Operating transformer fault studying and judging method and system based on mechanism data fusion calculation

Abstract

The invention discloses an operation transformer fault research and judgment method and system based on mechanism data fusion calculation, and belongs to the technical field of power equipment fault research and judgment. The method comprises the following steps: acquiring original operation data of the power transformer under various working conditions, and treating the original data to construct a fault research and judgment data set; establishing a neural network model, training the neural network model according to the fault research and judgment data set, and generating an evaluation model; and obtaining operation data of a target transformer in real time, and according to the operation data, based on the evaluation model, studying and judging a transformer fault. According to the method, the evolution result of the multi-physical field complex system of the operating transformer can be reliably predicted, common problems of parameter randomness, data noise, difficulty in model quantification and the like in fault research and judgment problems of the actually operating transformer can be effectively solved, and a powerful guarantee is provided for safe and stable operation of a power system.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault analysis and judgment, and more specifically, to a method and system for analyzing and judging a running transformer fault based on mechanism data fusion calculation. Background Art

[0002] As the core equipment in the power system, the transformer undertakes the key tasks of power conversion and distribution. Its stability and reliability directly affect the safe operation of the power system. Therefore, early diagnosis of transformer faults is of great significance for preventing large-scale power outages and reducing economic losses. At present, transformer fault diagnosis mainly relies on manual experience or simple threshold judgment, which often lacks accuracy and generalization ability. Traditional transformer fault diagnosis methods mainly rely on dissolved gas analysis (DGA), frequency response analysis (FRA), thermal imaging scanning and other technologies. These methods can reveal the operating status of the transformer to a certain extent, but they usually require a lot of manual intervention and are not sensitive enough to identify certain fault characteristics.

[0003] With the development of artificial intelligence technology, machine learning and deep learning methods have gradually been introduced into the field of transformer fault analysis. For example, methods such as support vector machine (SVM), convolutional neural network (CNN), and deep belief network (DBN) are used to automatically extract fault features and classify them. These methods are usually able to process large amounts of data, which improves the efficiency of fault analysis to a certain extent. However, transformer failure modes are diverse, and fault characteristics may show large differences in different environments and working conditions, which increases the difficulty of model training. Secondly, the cost of acquiring transformer fault data is high, and fault samples are scarce, which limits the generalization ability of the model. In addition, how to effectively integrate the physical characteristics and fault mechanisms of the transformer into the data-driven model to improve the interpretability and credibility of the model is also a pain point that needs to be solved urgently. Summary of the invention

[0004] In view of the above problems, the present invention proposes a method for fault analysis of operating transformers based on mechanism data fusion calculation, including:

[0005] Collect the original operating data of the transformer under various working conditions, and manage the original data to construct a fault analysis data set;

[0006] Establishing a neural network model, training the neural network model with the fault analysis data set, and generating an analysis model;

[0007] The operating data of the target transformer is acquired in real time, and the transformer operating fault is judged based on the judgment model.

[0008] Optionally, there are multiple operating conditions of the transformer, including: normal operation, overheating fault, discharge fault, and short - circuit fault.

[0009] Optionally, the original operation data includes: data on the content of dissolved gases in transformer oil, infrared image data, optical fiber temperature measurement signal data, partial discharge signal data, vibration data, and noise signal data.

[0010] Optionally, the governance of the original data includes:

[0011] Eliminating redundant and abnormal samples from the original operation data, and performing denoising and normalization processing on the remaining sample data.

[0012] Optionally, the neural network model includes:

[0013] An input layer, a feature extraction layer, a hidden layer, a dimensionality reduction layer, an output layer, and a custom loss function;

[0014] The feature extraction layer includes:

[0015] A recurrent neural network layer, multiple fully - connected layers, and a 1×1 convolutional layer;

[0016] The custom loss function includes:

[0017] A data - driven cross - entropy loss term, with an expression or corresponding description attached;

[0018] A mechanism - rule - driven regularization term, with an expression or corresponding description attached.

[0019] Optionally, using the fault judgment data set to train the neural network model to generate a judgment model, including:

[0020] Marking the fault data in the fault judgment data set, and dividing the marked fault judgment data set into a training set, a validation set, and a test set according to a preset ratio;

[0021] Training the neural network model with the training set and the validation set to generate a judgment model;

[0022] Using the validation set to verify the performance of the judgment model, and adjusting the model parameters of the evaluation model based on the verification results to optimize the performance of the judgment model.

[0023] Optionally, training the neural network model with the training set and the validation set to generate a judgment model, including:

[0024] When training the neural network model with the training set and the validation set, based on the custom loss function calculated by introducing mechanism data fusion, the neural network model is iteratively trained with the training set, and the performance of the neural network model is verified using the validation set during the iterative training process. If the performance does not improve for multiple consecutive cycles, the iterative training is stopped to generate a judgment model.

[0025] Optionally, during the iterative training process, adjust the model parameters of the neural network model to minimize the loss value of the custom loss function, and balance the importance of the custom loss function by adjusting the weight coefficient of the custom loss function.

[0026] Optionally, the custom loss function includes:

[0027] Balanced classification loss function and mechanism rule loss function;

[0028] The custom loss function is as follows:

[0029] L(θ) = L d (θ) + L p (θ)

[0030] The balanced classification loss function is as follows:

[0031]

[0032] Among them, θ is the model parameter to be optimized, including: the weights and biases of the network,

[0033] N is the total number of input operating transformer state samples;

[0034] L d (θ) is the total expression of the loss function, representing a function of the model parameter θ;

[0035] C is the operating state of a single category of transformer, representing the total number of operating states, including: different types of fault states or normal states;

[0036] y ic is the true label of whether the i-th group of samples belongs to the operating state c. If it belongs, then y ic = 1, otherwise y ic = 0;

[0037] w c is the weight of the operating state c;

[0038] α is a balance parameter used to control the influence degree of the focal loss part;

[0039] γ is a regulation parameter in the focal loss, used to reduce the weight of easy-to-classify samples;

[0040] p i (θ) is the probability that the model predicts the i-th group of samples to be in the operating state c;

[0041] The mechanism rule loss function is as follows:

[0042]

[0043] Among them, λ i is the weight coefficient;

[0044] f(x i ; θ) is the predicted output of the model for the i-th group of samples x i ;

[0045] r(f(x i ; θ), y i ) is a penalty function based on mechanism rules, which outputs a large value when the model prediction is inconsistent with the mechanism rules;

[0046] The definition of the penalty function is as follows:

[0047]

[0048] Among them, ò is a preset threshold for determining the boundary condition for triggering the penalty.

[0049] Optionally, based on the research and judgment model, the operation data of the target transformer is obtained in real time, and based on the operation data, the operation faults of the transformer are researched and judged, including:

[0050] Deploy the research and judgment model in the real-time data acquisition and processing system of the target transformer, so as to obtain the operation data of the target transformer in real time through the real-time data acquisition and processing system, and based on the operation data, predict the operation state of the target transformer, and based on the prediction result, research and judge the operation faults of the transformer.

[0051] On the other hand, the present invention also proposes an operation transformer fault research and judgment system based on mechanism data fusion calculation, including:

[0052] A data acquisition unit for acquiring the original operation data of the transformer under various working conditions and processing the original data to construct a fault research and judgment data set;

[0053] A model building unit for building a neural network model and training the neural network model with the fault research and judgment data set to generate a research and judgment model;

[0054] A research and judgment unit for obtaining the operation data of the target transformer in real time and researching and judging the operation faults of the transformer based on the research and judgment model.

[0055] Optionally, there are multiple operating conditions of the transformer, including: normal operation, overheating fault, discharge fault, and short-circuit fault.

[0056] Optionally, the original operation data includes: dissolved gas content data in transformer oil, infrared image data, optical fiber temperature measurement signal data, partial discharge signal data, vibration data, and noise signal data.

[0057] Optionally, governing the original data includes:

[0058] Eliminating redundant and abnormal samples in the original operation data, and performing denoising and normalization processing on the remaining sample data.

[0059] Optionally, the neural network model includes:

[0060] Input layer, feature extraction layer, hidden layer, dimensionality reduction layer, output layer, and custom loss function;

[0061] The feature extraction layer includes:

[0062] Recurrent neural network layer, multiple fully connected layers, 1×1 convolutional layer;

[0063] The custom loss function includes:

[0064] Data-driven cross-entropy loss term, with an expression or corresponding description attached;

[0065] Mechanism rule-driven regularization term, with an expression or corresponding description attached.

[0066] Optionally, training the neural network model with the fault judgment data set to generate a judgment model, including:

[0067] Marking the fault data in the fault judgment data set, and dividing the marked fault judgment data set into a training set, a validation set, and a test set according to a preset ratio;

[0068] Training the neural network model with the training set and the validation set to generate a judgment model;

[0069] Using the validation set to verify the performance of the judgment model, and adjusting the model parameters of the evaluation model based on the verification result to optimize the performance of the judgment model.

[0070] Optionally, training the neural network model with the training set and the validation set to generate a judgment model, including:

[0071] When training the neural network model with the training set and the validation set, based on the custom loss function calculated by introducing mechanism data fusion, the neural network model is iteratively trained with the training set, and the performance of the neural network model is verified using the validation set during the iterative training process. If the performance does not improve for multiple consecutive cycles, the iterative training is stopped to generate a judgment model.

[0072] Optionally, during the iterative training process, adjust the model parameters of the neural network model to minimize the loss value of the custom loss function, and balance the importance of the custom loss function by adjusting the weight coefficient of the custom loss function.

[0073] Optionally, the custom loss function includes:

[0074] Balanced classification loss function and mechanism rule loss function;

[0075] The custom loss function is as follows:

[0076] L(θ) = L d (θ) + L p (θ)

[0077] The balanced classification loss function is as follows:

[0078]

[0079] Among them, θ is the model parameter to be optimized, including: the weights and biases of the network,

[0080] N is the total number of input operation transformer state samples;

[0081] L d (θ) is the total expression of the loss function, representing a function of the model parameter θ;

[0082] C is the operation state of a single category of transformer, representing the total number of operation states, including: different types of fault states or normal states;

[0083] y ic is the true label indicating whether the i-th group of samples belongs to the operation state c. If it belongs, then y ic = 1, otherwise y ic = 0;

[0084] w c is the weight of the operation state c;

[0085] α is a balancing parameter used to control the influence degree of the focal loss part;

[0086] γ is a regulation parameter in the focal loss, used to reduce the weight of easily classified samples;

[0087] p i (θ) is the probability that the model predicts the i-th group of samples to be in the operating state c;

[0088] The mechanism rule loss function is as follows:

[0089]

[0090] where λ i is the weight coefficient;

[0091] f(x i ; θ) is the prediction output of the model for the i-th group of samples x i ;

[0092] r(f(x i ; θ), y i ) is a penalty function based on mechanism rules, which outputs a larger value when the model prediction is inconsistent with the mechanism rules;

[0093] The definition of the penalty function is as follows:

[0094]

[0095] where ò is a preset threshold for determining the boundary condition for triggering the penalty.

[0096] Optionally, based on the judgment model, the operation data of the target transformer is obtained in real time, and based on the operation data, the operation faults of the transformer are judged, including:

[0097] Deploy the judgment model in the real-time data acquisition and processing system of the target transformer, so as to obtain the operation data of the target transformer in real time through the real-time data acquisition and processing system, and based on the operation data, predict the operation state of the target transformer, and judge the operation faults of the transformer based on the prediction results.

[0098] On the other hand, the present invention also provides a computing device, including: one or more processors;

[0099] The processor is used to execute one or more programs;

[0100] When the one or more programs are executed by the one or more processors, the method as described above is implemented.

[0101] On the other hand, the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed, the method as described above is implemented.

[0102] Compared with the prior art, the beneficial effects of the present invention are:

[0103] The present invention provides a method for fault diagnosis of an operating transformer based on mechanism data fusion calculation, including: collecting the original operating data of the transformer under various working conditions, and processing the original data to construct a fault diagnosis data set; establishing a neural network model, training the neural network model with the fault diagnosis data set to generate a diagnosis model; and obtaining the operating data of the target transformer in real time, and diagnosing the operating faults of the transformer based on the diagnosis model. The present invention can reliably predict the evolution results of a multi-physical-field complex system of an operating transformer, effectively address common problems such as parameter randomness, data noise, and difficulty in quantifying the model in the fault diagnosis problem of an actual operating transformer, and provide a strong guarantee for the safe and stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] Figure 1 is a flow chart of the method of the present invention;

[0105] Figure 2 is a flow chart of an embodiment of the method of the present invention;

[0106] Figure 3 is a block diagram of the network model of an embodiment of the method of the present invention;

[0107] Figure 4 is a confusion matrix diagram of the model prediction results of an embodiment of the method of the present invention;

[0108] Figure 5 is a confusion matrix diagram of the prediction results of the comparative model of an embodiment of the method of the present invention;

[0109] Figure 6 is a structural diagram of the system of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0110] Now, exemplary embodiments of the present invention will be described with reference to the accompanying drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. These embodiments are provided to disclose the present invention in detail and completely, and to fully convey the scope of the present invention to those skilled in the art. The terms used in the exemplary embodiments shown in the drawings are not intended to limit the present invention. In the drawings, the same unit / element is denoted by the same reference numeral.

[0111] Unless otherwise specified, the terms (including scientific and technical terms) used herein have the ordinary meaning understood by those skilled in the art. In addition, it can be understood that the terms defined in a commonly used dictionary should be understood as having a meaning consistent with the context of their related fields, and should not be understood as having an idealized or overly formal meaning.

[0112] Embodiment 1:

[0113] The present invention proposes a method for judging faults of an operating transformer based on mechanism data fusion calculation, as Figure 1 shown, including:

[0114] Step 1, collect the original operation data of the transformer under various working conditions, and process the original data to construct a fault judgment data set;

[0115] Step 2, establish a neural network model, and train the neural network model with the fault judgment data set to generate a judgment model;

[0116] Step 3, obtain the operation data of the target transformer in real time, and judge the operation faults of the transformer based on the judgment model.

[0117] Among them, the various working conditions of the transformer include: normal operation, overheating fault, discharge fault, short - circuit fault.

[0118] Among them, the original operation data includes: the data of dissolved gas content in transformer oil, infrared image data, optical fiber temperature measurement signal data, partial discharge signal data, vibration data and noise signal data.

[0119] Among them, processing the original data includes:

[0120] Eliminate redundant and abnormal samples in the original operation data, and perform denoising and normalization processing on the remaining sample data.

[0121] Among them, the neural network model includes:

[0122] Input layer, feature extraction layer, hidden layer, dimensionality reduction layer, output layer, custom loss function;

[0123] The feature extraction layer includes:

[0124] Recurrent neural network layer, multiple fully - connected layers, 1×1 convolutional layer;

[0125] The custom loss function includes:

[0126] Data - driven cross - entropy loss term, with expression or corresponding description;

[0127] Mechanism - rule - driven regularization term, with expression or corresponding description.

[0128] Among them, training the neural network model with the fault judgment data set to generate a judgment model includes:

[0129] Mark the fault data in the fault judgment data set, and divide the marked fault judgment data set into a training set, a validation set and a test set according to a preset ratio;

[0130] Train the neural network model with the training set and the validation set to generate a judgment model;

[0131] Use the validation set to verify the performance of the judgment model, and adjust the model parameters of the evaluation model based on the verification result to optimize the performance of the judgment model.

[0132] Among them, training the neural network model with the training set and the validation set to generate a judgment model includes:

[0133] When training the neural network model with the training set and the validation set, based on the custom loss function introduced by the mechanism data fusion calculation, iteratively train the neural network model with the training set, and use the validation set to verify the performance of the neural network model during the iterative training process. If the performance does not improve for multiple consecutive cycles, stop the iterative training to generate a judgment model.

[0134] Among them, during the iterative training process, adjust the model parameters of the neural network model to minimize the loss value of the custom loss function, and balance the importance of the custom loss function by adjusting the weight coefficient of the custom loss function.

[0135] Among them, the custom loss function includes:

[0136] Balanced classification loss function and mechanism rule loss function;

[0137] The custom loss function is as follows:

[0138] L(θ) = L d (θ) + L p (θ)

[0139] The balanced classification loss function is as follows:

[0140]

[0141] Among them, θ is the model parameter to be optimized, including: the weights and biases of the network,

[0142] N is the total number of input operation transformer state samples;

[0143] L d (θ) is the total expression of the loss function, representing a function of the model parameter θ;

[0144] C is the operation state of a single category of transformer, representing the total number of operation states, including: different types of fault states or normal states;

[0145] y ic$y_{i,c}$ is the true label indicating whether the $i$-th group of samples belongs to the operating state $c$. If it belongs, then $y_{i,c} = 1$; otherwise, $y_{i,c} = 0$. ic $y_{i,c}$ ic $= 0$;

[0146] $w_c$ c is the weight of the operating state $c$;

[0147] $\alpha$ is the balance parameter used to control the influence degree of the focal loss part;

[0148] $\gamma$ is the adjustment parameter in the focal loss, used to reduce the weight of easily classified samples;

[0149] $p_i(\theta)$ i is the probability that the model predicts the $i$-th group of samples to be in the operating state $c$;

[0150] The mechanism rule loss function is as follows:

[0151]

[0152] where $\lambda$ i is the weight coefficient;

[0153] $f(x_i$ i ; $\theta)$ is the prediction output of the model for the $i$-th group of samples $x_i$ i ;

[0154] $r(f(x_i$ i ; $\theta), y_{i,c})$ i is a penalty function based on mechanism rules, which outputs a large value when the model prediction is inconsistent with the mechanism rules;

[0155] The definition of the penalty function is as follows:

[0156]

[0157] where $\delta$ is a preset threshold used to determine the boundary condition for triggering the penalty.

[0158] Among them, the judgment model obtains the operation data of the target transformer in real time, and judges the operation faults of the transformer according to the operation data, including:

[0159] Deploy the judgment model in the real-time data acquisition and processing system of the target transformer, so as to obtain the operation data of the target transformer in real time through the real-time data acquisition and processing system, and predict the operation state of the target transformer according to the operation data, and judge the operation faults of the transformer based on the prediction results.

[0160] The present invention will be further described below with specific embodiments:

[0161] The specific case process is as follows Figure 2 shown as follows, including:

[0162] Step 1, data governance: First, govern the multi-source heterogeneous raw data of the operating transformer collected, including steps such as signal denoising and normalization, in order to improve the training efficiency and performance of the model. According to the experience in the field of transformer fault judgment, the raw data can select typical operating data such as the content of dissolved gases in transformer oil, optical fiber temperature measurement signals, partial discharge signals, vibration and noise signals, etc.

[0163] Step 2, construction of neural network model: Secondly, design a deep learning model based on neural network so that the model can extract the complex features contained in the original operating data of the transformer. Considering that most of the existing research on automatic transformer fault judgment uses multi-layer perceptron (MLP), the present invention further explores other advanced model structures, such as recurrent neural network (RNN), the structure is as Figure 2 shown, making it perform better in processing time series data with complex features. The detailed design of the model is as follows:

[0164] Input layer:

[0165] Multiple types of data of the transformer operating under different working conditions after governance, including but not limited to the content of dissolved gases in oil, partial discharge signals, vibration and noise signals, optical fiber temperature measurement signals, operating current and voltage, etc.

[0166] Intermediate layer:

[0167] The first intermediate layer: Feature extraction layer, using recurrent neural network (RNN) to automatically extract the key features in the input signal.

[0168] Hidden layer: Use multiple fully connected layers (Dense Layer) to process the features extracted from the RNN.

[0169] The second intermediate layer: Dimensionality reduction layer, further compressing the feature space while retaining as much fault-related information as possible.

[0170] Nonlinear activation functions such as ReLU (Rectified Linear Unit) function are used in each intermediate layer to increase the expression ability of the model and accelerate the training process of the model.

[0171] Output layer:

[0172] Output the automatic fault judgment result. Determine the number of neurons in the output layer according to the number of fault categories. Use the softmax activation function in the output layer so that the output layer can output the probability distribution of each category.

[0173] Step 3, model training optimization method based on mechanism data fusion calculation:

[0174] The proposed method is mainly implemented by designing a custom loss function that includes a mechanism rule-driven regularization term. This loss function not only includes traditional classification losses (such as cross-entropy loss), but also includes a mechanism rule-driven regularization term. The mechanism rule-driven regularization term is used to penalize the model for generating prediction results that violate the physical characteristics and fault mechanisms of operating transformers.

[0175] During the model training process, a custom loss function that integrates mechanism rules is introduced: (1) An improved cross-entropy loss is used to measure the difference between the model's predicted output and the actual fault labels. (2) Constraints that reflect the physical characteristics and fault mechanisms of operating transformers are incorporated into the loss function.

[0176] The specific design of the custom loss function is as follows:

[0177] (1) Data-driven loss term: An improved cross-entropy loss function is used to measure the difference between the model output and the actual transformer fault labels, which is the basis for achieving high-accuracy classification.

[0178]

[0179] θ is the model parameter to be optimized, mainly the weights and biases of the network.

[0180] N represents the total number of input samples of the operating transformer status.

[0181] L d (θ) represents the total expression of the loss function, which is a function of the model parameter θ.

[0182] c represents the operating status of a single category of transformer, and C represents the total number of operating statuses (including different types of fault statuses or normal statuses).

[0183] y ic is the true label indicating whether the i-th group of samples belongs to the operating status c. If it belongs, then y ic = 1; otherwise, y ic = 0.

[0184] w c is the weight of the operating status c, which can be adjusted according to the number of samples in different categories to solve the problem of class imbalance. If the number of samples in a certain category is small, a higher weight can be given.

[0185] α is a balance parameter used to control the influence degree of the focal loss part.

[0186] γ is a regulation parameter in the focal loss used to reduce the weight of easily classified samples.

[0187] p iP(θ) is the probability that the model predicts the i-th group of samples to be in the operating state c, which can be obtained through forward propagation during the model training process.

[0188] It can be seen that the first term in the formula is the standard weighted cross-entropy loss part, which is used to measure the difference between the model prediction result and the true label. Among them, the weighting term w c allows different weights to be given to the losses of different classes. The second term is the focal loss part, which encourages the model to better learn these samples by increasing the loss weights of difficult-to-classify samples.

[0189] The design purpose of the above improved cross-entropy loss function is to handle the problem of sample imbalance and improve the model's recognition ability for difficult-to-classify samples through focal loss. Therefore, the above design is particularly applicable to the transformer fault judgment and judgment scenarios where the class imbalance is serious and the model needs to have good recognition ability for all classes. By minimizing the above custom loss function, the model can learn more accurate classification boundaries through training, thereby improving its performance in actual engineering applications.

[0190] (2) Mechanism rule-driven regularization term: Design the regularization term based on the physical characteristics of the transformer and the known fault modes based on domain expert experience. The regularization term is formed by comparing the difference between the predicted output of the model and the expected output based on mechanism rules, so as to ensure that the prediction result of the model conforms to the physical characteristics and fault mechanism of the transformer. The specific definition is as follows:

[0191]

[0192] Among them:

[0193] λ i is the weight coefficient, which can be set according to the importance of different mechanism rules.

[0194] f(x i ; θ) is the predicted output of the model for the i-th group of samples x i .

[0195] r(f(x i ; θ), y i ) is a penalty function based on mechanism rules, which outputs a larger value when the model prediction is inconsistent with the mechanism rules. The penalty function can be specifically defined as:

[0196]

[0197] Among them, ò is a preset threshold used to determine the boundary condition for triggering the penalty.

[0198] The above custom loss function can finally be expressed as the sum of the original loss function and the regularization term:

[0199] L(θ) = L d (θ) + L p (θ) where:

[0200] L d (θ) is the improved cross - entropy loss, which is used to measure the difference between the model prediction and the actual label.

[0201] L p (θ) is the regularization term, which is used to ensure that the prediction result of the model conforms to the physical characteristics and fault mechanisms of the transformer.

[0202] Through the above loss function design, it is ensured that the model not only focuses on the data - driven prediction accuracy during the training process, but also takes into account the physical characteristics and fault mechanisms of the transformer, thereby improving the generalization ability and prediction accuracy of the model.

[0203] Step 4, Model training and optimization: Use the labeled fault data to train the deep - learning model, and select an appropriate batch size to balance memory consumption and training efficiency. During the training process, continuously adjust and optimize the model parameters to make the loss function value as small as possible; at the same time, adjust the weight coefficients in the loss function to balance the importance of the classification loss and the mechanism rule loss. Set a sufficient number of iterations to ensure that the model fully learns, and at the same time monitor the possible over - fitting situation. If the performance on the validation set does not improve significantly in consecutive multiple epochs, stop the training. In addition, various optimization techniques can be adopted to accelerate convergence and improve the model performance.

[0204] Step 5, Model validation: Evaluate the performance of the model on independent validation sets and test sets. Classic metrics such as accuracy, recall, and F1 - score can be used to measure the accuracy of the judgment results, and k - fold cross - validation can be used to evaluate the generalization ability of the model.

[0205] Step 6, Model deployment: Deploy the trained offline model to the real - time data acquisition and processing system to realize the real - time automatic judgment of transformer faults.

[0206] Through the above - mentioned elaborate design, the automatic transformer fault judgment network model based on mechanism data fusion calculation can make full use of and learn the physical characteristics and fault mechanisms of the operating transformer, thereby improving the accuracy and reliability of the fault judgment results, providing timely decision - making support for the on - line monitoring and operation and maintenance of the transformer, and having important engineering significance.

[0207] By integrating mechanism rules to constrain the constructed neural network model, the present invention significantly reduces the number of training samples required for high-precision judgment, and particularly demonstrates strong generalization ability in the case of unbalanced data and small samples. In addition, the present invention clarifies the internal mechanism of automatic transformer fault judgment based on deep learning, which is conducive to the model learning the sample distribution of high-dimensional data. By introducing a mechanism rule-driven regularization term into the custom loss function, the prediction result of the model not only conforms to the sample distribution characteristics, but also follows the physical behavior of the transformer, thereby improving the reliability and generalization ability of the fault judgment result; at the same time, the interpretability of the model is enhanced, facilitating domain experts to understand and trust the judgment result.

[0208] The following takes an in-service transformer as an example for illustration:

[0209] First, collect the original data of the in-service transformer, mainly focusing on three types of transformer operation data, namely, dissolved gases in transformer oil, partial discharge, and infrared images. Then, according to expert experience, label the collected original data samples with operation status classification labels. Referring to the relevant standards and case compilations in the transformer industry, the fault states of the in-service transformer are specifically divided into three major types, namely, overheating fault, discharge fault, and short-circuit fault. After adding normal state samples, there are a total of four different types of operation states for the in-service transformer, and the corresponding labels are shown in Table 1 below:

[0210] Table 1

[0211]

[0212] In the stage of obtaining the original data, extract three types of data, namely, gas content, partial discharge statistical characteristic parameters, and electrical test data, and the corresponding feature vectors. After data governance and elimination of redundant and abnormal samples, 1200 groups of effective samples are sorted out from the original dataset. The distribution of various transformer states is shown in Table 1. To verify the performance of the model, the 1200 groups of operation transformer state samples are divided into a training set and a validation set according to a preset ratio. Specifically, 75% (900 samples) are randomly divided as the training set, and 25% (300 samples) are used as the test set to ensure that the model learns on sufficient training data and evaluates the accuracy and reliability of the automatic judgment result of the neural network on an independent test set. During the model verification process, record the distribution of the predicted samples in the test set.

[0213] In the data analysis stage, introduce a confusion matrix to assist in evaluating the accuracy of the automatic judgment result of the neural network and calculate statistical indicators. Use the following symbols to label the prediction results: True positive N TP represents the number of samples accurately determined to be in the normal state; False positive N FP refers to the number of samples misjudged as being in the normal state; True negative N TNRepresents the number of samples accurately determined to be in a fault state; false negative cases N FN Refers to the number of samples determined to be in an incorrect fault state. The precision rate λ is generated based on the label comparison result p , recall rate λ r , accuracy rate λ a , F1 score λ F1 and other 4 evaluation indicators. The calculation formulas are as follows:

[0214]

[0215]

[0216]

[0217]

[0218] To further verify the effectiveness of the proposed model, a transformer fault judgment model based on the proposed model and a traditional CNN network is established respectively, and the judgment performance is compared. After several repeated tests, when the model performance is stable, the confusion matrices of the proposed model and the comparison model are respectively as Figure 4 , Figure 5 shown, and the statistical evaluation indicators of the fault judgment results are shown in Table 2. Comparing Figure 4 , Figure 5 the calculation results in the confusion matrix, it can be seen that the judgment effect of the proposed model is generally better than that of the comparison machine learning model, especially in the case of unbalanced training sample quantities. Table 2 shows the statistical evaluation indicators of the above models, and each indicator of the proposed model is better than that of the comparison model

[0219] Table 2

[0220]

[0221] After training, compared with the traditional automatic transformer fault judgment model, the proposed model also has a large improvement in calculation speed. By controlling the model complexity, it is ensured that the models participating in the comparison have the same number of neurons. Under this standardized condition, the average processing speed of the proposed model is 12 milliseconds / sample, compared with 37 milliseconds / sample of the traditional multi-modal convolutional neural network (CNN) model, showing a significant calculation speed advantage. It can be seen that on the premise of maintaining the model complexity unchanged, the proposed model has effectively improved the sample processing efficiency, thus showing high potential application value in real-time transformer fault judgment and online monitoring of operating status

[0222] Example 2:

[0223] The present invention also proposes an operating transformer fault judgment system 200 based on mechanism data fusion calculation, as Figure 6As shown in the figure, it includes:

[0224] A data acquisition unit 201, configured to acquire the original operation data of the transformer under various working conditions, and process the original data to construct a fault judgment data set;

[0225] A model building unit 202, configured to establish a neural network model, and train the neural network model with the fault judgment data set to generate a judgment model;

[0226] A judgment unit 203, configured to acquire the operation data of the target transformer in real time, and judge the operation faults of the transformer based on the judgment model.

[0227] Among them, the various working conditions of the transformer include: normal operation, overheating fault, discharge fault, and short - circuit fault.

[0228] Among them, the original operation data includes: the data of the dissolved gas content in the transformer oil, infrared image data, optical fiber temperature measurement signal data, partial discharge signal data, vibration data, and noise signal data.

[0229] Among them, the processing of the original data includes:

[0230] Eliminating redundant and abnormal samples in the original operation data, and performing denoising and normalization processing on the remaining sample data.

[0231] Among them, the neural network model includes:

[0232] An input layer, a feature extraction layer, a hidden layer, a dimensionality reduction layer, an output layer, and a custom loss function;

[0233] The feature extraction layer includes:

[0234] A recurrent neural network layer, multiple fully - connected layers, and a 1×1 convolutional layer;

[0235] The custom loss function includes:

[0236] A data - driven cross - entropy loss term, with an expression or corresponding description attached;

[0237] A mechanism - rule - driven regularization term, with an expression or corresponding description attached.

[0238] Among them, training the neural network model with the fault judgment data set to generate a judgment model includes:

[0239] Marking the fault data in the fault judgment data set, and dividing the marked fault judgment data set into a training set, a validation set, and a test set according to a preset ratio;

[0240] Train the neural network model with the training set and the validation set to generate a judgment model;

[0241] Use the validation set to verify the performance of the judgment model, and adjust the model parameters of the evaluation model based on the verification results to optimize the performance of the judgment model.

[0242] Among them, training the neural network model with the training set and the validation set to generate a judgment model includes:

[0243] When training the neural network model with the training set and the validation set, based on the custom loss function calculated by introducing mechanism data fusion, iteratively train the neural network model with the training set, and use the validation set to verify the performance of the neural network model during the iterative training process. If the performance does not improve within multiple consecutive cycles, stop the iterative training to generate a judgment model.

[0244] Among them, during the iterative training process, adjust the model parameters of the neural network model to minimize the loss value of the custom loss function, and balance the importance of the custom loss function by adjusting the weight coefficient of the custom loss function.

[0245] Among them, the custom loss function includes:

[0246] Balanced classification loss function and mechanism rule loss function;

[0247] The custom loss function is as follows:

[0248] L(θ) = L d (θ) + L p (θ)

[0249] The balanced classification loss function is as follows:

[0250]

[0251] Among them, θ is the model parameter to be optimized, including: the weights and biases of the network,

[0252] N is the total number of input operating transformer state samples;

[0253] L d (θ) is the total expression of the loss function, representing a function of the model parameter θ;

[0254] C is the operating state of the transformer for a single category, representing the total number of operating states, including: different types of fault states or normal states;

[0255] y ic is the true label of whether the i-th group of samples belongs to the operating state c. If it belongs, then yic = 1, otherwise y ic = 0;

[0256] w c is the weight of the operating state c;

[0257] α is a balance parameter used to control the influence degree of the focal loss part;

[0258] γ is a regulation parameter in the focal loss, used to reduce the weight of easy-to-classify samples;

[0259] p i (θ) is the probability that the model predicts the i-th group of samples to be in the operating state c;

[0260] The mechanism rule loss function is as follows:

[0261]

[0262] where λ i is the weight coefficient;

[0263] f(x i ; θ) is the prediction output of the model for the i-th group of samples x i ;

[0264] r(f(x i ; θ), y i ) is a penalty function based on the mechanism rule, which outputs a large value when the model prediction is inconsistent with the mechanism rule;

[0265] The definition of the penalty function is as follows:

[0266]

[0267] where ò is a preset threshold used to determine the boundary condition for triggering the penalty.

[0268] Among them, based on the research and judgment model, the operation data of the target transformer is obtained in real time, and based on the operation data, the operation faults of the transformer are researched and judged, including:

[0269] Deploy the research and judgment model in the real-time data acquisition and processing system of the target transformer, so as to obtain the operation data of the target transformer in real time through the real-time data acquisition and processing system, and based on the operation data, predict the operation state of the target transformer, and based on the prediction result, research and judge the operation faults of the transformer.

[0270] The present invention can reliably predict the evolution results of the multi-physical field complex system of an operating transformer, effectively address the common problems such as parameter randomness, data noise, and difficulty in quantifying models in the actual operation of transformer fault judgment problems, and provide a strong guarantee for the safe and stable operation of the power system.

[0271] Embodiment 3:

[0272] Based on the same inventive concept, the present invention also provides a computer device, which includes a processor and a memory. The memory is used to store a computer program, and the computer program includes program instructions. The processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding function, so as to implement the steps of the method in the above embodiment.

[0273] Embodiment 4:

[0274] Based on the same inventive concept, the present invention also provides a storage medium, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a computer device and is used to store programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and, of course, the extended storage medium supported by the computer device. The computer-readable storage medium provides a storage space, and this storage space stores the operating system of the terminal. And, one or more instructions suitable for being loaded and executed by the processor are also stored in this storage space. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. One or more instructions stored in the computer-readable storage medium can be loaded and executed by the processor to implement the steps of the method in the above embodiment.

[0275] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) that contain computer-usable program code. The solutions in the embodiments of the present invention can be implemented in various computer languages. For example, object-oriented programming languages such as Java and interpreted scripting languages such as JavaScript, etc.

[0276] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0277] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing devices to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0278] These computer program instructions can also be loaded onto a computer or other programmable data processing devices, such that a series of operation steps are executed on the computer or other programmable devices to generate a computer-implemented process, so that the instructions executed on the computer or other programmable devices provide steps for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0279] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concepts. Therefore, the appended claims are intended to be construed as including the preferred embodiments and all changes and modifications falling within the scope of the present invention.

[0280] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A method for fault diagnosis of an operating transformer based on mechanism data fusion calculation, characterized in that, Including: Collecting the original operation data of the transformer under various working conditions, and governing the original data to construct a fault judgment data set; Establishing a neural network model, and training the neural network model with the fault judgment data set to generate a judgment model; Obtaining the operation data of the target transformer in real time, and judging the operation faults of the transformer based on the judgment model.

2. The method for judging the faults of an operating transformer according to claim 1, wherein The various working conditions of the transformer include: normal operation, overheating fault, discharge fault, and short-circuit fault.

3. The method for judging the faults of an operating transformer according to claim 1, wherein The original operation data includes: the content data of dissolved gases in transformer oil, infrared image data, optical fiber temperature measurement signal data, partial discharge signal data, vibration data, and noise signal data.

4. The method for judging the faults of an operating transformer according to claim 1, characterized in that The governing of the original data includes: Eliminating redundant and abnormal samples in the original operation data, and performing denoising and normalization processing on the remaining sample data.

5. The method for judging the faults of an operating transformer according to claim 1, characterized in that The neural network model includes: Input layer, feature extraction layer, hidden layer, dimensionality reduction layer, output layer, and custom loss function; The feature extraction layer includes: Recurrent neural network layer, multiple fully connected layers, 1×1 convolutional layer; The custom loss function includes: Data-driven cross-entropy loss term, with expression or corresponding description attached; Mechanism rule-driven regularization term, with expression or corresponding description attached.

6. The method for judging the faults of an operating transformer according to claim 1, wherein, The training of the neural network model with the fault judgment data set to generate a judgment model includes: Marking the fault data in the fault judgment data set, and dividing the marked fault judgment data set into a training set, a validation set, and a test set according to a preset ratio; Training the neural network model with the training set and the validation set to generate a judgment model; Using the validation set to verify the performance of the judgment model, and adjusting the model parameters of the evaluation model based on the verification result to optimize the performance of the judgment model.

7. The method for judging the faults of an operating transformer according to claim 6, wherein, The training of the neural network model with the training set and the validation set to generate a judgment model includes: When training the neural network model with the training set and the validation set, based on the custom loss function introduced by the mechanism data fusion calculation, iteratively training the neural network model with the training set, and using the validation set to verify the performance of the neural network model during the iterative training process. If the performance does not improve in consecutive multiple cycles, stop the iterative training to generate a judgment model.

8. The method for judging the fault of an operating transformer according to claim 7, characterized in that, During the iterative training process, adjusting the model parameters of the neural network model to minimize the loss value of the custom loss function, and balancing the importance of the custom loss function by adjusting the weight coefficient of the custom loss function.

9. The method for judging the faults of an operating transformer according to claim 8, wherein, The custom loss function includes: Balanced classification loss function and mechanism rule loss function; The custom loss function is as follows: L(θ) = L d (θ) + L p (θ) The balanced classification loss function is as follows: Where θ is the model parameter to be optimized, including: the weights and biases of the network, N is the total number of input operation transformer state samples; L d (θ) is the total expression of the loss function, representing a function of the model parameter θ; C is the transformer operation state of a single category, representing the total number of operation states, including: different types of fault states or normal states; y ic is the true label indicating whether the i-th group of samples belongs to the operating state c. If it belongs, then y ic = 1; otherwise y ic = 0; w c is the weight for operating state c; α is a balance parameter used to control the influence degree of the focal loss part; γ is a regulation parameter in the focal loss, which is used to reduce the weight of easily classified samples; p i (θ) is the probability that the model predicts the i-th group of samples to be in the operating state c; The mechanism rule loss function is as follows: Among them, λ i is the weight coefficient; f(x i ; θ) is the predicted output of the model for the i-th group of samples x i ; r(f(x i ; θ), y i ) is a penalty function based on mechanism rules. When the model prediction is inconsistent with the mechanism rules, this function outputs a relatively large value; The definition of the penalty function is as follows: where ò is a preset threshold for determining the boundary condition for triggering the penalty.

10. The transformer fault judgment method according to claim 1, characterized in that, Based on the judgment model, the operation data of the target transformer is obtained in real time, and based on the operation data, the operation faults of the transformer are judged, including: Deploy the judgment model in the real-time data acquisition and processing system of the target transformer, so as to obtain the operation data of the target transformer in real time through the real-time data acquisition and processing system, and based on the operation data, predict the operation state of the target transformer, and based on the prediction result, judge the operation faults of the transformer.

11. An operating transformer fault judgment system based on mechanism data fusion calculation, characterized in that, Including: The data acquisition unit is used to collect the original operation data of the transformer under various working conditions, and manage the original data to construct a fault judgment data set; The model building unit is used to build a neural network model, and train the neural network model with the fault judgment data set to generate a judgment model; The judgment unit is used to obtain the operation data of the target transformer in real time, and judge the operation faults of the transformer based on the judgment model.

12. The operating transformer fault judgment system according to claim 11, wherein The various working conditions of the transformer include: normal operation, overheat fault, discharge fault, short circuit fault.

13. The operating transformer fault judgment system according to claim 11, wherein The original operation data includes: the content data of dissolved gases in transformer oil, infrared image data, optical fiber temperature measurement signal data, partial discharge signal data, vibration data and noise signal data.

14. The operating transformer fault judgment system according to claim 11, wherein The management of the original data includes: Eliminating redundant and abnormal samples in the original operation data, and performing denoising and normalization processing on the remaining sample data.

15. The operating transformer fault judgment system according to claim 11, wherein, The neural network model includes: Input layer, feature extraction layer, hidden layer, dimensionality reduction layer, output layer, custom loss function; The feature extraction layer includes: Recurrent neural network layer, multiple fully connected layers, 1×1 convolutional layer; The custom loss function includes: Data-driven cross-entropy loss term, with expression or corresponding description; Mechanism rule-driven regularization term, with expression or corresponding description.

16. The operating transformer fault judgment system according to claim 11, characterized in that, Training the neural network model with the fault judgment data set to generate a judgment model includes: Marking the fault data in the fault judgment data set, and dividing the marked fault judgment data set into a training set, a validation set and a test set according to a preset ratio; Training the neural network model with the training set and the validation set to generate a judgment model; Using the validation set to check the performance of the judgment model, and adjusting the model parameters of the evaluation model based on the check result to optimize the performance of the judgment model.

17. The operating transformer fault judgment system according to claim 16, wherein, Training the neural network model with the training set and the validation set to generate a judgment model includes: When training the neural network model with the training set and the validation set, based on the custom loss function calculated by introducing mechanism data fusion, the neural network model is iteratively trained with the training set, and the performance of the neural network model is verified using the validation set during the iterative training process. If the performance does not improve for multiple consecutive cycles, the iterative training is stopped to generate a judgment model.

18. The operating transformer fault judgment system according to claim 17, wherein During the iterative training process, the model parameters of the neural network model are adjusted to minimize the loss value of the custom loss function, and the importance of the custom loss function is balanced by adjusting the weight coefficients of the custom loss function.

19. The operating transformer fault judgment system according to claim 18, characterized in that, The custom loss function includes: Balanced classification loss function and mechanism rule loss function; The custom loss function is as follows: L(θ) = L d (θ) + L p (θ) The balanced classification loss function is as follows: Where θ is the model parameter to be optimized, including the weights and biases of the network, N is the total number of input operating transformer state samples; L d (θ) is the total expression of the loss function, representing a function of the model parameter θ; C is the operating state of a single category of transformer, representing the total number of operating states, including different types of fault states or normal states; y ic is the true label indicating whether the i-th group of samples belongs to the operating state c. If it belongs, then y ic = 1, otherwise y ic = 0; w c The weight for the operating state c; α is a balance parameter used to control the influence degree of the focal loss part; γ is a regulation parameter in the focal loss used to reduce the weight of easily classified samples; p i (θ) is the probability that the model predicts the i-th group of samples to be in the operating state c; The mechanism rule loss function is as follows: Among them, λ i is the weight coefficient; f(x i ; θ) is the predicted output of the model for the i-th group of samples x i ; r(f(x i ; θ), y i ) is a penalty function based on mechanism rules. When the model prediction is inconsistent with the mechanism rules, this function outputs a relatively large value; The definition of the penalty function is as follows: Where ò is a preset threshold used to determine the boundary condition for triggering the penalty.

20. The transformer fault judgment system according to claim 11, characterized in that Based on the judgment model, the operating data of the target transformer is obtained in real time, and based on the operating data, the operating faults of the transformer are judged, including: Deploy the judgment model in the real-time data acquisition and processing system of the target transformer, so as to obtain the operating data of the target transformer in real time through the real-time data acquisition and processing system, and based on the operating data, predict the operating state of the target transformer, and judge the operating faults of the transformer based on the prediction results.

21. A computer device, characterized in that, Including: One or more processors; The processor is used to execute one or more programs; When the one or more programs are executed by the one or more processors, the method described in any one of claims 1-10 is implemented.

22. A computer-readable storage medium, characterized in that, There is a computer program stored thereon, and when the computer program is executed, the method described in any one of claims 1-10 is implemented.

Images