Transformer overheat defect detection positioning method and system based on transfer learning

Abstract

The present application relates to the technical field of machine learning, in particular to a transformer overheat defect detection and positioning method and system based on transfer learning, the method comprising: establishing a transformer experiment model containing an overheat defect; establishing a three-dimensional simulation model of the transformer experiment model; constructing a transformer overheat defect diagnosis model, pre-training the transformer overheat defect diagnosis model based on obtained acoustic sample data; fine-tuning the transformer overheat defect diagnosis model through transfer learning; constructing a temperature field inversion model; obtaining an ultrasonic signal of a transformer to be measured, inputting the ultrasonic signal into the trained transformer overheat defect diagnosis model for judgment; according to the judgment result of the transformer overheat diagnosis model, performing temperature field inversion on the winding inside the transformer to be measured through the temperature field inversion model, and positioning the overheat defect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, in particular to a transformer overheat defect detection and positioning method and system based on transfer learning. BACKGROUND

[0002] Real-time monitoring of internal overheat defects of a transformer plays a crucial role in prolonging the service life of the equipment, ensuring the safe and stable operation of the power system, and preventing sudden failures. In the diagnosis of internal overheat defects of a transformer based on machine learning algorithms, obtaining sufficient fault sample data to train a machine learning diagnosis model is a key issue. However, actual internal overheat defects of a transformer have the characteristics of occasionality and randomness. Adding overheat defects to an actual transformer poses a safety risk and a risk of retirement of the transformer. Therefore, some researchers choose a finite element simulation model as an alternative solution. In the simulation model, it is very convenient to arrange overheat defects. By arbitrarily traversing any position inside the transformer and simulating any temperature of overheat faults, acoustic sample data is obtained to train the machine learning diagnosis model.

[0003] However, in the simulation process, in order to reduce the complexity of simulation calculation, the internal structure of the transformer is usually simplified, which leads to a deviation between the obtained simulation data and the actual data, i.e. the risk of deviation from the actual situation of the obtained simulation sample. This deviation leads to poor performance of the machine learning algorithm model trained based on simulation data in subsequent actual application, thereby affecting the accuracy of the diagnosis result. SUMMARY

[0004] To solve the above problems, the first aspect of the present application provides a transformer overheat defect detection and positioning method based on transfer learning, comprising:

[0005] establishing a transformer experimental model containing overheat defects;

[0006] establishing a three-dimensional simulation model of the transformer experimental model, performing grid division on the three-dimensional simulation model, and setting the material physical parameters of each structure in the three-dimensional simulation model;

[0007] simulate overheat defects at different positions through the three-dimensional simulation model of the transformer, obtain acoustic sample data at different temperature levels for each overheat defect, simulate acoustic sample data without overheat defects, construct a transformer overheat defect diagnosis model, and pre-train the transformer overheat defect diagnosis model based on the obtained acoustic sample data;

[0008] collect ultrasonic sample data of overheat defects at different temperature levels at various positions using the transformer experimental model, fine-tune train the transformer overheat defect diagnosis model based on the ultrasonic sample data through transfer learning;

[0009] The temperature field inversion model is constructed, the overheat defects at different positions are simulated through the transformer experimental model and the three-dimensional simulation model respectively, the temperature field distribution signals of the overheat defects at each position are obtained, and a data set required by the temperature field inversion model is established based on the temperature field distribution signals, and the temperature field inversion model is trained through the data set;

[0010] An ultrasonic signal of the transformer to be measured is obtained, and the ultrasonic signal is input into the trained transformer overheat defect diagnosis model for judgment; according to the judgment result of the transformer overheat diagnosis model, the temperature field inversion of the winding inside the transformer to be measured is performed through the temperature field inversion model to locate the overheat defect.

[0011] The transformer overheat defect diagnosis model comprises a one-dimensional convolution layer, a ConvMixer module, a global average pooling layer and a full connection layer; the ConvMixer module comprises two one-dimensional convolution layers, a residual connection, batch normalization and an activation function.

[0012] The transformer overheat defect diagnosis model is trained based on the obtained acoustic sample data, and the specific operation is as follows:

[0013] The preprocessed acoustic sample data is input into the transformer overheat defect diagnosis model;

[0014] Through the cooperative action of the one-dimensional convolution layer and the ConvMixer module, deep feature extraction and representation of the original time domain signal of the acoustic sample data are performed, and key feature information related to the overheat defect category in the ultrasonic signal is obtained;

[0015] The key feature information is mapped to a fixed-dimensional classification vector through the full connection layer, and the probability distribution corresponding to each category of overheat defect is output through the activation function;

[0016] According to the probability distribution, the overheat defect category with the maximum probability value is selected to output the diagnosis result.

[0017] The transformer overheat defect diagnosis model is fine-tuned through transfer learning, and the specific method is as follows:

[0018] All convolution layers of the transformer overheat defect diagnosis model are frozen, the weights of the convolution layers are fixed, and only the parameters of the full connection layer are updated for fine-tuning;

[0019] A preset training iteration number is set, the ultrasonic sample data under the experimental condition is used to perform forward propagation on the transformer overheat defect diagnosis model to obtain an output prediction result; the output prediction result is compared with the actual experimental condition information, and a loss value is calculated;

[0020] According to the loss value, the parameter gradient of the full connection layer is calculated by using back propagation;

[0021] adjust the weights of the full connection layer according to the calculated parameter gradient;

[0022] determine whether the current iteration number reaches the preset training iteration number, if not, continue to use the ultrasound sample data for forward propagation, and adjust the weights of the full connection layer according to the new loss value;

[0023] if the current iteration number reaches the preset training iteration number, stop training and save the trained transformer overheating defect diagnosis model.

[0024] The temperature field inversion model comprises a SE attention mechanism module, a first convolutional neural network, a Transformer encoder and a second convolutional neural network connected in sequence.

[0025] The transformer overheating defect detection and positioning method further comprises three-dimensional visualization of the overheating defect positioning of the temperature inversion model, and the specific method is as follows:

[0026] extract the spatial coordinates (x, y, z) information of all temperature measuring points and the corresponding temperature prediction results, and convert the rectangular coordinates into polar coordinate form; x, y, z

[0027] generate uniformly distributed sampling points in the angle θ and height Z directions, and generate a two-dimensional grid;

[0028] perform cubic spline interpolation processing on the temperature data, convert the interpolated grid back to the rectangular coordinate system, draw the temperature field surface, and perform color mapping to associate the predicted temperature data with the color;

[0029] perform normalization processing on the temperature data, add a temperature scale, and adjust the viewing angle to optimize the three-dimensional visualization effect.

[0030] In the specific embodiment, the grid division of the three-dimensional simulation model specifically includes heating sheet division, heating rod division, solid region division, and geometric entity layer division.

[0031] The second aspect provides a transformer overheating defect detection and positioning system based on transfer learning, which applies the transformer overheating defect detection and positioning method based on transfer learning as described above, and comprises:

[0032] a simulation module for establishing a three-dimensional simulation model of the transformer experiment model, performing grid division on the three-dimensional simulation model, and setting the material physical parameters of each structure in the three-dimensional simulation model;

[0033] ​The diagnostic module is used for simulating overheat defects at different positions through a three-dimensional simulation model of the transformer, obtaining acoustic sample data at different temperature levels of each overheat defect, and simulating to obtain acoustic sample data without overheat defects; a transformer overheat defect diagnosis model is constructed, and the transformer overheat defect diagnosis model is pre-trained based on the obtained acoustic sample data;

[0034] The training module is used for collecting ultrasonic sample data of overheat defects at different temperature levels of various positions by using a transformer experimental model, and fine-tuning training of the transformer overheat defect diagnosis model through transfer learning based on the ultrasonic sample data;

[0035] The inversion module is used for constructing a temperature field inversion model, simulating overheat defects at different positions through the transformer experimental model and the three-dimensional simulation model respectively, obtaining temperature field distribution signals of overheat defects at various positions, establishing a data set required by the temperature field inversion model based on the temperature field distribution signals, and training the temperature field inversion model through the data set;

[0036] The processing module is used for obtaining an ultrasonic signal of a transformer to be measured, inputting the ultrasonic signal into the trained transformer overheat defect diagnosis model for judgment, and positioning overheat defects by performing temperature field inversion on the winding inside the transformer to be measured through the temperature field inversion model according to the judgment result of the transformer overheat diagnosis model.

[0037] The third aspect provides a transformer overheat defect detection and positioning device based on transfer learning, including a processor and a memory, wherein the processor implements the transfer learning transformer overheat defect detection and positioning method as described above when executing the computer program stored in the memory.

[0038] The fourth aspect provides a computer readable storage medium for storing a computer program, wherein the computer program is executed by a processor to implement the transfer learning transformer overheat defect detection and positioning method as described above.

[0039] Beneficial effects: The present application is a transformer overheat defect detection and positioning method and system based on transfer learning, which not only pre-trains a deep hybrid convolutional neural network with simulation working condition data, but also fine-tunes the pre-trained model with sample data under actual experimental working conditions through transfer learning. In addition, when constructing the temperature field inversion model, the transformer experimental model and the three-dimensional simulation model are improved to obtain more diversified data sets for training. The pure simulation model method may only rely on simulation data, lack the transfer learning process from simulation to actual working conditions, and lack the targeted improvement of the experimental model.

[0040] Through transfer learning, the model is pre-trained by a large amount of data under the simulation working condition, and is fine-tuned under the actual experimental working condition, so as to solve the problem of sample data scarcity under the actual working condition, and realize accurate diagnosis of the internal overheating defect of the transformer under the actual working condition. When constructing the temperature field inversion model, the experimental model and the simulation model are improved, the opportunity of the model to contact different temperature fields is increased, the model learns the temperature change rule of ultrasonic waves, and the prediction accuracy under different overheating defect conditions is improved. By obtaining the ultrasonic signal of the transformer to be measured, the thermal defect diagnosis model is first judged, and then the temperature field inversion model is used for temperature field inversion, and finally the temperature distribution and overheating fault position in the transformer are clearly displayed through three-dimensional visualization, so as to realize accurate positioning. BRIEF DESCRIPTION OF DRAWINGS

[0041] The solutions and advantages of the present application will become clear to those skilled in the art after reading the detailed description of the preferred embodiments. The drawings are only for the purpose of illustrating the preferred embodiments and are not considered as limiting the present application.

[0042] In the drawings:

[0043] Figure 1 Flow chart of the transformer overheating defect detection and positioning method based on transfer learning;

[0044] Figure 2 Three-dimensional simulation model of the transformer;

[0045] Figure 3 Sound field grid subdivision diagram, wherein (a) is a transformer overall structure subdivision network; (b) is a winding, heating sheet and heating rod subdivision network;

[0046] Figure 4 Probe arrangement diagram in the three-dimensional simulation model domain;

[0047] Figure 5 Details of the probe arrangement in the domain, wherein (a) is a probe arrangement diagram in a region not involving a heating sheet; (b) is a probe arrangement diagram in a region involving a heating sheet. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art belong to the scope of protection of the embodiments of the present application.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0050] ReferenceFigure 1 This embodiment provides a method for detecting and locating transformer overheating defects based on transfer learning. The specific implementation steps of this method are as follows:

[0051] S1. Establish an experimental model of a transformer that includes overheating defects;

[0052] A transformer experimental model with a structure similar to that of an actual transformer was established. The transformer experimental model included a tank, transformer windings, transformer oil, heating elements, heating rods, ultrasonic transceiver sensors, and a data acquisition card. The tank was made of steel and filled with transformer oil. The transformer windings were placed at the center of the tank, and sheet metal was embedded inside the windings to simulate the transformer core.

[0053] Meanwhile, the heating element is an alumina high-temperature ceramic heating element, and multiple alumina high-temperature ceramic heating elements are respectively set on the transformer winding to simulate the overheating defect of the winding.

[0054] To ensure that the heating of the alumina high-temperature ceramic heating element can be regulated, a patch-type temperature sensor and a thermocouple are also installed on the alumina high-temperature ceramic heating element. The patch-type temperature sensor is connected to a temperature controller to regulate the heating temperature of the alumina high-temperature ceramic heating element; the thermocouple is used to detect the temperature of the alumina high-temperature ceramic heating element and verify the accuracy of the temperature measurement by the patch-type temperature sensor.

[0055] The heating rod is located at the bottom corner inside the enclosure to simulate the background temperature field during transformer operation.

[0056] The ultrasonic transceiver sensor is attached to the outside of the enclosure and connected to the data acquisition card. The ultrasonic transceiver sensor is used to receive ultrasonic signals propagating inside the transformer experimental model.

[0057] S2. Establish a three-dimensional simulation model of the transformer experimental model, mesh the three-dimensional simulation model, and set the material physical parameters of each structure in the three-dimensional simulation model;

[0058] Based on the transformer experimental model, such as Figure 2 As shown, a three-dimensional simulation model including the enclosure, transformer windings, transformer oil, heating elements, heating rods, and ultrasonic transceiver sensors is established through finite element simulation. The specific method is as follows:

[0059] S201. Model and simplify the transformer experimental model;

[0060] In finite element simulation, comprehensive and detailed modeling significantly increases the difficulty of mesh generation and simulation calculation. To balance model complexity and solution efficiency, the structure of the transformer's 3D simulation model is simplified as follows: the central crossbeam of the casing is ignored; the main insulation between the high and low voltage windings is ignored; the outer insulating paper and support bars of the transformer windings are ignored; the interlayer spacers of the transformer windings are ignored; the metal clamps at the top and bottom of the windings are ignored; and the internal core model structure of the windings is ignored.

[0061] S202. Model the propagation medium below the transformer oil level inside the tank;

[0062] When both oil and air are present in the transformer experimental model, phenomena such as multiphase flow coupling, interface tracking, and mutual heat transfer significantly increase the complexity of the simulation solution and the computational resource requirements. In this embodiment, the finite element simulation only models the portion of the transformer below the oil level, ignoring the air region within the tank.

[0063] S203. Determine the dimensional parameters of each structure in the three-dimensional simulation model of the transformer.

[0064] Meanwhile, mesh generation plays a crucial role in data simulation, directly affecting computational accuracy and solution efficiency. (See also...) Figure 3 The meshing of the three-dimensional simulation model specifically includes the meshing of heating elements, heating rods, solid regions, and geometric entity layers.

[0065] Among them, the heating element is meshed with an extremely fine mesh, with the maximum element size set to be less than 0.05 m and the minimum element size set to be greater than 0.001 m. At the same time, an appropriate curvature factor and narrow region resolution are set to improve the meshing accuracy of the near field of the boundary, which is used to accurately characterize the temperature distribution characteristics near the ceramic heating element.

[0066] The heating rods are divided into ultra-fine sections, with the largest unit size less than 0.1 m and the smallest unit size greater than 0.005 m;

[0067] For solid regions, a coarse mesh is used, with the maximum element size less than 0.5 m and the minimum element size greater than 0.05 m;

[0068] The remaining portion of the geometric solid layer is meshed using free tetrahedral meshes. Tetrahedral elements are automatically generated based on the model's geometry, boundary curvature, and set global element size parameters to fill the blank areas of the 3D simulation model. Furthermore, to ensure the overall mesh quality and simulation stability, mesh growth rate and curvature control parameters are set to ensure a smooth transition between the free tetrahedral mesh and the surrounding refined regions, and to maintain regular element shapes, avoiding highly distorted or excessively small elements.

[0069] In the sound field simulation calculation, in order to balance the accuracy and calculation efficiency of the simulation, it is usually necessary to find a suitable compromise in the setting of the grid size. If the grid is too coarse, the simulation result cannot meet the requirement of solving accuracy, and if the grid is too fine, it will lead to too long solving time.

[0070] When simulating the problem of ultrasonic wave propagation, it is crucial to ensure that the grid size can reasonably resolve the wavelength. Setting the partitioned grid to be 1 / 5 of the ultrasonic wavelength not only ensures the solving accuracy of the model, but also avoids wasting computing resources due to too fine grid. Setting the partitioned grid of the boundary to be 1 / 20 of the wavelength, i.e. ensuring that each wavelength is covered by at least 20 grid cells, provides sufficient resolution to accurately describe the propagation process of the wave at the boundary.

[0071] S3, simulate overheating defects at different positions through a three-dimensional simulation model of the transformer, obtain acoustic sample data at different temperature levels of each overheating defect, and simulate to obtain acoustic sample data without overheating defects; construct a transformer overheating defect diagnosis model, and pre-train the transformer overheating defect diagnosis model based on the obtained acoustic sample data;

[0072] The transformer overheating defect diagnosis model comprises a one-dimensional convolution layer, a ConvMixer module, a global average pooling layer and a fully connected layer.

[0073] The input channel of the one-dimensional convolution layer is equal to the number of ultrasonic receiving sensors, which is used for convolution processing of the time series data of the ultrasonic receiving signal.

[0074] The ConvMixer module comprises two one-dimensional convolution layers, residual connection, batch normalization and an activation function. The first one-dimensional convolution layer is used to process the spatial information of each channel separately and extract spatial features; the second one-dimensional convolution layer fuses and recombines the channel outputs of the first one-dimensional convolution layer; at the same time, the residual connection connects the input and output of the first one-dimensional convolution layer; the batch normalization is used to accelerate the convergence of the training process; and the activation function is used to enhance the fitting ability of the model.

[0075] The transformer overheating defect diagnosis model is trained based on the obtained acoustic sample data, and the specific operation is as follows:

[0076] The preprocessed acoustic sample data is input into the transformer overheating defect diagnosis model.

[0077] Through the cooperative action of the one-dimensional convolution layer and the ConvMixer module, deep feature extraction and representation of the original time domain signal of the acoustic sample data are performed, and key feature information related to the overheating defect category in the ultrasonic signal is obtained.

[0078] The key feature information is mapped to a fixed-dimensional classification vector through a full connection layer, and an activation function outputs a probability distribution corresponding to each category of the overheating defect;

[0079] According to the probability distribution, an overheating defect category with the maximum probability value is selected to output a diagnosis result.

[0080] S4, collect ultrasonic sample data of overheating defects at different temperature levels in various positions by using a transformer experimental model, and fine-tune the transformer overheating defect diagnosis model through transfer learning based on the ultrasonic sample data;

[0081] For this embodiment, the source domain is a simulated working condition, and the target domain is an actual experimental working condition. It is difficult to arrange overheating defects in the transformer in the experimental condition, so the target domain sample data is scarce. Finite element simulation can very conveniently add overheating defects in the transformer, and a large amount of sample data can be obtained, so the acoustic sample data obtained in the simulation condition is used to pre-train the deep hybrid convolutional neural network, so that it can effectively determine whether there is an overheating defect in the transformer. Then, the pre-training model structure and the learned weight parameters are migrated to the target domain, and the ultrasonic sample data of the transformer normal background temperature collected in the experimental working condition is used to fine-tune the parameters of the pre-training model, to generate a target domain specific model, to realize accurate diagnosis of the overheating defect in the transformer in the actual working condition.

[0082] The fine-tuning training of the transformer overheating defect diagnosis model through transfer learning has the following specific methods:

[0083] S401, freeze all convolutional layers of the transformer overheating defect diagnosis model, fix the weights of the convolutional layers, and only update the parameters of the fine-tuning full connection layer; the relationship between the input features and the output labels can be remapped by fine-tuning the weight parameters of the full connection layer;

[0084] S402, preset the number of training iterations, and perform forward propagation on the transformer overheating defect diagnosis model using the ultrasonic sample data in the experimental working condition to obtain an output prediction result; compare the output prediction result with the actual experimental working condition information, and calculate a loss value; wherein the loss value usually uses cross-entropy loss, which belongs to the prior art, and will not be described here;

[0085] S403, according to the loss value, calculate the parameter gradient of the full connection layer using back propagation; back propagation propagates the error from the output layer to each layer of the network through the chain method. In this process, the weights of the convolutional layers will not be updated because their properties are frozen, and only the parameters of the full connection layer will be subjected to gradient calculation;

[0086] S404, adjust the weights of the full connection layer according to the calculated parameter gradient;

[0087] S405, judge whether the current iteration number reaches the preset training iteration number, if not, return to step S402, continue to use the ultrasonic sample data for forward propagation, and adjust the weights of the full connection layer according to the new loss value; with the training, the model will gradually adjust the parameters of the full connection layer to minimize the loss function, so that the prediction result of the transformer overheating defect diagnosis model is closer to the actual experimental working condition information, thereby improving the diagnosis accuracy of the model;

[0088] S406, if the current iteration number reaches the preset training iteration number, stop training, and save the trained transformer overheating defect diagnosis model; at this time, the parameters of the full connection layer of the model have been fully optimized, and can accurately map the relationship between the input features and the output labels, so as to realize effective diagnosis of the transformer overheating defect.

[0089] S5, construct a temperature field inversion model, simulate overheating defects at different positions through the transformer experimental model and the three-dimensional simulation model respectively, obtain the temperature field distribution signals of the overheating defects at each position, and establish a data set required by the temperature field inversion model based on the temperature field distribution signals, and train the temperature field inversion model through the data set;

[0090] In order to improve the accuracy and generalization ability of the model, the transformer experimental model and the three-dimensional simulation model are improved in this chapter, and more overheating defects at different positions are added in the transformer, so as to simulate more complex temperature field distribution. By allowing the model to contact more temperature fields in different situations, on the one hand, the model can learn more rules of ultrasonic changes with temperature, thereby improving its prediction accuracy in different overheating defect situations; on the other hand, the model can better reason when encountering new and unseen temperature distributions, thereby improving the generalization performance of the model.

[0091] Before simulating the overheating defects of the transformer experimental model and the three-dimensional simulation model, the transformer experimental model and the three-dimensional simulation model need to be improved, which is as follows:

[0092] The transformer experimental model is improved. Four position overheating defects are set for each of the upper, middle and lower three layers of the transformer winding, totaling 12 positions. Each position overheating defect is subdivided into 80℃, 100℃, 120℃, 140℃, 160℃ and 180℃, six temperature levels. In actual experiments or simulations, only one position heating sheet is heated at a certain temperature each time, so there are a total of 72 temperature fields. Adding the normal background temperature and the case without overheating, there are 73 types of temperature fields.

[0093] The three-dimensional simulation model is improved. The domain point probe is added to simulate the thermocouple to obtain the internal temperature information of the transformer. The arrangement of the domain point probe in the simulation model is as follows: Figure 4As shown in the diagram. A detailed schematic of the domain point probe arrangement is shown below. Figure 5 As shown. For the upper, middle, and lower winding layers, three domain point probes are set for each winding layer. In order to better capture the temperature distribution details around the ceramic heating element and enable the model to effectively predict the temperature of the overheated area in the subsequent temperature inversion process, 32 temperature measurement points are arranged per turn in the area involving the heating element, and 24 temperature measurement points are arranged per turn in the area not involving the heating element. Therefore, for a sample, the output temperature labels are 240.

[0094] A temperature field inversion model is constructed, which includes a convolutional neural network and a Transformer model.

[0095] The Transformer model relies on a multi-head attention mechanism, and its computational complexity is as follows:

[0096]

[0097] Where O represents the Order of Magnitude, N is the length of the ultrasound signal input sequence, d is the number of input data channels (i.e., the number of ultrasound sensors), and h is the number of attention heads. In this paper, the length of the acquired ultrasound signal data is 10000, the number of sensors in the ultrasound array is 13, and the number of attention heads is 4. Therefore, the computational complexity of directly using Transformer to process the acoustic data is 10000 × 10000 × 13 × 4, totaling 5.2 × 10⁻⁶. 9 .

[0098] Convolutional neural networks (CNNs) compute from input data through local receptive fields. The computational complexity is related to the kernel size *k*. If a one-dimensional convolution is used, the computational complexity of a CNN is as follows:

[0099]

[0100] Substituting into the actual situation, the size of the one-dimensional convolution kernel k =3, its computational complexity is 3.9 × 10 5 As the above analysis shows, ultrasound received signal data is typically high-sampling-rate data. Directly processing the raw data using a Transformer model incurs extremely high computational costs, while using a CNN can significantly reduce these costs. Furthermore, CNNs can continuously reduce feature dimensions through convolutional and pooling layers; therefore, adopting a CNN-Transformer structure effectively reduces the computational burden on the Transformer, improving computational efficiency.

[0101] In addition to considering the computational cost, combining CNN with Transformer can also play to their respective strengths. CNN is computationally efficient and suitable for feature extraction of large-scale data, and can capture the local changes of ultrasonic waveform signals in a short time. However, CNN can only gradually expand the receptive field through pooling operations and cannot directly focus on the entire input signal sequence. Transformer is not sufficient for low-level time series feature extraction and has relatively high computational cost, but it can focus on the global information of the entire waveform signal through the multi-head attention mechanism, so that the algorithm model is no longer limited to local features. By using the CNN-Transformer architecture, both short-term change features of ultrasonic signals and global dependency relationships can be captured, and the overall trend of long-term signals can be understood, while effectively relieving the computational pressure. In summary, the network structure combining CNN and Transformer is adopted.

[0102] Transformer was originally designed for machine translation, so it usually includes an encoder module, a decoder module, and position encoding. In theory, position encoding is a necessary part of the Transformer model, because the Transformer does not have sequence dependence like the recurrent neural network (RNN) algorithm model, and its structure is unordered, so position encoding is needed to provide sequential information. However, with the CNN-Transformer structure, the input data for the Transformer is the features after one-dimensional convolution and pooling operations, and CNN has extracted certain local time series information through convolution operations, so position encoding is no longer a necessary structure.

[0103] In addition, since the output target is the temperature prediction of multiple measuring points inside the transformer, which belongs to a regression task and is not a sequence generation task, a decoder module is not needed to generate the target sequence. For classification or regression prediction tasks, it is reasonable to only use the encoder part of the Transformer, and BERT algorithm and ViT (Vision Transformer) are both based on the encoder structure. Therefore, the encoder module is only retained in the Transformer part of the present embodiment.

[0104] To further refine the high-level abstract features extracted by the Transformer by integrating global information, a layer of CNN structure is added after the Transformer encoder. This can help the model extract more detailed features, while also reducing the feature dimension through convolution and pooling operations, thereby reducing the computational burden of subsequent fully connected layers and improving the accuracy of temperature field prediction. To enhance the dependency between feature channels, enable the model to dynamically adjust the weights of each feature channel, and reduce the interference of redundant feature information, an SE attention mechanism module is added to the CNN structure.

[0105] The overall design of the transformer internal temperature field inversion model is SE-CNN-Transformer-cnn structure, with the lower case cnn indicating that the module layer is only one layer, which is distinguished from the previous CNN of the Transformer. This structure fully combines the advantages of CNN and Transformer. The feature extraction process can be divided into the following stages. The first stage, two layers of CNN modules extract local features from the original acoustic signal; the second stage, the Transformer captures the long-distance dependencies between parts of the sequence through the self-attention mechanism, establishing relationships between each data of the input sequence and other data of the entire sequence, to help the model learn the overall trend of the signal; the third stage, the CNN module refines the features again to ensure the accuracy of temperature prediction. Through this multi-stage feature extraction from local features to global features and then to refined features, the performance of the algorithm model in ultrasonic data analysis and temperature field prediction tasks can be significantly improved.

[0106] The signal waveform received by the ultrasonic array is used as the input of the algorithm model. In each CNN module, ReLU is used as the activation function after convolution operation, and SE module is introduced between convolution operation and pooling operation. Max pooling is used to reduce the dimension of the features. The functions of the components in the Transformer encoder module are not described again, and the number of attention heads is set to 4. The Flatten layer is used to flatten the extracted feature data, i.e., to one-dimensionalize the multi-dimensional input. The input features are mapped to the transformer temperature field information through the fully connected layer (Dense), and the ReLU activation function is used in the internal hidden layer, and the Linear activation function is used in the output layer. The Linear activation function is very suitable for regression tasks, and can convert the output of the fully connected layer into the target prediction result. Finally, the model outputs the temperature prediction results of 240 measuring points, realizes the three-dimensional display of the internal temperature distribution of the transformer, and can effectively locate the position of the overheated area while obtaining the temperature information.

[0107] S6, obtaining an ultrasonic signal of the transformer to be measured, inputting the ultrasonic signal into the transformer overheating defect diagnosis model after migration learning for judgment; according to the judgment result of the transformer overheating diagnosis model, the temperature field of the winding inside the transformer to be measured is inversed through the temperature field inversion model, and the overheating defect is located.

[0108] By installing ultrasonic transceiving sensors on the side of the actual transformer tank to be measured, the ultrasonic signal inside the transformer tank is obtained. The collected ultrasonic signal is input into the transformer overheating defect diagnosis model. The model is trained based on a large number of transformer overheating defect samples and has effective recognition ability for overheating defects. The model analyzes the frequency, amplitude and phase of the ultrasonic signal to determine whether the transformer has overheating defects.

[0109] When the transformer overheating defect diagnosis model determines that the transformer to be measured has overheating defects, the temperature field inversion model is started to inverse the temperature field of the internal winding of the transformer to be measured.

[0110] In order to intuitively display the temperature distribution of the transformer winding and accurately locate the specific position of the overheating defect, the method further includes three-dimensional visualization of the overheating defect positioning of the temperature inversion model. The three-dimensional visualization can reconstruct the discrete predicted temperature results into a continuous temperature field distribution through interpolation, and map it to the geometric structure of the transformer winding, so as to clearly display the internal temperature distribution of the transformer and the position of the overheating fault. The specific operation is as follows:

[0111] Extract the spatial coordinates (x, y, z) information of all temperature measurement points and their corresponding temperature prediction results, and convert the rectangular coordinates into polar coordinate form; x, y, z

[0112] Generate uniformly distributed sampling points in the angle θ and height Z direction, and generate a two-dimensional grid;

[0113] Perform cubic spline interpolation processing on the temperature data, convert the interpolated grid back to the rectangular coordinate system, draw the temperature field surface, and perform color mapping to associate the predicted temperature data with the color;

[0114] Normalize the temperature data, add a temperature scale, adjust the viewing angle to optimize the three-dimensional visualization effect, so that the final image can clearly display the internal temperature field distribution of the transformer.

[0115] Then, the embodiment also provides a transformer overheating defect detection and positioning system based on migration learning, which applies the transformer overheating defect detection and positioning method based on migration learning as described above, comprising:

[0116] ​The simulation module is configured to establish a three-dimensional simulation model of the transformer experiment model, perform meshing on the three-dimensional simulation model, and set material physical parameters of each structure in the three-dimensional simulation model;

[0117] The diagnosis module is configured to simulate overheat defects at different positions through the three-dimensional simulation model of the transformer, acquire acoustic sample data at different temperature levels of each overheat defect, and simulate acquisition of acoustic sample data without overheat defects; construct a transformer overheat defect diagnosis model, and pre-train the transformer overheat defect diagnosis model based on the acquired acoustic sample data;

[0118] The training module is configured to collect ultrasonic sample data of overheat defects at different temperature levels of each position by using the transformer experiment model, and fine-tune train the transformer overheat defect diagnosis model through transfer learning based on the ultrasonic sample data;

[0119] The inversion module is configured to construct a temperature field inversion model, simulate overheat defects at different positions through the transformer experiment model and the three-dimensional simulation model respectively, acquire temperature field distribution signals of overheat defects at each position, establish a data set required by the temperature field inversion model based on the temperature field distribution signals, and train the temperature field inversion model through the data set;

[0120] The processing module is configured to acquire an ultrasonic signal of a to-be-tested transformer, input the ultrasonic signal into the trained transformer overheat defect diagnosis model for judgment, and perform temperature field inversion on internal windings of the to-be-tested transformer through the temperature field inversion model according to a judgment result of the transformer overheat diagnosis model to locate overheat defects.

[0121] In addition, a transformer overheat defect detection and positioning device based on transfer learning is also provided, which includes a processor and a memory, wherein the processor implements the transfer learning-based transformer overheat defect detection and positioning method as described above when executing a computer program saved in the memory.

[0122] Finally, a computer-readable storage medium for storing a computer program is also provided, wherein the computer program implements the transfer learning-based transformer overheat defect detection and positioning method as described above when executed by a processor.

[0123] Those skilled in the art will understand that embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0124] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0125] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0126] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those of skill in the art once they have the benefit of the present disclosure. Therefore, the appended claims are intended to cover all such variations and modifications as falling within the scope of the application.

Claims

1. A transformer overheat defect detection and positioning method based on transfer learning, characterized in that, The application relates to a transformer overheat defect diagnosis method and device. An experimental model of a transformer containing an overheat defect is established; A three-dimensional simulation model of the experimental model of the transformer is established, the three-dimensional simulation model is meshed, and the material physical parameters of each structure in the three-dimensional simulation model are set; Different overheat defects at different positions are simulated through the three-dimensional simulation model of the transformer, acoustic sample data at different temperature levels of each overheat defect is obtained, acoustic sample data without an overheat defect is simulated and obtained, a transformer overheat defect diagnosis model is constructed, the transformer overheat defect diagnosis model comprises a one-dimensional convolution layer, a ConvMixer module, a global average pooling layer and a full connection layer, the ConvMixer module comprises two one-dimensional convolution layers, a residual connection, batch normalization and an activation function, and the transformer overheat defect diagnosis model is pre-trained based on the obtained acoustic sample data; Ultrasonic sample data of overheat defects at different temperature levels at various positions is collected through the experimental model of the transformer, and the transformer overheat defect diagnosis model is fine-tuned through transfer learning based on the ultrasonic sample data; A temperature field inversion model is constructed, the temperature field inversion model comprises a SE attention mechanism module, a first convolutional neural network, a Transformer encoder and a second convolutional neural network which are sequentially connected, temperature field distribution signals of overheat defects at various positions are obtained by simulating overheat defects at different positions through the experimental model of the transformer and the three-dimensional simulation model, a data set required by the temperature field inversion model is established based on the temperature field distribution signals, and the temperature field inversion model is trained through the data set; Ultrasonic signals of a transformer to be measured are obtained, the ultrasonic signals are input into the trained transformer overheat defect diagnosis model for judgment, and the temperature field inversion model is used to perform temperature field inversion on the winding inside the transformer to be measured according to the judgment result of the transformer overheat diagnosis model, so that the overheat defect is located.

2. The method of claim 1, wherein, The transformer overheat defect diagnosis model is trained based on the obtained acoustic sample data, and the specific operation is as follows: The preprocessed acoustic sample data is input into the transformer overheat defect diagnosis model; Deep feature extraction and representation of original time domain signals of the acoustic sample data are realized through the cooperation of the one-dimensional convolution layer and the ConvMixer module, and key feature information related to the overheat defect category in the ultrasonic signal is obtained; The key feature information is mapped into a classification vector with a fixed dimension through the full connection layer, and the probability distribution corresponding to each category of the overheat defect is output through the activation function; The overheat defect category with the maximum probability value is selected according to the probability distribution, and a diagnosis result is output.

3. The method of claim 1, wherein, The transformer overheat defect diagnosis model is fine-tuned through transfer learning, and the specific method is as follows: All convolution layers of the transformer overheat defect diagnosis model are frozen, the weights of the convolution layers are fixed, and only the parameters of the full connection layer are updated; A preset training iteration number is set, the ultrasonic sample data under the experimental condition is used for forward propagation of the transformer overheat defect diagnosis model, and an output prediction result is obtained; the output prediction result is compared with actual experimental condition information, and a loss value is calculated; The parameter gradient of the full connection layer is calculated through back propagation according to the loss value; The weights of the full connection layer are adjusted according to the calculated parameter gradient. determine whether the current iteration number reaches the preset training iteration number, if not, continue to use the ultrasonic sample data for forward propagation, and adjust the weights of the full connection layer according to the new loss value; If the current iteration number reaches the preset training iteration number, stop training and save the trained transformer overheating defect diagnosis model.

4. The method of claim 1, wherein, It also includes three-dimensional visualization of the temperature inversion model for overheating defect positioning, and the specific method is: Extract the spatial coordinates of all temperature measurement points. x, y, z The information and its corresponding temperature prediction results are converted from rectangular coordinates to polar coordinates. at an angle θ and a height Z The direction generates the sampling points of uniform distribution, and generates two-dimensional grid; The temperature data is processed by cubic spline interpolation, the grid after interpolation is converted back to the rectangular coordinate system, the temperature field surface is drawn, and the color mapping is performed, the predicted temperature data is associated with the color; The temperature data is normalized, a temperature scale is added, and the viewing angle is adjusted to optimize the three-dimensional visualization effect.

5. The method of claim 1, wherein, The grid division of the three-dimensional simulation model specifically includes heating sheet division, heating rod division, solid region division, and geometric entity layer division.

6. A transformer overheat defect detection and positioning system based on transfer learning, applying the transformer overheat defect detection and positioning method based on transfer learning of claim 1, characterized in that, It includes: A simulation module is used to establish a three-dimensional simulation model of a transformer experimental model, divide the three-dimensional simulation model into grids, and set the material physical parameters of each structure in the three-dimensional simulation model; A diagnosis module is used to simulate overheating defects at different positions through the three-dimensional simulation model of the transformer, obtain acoustic sample data at different temperature levels for each overheating defect, and simulate acoustic sample data without overheating defects; A transformer overheating defect diagnosis model is constructed, and the transformer overheating defect diagnosis model is pre-trained based on the obtained acoustic sample data; A training module is used to collect ultrasonic sample data of overheating defects at different temperature levels in various positions by using a transformer experimental model, and to fine-tune the transformer overheating defect diagnosis model through transfer learning based on the ultrasonic sample data; An inversion module is used to construct a temperature field inversion model, simulate overheating defects at different positions through the transformer experimental model and the three-dimensional simulation model, obtain temperature field distribution signals of overheating defects at each position, establish a data set required by the temperature field inversion model based on the temperature field distribution signals, and train the temperature field inversion model through the data set; A processing module is used to obtain an ultrasonic signal of a transformer to be tested, input the ultrasonic signal into the trained transformer overheating defect diagnosis model for judgment, and locate the overheating defect through temperature field inversion of the internal winding of the transformer to be tested based on the judgment result of the transformer overheating diagnosis model.

7. A transformer overheat defect detection and positioning device based on transfer learning, characterized in that, It includes a processor and a memory, wherein the processor executes the computer program saved in the memory to realize the transfer learning transformer overheating defect detection and positioning method of any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, It is used to store a computer program, wherein the computer program is executed by a processor to realize the transfer learning transformer overheating defect detection and positioning method of any one of claims 1-5.

Images