Oil-immersed transformer turn-to-turn short circuit fault identification method based on convolutional neural network

Abstract

The invention relates to the technical field of power equipment fault detection, in particular to an oil-immersed transformer turn-to-turn short circuit fault identification method based on a convolutional neural network, and the method comprises the steps: collecting the multi-parameter characteristics of an internal fault of an oil-immersed transformer through a TMR current sensor and an optical infrared sensor; a convolutional neural network of a Swin Transform architecture is adopted to carry out data set image processing; and an AdamW optimizer is adopted to carry out model optimization and fault identification on the Swin Transform. The method has the beneficial effects that the actual engineering requirements are combined, and non-electrical parameters such as transformer main loop current, loss electrical parameters and transformer shell temperature are taken as the basis for fault identification, so that the interference resistance of the method is remarkably improved; the turn-to-turn short circuit fault of the oil-immersed transformer winding is identified by jointly utilizing a plurality of parameters, and the limitation of only depending on a single electrical parameter is avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault detection, and in particular to a method for identifying an oil-immersed transformer turn-to-turn short-circuit fault based on a convolutional neural network. Background Art

[0002] Most of the current methods for identifying inter-turn short-circuit faults in oil-immersed transformers are based on circuit models and single electrical parameters for research and analysis. However, with the occurrence of faults of varying degrees, fault identification based on a single electrical parameter has limitations and is not accurate enough. The coupling algorithm commonly used in current fault identification technology has poor adaptability. When the current fault identification method uses a classification algorithm, the scores of similar categories will not differ much (for example, 1% and 2% are 0.78 and 0.77 respectively). This is because the differences between different categories in the data set are not particularly large. Commonly used coupling algorithms may be adapted to target detection or other classification tasks, and cannot cope well with situations where the characteristics of power parameter categories are similar (similar scores between categories). With the expansion of the scale and increase in complexity of power systems, accurate and timely fault detection is crucial to the reliability of equipment operation. Summary of the invention

[0003] In view of the above problems or problems existing in the prior art, the present invention is proposed.

[0004] Therefore, the purpose of the present invention is to provide a method for identifying inter-turn short-circuit faults of oil-immersed transformers based on convolutional neural networks, which can realize high-precision intelligent detection of transformers, and combine with engineering practice to realize multi-parameter joint fault identification, avoiding the one-sidedness of identification methods relying solely on a single parameter or electrical parameter, and improving the anti-interference ability of the identification method.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: a method for identifying inter-turn short-circuit faults of oil-immersed transformers based on convolutional neural networks, which includes collecting multi-parameter features after internal faults of oil-immersed transformers through TMR current sensors and optical infrared sensors;

[0006] The convolutional neural network with Swin Transformer architecture is used to process the image data set;

[0007] AdamW optimizer is used to perform model optimization and fault identification on Swin Transformer.

[0008] As a preferred solution of the oil-immersed transformer inter-turn short-circuit fault identification method based on convolutional neural network of the present invention, TMR current sensor and optical infrared sensor collect a certain number of transformer winding main loop currents under different inter-turn short-circuit fault degrees and calculate the loss, and the infrared image of the transformer casing is used as a data set sample.

[0009] As a preferred solution of the oil-immersed transformer inter-turn short-circuit fault identification method based on convolutional neural network of the present invention, the TMR current sensor includes a primary coil and a secondary coil, the primary coil and the secondary coil are coupled through a magnetic core, the TMR current sensor detects the current change passing through the wire and generates a corresponding induced current in the secondary coil, and forms a voltage signal by measuring the induced current in the secondary coil. The TMR current sensor transmits the voltage signal to a data acquisition system, and the data acquisition system reads and records the voltage signal.

[0010] As a preferred solution of the oil-immersed transformer inter-turn short-circuit fault identification method based on convolutional neural network of the present invention, the current loss calculation includes resistance loss and iron loss, and the current change is continuously monitored by TMR current sensor. In combination with the electrical parameters of the transformer, the power loss of the transformer is calculated, wherein the resistance loss formula is as follows:

[0011] P loss =I 2 R.

[0012] As a preferred solution of the oil-immersed transformer inter-turn short-circuit fault identification method based on convolutional neural network of the present invention, an optical infrared sensor measures the surface temperature of the transformer housing, and the optical infrared sensor includes an infrared detector, and the measurement process is as follows:

[0013] The infrared detector detects the infrared radiation emitted by the transformer housing;

[0014] According to the intensity of infrared radiation received by the infrared detector, the optical infrared sensor calculates the surface temperature of the transformer casing;

[0015] The infrared sensor converts the detected temperature information into an electrical signal and transmits the electrical signal to the data acquisition system.

[0016] As a preferred solution of the oil-immersed transformer inter-turn short-circuit fault identification method based on convolutional neural network of the present invention, wherein: the convolutional neural network of Swin Transformer architecture, the network structure is Transformer.

[0017] As a preferred solution of the oil-immersed transformer inter-turn short-circuit fault identification method based on convolutional neural network of the present invention, the data set image processing process includes input image division, linear embedding, hierarchical window attention, window shifting, and stacked blocks.

[0018] As a preferred solution of the oil-immersed transformer inter-turn short-circuit fault identification method based on convolutional neural network of the present invention, the AdamW optimizer combines the Adam optimizer and weight decay, which can reduce the model parameter gradient, including initializing two 0 matrices v and m, the learning rate α, the decay rates β1 and β2, and the smoothing term ε.

[0019] As a preferred solution of the oil-immersed transformer inter-turn short-circuit fault identification method based on convolutional neural network of the present invention, wherein: AdamW optimizer optimizes the Swin Transformer model, wherein the algorithm steps of each iteration are as follows:

[0020] Calculate the current gradient g t , the g t represents the gradient at the tth iteration, that is, the derivative of the loss function with respect to the model parameters;

[0021] Update the first moment estimate:

[0022] m t =β1·m t-1 +(1-β1)·g t

[0023] Update the second moment estimate:

[0024] m t =β1·m t-1 +(1-β1)·g t

[0025] Perform bias correction on mt and vt:

[0026]

[0027] Among them, m t is the exponentially weighted average of the gradient, v t is the exponentially weighted average of the squared gradients, and represents the bias-corrected first- and second-order moment estimates;

[0028] Implement weight decay on θ:

[0029] θ t ' = θ t-1 -ηλθ t-1

[0030]

[0031] Among them, θ t represents the parameters after weight decay is applied, θ t-1 represents the model parameters after the (t-1)th iteration, θt+1 represents the model parameters after the t+1th iteration, η represents the learning rate, λ is the weight decay rate, is the L2 regularization coefficient, and β1 and β2 are the decay rates.

[0032] As a preferred solution of the oil-immersed transformer inter-turn short-circuit fault identification method based on convolutional neural network of the present invention, AdamW optimizer optimizes the Swin Transformer model, wherein the specific training process is as follows:

[0033] Set respective category labels for inter-turn short-circuit faults of different degrees, and process all collected transformer main circuit current, loss waveform and transformer shell infrared temperature images;

[0034] The processed data set is randomly shuffled and all images are divided into training set, validation set and test set in a ratio of 8:1:1;

[0035] Secondly, the convolutional neural network is trained with the training set, the number of iterations is set, and it is iterated multiple times to output the trained model;

[0036] Finally, the trained model is used to perform forward propagation on the test set to obtain the prediction results, and evaluation criteria, including accuracy, precision, and recall, are used to measure the performance of the model on the test set.

[0037] Beneficial effects of the invention: The present invention combines actual engineering needs and uses non-electrical parameters such as transformer main circuit current, loss electrical parameters and transformer casing temperature as the basis for fault identification, which significantly improves the method's resistance to interference and realizes the combined use of multiple parameters to identify turn-to-turn short-circuit faults in oil-immersed transformer windings, avoiding the limitation of relying solely on a single electrical parameter. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0039] Figure 1 Schematic diagram of the winding current waveform of the oil-immersed transformer turn-to-turn short-circuit fault identification method based on convolutional neural network.

[0040] Figure 2 Schematic diagram of loss waveform of oil-immersed transformer turn-to-turn short-circuit fault identification method based on convolutional neural network.

[0041] Figure 3Schematic diagram of transformer casing temperature for oil-immersed transformer turn-to-turn short-circuit fault identification method based on convolutional neural network.

[0042] Figure 4 Schematic diagram of the SwinTransformer model structure for the oil-immersed transformer inter-turn short-circuit fault identification method based on convolutional neural network.

[0043] Figure 5 This is a schematic diagram of the model identification process of the oil-immersed transformer turn-to-turn short-circuit fault identification method based on convolutional neural network.

[0044] Figure 6 Schematic diagram of the data set test results of the oil-immersed transformer inter-turn short-circuit fault identification method based on convolutional neural network. DETAILED DESCRIPTION

[0045] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0046] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0047] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0048] Example 1

[0049] Reference Figure 1 to Figure 6 , is an embodiment of the present invention, which provides an oil-immersed transformer inter-turn short-circuit fault identification method based on convolutional neural network, which can realize high-precision intelligent detection of transformers, and combine with engineering practice to realize multi-parameter joint fault identification, avoiding the one-sidedness of the identification method relying only on a single parameter or electrical parameter, and improving the anti-interference ability of the identification method.

[0050] Specifically, the multi-parameter characteristics of the oil-immersed transformer after the internal fault are collected through the transformer coupled magnetoresistance sensor (TMR current sensor) and the optical infrared sensor;

[0051] The convolutional neural network with Swin Transformer architecture is used to process the image data set;

[0052] AdamW optimizer is used to perform model optimization and fault identification on Swin Transformer.

[0053] Furthermore, the TMR current sensor and the optical infrared sensor collect a certain number of transformer winding main circuit currents under different degrees of inter-turn short-circuit faults and calculate the losses, as well as the infrared images of the transformer casing as data set samples.

[0054] Furthermore, the TMR current sensor includes a primary coil and a secondary coil. The primary coil and the secondary coil are coupled through a magnetic core. The TMR current sensor detects the current change passing through the wire and generates a corresponding induced current in the secondary coil. A voltage signal is formed by measuring the induced current in the secondary coil. The TMR current sensor transmits the voltage signal to a data acquisition system, and the data acquisition system reads and records the voltage signal.

[0055] Furthermore, the current loss calculation includes resistance loss and iron loss. The current change is continuously monitored by the TMR current sensor, and the power loss of the transformer is calculated in combination with the electrical parameters of the transformer. The resistance loss formula is as follows:

[0056] P loss =I 2 R.

[0057] Furthermore, the optical infrared sensor measures the surface temperature of the transformer housing. The optical infrared sensor includes an infrared detector, and the measurement process is as follows:

[0058] The infrared detector detects the infrared radiation emitted by the transformer housing;

[0059] According to the intensity of infrared radiation received by the infrared detector, the optical infrared sensor calculates the surface temperature of the transformer casing;

[0060] The infrared sensor converts the detected temperature information into an electrical signal and transmits the electrical signal to the data acquisition system.

[0061] Furthermore, the convolutional neural network of the Swin Transformer architecture has a Transformer network structure.

[0062] Furthermore, the dataset image processing process includes input image partitioning, linear embedding, hierarchical window attention, window shifting, and cascading blocks.

[0063] Specifically, input image division: the input image is divided into small blocks of fixed size, such as 4x4 image blocks;

[0064] Linear embedding: Linearly embed each image block, flatten it and map it to a high-dimensional feature space to form an initial feature representation;

[0065] Hierarchical Window Attention: The window attention mechanism is applied to the image blocks in a hierarchical manner. At each layer, the image blocks are divided into multiple windows, and self-attention calculation is performed in each window to capture local features.

[0066] Window shift: At a specific level, the window position is shifted to ensure feature interaction across windows, thereby capturing a wider range of contextual information;

[0067] Cascading blocks: Swin Transformer consists of multiple cascading blocks, each of which contains multiple layers of window attention and window shift operations. With each cascading block, the feature representation becomes richer and more abstract.

[0068] Furthermore, the AdamW optimizer combines the Adam optimizer with weight decay to reduce the model parameter gradients, including initializing two zero matrices v and m, three hyperparameters of the learning rate α, decay rates β1 and β2, and a smoothing term ε. The two zero matrices v and m are used to store the exponentially decayed average and square of past gradients, respectively. The learning rate α, decay rates β1 and β2 are initialized to α=0.001, β1=0.9, and β2=0.999, respectively. The smoothing term ε is usually set to a very small number to prevent division by zero.

[0069] Furthermore, the AdamW optimizer optimizes the Swin Transformer model, where the algorithm steps for each iteration are as follows:

[0070] Calculate the current gradient g t , the g t represents the gradient at the tth iteration, that is, the derivative of the loss function with respect to the model parameters;

[0071] Update the first moment estimate:

[0072] m t =β1·m t-1 +(1-β1)·g t

[0073] Update the second moment estimate:

[0074] m t =β1·m t-1 +(1-β1)·g t

[0075] Perform bias correction on mt and vt:

[0076]

[0077]

[0078] Among them, m t is the exponentially weighted average of the gradient, v t is the exponentially weighted average of the squared gradients, and represents the bias-corrected first- and second-order moment estimates;

[0079] Implement weight decay on θ:

[0080] θ t ' = θ t-1 -ηλθ t-1

[0081]

[0082] Among them, θ t represents the parameters after weight decay is applied, θ t-1 represents the model parameters after the (t-1)th iteration, θ t+1 represents the model parameters after the t+1th iteration, η represents the learning rate, λ is the weight decay rate, is the L2 regularization coefficient, and β1 and β2 are the decay rates.

[0083] Furthermore, the AdamW optimizer optimizes the Swin Transformer model. The specific training process is as follows:

[0084] Set respective category labels for inter-turn short-circuit faults of different degrees, and process all collected transformer main circuit current, loss waveform and transformer shell infrared temperature images;

[0085] The processed data set is randomly shuffled and all images are divided into training set, validation set and test set in a ratio of 8:1:1;

[0086] Secondly, the convolutional neural network is trained with the training set, the number of iterations is set, and it is iterated multiple times to output the trained model;

[0087] Finally, the trained model is used to perform forward propagation on the test set to obtain the prediction results, and evaluation criteria, including accuracy, precision, and recall, are used to measure the performance of the model on the test set.

[0088] It should be noted that after setting respective category labels for inter-turn short-circuit faults of different degrees, the corresponding categories and label contents of each degree are summarized and sorted, and the table is as follows:

[0089]

[0090]

[0091] Table 1 Fault severity category labels

[0092] Specifically, Figure 6 As shown in the figure, the data set samples are divided into six categories, namely normal samples and samples with 1-5% winding inter-turn short circuit faults. Except for the different image contents, the three data sets have the same configuration. The image format is kept in the RGB color domain and the size is uniformly scaled to 224×224 to facilitate the unified training of the model. 600 groups of samples are collected for each category, so the total number of samples in a single data set is 3600, and then the training set, validation set and test set are evenly divided according to the ratio of 8:1:1.

[0093] By adopting an adaptive coupling strategy, multi-model coupling detection optimization is achieved. This strategy will first normalize the confidence of each model, and then compare the class with the greatest confidence as the prediction class, taking the maximum value of all models to determine the category as the final prediction category. This coupling method helps trustworthy models gain more support and gives full play to the unique characteristic attributes of each data set. Compared with the 1% accuracy improvement of other coupling strategies, the improvement of the adaptive coupling strategy is as high as 2.5%. Based on the adaptive coupling algorithm, this strategy achieved a classification accuracy of 90.1% in identifying the degree of short-circuit faults between turns of windings of oil-immersed transformers.

[0094] In summary, the present invention combines actual engineering needs and uses non-electrical parameters such as transformer main circuit current, loss electrical parameters and transformer casing temperature as the basis for fault identification, which significantly improves the method's resistance to interference and realizes the combined use of multiple parameters to identify turn-to-turn short-circuit faults in oil-immersed transformer windings, avoiding the limitation of relying solely on a single electrical parameter.

[0095] Importantly, it should be noted that the construction and arrangement of the present application shown in a plurality of different exemplary embodiments are only exemplary. Although only a few embodiments are described in detail in this disclosure, it should be readily understood by those who refer to this disclosure that many modifications are possible without substantially departing from the novel teachings and advantages of the subject matter described in the application (e.g., mounting arrangements, use of materials, color, changes in orientation, etc.). For example, the element shown as integrally formed may be composed of a plurality of parts or elements, the position of the element may be inverted or otherwise changed, and the nature or number or position of the discrete element may be altered or changed. Therefore, all such modifications are intended to be included within the scope of the present invention. The order or sequence of any process or method steps may be changed or reordered according to an alternative embodiment. In the claims, any "bracket plus function" clause is intended to cover the structure of the execution function described herein, and is not only structurally equivalent but also equivalent structure. Without departing from the scope of the present invention, other replacements, modifications, changes and omissions may be made in the design, operating conditions and arrangement of the exemplary embodiments. Therefore, the present invention is not limited to a specific embodiment, but extends to a variety of modifications that still fall within the scope of the appended claims.

[0096] Additionally, in an effort to provide a concise description of example embodiments, all features of an actual implementation may not be described.

[0097] It should be understood that in the development of any actual implementation, as in any engineering or design project, numerous implementation-specific decisions may be made. Such a development effort may be complex and time-consuming, but for those of ordinary skill having the benefit of this disclosure, the development effort will be a routine task of design, fabrication, and production without undue experimentation.

[0098] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for identifying turn-to-turn short-circuit faults in oil-immersed transformers based on convolutional neural networks, characterized in that: include, The multi-parameter characteristics of the oil-immersed transformer after internal fault are collected through TMR current sensor and optical infrared sensor; The convolutional neural network with Swin Transformer architecture is used to process the image data set; AdamW optimizer is used to perform model optimization and fault identification on Swin Transformer.

2. The method for identifying turn-to-turn short-circuit faults of oil-immersed transformers based on convolutional neural networks according to claim 1, characterized in that: The TMR current sensor and the optical infrared sensor collect a certain number of transformer winding main loop currents under different turn-to-turn short-circuit fault levels and calculate the losses, as well as the infrared images of the transformer housing as data set samples.

3. The method for identifying turn-to-turn short-circuit faults of oil-immersed transformers based on convolutional neural networks according to claim 2, characterized in that: The TMR current sensor includes a primary coil and a secondary coil, wherein the primary coil and the secondary coil are coupled via a magnetic core. The TMR current sensor detects the current change passing through the conductor and generates a corresponding induced current in the secondary coil. A voltage signal is formed by measuring the induced current in the secondary coil. The TMR current sensor transmits the voltage signal to a data acquisition system, and the data acquisition system reads and records the voltage signal.

4. The method for identifying turn-to-turn short-circuit faults of oil-immersed transformers based on convolutional neural networks according to claim 3, characterized in that: The current loss calculation includes resistance loss and iron loss. The TMR current sensor continuously monitors the current change and calculates the power loss of the transformer in combination with the electrical parameters of the transformer. The resistance loss formula is as follows: P loss =I 2 R。 5. The method for identifying turn-to-turn short-circuit faults of oil-immersed transformers based on convolutional neural networks according to any one of claims 2 to 4, characterized in that: The optical infrared sensor measures the surface temperature of the transformer housing. The optical infrared sensor includes an infrared detector. The measurement process is as follows: The infrared detector detects the infrared radiation emitted by the transformer housing; The optical infrared sensor calculates the surface temperature of the transformer casing according to the intensity of the infrared radiation received by the infrared detector; The infrared sensor converts the detected temperature information into an electrical signal and transmits the electrical signal to the data acquisition system.

6. The method for identifying turn-to-turn short-circuit faults of oil-immersed transformers based on convolutional neural networks according to claim 1, characterized in that: The convolutional neural network of the Swin Transformer architecture has a Transformer network structure.

7. The method for identifying turn-to-turn short-circuit faults of oil-immersed transformers based on convolutional neural networks according to claim 6, characterized in that: The image processing process of the dataset includes input image partitioning, linear embedding, hierarchical window attention, window shifting, and stacking blocks.

8. The method for identifying turn-to-turn short-circuit faults of oil-immersed transformers based on convolutional neural networks according to claim 1, characterized in that: The AdamW optimizer combines the Adam optimizer with weight decay to reduce the model parameter gradients, including initializing two zero matrices v and m, three hyperparameters of the learning rate α, decay rates β1 and β2, and the smoothing term ε.

9. The method for identifying turn-to-turn short-circuit faults of oil-immersed transformers based on convolutional neural networks according to claim 8, characterized in that: The AdamW optimizer performs model optimization on the Swin Transformer, wherein the algorithm steps of each iteration are as follows: Calculate the current gradient g t , the g t represents the gradient at the tth iteration, that is, the derivative of the loss function with respect to the model parameters; Update the first moment estimate: m t =β1·m t-1 +(1-β1)·g t Update the second moment estimate: m t =β1·m t-1 +(1-β1)·g t For m t and v t To perform bias correction: Among them, m t is the exponentially weighted average of the gradient, v t is the exponentially weighted average of the squared gradients, and represents the bias-corrected first- and second-order moment estimates; Implement weight decay on θ: i t '=θ t-1 -elth t-1 Among them, θ t represents the parameters after weight decay is applied, θ t-1 represents the model parameters after the (t-1)th iteration, θ t+1 represents the model parameters after the t+1th iteration, η represents the learning rate, λ is the weight decay rate, is the L2 regularization coefficient, and β1 and β2 are the decay rates.

10. The method for identifying turn-to-turn short-circuit faults of oil-immersed transformers based on convolutional neural networks according to claim 8 or 9, characterized in that: The AdamW optimizer performs model optimization on the Swin Transformer, wherein the specific training process is as follows: Set respective category labels for inter-turn short-circuit faults of different degrees, and process all collected transformer main circuit current, loss waveform and transformer shell infrared temperature images; The processed data set is randomly shuffled and all images are divided into training set, validation set and test set in a ratio of 8:1:1; Secondly, the convolutional neural network is trained with the training set, the number of iterations is set, and it is iterated multiple times to output the trained model; Finally, the trained model is used to perform forward propagation on the test set to obtain the prediction results, and evaluation criteria, including accuracy, precision, and recall, are used to measure the performance of the model on the test set.

Images