Transformer fault identification method, related device and computer storage medium

Abstract

The invention provides a transformer fault identification method, a related device and a computer storage medium, and the method comprises the steps: generating a frequency response curve of a transformer according to a pulse response signal of the transformer after obtaining the pulse response signal of the transformer; inputting the frequency response curve of the transformer into a fault identification model, and outputting to obtain the fault type of the transformer; wherein the fault recognition model is obtained by training a deep residual neural network through a sample set; the sample set comprises a frequency response curve of the transformer in a superposed fault state. Through the pre-trained deep residual neural network (fault recognition model), the accuracy and efficiency of fault recognition are significantly improved, and a powerful guarantee is provided for rapid positioning and processing of faults.

Description

Technical Field

[0001] The present application relates to the technical field of power systems, and in particular to a transformer fault identification method, related devices, and computer storage media. Background Art

[0002] Power transformers, as key equipment in power systems, shoulder the crucial responsibilities of power transmission and voltage conversion. However, due to the complexity of the power system operating environment, transformer windings can suffer from a variety of mechanical failures, with winding deformation being the most common. Winding deformation failures have a cumulative effect. If not promptly detected and repaired, they can trigger other transformer failures and even lead to severe inter-turn short circuits, posing a serious threat to the safety and stability of the power system transmission network. Summary of the Invention

[0003] In view of this, the present application provides a transformer fault identification method, related devices and computer storage medium, which significantly improve the accuracy and efficiency of fault identification and provide strong guarantees for the rapid location and handling of faults.

[0004] A first aspect of the present application provides a method for identifying a transformer fault, comprising:

[0005] Obtaining the impulse response signal of the transformer;

[0006] generating a frequency response curve of the transformer according to the impulse response signal of the transformer;

[0007] The frequency response curve of the transformer is input into a fault identification model, and the fault type of the transformer is obtained as an output; wherein the fault identification model is obtained by training a deep residual neural network with a sample set; the sample set includes the frequency response curve of the transformer under a superimposed fault state.

[0008] Optionally, the method for constructing the fault identification model includes:

[0009] Classify the frequency response curve of the transformer under the superimposed fault state to obtain the fault category;

[0010] For each fault category, using the data of the fault category as a training sample set of the fault category;

[0011] The data in the training sample set is input into the deep residual neural network, and the predicted fault category is output;

[0012] Based on the predicted fault category and the actual fault category to which the data in the training sample set belongs, the parameters in the deep residual neural network are adjusted until the error between the predicted fault category and the actual fault category to which the data in the training sample set belongs meets the preset convergence condition, and the deep residual neural network is used as the fault recognition model.

[0013] Optionally, the step of inputting the data in the training sample set into a deep residual neural network and outputting a predicted fault category includes:

[0014] After the deep residual neural network receives the data in the training sample set, it extracts features from the data in the training sample set to obtain training sample features;

[0015] Perform global average pooling based on the training sample features to obtain global average pooling data;

[0016] Performing full connection on the global average pooling data to obtain fully connected data;

[0017] A normalized exponential function operation is performed on the fully connected data to obtain a predicted fault category.

[0018] Optionally, adjusting parameters in a deep residual neural network based on the predicted fault category and the actual fault category to which the data in the training sample set belongs includes:

[0019] Based on the predicted fault category and the actual fault category to which the data in the training sample set belongs, the learnable parameters in the convolution kernel are adjusted through a back-propagation algorithm.

[0020] A second aspect of the present application provides a transformer fault identification device, comprising:

[0021] A signal acquisition unit, used to acquire an impulse response signal of the transformer;

[0022] A curve generating unit, configured to generate a frequency response curve of the transformer according to an impulse response signal of the transformer;

[0023] A fault identification unit is used to input the frequency response curve of the transformer into a fault identification model and output the fault type of the transformer; wherein the fault identification model is obtained by training a deep residual neural network with a sample set; the sample set includes the frequency response curve of the transformer in a superimposed fault state.

[0024] Optionally, the construction unit of the fault identification model includes:

[0025] A classification unit, used to classify the frequency response curve of the transformer under the superimposed fault state to obtain the fault category;

[0026] A training sample set construction unit, configured to use, for each fault category, the data of the fault category as a training sample set of the fault category;

[0027] The prediction unit is used to input the data in the training sample set into the deep residual neural network and output the predicted fault category;

[0028] An adjustment unit is used to adjust the parameters in the deep residual neural network based on the predicted fault category and the actual fault category to which the data in the training sample set belongs, until the error between the predicted fault category and the actual fault category to which the data in the training sample set belongs meets a preset convergence condition, and the deep residual neural network is used as a fault recognition model.

[0029] Optionally, the prediction unit includes:

[0030] A feature extraction unit is used to extract features from the data in the training sample set after the deep residual neural network receives the data in the training sample set, so as to obtain training sample features;

[0031] The pooling unit is used to perform global average pooling based on the training sample features to obtain global average pooling data;

[0032] A fully connected unit, configured to perform a full connection on the global average pooled data to obtain fully connected data;

[0033] The function operation unit is used to perform a normalized exponential function operation on the fully connected data to obtain a predicted fault category.

[0034] Optionally, the adjustment unit includes:

[0035] The adjustment subunit is used to adjust the learnable parameters in the convolution kernel through a back propagation algorithm based on the predicted fault category and the actual fault category to which the data in the training sample set belongs.

[0036] As can be seen from the above scheme, this application provides a transformer fault identification method, related device, and computer storage medium. After obtaining the transformer's pulse response signal, a transformer frequency response curve is generated based on the transformer's pulse response signal. This transformer frequency response curve is then input into a fault identification model, and the output is the transformer fault type. The fault identification model is obtained by training a deep residual neural network using a sample set; the sample set includes the frequency response curve of the transformer in a superimposed fault state. The pre-trained deep residual neural network (fault identification model) significantly improves the accuracy and efficiency of fault identification, providing a strong guarantee for rapid fault location and resolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0038] Figure 1 A specific flow chart of a transformer fault identification method provided in an embodiment of the present application;

[0039] Figure 2 A schematic diagram of a transformer provided in another embodiment of the present application;

[0040] Figure 3 This is a frequency response curve diagram of a normal state of a disk A and an axial concave deformation fault superimposed on a 100pF capacitor provided by another embodiment of the present application;

[0041] Figure 4 This is a frequency response curve diagram of a normal state of a disk A and an axial concave deformation fault superimposed with 100 and 800 pF capacitors provided by another embodiment of the present application;

[0042] Figure 5 A frequency response curve diagram of a normal state and a superimposed axial concave deformation fault under a pancake spacing variation fault provided by another embodiment of the present application;

[0043] Figure 6 A frequency response curve diagram of a superposition of an axial concave deformation fault and a fault of varying degrees of inter-pancake spacing provided by another embodiment of the present application;

[0044] Figure 7 A frequency response curve diagram of a radial deformation fault superimposed on a slight pancake spacing change fault provided by another embodiment of the present application;

[0045] Figure 8 A frequency response curve diagram of a radial deformation fault superimposed with a fault of varying degrees of inter-pancake spacing provided by another embodiment of the present application;

[0046] Figure 9 A flowchart of a method for constructing a fault identification model provided in another embodiment of the present application;

[0047] Figure 10 A flowchart of a deep residual neural network processing method provided by another embodiment of the present application;

[0048] Figure 11 A loss function diagram at a 224*224 input resolution provided by another embodiment of the present application;

[0049] Figure 12 A schematic diagram of a transformer fault identification device provided by another embodiment of the present application;

[0050] Figure 13 A schematic diagram of an electronic device for implementing a transformer fault identification method provided in another embodiment of the present application. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0052] As used herein, the term "including" and its variations are open-ended, i.e., "including but not limited to." The term "based on" means "based, at least in part, on." The term "one embodiment" means "at least one embodiment," the term "another embodiment" means "at least one additional embodiment," and the term "some embodiments" means "at least some embodiments." Other terms are defined in the following description.

[0053] It should be noted that the information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions.

[0054] It should be noted that the concepts of "first" and "second" mentioned in this application are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0055] It should be noted that the modifications of "one" and "multiple" mentioned in this application are illustrative rather than restrictive. Those skilled in the art should understand that unless otherwise clearly indicated in the context, they should be understood as "one or more".

[0056] The present application embodiment provides a method for identifying transformer faults, such as Figure 1 As shown, the specific steps include:

[0057] S101. Obtain an impulse response signal of a transformer.

[0058] In the specific implementation process of the present application, the pulse response signal of the transformer can be detected by, but not limited to, injecting a pulse signal into the transformer, which is not limited here.

[0059] S102: Generate a frequency response curve of the transformer according to the impulse response signal of the transformer.

[0060] In the specific implementation process of the present application, the frequency response curve of the transformer can be generated by, but not limited to, performing Fourier transform on the impulse response signal, which is not limited here.

[0061] S103: Input the frequency response curve of the transformer into the fault identification model, and output the fault type of the transformer.

[0062] Among them, the fault identification model is obtained by training the deep residual neural network with a sample set; the sample set includes the frequency response curve of the transformer under the superimposed fault state.

[0063] First of all, it should be noted that transformer winding faults include radial deformation faults, inter-pancake spacing variation faults, inter-pancake short circuit faults, and winding axial phase change faults. In practical applications, transformer winding faults usually do not occur in the form of a single fault occurring independently. The present invention conducts a superposition experiment including transformer winding axial deformation faults, winding radial deformation faults, and inter-pancake spacing variation faults to explore the impact of multiple fault superposition states on the frequency response curve. Based on the experimental results, a sample set is generated for training a deep residual neural network.

[0064] Specifically, there are four common winding deformation faults: radial deformation fault (RD), axial deformation fault (AD), inter-pancake spacing variation fault (DSV), and inter-pancake short circuit fault (IDSC). If the two are superimposed on each other, due to the diversity of fault severity, fault location, and superposition method, the experiment will not be able to cover all types of fault superposition states. Therefore, on the basis of the normal state of the transformer, axial deformation fault and radial deformation fault are added, and then inter-pancake spacing variation fault is added on this basis. By setting a limited number of fault types, fault severity, and fault locations, the law of the frequency response curve when fault superposition occurs is preliminarily explored.

[0065] The transformer used in the experiment was a custom-made 10kV model transformer with Y-type connections on both the high-voltage and a-sides. The high-voltage winding of this transformer is designed as a pancake winding, divided into three sections: upper, middle, and lower, with 10 pancakes in each section, for a total of 30 pancakes. The upper and lower sections are fixed, non-removable, conventional tangled windings. The middle 10 pancakes are grouped into five groups of two pancakes each, all sequentially wound and removable and replaceable. The low-voltage winding is a six-layer, fixed, non-removable layered winding. Both the high and low voltage windings have insulated end rings at the top and bottom for insulation and support. To simulate winding failures under different fault conditions, custom windings with varying fault types and fault severity were used to replace the 10 replaceable middle pancakes, enabling the configuration of fault types.

[0066] like Figure 2 The figure shows a schematic diagram of a transformer provided by an embodiment of the present application. The high-voltage winding of the transformer is designed as a pancake winding, which is divided into three parts: upper, middle and lower. Each part has 10 pancakes, for a total of 30 pancakes. The upper and lower parts are fixed, non-detachable, ordinary tangled windings. The middle 10 pancakes are grouped into two groups, for a total of five groups. All of them are sequentially wound windings and can be disassembled and replaced. The low-voltage winding is a 6-layer fixed, non-detachable layer winding. The top and bottom of the high and low voltage windings have insulating end rings for insulation and support. In order to simulate winding failures under different fault conditions, windings with different fault types and fault degrees are customized to replace the 10 replaceable windings in the middle, thereby realizing the setting of the fault type.

[0067] In actual use, the transformer windings should be completely immersed in transformer insulating oil. For operability during the experiment, the transformer body was lifted out of the oil tank and fixed to a steel base, keeping it stationary. To simulate the oil tank to the greatest extent possible, an openable metal shell (a removable transformer shell) was designed and manufactured. The material can be stainless steel 304, and its size and appearance are consistent with the transformer box. The 10 windings in the middle are marked as five groups: A, B, C, D, and E. Each group of windings has two lead connectors, marked 1, 2, 3, 4, 5, and 6 respectively. Due to the symmetry of the transformer's replaceable windings, the experimental results of groups A and E, and groups B and D are the same during the experiment. Therefore, the test can be performed only on the three groups A, B, and C.

[0068] Offline experiment on superposition of transformer winding axial deformation fault and inter-panel spacing change fault:

[0069] To control variables, the experiment set up a normal-axial deformation comparison for the winding shape. By superimposing a change in inter-pancake spacing fault with a normal-axial deformation fault, the impact of the inter-pancake spacing change fault on the frequency response curve was compared with the existing axial deformation fault. The experiment explored the impact of inter-pancake spacing changes with a small fault level (100pF capacitor in parallel), a medium fault level (400pF capacitor in parallel), and a large fault level (800pF capacitor in parallel) on the frequency response curve when superimposed with the axial deformation fault. The experimental setup is shown in Table 1.

[0070] Table 1

[0071] Serial number Fault 1 Fault 2 Fault location (connector number) Failure severity 1 / / / / 2 / DSV 1-2 100pF 3 / DSV 1-2 400pF 4 / DSV 1-2 800pF 5 / DSV 5-6 100pF 6 / DSV 5-6 400pF 7 / DSV 5-6 800pF 8 AD / / / 9 AD DSV 1-2 100pF 10 AD DSV 1-2 400pF 11 AD DSV 1-2 800pF 12 AD DSV 5-6 100pF 13 AD DSV 5-6 400pF 14 AD DSV 5-6 800pF

[0072] During the experiment, only the inter-pancake spacing change fault was set on the A and E groups of windings, and the axial deformation fault was set only on the A group of windings. Due to the symmetry of the upper and lower structures of the windings, the same fault set at the 1-2 joint and the 5-6 joint should have the same impact. When collecting data in the experiment, the oscilloscope used the averaging mode, with an average of 128 times. The experiment was repeated 500 times while maintaining the same operating conditions. A comparative analysis will be conducted below.

[0073] The experiment first measured the frequency response curves under four conditions: normal state of the transformer, fault of change in spacing between windings of group A of the transformer (parallel 100pF capacitor fault), axial concave deformation fault of winding of group A, and superposition of axial concave deformation fault of winding of group A and fault of change in spacing between windings of group A.

[0074] pass Figure 3 From the frequency response curve shown, it can be seen that when a minor degree of inter-pancake spacing change fault occurs, the frequency response curve at 700kHz undergoes a resonant peak-valley reversal, while a new resonant peak is added to the high-frequency response curve at 950kHz; when an inter-pancake spacing change fault occurs, regardless of whether there is an axial deformation fault of the winding, the change in the frequency response curve is reflected as a resonant peak-valley change at 700kHz and 950kHz, that is, the main cause affecting the curve is the inter-pancake spacing change fault.

[0075] Comparing the frequency response curves of the change in the spacing between the pancakes when the transformer is in normal state and when the winding axial deformation occurs, the comparison results are as follows: Figure 4 shown.

[0076] Experiments show that the frequency response curves of the transformer in normal state and when the same inter-panel spacing change fault occurs when an axial concave deformation fault occurs are similar. Therefore, when the two faults occur at the same time, depending on the different ways in which different fault types affect the frequency response curve, one fault plays a major role within the corresponding impact frequency range. Based on the two curves with slight changes in the approximate frequency response curves, it can be considered that other faults have occurred at the same time in the transformer.

[0077] Due to the symmetrical structure of the transformer, when only the axial concave deformation is set on the A cake, the same inter-cake spacing change fault is set on the two symmetrical groups of windings A and E cakes. The frequency response curve is affected similarly. The same result as the setting of different degrees of inter-cake spacing change faults on the A cake is that the inter-cake spacing fault has a dominant influence on the frequency response curve. Therefore, whether the axial concave deformation fault with a lighter impact on the frequency response curve is set when the inter-cake spacing change fault occurs, the impact on the curve is also small. The comparison results are shown in Figure 2. Figure 5 and Figure 6 shown.

[0078] Offline experiment for superposition of transformer winding radial deformation fault and inter-panel spacing change fault:

[0079] In order to control variables, the experiment sets a normal-radial deformation control on the winding shape. By setting the superposition of the inter-cake spacing change fault and the normal-radial deformation fault, the influence of the inter-cake spacing change fault on the frequency response curve based on the existing radial deformation fault is compared. During the experiment, radial deformation faults are set on the A and B cakes to compare the influence of different radial deformation positions on the fault superposition experiment. The radial deformation degrees of 3%, 5%, 7%, and 10% are set respectively to compare the influence of different radial deformation degrees on the fault superposition experiment. The inter-cake spacing change fault equivalent to 100pF, 400pF, and 800pF capacitors are set in parallel to compare the influence of different inter-cake spacing change faults on the fault superposition experiment. The experimental settings are shown in Table 2:

[0080] Table 2

[0081] Serial number Fault 1 Fault 2 Fault location Radial deformation failure degree Connector number Fault severity of changes in inter-pane spacing 1 / / / / / / 2 RD / A, B 3% / / 3 RD DSV A, B 3% 1-2 100pF 4 RD DSV A, B 3% 1-2 400pF 5 RD DSV A, B 3% 1-2 800pF 6 RD / A, B 5% / / 7 RD DSV A, B 5% 1-2 100pF 8 RD DSV A, B 5% 1-2 400pF 9 RD DSV A, B 5% 1-2 800pF 10 RD / A, B 7% / / 11 RD DSV A, B 7% 1-2 100pF 12 RD DSV A, B 7% 1-2 400pF 13 RD DSV A, B 7% 1-2 800pF 14 RD / A, B 10% / / 15 RD DSV A, B 10% 1-2 100pF 16 RD DSV A, B 10% 1-2 400pF 17 RD DSV A, B 10% 1-2 800pF

[0082] The experiment first measured the frequency response curves under four conditions: normal state of the transformer, fault of change in spacing between windings of group A of the transformer (parallel 100pF capacitor fault), fault of radial deformation of group A windings to varying degrees, and superposition of radial deformation fault of group A windings and fault of change in spacing between windings of group A.

[0083] pass Figure 7It can be observed that the radial deformation fault is mainly manifested in the amplitude difference change of the frequency response curve. At 700kHz and 950kHz, the resonance peak and valley changes are still mainly dominated by the inter-pancake spacing change fault, while the degree of radial deformation has little effect on the frequency response curve. It is mainly manifested in the different amplitude offsets of the frequency response curve when different degrees of radial deformation fault occur.

[0084] On the basis of setting the same degree (10%) of radial deformation fault on Pancake A, different degrees of inter-pancake distance change fault are added to compare the influence of different degrees of inter-pancake distance change fault on the frequency response curve based on the radial deformation fault. The results are as follows: Figure 8 shown.

[0085] For faults with the same degree of inter-pancake spacing variation, the curve for the radial deformation fault associated with the additional winding exhibits an amplitude shift, consistent with the curve for radial deformation of the additional winding under normal transformer conditions. The significant amplitude shift in the frequency response curve and the changes in resonant peaks and valleys within certain frequency bands indicate that, when a fault significantly impacts the frequency response curve, it is possible to infer whether a radial deformation fault or an inter-pancake spacing variation fault has occurred by comparing the frequency response curve with the normal state.

[0086] Through experiments and data analysis, the present invention reveals which faults have a greater impact on a specific frequency range and which have a smaller impact when different faults are superimposed, as well as the law that the effects of multiple faults on the frequency response curve are superimposed, providing strong support for subsequent fault detection and classification.

[0087] Optionally, in another embodiment of the present application, an implementation method of a fault identification model construction method is as follows: Figure 9 Shown, including:

[0088] S901. Classify the frequency response curve of the transformer under the superimposed fault state to obtain a fault category.

[0089] Among them, the fault categories include but are not limited to short circuit faults between pancakes, change in spacing between pancakes, radial deformation faults of windings, axial deformation faults of windings, and superposition faults of radial deformation faults and change in spacing between pancakes, which are not limited here.

[0090] In the specific implementation process of this application, through the transformer winding deformation multiple fault curve classification method, it is possible to accurately determine whether the transformer winding has a deformation fault, as well as the type and degree of the fault, providing strong support for subsequent maintenance and replacement.

[0091] S902 : For each fault category, use the data of the fault category as a training sample set of the fault category.

[0092] It is understandable that in the actual application of this application, a part (such as 80%) of the fault category data can be selected as the training sample set of the fault category, and a part (such as 20%) of the fault category data can be selected as the test sample set of the fault category. There is no limitation here.

[0093] S903: Input the data in the training sample set into the deep residual neural network, and output the predicted fault category.

[0094] The deep residual neural network may be but is not limited to Resnet50, which is not limited here.

[0095] Resnet50 introduces the residual module, which solves the vanishing gradient problem in deep network training through residual connections, allowing for the training of deeper networks. It consists of multiple residual modules, each of which contains multiple convolutional layers, including a convolutional layer with downsampling to match the dimensions of the input and output.

[0096] Specifically, the deep residual neural network performs convolution, maximum pooling, and network function operations on the input curve, and outputs the predicted fault category.

[0097] In the specific implementation process of this application, the construction method of the Resnet50 model can be as follows:

[0098] During the initialization process, the default normalization layer is used, and the grouping and basic width of the control convolution are checked and limited to 1 and 64 respectively. Finally, two convolutional layers, two normalization layers and an activation function are initialized, and the downsampling layer and step size are set; then a forward propagation function is used to define a residual block to build the basic unit in restnet50. In this process, the input variable will be copied to the identity variable for residual connection. This step helps to alleviate the gradient disappearance problem when training deep neural networks and speed up training; then a module is defined to reduce the number of network parameters and the amount of calculation. It contains three convolutional layers and three batch normalization layers and an activation function. The first convolutional layer is a 1x1 convolution for reducing the dimension of the input feature map, the second convolutional layer is a 3x3 convolution for feature extraction, and the third convolutional layer is a 1x1 convolution for restoring the number of channels of the feature map to the original channel dimension. Assume that the input image size is H*W and the convolution kernel (pooling window) size is k h *k w The first convolutional layer reduces the number of channels of the input feature map to width.

[0099] Then, using the forward propagation method, the input variables are passed through a series of convolutions, batch normalization, and activation functions, and the result is returned. The whole process consists of convolution, activation functions, maximum pooling, full connection, and softmax. Assume that the input image size is H*W and the convolution kernel (pooling window) size is k h *k w , filled with (P h , P w ), the stride is (S h , S w ), then the size of the convolutional layer output feature map is:

[0100] ; .

[0101] In two-dimensional convolution, multiple convolution kernels are typically used to extract different features. Each convolution kernel contains several learnable parameters that can be optimized through the back-propagation algorithm. In addition to the convolution operation, two-dimensional convolution can also use pooling, padding and other techniques to adjust the size and shape of the output feature map, thereby better adapting to different image processing tasks.

[0102] Optionally, in another embodiment of the present application, an implementation of step S903 is as follows: Figure 10 Shown, including:

[0103] S1001. After the deep residual neural network receives the data in the training sample set, feature extraction is performed on the data in the training sample set to obtain training sample features.

[0104] In the specific implementation process of this application, for the first make_layer (make_layer is mainly used to define each layer in the restnet50 network, usually including convolution layer, batch normalization layer and activation function. Through make_layer, the entire network structure can be easily constructed without writing code layer by layer.), the input feature layer size is 56*56*64, the initial default step size is 1*1, so the height and width of the input image will not be compressed, the number of output channels is 64*4, so the output feature layer is 56*56*256. After completing the first m After the make_layer layer is constructed, multiple make_layer layers still need to be constructed; specify the second make_layer layer step size of 2*2 and the number of channels of 128, so the height and width of the input feature layer will be compressed, and the number of channels will be expanded to 128*4, resulting in an output feature layer of 28*28*512; the third make_layer layer step size of 2*2 and the number of channels of 256, thus obtaining an output feature layer of 14*14*1024; the fourth make_layer layer step size of 2*2 and the number of channels of 512, thus obtaining an output feature layer of 7*7*2048.

[0105] S1002: Perform global average pooling based on the training sample features to obtain global average pooling data.

[0106] Continuing with the above example, global average pooling is performed based on the training sample features to obtain a strip of length 2048, namely the global average pooling data.

[0107] Among them, the pooling operation mainly plays two roles: ① Feature compression: The pooling operation can convert each sub-region (also called pooling window) in the input data into a separate numerical value, thereby compressing the information in each sub-region into a representative numerical value. This helps to reduce redundant information in the feature map and make the model more efficient; ② Position invariance: Since the pooling operation only considers the statistics of each sub-region in the input data (such as the maximum value or average value), they are independent of the position of the input data. Therefore, the pooling layer can enhance the robustness of the model to transformations such as translation and rotation, thereby improving the generalization ability of the model. Average pooling calculates the average of all values in each sub-region as the output.

[0108] Since the stride of the maximum pooling is 2*2, the height and width of the input feature layer are compressed again, while the number of features remains unchanged, resulting in a 56*56*64 feature layer. Next, the second make layer is constructed (one Conv Block and two Identity Blocks). The first Conv Block changes the dimension of the input feature layer, so its number of channels and stride can be specified, thereby changing the number of channels and stride of the output feature layer. The Identity Block deepens the network, so it can be simply stacked twice.

[0109] S1003: Perform full connection on the global average pooling data to obtain fully connected data.

[0110] Among them, the number of fully connected neurons is the number of faults, which is used as the number of fault classification types. Assuming it is N, it is equivalent to classifying the input data into N types.

[0111] S1004: Perform a normalized exponential function operation on the fully connected data to obtain a predicted fault category.

[0112] The normalized exponential function may be, but is not limited to, softmax, and is not limited here.

[0113] S904. Based on the predicted fault category and the actual fault category to which the data in the training sample set belongs, adjust the parameters in the deep residual neural network until the error between the predicted fault category and the actual fault category to which the data in the training sample set belongs meets the preset convergence condition, and use the deep residual neural network as a fault recognition model.

[0114] The preset convergence conditions are pre-set and modified by technical personnel or authorized staff and are not limited here.

[0115] It should be noted that, in the specific implementation process of this application, it is not limited to using the preset convergence conditions to train the model, but the preset number of iterations can also be used to train the model, which is also not limited here.

[0116] In the specific implementation process of the present application, the learnable parameters in the convolution kernel can be adjusted through the back propagation algorithm based on, but not limited to, the predicted fault category and the actual fault category to which the data in the training sample set belongs.

[0117] In the specific implementation process of this application, after building the Resnet50 network, the classification effect of the transformer fault curve will be verified, and each set of experimental data will be converted into a frequency response curve for input, of which 80% of the experimental data will be used as a training set and the remaining 20% of the data will be used as a test set. Then, one curve of each type of experiment will be selected as a label, for a total of 14 sets of data. The input curve format is as follows: After performing the above operations, the parameters of each module in the constructed Resnet model will be initialized. First, the weights of all convolutional layers will be initialized using the initialization method, and then the weights of the batch normalization layer and the group normalization layer will be initialized to 1, and the bias will be initialized to 0, thereby achieving an initialization function that helps ensure that the model has a good initial state during training, which is conducive to faster convergence of the model and better performance. The loss function diagram when the input resolution is 224*224 is as follows. Figure 11 As shown in the figure, after 150 iterations, both the train loss and the val loss tend to be moderate, and the model training effect is good.

[0118] By training the corresponding fault classification model, the final output results are as follows when the resolution is 360*360:

[0119] Mean Recall: 99.35%. This represents the average recall of the model for all classes. Recall is the ratio of the number of samples correctly predicted as positive to the total number of actual positive samples.

[0120] Mean Precision: 99.35%. This represents the average precision of the model for all classes. Precision refers to the ratio of the number of examples correctly predicted as positive to the total number of examples predicted as positive.

[0121] Top-1 Accuracy: 99.29%. This represents the percentage of correctly predicted categories among all predictions. This refers to the percentage of cases where the model's highest possible prediction matches the actual label.

[0122] Top-5 Accuracy: 99.29%. This indicates the percentage of times the model ranks the correct label among the top five most likely categories. This metric is often used to assess a model's ability to distinguish difficult categories.

[0123] The classification effect of the same data set was then verified using the classification learner in Matlab. The model training and testing were performed using linear SVM, quadratic SVM, cubic SVM, fine Gaussian SVM, medium Gaussian SVM, and coarse Gaussian SVM. The training results are shown in Table 3:

[0124] Table 3

[0125]

[0126] Table 3 shows that the cubic SVM model has the highest accuracy. The results show that the cubic SVM model with the highest accuracy is approximately 5 percentage points lower than the Resnet50 model. In some fault categories where the SVM model test results were poor, the Resnet50 model still performed well.

[0127] As can be seen from the above scheme, this application provides a method for identifying transformer faults. After obtaining the transformer's impulse response signal, a frequency response curve of the transformer is generated based on the impulse response signal. The frequency response curve of the transformer is then input into a fault identification model, and the output is the transformer fault type. The fault identification model is obtained by training a deep residual neural network using a sample set; the sample set includes the frequency response curve of the transformer under superimposed fault conditions. The pre-trained deep residual neural network (fault identification model) significantly improves the accuracy and efficiency of fault identification, providing a strong guarantee for rapid fault location and resolution.

[0128] Another embodiment of the present application provides a device for identifying transformer faults, such as Figure 12 As shown, specifically including:

[0129] The signal acquisition unit 1201 is configured to acquire an impulse response signal of the transformer.

[0130] The curve generating unit 1202 is configured to generate a frequency response curve of the transformer according to the impulse response signal of the transformer.

[0131] The fault identification unit 1203 is configured to input the frequency response curve of the transformer into the fault identification model, and output the fault type of the transformer.

[0132] Among them, the fault identification model is obtained by training the deep residual neural network with a sample set; the sample set includes the frequency response curve of the transformer under the superimposed fault state.

[0133] The specific working process of the units disclosed in the above embodiments of this application can be found in the corresponding method embodiments, such as Figure 1 As shown, no further details are given here.

[0134] Optionally, in another embodiment of the present application, an implementation of a construction unit of a fault identification model includes:

[0135] The classification unit is used to classify the frequency response curve of the transformer under the superimposed fault state to obtain the fault category.

[0136] The training sample set construction unit is used to use the data of each fault category as the training sample set of the fault category.

[0137] The prediction unit is used to input the data in the training sample set into the deep residual neural network and output the predicted fault category.

[0138] An adjustment unit is used to adjust the parameters in the deep residual neural network based on the predicted fault category and the actual fault category to which the data in the training sample set belongs, until the error between the predicted fault category and the actual fault category to which the data in the training sample set belongs meets the preset convergence condition, and the deep residual neural network is used as a fault recognition model.

[0139] For the specific working process of the units disclosed in the above embodiments of the present application, please refer to the corresponding method embodiments, such as Figure 9 As shown, no further details are given here.

[0140] Optionally, in another embodiment of the present application, an implementation of the prediction unit includes:

[0141] The feature extraction unit is used to extract features from the data in the training sample set after the deep residual neural network receives the data in the training sample set to obtain training sample features.

[0142] The pooling unit is used to perform global average pooling based on the training sample features to obtain global average pooling data.

[0143] The fully connected unit is used to fully connect the global average pooling data to obtain fully connected data.

[0144] The function operation unit is used to perform a normalized exponential function operation on the fully connected data to obtain the predicted fault category.

[0145] For the specific working process of the units disclosed in the above embodiments of the present application, please refer to the corresponding method embodiments, such as Figure 10 As shown, no further details are given here.

[0146] Optionally, in another embodiment of the present application, an implementation of the adjustment unit includes:

[0147] The adjustment subunit is used to adjust the learnable parameters in the convolution kernel through the back propagation algorithm based on the predicted fault category and the actual fault category to which the data in the training sample set belongs.

[0148] The specific working process of the units disclosed in the above embodiments of this application can be found in the corresponding method embodiments and will not be repeated here.

[0149] As can be seen from the above scheme, this application provides a transformer fault identification device. After obtaining the transformer's impulse response signal, the device generates a transformer frequency response curve based on the transformer's impulse response signal. The transformer frequency response curve is then input into a fault identification model, and the output is the transformer fault type. The fault identification model is obtained by training a deep residual neural network using a sample set; the sample set includes the frequency response curve of the transformer under superimposed fault conditions. The pre-trained deep residual neural network (fault identification model) significantly improves the accuracy and efficiency of fault identification, providing a strong guarantee for rapid fault location and resolution.

[0150] The functions described above herein may be performed, at least in part, by one or more hardware logic components. For example, and without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0151] Another embodiment of the present application provides an electronic device, such as Figure 13 Shown, including:

[0152] One or more processors 1301.

[0153] The storage device 1302 stores one or more programs.

[0154] When the one or more programs are executed by the one or more processors 1301 , the one or more processors 1301 implement the transformer fault identification method as described in the above embodiment.

[0155] Another embodiment of the present application provides a computer storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the transformer fault identification method as described in the above embodiment is implemented.

[0156] In the context of this application, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or apparatus. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of machine-readable storage media may include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), optical fibers, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0157] It should be noted that the computer-readable medium referred to in this application may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. Computer-readable storage media may include, but are not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or components, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, an electrical connection having one or more conductors, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device, or component. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. This propagated data signal may take a variety of forms, including, but not limited to, electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wire, optical cable, RF (radio frequency), or any suitable combination thereof.

[0158] The computer-readable medium may be included in the electronic device, or may exist independently without being incorporated into the electronic device.

[0159] Another embodiment of the present application provides a computer program product. When the computer program product is executed, it is used to perform the above-mentioned transformer fault identification method.

[0160] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a non-transitory computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processing device, the above-mentioned functions defined in the method of the embodiment of the present application are performed.

[0161] Although the subject matter has been described in terms of structural features and / or method logic actions, it should be understood that the subject matter defined in this application is not necessarily limited to the specific features or actions described above. On the contrary, the specific features and actions described above are merely example forms of implementing this application.

[0162] Although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this application. Certain features described in the context of separate embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented in multiple embodiments individually or in any suitable sub-combination.

[0163] The above description is merely a preferred embodiment of the present application and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of application of this application is not limited to technical solutions formed by a specific combination of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned application concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions claimed in this application.

Claims

1. A method for identifying transformer faults, characterized in that: include: Obtaining the impulse response signal of the transformer; generating a frequency response curve of the transformer according to the impulse response signal of the transformer; The frequency response curve of the transformer is input into a fault identification model, and the fault type of the transformer is obtained as an output; wherein the fault identification model is obtained by training a deep residual neural network with a sample set; the sample set includes the frequency response curve of the transformer under a superimposed fault state.

2. The transformer fault identification method according to claim 1, characterized in that: The method for constructing the fault identification model includes: Classify the frequency response curve of the transformer under the superimposed fault state to obtain the fault category; For each fault category, using the data of the fault category as a training sample set of the fault category; The data in the training sample set is input into the deep residual neural network, and the predicted fault category is output; Based on the predicted fault category and the actual fault category to which the data in the training sample set belongs, the parameters in the deep residual neural network are adjusted until the error between the predicted fault category and the actual fault category to which the data in the training sample set belongs meets the preset convergence condition, and the deep residual neural network is used as the fault recognition model.

3. The transformer fault identification method according to claim 2, characterized in that: The data in the training sample set is input into the deep residual neural network, and the output is the predicted fault category, including: After the deep residual neural network receives the data in the training sample set, it extracts features from the data in the training sample set to obtain training sample features; Perform global average pooling based on the training sample features to obtain global average pooling data; Performing full connection on the global average pooling data to obtain fully connected data; A normalized exponential function operation is performed on the fully connected data to obtain a predicted fault category.

4. The transformer fault identification method according to claim 2, characterized in that: The adjusting of parameters in the deep residual neural network based on the predicted fault category and the actual fault category to which the data in the training sample set belongs includes: Based on the predicted fault category and the actual fault category to which the data in the training sample set belongs, the learnable parameters in the convolution kernel are adjusted through a back-propagation algorithm.

5. A transformer fault identification device, characterized in that: include: A signal acquisition unit, used to acquire an impulse response signal of the transformer; A curve generating unit, configured to generate a frequency response curve of the transformer according to an impulse response signal of the transformer; A fault identification unit is used to input the frequency response curve of the transformer into a fault identification model and output the fault type of the transformer; wherein the fault identification model is obtained by training a deep residual neural network with a sample set; the sample set includes the frequency response curve of the transformer in a superimposed fault state.

6. The transformer fault identification device according to claim 5, characterized in that: The construction unit of the fault identification model includes: A classification unit, used to classify the frequency response curve of the transformer under the superimposed fault state to obtain the fault category; A training sample set construction unit, configured to use, for each fault category, the data of the fault category as a training sample set of the fault category; The prediction unit is used to input the data in the training sample set into the deep residual neural network and output the predicted fault category; An adjustment unit is used to adjust the parameters in the deep residual neural network based on the predicted fault category and the actual fault category to which the data in the training sample set belongs, until the error between the predicted fault category and the actual fault category to which the data in the training sample set belongs meets a preset convergence condition, and the deep residual neural network is used as a fault recognition model.

7. The transformer fault identification device according to claim 6, characterized in that: The prediction unit includes: A feature extraction unit is used to extract features from the data in the training sample set after the deep residual neural network receives the data in the training sample set, so as to obtain training sample features; The pooling unit is used to perform global average pooling based on the training sample features to obtain global average pooling data; A fully connected unit, configured to perform a full connection on the global average pooled data to obtain fully connected data; The function operation unit is used to perform a normalized exponential function operation on the fully connected data to obtain a predicted fault category.

8. The transformer fault identification device according to claim 6, characterized in that: The adjustment unit includes: The adjustment subunit is used to adjust the learnable parameters in the convolution kernel through a back propagation algorithm based on the predicted fault category and the actual fault category to which the data in the training sample set belongs.

9. An electronic device, characterized in that: include: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the transformer fault identification method according to any one of claims 1 to 4.

10. A computer storage medium, characterized in that A computer program is stored thereon, wherein when the computer program is executed by a processor, the transformer fault identification method according to any one of claims 1 to 4 is implemented.

Images