A transformer fault diagnosis method and system

Abstract

The present invention proposes a transformer fault diagnosis method and system, which includes the following steps: obtaining a transformer historical fault data set; setting a fault label according to the fault type, and correspondingly forming a number of fault sample data; fusing an infrared image and a vibration signal to obtain a fused image; using the fused image data to train a double-layer convolutional neural network, and calculating the characteristic index of the fused image and distinguishing difficult and non-difficult samples, respectively training the classifier A and classifier B of the double-layer convolutional neural network to obtain a trained double-layer convolutional neural network; migrating the trained classifier A to an MLP perceptron to generate a multi-source fusion network; using a pre-processed sound signal to train the multi-source fusion network to obtain a trained multi-source fusion network; collecting signals in real time, inputting them into an integrated network, and obtaining a fault diagnosis output result. The present invention reduces the amount of data processing, improves the data processing speed, and improves the accuracy of fault identification.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power equipment, and in particular to a transformer fault diagnosis method and system. Background Art

[0002] Power transformers are one of the key equipment in the power system. Since power transformers are located in a key position in the power grid and have a complex operating environment, once a failure occurs, it will cause problems such as cascading failures in the power grid and large-scale power outages. What's worse, it will cause serious accidents such as explosions and fires, which will cause huge direct or indirect economic losses. Therefore, the operating status of the transformer determines the stability and safety of the power system, and the safe, reliable and stable operation of the power transformer is crucial.

[0003] With the continuous development of sensor monitoring equipment, as well as its large-scale deployment and application on the surface and inside of power transformers, the requirements for the intelligent level of power transformer equipment monitoring and positioning have also been greatly improved. The transformer has obvious parameter changes and abnormal signals before a fault occurs. When diagnosing a transformer fault, the abnormal condition of the transformer should be detected. However, the causes of internal transformer failures are very complex. When diagnosing a fault, a large amount of data and signals need to be collected for comprehensive judgment and analysis. In traditional transformer fault diagnosis, engineers need to have rich experience. Different engineers will make different subjective judgments based on data and signals, which restricts the improvement of transformer fault diagnosis accuracy and speed.

[0004] The main methods for transformer fault diagnosis include three ratio method, expert system method and neural network method. Among them, the three ratio method uses ratios for linear classification, which is prone to the problem of too absolute classification boundaries; the accuracy of the expert system method is too dependent on the experience of experts in the knowledge base; the neural network method has a reasonable structure and has the ability of continuous self-learning. Comprehensive analysis shows that the neural network method can better meet the needs of transformer fault diagnosis.

[0005] At present, the most commonly used technology for transformer fault diagnosis is to analyze the dissolved gas in transformer oil using the neural network method, but this technology still has some defects, such as: it is difficult to extract features based on sample data, the corresponding mechanism between fault type and gas in oil is unclear, and the accuracy of gas distribution measurement is insufficient. In addition, since the fault type is indirectly inferred by monitoring the generated products, there is an obvious lag, which leads to problems such as over-repair or disrepair, resulting in low evaluation accuracy. Summary of the invention

[0006] In view of the above problems, the present invention provides a transformer fault diagnosis method and system, which solves the problems in the prior art of difficult feature extraction, insufficient measurement accuracy, and obvious hysteresis.

[0007] To solve the above technical problems, the technical solution adopted by the present invention is: a transformer fault diagnosis method, comprising the following steps: obtaining a transformer historical fault data set, the historical fault data set including infrared images, vibration signals and sound signals, and corresponding fault types; setting fault labels according to fault types, and corresponding the fault labels with infrared images, vibration signals and sound signals to form a number of fault sample data; after converting the infrared images and vibration signals into grayscale images of the same size, performing image data fusion to obtain a fused image; using the fused image data to train a double-layer convolutional neural network, and calculating the characteristic index of the fused image at the same time, distinguishing difficult-to-identify and non-difficult samples according to the characteristic index, and using the difficult-to-identify and non-difficult samples to train classifiers A and classifiers B of the double-layer convolutional neural network respectively to obtain a trained double-layer convolutional neural network; migrating the trained classifier A to an MLP perceptron to generate a multi-source fusion network; using the preprocessed sound signal to train the multi-source fusion network to obtain a trained multi-source fusion network; real-time acquisition of infrared images, vibration signals and sound signals of the transformer, inputting them into an integrated network composed of a double-layer convolutional neural network and a multi-source fusion network, and obtaining a fault diagnosis output result.

[0008] As a preferred solution, the step of calculating the characteristic index of the fused image and distinguishing between difficult-to-identify samples and non-difficult-to-identify samples according to the characteristic index includes:

[0009] Suppose the features of the fused image after the convolution layer for:

[0010] ;

[0011] Where, i=1,2,3,…N, j=1,2,3,…M;

[0012] Calculate the mean and variance of each dimension feature, and then calculate the normalized feature index ;

[0013] ;

[0014] In the above formula, is the mean, is the variance, N is the number of images, and M is the dimension of the convolution feature;

[0015] like If it is greater than the index threshold, it is a difficult sample to identify, otherwise it is a non-difficult sample to identify.

[0016] As a preferred solution, the two-layer convolutional neural network model includes classifier A and classifier B, wherein classifier A is a SoftMax classifier, and classifier B is an SVM classifier, and both share multiple convolutional layers and pooling layers to achieve the extraction of common features.

[0017] As a preferred solution, the image data fusion is performed to obtain a fused image, specifically: the read infrared image is converted into a grayscale image with a size of [m, n, 1]; the vibration signal is normalized and converted into a grayscale image with a size of [m, n, 1]; the two matrices are fused into a matrix of [m, n, 2].

[0018] As a preferred solution, it also includes denoising the original vibration signal, specifically: performing adaptive signal decomposition on the original vibration signal, extracting the kurtosis value of each IMF component after decomposition, arranging the IMF components in descending order according to the kurtosis value, reconstructing the IMF components whose kurtosis value ratio is greater than the kurtosis threshold, and generating a denoised vibration signal.

[0019] As a preferred solution, assuming that the kurtosis value ratio is CPK, the calculation formula is as follows:

[0020] ;

[0021] ;

[0022] In the above formula, For time series The kurtosis value of For signal The mean of For signal The standard deviation of, E is the expectation, is the kurtosis value of the rth IMF component, and P is the number of IMF components after signal decomposition.

[0023] As a preferred solution, the fault types include winding fault, core fault, tap changer fault and oil tank fault.

[0024] The present invention also discloses a transformer fault diagnosis system, including: a data acquisition module, used to acquire a transformer historical fault data set, the historical fault data set including infrared images, vibration signals and sound signals, and corresponding fault types; a label module, used to set fault labels according to fault types, and form a number of fault sample data by corresponding the fault labels with infrared images, vibration signals and sound signals; an image fusion module, used to convert the infrared images and vibration signals into grayscale images of the same size, and then fuse the image data to obtain a fused image; a first training module, using the fused image data to train a double-layer convolutional neural network, and at the same time calculating the fusion image A feature index is used to distinguish between difficult-to-identify and non-difficult-to-identify samples according to the feature index, and the difficult-to-identify and non-difficult-to-identify samples are used to train classifiers A and B of the double-layer convolutional neural network respectively to obtain a trained double-layer convolutional neural network; a migration module is used to migrate the trained classifier A to the MLP perceptron to generate a multi-source fusion network; a second training module is used to train the multi-source fusion network using the pre-processed sound signal to obtain a trained multi-source fusion network; a diagnosis output module is used to collect infrared images, vibration signals and sound signals of the transformer in real time, and input them into an integrated network composed of a double-layer convolutional neural network and a multi-source fusion network to obtain a fault diagnosis output result.

[0025] Compared with the prior art, the beneficial effects of the present invention include: constructing an integrated network composed of a double-layer convolutional neural network and a multi-source fusion network, fusing infrared images and vibration signals for training, distinguishing difficult and non-difficult samples according to feature indexes, using different classifiers for training, and then combining difficult sample data with sound signals to train the multi-source fusion network. Taking advantage of the different sensitivities of infrared images, vibration signals, and sound signals to different fault types, and their complementarity, the trained integrated network has a higher diagnostic accuracy and can achieve a more accurate judgment of the fault type. Compared with a single feature, the three data features can significantly improve the accuracy of fault identification, reduce the amount of data processing, increase the data processing speed, improve the real-time performance of fault identification, and have good diagnostic performance in a small sample background. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the accompanying drawings, the same reference numerals are used to refer to the same components. Among them:

[0027] Figure 1 A schematic diagram of a transformer fault diagnosis method according to an embodiment of the present invention;

[0028] Figure 2 Another schematic diagram of a transformer fault diagnosis method according to an embodiment of the present invention;

[0029] Figure 3 It is a schematic diagram of the structure of a transformer fault diagnosis system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific implementation modes and drawings are only exemplary descriptions of the technical solution of the present invention, and should not be regarded as the whole of the present invention or as a limitation or restriction to the technical solution of the present invention.

[0031] According to one embodiment of the present invention, Figure 1 A transformer fault diagnosis method comprises the following steps:

[0032] S101, obtaining a transformer historical fault data set, where the historical fault data set includes infrared images, vibration signals, sound signals, and corresponding fault types.

[0033] S102, setting a fault label according to the fault type, and matching the fault label with the infrared image, the vibration signal and the sound signal to form a number of fault sample data.

[0034] Specifically, the fault types include winding fault, core fault, tap changer fault and oil tank fault, and their fault labels are set as error01, error02, error03, error04 respectively. Each fault type corresponds to the infrared image, vibration signal and sound signal data collected at the time of the fault.

[0035] S103, after converting the infrared image and the vibration signal into grayscale images of the same size, image data fusion is performed to obtain a fused image.

[0036] The image data is fused to obtain a fused image. Specifically, the infrared image is converted into a grayscale image with a size of [m, n, 1]; the vibration signal is normalized and converted into a grayscale image with a size of [m, n, 1]; the two matrices are fused into a matrix of [m, n, 2]. For example, if the grayscale images of the infrared image and the vibration signal are of size [64, 32, 1], the size of the fused image is [64, 32, 2].

[0037] S104, using the fused image data to train a two-layer convolutional neural network, and calculating the characteristic index of the fused image at the same time, distinguishing difficult-to-recognize and non-difficult-to-recognize samples according to the characteristic index, and using the difficult-to-recognize and non-difficult-to-recognize samples to train classifiers A and classifiers B of the two-layer convolutional neural network respectively, to obtain a trained two-layer convolutional neural network.

[0038] In the embodiment of the present invention, the above-mentioned calculation of the characteristic index of the fused image and distinguishing the difficult-to-identify samples and the non-difficult-to-identify samples according to the characteristic index include:

[0039] Suppose the features of the fused image after the convolution layer for:

[0040] ;

[0041] Where, i=1,2,3,…N, j=1,2,3,…M;

[0042] Calculate the mean and variance of each dimension feature, and then calculate the normalized feature index ;

[0043] ;

[0044] In the above formula, is the mean, is the variance, N is the number of images, and M is the dimension of the convolution feature;

[0045] like If it is greater than the index threshold, it is a difficult sample to identify, otherwise it is a non-difficult sample to identify.

[0046] In an embodiment of the present invention, the above-mentioned two-layer convolutional neural network model includes convolution layer 1, convolution layer 2, average pooling layer, classifier A and classifier B. The two layers of convolution respectively extract features from the input data, and the last convolution output layer uses an average pooling layer instead of a fully connected layer, which can reduce the total amount of training parameters and retain the feature information of the fused image. Classifier A is a SoftMax classifier, and classifier B is an SVM classifier, and the two share multiple convolutional layers and pooling layers to achieve the extraction of common features. By utilizing the high accuracy and precision of the SoftMax classifier, processing samples that are difficult to identify can significantly improve the classification accuracy of the model; by utilizing the fast calculation speed of the SVM classifier, processing samples that are not difficult to identify can significantly improve the classification speed of the model, thereby achieving the overall rapidity and high precision requirements of the classification model.

[0047] S105, migrating the trained classifier A to the MLP perceptron to generate a multi-source fusion network.

[0048] It should be understood that MLP perceptron, or multilayer perceptron (MLP), is a feedforward artificial neural network model that can have multiple hidden layers in addition to the input and output layers. For example, the structure of MLP perceptron is: fully connected layer FC1 → fully connected layer FC2. Transfer, or transfer learning, is to use the trained classifier A as the starting point and reuse it in the MLP perceptron.

[0049] S106, using the preprocessed sound signal to train a multi-source fusion network to obtain a trained multi-source fusion network.

[0050] Before model training, infrared images, vibration signals and sound signals need to be preprocessed, specifically: adding noise, cutting, transforming, rotating and mirroring the grayscale images to expand the samples to be trained.

[0051] It also includes noise reduction processing for infrared images, vibration signals and sound signals, for example: adaptive signal decomposition of the original vibration signal, extraction of the kurtosis value of each IMF component after decomposition, arranging the IMF components in descending order according to the kurtosis value, reconstructing the IMF components whose kurtosis value ratio is greater than the kurtosis threshold, and generating a noise-reduced vibration signal.

[0052] Assuming the proportion of kurtosis value is CPK, the calculation formula is as follows:

[0053] ;

[0054] ;

[0055] In the above formula, For time series The kurtosis value of For signal The mean of For signal The standard deviation of, E is the expectation, is the kurtosis value of the rth IMF component, and P is the number of IMF components after signal decomposition.

[0056] It should be noted that in model training and actual testing, the gradient back propagation method used by the two-layer convolutional neural network and the multi-source fusion network is a pre-training method, that is, the back propagation path will pass through the MLP layer and the convolutional neural network layer at the same time, so that each parameter data can correct the convolutional neural network to a certain extent. The loss function of the two-layer convolutional neural network and the multi-source fusion network is the cross entropy cost function. When the loss function is minimized and the training set results are close to the verification set results, the training is completed.

[0057] S107, real-time acquisition of infrared images, vibration signals and sound signals of the transformer, input into an integrated network consisting of a double-layer convolutional neural network and a multi-source fusion network, and obtain a fault diagnosis output result. Generally, the fault diagnosis results output by the double-layer convolutional neural network and the multi-source fusion network are of the same fault type. If the output results are different, manual intervention is required for judgment.

[0058] It can be understood that in the embodiment of the present invention, the historical fault data set and the real-time collected infrared images, vibration signals and sound signals are obtained through multiple sensor groups installed on the transformer casing, and each sensor group includes an infrared imaging sensor, a vibration sensor and a sound sensor.

[0059] See also Figure 3 The present invention also discloses a transformer fault diagnosis system, comprising:

[0060] The data acquisition module 101 is used to acquire a transformer historical fault data set, where the historical fault data set includes infrared images, vibration signals, sound signals, and corresponding fault types.

[0061] The label module 102 is used to set a fault label according to the fault type, and to correspond the fault label with the infrared image, the vibration signal and the sound signal to form a number of fault sample data.

[0062] The image fusion module 103 is used to convert the infrared image and the vibration signal into grayscale images of the same size, and then perform image data fusion to obtain a fused image.

[0063] The first training module 104 uses the fused image data to train a two-layer convolutional neural network, and simultaneously calculates the feature index of the fused image, distinguishes between difficult-to-identify and non-difficult-to-identify samples according to the feature index, and respectively uses the difficult-to-identify and non-difficult-to-identify samples to train classifiers A and B of the two-layer convolutional neural network to obtain a trained two-layer convolutional neural network.

[0064] The migration module 105 migrates the trained classifier A to the MLP perceptron to generate a multi-source fusion network.

[0065] The second training module 106 uses the pre-processed sound signal to train the multi-source fusion network to obtain a trained multi-source fusion network.

[0066] The diagnosis output module 107 collects infrared images, vibration signals and sound signals of the transformer in real time, and inputs them into an integrated network composed of a double-layer convolutional neural network and a multi-source fusion network to obtain a fault diagnosis output result.

[0067] In summary, the beneficial effects of the present invention include: constructing an integrated network composed of a double-layer convolutional neural network and a multi-source fusion network, fusing infrared images and vibration signals for training, distinguishing difficult and non-difficult samples according to feature indexes, using different classifiers for training, and then combining difficult sample data with sound signals to train the multi-source fusion network. Taking advantage of the different sensitivities of infrared images, vibration signals, and sound signals to different fault types, and their complementarity, the trained integrated network has a higher diagnostic accuracy and can achieve a more accurate judgment of the fault type. Compared with a single feature, the three data features can significantly improve the accuracy of fault identification, reduce the amount of data processing, increase the data processing speed, improve the real-time performance of fault identification, and have good diagnostic performance in a small sample background.

[0068] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0069] It should be understood that if the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.

[0070] The technical scope of the present invention is not limited to the contents in the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of ​​the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. A transformer fault diagnosis method, characterized in that: The steps include: Acquire a transformer historical fault data set, wherein the historical fault data set includes infrared images, vibration signals, sound signals, and corresponding fault types; Setting a fault label according to the fault type, and corresponding the fault label with the infrared image, vibration signal and sound signal to form a number of fault sample data; After converting the infrared image and the vibration signal into grayscale images of the same size, image data fusion is performed to obtain a fused image; The double-layer convolutional neural network is trained by using the fused image data, and the characteristic index of the fused image is calculated at the same time. According to the characteristic index, the samples with difficulty in identification and those with no difficulty in identification are distinguished, and the classifiers A and B of the double-layer convolutional neural network are trained by using the samples with difficulty in identification and those with no difficulty in identification, respectively, to obtain a trained double-layer convolutional neural network; Migrate the trained classifier A to the MLP perceptron to generate a multi-source fusion network; Using the preprocessed sound signal to train the multi-source fusion network to obtain a trained multi-source fusion network; The infrared image, vibration signal and sound signal of the transformer are collected in real time and input into an integrated network composed of a double-layer convolutional neural network and a multi-source fusion network to obtain the fault diagnosis output results.

2. The transformer fault diagnosis method according to claim 1, characterized in that: The step of calculating the characteristic index of the fused image and distinguishing the difficult-to-identify samples and the non-difficult-to-identify samples according to the characteristic index comprises: Suppose the features of the fused image after the convolution layer for: ; Where, i=1,2,3,…N, j=1,2,3,…M; Calculate the mean and variance of each dimension feature, and then calculate the normalized feature index ; ； In the above formula, is the mean, is the variance, N is the number of images, and M is the dimension of the convolution feature; like If it is greater than the index threshold, it is a difficult sample to identify, otherwise it is a non-difficult sample to identify.

3. The transformer fault diagnosis method according to claim 1, characterized in that: The double-layer convolutional neural network model includes classifier A and classifier B, wherein classifier A is a SoftMax classifier and classifier B is a SVM classifier, and both share multiple convolutional layers and pooling layers to realize the extraction of common features.

4. The transformer fault diagnosis method according to claim 1, characterized in that: The image data fusion is performed to obtain a fused image, specifically: Convert the read infrared image into a grayscale image with a size of [m, n, 1]; Normalize the vibration signal and convert it into a grayscale image with a size of [m, n, 1]; Merge the two matrices into a [m, n, 2] matrix.

5. The transformer fault diagnosis method according to claim 1, characterized in that: The method also includes performing noise reduction processing on the original vibration signal, specifically: performing adaptive signal decomposition on the original vibration signal, extracting the kurtosis value of each IMF component after decomposition, arranging each IMF component in descending order according to the kurtosis value, reconstructing the IMF component whose kurtosis value accounts for a proportion greater than the kurtosis threshold, and generating a noise-reduced vibration signal.

6. The transformer fault diagnosis method according to claim 5, characterized in that: Assuming the proportion of kurtosis value is CPK, the calculation formula is as follows: ； ； In the above formula, For time series The kurtosis value of For signal The mean of For signal The standard deviation of, E is the expectation, is the kurtosis value of the rth IMF component, and P is the number of IMF components after signal decomposition.

7. The transformer fault diagnosis method according to claim 1, characterized in that: The fault types include winding fault, core fault, tap changer fault and oil tank fault.

8. A transformer fault diagnosis system, characterized in that: include: A data acquisition module, used to acquire a transformer historical fault data set, wherein the historical fault data set includes infrared images, vibration signals and sound signals, and corresponding fault types; A label module is used to set a fault label according to the fault type, and form a number of fault sample data by matching the fault label with the infrared image, vibration signal and sound signal; An image fusion module, used for converting the infrared image and the vibration signal into grayscale images of the same size, and then fusing the image data to obtain a fused image; The first training module uses the fused image data to train a two-layer convolutional neural network, and simultaneously calculates the characteristic index of the fused image, distinguishes between difficult-to-identify and non-difficult-to-identify samples according to the characteristic index, and respectively uses the difficult-to-identify and non-difficult-to-identify samples to train classifiers A and B of the two-layer convolutional neural network to obtain a trained two-layer convolutional neural network; The migration module migrates the trained classifier A to the MLP perceptron to generate a multi-source fusion network; The second training module uses the preprocessed sound signal to train the multi-source fusion network to obtain a trained multi-source fusion network; The diagnostic output module collects infrared images, vibration signals and sound signals of the transformer in real time, and inputs them into an integrated network composed of a double-layer convolutional neural network and a multi-source fusion network to obtain fault diagnosis output results.

Images