A Fault Diagnosis Method for Power Transformers Considering Feature Coupling Relationships

Abstract

The present invention discloses a power transformer fault diagnosis method considering feature coupling relationships, including: selecting main features related to transformer operation faults; determining a feature state transition sequence according to the power transformer fault mechanism; analyzing and optimizing the features included in the feature state transition sequence to obtain feature data most relevant to transformer faults; performing standardization processing on the feature data; constructing a Bi-LSTM fault diagnosis model based on a deep neural network and training the Bi-LSTM fault diagnosis model; and using the trained Bi-LSTM fault diagnosis model to evaluate the current operation state of the transformer. The present invention comprehensively considers the internal mechanism of the power transformer and the data fitting ability of the deep neural network, considers the feature coupling relationships inside the equipment while using the deep neural network for transformer fault diagnosis, realizes a reliable mapping between data features and equipment fault types, conducts refined state evaluation, and improves the fault diagnosis accuracy of the power transformer.

Description

Technical Field

[0001] The present invention relates to the technical field of power transformer fault diagnosis, and particularly to a power transformer fault diagnosis method considering feature coupling relationships. Background Art

[0002] As a core device for power transmission and conversion in a power system, how to perform precise reliability maintenance on a power transformer is one of the issues widely concerned by the current power industry and academia. Power transformer fault diagnosis methods are mainly divided into two categories. One category is physical operation modeling or constructing an expert system based on equipment mechanism models and expert knowledge. The other category is machine learning or deep learning methods based on data models. Driven by the new generation of artificial intelligence theory and power big data, fault diagnosis methods based on machine learning or deep learning have been widely studied and applied. It can effectively promote the transformation of the equipment maintenance method from planned maintenance to condition-based maintenance and predictive maintenance, and solve problems such as insufficient maintenance or over-maintenance existing in the traditional planned maintenance process.

[0003] Compared with machine learning fault diagnosis methods such as Support Vector Machine (SVM), Particle Swarm Optimization-SVM (PSO-SVM), and AdaBoost, deep learning diagnosis methods mainly based on deep neural networks have a strong high-dimensional non-linear mapping ability and can perform more in-depth analysis on equipment operation data. Deep neural networks such as Back Propagation Network (BPNet) and Deep Belief Network (DBN) can directly learn after receiving historical equipment data, adjust the transfer parameters between model units through iterative calculations, and fit the distribution law of equipment monitoring data. In addition, there are also solutions that propose using Bi-directional Long Short-Term Memory (Bi-LSTM) for transformer fault diagnosis. It constructs network models for historical time series of different data feature parameters respectively, and obtains the fault diagnosis result through BPNet after obtaining the output at the current moment from each Bi-LSTM network respectively.

[0004] Inside a large power transformer is a multi - physical - field coupling environment where the electromagnetic field, temperature field, and flow field interact with each other. This multi - physical - field coupling environment strengthens the dependence relationship of state characteristics, resulting in an induced or compliant causal relationship in the change of characteristic data. Traditional methods such as physical operation modeling or constructing expert systems based on equipment mechanism models and expert knowledge are limited by data analysis capabilities and are no longer applicable to the current situation where the integration of transformer equipment is continuously increasing and the component coupling is continuously strengthening; fault diagnosis methods based on machine learning only predict data by fitting the data distribution law from historical data and lack in - depth analysis of features; although the widely used deep - learning fault diagnosis methods currently have strong high - dimensional non - linear mapping capabilities, they still perform fusion analysis and model construction from the data level without considering the equipment fault mechanism and the internal characteristic action mode of the data, which may lead to a reduction in the reliability and accuracy of equipment state assessment. Therefore, it is necessary to adopt a hybrid - driven method of mechanism model and data model. Based on the actual operation of the equipment, use a deep neural network to adaptively extract the complex interaction relationships inside the equipment and construct a transformer fault diagnosis model that organically combines data - driven and knowledge - guided. Summary of the Invention

[0005] The purpose of the present invention is to provide a power transformer fault diagnosis method considering characteristic coupling relationships, comprehensively considering the internal mechanism of the power transformer and the data - fitting ability of the deep neural network. When using the deep neural network for fault diagnosis of the transformer, consider the characteristic coupling relationships inside the equipment, realize a reliable mapping between data characteristics and equipment fault types, conduct refined state assessment, and improve the fault diagnosis accuracy of the power transformer.

[0006] To achieve the above - mentioned purpose, the present invention provides the following solutions:

[0007] A power transformer fault diagnosis method considering characteristic coupling relationships, comprising the following steps:

[0008] Step S1, select the main characteristics related to the operation faults of the transformer;

[0009] Step S2, determine the characteristic state transition sequence according to the power transformer fault mechanism;

[0010] Step S3, analyze and optimize the characteristics contained in the characteristic state transition sequence according to the relationship between the selected characteristics and the transformer fault types and the interaction between the characteristics to obtain the characteristic data most relevant to the transformer faults;

[0011] Step S4, perform standardization processing on the characteristic data;

[0012] Step S5: Based on the deep neural network, construct a Bi-LSTM fault diagnosis model to mine the coupling relationship between feature data, and on this basis, fit the mapping relationship between the feature action result and the transformer fault type;

[0013] Step S6: Train the Bi-LSTM fault diagnosis model;

[0014] Step S7: When new power transformer monitoring data is obtained, input it into the trained Bi-LSTM fault diagnosis model to evaluate the current operating state of the transformer.

[0015] Further, in step S2, determine the characteristic state transition sequence according to the power transformer fault mechanism, which specifically includes: for the power transformer, obtain the characteristics related to the fault mechanism, as well as the causal relationship, sequence relationship, and dependency relationship between the characteristics through prior knowledge and expert experience, and construct the characteristic state transition sequence.

[0016] Further, in step S4, perform standardization processing on the feature data, which specifically includes: mapping the feature data to the same scale through the Z-score standardization method.

[0017] Further, in step S5, based on the deep neural network, construct a Bi-LSTM fault diagnosis model to mine the coupling relationship between feature data, and on this basis, fit the mapping relationship between the feature action result and the transformer fault type,

[0018] Step S51: Select a typical recurrent neural network as the basic structure of the deep neural network and construct a Bi-LSTM fault diagnosis model;

[0019] Step S52: Select the state transfer method between network units in the deep neural network;

[0020] Step S53: Improve the deep neural network fault diagnosis model. On the basis of the original LSTM, flexibly process the feature sequence by adding network units that transfer information from back to front. The input state of each feature in the network at the current moment depends on the feature subsequences input before and after this moment, and automatically mine the coupling action mode through multiple gated units and two calculation directions.

[0021] Further, in step S6, train the Bi-LSTM fault diagnosis model, which specifically includes:

[0022] Step S61: Encode the power transformer fault types;

[0023] Step S62: Determine the data input matrix according to the characteristic state transition sequence order and the neural network model structure;

[0024] Step S63: Divide the training data into a training set and a validation set, and set a set of neural network hyperparameters to verify the fault diagnosis effect of the trained models under different hyperparameters. Among them, the training set is used to determine the neural network model obtained under a certain set of hyperparameters, and the validation set is used to verify the fault diagnosis effect of the neural network model trained under this set of hyperparameters.

[0025] Step S64: Select a set of hyperparameters from the set of hyperparameters for training the Bi-LSTM fault diagnosis model.

[0026] Step S65: Initialize the network model parameters.

[0027] Step S66: Batch input the training set data records into the network input layer.

[0028] Step S67: Update the network model parameters according to the network output loss through the gradient descent optimization algorithm.

[0029] Step S68: When all data records are input and the number of traversals of all data reaches the specified value, save the network model structure and the obtained model parameters.

[0030] Step S69: Input the validation set into the Bi-LSTM fault diagnosis model, verify the fault diagnosis effect of the model and record it.

[0031] Step S610: When all sets of hyperparameters in the set of hyperparameters have been used for model training, compare the fault diagnosis effects of the models under different neural network hyperparameters on the validation data set, and comprehensively select a set of hyperparameters that make the fault diagnosis model perform optimally after weighing.

[0032] Step S611: Set the Bi-LSTM fault diagnosis model with the determined hyperparameters, retrain the Bi-LSTM fault diagnosis model using all the training data in Step S62, perform iterative calculations according to Steps S65 - S67, and when all data records are input and the number of traversals of all data reaches the specified value, save the trained model for real-time fault diagnosis of power transformers.

[0033] Further, in Step S7, when new power transformer monitoring data is obtained, input it into the trained Bi-LSTM fault diagnosis model to evaluate the current operating state of the transformer, specifically including:

[0034] Step S71: Obtain the latest monitoring data of the power transformer.

[0035] Step S72: Screen the characteristic parameters of the monitoring data records according to the main characteristics related to the transformer operation faults determined in Step S1.

[0036] Step S73: Rearrange the feature parameters in the data record according to the feature order in the feature state transition sequence determined in Step S2;

[0037] Step S74: Process the feature parameters according to the feature optimization and feature standardization methods described in Step S3;

[0038] Step S75: Input the monitored feature data into the trained Bi-LSTM fault diagnosis model in the order of the feature sequence;

[0039] Step S76: Obtain the model prediction result and determine the operating state of the transformer based on the fault code.

[0040] According to the specific embodiments provided by the present invention, the following technical effects are disclosed: The power transformer fault diagnosis method considering the feature coupling relationship provided by the present invention deeply mines the transformer monitoring data. On the basis of directly analyzing the data record, it effectively integrates the feature action state transition relationship and the deep neural network with strong representation ability at the data feature level, clarifies the feature action mode, obtains the uncertain factors affecting the operating state of the transformer, and accurately and reliably analyzes the faults of the power transformer; In the present invention, first, the initial feature state transition sequence of the data is determined according to the equipment fault mechanism, and then a Bi-LSTM fault diagnosis model is constructed to be deeply integrated with and support each other with the feature sequence. While mining the multi-feature coupling effect, the fault diagnosis effect of the power transformer is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0042] Figure 1 It is a schematic diagram of the internal coupling relationship of the power transformer in the embodiment of the present invention;

[0043] Figure 2 It is a schematic diagram of the construction process of the feature state transition sequence in the embodiment of the present invention;

[0044] Figure 3 It is the RNN fault diagnosis model structure in the embodiment of the present invention;

[0045] Figure 4 It is the internal structure of the LSTM network unit in the embodiment of the present invention;

[0046] Figure 5 It is the Bi-LSTM fault diagnosis model structure in the embodiment of the present invention;

[0047] Figure 6 This is the flowchart of the training process of the Bi-LSTM fault diagnosis model according to the embodiments of the present invention;

[0048] Figure 7 This is the flowchart of the Bi-LSTM fault diagnosis algorithm for fusing feature sequences according to the embodiments of the present invention. Detailed implementation manners

[0049] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0050] The object of the present invention is to provide a power transformer fault diagnosis method considering the feature coupling relationship, comprehensively considering the internal mechanism of the power transformer and the data fitting ability of the deep neural network, considering the feature coupling relationship inside the equipment while using the deep neural network to diagnose the transformer fault, realizing a reliable mapping between data features and equipment fault types, performing refined state evaluation, and improving the fault diagnosis accuracy of the power transformer.

[0051] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0052] As Figure 1 and Figure 7 shown, the power transformer fault diagnosis method considering the feature coupling relationship provided by the present invention includes:

[0053] Step S1, select the main features related to the transformer operation fault; the main features are obtained according to the transformer fault mechanism and all have a direct or indirect relationship with the transformer fault state. Among them, when selecting the main features related to the equipment fault, the influence of various factors needs to be considered, such as the external environment;

[0054] Step S2, determine the feature state transition sequence according to the power transformer fault mechanism;

[0055] Among them, inside a large power transformer is a multi-physical field coupling environment where the electromagnetic field, temperature field, and flow field interact with each other. The multi-field coupling environment strengthens the dependence relationship of the state features, bringing a causal relationship with the feature data change being induced or compliant. The schematic diagram of the internal coupling relationship of the power transformer is as Figure 1As shown in the figure. The operating state of the equipment in the physical space of the transformer is mapped to the information space through sensing and monitoring equipment, and the modeling and analysis in the information space can support the intelligent operation and maintenance of the equipment. Inside the power transformer, the manifestation of one characteristic may be affected by multiple other characteristic factors and act on other related characteristics in turn.

[0056] Among them, under the multi-field coupling effect in the power transformer, the manifestation of one characteristic may be affected by multiple other characteristic factors and act on other related characteristics in turn. The characteristic state transition sequence is initially constructed according to the causal relationship, sequence relationship, dependence relationship, etc. between characteristic effects. The schematic diagram of the construction process of the characteristic state transition sequence is as Figure 2 shown. Under the multi-field coupling effect of the complex transformer, the characteristic effect relationship in the equipment operation data is initially obtained according to the fault mechanism, and the initial characteristic state transition sequence is determined.

[0057] Among them, since the change of the characteristic state of different physical fields in the same equipment can be reflected in the change of the characteristics in a single physical field through the transmission of energy flow, when the characteristic parameters related to the equipment fault mechanism cannot be comprehensively obtained or analyzed, that is, when all the characteristics in multiple physical fields cannot be obtained at the same time or the dependence relationship between the characteristics in multiple physical fields cannot be obtained, the information model can be simplified, and the fault characteristics in a single physical field can be selected for characteristic coupling analysis and equipment state evaluation.

[0058] Among them, for complex power transformers, the characteristics related to the fault mechanism and the causal relationship, sequence relationship, dependence relationship, etc. between the characteristics need to be obtained through prior knowledge, expert experience, etc.

[0059] Among them, the characteristic sequence is used to rearrange the characteristic parameters of the monitoring data.

[0060] Step S3, characteristic analysis and optimization. According to the relationship between the selected characteristics and the transformer fault types and the interaction between the characteristics, further analyze and optimize the characteristics contained in the characteristic sequence to obtain the data characteristics most relevant to the transformer fault, so as to improve the stability of the fault diagnosis algorithm;

[0061] Step S4, standardize the characteristic data. In order to improve the accuracy of the algorithm, it is necessary to standardize the characteristic data before the fault assessment. The characteristic data is mapped to the same scale through the Z-score standardization method to avoid some characteristics forming a dominant role due to different dimensions. The specific calculation method is as shown in Equation (1):

[0062]

[0063] In the formula, x represents the characteristic parameter in each data record, and mean(x) and σ represent the mean and standard deviation of each characteristic parameter respectively.

[0064] Step S5: Based on the deep neural network, construct a Bi-LSTM fault diagnosis model to mine the coupling relationship between feature data, and on this basis, fit the mapping relationship between the feature action result and the transformer fault type. Specifically, it includes:

[0065] Step S51: Select the basic structure of the deep neural network.

[0066] The feature sequence containing dependency logic has a formal temporal structure. Among the existing deep neural network structures, the recurrent neural network can better meet the needs of calculating temporal data. The fault diagnosis model constructed using the typical Recurrent Neural Network (RNN) is as Figure 3 shown. In the model, the number of input layers, hidden layers is the same as the number of input features. The input of the previous moment's feature will affect the input state of the current feature data through the calculation of the hidden layer, and they will jointly act on the feature input at the next moment. After the last hidden layer obtains the cumulative action state of all feature gases, the fault evaluation result is calculated through the added fully connected network layer.

[0067] Step S52: Select the state transfer method between network units in the deep neural network.

[0068] The Long Short-Term Memory (LSTM) is an improvement of the traditional recurrent neural network. Compared with the original RNN network structure, it adds an input gate, a forget gate, and an output gate between each hidden state. It can adaptively monitor the change in the number of data features, control the flow of information through learnable gates, automatically attenuate the cumulative feature effects, and flexibly adjust the long and short feature dependency relationships, so as to autonomously learn the strength of the interaction between features. The specific LSTM network unit structure and information processing method are as Figure 4 shown. Compared with the original RNN network structure, it adds an input gate, a forget gate, and an output gate between each hidden state. The specific calculation process inside the network unit is shown in Step S52.

[0069] Among them, during the training process of the transformer fault diagnosis model using LSTM, the network sequentially transmits the feature effects through forward propagation and finally maps them to the transformer operating state. The gating unit learns the network parameters during the backpropagation of gradient descent to jointly control how the next hidden state should be affected by the existing features. The forward propagation calculation process of the neural network after adding the gating unit is shown in Equation (2).

[0070]

[0071] where \(t\in[1,n]\), \(n\) is the number of feature variables; \(\sigma\) is the sigmoid activation function, which is used to perform a non-linear transformation on the variable value and map it to the interval \((0,1)\), and can improve the non-linear fitting ability of the network; tanh is the hyperbolic tangent function, which can control the variable value within the interval \((-1,1)\) and can increase the convergence speed of the network model; \(X\) t is the feature input matrix, which is determined according to the input data format and the network model structure; \(W\) and \(b\) are the weight parameter matrix and the bias vector between network units respectively; the memory cell state \(C\) at the previous moment t-1 and the forget gate \(F\) t , the candidate memory cell and the input gate \(I\) t are multiplied element-wise and summed respectively to obtain \(C\) t ; the hidden state \(H\) directly acting on the next input feature t is jointly determined by the output gate \(O\) t and the candidate memory cell . After the network model finishes traversing the feature sequence, the final action states of multiple features are saved in the hidden layer \(H\) n . Finally, the network prediction result \(Y\) is obtained under the calculation of the fully connected layer. The specific calculation process is shown in Equation (3).

[0072] \(Y = H\) n W hy + b h (3)

[0073] wherein, the LSTM fault diagnosis network can initially determine the relationship between network units through the feature state transition sequence, and then perform iterative calculations with the goal of minimizing the classification loss until the algorithm converges. Considering the differences in the effects of multiple features during the process of evaluating the operating state of the transformer can build a refined data-driven model for transformer fault diagnosis.

[0074] Step S53: Improve the deep neural network fault diagnosis model.

[0075] In the multi-physical field coupling environment of large power transformers, the effects between features are not necessarily one-way propagation. It may be a multi-directional interaction relationship of mutual interaction and influence. During the process of transformer fault diagnosis, if it is only assumed that the feature action states in the equipment are all transmitted from front to back based on the feature sequence, it has certain limitations. Therefore, the present invention selects to use a bidirectional LSTM (Bi-directional LSTM, Bi-LSTM) network to further explore the coupling relationship of characteristic gases, and on this basis, fit the mapping relationship between the feature action result and the equipment state type.

[0076] Among them, based on the original LSTM, Bi-LSTM flexibly processes the feature sequence by adding network units that transmit information from back to front. The input state of each feature in the network at the current moment depends on the subsequences of features input before and after this moment. The coupling action mode is automatically mined through multiple gating units and two calculation directions.

[0077] Among them, the Bi-LSTM fault diagnosis network first receives the device monitoring data records according to the feature sequence, then obtains the bidirectional cumulative action state through the bidirectional propagation of the hidden state by the LSTM unit, and connects to obtain the final state matrix. Finally, the feature action state is mapped to the device state type through the fusion calculation of the fully connected network layer to obtain the fault diagnosis result. The improved transformer fault diagnosis model structure is as Figure 5 shown.

[0078] Step S6: Train the Bi-LSTM fault diagnosis model.

[0079] Before using the Bi-LSTM fault diagnosis algorithm that fuses the prior feature sequence to evaluate the state of the transformer, it is necessary to train the Bi-LSTM fault diagnosis model to determine the calculation parameters in the model, specifically including:

[0080] Step S61: Encode the fault types of the power transformer.

[0081] Step S62: Determine the data input matrix according to the feature sequence order and the neural network model structure.

[0082] Step S63: Divide the training data into a training set and a validation set, and set the neural network hyperparameter set to verify the fault diagnosis effect of the trained model under different hyperparameters.

[0083] Among them, the training set is used to determine the neural network model obtained under a certain set of hyperparameters, and the validation set is used to verify the fault diagnosis effect of the neural network model trained under this set of hyperparameters.

[0084] Step S64: Select a set of hyperparameters from the hyperparameter set to train the Bi-LSTM fault diagnosis model.

[0085] Among them, a set of hyperparameters of the Bi-LSTM network model includes: learning rate, batch size, dropout ratio, weight decay system, model depth, model width, etc.

[0086] Step S65: Initialize the network model parameters.

[0087] Among them, model parameters are different from model hyperparameters. Model parameters are the parameters adjusted by fitting the training data, while hyperparameters are the parameters artificially adjusted before training.

[0088] Step S66: Batch input the training set data records into the network input layer.

[0089] Among them, after the data is input into the neural network input layer, the fault diagnosis network will perform a series of calculations, and the calculation process is as follows:

[0090] 1) The network cyclically forwards the feature action state.

[0091] 2) The network cyclically backwards the feature action state within the feature action step size.

[0092] 3) Connect the bidirectional hidden states to obtain the final feature state matrix and input it into the fully connected network.

[0093] 4) Calculate the loss function value between the network output result and the actual category of the sample.

[0094] Step S67: Update the network model parameters according to the network output loss through the gradient descent optimization algorithm.

[0095] Step S68: When all data records are completely input and the number of traversals of all data reaches the specified value, save the network model structure and the obtained model parameters.

[0096] Step S69: Input the validation set into the fault diagnosis model, verify the fault diagnosis effect of the model and record it.

[0097] Among them, the fault diagnosis effect of the model can be reflected by the final loss function value and the fault prediction accuracy rate.

[0098] Step S610: When all hyperparameter groups in the hyperparameter set have been used for model training, compare the fault diagnosis effects of the models under different neural network hyperparameters, and comprehensively measure and select a set of hyperparameters that make the fault diagnosis model perform optimally.

[0099] Step S611: Set the neural network fault diagnosis model with the determined hyperparameters, retrain the neural network model using all the training data in Step S62, perform iterative calculations according to Steps S65 - S67, and when all data records are completely input and the number of traversals of all data reaches the specified value, save the trained model for real-time fault diagnosis of power transformers.

[0100] Figure 6 It is the training process of the Bi-LSTM fault diagnosis model for the fusion feature sequence described in Steps S65 - S68 under a certain set of model hyperparameters.

[0101] Step S7, when new monitoring data of the power transformer is obtained, it is input into the trained Bi-LSTM fault diagnosis model to evaluate the current operating state of the transformer. Specifically, it includes:

[0102] Step S71, obtain the latest monitoring data of the device.

[0103] Step S72, screen the characteristic parameters recorded in the monitoring data according to the main characteristics related to the transformer operation faults determined in step S1.

[0104] Step S73, rearrange the characteristic parameters in the data record according to the order of the characteristics within the characteristic sequence determined in step S2.

[0105] Step S74, process the characteristic parameters according to the characteristic optimization and characteristic standardization methods described in step S3.

[0106] Step S75, input the monitoring characteristic data into the trained Bi-LSTM fault diagnosis model in the order of the characteristic sequence.

[0107] Step S76, obtain the model prediction result and determine the operating state of the transformer according to the fault code.

[0108] Figure 7 The flowchart of the Bi-LSTM fault diagnosis algorithm for the fused characteristic sequence as described in steps S2 - S7 after determining the main characteristics for evaluating the operating state of the transformer according to the actual situation.

[0109] In a specific embodiment, insulating oil is the main insulating material in the transformer insulation system. After being affected by temperature changes and discharges, it will age and generate gases through ionic reactions or oxidation reactions. The content of dissolved gases in the insulating oil can significantly reflect the operating conditions of the equipment components. And when too much gas decomposes, it will lead to the degradation of the electrical properties of the insulating oil, which directly affects the operating state and service life of the transformer.

[0110] Analyzing the components and content of the gases dissolved in the insulating oil is one of the most effective measures for monitoring the safe operation of power transformers. The method proposed in the present invention can be used to analyze the dissolved gases in the transformer oil for transformer fault diagnosis.

[0111] The first step, transformer fault feature selection.

[0112] In order to simplify the information model, in a multi-physical field environment, the dissolved gases in the oil in the flow field are selected for characteristic coupling analysis and equipment state evaluation. The effects generated by the electromagnetic field and the temperature field can be reflected in the content change of the characteristic gases through the oil flow circulation.

[0113] In the process of state assessment and fault diagnosis of power transformers by analyzing the types and contents of dissolved gases in oil, the oxygen concentration is different under different external environments, and oxygen atoms with too high a content will play a dominant role. Therefore, excluding carbon monoxide (CO) and carbon dioxide (CO2) after the oxidation reaction, among the seven gases in the oil, five gases, namely hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), and acetylene (C2H2), are selected as the main characteristic gases for evaluating the operating state of the transformer.

[0114] Step 2: Coding of transformer fault types.

[0115] In this embodiment, the fault types of the transformer can be divided into five types, namely low-temperature overheating, low-energy discharge, medium-temperature overheating, high-temperature overheating, and arc discharge. The operating state types of the transformer and the corresponding codes are shown in Table 1.

[0116] Table 1 Operating state types and codes of power transformers

[0117]

[0118] Step 3: Determine the characteristic state transition sequence.

[0119] Although it is difficult to directly model and analyze the complex coupling relationship of the internal characteristic gases of power transformers, the gas production principle of the equipment is basically the same. According to the current Dissolved Gas Analysis (DGA) and judgment guidelines for transformers, hydrogen atoms in insulating oil are relatively active and have a high content. Discharge or overheating faults cause the weakest C-H bonds in the chemical groups in the oil to break and recombine to produce hydrogen (H2) and low hydrocarbon gases (CH4, C2H6, C2H4, C2H2), and the formation of CH4, C2H6, C2H4, and C2H2 from hydrogen atoms and C-H, C-C, C=C, and C≡C bonds in the chemical groups requires higher temperatures and energies in turn. Therefore, the characteristic sequence [H2→CH4→C2H6→C2H4→C2H2] is initially determined to represent the transfer relationship of characteristic states, and the characteristic gases in the sequence have sequential dependence.

[0120] Step 4: Fault feature analysis and optimization.

[0121] Since the contents of characteristic gases are relatively dispersed and the gas contents vary greatly under different environments, in order to improve the stability of the fault diagnosis algorithm, the gas data corresponding to the five characteristics in this paper are reconstructed. Let C sum be the total content of the five gases, C TH be the total content of the four hydrocarbon gases except H2, and the relative content of H2 is C H2％ =C H2 / C sum , and the relative contents of the remaining gases are CCH4％ = C CH4 / C TH , C C2H6％ = C C2H6 / C TH , C C2H4％ = C C2H4 / C TH , C C2H2％ = C C2H2 / C TH , finally forming the characteristic sequence [H2% → CH4% → C2H6% → C2H4% → C2H2%].

[0122] Step 5, preprocessing of historical operation data.

[0123] 1) Select the gas contents of H2, CH4, C2H6, C2H4, and C2H2 and the corresponding transformer status types from the device historical operation data records to form training data.

[0124] 2) Reconstruct the training data according to the characteristic sequence [H2% → CH4% → C2H6% → C2H4% → C2H2%].

[0125] 3) Use the Z-score normalization method to normalize the reconstructed training data.

[0126] Step 6, verification of hyperparameters of the Bi-LSTM fault diagnosis model.

[0127] Determine the data input matrix according to the Bi-LSTM neural network model structure, and verify the model hyperparameters through steps S63 - S68. Select a set of hyperparameters that make the model perform optimally and record them.

[0128] Step 7, training of the Bi-LSTM fault diagnosis model.

[0129] Set the Bi-LSTM fault diagnosis model with the determined hyperparameters, batch input the preprocessed characteristic data obtained in the fifth step into the fault diagnosis model, retrain the neural network, and perform iterative calculations according to steps S65 - S67. When all data records are completed and the number of traversals of all data reaches the specified value, save the trained model.

[0130] Step 8, preprocessing of real-time monitoring data.

[0131] Obtain new monitoring data of dissolved gases in transformer oil, reconstruct the monitoring data in the form and order of [H2% → CH4% → C2H6% → C2H4% → C2H2%], and perform Z-score normalization on the reconstructed monitoring data according to the mean and standard deviation of different characteristics obtained in step 3) of the fifth step.

[0132] Step 9, Prediction of the Bi-LSTM Fault Diagnosis Model.

[0133] Input the processed characteristic gas monitoring data into the Bi-LSTM fault diagnosis model in the order of the characteristic sequence to obtain the model output result.

[0134] Step 10, Obtain the transformer fault diagnosis result.

[0135] The model output result is a fault code. According to the correspondence between the operation state types and codes of power transformers in Table 1, the operation state of the transformer is obtained.

[0136] The power transformer fault diagnosis method considering the characteristic coupling relationship provided by the present invention starts from the actual operation of the transformer, determines the characteristic state transition sequence that is easy to represent based on the equipment fault mechanism, then constructs a Bi-LSTM fault diagnosis network based on the characteristic sequence, further excavates the characteristic coupling relationship through multiple gating units and two calculation directions, clarifies the data characteristic action mode, and on this basis, fits the reliable mapping between the characteristic action state and the equipment state. Incorporating the characteristic sequence with a dependency relationship into the deep neural network can further sort out the complex mapping relationship between the characteristic data and the transformer state, construct a refined equipment diagnosis model in a hybrid driving manner of the mechanism model and the data model, improve the power transformer fault diagnosis accuracy, and thus provide a technical guarantee idea for the popularization and application of complex power equipment in the power system.

[0137] Each embodiment in this specification is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0138] In this application, specific examples are used to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A power transformer fault diagnosis method considering feature coupling relationship, characterized in that Including: Step S1, select the main features related to the operation faults of the transformer; Step S2, determine the characteristic state transition sequence according to the fault mechanism of the power transformer. Specifically: For the power transformer, obtain the features related to the fault mechanism, as well as the causal relationship, sequence relationship, and dependence relationship between the features through prior knowledge and expert experience, and construct the characteristic state transition sequence; Step S3, analyze and optimize the features contained in the characteristic state transition sequence according to the relationship between the selected features and the transformer fault types and the interaction between the features, and obtain the characteristic data most relevant to the transformer faults; Step S4, perform standardization processing on the characteristic data; Step S5, based on the deep neural network, construct a Bi-LSTM fault diagnosis model, mine the coupling relationship between the characteristic data, and on this basis, fit the mapping relationship between the characteristic action result and the transformer fault type; Step S6, train the Bi-LSTM fault diagnosis model, specifically including: Step S61, encode the power transformer fault types; Step S62, determine the data input matrix according to the characteristic state transition sequence order and the neural network model structure; Step S63, divide the training data into a training set and a validation set, and set the neural network hyperparameter set to verify the fault diagnosis effect of the trained model under different hyperparameters. Among them, the training set is used to determine the neural network model obtained under a certain set of hyperparameters, and the validation set is used to verify the fault diagnosis effect of the neural network model trained under this set of hyperparameters; Step S64, select a set of hyperparameters from the hyperparameter set to train the Bi-LSTM fault diagnosis model; Step S65, initialize the network model parameters; Step S66, batch input the training set data records into the network input layer; Step S67, update the network model parameters through the gradient descent optimization algorithm according to the network output loss; Step S68, when all data records are input and the number of traversal times of all data reaches the specified value, save the network model structure and the obtained model parameters; Step S69, input the validation set into the Bi-LSTM fault diagnosis model, verify the fault diagnosis effect of the model and record it; Step S610, when all hyperparameter groups in the hyperparameter set are used for model training, compare the fault diagnosis effects of the models under different neural network hyperparameters on the validation data set, and comprehensively measure and select a set of hyperparameters that make the fault diagnosis model perform optimally; Step S611, set the Bi-LSTM fault diagnosis model with the determined hyperparameters, retrain the Bi-LSTM fault diagnosis model using all the training data in Step S62, perform iterative calculations according to Steps S65 - S67, and when all data records are input and the number of traversal times of all data reaches the specified value, save the trained model for real-time fault diagnosis of the power transformer; Step S7, when new power transformer monitoring data is obtained, input it into the trained Bi-LSTM fault diagnosis model to evaluate the current operating state of the transformer.

2. The power transformer fault diagnosis method considering feature coupling relationships according to claim 1, characterized in that In step S4, the feature data is normalized, specifically including: mapping the feature data to the same scale by the Z-score normalization method.

3. The power transformer fault diagnosis method considering feature coupling relationships according to claim 1, characterized in that In step S5, based on the deep neural network, a Bi-LSTM fault diagnosis model is constructed to mine the coupling relationship between the feature data, and on this basis, the mapping relationship between the feature action result and the transformer fault type is fitted. Step S51: Select a typical recurrent neural network as the basic structure of the deep neural network to construct a Bi-LSTM fault diagnosis model. Step S52: Select the state transfer method between the network units in the deep neural network. Step S53: Improve the deep neural network fault diagnosis model. On the basis of the original LSTM, network units for transmitting information from back to front are added to flexibly process the feature sequence. The input state of each feature in the network at the current moment depends on the feature subsequences input before and after this moment. The coupling action mode is automatically mined through multiple gated units and two calculation directions.

4. The power transformer fault diagnosis method considering feature coupling relationships according to claim 1, characterized in that In step S7, when new power transformer monitoring data is obtained, it is input into the trained Bi-LSTM fault diagnosis model to evaluate the current operating state of the transformer, specifically including: Step S71: Obtain the latest monitoring data of the power transformer. Step S72: Screen the characteristic parameters recorded in the monitoring data according to the main characteristics related to the transformer operation faults determined in step S1. Step S73: Rearrange the characteristic parameters in the data record according to the feature order in the feature state transition sequence determined in step S2. Step S74: Process the characteristic parameters according to the feature optimization and feature normalization methods described in step S3. Step S75: Input the monitoring feature data into the trained Bi-LSTM fault diagnosis model in the order of the feature sequence. Step S76: Obtain the model prediction result and determine the operating state of the transformer according to the fault code.

Images