Hydropower station transformer fault diagnosis algorithm based on multi-channel neural network multi-source heterogeneous data feature fusion

Abstract

The invention discloses a hydropower station transformer fault diagnosis algorithm based on multi-channel neural network multi-source heterogeneous data feature fusion, and the method comprises the steps: collecting one-dimensional data, two-dimensional data and other multi-source heterogeneous data related to a transformer, and deleting the irrelevant data through a variance elimination method; the method comprises the following steps of: further screening data by using a Pearson correlation coefficient, classifying the screened data, respectively establishing adaptive models for different types of data, extracting data features, fusing the data features of different dimensions, continuously extracting the fused data features by using a deep residual network, and extracting the fused data features. And finally outputting a diagnosis result. The problems that in current hydropower station transformer fault detection, faults are affected by various parameters, fault data are multi-source heterogeneous data, and an independent fault analysis method is low in accuracy and one-sided in detection range are solved, and the fault diagnosis comprehensiveness and accuracy can be improved.

Description

Technical Field

[0001] The invention belongs to the technical field of hydropower station fault detection, and in particular relates to a hydropower station transformer fault diagnosis algorithm based on multi-channel neural network multi-source heterogeneous data feature fusion. Background Art

[0002] Power transformers are one of the key electrical equipment in hydropower stations. They undertake important tasks such as voltage level conversion and power transmission and distribution. Their operating status will directly affect the transmission of power in hydropower stations. For hydropower stations, power transformer fault diagnosis is particularly important. Currently, common transformer fault diagnosis methods include dissolved gas analysis in oil, sound-based transformer fault diagnosis methods, and artificial intelligence-based fault diagnosis methods.

[0003] The dissolved gas analysis (DGA) method detects the dissolved gas content (mainly H 2 , CH 4 , C 2 H 6 , C 2 H 4 , C 2 H 2 ) is a diagnostic method to analyze the operating status of the transformer. Although these methods are simple to operate, they are susceptible to limitations such as missing codes and absolute thresholds, making it difficult to determine the fault location of the transformer and the accuracy of fault diagnosis is limited.

[0004] The sound-based transformer fault diagnosis method analyzes the noise, ultrasonic waves and diagnostic signals in the transformer and establishes a typical spectrum for fault diagnosis. This method based on sound analysis is easily interfered by environmental sound sources. At the same time, the complex internal structure of the transformer leads to a low accuracy rate of the sound-based fault diagnosis method.

[0005] The transformer fault diagnosis method based on artificial intelligence collects and processes relevant data of the transformer, establishes and trains a fault diagnosis model to diagnose the transformer fault; however, in the existing methods, the model is often only established based on a certain type of data, and the data type is single and the feature extraction is incomplete, resulting in low fault diagnosis accuracy and inability to accurately locate the fault site.

[0006] Therefore, it is necessary to design a hydropower station transformer fault diagnosis algorithm based on multi-channel neural network and multi-source heterogeneous data feature fusion to solve the above problems. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a hydropower station transformer fault diagnosis algorithm based on multi-channel neural network multi-source heterogeneous data feature fusion. The method aims to solve the current hydropower station transformer fault detection, in which the fault is affected by multiple parameters, the fault data is multi-source heterogeneous data, and the single fault analysis method has the problems of low accuracy and one-sided detection range. The method has the characteristics of being able to obtain multiple parameters related to transformer faults, obtaining data from multiple dimensions, selecting different neural networks to extract corresponding features according to the type of data, and being able to adaptively adjust the weights of various parameters.

[0008] In order to achieve the above technical effects, the technical solution adopted by the present invention is: A fault diagnosis algorithm for a hydropower station transformer based on multi-channel neural network and multi-source heterogeneous data feature fusion includes the following steps: S1, Data collection: Collect multi-source heterogeneous data, including one-dimensional data and two-dimensional data; One-dimensional data includes time series data and non-time series data. Time series data includes gas content in oil, temperature data, humidity data, current data and vibration data. Non-time series data includes temperature, humidity and air density of the transformer operating environment. The two-dimensional data includes infrared image data; More comprehensive data collection can effectively ensure more comprehensive feature extraction; S2, data preprocessing: The variance elimination method deletes data that is not related to the fault diagnosis results, calculates the variance of all input parameters, and deletes parameters with zero variance. A variance of zero means that the parameter has not changed in all categories. The variance elimination method formula is: ; In the formula, S 2 is the sample variance, X is the sample data, is the sample mean, N is the number of samples; calculate the Pearson correlation coefficient between each parameter and the fault diagnosis result, and retain the parameters with higher correlation; Find the absolute value of the Pearson correlation coefficient for each parameter: ; In the formula, x represents sample data, y represents fault label, n is the total number of samples, and subscript i represents the i-th sample; Taking the mean as the threshold, the absolute value of the Pearson coefficient of each parameter is compared with the threshold, and the parameters with an absolute value of the Pearson coefficient higher than the threshold are retained; the data is normalized to the mean and the data set is divided into a training set and a test set in a ratio of 8:2. The mean calculation method is: ; in is the data obtained after mean normalization, x is the sample data, u is the sample mean, and S is the standard deviation; Deleting some parameters with less correlation can improve the model training speed and fault diagnosis accuracy; S3, build a fault diagnosis model; S4, perform fault diagnosis.

[0009] Preferably, in step S3, building a fault diagnosis model includes: Construct a compressed excitation network SEnet, which includes two one-dimensional compressed excitation networks 1DSEnet and one two-dimensional compressed excitation network 2DSEnet; The one-dimensional data consisting of the preprocessed time series data and non-time series data is used as the input of 1DSEnet, and the two-dimensional data consisting of the preprocessed infrared image data is used as the input of 2DSEnet; Use SEnet to adaptively adjust the weights of each input parameter and replace the relu function in the SEnet fully connected layer with the tanh function: ; ; In the formula, x is the input sample; in relu, features less than 0 will be set to 0, which will inactivate some neurons and cause the loss of data features, while the use of tanh function can ensure the integrity of data features; Three neural networks, namely one-dimensional convolutional neural network 1DCNN, two-dimensional convolutional neural network 2DCNN and gated recurrent unit GRU, are established respectively, and the time series data, non-time series data and two-dimensional data after weight adjustment are used as the input of GRU, 1DCNN and 2DCNN respectively; among them, 1DCNN is good at processing non-time series data, 2DCNN can process two-dimensional image data, and GRU is good at processing time series data; Improve GRU, use bidirectional GRU for feature extraction, introduce dropout layer to prevent bidirectional GRU from overfitting, introduce spatially separable convolution in 2DCNN to improve network structure, reduce model training parameters, and increase training speed; The bidirectional GRU is used to extract the temporal features in the data, the 1DCNN is used to extract the non-temporal features in the data, and the 2DCNN is used to extract the two-dimensional image features in the data. The extracted features of the three dimensions are accumulated for feature fusion: All data features are one-dimensionalized and used as the input of the fully connected layer. Multiple fully connected layers are used to extract the fused data features. The residual neural network ResNet is introduced to improve the DNN network structure, and the DNN is connected across layers to avoid the loss of early data features. At the same time, deeper features can be extracted. The batch normalization layer Batch Normalization is introduced to standardize the distribution of data features. Finally, the data features are input into the classifier for fault diagnosis.

[0010] Preferably, in step S4, performing fault diagnosis includes performing fault diagnosis through a classifier in the model constructed in step S3, and the specific method is: The fault categories are divided into overheating faults, discharge faults and mechanism faults, and the specific locations corresponding to the fault categories are located; The fault data is labeled, the fault type is encoded using the One-Hot method, the categorical variables are represented as binary vectors, and the discrete fault categories are expanded to the Euclidean space, which is beneficial for fault identification.

[0011] The model constructed in step S3 is trained using the labeled data to obtain a trained model, and then the unlabeled data is used as the input of the model, and finally the category corresponding to the data is input, so that the fault category is more comprehensive and the fault location is more accurately located.

[0012] Furthermore, each type of fault can be located at the location where the fault occurs. For example, an overheating fault can be located at the current loop, iron core, and switch, and a discharge fault can be located at the switch, winding, and iron core.

[0013] The beneficial effects of the present invention are as follows: 1. Using variance elimination method and Pearson correlation coefficient to screen data can reduce the impact of irrelevant data on diagnostic results. Using SEnet to process data and adaptively adjust the weights of each parameter can help the model focus on more important parameters.

[0014] 2. The input of the algorithm is multi-source heterogeneous data. Using appropriate models to extract features from different types of data can make the extracted features more comprehensive and accurate.

[0015] 3. Fuse different types of features and use a deep residual network to extract the fused data features to further improve the accuracy of fault diagnosis.

[0016] 4. Locate the fault type to a specific location so that the location of the fault can be determined during fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of a flow chart in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] Embodiment 1: A fault diagnosis algorithm for a hydropower station transformer based on multi-channel neural network and multi-source heterogeneous data feature fusion includes the following steps: S1, Data collection: Collect multi-source heterogeneous data, including one-dimensional data and two-dimensional data; One-dimensional data includes time series data and non-time series data. Time series data includes gas content in oil, temperature data, humidity data, current data and vibration data. Non-time series data includes temperature, humidity and air density of the transformer operating environment. The two-dimensional data includes infrared image data; More comprehensive data collection can effectively ensure more comprehensive feature extraction; S2, data preprocessing: S3, build a fault diagnosis model; S4, perform fault diagnosis.

[0019] Embodiment 2: like Figure 1 As shown, the specific implementation process of the present invention includes: 1. Collect multi-source heterogeneous data and perform data labeling; 2. Use variance elimination method to remove irrelevant data; 3. Use Pearson correlation coefficient to filter the data; 4. Normalize the data of different dimensions respectively; 5. Establish adaptive adjustment parameter weights for 1DSEnet and 2DSEnet respectively; 6. Establish bidirectional GRU, 1DCNN, and 2DCNN respectively. The input of the model is time series data, non-time series data, and two-dimensional image data respectively; 7. Introduce the dropout layer in the bidirectional GRU model to prevent the bidirectional GRU from overfitting; 8. Introducing spatially separable convolution into 2DCNN to reduce the training parameters of the model and improve the training speed; 9. Accumulate and fuse the features output by the three models; 10. Build a deep residual neural network to learn the fused data features; 11. Output fault diagnosis results.

[0020] Data preprocessing using variance elimination to remove irrelevant data includes: The variance elimination method deletes data that is not related to the fault diagnosis results, calculates the variance of all input parameters, and deletes parameters with zero variance. A variance of zero means that the parameter has not changed in all categories. The variance elimination method formula is: ; In the formula, S 2 is the sample variance, X is the sample data, is the sample mean, N is the number of samples; calculate the Pearson correlation coefficient between each parameter and the fault diagnosis result, and retain the parameters with higher correlation; Find the absolute value of the Pearson correlation coefficient for each parameter: ; In the formula, x represents sample data, y represents fault label, n is the total number of samples, and subscript i represents the i-th sample; Taking the mean as the threshold, the absolute value of the Pearson coefficient of each parameter is compared with the threshold, and the parameters with an absolute value of the Pearson coefficient higher than the threshold are retained; the data is normalized to the mean and the data set is divided into a training set and a test set in a ratio of 8:2. The mean calculation method is: ; in is the data obtained after mean normalization, x is the sample data, u is the sample mean, and S is the standard deviation; Deleting some parameters with less correlation can improve the model training speed and fault diagnosis accuracy; Building a fault diagnosis model includes: Construct a compressed excitation network SEnet, which includes two one-dimensional compressed excitation networks 1DSEnet and one two-dimensional compressed excitation network 2DSEnet; The one-dimensional data consisting of the preprocessed time series data and non-time series data is used as the input of 1DSEnet, and the two-dimensional data consisting of the preprocessed infrared image data is used as the input of 2DSEnet; Use SEnet to adaptively adjust the weights of each input parameter and replace the relu function in the SEnet fully connected layer with the tanh function: ; ; In the formula, x is the input sample; in relu, features less than 0 will be set to 0, which will inactivate some neurons and cause the loss of data features, while the use of tanh function can ensure the integrity of data features; Three neural networks, namely one-dimensional convolutional neural network 1DCNN, two-dimensional convolutional neural network 2DCNN and gated recurrent unit GRU, are established respectively, and the time series data, non-time series data and two-dimensional data after weight adjustment are used as the input of GRU, 1DCNN and 2DCNN respectively; among them, 1DCNN is good at processing non-time series data, 2DCNN can process two-dimensional image data, and GRU is good at processing time series data; Improve GRU, use bidirectional GRU for feature extraction, introduce dropout layer to prevent bidirectional GRU from overfitting, introduce spatially separable convolution in 2DCNN to improve network structure, reduce model training parameters, and increase training speed; The bidirectional GRU is used to extract the temporal features in the data, the 1DCNN is used to extract the non-temporal features in the data, and the 2DCNN is used to extract the two-dimensional image features in the data. The extracted features of the three dimensions are accumulated for feature fusion: All data features are one-dimensionalized and used as the input of the fully connected layer. Multiple fully connected layers are used to extract the fused data features. The residual neural network ResNet is introduced to improve the DNN network structure, and the DNN is connected across layers to avoid the loss of early data features. At the same time, deeper features can be extracted. The batch normalization layer Batch Normalization is introduced to standardize the distribution of data features. Finally, the data features are input into the classifier for fault diagnosis.

[0021] Fault diagnosis includes fault diagnosis through the classifier in the constructed model. The specific method is as follows: The fault categories are divided into overheating faults, discharge faults and mechanism faults, and the specific locations corresponding to the fault categories are located; The fault data is labeled, the fault type is encoded using the One-Hot method, the categorical variable is represented as a binary vector, and the discrete fault category is expanded to the Euclidean space, which is conducive to fault identification; the embodiment includes a total of 10 fault categories and 1 normal state, such as Figure 1 As shown, they are normal state, current loop overheat fault, iron core overheat fault, switch overheat fault, winding partial discharge fault, iron core partial discharge fault, oil discharge fault, switch discharge fault, cooling system fault, winding resonance fault and insulating oil fault.

[0022] The model constructed in step S3 is trained using the labeled data to obtain a trained model, and then the unlabeled data is used as the input of the model, and finally the category corresponding to the data is input, so that the fault category is more comprehensive and the fault location is more accurately located.

[0023] Furthermore, each type of fault can be located at the location where the fault occurs. For example, an overheating fault can be located at the current loop, iron core, and switch, and a discharge fault can be located at the switch, winding, and iron core.

Claims

1. A fault diagnosis algorithm for hydropower station transformer based on multi-channel neural network and multi-source heterogeneous data feature fusion, characterized in that: The following steps are involved: S1, Data collection: Collect multi-source heterogeneous data, including one-dimensional data and two-dimensional data; One-dimensional data includes time series data and non-time series data. Time series data includes gas content in oil, temperature data, humidity data, current data and vibration data. Non-time series data includes temperature, humidity and air density of the transformer operating environment. Two-dimensional data includes infrared image data; S2, data preprocessing: The variance elimination method deletes data that is not related to the fault diagnosis results, calculates the variance of all input parameters, and deletes parameters with zero variance. A variance of zero means that the parameter has not changed in all categories. The variance elimination method formula is: ； In the formula, S 2 is the sample variance, X is the sample data, is the sample mean, N is the number of samples; Calculate the Pearson correlation coefficient between each parameter and the fault diagnosis result, and retain the parameters with higher correlation; Find the absolute value of the Pearson correlation coefficient for each parameter: ； In the formula, x represents sample data, y represents fault label, n is the total number of samples, and subscript i represents the i-th sample; Taking the mean as the threshold, the absolute value of the Pearson coefficient of each parameter is compared with the threshold, and the parameters with the absolute value of the Pearson coefficient higher than the threshold are retained; the data is normalized to the mean and the data set is divided into a training set and a test set according to a fixed ratio. The mean calculation method is: ； in is the data obtained after mean normalization, x is the sample data, u is the sample mean, and S is the standard deviation; S3, build a fault diagnosis model; S4, perform fault diagnosis.

2. According to claim 1, a hydropower station transformer fault diagnosis algorithm based on multi-channel neural network multi-source heterogeneous data feature fusion is characterized in that: In step S3, constructing a fault diagnosis model includes: Construct a compressed excitation network SEnet, which includes two one-dimensional compressed excitation networks 1DSEnet and one two-dimensional compressed excitation network 2DSEnet; The one-dimensional data consisting of the preprocessed time series data and non-time series data is used as the input of 1DSEnet, and the two-dimensional data consisting of the preprocessed infrared image data is used as the input of 2DSEnet; Use SEnet to adaptively adjust the weights of each input parameter and replace the relu function in the SEnet fully connected layer with the tanh function: ； ； In the formula, x is the input sample; Three neural networks, namely one-dimensional convolutional neural network 1DCNN, two-dimensional convolutional neural network 2DCNN and gated recurrent unit GRU, are established respectively, and the time series data, non-time series data and two-dimensional data after weight adjustment are used as the input of GRU, 1DCNN and 2DCNN respectively; Improve GRU, use bidirectional GRU for feature extraction, introduce dropout layer to prevent bidirectional GRU from overfitting, and introduce spatially separable convolution in 2DCNN to improve the network structure; The bidirectional GRU is used to extract the temporal features in the data, the 1DCNN is used to extract the non-temporal features in the data, and the 2DCNN is used to extract the two-dimensional image features in the data. The extracted features of the three dimensions are accumulated for feature fusion: All data features are one-dimensionalized and used as the input of the fully connected layer. Multiple fully connected layers are used to extract the fused data features. The residual neural network ResNet is introduced to improve the DNN network structure, and the DNN is connected across layers. The batch normalization layer is introduced to standardize the distribution of data features. Finally, the data features are input into the classifier for fault diagnosis.

3. According to claim 2, a hydropower station transformer fault diagnosis algorithm based on multi-channel neural network multi-source heterogeneous data feature fusion is characterized in that: In step S4, fault diagnosis includes performing fault diagnosis through the classifier in the model constructed in step S3, and the specific method is as follows: The fault categories are divided into overheating faults, discharge faults and mechanism faults, and the specific locations corresponding to the fault categories are located; Label the fault data, encode the fault type using the One-Hot method, represent the categorical variables as binary vectors, and expand the discrete fault categories to the Euclidean space; The model constructed in step S3 is trained using the labeled data to obtain a trained model, and then the unlabeled data is used as the input of the model, and finally the category corresponding to the data is input.

Images