A method for automatic generation of transformer frequency response data based on deep learning
Through deep learning-based methods, a large data set of transformer winding frequency response is automatically generated, which solves the problems of high cost and complexity of fault detection in the existing technology, and realizes efficient and intelligent fault diagnosis and health status monitoring.
Patent Information
- Application Number
- CN202411801746.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-09
- Publication Date
- 2025-04-01
AI Technical Summary
The existing power transformer winding fault detection methods are costly and risky, and cannot effectively obtain frequency response data of different transformers. The modeling process is complex, making it difficult to meet the needs of efficient and intelligent fault detection.
The automatic generation method of transformer frequency response data based on deep learning is adopted. By constructing a transformer finite element model, winding deformation faults are simulated, high-frequency equivalent circuit topology is constructed, frequency response curves are generated, and a network training data set is automatically generated through the differential discrimination network.
It realizes efficient automatic generation of transformer winding frequency response data, avoids cost and complexity problems in traditional methods, and provides an efficient and scalable data support solution for transformer fault diagnosis and health status monitoring.
Smart Images
Figure CN119740430A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of power system and equipment fault diagnosis and artificial intelligence, and specifically to a method for automatically generating transformer frequency response data based on deep learning. Background Art
[0002] As the core equipment of the power system, power transformers undertake the key tasks of power transmission and conversion. The reliability of their operation directly affects the safety and stability of the power grid. However, under long-term high-load operation and complex environmental influences, transformer windings are susceptible to electrical and mechanical stress, resulting in winding deformation, loosening and other faults. Once these faults occur, they may cause equipment damage, power outages and even cause larger-scale power failures. Therefore, early fault detection and preventive maintenance of transformers are particularly important.
[0003] Existing power transformer winding fault detection mainly relies on frequency response analysis (FRA) technology. The methods for acquiring winding frequency response data are mainly divided into three categories: 1) Experimental-based winding frequency response data acquisition; 2) Winding frequency response data acquisition based on equivalent model construction; 3) Frequency response data expansion based on artificial intelligence. The experimental-based frequency response data acquisition method is expensive and risky, and it is impossible to collect a large number of frequency response data of different transformer models; the frequency response data acquisition method based on equivalent model construction has a complex modeling process and requires precise geometric structure, material and other information of the transformer;
[0004] In view of the limitations of these traditional methods, a more efficient and intelligent method for acquiring winding frequency response data is urgently needed. Summary of the invention
[0005] The object of the present invention is to provide a method for automatically generating transformer frequency response data based on deep learning, comprising the following steps:
[0006] 1) Construct a finite element model of the transformer;
[0007] 2) Based on the transformer finite element model, the winding deformation fault is simulated to obtain the influence of different winding deformation faults on the electrical parameters of the winding high-frequency equivalent circuit, so as to construct an electrical parameter change scheme for simulating different fault conditions;
[0008] 3) Construct the high-frequency equivalent circuit topology of the transformer winding and calculate the high-frequency equivalent circuit parameters of the healthy winding to construct the transformer frequency response equation;
[0009] 4) Generate multiple groups of high-frequency equivalent circuit parameters by random method, construct frequency response curves corresponding to each group of high-frequency equivalent circuit parameters, and label these frequency response curves; the label content is whether there is a significant difference or not; the frequency response curve without significant difference is the frequency response curve of the fault-free transformer, and the frequency response curve with significant difference is the frequency response curve of the faulty transformer;
[0010] 5) Obtain the actual frequency response curve of the transformer and mark it with a label where there is no significant difference;
[0011] 6) constructing a training data set for a frequency response curve difference discrimination network using the frequency response curves of step 4) and step 5);
[0012] 7) Use the training data set to train different image classification deep learning models to obtain multiple difference discrimination networks, compare the classification accuracy of these difference discrimination networks, and determine the best difference discrimination network;
[0013] 8) The healthy winding parameters are adjusted using electrical parameter change schemes under different fault conditions to obtain high-frequency equivalent circuit parameters under different fault conditions, and the corresponding frequency response data are generated. These frequency response data are used as a large frequency response data set for winding deformation fault diagnosis; the labels of the frequency response data are determined by the optimal difference discrimination network.
[0014] Furthermore, in step 1), the step of constructing a finite element model of the transformer is: using a finite element analysis tool to perform two-dimensional or three-dimensional modeling simulation on the transformer, thereby constructing a finite element model of the transformer.
[0015] Furthermore, electrical parameter variation schemes for simulating different fault conditions include a winding short circuit fault simulation scheme, a radial deformation fault simulation scheme, and a winding spacing variation fault simulation scheme.
[0016] Furthermore, the inter-winding short-circuit fault simulation scheme is: replace G in the KCL equation from winding unit a to winding unit b with h , C h , and R in the KVL equation h , L h The value is set to zero to simulate a short circuit fault from winding unit a to winding unit b;
[0017] The simulation scheme for radial deformation fault is: increase the KCL equation of winding unit a and C in KVL equation gh Increase and decrease C hl , to simulate the deformation fault of winding unit a, where C gh The highest percentage of increase was 42%, C hl The highest percentage reduction was 40.2%;
[0018] The simulation scheme for the fault of winding spacing change is: increase C in the KCL equation of winding unit a h , and C in the KVL equation for winding unit a+1 h , to simulate the fault of the change in the spacing between windings, where C h The highest percentage of enlargement was 75%.
[0019] When the lowest percentage of the adjustment value is 0, the parameter is not adjusted, that is, the parameter is in a healthy state.
[0020] Further, the KCL equation is as follows:
[0021]
[0022] In the formula, G gh , G hl , G h is the inductance; R0 is the resistance; is the voltage; is the current; C gh , C hl , C h is the capacitance; w is the angular frequency of the external sinusoidal excitation applied to the winding; the superscript “.” indicates phasor operation.
[0023] Further, the KVL equation is as follows:
[0024]
[0025] In the formula, Rh is the resistance; L h For inductance.
[0026] Further, the transformer frequency response equation is shown below:
[0027]
[0028] Where H(f) is the frequency response of the winding when the frequency of the external sinusoidal excitation is f; x(f) is the intermediate matrix.
[0029] Furthermore, when training the image classification deep learning model, the loss function L used is as follows:
[0030]
[0031] Where: p represents the predicted probability that the sample belongs to the positive class; y represents the true label of the sample, 1 represents "no significant difference", and 0 represents "significant difference"; α is the class balance coefficient, which is used to adjust the contribution of different categories to the loss function; γ is the focusing parameter used to adjust the weights of easy-to-classify and difficult-to-classify samples.
[0032] Furthermore, the input of the image classification deep learning model is a frequency response curve in image format, and the output is a label of the frequency response curve.
[0033] Furthermore, in step 4), the frequency response curve with significant difference refers to a frequency response curve with less than 6 or more than 20 resonance points in the frequency range of 1 kHz to 1 MHz.
[0034] The technical effect of the present invention is unquestionable. The present invention proposes a method for generating transformer frequency response data. The method is based on deep learning and aims to automatically generate a large data set of transformer winding frequency response. Unlike traditional methods that pursue accurate reproduction of actual transformer frequency response, this method focuses more on identifying and capturing the changing pattern of frequency response under different fault conditions. In this way, not only can the cost and complexity problems in traditional methods be effectively avoided, but also an efficient and scalable data support solution is provided for transformer fault diagnosis and health status monitoring. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 is the overall flow chart of the method;
[0036] Figure 2 A mind map automatically generated for transformer frequency response big data based on deep learning;
[0037] Figure 3 It is the finite element simulation model of transformer;
[0038] Figure 4 is the high frequency equivalent circuit topology of the transformer winding;
[0039] Figure 5 It is a typical radial deformation fault simulation model;
[0040] Figure 6 This is a case of "significantly different" frequency response;
[0041] Figure 7 This is a frequency response case of “no significant difference”;
[0042] Figure 8 This is the frequency response curve finally generated by the discriminant model. DETAILED DESCRIPTION
[0043] The present invention is further described below in conjunction with the embodiments, but it should not be understood that the above subject matter of the present invention is limited to the following embodiments. Without departing from the above technical ideas of the present invention, various substitutions and changes are made according to the common technical knowledge and customary means in the art, which should all be included in the protection scope of the present invention.
[0044] Embodiment 1:
[0045] See also Figures 1 to 8 , a method for automatically generating transformer frequency response data based on deep learning, comprising the following steps:
[0046] 1) Construct a finite element model of the transformer;
[0047] 2) Based on the transformer finite element model, the winding deformation fault is simulated to obtain the influence of different winding deformation faults on the electrical parameters of the winding high-frequency equivalent circuit, so as to construct an electrical parameter change scheme for simulating different fault conditions;
[0048] 3) Construct the high-frequency equivalent circuit topology of the transformer winding and calculate the high-frequency equivalent circuit parameters of the healthy winding to construct the transformer frequency response equation;
[0049] 4) Generate multiple groups of high-frequency equivalent circuit parameters by random method, construct frequency response curves corresponding to each group of high-frequency equivalent circuit parameters, and label these frequency response curves; the label content is whether there is a significant difference or not; the frequency response curve without significant difference is the frequency response curve of the fault-free transformer, and the frequency response curve with significant difference is the frequency response curve of the faulty transformer;
[0050] 5) Obtain the actual frequency response curve of the transformer and mark it with a label where there is no significant difference;
[0051] 6) constructing a training data set for a frequency response curve difference discrimination network using the frequency response curves of step 4) and step 5);
[0052] 7) Use the training data set to train different image classification deep learning models to obtain multiple difference discrimination networks, compare the classification accuracy of these difference discrimination networks (the classification accuracy of each evaluation model when the model is trained using the focal loss function and the traditional binary cross entropy loss function), and determine the best difference discrimination network;
[0053] 8) The healthy winding parameters are adjusted using electrical parameter change schemes under different fault conditions to obtain high-frequency equivalent circuit parameters under different fault conditions, and the corresponding frequency response data are generated. These frequency response data are used as a large frequency response data set for winding deformation fault diagnosis; the labels of the frequency response data are determined by the optimal difference discrimination network.
[0054] In step 1), the step of constructing a finite element model of the transformer is: using a finite element analysis tool to perform two-dimensional or three-dimensional modeling simulation on the transformer, thereby constructing a finite element model of the transformer.
[0055] The electrical parameter variation schemes used to simulate different fault conditions include the inter-winding short-circuit fault simulation scheme, the radial deformation fault simulation scheme, and the inter-winding spacing change fault simulation scheme.
[0056] The inter-winding short-circuit fault simulation scheme is: replace G in the KCL equation from winding unit a to winding unit b h , C h , and R in the KVL equation h , L h The value is set to zero to simulate a short circuit fault from winding unit a to winding unit b;
[0057] The simulation scheme for radial deformation fault is: increase the KCL equation of winding unit a and C in KVL equation gh Increase and decrease C hl , to simulate the deformation fault of winding unit a, where C gh The highest percentage of increase was 42%, C hl The highest percentage reduction was 40.2%;
[0058] The simulation scheme for the fault of winding spacing change is: increase C in the KCL equation of winding unit a h , and C in the KVL equation for winding unit a+1 h , to simulate the fault of the change in the spacing between windings, where C h The highest percentage of enlargement was 75%.
[0059] When the lowest percentage of the adjustment value is 0, the parameter is not adjusted, that is, the parameter is in a healthy state.
[0060] The KCL equation is shown below:
[0061]
[0062] In the formula, G gh , G hl , G h is the inductance; R0 is the resistance; is the voltage; is the current; C gh , C hl , C h is the capacitance; w is the angular frequency of the external sinusoidal excitation applied to the winding; the superscript “.” indicates phasor operation.
[0063] The KVL equation is shown below:
[0064]
[0065] In the formula, R h is the resistance; L h For inductance.
[0066] The transformer frequency response equation is shown below:
[0067]
[0068] Where H(f) is the frequency response of the winding when the frequency of the external sinusoidal excitation is f; x(f) is the intermediate matrix.
[0069] When training the image classification deep learning model, the loss function L used is as follows:
[0070]
[0071] Where: p represents the predicted probability that the sample belongs to the positive class; y represents the true label of the sample, 1 represents "no significant difference", and 0 represents "significant difference"; α is the class balance coefficient, which is used to adjust the contribution of different categories to the loss function; γ is the focusing parameter used to adjust the weights of easy-to-classify and difficult-to-classify samples.
[0072] The input of the image classification deep learning model is the frequency response curve in image format, and the output is the label of the frequency response curve.
[0073] In step 4), the frequency response curve with significant difference refers to a frequency response curve with less than 6 or more than 20 resonance points in the frequency range of 1 kHz to 1 MHz.
[0074] Image classification deep learning models include convolutional neural networks AlexNet and VGG16;
[0075] The convolutional neural network AlexNet includes five convolutional layers, a maximum pooling layer after the partial convolutional layer, and three fully connected layers. Finally, the network uses SoftMax to output the probability of each category.
[0076] The VGG16 contains a total of 16 weight layers, 13 of which are convolutional layers and 3 are fully connected layers. In the convolutional layers of VGG16, small 3x3 convolution kernels are used, and maximum pooling layers are inserted between consecutive convolutional layers to reduce the spatial dimension of the feature map. The three fully connected layers in the latter section convert the high-dimensional feature maps from the convolutional layers into one-dimensional feature vectors for final decision making. After the fully connected layer, a SoftMax output layer follows to predict the category to which the image belongs and provide probabilistic output for classification tasks.
[0077] Embodiment 2:
[0078] A method for automatically generating transformer frequency response data based on deep learning, comprising the following steps:
[0079] 1) Construct a finite element model of the transformer;
[0080] 2) Based on the transformer finite element model, the winding deformation fault is simulated to obtain the influence of different winding deformation faults on the electrical parameters of the winding high-frequency equivalent circuit, so as to construct an electrical parameter change scheme for simulating different fault conditions;
[0081] 3) Construct the high-frequency equivalent circuit topology of the transformer winding and calculate the high-frequency equivalent circuit parameters of the healthy winding to construct the transformer frequency response equation;
[0082] 4) Generate multiple groups of high-frequency equivalent circuit parameters by random method, construct frequency response curves corresponding to each group of high-frequency equivalent circuit parameters, and label these frequency response curves; the label content is whether there is a significant difference or not; the frequency response curve without significant difference is the frequency response curve of the fault-free transformer, and the frequency response curve with significant difference is the frequency response curve of the faulty transformer;
[0083] 5) Obtain the actual frequency response curve of the transformer and mark it with a label where there is no significant difference;
[0084] 6) constructing a training data set for a frequency response curve difference discrimination network using the frequency response curves of step 4) and step 5);
[0085] 7) Use the training data set to train different image classification deep learning models to obtain multiple difference discrimination networks, compare the classification accuracy of these difference discrimination networks (that is, the classification accuracy of each evaluation model when the focal loss function and the traditional binary cross entropy loss function are used to train the model), and determine the best difference discrimination network;
[0086] 8) The healthy winding parameters are adjusted using electrical parameter change schemes under different fault conditions to obtain high-frequency equivalent circuit parameters under different fault conditions, and the corresponding frequency response data are generated. These frequency response data are used as a large frequency response data set for winding deformation fault diagnosis; the labels of the frequency response data are determined by the optimal difference discrimination network.
[0087] Embodiment 3:
[0088] A method for automatically generating transformer frequency response data based on deep learning, the technical content of which is the same as that of Example 2. Furthermore, in step 1), the step of constructing a finite element model of the transformer is: using a finite element analysis tool to perform two-dimensional or three-dimensional modeling simulation on the transformer, thereby constructing a finite element model of the transformer.
[0089] Embodiment 4:
[0090] A method for automatically generating transformer frequency response data based on deep learning, the technical content of which is the same as any one of Examples 2-3. Furthermore, electrical parameter change schemes for simulating different fault conditions include a winding short-circuit fault simulation scheme, a radial deformation fault simulation scheme, and a winding spacing change fault simulation scheme.
[0091] Embodiment 5:
[0092] A method for automatically generating transformer frequency response data based on deep learning, the technical content is the same as any one of embodiments 2-4, further, the inter-winding short-circuit fault simulation scheme is: replace G in the KCL equation from winding unit a to winding unit b h , C h , and R in the KVL equation h , L h The value is set to zero to simulate a short circuit fault from winding unit a to winding unit b;
[0093] The simulation scheme for radial deformation fault is: increase the KCL equation of winding unit a and C in KVL equation gh Increase and decrease C hl , to simulate the deformation fault of winding unit a, where C gh The highest percentage of increase was 42%, C hl The highest percentage reduction was 40.2%;
[0094] The simulation scheme for the fault of winding spacing change is: increase C in the KCL equation of winding unit a h , and C in the KVL equation for winding unit a+1 h , to simulate the fault of the change in the spacing between windings, where C h The highest percentage of enlargement was 75%.
[0095] When the lowest percentage of the adjustment value is 0, the parameter is not adjusted, that is, the parameter is in a healthy state.
[0096] Embodiment 6:
[0097] A method for automatically generating transformer frequency response data based on deep learning, the technical content is the same as any one of embodiments 2-5, and further, the KCL equation is as follows:
[0098]
[0099] In the formula, G gh , G hl , G h is the inductance; R0 is the resistance; is the voltage; is the current; C gh , C hl , C h is the capacitance; w is the angular frequency of the external sinusoidal excitation applied to the winding; the superscript “.” indicates phasor operation.
[0100] Embodiment 7:
[0101] A method for automatically generating transformer frequency response data based on deep learning, the technical content is the same as any one of Embodiments 2-6, and further, the KVL equation is as follows:
[0102]
[0103] Embodiment 8:
[0104] A method for automatically generating transformer frequency response data based on deep learning, the technical content is the same as any one of embodiments 2-7, and further, the transformer frequency response equation is as follows:
[0105]
[0106] In the formula, h=[0,…,0,1,0,…,0]; H(f) is the frequency response of the winding when the frequency of the external sinusoidal excitation is f.
[0107] Embodiment 9:
[0108] A method for automatically generating transformer frequency response data based on deep learning, the technical content is the same as any one of embodiments 2-8, and further, when training the image classification deep learning model, the loss function used is as follows:
[0109]
[0110] Where: p represents the predicted probability that the sample belongs to the positive class; y represents the true label of the sample, 1 represents "no significant difference", and 0 represents "significant difference"; α is the class balance coefficient, which is used to adjust the contribution of different categories to the loss function; γ is the focusing parameter used to adjust the weights of easy-to-classify and difficult-to-classify samples.
[0111] Embodiment 10:
[0112] A method for automatically generating transformer frequency response data based on deep learning, the technical content of which is the same as any one of Examples 2-9, and further, the input of the image classification deep learning model is a frequency response curve in image format, and the output is a label of the frequency response curve.
[0113] Embodiment 11:
[0114] A method for automatically generating transformer frequency response data based on deep learning, the technical content of which is the same as any one of Examples 2-10, and further, in step 4), the frequency response curve with significant differences refers to a frequency response curve with less than 6 or more than 20 resonance points in the frequency range of 1kHz to 1MHz.
[0115] Embodiment 12:
[0116] A method for automatically generating transformer frequency response data based on deep learning, the technical content of which is the same as any one of Embodiments 2-11, and further, the image classification deep learning model includes convolutional neural networks AlexNet and VGG16;
[0117] The convolutional neural network AlexNet includes five convolutional layers, a maximum pooling layer after the partial convolutional layer, and three fully connected layers. Finally, the network uses SoftMax to output the probability of each category.
[0118] The VGG16 contains a total of 16 weight layers, 13 of which are convolutional layers and 3 are fully connected layers. In the convolutional layers of VGG16, small 3x3 convolution kernels are used, and maximum pooling layers are inserted between consecutive convolutional layers to reduce the spatial dimension of the feature map. The three fully connected layers in the latter section convert the high-dimensional feature maps from the convolutional layers into one-dimensional feature vectors for final decision making. After the fully connected layer, a SoftMax output layer follows to predict the category to which the image belongs and provide probabilistic output for classification tasks.
[0119] Embodiment 13:
[0120] A method for automatically generating transformer frequency response data based on deep learning, the steps are as follows:
[0121] Step 1: Based on the finite element analysis tool, perform two-dimensional or three-dimensional modeling and simulation on the actual transformer. Simulate different types and degrees of winding deformation faults in the finite element model, summarize the impact of different winding deformation faults on the electrical parameters of the winding high-frequency equivalent circuit, and reveal the main change rules of electrical parameters under different fault conditions.
[0122] Step 2: Construct the high-frequency equivalent circuit topology of the transformer winding. Based on the calculation results of step 1, calculate the high-frequency equivalent circuit parameters of the healthy winding. Derive the mathematical and programmatic expression from the high-frequency equivalent circuit to the frequency response data, laying the foundation for the subsequent automatic generation of winding frequency response big data.
[0123] Step 3: Generate multiple frequency response curves of the transformer. For the preset electrical component parameters of the equivalent circuit topology in step 2, generate multiple high-frequency equivalent circuit parameter groups corresponding to multiple transformers by random assignment within the preset parameter range, and draw the corresponding frequency response curve for each group of parameters. These frequency response curves corresponding to different random parameter groups are divided into two categories: "significant difference exists" and "no significant difference exists".
[0124] Step 4: Train the difference discrimination model. Through manual labeling, establish a training data set for the frequency response curve difference discrimination network, and introduce the actual frequency response curve (marked as "no significant difference") to expand the data set. Then, use different image classification deep learning models to train the difference discrimination network, and compare the performance of different models to obtain the final difference discrimination model.
[0125] Step 5: Obtain the frequency response of the transformer under different faults. Re-generate multiple parameter groups on the equivalent circuit topology in step 2, corresponding to multiple transformers, and draw the corresponding frequency response curves, which are then input into the difference discrimination model. Keep the data with "no significant difference" as the parameters of the healthy winding. Next, combined with the electrical parameter change law in step 1, the frequency response data under various winding states corresponding to different parameter groups are generated through the programmed expression from the winding high-frequency equivalent circuit to the frequency response. After sorting, a large frequency response data set for winding deformation fault diagnosis is finally formed.
[0126] Figure 2 Flowchart automatically generated for deep learning-based transformer frequency response data.
[0127] In step 1, this example uses a 10kVA, 11.55kV / 412V oil-immersed transformer as the experimental object, with a rated frequency of 50Hz and a double winding structure, which are high-voltage winding and low-voltage winding. Two-dimensional or three-dimensional modeling is performed on this transformer on the finite element analysis simulation software. By simulating different types of winding deformation of the transformer in the finite element model, the influence of different winding deformation faults on the resistance, capacitance, and inductance parameters of the high-frequency equivalent circuit of the winding is calculated, and the main change rules of the electrical parameters under different fault conditions are obtained.
[0128] In the step 2, a high-frequency equivalent circuit model topology of the transformer winding is established, including inductance L, capacitance C, resistance R and mutual inductance M. Relying on the simulation model of step 1, the high-frequency equivalent circuit parameters of the healthy winding are calculated, and the mathematical and programmatic expression from the high-frequency equivalent circuit to the frequency response data is derived.
[0129] In the step three, electrical component parameters are preset for the equivalent circuit model in the step two, and within the preset parameter range, multiple groups of parameter combinations are randomly generated, and according to the expression from the high-frequency equivalent circuit to the frequency response data in step two, the corresponding frequency response image is drawn through numerical calculation software, and according to the actual frequency response characteristics (such as the number and distribution of resonance points), the frequency response image is divided into two types of images: "there are significant differences" and "there are no significant differences".
[0130] In step 4, based on step 3, in order to expand the data set, the actual frequency response curve is introduced and marked as "no significant difference" as an additional sample to ensure the diversity and representativeness of the training data. Together with the samples in step 3, a training data set for the frequency response curve difference discrimination network is established by manual labeling. The frequency response curve is converted into a standardized image format and input into a variety of deep learning image classification models for training a difference discrimination network. The performance of these deep learning models is compared to obtain the final difference discrimination model.
[0131] In the above step five, the parameter groups of multiple winding high-frequency equivalent circuits are continuously generated on the equivalent circuit of step two using a random generation algorithm, corresponding to multiple transformers, and the corresponding frequency response curves are drawn, and the generated frequency response curves are input into the difference discrimination network in step four. The high-frequency equivalent circuit parameter group corresponding to the "significant difference" image is discarded, and the high-frequency equivalent circuit parameter group corresponding to the "no significant difference" image is retained and used as the healthy parameter group. Further, combined with the electrical parameter change law of step one, the parameter groups of these transformers in different winding fault states are obtained, and the fault frequency response data are generated using a programmed method. All the generated frequency response data are sorted out to eventually form a frequency response data set covering a variety of fault scenarios for intelligent diagnosis of winding deformation faults.
[0132] Embodiment 14:
[0133] A method for automatically generating transformer frequency response data based on deep learning, the steps are as follows:
[0134] Step 1: Based on the finite element analysis tool, perform two-dimensional or three-dimensional modeling and simulation on the actual transformer. Simulate different types and degrees of winding deformation faults in the finite element model, summarize the impact of different winding deformation faults on the electrical parameters of the winding high-frequency equivalent circuit, and reveal the main change rules of electrical parameters under different fault conditions.
[0135] Step 2: Construct the high-frequency equivalent circuit topology of the transformer winding. Based on the calculation results of step 1, calculate the high-frequency equivalent circuit parameters of the healthy winding. Derive the mathematical and programmatic expression from the high-frequency equivalent circuit to the frequency response data, laying the foundation for the subsequent automatic generation of winding frequency response big data.
[0136] Step 3: Generate multiple frequency response curves of the transformer. For the preset electrical component parameters of the equivalent circuit topology in step 2, generate multiple high-frequency equivalent circuit parameter groups corresponding to multiple transformers by random assignment within the preset parameter range, and draw the corresponding frequency response curve for each group of parameters. These frequency response curves corresponding to different random parameter groups are divided into two categories: "significant difference exists" and "no significant difference exists".
[0137] Step 4: Train the difference discrimination model. Through manual labeling, establish a training data set for the frequency response curve difference discrimination network, and introduce the actual frequency response curve (marked as "no significant difference") to expand the data set. Then, use different image classification deep learning models to train the difference discrimination network, and compare the performance of different models to obtain the final difference discrimination model.
[0138] Step 5: Obtain the frequency response of the transformer under different faults. Re-generate multiple parameter groups on the equivalent circuit topology in step 2, corresponding to multiple transformers, and draw the corresponding frequency response curves, which are then input into the difference discrimination model. Keep the data with "no significant difference" as the parameters of the healthy winding. Next, combined with the electrical parameter change law in step 1, the frequency response data under various winding states corresponding to different parameter groups are generated through the programmed expression from the winding high-frequency equivalent circuit to the frequency response. After sorting, a large frequency response data set for winding deformation fault diagnosis is finally formed.
[0139] In step 1, this example uses a 10kVA, 11.55kV / 412V oil-immersed transformer as the experimental object, with a rated frequency of 50Hz and a double winding structure, including a high-voltage winding and a low-voltage winding. A two-dimensional or three-dimensional model is constructed on the finite element analysis simulation software. The constructed model is as follows: Figure 3 shown.
[0140] By simulating different types of transformer winding deformation in the finite element model, the influence of different winding deformation faults on the resistance, capacitance, and inductance parameters of the winding high-frequency equivalent circuit is calculated, and the main change rules of electrical parameters under different fault conditions are obtained:
[0141] 1) The influence of the inter-cake spacing change fault on parameter changes is mainly reflected in the parameters of the moving cake and the cake whose distance from the moving cake has changed, including C hl , C gh , C h , while L h The changes in all pies were small and irregular.
[0142] 2) The main cause of the change in the spacing between cakes is C h After the failure occurs, the C of the cake that moves and the cake that increases the distance between them h The changes in the two parameters are similar, and both are more than six times the changes in other parameters.
[0143] 3) By increasing C h To simulate the fault of the spacing change between winding discs.
[0144] Figure 4 This is a typical radial deformation simulation diagram of a transformer.
[0145] In the step 2, a high-frequency equivalent circuit topology of the transformer winding is established, including inductance L, capacitance C, resistance R and mutual inductance M. The established circuit topology is shown in FIG. Figure 5 As shown. Relying on the simulation model of step 1, the high-frequency equivalent circuit parameters of the healthy winding are calculated. The mathematical and programmatic expressions from the high-frequency equivalent circuit to the frequency response data are derived. Taking node 1 and inductor branch 1 as an example, the relationship between current and voltage is analyzed:
[0146]
[0147] Through KCL and KVL, a comprehensive matrix equation including all node voltages and inductor currents is derived, which is expressed as formula (2):
[0148]
[0149] Further transform (2) into a matrix, expressed as formula (3)
[0150]
[0151] Input different f values into equation (3) to obtain the FRA of transformer windings at different frequencies.
[0152] According to the main variation rules of electrical parameters under different fault conditions in step 1, the frequency response calculation method under each fault state is introduced: ① Inter-pane short circuit fault: When a short circuit occurs between the ath unit and the bth unit of the high-voltage winding, the G in the ath to bth KCL equation can be replaced by h , C h and R in the KVL equations a to b h , L h ② Radial deformation fault: When the ath unit of the high-voltage winding has a radial deformation fault, the ath KCL equation and the C in the ath KVL equation can be gh Increase C by a certain percentage hl ③ Inter-pane spacing change fault: When the inter-pane spacing change fault occurs in the ath unit of the high-voltage winding, the C in the ath KCL equation and the a+1th KVL equation can be h Increase by a certain percentage.
[0153] In the step three, electrical component parameters are preset for the equivalent circuit topology of step two, and within the preset parameter range (the value range of the conductivity parameter is 0.0001-5000MS, the value range of the inductance parameter is 0.0001-5000mH, the value range of the capacitance parameter is 0.00001-5000uF, and the value range of the resistance parameter is 0-500Ω), multiple groups of parameter combinations are randomly generated, and according to the expression from the high-frequency equivalent circuit to the frequency response data in step two, the corresponding frequency response image is drawn by numerical calculation software, and according to the actual frequency response characteristics (within the frequency range of 1kHz to 1MHz, the generated frequency response curve with less than 2 or more than 30 resonance points is regarded as "significant difference"), the frequency response image is divided into two types of images: "significant difference" and "no significant difference". Figure 6 , 7 These are the "significant difference" and "no significant difference" graphs.
[0154] In step 4, based on step 3, in order to expand the data set, the actual frequency response curve is introduced and marked as "no significant difference" as an additional sample to ensure the diversity and representativeness of the training data. Together with the samples in step 3, a training data set for the frequency response curve difference discrimination network is established by manual labeling. The training data set is shown in Table 1.
[0155] Table 1 Training dataset
[0156]
[0157] The frequency response curve is converted into a standardized image format and input into the deep learning image classification model to train the difference discrimination network. In order to improve the classification accuracy of all samples and solve the problem caused by data imbalance, this example selects focal loss as the loss function, and the expression is shown in (4)
[0158]
[0159] Where: p represents the predicted probability that the sample belongs to the positive class, y represents the true label of the sample (1 represents "no significant difference", 0 represents "significant difference"). α is the class balance coefficient, which is used to adjust the contribution of different classes to the loss function. γ is the focusing parameter used to adjust the weights of easy-to-classify and difficult-to-classify samples.
[0160] A variety of deep learning image classification models were input as the basic architecture of the difference discrimination network. To ensure the reliability of the experimental results, the model under each configuration was repeated 10 times to calculate the average accuracy, thereby reducing the impact of chance. The performance of these models was compared to obtain the final difference discrimination model. The model performance comparison is shown in Table 2.
[0161] Table 2
[0162]
[0163] In the above step 5, within the preset parameter range, the random generation algorithm is reused to continuously generate parameter groups of multiple winding high-frequency equivalent circuits corresponding to multiple transformers, and the corresponding frequency response curves are drawn. The generated frequency response curves are input into the difference discrimination network in step 4, and the high-frequency equivalent circuit parameter groups corresponding to the "significant difference" images are discarded, and the high-frequency equivalent circuit parameter groups corresponding to the "no significant difference" images are retained and used as healthy parameter groups. Further, combined with the electrical parameter change law of step 1, the parameter groups of these transformers in different winding fault states are obtained, and the fault frequency response data are generated using a programmed method. All the generated frequency response data are sorted out to eventually form a frequency response data set covering a variety of fault scenarios, which can be used for intelligent diagnosis of winding deformation faults.
Claims
1. A method for automatically generating transformer frequency response data based on deep learning, characterized in that: The following steps are involved: 1) Construct a finite element model of the transformer. 2) Based on the transformer finite element model, the winding deformation fault is simulated to obtain the influence of different winding deformation faults on the electrical parameters of the winding high-frequency equivalent circuit, so as to construct an electrical parameter change scheme for simulating different fault conditions; 3) Construct the high-frequency equivalent circuit topology of the transformer winding and calculate the high-frequency equivalent circuit parameters of the healthy winding to construct the transformer frequency response equation; 4) Generate multiple groups of high-frequency equivalent circuit parameters by random method, construct frequency response curves corresponding to each group of high-frequency equivalent circuit parameters, and label these frequency response curves; the label content is whether there is a significant difference or not; the frequency response curve without significant difference is the frequency response curve of the fault-free transformer, and the frequency response curve with significant difference is the frequency response curve of the faulty transformer; 5) Obtain the actual frequency response curve of the transformer and mark it with a label where there is no significant difference; 6) constructing a training data set for a frequency response curve difference discrimination network using the frequency response curves of step 4) and step 5); 7) Use the training data set to train different image classification deep learning models to obtain multiple difference discrimination networks, compare the classification accuracy of these difference discrimination networks, and determine the best difference discrimination network; 8) The healthy winding parameters are adjusted using electrical parameter change schemes under different fault conditions to obtain high-frequency equivalent circuit parameters under different fault conditions, and the corresponding frequency response data are generated. These frequency response data are used as a large frequency response data set for winding deformation fault diagnosis; the labels of the frequency response data are determined by the optimal difference discrimination network.
2. The method for automatically generating transformer frequency response data based on deep learning according to claim 1, characterized in that: In step 1), the step of constructing a finite element model of the transformer is: using a finite element analysis tool to perform two-dimensional or three-dimensional modeling simulation on the transformer, thereby constructing a finite element model of the transformer.
3. The method for automatically generating transformer frequency response data based on deep learning according to claim 1 is characterized in that: The electrical parameter variation schemes used to simulate different fault conditions include the inter-winding short-circuit fault simulation scheme, the radial deformation fault simulation scheme, and the inter-winding spacing change fault simulation scheme.
4. The method for automatically generating transformer frequency response data based on deep learning according to claim 3 is characterized in that: The inter-winding short-circuit fault simulation scheme is: replace G in the KCL equation from winding unit a to winding unit b h , C h , and R in the KVL equation h , L h The value is set to zero to simulate a short-circuit fault from winding unit a to winding unit b; The simulation scheme for radial deformation fault is: increase the KCL equation of winding unit a and C in KVL equation gh Increase and decrease C hl , to simulate the deformation fault of winding unit a, where C gh The highest percentage of increase was 42%, C hl The highest percentage reduction was 40.2%; The simulation scheme for the fault of the change in the spacing between windings is as follows: increase the C in the KCL equation of the winding unit a. h , and C in the KVL equation for winding unit a+1 h , to simulate the fault of the winding spacing change, where C h The highest percentage of enlargement was 75%.
5. The method for automatically generating transformer frequency response data based on deep learning according to claim 4, characterized in that: The KCL equation is shown below: In the formula, G gh , G hl , G h is the inductance; R0 is the resistance; is the voltage; is the current; C gh , C hl , C h is the capacitance; w is the angular frequency of the external sinusoidal excitation applied to the winding; the superscript ". " indicates phasor operation.
6. The method for automatically generating transformer frequency response data based on deep learning according to claim 4, characterized in that: The KVL equation is shown below: In the formula, Rh is the resistance; L h For inductance.
7. The method for automatically generating transformer frequency response data based on deep learning according to claim 1, characterized in that: The transformer frequency response equation is shown below: Where H(f) is the frequency response of the winding when the frequency of the external sinusoidal excitation is f; x(f) is the intermediate matrix.
8. The method for automatically generating transformer frequency response data based on deep learning according to claim 1, characterized in that: When training the image classification deep learning model, the loss function L used is as follows: Where: p represents the predicted probability that the sample belongs to the positive class; y represents the true label of the sample, 1 represents "no significant difference", and 0 represents "significant difference"; α is the class balance coefficient, which is used to adjust the contribution of different categories to the loss function; γ is the focusing parameter used to adjust the weights of easy-to-classify and difficult-to-classify samples.
9. The method for automatically generating transformer frequency response data based on deep learning according to claim 1, characterized in that: The input of the image classification deep learning model is the frequency response curve in image format, and the output is the label of the frequency response curve.
10. The method for automatically generating transformer frequency response data based on deep learning according to claim 1, characterized in that: In step 4), the frequency response curve with significant difference refers to a frequency response curve with less than 6 or more than 20 resonance points in the frequency range of 1 kHz to 1 MHz.