Transformer winding fault diagnosis method, device and equipment and storage medium

Abstract

The invention discloses a transformer winding fault diagnosis method, device and equipment and a storage medium, which are applied to the field of fault diagnosis, and comprise the following steps: obtaining a frequency response curve graph of a to-be-detected transformer winding; inputting the frequency response curve graph into a pre-trained transformer winding fault diagnosis model, and outputting a diagnosis result of the to-be-detected transformer winding; wherein the transformer winding fault diagnosis model is a deep convolutional neural network comprising a residual module; the pre-trained transformer winding fault diagnosis model is obtained by training based on a training database comprising a normal frequency response curve graph, a single-fault frequency response curve graph and a multi-fault superposition frequency response curve graph. Even under the complex working condition of multi-fault superposition, the fault can be accurately identified and judged, the early-stage superposition fault can be timely found, intervention measures are taken, large-area power failure caused by fault expansion is avoided, the safety and stability of power grid operation are remarkably improved, and the economic loss is reduced.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis, and in particular to a method, apparatus, equipment and storage medium for diagnosing transformer winding faults. Background Technology

[0002] As the core equipment for power transmission and voltage transformation in a power system, the operating status of power transformers directly determines the safety and stability of the power grid. Transformer windings, subjected to long-term electrodynamic forces, thermal stress, and mechanical vibration, are prone to mechanical faults such as deformation, inter-turn short circuits, and changes in inter-turn spacing. These faults have a cumulative effect. If initial minor deformations are not diagnosed in time, they can gradually develop into serious accidents such as inter-turn short circuits and winding burnout, leading to widespread power outages and huge economic losses. Current transformer winding fault diagnosis methods mostly focus on single faults, lacking comprehensive fault data to effectively diagnose overlapping faults.

[0003] Therefore, how to achieve accurate diagnosis of superimposed faults and meet the monitoring needs of the complex and ever-changing operating conditions of modern power systems is a technical problem that urgently needs to be solved. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method, device, equipment and storage medium for diagnosing transformer winding faults, which solves the problem that the existing technology does not provide comprehensive diagnosis and does not meet the requirements of monitoring the complex and ever-changing operating conditions of modern power systems.

[0005] To solve the above-mentioned technical problems, the present invention provides a method for diagnosing transformer winding faults, comprising:

[0006] Obtain the frequency response curve of the transformer winding under test;

[0007] The frequency response curve is input into a pre-trained transformer winding fault diagnosis model, and the diagnosis result of the transformer winding to be tested is output.

[0008] The transformer winding fault diagnosis model is a deep convolutional neural network containing residual modules; the pre-trained transformer winding fault diagnosis model is trained based on a training database containing normal frequency response curves, single fault frequency response curves, and multi-fault superimposed frequency response curves.

[0009] Optionally, the single-fault frequency response curve includes a frequency response curve for a short-circuit fault between winding discs, a frequency response curve for a fault due to a change in the spacing between winding discs, a frequency response curve for a fault due to radial deformation of the winding, and a frequency response curve for a fault due to axial deformation of the winding. The multi-fault superimposed frequency response curve includes a frequency response curve for a fault due to axial deformation of the winding superimposed with a fault due to a change in the spacing between winding discs, and a frequency response curve for a fault due to radial deformation of the winding superimposed with a fault due to a change in the spacing between winding discs.

[0010] Optionally, the process of constructing the training database includes:

[0011] When the terminals at both ends of the winding are short-circuited by a wire, the frequency response curve of the short-circuit fault between the winding discs is obtained.

[0012] When simulating a fault in the spacing between the winding discs by connecting capacitors of different sizes in parallel, the frequency response curve of the fault in the spacing between the winding discs is obtained.

[0013] Collect the frequency response curves of the radial deformation fault of the winding corresponding to the winding with different degrees of radial deformation;

[0014] The axial deformation of the winding was simulated using an axially concave winding, and the frequency response curve of the axial deformation fault of the winding was obtained.

[0015] When the axial concave winding is used to simulate the axial deformation of the winding, and capacitors of different sizes are connected in parallel to simulate the fault of the change in the spacing between the winding discs, the frequency response curve of the axial deformation fault superimposed on the fault of the change in the spacing between the winding discs is obtained.

[0016] When the radial deformation of the windings is simulated by connecting capacitors of different sizes in parallel to simulate the fault of the spacing between the winding discs, the frequency response curve of the radial deformation fault of the windings superimposed on the fault of the spacing between the winding discs is obtained.

[0017] Optionally, the frequency response curve of the transformer winding under test is obtained, including:

[0018] A voltage detection signal with preset parameters is generated by a nanosecond pulse generator, and the voltage detection signal is transmitted to the protection circuit via a coaxial cable; the output terminal of the protection circuit is connected to the input terminal of the bushing capacitive coupling sensor.

[0019] The voltage detection signal, adjusted by the protection circuit, is coupled and injected into the first end of the winding of the transformer under test through the bushing capacitive coupling sensor. The adjusted voltage detection signal propagates along the axial direction of the winding of the transformer under test to the end of the winding.

[0020] When the end of the transformer winding under test is directly grounded, the response current signal propagated by the transformer winding under test is measured by a broadband current sensor connected in series with the grounding circuit.

[0021] When the end of the transformer winding to be tested is not directly grounded, the response voltage signal propagated by the transformer winding to be tested is measured by the bushing capacitance coupling sensor coupled between the end of the transformer winding to be tested and the ground.

[0022] The response current signal and response voltage signal in the time domain are converted into frequency domain data by fast Fourier transform, and the amplitude corresponding to each frequency point is extracted to obtain amplitude-frequency characteristic data.

[0023] The frequency response curve is plotted based on the amplitude-frequency characteristic data; the horizontal axis of the frequency response curve is frequency, and the vertical axis is amplitude.

[0024] Optionally, before inputting the frequency response curve into a pre-trained transformer winding fault diagnosis model and outputting the diagnostic results of the transformer winding under test, the method further includes:

[0025] The frequency response curves of each type in the training database are evenly divided into training set and test set;

[0026] A transformer winding fault diagnosis model is constructed, which includes convolutional blocks and identity blocks. Both the convolutional blocks and the identity blocks are designed based on the residual learning mechanism and are configured with parallel backbone feature extraction paths and residual edge paths.

[0027] The frequency response curves of each type in the training set are input into the transformer winding fault diagnosis model. The learnable parameters of the convolutional layer in the transformer winding fault diagnosis model are optimized by the backpropagation algorithm to complete the model training and obtain the trained transformer winding fault diagnosis model.

[0028] The frequency response curves of each type in the test set are input into the trained transformer winding fault diagnosis model for testing. By comparing the diagnostic results output by the model with the actual fault types of the samples in the test set, the fault classification accuracy and generalization ability of the model are evaluated.

[0029] When the model's fault classification accuracy and generalization ability reach the preset standard, the pre-trained transformer winding fault diagnosis model is obtained.

[0030] Optionally, the main feature extraction paths of the convolutional block and the identity block are sequentially configured with three sets of processing units; the first set of processing units includes a first two-dimensional convolutional layer, a batch normalization layer, and an activation function layer; the second set of processing units includes a second two-dimensional convolutional layer, a batch normalization layer, and an activation function layer; and the third set of processing units includes a third two-dimensional convolutional layer and a batch normalization layer.

[0031] The first two-dimensional convolutional layer is a 1x1 convolution used to reduce the dimension of the input feature map; the second two-dimensional convolutional layer is a 3x3 convolution used for feature extraction; and the third two-dimensional convolutional layer is a 1x1 convolution used to restore the number of channels in the feature map to the original channel dimension.

[0032] Optionally, the residual edge path of the convolutional block is configured with a two-dimensional convolutional layer and a batch normalization layer to match the output dimension of the backbone feature extraction path of the convolutional block; the residual edge path of the identity block directly performs identity mapping on the input.

[0033] The present invention also provides a transformer winding fault diagnosis device, comprising:

[0034] The acquisition module is used to acquire the frequency response curve of the transformer winding under test;

[0035] The diagnostic module is used to input the frequency response curve into a pre-trained transformer winding fault diagnosis model and output the diagnostic results of the transformer winding to be tested.

[0036] The transformer winding fault diagnosis model is a deep convolutional neural network containing residual modules; the pre-trained transformer winding fault diagnosis model is trained based on a training database containing normal frequency response curves, single fault frequency response curves, and multi-fault superimposed frequency response curves.

[0037] The present invention also provides a transformer winding fault diagnosis device, comprising:

[0038] Memory, used to store computer programs;

[0039] A processor is used to implement the transformer winding fault diagnosis method described above when executing the computer program.

[0040] The present invention also provides a computer-readable storage medium storing computer-executable instructions, which, when loaded and executed by a processor, implement the transformer winding fault diagnosis method described above.

[0041] As can be seen, this invention obtains the frequency response curve of the transformer winding to be tested; inputs the frequency response curve into a pre-trained transformer winding fault diagnosis model, and outputs the diagnosis result of the transformer winding to be tested; wherein, the transformer winding fault diagnosis model is a deep convolutional neural network containing a residual module; the pre-trained transformer winding fault diagnosis model is trained based on a training database containing normal frequency response curves, single fault frequency response curves, and multi-fault superposition frequency response curves. The training database of this method covers the full-scene feature data of multi-fault superposition, enabling the model to fully learn and master the feature coupling law under superposition faults, and can accurately distinguish the differences between normal state, single fault and multi-fault superposition state, effectively breaking through the limitations of existing methods in the field of superposition fault diagnosis; at the same time, with the deep convolutional neural network containing a residual module, it can further explore the subtle features of superposition faults, ensuring that even under the complex working conditions of multi-fault superposition, accurate identification and judgment of faults can be achieved, helping to detect the initial superposition faults in a timely manner and take intervention measures, avoiding the fault from expanding and causing large-scale power outages, significantly improving the safety and stability of power grid operation, and reducing economic losses.

[0042] In addition, the present invention also provides a transformer winding fault diagnosis device, equipment and storage medium, which also have the above-mentioned beneficial effects. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0044] Figure 1 A flowchart of a transformer winding fault diagnosis method provided in an embodiment of the present invention;

[0045] Figure 2 A frequency response curve of a short circuit fault between winding discs provided in an embodiment of the present invention;

[0046] Figure 3 A fault frequency response curve for varying winding disc spacing is provided in an embodiment of the present invention.

[0047] Figure 4 A frequency response curve of a winding radial deformation fault is provided in an embodiment of the present invention;

[0048] Figure 5 A frequency response curve of a winding axial deformation fault is provided in an embodiment of the present invention;

[0049] Figures 6(a) and 6(b) are frequency response curves of a winding axial deformation fault superimposed with a winding disc spacing change fault provided by an embodiment of the present invention.

[0050] Figure 7 This invention provides a frequency response curve of a winding radial deformation fault superimposed with a change in the spacing between winding discs;

[0051] Figure 8 A Conv Block structure diagram provided in an embodiment of the present invention;

[0052] Figure 9 A structural diagram of an Identity Block provided in an embodiment of the present invention;

[0053] Figure 10 An average accuracy graph of training results provided in an embodiment of the present invention;

[0054] Figure 11 This is a schematic diagram of the structure of a transformer winding fault diagnosis device provided in an embodiment of the present invention;

[0055] Figure 12 This is a schematic diagram of the structure of a transformer winding fault diagnosis device provided in an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] As the core equipment for power transmission and voltage transformation in a power system, the operating status of power transformers directly determines the safety and stability of the power grid. Transformer windings, subjected to electrodynamic forces, thermal stress, and mechanical vibration over long periods, are prone to mechanical faults such as radial deformation, axial deformation, inter-turn short circuits, and changes in inter-turn spacing. These faults have a cumulative effect—if initial minor deformations are not diagnosed in time, they can gradually develop into serious accidents such as inter-turn short circuits and winding burnout, leading to widespread power outages and causing huge economic losses.

[0058] Currently, transformer winding deformation fault diagnosis technology is mainly divided into two categories: offline and online. Offline methods (such as short-circuit impedance method, low-voltage pulse method, and traditional frequency response analysis method) require the transformer to be shut down. Although the principle is mature, it cannot achieve real-time monitoring and has low sensitivity to slight deformation. Online methods (such as vibration analysis method and partial discharge monitoring method) can detect during transformer operation, but are greatly affected by on-site electromagnetic interference, resulting in insufficient diagnostic accuracy, especially in dealing with multiple fault superposition scenarios. In addition, existing diagnostic technologies also have the following problems: (1) Limitations of single fault diagnosis: Existing research and applications mostly focus on the diagnosis of single faults (such as only radial deformation or only inter-panel short circuit), ignoring the superposition characteristics of transformer winding faults in actual engineering scenarios. Due to the complexity of fault causes (such as short-circuit current impact may simultaneously cause radial deformation and inter-panel spacing changes), winding faults often exist in the form of multiple fault superposition. However, existing methods have not established a frequency response curve analysis mechanism under multiple fault superposition conditions, and cannot distinguish between single faults and multiple fault superposition, which easily leads to missed judgments and misjudgments. (2) Confusion of multiple fault curve features: The effects of different types of faults on the frequency response curves overlap—for example, radial deformation mainly affects the curve amplitude, and the change in the spacing between the discs mainly affects the resonant peak and valley frequencies. When the two are superimposed, the curve features are mixed. Existing technologies lack research on the correspondence between the dominant fault and the characteristic frequency band, and cannot extract the feature information of each individual fault from the superimposed curves, resulting in a significant decrease in diagnostic accuracy. (3) Lack of multi-fault data support and standardized methods: Existing fault databases are mostly built based on single-fault simulation experiments and do not cover experimental data of multiple fault superposition. At the same time, the industry has not yet formed diagnostic standards for multi-fault superposition scenarios (such as curve difference thresholds and fault type judgment rules). When relying on expert experience for interpretation, the subjectivity is strong and it is difficult to promote to engineering practice.

[0059] In summary, existing diagnostic methods for transformer windings with multiple superimposed faults suffer from technical limitations, including the inability to identify superimposed faults, difficulties in analyzing curve features, and a lack of standardized methods. There is an urgent need for a method that can accurately extract frequency response features under superimposed fault conditions and achieve reliable diagnosis, in order to fill the gap in engineering applications and ensure the safe and stable operation of power transformers.

[0060] This invention, based on the pulse-coupled injection method, utilizes bipolar pulses as excitation. Fault simulation experiments demonstrate that when different faults are superimposed, there exist faults with significant and minor impacts on a specific frequency range, and the influence of different faults on the frequency response curve exhibits a superposition state. Furthermore, the variation patterns of the frequency response curves under single-fault and multi-fault conditions of transformers are investigated, and a deep convolutional neural network including a residual module is introduced to identify and classify different faults. Therefore, this method can address both single-fault and multi-fault identification in transformers, providing technical support for the diagnosis of multiple fault conditions in power transformers.

[0061] Please refer to Figure 1 , Figure 1 A flowchart illustrating a transformer winding fault diagnosis method provided in an embodiment of the present invention. The method may include:

[0062] S101: Obtain the frequency response curve of the transformer winding to be tested.

[0063] The execution subject in this embodiment is a terminal. This embodiment does not limit the type of terminal, as long as it can perform the operation of the transformer winding fault diagnosis method. In this embodiment, the object to be detected is the winding of a running or under-maintenance power transformer, and the goal is to obtain a frequency response curve that reflects the structural state of the winding. This curve uses frequency as the abscissa and response parameters such as the amplitude ratio and phase difference of the winding at the corresponding frequency as the ordinate. The curve shape is directly related to the physical structure of the winding (such as the number of turns, inter-turn spacing, and degree of winding deformation): the frequency response curves of a normal winding and those with faults such as deformation or short circuits (including superimposed faults) will exhibit characteristic differences.

[0064] Furthermore, the acquisition of the frequency response curve of the transformer winding under test can specifically include: generating a voltage detection signal with preset parameters through a nanosecond pulse generator; transmitting the voltage detection signal to a protection circuit via a coaxial cable; connecting the output of the protection circuit to the input of a bushing capacitive coupling sensor; coupling the voltage detection signal, regulated by the protection circuit, to the beginning of the transformer winding under test through the bushing capacitive coupling sensor; propagating the regulated voltage detection signal along the axial direction of the transformer winding under test to the end of the transformer winding under test; when the end of the transformer winding under test is directly grounded, measuring the response current signal propagated by the transformer winding under test through a broadband current sensor connected in series with the grounding loop; when the end of the transformer winding under test is not directly grounded, measuring the response voltage signal propagated by the transformer winding under test through a bushing capacitive coupling sensor coupled between the end of the transformer winding under test and ground; converting the time-domain response current signal and response voltage signal into frequency-domain data through a fast Fourier transform, extracting the amplitude corresponding to each frequency point, and obtaining amplitude-frequency characteristic data; and plotting a frequency response curve based on the amplitude-frequency characteristic data; the horizontal axis of the frequency response curve is frequency, and the vertical axis is amplitude. The protection circuit in this embodiment enables lossless forward injection of voltage detection signals and interface conversion. It also filters out some high-amplitude voltage signals, achieves impedance matching to ensure lossless forward signal transmission, and blocks high-amplitude reverse signals reflected from the transformer winding under test, preventing damage to the nanosecond pulse generator. The voltage detection signal is adjusted to a commonly used amplitude and pulse width, resulting in a waveform with good square wave characteristics, effectively meeting the excitation supply requirements of the pulse coupling injection method. The pulse is injected to the beginning and end of the winding using a bushing capacitive coupling sensor. Based on the transformer bushing dimensions, a 3cm wide copper metal tape can be selected as the bushing capacitive coupling sensor.

[0065] S102: Input the frequency response curve into the pre-trained transformer winding fault diagnosis model and output the diagnosis results of the transformer winding to be tested.

[0066] The transformer winding fault diagnosis model in this embodiment is a deep convolutional neural network (DCNN) containing residual modules. DCNNs can automatically extract high-level features through multi-layer nonlinear transformations, reducing the workload of manual feature design. The multi-layered structure of a DCNN enables the model to extract features layer by layer from low-level features (such as simple local patterns) to high-level features (such as global patterns). Features extracted from different layers can also be combined to form a comprehensive understanding of the data by the network, enabling the identification of more complex patterns.

[0067] In this embodiment, the pre-trained transformer winding fault diagnosis model is trained based on a training database containing normal frequency response curves, single-fault frequency response curves, and multi-fault superimposed frequency response curves. Furthermore, the single-fault frequency response curves in this embodiment can include frequency response curves for winding disc short-circuit faults, winding disc spacing variation faults, winding radial deformation faults, and winding axial deformation faults. The multi-fault superimposed frequency response curves can include frequency response curves for winding axial deformation faults superimposed with winding disc spacing variation faults and winding radial deformation faults superimposed with winding disc spacing variation faults. This embodiment does not limit the method of obtaining each type of curve.

[0068] Furthermore, the construction process of the training database in this embodiment may include: obtaining a frequency response curve of short-circuit fault between winding discs when the terminals at both ends of the winding are short-circuited by a wire; obtaining a frequency response curve of fault of varying spacing between winding discs when simulating fault of varying spacing between winding discs by connecting capacitors of different sizes in parallel; collecting frequency response curves of radial deformation faults of windings corresponding to windings with different degrees of radial deformation; obtaining a frequency response curve of axial deformation fault of windings by simulating axial deformation of windings using axially concave windings; obtaining a frequency response curve of axial deformation fault superimposed on fault of varying spacing between winding discs when axial deformation of windings is simulated by simulating fault of varying spacing between winding discs by connecting capacitors of different sizes in parallel; and obtaining a frequency response curve of radial deformation fault superimposed on fault of varying spacing between winding discs when simulating fault of varying spacing between winding discs based on windings with different degrees of radial deformation by connecting capacitors of different sizes in parallel.

[0069] For specific curve charts, please refer to [the relevant information]. Figures 2 to 7 , Figure 2 A frequency response curve of a short circuit fault between winding discs provided in an embodiment of the present invention; Figure 3A fault frequency response curve for varying winding disc spacing is provided in an embodiment of the present invention. Figure 4 A frequency response curve of a winding radial deformation fault is provided in an embodiment of the present invention; Figure 5 Figure 6(a) and Figure 6(b) are frequency response curves of a winding axial deformation fault superimposed with the change in the spacing between winding discs, provided in an embodiment of the present invention. Figure 7 This is a frequency response curve diagram of a winding radial deformation fault superimposed with a change in the spacing between winding discs, provided as an embodiment of the present invention. Figure 2 It can be observed that when this type of fault occurs, the resonant peaks and valleys in the low-frequency range of the frequency response curve show a slight shift, the resonant peaks and valleys in the mid-frequency range show a significant shift, and the amplitude of the resonant peaks and valleys in the high-frequency range changes significantly. Comparing the waveforms of short-circuit faults of different degrees, it can be observed that as the degree of short-circuit fault increases, the frequency response curve shows a rightward shift in the mid-to-high frequency range. Figure 3 It can be observed that the frequency response curve shows a slight shift in the resonant peaks and valleys in the low-frequency range, while the resonant peaks and valleys in the mid-to-high-frequency range exhibit a significant shift. Figure 4 It can be observed that when this type of fault occurs, the overall frequency band of the curve shows an amplitude shift, particularly noticeable in the mid-range and high-frequency bands where amplitude changes occur. Figure 5 It can be observed that in the offline state, the frequency response curve exhibits amplitude shift below 200kHz and varying degrees of amplitude shift above 300kHz. In the online state, the frequency response curve shows a similar change to the offline state, generally exhibiting amplitude shift. Figures 6(a) and 6(b) show that due to the transformer's symmetrical structure, when axial concave deformation is only applied to winding A, the same inter-winding spacing variation fault on both windings A and E results in similar effects on the frequency response curve. Similar to the results when different degrees of inter-winding spacing variation faults are applied to winding A, the inter-winding spacing fault has a dominant influence on the frequency response curve. Therefore, whether or not an axial concave deformation fault, which has a less significant impact on the frequency response curve, is applied to the corresponding joint where the inter-winding spacing variation fault occurs, also has a relatively small impact on the curve. Figure 7 It can be observed that, based on faults involving changes in the spacing between windings of the same degree, the change in the frequency response curve due to radial deformation of the added windings is manifested as an amplitude shift. This is consistent with the change pattern of the frequency response curve under normal transformer conditions. The significant amplitude shift and changes in resonant peaks and valleys within certain frequency bands in the frequency response curve suggest that, when faults significantly affect the frequency response curve, the presence of radial deformation or spacing change faults can be inferred by comparing the frequency response curve with that under normal conditions. This also confirms the "hybridized" characteristic of the frequency response curve when multiple faults are superimposed.

[0070] Furthermore, before inputting the frequency response curves into the pre-trained transformer winding fault diagnosis model and outputting the diagnostic results of the transformer winding to be tested, the process may further include: uniformly dividing the frequency response curves of various types in the training database into training and test sets; constructing a transformer winding fault diagnosis model containing convolutional blocks and identity blocks, both of which are designed based on residual learning mechanisms and configured with parallel backbone feature extraction paths and residual edge paths; inputting the frequency response curves of various types in the training set into the transformer winding fault diagnosis model, optimizing the learnable parameters of the convolutional layers in the transformer winding fault diagnosis model through backpropagation algorithm, completing model training, and obtaining the trained transformer winding fault diagnosis model; inputting the frequency response curves of various types in the test set into the trained transformer winding fault diagnosis model for testing, evaluating the model's fault classification accuracy and generalization ability by comparing the diagnostic results output by the model with the actual fault types of the samples in the test set; and obtaining the pre-trained transformer winding fault diagnosis model when the model's fault classification accuracy and generalization ability reach the preset standard.

[0071] This embodiment addresses the vanishing gradient problem in deep network training through residual connections, enabling the training of deeper networks. The aforementioned residual connection design alleviates the vanishing gradient problem during deep network training, ensuring the model's ability to extract deep features from the frequency response curve. The transformer winding fault diagnosis model in this embodiment includes convolutional blocks (ConvBlocks) and identity blocks (Identity Blocks), which respectively change the network's dimensionality and deepen the network. The structure of the ConvBlock is as follows... Figure 8 As shown, Figure 8 A Conv Block structure diagram is provided for an embodiment of the present invention; the structure of the Conv Block is as follows: Figure 9 As shown, Figure 9 This is a structural diagram of an Identity Block provided in an embodiment of the present invention. Figure 8 and Figure 9As can be seen, both blocks are divided into two parts: the left part is the main part, and the right part is the residual edge part. ConvBlock can be used to change the width, height, and number of channels of the output feature layer. The main feature extraction path of the above convolutional block and identity block is set with three sets of processing units in sequence; the first set of processing units includes the first two-dimensional convolutional layer (Conv2d), the batch normalization layer (BatchNorm), and the activation function layer (ReLU); the second set of processing units includes the second two-dimensional convolutional layer, the batch normalization layer, and the activation function layer; the third set of processing units includes the third two-dimensional convolutional layer and the batch normalization layer; wherein, the first two-dimensional convolutional layer is a 1x1 convolution, used to reduce the dimension of the input feature map; the second two-dimensional convolutional layer is a 3x3 convolution, used for feature extraction; the third two-dimensional convolutional layer is a 1x1 convolution, used to restore the number of channels of the feature map to the original channel dimension. The residual side paths of the aforementioned convolutional blocks are configured with two-dimensional convolutional layers and batch normalization layers to match the output dimension of the main feature extraction path of the convolutional block; the residual side paths of the identity blocks directly perform identity mapping on the input. Finally, the output of the main feature extraction path of the convolutional block is added to the output of the residual side paths of the convolutional block, and the addition result is processed using an activation function to obtain the output features of the convolutional block; the processing result of the main feature extraction path of the identity block is added to the identity mapping result of the residual side paths of the identity block, and then processed by an activation function to obtain the output of the identity block.

[0072] This embodiment introduces a deep convolutional neural network including a residual module to classify different faults, verifying the effectiveness of the convolutional neural network in fault classification, with an accuracy of 94.4%. (For reference...) Figure 10 , Figure 10 This is an average accuracy map of the training results provided in an embodiment of the present invention. By employing a deep convolutional neural network with residual modules, the gradient vanishing problem in deep networks is effectively alleviated by the residual connections. This allows for in-depth mining of subtle features in the frequency response curve that reflect the normal state of the winding, single faults, and multiple fault superpositions. In particular, it achieves more accurate feature differentiation than traditional shallow networks for complex feature coupling relationships under superposition faults.

[0073] The transformer winding fault diagnosis method provided in this embodiment of the invention is as follows: S101, the frequency response curve of the transformer winding to be tested is obtained; S102, the frequency response curve is input into a pre-trained transformer winding fault diagnosis model, and the diagnosis result of the transformer winding to be tested is output; wherein, the transformer winding fault diagnosis model is a deep convolutional neural network containing residual modules; the pre-trained transformer winding fault diagnosis model is trained based on a training database containing normal frequency response curves, single fault frequency response curves, and multi-fault superimposed frequency response curves. This method's training database covers full-scenario feature data of multiple fault superposition, enabling the model to fully learn and master the feature coupling rules under superposition faults. It can accurately distinguish the differences between normal states, single faults, and multiple fault superposition states, effectively overcoming the limitations of existing methods in the field of superposition fault diagnosis. Simultaneously, by combining it with a deep convolutional neural network containing residual modules, it can further mine subtle features of superposition faults, ensuring accurate fault identification and judgment even under complex operating conditions of multiple fault superposition. This helps to promptly detect initial superposition faults and take intervention measures, preventing fault escalation and large-scale power outages, significantly improving the safety and stability of power grid operation, and reducing economic losses. Furthermore, by establishing a training database of frequency response curves for single-fault to multi-fault superposition (including combinations such as radial / axial deformation and inter-pie spacing changes), a multi-feature mapping system is formed. It can also identify the dominant fault-frequency band correspondence when multiple faults are superpositioned (e.g., inter-pie spacing changes dominate the 700-950kHz frequency band). Compared with previous transformer winding fault identification technologies, it can handle more complex fault conditions and multiple superposition fault diagnosis, providing more sufficient information for power transformer operation and maintenance.

[0074] To facilitate understanding, the transformer winding fault diagnosis model is described in detail below: it consists of multiple residual modules, each containing multiple convolutional layers, including one convolutional layer with downsampling to match the dimensions of the input and output. To achieve fault classification, all curves in the training database are first initially classified, with 500 data points in each category. 80% of these are used as the training set and 20% as the test set, inputting them into the transformer winding fault diagnosis model. Then, convolution, max pooling, and network function operations are performed on the input curves. It is understood that when dividing the training and test sets, the curves of each type should be evenly distributed to ensure that both the training and test sets contain curves of each type.

[0075] During initialization, the default normalization layer is used to check and limit the grouping and basic width of the control convolution to 1 and 64 respectively. Finally, two convolutional layers, two normalization layers, and one activation function are initialized, and the downsampling layer and stride are set. Then, a residual block is defined using the forward propagation function to construct the basic unit in the transformer winding fault diagnosis model. In this process, the input variables are copied to the identity variable for residual connections. This step helps alleviate the vanishing gradient problem and speeds up training when training deep neural networks. Next, a module is defined to reduce the number of network parameters and computational cost. It contains three convolutional layers, three batch normalization layers, and one activation function. The first convolutional layer is a 1x1 convolution to reduce the dimensionality of the input feature map, the second convolutional layer is a 3x3 convolution for feature extraction, and the third convolutional layer is a 1x1 convolution to restore the number of channels in the feature map to the original channel dimension. Assume the input curve image size is H×W, and the convolution kernel (pooling window) size is... × The first convolutional layer reduces the number of channels in the input feature map to width. Then, using forward propagation, the input variables are processed through a series of convolutions, batch normalization, and activation functions before the result is returned. The entire process consists of convolutions, activation functions, max pooling, fully connected layers, and softmax (normalization exponential function). Assuming the input image size is H×W, and the convolution kernel (pooling window) size is... × Fill in with ( , ), stride is ( , If the convolutional layer output feature map is then:

[0076] ;

[0077] .

[0078] in, This indicates the height of the output feature map of the convolutional layer; This indicates the width of the feature map output by the convolutional layer; W represents the height of the input feature map (or input data) of the convolutional layer; W represents the width of the input feature map (or input data) of the convolutional layer. This indicates the amount of padding (i.e., the number of extra pixels added on both sides of the input height direction) in the input feature map during the convolution operation. This indicates the amount of padding applied in the width direction of the input feature map during the convolution operation; This represents the size of the convolution kernel in the height direction; This indicates the size of the convolution kernel in the width direction; This represents the stride of the convolution operation in the height direction (i.e., the number of pixels the convolution kernel slides in the height direction each time). This indicates the stride of the convolution operation in the width direction.

[0079] Convolution operations performed on two-dimensional images are called 2D convolution, commonly used for image feature extraction in computer vision tasks. In 2D convolution, the convolution kernel is a two-dimensional matrix that slides along the height and width directions of the image and calculates the dot product with the input data, thereby extracting local features of the image. Multiple convolution kernels are typically used to extract different features, and each kernel contains several learnable parameters that can be optimized using the backpropagation algorithm. Besides convolution operations, 2D convolution can also use techniques such as pooling and padding to adjust the size and shape of the output feature map, thus better adapting to different image processing tasks.

[0080] The transformer winding fault diagnosis device provided in the embodiments of the present invention will be described below. The transformer winding fault diagnosis device described below and the transformer winding fault diagnosis method described above can be referred to in correspondence.

[0081] Please refer to the details. Figure 11 , Figure 11 A schematic diagram of a transformer winding fault diagnosis device provided in an embodiment of the present invention may include:

[0082] The acquisition module 100 is used to acquire the frequency response curve of the transformer winding to be tested.

[0083] The diagnostic module 200 is used to input the frequency response curve into a pre-trained transformer winding fault diagnosis model and output the diagnostic results of the transformer winding to be tested.

[0084] The transformer winding fault diagnosis model is a deep convolutional neural network containing residual modules; the pre-trained transformer winding fault diagnosis model is trained based on a training database containing normal frequency response curves, single fault frequency response curves, and multi-fault superimposed frequency response curves.

[0085] Furthermore, based on the above embodiments, the single-fault frequency response curve may include a frequency response curve for a short-circuit fault between winding discs, a frequency response curve for a fault due to a change in the spacing between winding discs, a frequency response curve for a radial deformation fault in the winding, and a frequency response curve for a axial deformation fault in the winding. The multi-fault superimposed frequency response curve may include a frequency response curve for a fault due to a axial deformation fault superimposed on a fault due to a change in the spacing between winding discs, and a frequency response curve for a fault due to a radial deformation fault superimposed on a fault due to a change in the spacing between winding discs.

[0086] Furthermore, the process of constructing the aforementioned training database may include:

[0087] The first acquisition unit is used to acquire the frequency response curve of the short circuit fault between the winding discs when the terminals at both ends of the winding are short-circuited by a wire.

[0088] The second acquisition unit is used to acquire the frequency response curve of the winding inter-panel spacing change fault when simulating the inter-panel spacing change fault by connecting capacitors of different sizes in parallel.

[0089] The third acquisition unit is used to collect the frequency response curves of the radial deformation fault of the winding corresponding to the winding with different degrees of radial deformation.

[0090] The fourth acquisition unit is used to simulate the axial deformation of the winding using the axial concave winding and acquire the axial deformation fault frequency response curve of the winding.

[0091] The fifth acquisition unit is used to acquire the frequency response curve of the winding axial deformation fault superimposed with the winding inter-panel spacing change fault when the axial concave winding is used to simulate the winding axial deformation and the inter-panel spacing change fault is simulated by connecting capacitors of different sizes in parallel.

[0092] The sixth acquisition unit is used to acquire the frequency response curve of the radial deformation fault superimposed on the inter-panel spacing fault when the radial deformation windings of different degrees are simulated by connecting capacitors of different sizes in parallel.

[0093] Furthermore, based on the above embodiments, the acquisition module 100 may include:

[0094] A transmission unit is used to generate a voltage detection signal with preset parameters through a nanosecond pulse generator. The voltage detection signal is transmitted to the protection circuit via a coaxial cable. The output terminal of the protection circuit is connected to the input terminal of the bushing capacitive coupling sensor.

[0095] An injection unit is used to couple and inject the voltage detection signal regulated by the protection circuit into the first end of the winding of the transformer under test through the bushing capacitance coupling sensor. The regulated voltage detection signal propagates along the axial direction of the winding of the transformer under test to the end of the winding of the transformer under test.

[0096] The seventh acquisition unit is used to measure the response current signal propagated by the transformer winding under test by a broadband current sensor connected in series with the grounding circuit when the end of the winding of the transformer under test is directly grounded.

[0097] The eighth acquisition unit is used to measure the response voltage signal propagated by the transformer winding under test by means of the bushing capacitance coupling sensor coupled between the end of the transformer winding under test and the ground when the end of the transformer winding under test is not directly grounded.

[0098] The time-frequency conversion unit is used to convert the response current signal and response voltage signal in the time domain into frequency domain data through fast Fourier transform, extract the amplitude corresponding to each frequency point, and obtain amplitude-frequency characteristic data.

[0099] A plotting unit is used to plot the frequency response curve based on the amplitude-frequency characteristic data; the horizontal axis of the frequency response curve is frequency, and the vertical axis is amplitude.

[0100] Furthermore, based on the above embodiments, the transformer winding fault diagnosis device may further include:

[0101] The dataset partitioning module is used to evenly divide the frequency response curves of each type in the training database into training set and test set before inputting the frequency response curves into the pre-trained transformer winding fault diagnosis model and outputting the diagnosis results of the transformer winding to be tested.

[0102] The model building module is used to build a transformer winding fault diagnosis model containing convolutional blocks and identity blocks. Both the convolutional blocks and the identity blocks are designed based on the residual learning mechanism and are configured with parallel backbone feature extraction paths and residual edge paths.

[0103] The training module is used to input the frequency response curves of various types in the training set into the transformer winding fault diagnosis model, optimize the learnable parameters of the convolutional layer in the transformer winding fault diagnosis model through the backpropagation algorithm, complete the model training, and obtain the trained transformer winding fault diagnosis model.

[0104] The testing module is used to input the frequency response curves of various types in the test set into the trained transformer winding fault diagnosis model for testing. By comparing the diagnostic results output by the model with the actual fault types of the samples in the test set, the fault classification accuracy and generalization ability of the model are evaluated.

[0105] The determination module is used to obtain the pre-trained transformer winding fault diagnosis model when the fault classification accuracy and generalization ability of the model reach the preset standard.

[0106] Furthermore, based on the above embodiments, the main feature extraction paths of the convolutional block and the identity block are sequentially configured with three sets of processing units; the first set of processing units includes a first two-dimensional convolutional layer, a batch normalization layer, and an activation function layer; the second set of processing units includes a second two-dimensional convolutional layer, a batch normalization layer, and an activation function layer; and the third set of processing units includes a third two-dimensional convolutional layer and a batch normalization layer.

[0107] The first two-dimensional convolutional layer is a 1x1 convolution used to reduce the dimension of the input feature map; the second two-dimensional convolutional layer is a 3x3 convolution used for feature extraction; and the third two-dimensional convolutional layer is a 1x1 convolution used to restore the number of channels in the feature map to the original channel dimension.

[0108] Furthermore, based on the above embodiments, the residual edge path of the convolutional block is configured with a two-dimensional convolutional layer and a batch normalization layer to match the output dimension of the backbone feature extraction path of the convolutional block; the residual edge path of the identity block directly performs identity mapping on the input.

[0109] It should be noted that the order of the modules and units in the above-mentioned transformer winding fault diagnosis device can be changed without affecting the logic.

[0110] The transformer winding fault diagnosis device provided in this embodiment of the invention uses an acquisition module 100 to acquire the frequency response curve of the transformer winding to be tested; and a diagnosis module 200 to input the frequency response curve into a pre-trained transformer winding fault diagnosis model and output the diagnosis result of the transformer winding to be tested. The transformer winding fault diagnosis model is a deep convolutional neural network containing a residual module; the pre-trained transformer winding fault diagnosis model is trained based on a training database containing normal frequency response curves, single-fault frequency response curves, and multi-fault superimposed frequency response curves. The training database of this device covers feature data of multiple fault superposition scenarios, enabling the model to fully learn and master the feature coupling rules under superposition faults. It can accurately distinguish the differences between normal state, single fault and multiple fault superposition state, effectively breaking through the limitations of existing methods in the field of superposition fault diagnosis. At the same time, with the addition of a deep convolutional neural network with residual modules, it can further explore the subtle features of superposition faults, ensuring that even under the complex working conditions of multiple fault superposition, it can achieve accurate identification and judgment of faults, help to detect the initial superposition faults in time and take intervention measures, avoid the fault from expanding and causing large-scale power outages, significantly improve the safety and stability of power grid operation, and reduce economic losses.

[0111] The transformer winding fault diagnosis device provided in the embodiments of the present invention is described below. The transformer winding fault diagnosis device described below and the transformer winding fault diagnosis method described above can be referred to in correspondence.

[0112] Please refer to Figure 12 , Figure 12 A schematic diagram of a transformer winding fault diagnosis device provided in an embodiment of the present invention may include:

[0113] Memory 10 is used to store computer programs;

[0114] The processor 20 is used to execute computer programs to implement the above-described transformer winding fault diagnosis method.

[0115] The memory 10, processor 20, and communication interface 31 all communicate with each other through the communication bus 32.

[0116] In this embodiment of the invention, the memory 10 is used to store one or more programs. The programs may include program code, which includes computer operation instructions. In this embodiment of the invention, the memory 10 may store programs for implementing the following functions:

[0117] Obtain the frequency response curve of the transformer winding under test;

[0118] The frequency response curve is input into the pre-trained transformer winding fault diagnosis model, and the diagnosis result of the transformer winding to be tested is output.

[0119] Among them, the transformer winding fault diagnosis model is a deep convolutional neural network containing residual modules; the pre-trained transformer winding fault diagnosis model is trained based on a training database containing normal frequency response curves, single fault frequency response curves, and multi-fault superimposed frequency response curves.

[0120] In one possible implementation, the memory 10 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; and the data storage area may store data created during use.

[0121] Furthermore, memory 10 may include read-only memory and random access memory, providing instructions and data to the processor. A portion of the memory may also include NVRAM. The memory stores operating systems and operating instructions, executable modules, or data structures, or subsets thereof, or extended sets thereof, wherein the operating instructions may include various operating instructions for implementing various operations. The operating system may include various system programs for implementing various basic tasks and handling hardware-based tasks.

[0122] Processor 20 can be a central processing unit (CPU), an application-specific integrated circuit, a digital signal processor, a field-programmable gate array, or other programmable logic device. Processor 20 can be a microprocessor or any conventional processor. Processor 20 can call programs stored in memory 10.

[0123] Communication interface 31 can be an interface for the communication module, used to connect with other devices or systems.

[0124] Of course, it should be noted that, Figure 12 The structure shown does not constitute a limitation on the transformer winding fault diagnosis device in the embodiments of the present invention. In practical applications, the transformer winding fault diagnosis device may include more than Figure 12 More or fewer components as shown, or combinations of certain components.

[0125] The readable storage medium provided in the embodiments of the present invention is described below. The readable storage medium described below can be referred to in correspondence with the transformer winding fault diagnosis method described above.

[0126] The present invention also provides a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described transformer winding fault diagnosis method.

[0127] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0128] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0129] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0130] Finally, it should be noted that in this document, relationships such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0131] The present invention provides a detailed description of a transformer winding fault diagnosis method, apparatus, device, and computer-readable storage medium. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for diagnosing transformer winding faults, characterized in that, include: Obtain the frequency response curve of the transformer winding under test; The frequency response curve is input into a pre-trained transformer winding fault diagnosis model, and the diagnosis result of the transformer winding to be tested is output. The transformer winding fault diagnosis model is a deep convolutional neural network containing residual modules; the pre-trained transformer winding fault diagnosis model is trained based on a training database containing normal frequency response curves, single fault frequency response curves, and multi-fault superimposed frequency response curves.

2. The transformer winding fault diagnosis method according to claim 1, characterized in that, The single-fault frequency response curves include frequency response curves for short-circuit faults between winding discs, frequency response curves for faults involving changes in the spacing between winding discs, frequency response curves for radial deformation faults, and frequency response curves for axial deformation faults. The multi-fault superimposed frequency response curves include frequency response curves for axial deformation faults superimposed on faults involving changes in the spacing between winding discs and frequency response curves for radial deformation faults superimposed on faults involving changes in the spacing between winding discs.

3. The transformer winding fault diagnosis method according to claim 2, characterized in that, The process of building the training database includes: When the terminals at both ends of the winding are short-circuited by a wire, the frequency response curve of the short-circuit fault between the winding discs is obtained. When simulating a fault in the spacing between the winding discs by connecting capacitors of different sizes in parallel, the frequency response curve of the fault in the spacing between the winding discs is obtained. Collect the frequency response curves of the radial deformation fault of the winding corresponding to the winding with different degrees of radial deformation; The axial deformation of the winding was simulated using an axially concave winding, and the frequency response curve of the axial deformation fault of the winding was obtained. When the axial concave winding is used to simulate the axial deformation of the winding, and capacitors of different sizes are connected in parallel to simulate the fault of the change in the spacing between the winding discs, the frequency response curve of the axial deformation fault superimposed on the fault of the change in the spacing between the winding discs is obtained. When the radial deformation of the windings is simulated by connecting capacitors of different sizes in parallel to simulate the fault of the spacing between the winding discs, the frequency response curve of the radial deformation fault of the windings superimposed on the fault of the spacing between the winding discs is obtained.

4. The transformer winding fault diagnosis method according to claim 1, characterized in that, Obtain the frequency response curve of the transformer winding under test, including: A voltage detection signal with preset parameters is generated by a nanosecond pulse generator, and the voltage detection signal is transmitted to the protection circuit via a coaxial cable; the output terminal of the protection circuit is connected to the input terminal of the bushing capacitive coupling sensor. The voltage detection signal, adjusted by the protection circuit, is coupled and injected into the first end of the winding of the transformer under test through the bushing capacitive coupling sensor. The adjusted voltage detection signal propagates along the axial direction of the winding of the transformer under test to the end of the winding. When the end of the transformer winding under test is directly grounded, the response current signal propagated by the transformer winding under test is measured by a broadband current sensor connected in series with the grounding circuit. When the end of the transformer winding to be tested is not directly grounded, the response voltage signal propagated by the transformer winding to be tested is measured by the bushing capacitance coupling sensor coupled between the end of the transformer winding to be tested and the ground. The response current signal and response voltage signal in the time domain are converted into frequency domain data by fast Fourier transform, and the amplitude corresponding to each frequency point is extracted to obtain amplitude-frequency characteristic data. The frequency response curve is plotted based on the amplitude-frequency characteristic data; the horizontal axis of the frequency response curve is frequency, and the vertical axis is amplitude.

5. The transformer winding fault diagnosis method according to claim 1, characterized in that, Before inputting the frequency response curve into the pre-trained transformer winding fault diagnosis model and outputting the diagnosis result of the transformer winding under test, the process also includes: The frequency response curves of each type in the training database are evenly divided into training set and test set; A transformer winding fault diagnosis model is constructed, which includes convolutional blocks and identity blocks. Both the convolutional blocks and the identity blocks are designed based on the residual learning mechanism and are configured with parallel backbone feature extraction paths and residual edge paths. The frequency response curves of each type in the training set are input into the transformer winding fault diagnosis model. The learnable parameters of the convolutional layer in the transformer winding fault diagnosis model are optimized by the backpropagation algorithm to complete the model training and obtain the trained transformer winding fault diagnosis model. The frequency response curves of each type in the test set are input into the trained transformer winding fault diagnosis model for testing. By comparing the diagnostic results output by the model with the actual fault types of the samples in the test set, the fault classification accuracy and generalization ability of the model are evaluated. When the model's fault classification accuracy and generalization ability reach the preset standard, the pre-trained transformer winding fault diagnosis model is obtained.

6. The transformer winding fault diagnosis method according to claim 5, characterized in that, The main feature extraction paths of the convolutional block and the identity block are sequentially configured with three sets of processing units; the first set of processing units includes a first two-dimensional convolutional layer, a batch normalization layer, and an activation function layer; the second set of processing units includes a second two-dimensional convolutional layer, a batch normalization layer, and an activation function layer; the third set of processing units includes a third two-dimensional convolutional layer and a batch normalization layer. The first two-dimensional convolutional layer is a 1x1 convolution used to reduce the dimension of the input feature map; the second two-dimensional convolutional layer is a 3x3 convolution used for feature extraction; and the third two-dimensional convolutional layer is a 1x1 convolution used to restore the number of channels in the feature map to the original channel dimension.

7. The transformer winding fault diagnosis method according to claim 5, characterized in that, The residual edge path of the convolutional block is configured with a two-dimensional convolutional layer and a batch normalization layer to match the output dimension of the backbone feature extraction path of the convolutional block; the residual edge path of the identity block directly performs identity mapping on the input.

8. A transformer winding fault diagnosis device, characterized in that, include: The acquisition module is used to acquire the frequency response curve of the transformer winding under test; The diagnostic module is used to input the frequency response curve into a pre-trained transformer winding fault diagnosis model and output the diagnostic results of the transformer winding to be tested. The transformer winding fault diagnosis model is a deep convolutional neural network containing residual modules; the pre-trained transformer winding fault diagnosis model is trained based on a training database containing normal frequency response curves, single fault frequency response curves, and multi-fault superimposed frequency response curves.

9. A transformer winding fault diagnosis device, characterized in that, include: Memory, used to store computer programs; A processor is configured to implement the transformer winding fault diagnosis method as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when loaded and executed by a processor, implement the transformer winding fault diagnosis method as described in any one of claims 1 to 7.

Images