Fault diagnosis method for voltage transformer

Abstract

The invention relates to the technical field of voltage transformer fault diagnosis, in particular to a voltage transformer fault diagnosis method, which comprises the following steps of data acquisition, data preprocessing, feature selection, fault classification, model evaluation and practical application. According to the method, short-time Fourier transform, a maximum information coefficient, a convolutional neural network and a long short-term memory network are effectively combined by adopting a predictive maintenance method based on hybrid driving, so that fault diagnosis of a capacitor breakdown fault, a turn-to-turn short circuit fault, a ferromagnetic resonance fault and an oil leakage fault of the voltage transformer is realized; and in the data preprocessing stage, a simulated three-phase voltage signal is calculated through a discrete time model formula, the real three-phase voltage signal is replaced to a certain extent, and abnormal points are manufactured on the generated data, so that the problem of less three-phase voltage signal data is solved, and the diagnosis accuracy is improved.

Description

Technical Field

[0001] The invention relates to the technical field of voltage transformer fault diagnosis, and in particular to a voltage transformer fault diagnosis method. Background Art

[0002] Capacitor voltage transformer is a special voltage transformer. It divides the voltage through series capacitors, and then reduces the voltage and isolates it through electromagnetic transformers to provide voltage signals for meters, relay protection, etc. Capacitor voltage transformer has the advantages of high insulation strength, small size and light weight. Therefore, in high-voltage systems, this type of voltage transformer is a more suitable choice.

[0003] In order to ensure the normal operation of the capacitor voltage transformer and the stability of the power grid, the staff needs to regularly perform fault diagnosis on the capacitor voltage transformer, which is time-consuming and labor-intensive. Therefore, there are capacitor voltage transformer fault diagnosis methods and devices disclosed in the invention patent with publication number CN117726910A and a voltage transformer fault diagnosis method, equipment and storage medium based on data enhancement disclosed in the invention patent with publication number CN117171616A, etc., to perform fault diagnosis on the capacitor voltage transformer.

[0004] However, during the diagnosis process, it was found that the existing capacitive voltage transformer fault diagnosis method has low diagnostic accuracy and is inconvenient to obtain three-phase voltage signal data, which also affects the diagnostic accuracy and leads to poor practicality. Therefore, a voltage transformer fault diagnosis method is urgently needed to improve the above problems. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a predictive maintenance method based on hybrid drive, which effectively combines short-time Fourier transform, maximum information coefficient, convolutional neural network and long short-term memory network to realize fault diagnosis of capacitor breakdown fault, inter-turn short circuit fault, ferromagnetic resonance fault and oil leakage fault of voltage transformer, and in the data preprocessing stage, calculates the simulated three-phase voltage signal through discrete time model formula, replaces the real three-phase voltage signal to a certain extent, and creates abnormal points on the generated data, solves the problem of less three-phase voltage signal data, and improves the diagnosis accuracy. A fault diagnosis method for voltage transformer.

[0006] A fault diagnosis method for a voltage transformer of the present invention comprises the following steps:

[0007] Step 1: Data acquisition:

[0008] Obtaining voltage transformer data and three-phase voltage signal data;

[0009] Step 2: Data preprocessing:

[0010] The simulated three-phase voltage signal is calculated through the discrete time model formula to replace part of the real three-phase voltage signal, and abnormal points are created on the generated data. The data is then labeled and encoded. Since the proportion of fault type data in the entire data set is relatively scarce, the SMOTE sampling strategy is used to balance it.

[0011] Step 3: Feature selection:

[0012] The data were screened using short-time Fourier transform and maximum information coefficient;

[0013] Step 4: Fault classification: Build a CNN-LSTM model architecture to handle the fault diagnosis task here, and finally use the softmax activation function for multi-classification to calculate the probability distribution of the fault category label;

[0014] Step 4: Model evaluation:

[0015] After model training and prediction are completed, the performance of the model is evaluated through confusion matrix, classification report and ROC curve;

[0016] Step 5: Practical application:

[0017] After the model evaluation is passed, the model is applied to the capacitor voltage transformer to perform store diagnosis on the capacitor voltage transformer, which is convenient for the staff to perform predictive maintenance on the capacitor voltage transformer;

[0018] By adopting a predictive maintenance method based on hybrid drive, short-time Fourier transform, maximum information coefficient, convolutional neural network and long short-term memory network are effectively combined to realize fault diagnosis of capacitor breakdown fault, inter-turn short circuit fault, ferromagnetic resonance fault and oil leakage fault of voltage transformer. In the data preprocessing stage, the simulated three-phase voltage signal is calculated by discrete time model formula, which replaces the real three-phase voltage signal to a certain extent, and creates abnormal points on the generated data to solve the problem of less three-phase voltage signal data and improve the diagnosis accuracy.

[0019] Preferably, the data preprocessing specifically includes the following steps:

[0020] S1. Select an appropriate discrete-time model formula;

[0021] S2, generating abnormal points, by adjusting the amplitude of the three-phase voltage fundamental component, adding noise and signal phase, to create abnormal points of the three-phase voltage signal;

[0022] S3. Label and encode the data. Since the proportion of fault type data in the entire data set is relatively scarce, the SMOTE sampling strategy is used to balance it.

[0023] Preferably, the selected discrete time model formula is as follows:

[0024]

[0025] Where V am , V bm , V cm is the amplitude of the fundamental component of the three-phase voltage, a N is the amplitude ratio of the Nth harmonic to the fundamental wave, ε ak , ε bk , ε ck is the noise term, ω represents the angular frequency of the sinusoidal signal, is the signal phase, T s is the sampling interval, and k is the sampling time.

[0026] Preferably, the specific method of using the SMOTE sampling strategy for balancing includes:

[0027] The normal data and the remaining four types of fault data are labeled with the letters "AE" respectively, and then digital encoding is used to convert different labels into numbers "0-4". Then SMOTE is further used to synthesize new minority class abnormal samples to increase the number of our minority classes and balance the number of samples in each category. This is beneficial for subsequent models to diagnose and identify data of these fault types with only a small number of samples, thereby improving the accuracy.

[0028] Preferably, the specific method of feature selection includes:

[0029] Dividing the three-phase voltage signal into several small segments and applying Fourier transform on each segment can simultaneously obtain the time and frequency information of the signal, improve the recognition ability of the subsequent fault diagnosis model, and apply the maximum information coefficient to find the relationship between the amplitude of the three-phase voltage signal and the fault label, the relationship between the voltage ratio difference and the fault label, and the relationship between the phase angle difference and the fault label.

[0030] Preferably, the specific method of the fault diagnosis includes:

[0031] Get the result of the previous feature extraction part, and input the three-phase voltage signal after STFT conversion into the CNN part of the model. Here, the input is a four-dimensional tensor. In the convolution layer of the CNN model, set the convolution kernel size to (3,3), and use Conv2D to extract the local features of the frequency domain signal. In order to prevent overfitting, Dropout is added to discard a certain proportion of neurons to improve the generalization ability of the model. Then the output of the convolution layer is flattened, and the features are further extracted through neurons in the fully connected layer. The remaining time series data filtered by MIC is handed over to the LSTM part for processing. Here, a three-dimensional tensor is input, and then a certain number of units are set in the LSTM layer to capture long-term dependencies. Dropout is also used to prevent overfitting. Finally, the outputs of CNN and LSTM are merged to form a new feature display. In this way, the frequency domain and time series features are considered at the same time, and the softmax activation function is used for multi-classification to calculate the probability distribution of the fault category label.

[0032] Preferably, the data obtained from the voltage transformer includes: load data, primary side voltage data and specification attribute data of the voltage transformer itself.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. By adopting a predictive maintenance method based on hybrid drive, short-time Fourier transform, maximum information coefficient, convolutional neural network and long short-term memory network are effectively combined to realize fault diagnosis of capacitor breakdown fault, turn-to-turn short circuit fault, ferromagnetic resonance fault and oil leakage fault of voltage transformer;

[0035] 2. In the data preprocessing stage, the simulated three-phase voltage signal is calculated through the discrete time model formula, which replaces the real three-phase voltage signal to a certain extent, and creates abnormal points on the generated data to solve the problem of less three-phase voltage signal data and improve the diagnostic accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the fault diagnosis process of the voltage transformer of the present invention;

[0037] Figure 2 It is a schematic flow chart of the data preprocessing steps of the present invention;

[0038] Figure 3 It is a schematic diagram of the feature selection process of the present invention;

[0039] Figure 4 It is a schematic diagram of the fault diagnosis process of the present invention;

[0040] Figure 5It is a confusion matrix diagram of the present invention;

[0041] Figure 6 It is the classification report diagram of the present invention;

[0042] Figure 7 It is the ROC curve diagram of the present invention. DETAILED DESCRIPTION

[0043] In order to facilitate understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. The present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0044] Example

[0045] like Figures 1 to 7 As shown, a fault diagnosis method for a voltage transformer comprises the following steps:

[0046] Step 1: Data acquisition:

[0047] Obtaining voltage transformer data and three-phase voltage signal data;

[0048] Step 2: Data preprocessing:

[0049] The simulated three-phase voltage signal is calculated through the discrete time model formula to replace part of the real three-phase voltage signal, and abnormal points are created on the generated data. The data is then labeled and encoded. Since the proportion of fault type data in the entire data set is relatively scarce, the SMOTE sampling strategy is used to balance it.

[0050] Step 3: Feature selection:

[0051] The data were screened using short-time Fourier transform and maximum information coefficient;

[0052] Step 4: Fault classification: Build a CNN-LSTM model architecture to handle the fault diagnosis task here, and finally use the softmax activation function for multi-classification to calculate the probability distribution of the fault category label;

[0053] Step 4: Model evaluation:

[0054] After model training and prediction are completed, the performance of the model is evaluated through confusion matrix, classification report and ROC curve;

[0055] Step 5: Practical application:

[0056] After the model evaluation is passed, the model is applied to the capacitor voltage transformer to perform store diagnosis on the capacitor voltage transformer, which is convenient for the staff to perform predictive maintenance on the capacitor voltage transformer;

[0057] The data preprocessing specifically includes the following steps:

[0058] S1. Select an appropriate discrete-time model formula;

[0059] S2, generating abnormal points, by adjusting the amplitude of the three-phase voltage fundamental component, adding noise and signal phase, to create abnormal points of the three-phase voltage signal;

[0060] S3: Label and encode the data. Since the proportion of fault type data in the entire data set is relatively scarce, the SMOTE sampling strategy is used to balance it.

[0061] The discrete time model formula of the selection is as follows:

[0062]

[0063] Where V am , V bm , V cm is the amplitude of the fundamental component of the three-phase voltage, a N is the amplitude ratio of the Nth harmonic to the fundamental wave, ε ak , ε bk , ε ck is the noise term, ω represents the angular frequency of the sinusoidal signal, is the signal phase, T s is the sampling interval, k is the sampling time;

[0064] The specific approach of using the SMOTE sampling strategy for balancing includes:

[0065] The normal data and the remaining four types of fault data are labeled with the letters "AE" respectively, and then digital encoding is used to convert different labels into numbers "0-4". Then SMOTE is further used to synthesize new minority class abnormal samples to increase the number of our minority classes and balance the number of samples in each category. This is conducive to the subsequent model's diagnosis and recognition of these fault type data with only a small number of samples, and improves the accuracy;

[0066] The specific method of feature selection includes:

[0067] Dividing the three-phase voltage signal into several small segments and applying Fourier transform on each segment can simultaneously obtain the time and frequency information of the signal, improve the recognition ability of the subsequent fault diagnosis model, and apply the maximum information coefficient to find the relationship between the amplitude of the three-phase voltage signal and the fault label, the relationship between the voltage ratio difference and the fault label, and the relationship between the phase angle difference and the fault label;

[0068] The specific methods of the fault diagnosis include:

[0069] Get the result of the previous feature extraction part, input the three-phase voltage signal after STFT conversion into the CNN part of the model, where the input is a four-dimensional tensor. In the convolution layer of the CNN model, set the convolution kernel size to (3,3), and use Conv2D to extract the local features of the frequency domain signal. In order to prevent overfitting, Dropout is added to discard a certain proportion of neurons to improve the generalization ability of the model, and then the output of the convolution layer is flattened. In the fully connected layer, the features are further extracted through neurons, and the remaining time series data filtered by MIC is handed over to the LSTM part for processing. Here, a three-dimensional tensor is input, and then a certain number of units are set in the LSTM layer to capture long-term dependencies. Dropout is also used to prevent overfitting. Finally, the outputs of CNN and LSTM are merged to form a new feature display. In this way, the frequency domain and time series features are considered at the same time, and then the softmax activation function is used for multi-classification to calculate the probability distribution of the fault category label.

[0070] The obtaining of data of the voltage transformer comprises:

[0071] Load data, primary side voltage data and specification attribute data of the voltage transformer itself.

[0072] Experimental example

[0073] A case analysis is conducted using sample data from a capacitive voltage transformer device in a 110kv third-generation smart substation;

[0074] The original data set was obtained from the historical detection data of the sensors of the substation, and some data was extracted. In addition, some three-phase voltage data was simulated and generated. Ten repeated tests were conducted, each with 1000 sample data, including data of healthy status, data with capacitor breakdown, inter-turn short circuit, oil leakage and ferromagnetic resonance. The distribution of each sample data is shown in Table 1 below:

[0075] Table 1 Data status statistics in the dataset

[0076] Data No. Current transformer status Sample size A Health status 800 B Capacitor breakdown 40 C Turn-to-turn short circuit 40 D Oil leak 80 E Ferromagnetic resonance 40

[0077] After data preprocessing, each piece of data includes

[0078] V ak 、V bk 、V ck , primary_voltage, oil_temp, HUFL, HULL, MUFL, MULL, LUFL, LULL, and Label attribute columns.

[0079] as follows Figure 5 As shown in FIG. 1 , the ROC curve results are obtained after the model is trained on the training set using the predictive maintenance technology and the classification prediction performance of the trained model is tested on the test set.

[0080] like Figure 5 As shown, in order to comprehensively evaluate the model and further understand the classification effect of the model, a confusion matrix is ​​drawn.

[0081] The final evaluation report is as follows: Figure 6 As shown;

[0082] Among them, A represents the healthy state, B represents the capacitor breakdown fault state, C represents the turn-to-turn short circuit fault state, D represents the oil leakage fault state, and E represents the ferromagnetic resonance fault state;

[0083] Analysis of the above results shows that the overall average prediction rate of the method proposed in this patent has reached 94%, especially the recognition effect of data in state A is very accurate, the accuracy of data recognition in state B is accurate, without misjudgment, and the recognition of data in state C is also relatively accurate. A good balance is achieved between accuracy and coverage, among which the coverage of state D is relatively high, reaching 95%, indicating that most samples can be correctly classified.

[0084] The above analysis combined with the example data shows that the fault diagnosis method for capacitive voltage transformers can effectively identify and predict whether the equipment has a fault based on historical status monitoring data to a certain extent.

[0085] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A fault diagnosis method for a voltage transformer, characterized in that: The following steps are involved: Step 1: Data acquisition: Obtaining voltage transformer data and three-phase voltage signal data; Step 2: Data preprocessing: The simulated three-phase voltage signal is calculated through the discrete time model formula to replace part of the real three-phase voltage signal, and abnormal points are created on the generated data. The data is then labeled and encoded. Since the proportion of fault type data in the entire data set is relatively scarce, the SMOTE sampling strategy is used to balance it. Step 3: Feature selection: The data were screened using short-time Fourier transform and maximum information coefficient; Step 4: Fault classification: Build a CNN-LSTM model architecture to handle the fault diagnosis task here, and finally use the softmax activation function for multi-classification to calculate the probability distribution of the fault category label; Step 4: Model evaluation: After model training and prediction are completed, the performance of the model is evaluated through confusion matrix, classification report and ROC curve; Step 5: Practical application: After the model evaluation is passed, the model is applied to the capacitor voltage transformer to perform store diagnosis on the capacitor voltage transformer, which is convenient for the staff to perform predictive maintenance on the capacitor voltage transformer.

2. A fault diagnosis method for a voltage transformer as claimed in claim 1, characterized in that: The data preprocessing specifically includes the following steps: S1. Select an appropriate discrete-time model formula; S2, generating abnormal points, by adjusting the amplitude of the three-phase voltage fundamental component, adding noise and signal phase, to create abnormal points of the three-phase voltage signal; S3. Label and encode the data. Since the proportion of fault type data in the entire data set is relatively scarce, the SMOTE sampling strategy is used to balance it.

3. A fault diagnosis method for a voltage transformer as claimed in claim 2, characterized in that: The discrete time model formula of the selection is as follows: Where V am , V bm , V cm is the amplitude of the fundamental component of the three-phase voltage, a N is the amplitude ratio of the Nth harmonic to the fundamental wave, ε ak , ε bk , ε ck is the noise term, ω represents the angular frequency of the sinusoidal signal, is the signal phase, T s is the sampling interval, and k is the sampling time.

4. A fault diagnosis method for a voltage transformer as claimed in claim 2, characterized in that: The specific approach of using the SMOTE sampling strategy for balancing includes: The normal data and the remaining four types of fault data are labeled with the letters "AE" respectively, and then digital encoding is used to convert different labels into numbers "0-4". Then SMOTE is further used to synthesize new minority class abnormal samples to increase the number of our minority classes and balance the number of samples in each category. This is beneficial for subsequent models to diagnose and identify data of these fault types with only a small number of samples, thereby improving the accuracy.

5. A fault diagnosis method for a voltage transformer as claimed in claim 1, characterized in that: The specific method of feature selection includes: Dividing the three-phase voltage signal into several small segments and applying Fourier transform on each segment can simultaneously obtain the time and frequency information of the signal, improve the recognition ability of the subsequent fault diagnosis model, and apply the maximum information coefficient to find the relationship between the amplitude of the three-phase voltage signal and the fault label, the relationship between the voltage ratio difference and the fault label, and the relationship between the phase angle difference and the fault label.

6. A fault diagnosis method for a voltage transformer as claimed in claim 1, characterized in that: The specific methods of the fault diagnosis include: Get the result of the previous feature extraction part, and input the three-phase voltage signal after STFT conversion into the CNN part of the model. Here, the input is a four-dimensional tensor. In the convolution layer of the CNN model, set the convolution kernel size to (3,3), and use Conv2D to extract the local features of the frequency domain signal. In order to prevent overfitting, Dropout is added to discard a certain proportion of neurons to improve the generalization ability of the model. Then the output of the convolution layer is flattened, and the features are further extracted through neurons in the fully connected layer. The remaining time series data filtered by MIC is handed over to the LSTM part for processing. Here, a three-dimensional tensor is input, and then a certain number of units are set in the LSTM layer to capture long-term dependencies. Dropout is also used to prevent overfitting. Finally, the outputs of CNN and LSTM are merged to form a new feature display. In this way, the frequency domain and time series features are considered at the same time, and the softmax activation function is used for multi-classification to calculate the probability distribution of the fault category label.

7. A fault diagnosis method for a voltage transformer as claimed in claim 1, characterized in that: The data obtained from the voltage transformer include: load data, primary side voltage data and specification attribute data of the voltage transformer itself.

Images