A power transformer fault diagnosis method based on DBN and improved ELM
By combining DBN and an improved ELM, and utilizing feature filtering and weight regularization techniques, the problems of slow speed and low accuracy in power transformer fault diagnosis are solved, achieving rapid and efficient fault identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-10
- Publication Date
- 2026-03-24
AI Technical Summary
Existing methods for diagnosing power transformer faults are slow and have low accuracy. Furthermore, due to limited fault data, they are prone to overfitting, resulting in low diagnostic accuracy.
Feature extraction is performed using a deep belief network (DBN) and combined with an improved extreme learning machine (ELM). Features are selected by sequential forwarding, input weights are generated using Circle chaotic mapping, and weight regularization is performed using an improved DropConnect algorithm and L1 and L2 norms to achieve fast and accurate fault diagnosis.
It improves the accuracy and efficiency of power transformer fault diagnosis, enabling rapid and accurate identification of fault types, reducing training time, and enhancing the model's classification accuracy.
Smart Images
Figure CN117034082B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of smart grids and relates to a power transformer fault diagnosis method based on DBN and improved ELM. BACKGROUND
[0002] A power transformer is an indispensable key electrical equipment in a box-type substation, plays an important role in voltage conversion and power transmission, and its operation state is directly related to the stability and safety of the entire smart grid. Due to long-term operation and complex working environment, the power transformer faces the risk of failure, and once the power transformer fails, it may cause significant economic loss and social impact. Therefore, it is of great significance to diagnose the fault type of the power transformer. Generally speaking, the transformer fault classification mode is uncertain, and the fault type obtained from the three ratios has errors, and based on the conventional neural network algorithm, the initial model lacks experience, and the model accuracy may be poor, so a certain amount of training samples are needed to obtain a relatively ideal prediction result. However, due to the relatively limited fault data of the power transformer, the existing fault diagnosis model is prone to overfitting problem, thereby resulting in low fault diagnosis accuracy. In addition, most of the current power transformer fault diagnosis models need to consume a large amount of time for training, which affects the efficiency of the fault diagnosis method.
[0003] In summary, there is an urgent need for a power transformer fault diagnosis method that can quickly and accurately diagnose the power transformer. SUMMARY
[0004] Therefore, the purpose of the present application is to provide a power transformer fault diagnosis method based on DBN and improved ELM to solve the problems of slow speed and low accuracy of the current power transformer fault diagnosis method, use DBN to extract features from power transformer operation state data, input the output results of DBN into the improved ELM for power transformer fault classification, and quickly and accurately realize the fault diagnosis of the power transformer.
[0005] To achieve the above purpose, the present application provides the following technical solutions:
[0006] A power transformer fault diagnosis method based on DBN and improved ELM, specifically comprising the following steps:
[0007] S1: Obtain power transformer operation state data, perform data cleaning and normalization processing, use the sequential forward method to screen features closely related to the operation state of the transformer, thereby reducing the dimension of the feature space and improving the training efficiency of the subsequent recognition model;
[0008] S2: input the features screened in step S1 into a deep belief network (DBN), train and learn the potential representation of the input features layer by layer through the DBN, capture the correlation between the transformer operating state data, i.e., learn the transformer operating state feature representation;
[0009] S3: fuse the DBN and the improved ELM, i.e., input the transformer operating state feature representation learned in step S2 into the improved ELM, generate the input weight of the ELM by using the Circle chaotic mapping, realize the regularization processing of the input weight and the output weight of the ELM based on the improved DropConnect algorithm and L1, L2 norms, which can improve the accuracy of the transformer operating state classification and provide a feasibility for accurately diagnosing the transformer fault, wherein the ELM is an extreme learning machine. The fusion of the DBN and the improved ELM can quickly and accurately realize the fault diagnosis of the power transformer.
[0010] Further, in step S1, the features closely related to the transformer operating state are screened by using the sequential forward method, which specifically includes the following steps:
[0011] S11: create an empty set for storing the features selected by the sequential forward method;
[0012] S12: for all unselected features, calculate their accuracy in the power transformer operating state classification respectively, select the feature with the highest classification accuracy, and add it to the created empty set;
[0013] S13: for the remaining unselected features, calculate their accuracy in the power transformer operating state classification together with the features already in the set respectively, select the unselected feature with the highest classification accuracy, and continue to add it to the set;
[0014] S14: repeat step S13 until the preset number of features or the accuracy in the power transformer operating state classification no longer improves, and in this way, the features with the highest classification accuracy in the power transformer operating state classification can be selected.
[0015] Further, in step S2, the DBN is composed of multiple restricted Boltzmann machines (RBMs), and each RBM is composed of a visible layer and a hidden layer;
[0016] The training and learning of the potential representation of the input features layer by layer through the DBN specifically includes the following steps:
[0017] S21: the RBMs pre-train the transformer operating state data layer by layer through maximum likelihood estimation, and can learn the high-level feature representation of the transformer operating state data step by step;
[0018] S22: After completing the layer-by-layer pre-training, the overall fine-tuning is performed, the entire DBN model is trained using back propagation, the parameters of the model are adjusted to minimize the loss function, and the quality of the feature representation and the performance of the transformer diagnosis model can be further improved.
[0019] Further, in step S3, the input weights of the ELM generated by the Circle chaotic mapping are generated by the sequence generated by the internal irrelevance of the Circle chaotic mapping instead of the random input weights of the ELM, which can enhance the difference between the hidden neurons in the transformer operating state classification model. Since the Circle chaotic mapping sequence itself contains offset and nonlinear characteristics, it can play the role of bias in the transformer operating state classification model, so there is no need to manually set the bias.
[0020] Further, in step S3, the input weight regularization processing of the ELM is implemented based on the improved DropConnect algorithm, which includes the following steps:
[0021] S31: By selecting the median value of the input weights generated by the Circle chaotic mapping as the threshold, the distribution of the connection weights can be more accurately reflected. Connection weights greater than the threshold are high-value connection weights, which contribute more to the performance of the transformer fault diagnosis network and play an important role in the decision of the model. On the contrary, connection weights less than the threshold are low-value connection weights, which contribute less to the performance of the transformer fault diagnosis network. By layering the connection weights, the performance of the transformer fault diagnosis model can be better optimized;
[0022] S32: The connection between neurons is controlled by the mask matrix. In the mask matrix, 1 indicates that the corresponding connection weight is retained, and 0 indicates that the corresponding connection weight is discarded. In order to make the transformer fault diagnosis model not overly dependent on specific connections, the mask matrix is set to follow a Gaussian distribution, and the median value of the Gaussian distribution is used as the threshold. The mask matrix corresponding to the high-value connection weight is composed of 1 for values greater than the threshold and 0 for values less than the threshold in the Gaussian distribution, which can help the network select and highlight the connections between neurons with more significant features and improve the sensitivity of the transformer fault diagnosis model to key features. Conversely, for the mask matrix corresponding to the low-value connection weight, it is composed of 1 for values less than the threshold and 0 for values greater than the threshold in the Gaussian distribution, which can reduce the complexity and redundancy of the model;
[0023] S33: Introducing Gaussian noise makes the connection weights in the network have a larger range of variation and diversity, which can better capture different features and patterns in the transformer operating state data;
[0024] S34: multiplying the connection weight with the mask matrix and adding the Gaussian noise multiplied with the mask matrix, the input weight regularized by the improved DropConnect algorithm can be obtained, and the transformer fault diagnosis model can be made more efficient and accurate by the regularization method.
[0025] Further, in step S3, the DBN and the improved ELM are fused by replacing the Softmax classifier in the DBN classification model with the improved ELM. When directly using the Softmax classifier to classify the transformer fault, the problems of gradient disappearance and long training time are prone to occur, and comparedly, the improved ELM does not need the back propagation algorithm, and the problems of gradient disappearance and long training time can be avoided when classifying the transformer fault. The selected original transformer operating state features are converted into more representative feature representations by using the DBN layer-by-layer training and learning, and the output results of the DBN are input into the improved ELM for power transformer fault classification, so that the power transformer fault diagnosis can be quickly and accurately realized.
[0026] The application has the advantages that: the application adopts the sequential forward method to screen the features closely related to the transformer operating state, the potential representation of the input features is trained and learned layer by layer by the DBN, the input weight of the ELM is generated by the Circle chaotic mapping, the regularization processing of the input weight and the output weight of the ELM is realized based on the improved DropConnect algorithm and L1, L2 norm, the accuracy of the transformer operating state classification can be improved, and the feasibility of accurately diagnosing the transformer fault is provided. The DBN and the improved ELM are fused, and the power transformer fault diagnosis can be quickly and accurately realized. The fusion method can fully utilize the feature extraction capability of the DBN and the fast training characteristics of the ELM, realize the efficient modeling and accurate classification of the transformer operating state, and thus an effective power transformer fault diagnosis scheme is provided.
[0027] Other advantages, objects, and features of the application will be set forth in part in the following specification taken in conjunction with the accompanying drawings, and in part will become apparent to those skilled in the art from a consideration of the following specification and drawings, or can be learned from the practice of the application. The advantages and objects of the application can be realized and attained by means of the instrumentalities and combinations pointed out in the following specification. BRIEF DESCRIPTION OF DRAWINGS
[0028] In order to make the objects, technical solutions and advantages of the application clearer, the preferred detailed description of the application will be made below in combination with the drawings, in which:
[0029] Fig. 1 FIG. 1 is a schematic diagram of the power transformer fault diagnosis method based on the DBN and the improved ELM of the application;
[0030] Fig. 2A schematic diagram of the improved DropConnect algorithm applied in a network; DETAILED DESCRIPTION
[0031] The present application is described herein with reference to particular embodiments thereof, which provide for a thorough and enabling disclosure of the application. The skilled person will readily appreciate that the application is not limited to the embodiments described herein, and that variations and modifications can be made to the embodiments described herein without departing from the spirit and scope of the application. The following examples are provided by way of illustration of the basic principles of the application, and the features of the examples can be combined with each other, without conflict, where appropriate.
[0032] Reference is made to Figs. 1-2 The application provides a power transformer fault diagnosis method based on DBN and improved ELM, specifically comprising the following steps:
[0033] S1: Obtain power transformer operating state data, perform data cleaning and normalization processing, and use sequential forward method to screen features closely related to transformer operating state, thereby reducing the dimension of feature space and improving the training efficiency of subsequent identification model.
[0034] In step S1, the sequential forward method is used to screen features closely related to transformer operating state, specifically comprising the following steps:
[0035] S11: Create an empty set for storing the features selected by the sequential forward method;
[0036] S12: Calculate the accuracy of each unselected feature in power transformer operating state classification, select the feature with the highest classification accuracy, and add it to the created empty set;
[0037] S13: Calculate the accuracy of each unselected feature together with the features already in the set in power transformer operating state classification, select the unselected feature with the highest classification accuracy, and continue to add it to the set;
[0038] S14: Repeat step S13 until the pre-set number of features or the accuracy in power transformer operating state classification no longer improves, and in this way, the features with the highest classification accuracy in power transformer operating state classification can be selected.
[0039] S2: Capture the association between transformer operating state data by training and learning the potential representation of input features layer by layer through DBN.
[0040] In step S2, the DBN trains and learns the potential representation of the input features layer by layer, specifically including the following steps:
[0041] S21: Since there are complex nonlinear factors in the transformer operating state, the DBN can capture the nonlinear relationship in the data without being limited to a linear model, which is crucial for the classification of the power transformer operating state. The DBN is composed of multiple restricted Boltzmann machines (RBM), and each RBM is composed of a visible layer and a hidden layer;
[0042] S22: The RBM pre-trains the transformer operating state data layer by layer through maximum likelihood estimation, and can gradually learn the high-level feature representation of the transformer operating state data;
[0043] S23: After completing the layer-by-layer pre-training, the overall fine-tuning is performed, the entire DBN model is trained using backpropagation, the parameters of the model are adjusted to minimize the loss function, and the quality of the feature representation and the performance of the transformer diagnosis model can be further improved.
[0044] S3: The DBN and the improved ELM are fused, that is, the transformer operating state feature representation learned in step S2 is input into the improved ELM, the input weights of the ELM are generated using Circle chaotic mapping, and the input weights and output weights of the ELM are regularized based on the improved DropConnect algorithm and L1, L2 norms, respectively, which can improve the accuracy of the transformer operating state classification and provide a feasibility for accurate diagnosis of transformer faults. The ELM is an extreme learning machine. The fusion of DBN and improved ELM can quickly and accurately realize the fault diagnosis of power transformers.
[0045] In step S3, the Circle chaotic mapping generates the input weights of the ELM, which uses the sequence z = (z1, z2, …) generated by the internal irrelevance of the Circle chaotic mapping to replace the random input weights of the ELM, which can enhance the differences between hidden neurons in the transformer operating state classification model. Since the Circle chaotic mapping sequence itself contains offset and nonlinear characteristics, it can act as a bias in the transformer operating state classification model, so there is no need to manually set the bias. The Circle chaotic mapping calculation formula is as follows:
[0046]
[0047] where z k is the value of the kth step, a and b are constants, n is the number of ELM input layers multiplied by the number of hidden layers, and the initial input weights W are obtained by arranging elements in the sequence Z, so W can be represented as:
[0048]
[0049] wherein K is the number of ELM input layer, L is the number of hidden layer.
[0050] The input weight regularization processing of the ELM is realized based on the improved DropConnect algorithm, and specifically includes the following steps:
[0051] S31: By selecting the median value of the input weight W generated by the Circle chaotic mapping as the threshold value, the distribution of the connection weight can be more accurately reflected. The connection weight W greater than the threshold value is a high-value connection weight, which contributes more to the performance of the transformer fault diagnosis network and plays an important role in the decision of the model. On the contrary, the connection weight W less than the threshold value is a low-value connection weight, which contributes less to the performance of the transformer fault diagnosis network. By performing hierarchical processing on the connection weight W, the performance of the transformer fault diagnosis model can be better optimized;
[0052] S32: By using the mask matrix M weight to control the connection between neurons, 1 in the mask matrix M weight indicates that the corresponding connection weight is retained, and 0 indicates that the corresponding connection weight W is discarded. In order to make the transformer fault diagnosis model not overly dependent on specific connections, the mask matrix M weight is set to be subject to Gaussian distribution, and the median value of the Gaussian distribution is used as the threshold value. The mask matrix M weight composed of 1 greater than the threshold value and 0 less than the threshold value in the Gaussian distribution can help the network to select and highlight the connection between neurons with more significant features, and improve the sensitivity of the transformer fault diagnosis model to key features. On the contrary, the mask matrix M weight composed of 1 less than the threshold value and 0 greater than the threshold value in the Gaussian distribution can reduce the complexity and redundancy of the model;
[0053] S33: The introduction of Gaussian noise ε makes the connection weight W in the network have a larger range of variation and diversity, so that it can better capture different features and patterns in the transformer operating state data;
[0054] S34: Multiplying the connection weight with the mask matrix M weight and the Gaussian noise ε respectively can obtain the input weight W' regularized by the improved DropConnect method. This regularization method can make the transformer fault diagnosis model more efficient and accurate. The specific calculation formula is as follows:
[0055] W'=(M weight +ε)*W
[0056] The DBN and the improved ELM are fused by replacing the Softmax classifier in the DBN classification model with the improved ELM. When the Softmax classifier is directly used to classify the transformer fault, the problems of gradient disappearance and long training time are prone to exist. In comparison, the improved ELM does not need the back propagation algorithm, and the problems of gradient disappearance and long training time can be avoided when classifying the transformer fault. The selected original transformer operating state features are converted into more representative features by using the DBN layer-by-layer training and learning. The output result of the DBN is input into the improved ELM to classify the power transformer fault, and the power transformer fault diagnosis can be quickly and accurately realized.
[0057] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the purpose and scope of the technical solutions, which should be covered in the scope of the claims of the present application.
Claims
1. A DBN and improved ELM-based power transformer fault diagnosis method, characterized in that, The method specifically comprises the following steps: S1: Obtain power transformer operating state data, perform data cleaning and normalization processing, and use the sequential forward method to screen features closely related to the transformer operating state; S2: Input the features screened in step S1 into a deep belief network (DBN), train and learn the potential representation of the input features layer by layer through the DBN, capture the correlation between the transformer operating state data, that is, learn the transformer operating state feature representation; S3: Fuse the DBN and the improved ELM, that is, input the transformer operating state feature representation learned in step S2 into the improved ELM, generate the input weights of the ELM by using the Circle chaotic mapping, realize the regularization processing of the input weights and the output weights of the ELM based on the improved DropConnect algorithm and L1 and L2 norms, and finally output the power transformer fault type, wherein the ELM is an extreme learning machine (ELM); The input weight regularization processing of the ELM is realized based on the improved DropConnect algorithm, and specifically comprises the following steps: S31: Select the median of the input weights generated by the Circle chaotic mapping as a threshold, the connection weights greater than the threshold are high-value connection weights, which contribute greatly to the performance of the transformer fault diagnosis network; on the contrary, the connection weights less than the threshold are low-value connection weights, which contribute little to the performance of the transformer fault diagnosis network, and the performance of the transformer fault diagnosis model is better optimized through hierarchical processing of the connection weights; S32: Control the connection between neurons through a mask matrix, 1 in the mask matrix indicates that the corresponding connection weight is retained, and 0 indicates that the corresponding connection weight is discarded; in order to make the transformer fault diagnosis model not overly dependent on specific connections, the mask matrix is set to be subject to a Gaussian distribution, and the median of the Gaussian distribution is taken as a threshold; the mask matrix corresponding to the high-value connection weights is composed of 1 greater than the threshold and 0 less than the threshold in the Gaussian distribution; on the contrary, the mask matrix corresponding to the low-value connection weights is composed of 1 less than the threshold and 0 greater than the threshold in the Gaussian distribution; S33: Multiply the connection weights by the mask matrix, add Gaussian noise, and multiply the mask matrix to obtain the input weights regularized by the improved DropConnect algorithm.
2. The power transformer fault diagnostic method according to claim 1, characterized by, In step S1, the sequential forward method is used to screen features closely related to the transformer operating state, and specifically comprises the following steps: S11: Create an empty set for storing the features selected by the sequential forward method; S12: Calculate the accuracy of each unselected feature in the power transformer operating state classification, select the feature with the highest classification accuracy, and add it to the created empty set; S13: Calculate the accuracy of each unselected feature together with the features already in the set in the power transformer operating state classification, select the unselected feature with the highest classification accuracy, and continue to add it to the set; S14: Repeat step S13 until the preset number of features or the accuracy in the power transformer operating state classification no longer improves.
3. The power transformer fault diagnostic method according to claim 1, characterized by, In step S2, the DBN is composed of multiple Restricted Boltzmann Machines (RBMs), and each RBM consists of a visible layer and a hidden layer; The process of training and learning the latent representation of input features layer by layer through DBN specifically includes the following steps: S21: RBM pre-trains the transformer operating status data layer by layer through maximum likelihood estimation, and gradually learns the high-level feature representation of the transformer operating status data. S22: After completing layer-by-layer pre-training, perform overall fine-tuning by using backpropagation to train the entire DBN model and adjust the model's parameters to minimize the loss function.
4. The power transformer fault diagnostic method according to claim 1, characterized by, In step S3, the step of generating the input weights of the ELM using the Circle chaotic map is to replace the random input weights of the ELM with a sequence generated by the uncorrelated sequence within the Circle chaotic map.
5. The power transformer fault diagnostic method according to claim 1, characterized by, In step S3, the fusion of DBN and the improved ELM involves replacing the Softmax classifier in the DBN classification model with the improved ELM.