Transformer fault intelligent diagnosis system and method based on K-Medoid and SMOTE optimization support vector machine

Abstract

The invention discloses a transformer fault intelligent diagnosis system and method based on a K-Medoid and SMOTE optimization support vector machine, and the system comprises a data collection module which is used for collecting dissolved gas and operation parameters in transformer oil in real time; the data preprocessing module is connected with the data acquisition module and is used for completing data standardization and abnormal value detection; the sample balancing module is connected with the data preprocessing module and used for reducing majority class redundancy through K-Medoid clustering and expanding minority classes through an SMOTE algorithm; the intelligent classification module is connected with the sample balance module and adopts a weighted SVM classifier to train and predict; and the result output and visualization module is connected with the intelligent classification module and is used for displaying the classification result, the confidence coefficient and the key performance indexes. According to the method, the problem of classification performance degradation caused by sample imbalance in the prior art is solved, and the recognition capability of the SVM model on minority class faults and the overall diagnosis precision are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of transformers, and particularly relates to a transformer fault intelligent diagnosis system and method based on K-Medoid and SMOTE optimized support vector machines. BACKGROUND

[0002] With the continuous expansion of the power system, the number and complexity of transformers are significantly improved. As the core equipment of the power system, the running state of the transformer directly affects the safety and stability of the system, and its failure may cause a chain effect, resulting in large-area power outages and serious social, economic and environmental losses.

[0003] The traditional dissolved gas analysis (DGA) method has low accuracy when dealing with complex nonlinear data, and it is difficult to identify minority class fault samples. With the development of artificial intelligence, machine learning methods based on support vector machines (SVM) have shown advantages in small sample and nonlinear classification. However, the number of samples of different fault types in actual transformer operation data is significantly different, which causes the SVM model to be biased towards the majority class during training, resulting in biased diagnosis results.

[0004] Existing research mainly focuses on parameter optimization and feature selection: optimizing SVM parameters through genetic algorithm, particle swarm, etc. to improve classification performance, but the sample imbalance problem is still prominent; using single SMOTE (English full name: Synthetic Minority Oversampling Technique) oversampling to expand minority class samples, but there are problems of distribution distortion and boundary ambiguity; combining clustering algorithms to reduce redundant samples, but unable to balance the dual goals of majority class compression and minority class expansion. Therefore, the current technology still has problems such as low accuracy, unstable boundary and low training efficiency in small sample and sample imbalance power equipment fault identification. SUMMARY

[0005] To solve the problems in the prior art, the purpose of the present application is to provide a transformer fault intelligent diagnosis system and method based on K-Medoid and SMOTE optimized support vector machines, which solves the problem of classification performance degradation caused by sample imbalance in the prior art, and improves the recognition ability of the SVM model for minority class faults and the overall diagnosis accuracy.

[0006] To achieve the above purposes and achieve the above technical effects, the technical solution adopted by the present application is: A transformer fault intelligent diagnosis system based on K-Medoid and SMOTE optimized support vector machines, comprising: a data acquisition module for real-time acquisition of transformer dissolved gases and operating parameters; A data preprocessing module is connected with the data acquisition module, and is used for completing data standardization and abnormal value detection. A sample balancing module is connected with the data preprocessing module, and is used for reducing majority class redundancy through K-Medoid clustering and expanding minority class through SMOTE algorithm. An intelligent classification module is connected with the sample balancing module, and adopts a weighted SVM classifier for training and prediction. A result output and visualization module is connected with the intelligent classification module, and is used for displaying classification results, confidence and key performance indicators.

[0007] The application further discloses a transformer fault intelligent diagnosis method based on K-Medoid and SMOTE optimized support vector machine. Step one: data acquisition The data acquisition module is used for collecting transformer oil dissolved gas and operation parameters in real time, forming an original sample data set, and outputting an original gas feature data table. Step two: data preprocessing The data preprocessing module is used for performing noise filtering, standardization and abnormal value detection on the collected original data. Step three: sample balancing The sample balancing module is used for correcting unbalanced data after preprocessing. Step four: model training and classification, which is executed by the intelligent classification module. Step five: result output and evaluation, which is executed by the result output and visualization module, and outputs classification results of each fault type, calculates Precision, Recall, F1 value and AUC performance indicators, visually displays diagnosis results through an interface, and outputs a transformer fault diagnosis report and model performance analysis results.

[0008] Further, step two includes the following steps: (1) data standardization and cleaning The original features x i,j % Z-score standardization is performed to make different gas feature dimensions consistent: wherein, x i,j is a standardized feature value, is a feature mean value, is a standard deviation, is a small constant to prevent the denominator from being zero. (2) Using Huber function clipping to remove abnormal samples: wherein, is a clipping threshold, is a sign function, take the smaller value of the two; (3) Output the balanced training sample set.

[0009] Further, step three includes the following steps: 1) Extract the majority class representative samples using K-Medoid clustering, and remove redundant data; 2) Synthesize new samples in the neighborhood of the minority class using the SMOTE algorithm; 3) Merge the generated new samples with the original minority class samples to obtain the expanded minority class sample set, and combine the majority class sample set to form the training balanced sample set .

[0010] Further, in step 1), the step of using K-Medoid clustering to extract majority class representative samples and remove redundant data includes: First, randomly select samples as initial medoids; then, calculate the distance between the remaining samples and each medoid, and divide each sample into the cluster corresponding to the nearest medoid, and in each cluster, sequentially consider all samples in the cluster as candidate medoids, calculate the total distance of all samples in the cluster to the medoid when the sample is a medoid, and select the sample with the smallest total distance as the new medoid; after completing the above steps, determine whether the medoid has changed: If the medoid position has not changed or the maximum number of iterations has been reached, stop iteration and output the final representative sample set ; otherwise, continue the next iteration until the convergence condition is met.

[0011] Further, in step 2), the step of synthesizing new samples in the neighborhood of the minority class using the SMOTE algorithm includes: First, for each sample in the minority class sample set, calculate the Euclidean distance between it and any other minority class sample : Then sort the distances from small to large, and select the nearest neighbor set of the sample as the nearest neighbor set ; After that, randomly select a sample from the set , as a reference point for interpolation generation; in the sample with his neighbors A location point is randomly selected from the given points as a new synthetic sample. The calculation formula is as follows: in, For the range of values ​​within A random number between these values ​​is used to control the interpolation ratio; when hour, When it overlaps with the original sample, hour, When the value is equal to the neighboring sample, a continuous sample between the two is generated when taking the median value. Repeat the above steps until the number of synthetic minority class samples generated meets the preset expansion ratio or is basically balanced with the number of majority class samples.

[0012] Furthermore, step four includes the following steps: s1 data input and partitioning The training balanced sample set output from step three Input to the intelligent classification module, where, The dataset is divided into training and validation sets proportionally for model training and performance evaluation. S2 Model Establishment and Optimization Objectives Weighted soft-margin support vector machines are used for classification to address the remaining issues of class imbalance and asymmetric loss due to misclassification. The objective function is: The constraints are: in, Let be the normal vector of the hyperplane. For bias terms, It is a nonlinear feature mapping function. As slack variables, , This is the penalty coefficient for positive and negative samples, used to balance the two types of error loss; In its dual form, the decision function of SVM is: in, For kernel function, For Lagrange multipliers; s3 kernel function and parameter selection Radial basis kernel function used: Among them, parameters Controlling the decay rate of similarity between samples; penalty coefficient With kernel parameters All models are automatically optimized through grid search and cross-validation to ensure that they achieve the best generalization ability across different categories. S4 model training Read training samples and labels, and initialize parameters. , Cross-validation is used to calculate classification performance metrics, the optimal parameter combination is selected, and the final model is retrained. The trained model parameters are then used to... Save it for subsequent classification prediction.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1) Two-layer sample balancing mechanism: Combining K-Medoid redundancy removal and SMOTE oversampling to achieve dual optimization of sample quantity and quality; 2) Adaptive parameter optimization of SVM model: Automatically optimizes C and γ parameters through cross-validation to improve generalization ability; 3) Modular design, which can be linked with the power DGA system to realize online intelligent diagnosis. Attached Figure Description

[0014] Figure 1 This is a transformer fault distribution diagram of Embodiment 1 of the present invention. Detailed Implementation

[0015] The present invention will now be described in detail so that its advantages and features can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0016] The following provides a brief overview of one or more aspects to offer a basic understanding of them. This overview is not an exhaustive summary of all conceived aspects, nor is it intended to identify key or decisive elements of all aspects, nor to define the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form to prepare for the more detailed descriptions that follow.

[0017] This invention discloses an intelligent transformer fault diagnosis system based on K-Medoid and SMOTE optimized support vector machine, comprising: The data acquisition module is used to collect dissolved gases (H2, CH4, C2H6, C2H4, C2H2) and operating parameters in transformer oil in real time; The data preprocessing module, connected to the data acquisition module, is used to perform data standardization and outlier detection; The sample balancing module is used to reduce redundancy in the majority class through K-Medoid clustering and to augment the minority class through the SMOTE algorithm. The intelligent classification module uses a weighted SVM classifier for training and prediction. The results output and visualization module is used to display classification results, confidence scores, and key performance indicators. The modules communicate with each other through a unified interface and can be deployed on substation intelligent monitoring terminals or cloud platforms.

[0018] This invention also discloses an intelligent fault diagnosis method for transformers based on K-Medoid and SMOTE optimized support vector machines, comprising the following steps: Step 1: Data Collection The data acquisition module collects dissolved gases (H2, CH4, C2H6, C2H4, C2H2) and operating parameters in transformer oil in real time, forming a raw sample dataset and outputting a raw gas characteristic data table to provide input for the next step.

[0019] Step 2: Data Preprocessing Performed by the data preprocessing module, this process involves noise filtering, standardization, and outlier detection on the collected raw data. Specifically, this includes: (1) Data standardization and cleaning For original features x i,j % Perform Z-score normalization to ensure consistency in the characteristic dimensions of different gases: in, x i,j These are the standardized eigenvalues. The characteristic mean, Standard deviation To prevent small constants with a denominator of zero; (2) Use the Huber function to prune and remove outlier samples: in, , which is the cropping threshold (Huber parameter). For symbolic functions, Take the smaller of the two values; (3) Output a balanced training sample set.

[0020] Step 3: Sample Balancing Performed by the sample balancing module, the imbalance correction is applied to the preprocessed data, specifically including: 1) Use K-Medoid clustering to extract representative samples of the majority class and remove redundant data. This specifically includes the following steps: First, select randomly. The first sample is used as the initial medoid (representative sample). Then, the distances between the remaining samples and each medoid are calculated, and each sample is assigned to the cluster corresponding to the nearest medoid. Within each cluster, all samples in the cluster are considered as candidate medoids, and the sum of the distances from all samples in the cluster to the selected sample when it becomes the medoid is calculated. The sample that minimizes this sum of distances is selected as the new medoid. After completing the above two steps of "sample allocation—medoid update," it is determined whether the medoid has changed: if the medoid position has not changed or the preset maximum number of iterations has been reached, the iteration stops, and the final representative sample set is output. Otherwise, continue with the next iteration until the convergence condition is met. 2) The Synthetic Minority Over-sampling Technique (SMOTE) algorithm is used to synthesize new samples within the minority class neighborhood. The specific process is as follows: First, for each sample in the minority class sample set... Calculate the Euclidean distance between it and other minority class samples: Sort the data by distance from smallest to largest, and select the closest one. Each sample is used as the nearest neighbor set of that sample. ; Then, from the set A sample is randomly selected from the middle. This serves as a reference point for interpolation generation. In the sample... with his neighbors A location point is randomly selected from the given points as a new synthetic sample. The calculation formula is as follows: in, For the range of values ​​within A random number between these values ​​is used to control the interpolation ratio. When hour, Overlapping with the original sample; when hour, It equals the neighboring samples; when taking the median value, it can generate a continuous sample between the two. Repeat the above steps until the number of synthetic minority class samples generated meets the preset expansion ratio or is basically balanced with the number of majority class samples. 3) Merge the newly generated samples with the original minority class samples to obtain an expanded minority class sample set, which, together with the majority class sample set, constitutes the training balanced sample set. This serves as the input data for training the subsequent intelligent classification module.

[0021] Step 4: Model training and classification, performed by the intelligent classification module. After sample balancing, an intelligent classification module is used to train and classify faults on the balanced data based on a weighted support vector machine (SVM). The main function of this module is to learn the nonlinear mapping relationship between gas characteristics and fault types, enabling accurate identification of unknown samples. Its execution process includes the following steps: s1 Data Input and Splitting. The training balanced sample set output from step three... Input to the intelligent classification module, where, Represents a feature vector (such as the gas concentration feature in oil). The corresponding fault category is indicated by the dataset being divided into a training set and a validation set (e.g., 8:2) for model training and performance evaluation. Model S2 Establishment and Optimization Objectives. A weighted soft-margin support vector machine (Weighted SVM) is used for classification to address the remaining class imbalance and asymmetric loss due to misclassification. The objective function is: The constraints are: in, Let be the normal vector of the hyperplane. For bias terms, It is a nonlinear feature mapping function. As slack variables, , where represents the penalty coefficients for positive and negative samples, used to balance the losses from the two types of errors. In the dual form, the decision function of SVM is: in, For kernel function, It is a Lagrange multiplier.

[0022] s3 kernel function and parameter selection. This invention uses the radial basis function (RBF): Among them, parameters Controlling the decay rate of similarity between samples; penalty coefficient With kernel parameters All models are automatically optimized using grid search and cross-validation (k-fold CV) to ensure that they achieve the best generalization ability across different classes. The training process for the S4 model is as follows: Read the training samples and labels, and initialize the parameters. , Cross-validation is used to calculate classification performance metrics, the optimal parameter combination is selected, and the final model is retrained. The trained model parameters are then used to... Save it for subsequent classification prediction.

[0023] Step 5: Result Output and Evaluation, performed by the Result Output and Visualization module. Real-time or test samples are input into the trained model for prediction, outputting classification results for each fault type; and calculating performance metrics such as Precision, Recall, F1 score, and AUC, visually displaying the diagnostic results through the interface. A transformer fault diagnosis report and model performance analysis results are output.

[0024] All of the text x Both subscripts represent samples, and different subscripts represent different samples, such as x i , x j This represents two different samples. x {i,1} Indicates sample x i The first characteristic.

[0025] Example 1 Step 1: Data Collection Experimental data from a power company's transformer fault monitoring system, which monitors a large number of transformers over a long period, were used. This resulted in a rich and comprehensive collection of data, including dissolved gas content in transformer oil and its corresponding fault type labels. Multiple transformer fault types are included in this dataset, such as partial discharge, overheating, and arc discharge. The dissolved gas content in the oil exhibits different characteristics corresponding to different fault types: a relative increase in hydrogen and methane content may be caused by partial discharge faults; significant changes in ethylene and ethane content may originate from overheating faults; and a significant increase in acetylene content often indicates the occurrence of arc discharge faults. The transformer fault distribution map is shown below. Figure 1 As shown.

[0026] This embodiment demonstrates that the number of samples varies significantly across different fault types: the large number of samples for some common fault types is obvious, while the relatively small number of samples for rare fault types is equally apparent. This sample imbalance, which is quite common in actual data collection, presents a significant challenge to the training process and performance evaluation of fault diagnosis models.

[0027] The gas composition and fault type are shown in Table 1.

[0028] Table 1 Step 2: Data Preprocessing The data is normalized to map the range of values ​​to [0,1]. The samples are then cleaned to remove noise and outliers.

[0029] Step 3: Sample Balancing The K-medoid algorithm is used to cluster the samples to find representative sample points for each category. Then, the SMOTE method is used to synthesize minority class samples, increasing the number of minority class samples and making the sample size of each category more balanced.

[0030] Step 4: Model Training and Classification Support vector machine training is implemented using Python's scikit-learn library. A penalty coefficient is defined. nuclear parameters The optimal parameter combination was determined through five-fold cross-validation.

[0031] The dataset was randomly divided into 640 training sets and 160 validation sets. A radial basis function (RBF) kernel was used, and the final parameters were determined as follows: The SVM decision function has 75 support vectors and a bias term. .

[0032] Step 5: Results Output and Evaluation The trained support vector machine model is used to diagnose faults on the test set, and the diagnostic results are obtained.

[0033] Imbalanced sample handling effect. After processing with K-medoid and SMOTE methods, the diagnostic accuracy of the model was significantly improved. Table 2 shows the comparison of the accuracy, recall, and F1 score of the model before and after processing.

[0034] Table 2 As shown in Table 2 above, all performance indicators show a significant improvement after the imbalanced sample handling is completed. Accuracy increased from 75.2% to 90.5% when classifying fault types. Examples show that misclassification is reduced. For the identification of positive class samples, especially minority class fault samples, recall increased from 72.1% to 88.7%, demonstrating a significant improvement in the model after processing. The F1 score, as the harmonic mean, is a comprehensive indicator reflecting classification performance, and its value increased from 73.6% to 89.6%. Therefore, the combined use of K-medoid and SMOTE methods produced the expected results in addressing the imbalanced sample problem. Overall classification performance improvement was achieved through the effective implementation of the above processing steps.

[0035] Any parts or structures not specifically described in this invention can be made using existing technologies or products, and will not be elaborated upon here.

[0036] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A transformer fault intelligent diagnosis system based on K-Medoid and SMOTE optimized support vector machine, characterized in that, include: The data acquisition module is used to collect dissolved gases and operating parameters in transformer oil in real time. The data preprocessing module, connected to the data acquisition module, is used to perform data standardization and outlier detection; The sample balancing module, connected to the data preprocessing module, is used to reduce redundancy in the majority class through K-Medoid clustering and to expand the minority class through the SMOTE algorithm. The intelligent classification module, connected to the sample balancing module, uses a weighted SVM classifier for training and prediction. The results output and visualization module is connected to the intelligent classification module to display classification results, confidence levels, and key performance indicators.

2. A transformer fault intelligent diagnosis method based on K-Medoid and SMOTE optimized support vector machine, characterized in that, The intelligent transformer fault diagnosis system based on K-Medoid and SMOTE optimized support vector machine as described in claim 1 is used for diagnosis, including the following steps: Step 1: Data Collection The data acquisition module performs real-time acquisition of dissolved gases and operating parameters in transformer oil, forming a raw sample dataset and outputting a raw gas characteristic data table. Step 2: Data Preprocessing The data preprocessing module performs noise filtering, standardization, and outlier detection on the collected raw data. Step 3: Sample Balancing This process is performed by the sample balancing module to correct imbalances in the preprocessed data. Step 4: Model training and classification, performed by the intelligent classification module; Step 5: Results Output and Evaluation. This step is performed by the results output and visualization module. It outputs the classification results for each fault type and calculates the Precision, Recall, F1 value, and AUC performance indicators. The diagnostic results are then visualized through the interface, and the transformer fault diagnosis report and model performance analysis results are output.

3. The transformer fault intelligent diagnosis method based on K-Medoid and SMOTE optimized support vector machine according to claim 2, characterized in that, Step two includes the following steps: (1) Data standardization and cleaning On the original features x i,j % Perform Z-score standardization to make different gas feature dimensions consistent: wherein, x i,j is the standardized feature value, is the feature mean value, is the standard deviation, is a small constant to prevent the denominator from being zero; (2) Use the Huber function to prune and remove outlier samples: wherein is a clipping threshold, is a sign function, takes the smaller value of both. (3) Output a balanced training sample set.

4. The transformer fault intelligent diagnosis method based on K-Medoid and SMOTE optimized support vector machine according to claim 2, characterized in that, Step three includes the following steps: 1) Use K-Medoid clustering to extract representative samples of the majority class and remove redundant data; 2) New samples are synthesized within the minority class neighborhood using the SMOTE algorithm; 3) merge the generated new samples with the original minority class samples to obtain an expanded minority class sample set, which, together with the majority class sample set, forms a balanced training sample set .

5. The transformer fault intelligent diagnosis method based on K-Medoid and SMOTE optimized support vector machine according to claim 4, characterized in that, Step 1), which involves using K-Medoid clustering to extract representative samples of the majority class and remove redundant data, includes the following steps: First, randomly select samples as initial medoids; then, calculate the distance between the rest of the samples and each medoid, and divide each sample into the cluster corresponding to the nearest medoid, and in each cluster, in turn, all samples in the cluster are regarded as candidate medoids, the distance sum of all samples in the cluster to the sample as medoid is calculated, and the sample with the minimum distance sum is selected as the new medoid; after the above steps, it is judged whether the medoid has changed: If the medoid position is not changed or reaches the preset maximum iteration number, the iteration is stopped, and the final representative sample set is output ; otherwise, the next round of iteration is continued until the convergence condition is met.

6. The transformer fault intelligent diagnosis method based on K-Medoid and SMOTE optimized support vector machine according to claim 4, characterized in that, Step 2), the steps of synthesizing new samples in the minority class neighborhood using the SMOTE algorithm include: First, for each sample in the minority class sample set , the Euclidean distance between it and any other minority class sample is computed: Then according to the distance from small to large order, select the nearest distance of the sample as the nearest neighbor set of the sample ;​ Then from the set A sample is randomly selected from the middle. , as a reference point for interpolation generation; in the sample with his neighbors A location point is randomly selected from the given points as a new synthetic sample. The calculation formula is as follows: wherein, is a random number with a value range between is used to control the interpolation ratio; when , coincides with the original sample, when , is equal to the neighboring sample, and the intermediate value is generated to generate a continuous sample between the two. Repeat the above steps until the number of synthetic minority class samples generated meets the preset expansion ratio or is basically balanced with the number of majority class samples.

7. The transformer fault intelligent diagnosis method based on K-Medoid and SMOTE optimized support vector machine according to claim 1, characterized in that, Step four includes the following steps: s1 data input and partitioning the training balanced sample set output from step three inputting the intelligent classification module, wherein the data set is proportionally divided into a training set and a validation set for model training and performance evaluation; S2 Model Establishment and Optimization Objectives Weighted soft-margin support vector machines are used for classification to address the remaining issues of class imbalance and asymmetric loss due to misclassification. The objective function is: The constraints are: wherein, is a hyperplane normal vector, is a bias term, is a nonlinear feature mapping function, is a slack variable, , are penalty coefficients for positive and negative samples, respectively, to balance the loss of two classes. In its dual form, the decision function of SVM is: wherein is a kernel function, is a Lagrange multiplier; s3 kernel function and parameter selection Radial basis kernel function used: where the parameters controlling the decay rate of the similarity between samples; penalty coefficient and the kernel parameters are automatically optimized by grid search and cross-validation to ensure the best generalization ability of the model between different categories; S4 model training Read training samples and labels, and initialize parameters. , Cross-validation is used to calculate classification performance metrics, the optimal parameter combination is selected, and the final model is retrained. The trained model parameters are then used to... Save it for subsequent classification prediction.

Images