

[0028] The embodiment of the present invention solves the technical problem of low transformer fault diagnosis accuracy in the prior art by providing a transformer oil chromatogram online fault monitoring method based on a fuzzy neural network algorithm.

[0029] The technical solution of the present invention is to solve the above technical problems, and the overall idea is as follows: A transformer oil chromatogram online fault monitoring method based on a fuzzy neural network algorithm comprises the following steps: obtaining concentration data of multiple gases at each monitoring point in the transformer oil chromatogram, and constructing a monitoring point cluster with the monitoring point as the center when the gas concentration data is greater than a preset concentration threshold corresponding to the gas; performing wavelet decomposition on the concentration time series data of each monitoring point in the monitoring point cluster to obtain cleaning concentration time series data of each monitoring point, wherein the time length of each cleaning concentration time series data is the same; obtaining template concentration time series data of the monitoring point cluster, and judging whether a monitoring point in the monitoring point cluster is outliers based on the template concentration time series data and the cleaning concentration time series data, and repairing the outliers; extracting time domain features, frequency domain features, and time-frequency domain features of the cleaning concentration time series data of each monitoring point in the monitoring point cluster, and constructing an enhancement vector of the monitoring point based on the cleaning concentration time series data of the monitoring point and the time domain features, frequency domain features, and time-frequency domain features of the cleaning concentration time series data; and inputting the enhancement vector of each monitoring point in the monitoring point cluster into a preset fuzzy neural network fault diagnosis model to obtain the fault type, fault level, and fault probability confidence of the transformer oil chromatogram.

[0030] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0031] First, the term "and / or" as used herein simply describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A alone, A and B together, or B alone. Furthermore, the character " / " in this document generally indicates an "or" relationship between the associated objects.

[0032] The present invention provides Figure 1 The transformer oil chromatogram online fault monitoring method based on the fuzzy neural network algorithm shown includes steps S11-S15: Step S11 , obtaining concentration data of multiple gases at each monitoring point in the transformer oil chromatogram, and building a monitoring point cluster with the monitoring point as the center when the concentration data of a gas is greater than a preset concentration threshold corresponding to the gas.

[0033] Specifically: The oil-gas separator in the online monitoring system continuously collects oil samples from the transformer in real time. The collection process remains closed and stable to ensure that the oil samples represent the current operating status.

[0034] The oil sample first enters the degassing device, which effectively separates the trace gases dissolved in the oil by heating and reducing pressure. The types of gases may include hydrogen, carbon monoxide, methane, acetylene, ethylene, etc.

[0035] The separated gases then enter the detection module, where the composite chromatographic column separates the gases according to their retention times. The sensor then accurately detects the concentration of each gas, ultimately generating concentration data with a timestamp.

[0036] When the concentration of (any) gas at a certain moment exceeds 1.5 times the historical average (preset concentration threshold), it means that the concentration data at the current monitoring point has experienced abnormal fluctuations. The abnormal fluctuations may be caused by partial discharge, temperature anomalies, or sudden aggravation of insulation aging.

[0037] At this time, a monitoring point cluster can be constructed with the monitoring point as the center. The number of monitoring points in the monitoring point cluster can be determined according to the actual situation of the transformer. The reference number provided by the present invention is 6-24.

[0038] Step S12 , performing wavelet decomposition on the concentration time series data of each monitoring point in the monitoring point cluster to obtain cleaning concentration time series data of each monitoring point, wherein the time length of each cleaning concentration time series data is the same.

[0039] In order to process the data more sensitively and finely and ensure the accuracy and stability of subsequent analysis results, the concentration time series data of each monitoring point in the monitoring point cluster is subjected to wavelet decomposition to obtain the cleaning concentration time series data of each monitoring point, including: The concentration time series data can be first subjected to wavelet decomposition, and appropriate wavelet basis functions can be selected to perform multi-scale decomposition to decompose a concentration time series data into several high-frequency detail components and a low-frequency approximate component.

[0040] Decompose the concentration time series data into high-frequency detail components and low-frequency approximate components, including:

[0041] in, for The concentration time series data at each moment, for Moment The low-frequency approximate component of the layer, for Moment The high-frequency detail components of the layer; In order to eliminate the random noise signal in the high-frequency component, the wavelet hard threshold denoising method can be used to denoise each high-frequency detail component, including:

[0042] in, for Moment The high-frequency detail component after layer denoising, is the preset vector; Each low-frequency approximation component Input the preset LSTM model to approximate the low-frequency components Error correction is performed, where the input formula of the preset LSTM model is:

[0043] in, for The current input gate state at time t, For the The weight matrix of the layer, is the hidden state at the previous moment, For the The bias of the layer, is the Sigmoid activation function; The corrected low-frequency approximate component It is spliced ​​with the corresponding denoised high-frequency detail components to obtain the cleaning concentration time series data.

[0044] The final output cleaning concentration time series data splices the detailed features retained in the previous high-frequency part into the cleaning results and serves as an important supplement to the subsequent feature vector. Step S12 ensures that the key abnormal features are retained during the data processing process and the overall signal quality is improved, providing more accurate and complete input information for subsequent fault diagnosis.

[0045] Step S13: obtaining the template concentration time series data of the monitoring point cluster, and judging whether a monitoring point in the monitoring point cluster is out of the cluster based on the template concentration time series data and the cleaning concentration time series data, and repairing the outlier monitoring point.

[0046] Obtain template concentration time series data for the monitoring point cluster, including: The historical concentration data of multiple gases at each monitoring point in the monitoring point cluster are intercepted to obtain several historical concentration time series data for each monitoring point, where the duration of the historical concentration time series data is the same as the duration of the cleaning concentration time series data; the mean of the historical concentration time series data of the monitoring points in the monitoring point cluster is used as the template concentration time series data of the monitoring point cluster.

[0047] Historical concentration data refers to the concentration data of the monitoring point at the time of the fault. Historical concentration time series data is a part of the concentration data. The number of historical concentration time series data for each monitoring point can be determined according to the accuracy requirements. The reference number provided by the present invention is 12-24. The duration of each historical concentration time series data is the same as the duration of the cleaning concentration time series data. The average value of each gas concentration in the historical concentration time series data is calculated to obtain the template concentration time series data of the monitoring point cluster.

[0048] Determine whether a monitoring point in a monitoring point cluster is out of the group, including: Determine the cumulative distance between the cleaning concentration time series data of the monitoring point and the template concentration time series data of the monitoring point cluster, including:

[0049] in, Cleaning concentration time series data Time series data with template concentration The cumulative distance between Cleaning concentration time series data Middle data and template concentration time series data The data, is an alignment path, is the set of all alignment paths; DTW is used to measure the minimum cumulative distance between two time series after nonlinear alignment on the time axis.

[0050]

[0051] in, For monitoring points The local reachable density of For monitoring points The neighborhood size between the nearest monitoring points, For monitoring points The nearest monitoring point, From the monitoring point The accessible distance to the monitoring point;

[0052] in, For monitoring points The local outlier factor, For monitoring points The local reachable density of If the local outlier factor of a monitoring point is greater than a preset value, the monitoring point is considered an outlier monitoring point.

[0053] Repairing outlier monitoring points includes: removing abnormal data from the cleaning concentration time series data of the outlier monitoring points; filling the outlier monitoring points with cubic spline interpolation. Methods for determining abnormal data in monitoring points.

[0054] To ensure the continuity of time series and the integrity of features, the cubic spline interpolation method can be used to repair abnormal data in outliers. A set of cubic continuous functions can be constructed based on known points, and the positions of abnormal points can be filled with smooth curves to ensure that the repaired sequence maintains continuity in the first-order derivative and the second-order derivative, avoiding the introduction of mutation errors, thereby ensuring that subsequent models can continue to receive input data with consistent structure, enhancing the robustness and abnormal adaptability of the overall monitoring system.

[0055] Step S14: extract the time domain features, frequency domain features, and time-frequency domain features of the cleaning concentration time series data of each monitoring point in the monitoring point cluster, and construct an enhancement vector for the monitoring point using the cleaning concentration time series data of the monitoring point, the time domain features, frequency domain features, and time-frequency domain features of the cleaning concentration time series data.

[0056] Three types of features are extracted from the cleaning concentration time series data: time domain features, frequency domain features, and time-frequency domain features. Time domain features can include the mean and variance of each gas within a certain time window, indicating the central tendency and fluctuation of the data, respectively.

[0057] Specifically include:

[0058] in, For the The gas concentration at each monitoring point, is the length of time for cleaning concentration time series data, is the central trend of gas concentration;

[0059] in, is the variance;

[0060] in, is the wavelet packet energy entropy, For the frequency bands, is the number of frequency bands.

[0061] An enhanced vector of the monitoring point is constructed using the cleaning concentration time series data of the monitoring point, the time domain features, the frequency domain features and the time-frequency domain features of the cleaning concentration time series data, including: splicing the cleaning concentration time series data of the monitoring point, the time domain features, the frequency domain features and the time-frequency domain features of the cleaning concentration time series data to obtain a high-dimensional feature vector; decomposing the high-dimensional feature vector through the eigenvalues ​​of the preset covariance matrix to obtain the principal component; fusing the preset sorted features in the principal component to obtain the enhanced vector of the monitoring point.

[0062] The above three types of features can extract a total of 10 dimensions of auxiliary features, which are then spliced ​​together with the cleaned concentration time series data to form a high-dimensional feature vector. In order to reduce dimensional redundancy, avoid information duplication and improve computational efficiency, the principal component analysis method can be introduced to reduce the dimensionality of the fused features. The principal component direction is retained through the eigenvalue decomposition of the covariance matrix, and the top 8 features with the highest cumulative variance contribution rate are selected to form the final enhanced feature vector.

[0063] In addition, in order to unify the input space, the enhanced feature vector can be normalized, and all dimensional features can be linearly mapped to the range of 0 to 1. The normalization calculation is:

[0064] in is the eigenvalue, and The above processing ensures that each type of feature has equal expressive power in the model input, improves the convergence speed and discrimination accuracy of the fuzzy neural network, and provides stable and reliable high-quality feature support for fault diagnosis.

[0065] Step S15 , inputting the enhanced vector of each monitoring point in the monitoring point cluster into a preset fuzzy neural network fault diagnosis model to obtain the fault type, fault level and fault probability confidence of the transformer oil chromatogram.

[0066] Specifically: The normalized enhanced feature vector can be input into the constructed five-layer fuzzy neural network. The five-layer fuzzy neural network consists of an input layer, a fuzzification layer, a rule layer, a normalization layer, and an output layer. Combining fuzzy logic and backpropagation mechanism, it can achieve accurate identification of nonlinear fault modes. The input layer receives an 8-dimensional standardized feature vector. The fuzzification layer uses a Gaussian membership function to perform a fuzzy transformation on each input, mapping continuous values ​​into fuzzy linguistic variables to express the uncertainty of the input. The fuzzy membership function is in the form of:

[0067] in is the eigenvalue, c is the membership function center, is the membership function width.

[0068] The rule layer constructs the reasoning path based on the fuzzy rule set generated by expert knowledge and training samples. Each rule corresponds to a specific combination of membership patterns. The network dynamically adjusts the weight coefficient of each rule during the training process, so that it can adaptively enhance effective rules and suppress invalid rules when the samples change.

[0069] The normalization layer normalizes all rule activation values ​​to eliminate the superposition amplification effect of strong rules and improve model stability. The output layer uses an error backpropagation algorithm with an adaptive learning rate to iteratively update the network weights. The weights are corrected layer by layer according to the error gradient to minimize the difference between the output and the target. The weight update formula is:

[0070] in represents the weight adjustment value from layer i to layer j, η is the dynamically adjusted learning rate, is the partial derivative of the error function with respect to the weight.

[0071] The final network output consists of three components: the fault type, fault severity, and fault probability confidence level. The fault type is used to locate the root cause of the problem, the fault severity reflects the system risk level, and the fault probability confidence level assists operations and maintenance personnel in developing proactive intervention strategies, enabling timely early warning and accurate assessment of potential transformer anomalies.

[0072] For example: First, a real-time threshold judgment is made on the concentration of key gases. When the acetylene concentration exceeds 1 microliter per liter or the total hydrocarbon concentration exceeds 150 microliters per liter, a level 1 warning is immediately triggered, marking the current state as an early stage of abnormality.

[0073] At this time, the fault probability confidence level output by the fault diagnosis module is called, and the short-term change rate of the gas production rate and the fault coding result generated based on the three-ratio method are calculated at the same time, and a comprehensive score is performed according to the set weight.

[0074] The failure probability confidence level reflects the model's confidence in the current abnormality and accounts for 60% of the decision-making weight. The gas production rate growth rate indicates the growth rate of various combustible gases in recent cycles and accounts for 20% of the weight, reflecting the severity of the changing trend. The three-ratio coding classifies and determines the concentration ratios of gases such as methane, ethylene, and acetylene based on standard methods, accounting for the remaining 20% ​​of the weight. These three factors are used to calculate a comprehensive diagnostic score. When the score exceeds 0.7, it is determined to be a serious fault state and the early warning escalation process is initiated.

[0075] In summary, this invention effectively addresses the challenge of processing nonstationary, fluctuating data through multidimensional data preprocessing, improving data accuracy and completeness and providing a reliable data foundation for subsequent fault diagnosis. Fuzzy neural networks, combined with multiple criteria, perform hierarchical fault diagnosis, fully accounting for the impact of various factors on fault diagnosis and improving its accuracy.

[0076] In addition, the present invention shows other feasibility examples in the data cleaning, outlier detection and feature fusion stages respectively: In dynamic adaptive data cleaning, the wavelet hard threshold denoising algorithm first performs wavelet decomposition on each time series signal, breaking it down into high-frequency detail components and low-frequency approximation components in different frequency bands. For the high-frequency components, the energy distribution of all high-frequency detail coefficients is calculated, and the 90th percentile is determined as the dynamic denoising threshold.

[0077] The threshold represents the upper limit of the current signal noise energy. Only the signal part with an amplitude exceeding the threshold is retained, and other components with smaller amplitudes and close to the background noise are directly set to zero, thereby effectively retaining the edge characteristics of the sudden abnormal signal while removing systematic noise interference.

[0078] After high-frequency denoising, the smoothed low-frequency components are fused with the denoised high-frequency residual signal as LSTM input to enhance the model's memory of fine-grained perturbations. During the LSTM modeling process, the model structure is dynamically reconstructed based on the current sliding window length. When the window is shortened from the normal 24 monitoring points to 6 monitoring points during periods of severe fluctuations, it indicates that the current data exhibits a high-frequency, low-stability trend. To avoid overfitting and slow learning, the system reduces the number of hidden neurons in each layer of the LSTM network by 20% to reduce model complexity and improve generalization capabilities. At the same time, the learning rate is increased by 50%, allowing the model to more efficiently complete weight adjustments in the face of rapidly changing data environments, thereby adapting to the feature learning needs of sudden changes.

[0079] In the feature fusion stage: First, the data within 72 hours before the current time point are extracted and divided into 12 cycles, each cycle is 6 hours long. The concentration monitoring points in each cycle are arranged in chronological order to construct a set of continuous historical reference templates.

[0080] During the outlier detection process, the dynamic time warping method is used to compare the current time series segment with the 12 periodic templates one by one. Dynamic programming is used to find the optimal matching path on the time axis and calculate the DTW distance between the two sequences. This distance can flexibly align local time drift and nonlinear deformation, thereby effectively characterizing the actual similarity of time series fluctuations.

[0081] To avoid computational errors caused by over-alignment during the path search process, the Sakoe-Chiba band constraint is introduced to restrict the DTW search path to a certain bandwidth near the main diagonal, allowing only paths with a slope of 1. That is, each point can only be matched with a point in the same direction as the time axis.

[0082] The bandwidth is set to 10% of the current sliding window length. This setting ensures alignment freedom while effectively reducing distortion caused by abnormal matching paths and improving the local stability of the match. After calculating the DTW distance between the current data and each template, the average is calculated, and twice the average is set as the outlier determination threshold. If the DTW distance of the current data exceeds this threshold, indicating that the current time series fluctuation pattern has significantly deviated from the historical feature distribution, the system marks this point as a non-stationary outlier and further triggers the interpolation repair mechanism to maintain time series continuity and sample integrity.

[0083] The dynamic rule adjustment method of the neural network is: The improved particle swarm optimization algorithm is introduced to dynamically adjust the weight coefficients of fuzzy rules. The algorithm simulates the group intelligence behavior of particles in the search space. Each particle corresponds to a weight configuration of a fuzzy rule. The current position of the particle represents the current weight solution, the speed represents the update step size, and the individual optimal position and the optimal position of the group To avoid falling into local optimality, a linear reduction strategy is implemented for the particle inertia weight, so that it maintains a strong global search capability in the early stage and enhances the local convergence effect in the later stage. The inertia weight gradually decreases from the initial value of 0.9 to 0.4 in the 50th iteration. In each round of update, the particle weight coefficient is adjusted by the following formula:

[0084] in is the weight of the i-th fuzzy rule at the t-th iteration, represents the optimal weight configuration that the particle has searched historically, The global optimal configuration for the current population is achieved. α, the individual learning factor, is set to 0.5, and β, the group learning factor, is set to 0.3. This structure allows each weight value to retain its own optimization history while drawing on the group's exploration experience, thereby achieving dynamic weight adjustment. The final converged weight distribution reflects the importance of each fuzzy rule under different fault modes, effectively improving the network's response accuracy to fuzzy inputs and the credibility of diagnostic results.

[0085] The technical solution provided by the present invention can adapt to various complex working conditions during the operation of the transformer. Whether it is a normal operating state or a sudden load change or a sudden change in ambient temperature, it can stably and accurately monitor and diagnose. In particular, in terms of processing non-stationary fluctuating data, the adaptability is improved by cleaning the data processing method, effectively improving the reliability and stability of the system.

[0086] Based on the same inventive concept, the present invention provides Figure 2 The transformer oil chromatogram online fault monitoring device based on fuzzy neural network algorithm shown in the figure includes: An acquisition module 21 is used to obtain concentration data of multiple gases at each monitoring point in the transformer oil chromatogram, and to construct a monitoring point cluster centered on the monitoring point when the concentration data of a gas is greater than a preset concentration threshold corresponding to the gas; A decomposition module 22 is used to perform wavelet decomposition on the concentration time series data of each monitoring point in the monitoring point cluster to obtain the cleaning concentration time series data of each monitoring point, where the time length of each cleaning concentration time series data is the same; The repair module 23 is used to obtain the template concentration time series data of the monitoring point cluster, and determine whether the monitoring point in the monitoring point cluster is out of the cluster based on the template concentration time series data and the cleaning concentration time series data, and repair the outlier monitoring point; The enhancement module 24 is used to extract the time domain features, frequency domain features, and time-frequency domain features of the cleaning concentration time series data of each monitoring point in the monitoring point cluster, and construct an enhancement vector for the monitoring point based on the cleaning concentration time series data of the monitoring point and the time domain features, frequency domain features, and time-frequency domain features of the cleaning concentration time series data; The output module 25 is used to input the enhanced vector of each monitoring point in the monitoring point cluster into a preset fuzzy neural network fault diagnosis model to obtain the fault type, fault level and fault probability confidence of the transformer oil chromatogram.

[0087] Based on the same inventive concept, the present invention further provides an electronic device, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to execute and implement the transformer oil chromatogram online fault monitoring method based on the fuzzy neural network algorithm as provided above.

[0088] Based on the same inventive concept, the present invention also provides a non-temporary computer-readable storage medium. When the instructions in the storage medium are executed by the processor of an electronic device, the electronic device is able to implement the transformer oil chromatography online fault monitoring method based on the fuzzy neural network algorithm as provided above.

[0089] Since the electronic device described in this embodiment is an electronic device used to implement the information processing method in the embodiment of the present invention, based on the information processing method described in the embodiment of the present invention, those skilled in the art will be able to understand the specific implementation of the electronic device of this embodiment and its various variations. Therefore, how the electronic device implements the method in the embodiment of the present invention will not be described in detail here. As long as the electronic device used by those skilled in the art to implement the information processing method in the embodiment of the present invention falls within the scope of protection of the present invention.

[0090] 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.

[0091] 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.

[0092] 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.

[0093] 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 1A step that specifies a function in one or more boxes.

[0094] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0095] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A transformer oil chromatogram online fault monitoring method based on fuzzy neural network algorithm, characterized in that: include: Acquire the concentration data of multiple gases at each monitoring point in the transformer oil chromatogram, and build a monitoring point cluster centered on the monitoring point when the gas concentration data is greater than the preset concentration threshold corresponding to the gas; Perform wavelet decomposition on the concentration time series data of each monitoring point in the monitoring point cluster to obtain the cleaning concentration time series data of each monitoring point. The time length of each cleaning concentration time series data is the same. Obtain the template concentration time series data of the monitoring point cluster, and determine whether the monitoring point in the monitoring point cluster is out of the cluster based on the template concentration time series data and the cleaning concentration time series data, and repair the outlier monitoring point; Extract the time domain features, frequency domain features and time-frequency domain features of the cleaning concentration time series data of each monitoring point in the monitoring point cluster, and construct the enhancement vector of the monitoring point based on the cleaning concentration time series data of the monitoring point and the time domain features, frequency domain features and time-frequency domain features of the cleaning concentration time series data; The enhanced vector of each monitoring point in the monitoring point cluster is input into the preset fuzzy neural network fault diagnosis model to obtain the fault type, fault level and fault probability confidence of the transformer oil chromatogram.

2. The transformer oil chromatogram online fault monitoring method based on fuzzy neural network algorithm as claimed in claim 1, characterized in that: The concentration time series data of each monitoring point in the monitoring point cluster is subjected to wavelet decomposition to obtain the cleaning concentration time series data of each monitoring point, including: Decompose the concentration time series data into high-frequency detail components and low-frequency approximate components, including: in, for The concentration time series data at each moment, for Moment The low-frequency approximate component of the layer, for Moment The high-frequency detail components of the layer; Denoise each high-frequency detail component, including: in, for Moment The high-frequency detail component after layer denoising, is the preset vector; Each low-frequency approximation component Input the preset LSTM model to approximate the low-frequency components Error correction is performed, where the input formula of the preset LSTM model is: in, for The current input gate state at time t, For the The weight matrix of the layer, is the hidden state at the previous moment, For the The bias of the layer, is the Sigmoid activation function; The corrected low-frequency approximate component It is spliced ​​with the corresponding denoised high-frequency detail components to obtain the cleaning concentration time series data.

3. The transformer oil chromatogram online fault monitoring method based on fuzzy neural network algorithm as claimed in claim 1, characterized in that: Obtain template concentration time series data for the monitoring point cluster, including: The historical concentration data of multiple gases at each monitoring point in the monitoring point cluster are intercepted to obtain several historical concentration time series data of each monitoring point, where the duration of the historical concentration time series data is the same as the duration of the cleaning concentration time series data; The mean of the historical concentration time series data of the monitoring points in the monitoring point cluster is used as the template concentration time series data of the monitoring point cluster.

4. The transformer oil chromatogram online fault monitoring method based on fuzzy neural network algorithm as claimed in claim 1, characterized in that: Determine whether a monitoring point in a monitoring point cluster is out of the group, including: Determine the cumulative distance between the cleaning concentration time series data of the monitoring point and the template concentration time series data of the monitoring point cluster, including: in, Cleaning concentration time series data Time series data with template concentration The cumulative distance between Cleaning concentration time series data Middle data and template concentration time series data The data, is an alignment path, is the set of all alignment paths; in, For monitoring points The local reachable density of For monitoring points The neighborhood size between the nearest monitoring points, For monitoring points The nearest monitoring point, From the monitoring point The accessible distance to the monitoring point; in, For monitoring points The local outlier factor, For monitoring points The local reachable density of If the local outlier factor of a monitoring point is greater than a preset value, the monitoring point is considered an outlier monitoring point.

5. The transformer oil chromatogram online fault monitoring method based on fuzzy neural network algorithm as claimed in claim 1 is characterized in that, Repair outlier monitoring points, including: Eliminate abnormal data from the cleaning concentration time series data of outlier monitoring points; The outlier monitoring points are filled using cubic spline interpolation.

6. The transformer oil chromatogram online fault monitoring method based on fuzzy neural network algorithm as claimed in claim 1 is characterized in that, Extract the time domain features, frequency domain features, and time-frequency domain features of the cleaning concentration time series data of each monitoring point in the monitoring point cluster, including: in, For the The gas concentration at each monitoring point, is the length of time for cleaning concentration time series data, is the central trend of gas concentration; in, is the variance; in, is the wavelet packet energy entropy, For the frequency bands, is the number of frequency bands.

7. The transformer oil chromatogram online fault monitoring method based on fuzzy neural network algorithm as claimed in claim 1 is characterized in that, The enhanced vector of the monitoring point is constructed based on the cleaning concentration time series data of the monitoring point, the time domain characteristics, frequency domain characteristics and time-frequency domain characteristics of the cleaning concentration time series data, including: The cleaning concentration time series data of the monitoring point, the time domain features, the frequency domain features and the time-frequency domain features of the cleaning concentration time series data are spliced ​​to obtain a high-dimensional feature vector; The principal components are obtained by decomposing the high-dimensional eigenvectors through the eigenvalues ​​of the preset covariance matrix; The preset sorted features in the principal component are fused to obtain the enhanced vector of the monitoring point.

8. A transformer oil chromatogram online fault monitoring device based on fuzzy neural network algorithm, characterized in that: include: An acquisition module is used to obtain the concentration data of multiple gases at each monitoring point in the transformer oil chromatogram, and when the concentration data of a gas is greater than a preset concentration threshold corresponding to the gas, a monitoring point cluster is constructed with the monitoring point as the center; A decomposition module is used to perform wavelet decomposition on the concentration time series data of each monitoring point in the monitoring point cluster to obtain the cleaning concentration time series data of each monitoring point, and the time length of each cleaning concentration time series data is the same; The repair module is used to obtain the template concentration time series data of the monitoring point cluster, and determine whether the monitoring point in the monitoring point cluster is out of the cluster based on the template concentration time series data and the cleaning concentration time series data, and repair the outlier monitoring point; The enhancement module is used to extract the time domain features, frequency domain features and time-frequency domain features of the cleaning concentration time series data of each monitoring point in the monitoring point cluster, and construct the enhancement vector of the monitoring point based on the cleaning concentration time series data of the monitoring point and the time domain features, frequency domain features and time-frequency domain features of the cleaning concentration time series data; The output module is used to input the enhanced vector of each monitoring point in the monitoring point cluster into the preset fuzzy neural network fault diagnosis model to obtain the fault type, fault level and fault probability confidence of the transformer oil chromatogram.

9. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute and implement the transformer oil chromatogram online fault monitoring method based on the fuzzy neural network algorithm according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that When the instructions in the non-transitory computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to implement the transformer oil chromatogram online fault monitoring method based on a fuzzy neural network algorithm as described in any one of claims 1 to 7.

Images