Transformer winding deformation fault diagnosis method, system, equipment and medium
By integrating multi-dimensional features and optimizing model parameters, the problems of incomplete feature information and insufficient model optimization in transformer winding fault diagnosis are solved, achieving accurate diagnosis of winding status and improving the accuracy and robustness of diagnosis.
Patent Information
- Application Number
- CN202511666904.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-14
- Publication Date
- 2026-02-17
AI Technical Summary
Existing methods for diagnosing transformer winding faults suffer from incomplete feature information, insufficient model optimization, and low diagnostic accuracy. They are unable to fully reflect the winding deformation state and are susceptible to interference, leading to misjudgments.
By integrating vibration, strain, and axial force signals from transformer windings, multiple features are extracted through modal decomposition. Particle swarm optimization is then used to optimize the parameters of the support vector machine model, and a multi-feature sample dataset is constructed to train a machine learning diagnostic model.
It significantly improves the accuracy and reliability of transformer winding fault diagnosis, and can accurately distinguish between the normal state of the winding, radial deformation and axial deformation, thereby improving the accuracy and real-time performance of the diagnosis.
Smart Images
Figure CN121542872A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of transformer monitoring technology, and in particular to a method, system, equipment and medium for diagnosing transformer winding deformation faults. Background Technology
[0002] As a core piece of equipment in a power system, transformers are prone to mechanical deformation of their windings under short-circuit impacts, and the cumulative effect can lead to serious faults. Winding fault diagnosis is a key means to ensure the reliable operation of transformers. In recent years, the rapid development of intelligent monitoring technology has provided new approaches to condition monitoring and fault prevention of power equipment. In particular, the integration of Internet of Things (IoT) technology enables real-time and remote monitoring of transformer health, significantly improving the timeliness and accuracy of fault diagnosis. However, current fault diagnosis methods still face many challenges.
[0003] Currently, transformer winding fault diagnosis methods mainly include offline monitoring (such as frequency response analysis and short-circuit impedance methods) and online monitoring (such as vibration signal analysis and fiber optic sensing methods). Offline monitoring methods require power outages, affecting the normal operation of the power system, and cannot capture the dynamic changes of winding faults in real time. Online monitoring often relies on a single feature quantity, which is difficult to comprehensively characterize the winding fault state and is easily affected by environmental interference, leading to misjudgment or missed faults. Current machine learning-based diagnostic models mainly rely on single physical quantities or traditional model analysis, which have significant shortcomings: First, feature extraction is limited, often relying on single time-domain or frequency-domain features of vibration signals, making it difficult to comprehensively reflect the winding deformation state; second, model parameters are not optimized, and the diagnostic accuracy of traditional SVM models is limited due to unreasonable penalty factors and kernel parameter settings; third, fault classification accuracy is low, with insufficient ability to distinguish between radial and axial deformation, making it susceptible to interference and misjudgment. Summary of the Invention
[0004] Therefore, the technical problem to be solved by the present invention is that existing diagnostic methods have incomplete feature information, insufficient model optimization, and low diagnostic accuracy.
[0005] The above-mentioned technical problems are solved by the following technical solutions: A method for diagnosing transformer winding deformation faults includes acquiring vibration signals, strain signals, and axial force signals of the transformer windings; performing modal decomposition on the vibration signals to obtain multiple modal components; extracting the strain peak value of the strain signals, obtaining the steady-state rate of change of the axial force signals, and obtaining the energy proportion of each mode of the multiple modal components to construct a multi-feature sample dataset; using an optimization algorithm to optimize the model parameters of a machine learning diagnostic model to obtain the optimal parameter combination; constructing and training the machine learning diagnostic model based on the optimal parameter combination and the multi-feature sample dataset; and using the trained machine learning diagnostic model to achieve fault diagnosis of normal state, radial deformation, and axial deformation of the transformer windings.
[0006] In a preferred embodiment of the transformer winding deformation fault diagnosis method of the present invention: obtaining the modal energy ratio of each of the plurality of modal components includes: calculating the total signal energy of the vibration signal; calculating the modal energy of each of the plurality of modal components respectively; calculating the ratio of the modal energy to the total signal energy to obtain the modal energy ratio.
[0007] In a preferred embodiment of the transformer winding deformation fault diagnosis method of the present invention: the extraction of the strain peak value of the strain signal and the steady-state rate of change of the axial force signal includes: determining the maximum instantaneous value of the strain signal during the short-circuit impact as the strain peak value; obtaining the initial axial force before the short-circuit impact and the steady-state axial force after the short-circuit impact; and calculating the rate of change of the steady-state axial force relative to the initial axial force as the steady-state rate of change.
[0008] In a preferred embodiment of the transformer winding deformation fault diagnosis method of the present invention: the step of using an optimization algorithm to optimize the model parameters of the machine learning diagnostic model to obtain the optimal parameter combination includes: setting the parameter range to be optimized for the machine learning diagnostic model; setting the iteration termination criterion of the optimization algorithm; iteratively searching within the parameter range using the optimization algorithm and evaluating the fitness of the search results; and determining the optimal parameter combination based on the search results with the best fitness.
[0009] In a preferred embodiment of the transformer winding deformation fault diagnosis method of the present invention: the optimization algorithm is a particle swarm optimization algorithm, and the optimization algorithm includes: initializing a particle swarm, with each particle corresponding to a set of model parameters; determining the global optimal position and the local optimal position according to the fitness of the model parameters; iteratively updating the particle velocity and position until the iteration termination criterion is met, and determining the optimal parameter combination with the global optimal position.
[0010] In a preferred embodiment of the transformer winding deformation fault diagnosis method of the present invention: the machine learning diagnostic model is a support vector machine model, and the construction and training of the machine learning diagnostic model based on the optimal parameter combination and the multi-feature sample dataset includes: selecting a kernel function to map the multi-feature sample dataset to a high-dimensional space; dividing the multi-feature sample dataset into a training set and a test set; configuring the support vector machine model using the optimal parameter combination, and training the machine learning diagnostic model using the training set.
[0011] In a preferred embodiment of the transformer winding deformation fault diagnosis method of the present invention: the construction of the multi-feature sample dataset further includes: acquiring sample data of the corresponding winding normal state, radial deformation and axial deformation; setting corresponding fault codes for the normal state, radial deformation and axial deformation; and using the fault codes to mark the sample dataset.
[0012] A transformer winding deformation fault diagnosis system includes a signal acquisition module, a vibration decomposition module, a feature extraction module, a sample construction module, a model optimization module, a model training module, and a fault diagnosis module.
[0013] In a preferred embodiment of the transformer winding deformation fault diagnosis system of the present invention: a signal acquisition module is used to acquire vibration signals, strain signals, and axial force signals of the transformer winding; a vibration decomposition module is used to perform modal decomposition processing on the vibration signals to obtain multiple modal components; a feature extraction module is used to extract the strain peak value of the strain signal, obtain the steady-state rate of change of the axial force signal, and obtain the energy proportion of each mode of the multiple modal components; a sample construction module is used to construct a multi-feature sample dataset based on the strain peak value, the steady-state rate of change, and the energy proportion of each mode; a model optimization module is used to optimize the model parameters of the machine learning diagnostic model using an optimization algorithm to obtain the optimal parameter combination; a model training module is used to construct and train the machine learning diagnostic model based on the optimal parameter combination and the multi-feature sample dataset; and a fault diagnosis module is used to use the trained machine learning diagnostic model to realize fault diagnosis of the normal state, radial deformation, and axial deformation of the transformer winding.
[0014] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the transformer winding deformation fault diagnosis method described above.
[0015] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a transformer winding deformation fault diagnosis method as described above.
[0016] The beneficial effects of this invention are as follows: by fusing strain signals, axial force signals, and vibration signals, and extracting their respective core features, a multi-feature quantity is constructed, which more comprehensively reflects the deformation state of the winding; by performing modal decomposition processing on the vibration signal and extracting its energy proportion, the identification of fault features is effectively enhanced; by using optimization algorithms to optimize the parameters of the machine learning diagnostic model, the final classification performance of the model is improved; this method enables the method to accurately distinguish the normal state, radial deformation, and axial deformation of the winding, significantly improving the accuracy and reliability of fault diagnosis. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Obviously, the drawings described below only relate to some embodiments of the present invention and are not intended to limit the present invention.
[0018] Figure 1 The waveform of positive impact strain is shown; Figure 2 The negative impact strain waveform is shown; Figure 3 The evolution process of the center frequencies of each mode is shown; Figure 4 A flowchart of a method for diagnosing transformer winding deformation faults is shown. Figure 5 The energy distribution of each mode along the x-axis after VMD decomposition is shown. Figure 6 The energy distribution of each mode along the y-axis after VMD decomposition is shown. Figure 7 The energy distribution of each mode along the z-axis after VMD decomposition is shown. Figure 8 The PSO optimization process is shown. Figure 9 The specific diagnostic results of the PSO-SVM model are shown. Detailed Implementation
[0019] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.
[0020] The terminology used in this invention is that which is currently widely used in the art in consideration of the function of the invention; however, these terms may vary according to the intent of those skilled in the art, precedent, or new technology in the art. Furthermore, specific terms may be chosen by the applicant, and in such cases, their detailed meanings will be described in the detailed description of the invention. Therefore, the terms used in this specification should not be construed as simple names, but rather based on their meanings and the overall description of the invention.
[0021] Example 1, referring to Figures 1 to 9 This embodiment provides a method for diagnosing transformer winding deformation faults, including: The vibration, strain, and axial force signals of the transformer windings are acquired. The strain, axial force, and vibration signals of the transformer windings under short-circuit impact are collected by a sensor array to establish a raw sample dataset for diagnostic analysis.
[0022] Modal decomposition is used to process vibration signals to obtain multiple modal components; different frequency components in the vibration signal are separated to improve the sensitivity of the feature quantity to the fault, while removing the interference of high-frequency noise and useful signal to enhance feature identification.
[0023] The strain peak value of the strain signal is extracted, the steady-state change rate of the axial force signal is obtained, and the modal energy ratio of multiple modal components is obtained to construct a multi-feature sample dataset; the multi-dimensional features of strain, axial force, and vibration are integrated to comprehensively reflect the winding deformation state and solve the problem of one-sided information from a single feature.
[0024] An optimization algorithm is used to find the optimal model parameters of the machine learning diagnostic model and obtain the optimal parameter combination. The particle swarm optimization (PSO) algorithm is used to optimize the penalty factor C and kernel parameter g of the support vector machine (SVM) model, so that the model classification performance is optimal, the diagnostic accuracy is improved, and the problems of unreasonable parameter settings and limited accuracy of traditional models are solved.
[0025] Based on the optimal parameter combination and multi-feature sample dataset, a machine learning diagnostic model is constructed and trained. According to the optimal parameter combination, including but not limited to the best kernel function type, C, and g values, the final PSO-SVM training model is constructed and trained by dividing the training set in a 7:3 ratio, thereby obtaining a fault diagnosis model with optimal classification performance that can adapt to complex working conditions. Figure 9 This provides the specific diagnostic results of the PSO-SVM model applied to a portion of the test set data. By utilizing a trained machine learning diagnostic model, fault diagnosis of normal winding conditions, radial deformation, and axial deformation of transformer windings can be achieved. After inputting the test set into the trained PSO-SVM model and verifying its accuracy, the transformer winding condition can be identified, enabling precise diagnosis of normal winding conditions, radial deformation, and axial deformation, thus improving diagnostic accuracy and real-time performance.
[0026] It should be noted that the shortcomings of the existing technical solutions are: transformer winding fault diagnosis mostly relies on a single feature or traditional machine learning model, which has problems such as low diagnostic accuracy, insufficient feature discrimination, insufficient model parameter optimization, difficulty in fully reflecting the winding fault state, and insufficient ability to distinguish between radial and axial deformation.
[0027] Therefore, to address the above issues, this method integrates multi-dimensional features and optimizes model parameters, significantly improving the accuracy and robustness of winding fault diagnosis. The average accuracy rate can reach 97.73%, enabling accurate identification of normal winding conditions, radial deformation, and axial deformation, and monitoring and diagnosis can be completed without power outages.
[0028] The specific comparison is shown in the table below: Table 1: Comparison of Performance Indicators of Different Diagnostic Models
[0029] Example 2, refer to Figures 1 to 9 This embodiment, based on the previous embodiment, provides a method for diagnosing transformer winding deformation faults, including: Specifically, the vibration signal, strain signal, and axial force signal of the transformer winding are acquired; the vibration signal is processed by modal decomposition to obtain multiple modal components; the strain peak value of the strain signal, the steady-state rate of change of the axial force signal, and the energy proportion of each mode of multiple modal components are extracted to construct a multi-feature sample dataset; the model parameters of the machine learning diagnostic model are optimized using an optimization algorithm to obtain the optimal parameter combination; based on the optimal parameter combination and the multi-feature sample dataset, the machine learning diagnostic model is constructed and trained; and the trained machine learning diagnostic model is used to realize fault diagnosis of the normal state, radial deformation, and axial deformation of the transformer winding.
[0030] The strain, axial force, and vibration signals of the transformer winding under short-circuit impact are collected by fiber Bragg grating sensors. Variational mode decomposition (VMD) is applied to the vibration signals for preprocessing. The peak strain, steady-state rate of change of axial force, and energy proportion of each VMD mode are extracted as multi-dimensional features. The parameters of the support vector machine (SVM) model are optimized using the particle swarm optimization (PSO) algorithm. The PSO-SVM model is constructed and trained using the optimal parameters to achieve accurate diagnosis of the normal state, radial deformation, and axial deformation of the winding, thereby improving the accuracy and reliability of diagnosis and providing precise decision support for the operation and maintenance of power equipment.
[0031] Detailed, such as Figure 3 The evolution of the center frequencies of each mode is plotted using a logarithmic horizontal axis to visually understand the convergence speed and stability of each mode during iteration. This also allows for a rough assessment of the separation between modes, leading to the selection of K=6 decomposition modes. The non-overlapping center frequency evolution is verified, along with the balance parameters. =2000, to avoid modal aliasing, the frequency initialization method is set to uniform distribution, and the modes are initialized. .
[0032] Calculate the vibration signal u using Hilbert transform. k The analytic signal Au corresponding to (t) k(t), and thus the formula for the one-sided spectrum is obtained: Then the complex exponent term With analytic signal Au k (t) Mix, and shift the spectrum of the mixed signal to the corresponding estimated center frequency. Then, the following constrained variational problem is established, with the norm squared L2 of the gradient as the optimization objective: In the formula, x(t) represents the original data; Represented as a Dirac distribution; Represents the convolution operator; This is the sum of all modes.
[0033] Then combine the quadratic penalty term and the Lagrange multiplier. A new constrained variational problem was constructed as follows: In the formula, These are the coefficients of the Lagrange multipliers, which can be set to 0 without strictly enforcing the constraints.
[0034] Then, the Alternating Direction Multiplier Method (ADMM) optimization algorithm is used to solve the problem and obtain the decomposition mode. and its center frequency The final expression is: Also, such as Figure 1 and Figure 2 The definition of peak strain is similar to that of peak short-circuit current. Peak short-circuit strain is defined as the maximum instantaneous value of the strain waveform measured during a short circuit, i.e., the highest point reached during the short-circuit duration. It may be the maximum value in the positive direction or the maximum value in the negative direction. If the peak short-circuit strain is the maximum value in the positive direction, it is called a positive impact; otherwise, it is a negative impact. Furthermore, the average strain value approximately 200 ms before the short-circuit impact is taken as the initial strain.
[0035] The steady-state rate of change is defined similarly to the short-circuit peak current. The transient peak value of the axial force is defined as the maximum instantaneous value of the axial force waveform measured during a short circuit, i.e., the highest point reached during the short-circuit duration. This value may be the maximum value in the positive or negative direction. Additionally, the average axial force value approximately 200 ms before the first short-circuit impact is taken as the initial axial force F0, and the average axial force value under each rated current steady-state test is taken as the steady-state axial force F. W The steady-state axial force F WThe steady-state rate of change is calculated as the ratio of the difference between the initial axial force F0 and the initial axial force F0. : In the formula, F W The average value of the periodic axial force was measured after each short-circuit impact on the test winding, with the rated current applied to the winding. Steady-state rate of change. It can reflect the accumulation of axial force on the winding, which is helpful for further analysis of the axial force accumulation effect of the winding.
[0036] Specifically, obtaining the modal energy percentage of each modal component includes: calculating the total signal energy of the vibration signal; calculating the modal energy of each modal component; and calculating the ratio of modal energy to the total signal energy to obtain the modal energy percentage.
[0037] Assuming the discrete sampled signal in the time domain of the vibration signal is a(n), and n=1,2,...,N, then the total energy of the sampled signal a(n) is: This method decomposes the vibration signal into 6 modes using VMD, thus allowing the determination of the energy percentage (ER) of each mode. k for: Among them, ER k E represents the modal energy ratio. k Let be the total energy of a certain mode, where k = 1, 2, 3, ..., 6.
[0038] like Figures 5 to 7 To maximize the utilization of vibration characteristics, a weighted average of the energy proportion data in each of the three directions was performed, with equal weights for each direction, ultimately yielding energy proportion values for six modes. This quantifies the vibration energy distribution across different modes, assessing the mechanical state of the winding and obtaining characteristic quantities with higher fault sensitivity.
[0039] Specifically, extracting the peak strain of the strain signal and the steady-state rate of change of the axial force signal includes: determining the maximum instantaneous value of the strain signal during the short-circuit impact as the peak strain; obtaining the initial axial force before the short-circuit impact and the steady-state axial force after the short-circuit impact; and calculating the rate of change of the steady-state axial force relative to the initial axial force as the steady-state rate of change. The peak strain reflects the maximum deformation of the winding under short-circuit impact; the steady-state rate of change reflects the accumulation of axial force on the winding, which is helpful for analyzing the cumulative effect of axial force on the winding.
[0040] The peak short-circuit strain is defined as the maximum instantaneous value of the strain waveform measured during a short circuit. The average strain value approximately 200 ms before the short-circuit impact is taken as the initial strain. The average axial force value approximately 200 ms before the first short-circuit impact is taken as the initial axial force. The average axial force value under each rated current steady-state test is taken as the steady-state axial force. The difference between the steady-state axial force and the initial axial force is calculated, and the ratio of this difference to the initial axial force is taken as the steady-state rate of change. The peak strain value reflects the maximum deformation of the winding under short-circuit impact; the steady-state rate of change reflects the accumulation of axial force on the winding, which is beneficial for analyzing the cumulative effect of axial force on the winding.
[0041] Specifically, the optimization algorithm is used to find the optimal parameter combination of the machine learning diagnostic model. This includes: setting the parameter range to be optimized for the machine learning diagnostic model; setting the iteration termination criterion of the optimization algorithm; iteratively searching within the parameter range using the optimization algorithm and evaluating the fitness of the search results; and determining the optimal parameter combination based on the search results with the best fitness.
[0042] like Figure 8 The penalty factor C and kernel parameter g of the SVM model were set as the parameters to be optimized. The search range of C was set to (0.1-100), and the range of g was set to (0.001-10). The iteration termination criterion of the PSO algorithm was reaching the set maximum number of iterations. The optimal parameter combination obtained after optimization was C=92.06822, g=7.92556520. This overcomes the problem of limited diagnostic accuracy caused by unreasonable parameter settings in traditional SVM models, and optimizes the model performance.
[0043] Specifically, the optimization algorithm is a particle swarm optimization algorithm. The optimization algorithm includes: initializing the particle population, with each particle corresponding to a set of model parameters; determining the global optimum and local optimum based on the fitness of the model parameters; iteratively updating the particle velocity and position until the iteration termination criterion is met, and determining the optimal parameter combination based on the global optimum position.
[0044] The population size for PSO is set to 50, the number of iterations to 100, and the inertia factor to 0.7. An initial particle population is established with corresponding velocities set. The fitness of each particle is evaluated, and the globally optimal PSO is determined. gbest and local optimum P lbest In each iteration, the particle velocity is updated according to the velocity update formula.
[0045] Speed update formula: In the formula, and They represent the first The position and velocity of the j-th particle in the next iteration; To represent the inertia weighting coefficient; The number of iterations is represented by r1 and r2, which represent random numbers between [0,1].
[0046] Then adjust the particle position according to the position update formula.
[0047] Position update formula: In this way, the global search capability of the PSO algorithm can be utilized to efficiently and quickly find the parameter combination that optimizes the performance of the SVM model within the parameter space.
[0048] Specifically, the machine learning diagnostic model is a support vector machine model. Based on the optimal parameter combination and a multi-feature sample dataset, the construction and training of the machine learning diagnostic model includes: selecting a kernel function to map the multi-feature sample dataset to a high-dimensional space; dividing the multi-feature sample dataset into a training set and a test set; configuring the support vector machine model with the optimal parameter combination; and using the training set to train the machine learning diagnostic model.
[0049] The Radial Basis Function (RBF) kernel is used to map features to a high-dimensional space. The sample dataset is divided into training and test sets in a 7:3 ratio. The optimal parameters C and g are used to construct the final SVM training model. The optimal hyperplane is determined by solving a convex quadratic optimization problem, and the decision function is finally obtained. Leveraging the advantages of SVM in small sample sizes and nonlinear problems, the kernel function is used to handle nonlinear classification problems, and a diagnostic model with strong classification ability and robustness is obtained using the training set.
[0050] Specifically, constructing a multi-feature sample dataset also includes: obtaining sample data of the normal state, radial deformation, and axial deformation of the corresponding winding; setting corresponding fault codes for the normal state, radial deformation, and axial deformation; and labeling the sample dataset using the fault codes.
[0051] The selected sample data must include samples of normal state, radial deformation fault, and axial deformation fault, with each type of sample having a roughly equal proportion. The fault codes are 0 for normal state, 1 for radial deformation, and 2 for axial deformation. This establishes a balanced and operable sample set that includes all target classification states, providing the foundational data for subsequent model training and validation.
[0052] Example 3, referring to Figure 4 This embodiment provides a transformer winding deformation fault diagnosis system, including: The signal acquisition module is used to acquire vibration, strain, and axial force signals from the transformer windings. This module includes a fiber optic sensor array consisting of strain sensors, axial force sensors, and vibration sensors. The sensors are responsible for real-time acquisition of multi-dimensional physical quantities generated by the transformer windings under short-circuit impacts, including strain, axial force, and vibration signals, providing the raw data foundation for subsequent diagnostics through multi-feature fusion.
[0053] The vibration decomposition module is used to process vibration signals by mode decomposition to obtain multiple modal components. Variational Mode Decomposition (VMD) preprocessing is applied to the acquired vibration signals. By setting the number of decomposition modes and balancing parameters, high-frequency noise and components of different frequencies in the vibration signal are effectively separated. This improves feature recognition and prepares for extracting the energy proportion of each mode.
[0054] The feature extraction module is used to extract the strain peak value of the strain signal, obtain the steady-state rate of change of the axial force signal, and obtain the modal energy proportion of multiple modal components; it performs feature quantity calculations: extracting the short-circuit strain peak value from the strain signal; calculating the steady-state rate of change from the axial force signal; and calculating the modal energy proportion of each mode from the modal components after VMD decomposition. The original signal is transformed into an eight-dimensional feature vector containing the strain peak value, axial force rate of change, and energy proportion of six modes.
[0055] The sample construction module is used to construct a multi-feature sample dataset based on peak strain, steady-state rate of change, and energy proportion of each mode. It standardizes the extracted multi-dimensional feature data and maps it to the [0,1] interval. Simultaneously, it is responsible for selecting and encoding sample data, including normal state 0, radial deformation 1, and axial deformation 2, and associating the multi-dimensional features with the corresponding fault labels to form a sample dataset that can be used for model training.
[0056] The model optimization module is used to optimize the model parameters of the machine learning diagnostic model using optimization algorithms to obtain the optimal parameter combination. This module uses the Particle Swarm Optimization (PSO) algorithm as the optimization algorithm. Based on set parameters such as population size, number of iterations, and inertia factor, it searches for the optimal combination of the penalty factor C and kernel parameter g of the Support Vector Machine (SVM) model. By iteratively updating the velocity and position of the particles, it ensures that the parameters that best enable SVM classification performance are found.
[0057] The model training module is used to build and train a machine learning diagnostic model based on the optimal parameter combination and a multi-feature sample dataset. This module uses SVM as the diagnostic model and employs the RBF kernel function to map features to a high-dimensional space. It divides the sample dataset into training and test sets according to a preset ratio. Using the training set and the optimal parameters determined by the model optimization module, the final PSO-SVM diagnostic model is trained.
[0058] The fault diagnosis module utilizes a trained machine learning diagnostic model to diagnose faults in the normal state, radial deformation, and axial deformation of transformer windings. This module inputs the data to be diagnosed into the trained PSO-SVM model. The model's decision function enables accurate identification and classification of winding fault states.
[0059] The overall accuracy of this system can reach 97.73%. This system can be used for online fault diagnosis of transformers in operation, ensuring the safe and stable operation of the power system.
[0060] Finally, it should be noted that the methods and devices described in detail above are merely embodiments, and those skilled in the art can modify these embodiments in different ways as long as they do not depart from the scope of the present invention.
Claims
1. A method for diagnosing transformer winding deformation faults, characterized in that, include: Acquire vibration signals, strain signals, and axial force signals of the transformer windings; The vibration signal is processed by modal decomposition to obtain multiple modal components; Extract the strain peak value of the strain signal, obtain the steady-state change rate of the axial force signal, and obtain the modal energy ratio of each of the multiple modal components to construct a multi-feature sample dataset; An optimization algorithm is used to optimize the model parameters of the machine learning diagnostic model and obtain the optimal parameter combination. Based on the optimal parameter combination and the multi-feature sample dataset, the machine learning diagnostic model is constructed and trained. The trained machine learning diagnostic model is used to diagnose faults in the normal state, radial deformation, and axial deformation of transformer windings.
2. The transformer winding deformation fault diagnosis method as described in claim 1, characterized in that: The process of obtaining the modal energy percentage of each of the multiple modal components includes: Calculate the total energy of the vibration signal; Calculate the modal energy of each of the multiple modal components; Calculate the ratio of the modal energy to the total signal energy to obtain the energy percentage of each modality.
3. The transformer winding deformation fault diagnosis method as described in claim 1, characterized in that: The extraction of the strain peak value of the strain signal and the steady-state rate of change of the axial force signal includes: The maximum instantaneous value of the strain signal during the short-circuit impact is determined as the strain peak value; Obtain the initial axial force before the short-circuit impact and the steady-state axial force after the short-circuit impact; The rate of change of the steady-state axial force relative to the initial axial force is calculated and used as the steady-state rate of change.
4. The transformer winding deformation fault diagnosis method as described in claim 1, characterized in that: The process of using optimization algorithms to optimize the model parameters of the machine learning diagnostic model to obtain the optimal parameter combination includes: Define the parameter range to be optimized for the machine learning diagnostic model; Define the iteration termination criterion for the optimization algorithm; Within the parameter range, the fitness of the search results is evaluated through iterative searching using the optimization algorithm. The optimal parameter combination is determined based on the search result with the best fitness.
5. The transformer winding deformation fault diagnosis method as described in claim 1 or 4, characterized in that: The optimization algorithm is a particle swarm optimization algorithm, and the optimization search using the optimization algorithm includes: Initialize the particle population, with each particle corresponding to a set of model parameters; Based on the fitness of the model parameters, determine the global optimum and the local optimum. The particle velocity and position are iteratively updated until the iteration termination criterion is met, and the optimal parameter combination is determined based on the global optimal position.
6. The transformer winding deformation fault diagnosis method as described in claim 5, characterized in that: The machine learning diagnostic model is a support vector machine model. The process of constructing and training the machine learning diagnostic model based on the optimal parameter combination and the multi-feature sample dataset includes: A kernel function is selected to map the multi-feature sample dataset to a high-dimensional space; The multi-feature sample dataset is divided into a training set and a test set; The support vector machine model is configured using the optimal parameter combination, and the machine learning diagnostic model is trained using the training set.
7. The transformer winding deformation fault diagnosis method as described in claim 1, characterized in that: The construction of the multi-feature sample dataset also includes: Obtain sample data of the normal state, radial deformation, and axial deformation of the corresponding winding; Set corresponding fault codes for the normal state, radial deformation, and axial deformation; The sample dataset is labeled using the fault codes.
8. A transformer winding deformation fault diagnosis system, employing the transformer winding deformation fault diagnosis method as described in any one of claims 1 to 7, characterized in that, include: The signal acquisition module is used to acquire vibration signals, strain signals, and axial force signals of the transformer windings. The vibration decomposition module is used to perform modal decomposition processing on the vibration signal to obtain multiple modal components; The feature extraction module is used to extract the strain peak value of the strain signal, obtain the steady-state change rate of the axial force signal, and obtain the energy ratio of each mode of the multiple modal components; The sample construction module is used to construct a multi-feature sample dataset based on the strain peak value, the steady-state change rate, and the energy proportion of each mode. The model optimization module is used to optimize the model parameters of the machine learning diagnostic model using optimization algorithms to obtain the optimal parameter combination; The model training module is used to construct and train the machine learning diagnostic model based on the optimal parameter combination and the multi-feature sample dataset. The fault diagnosis module is used to diagnose faults in the normal state, radial deformation, and axial deformation of transformer windings by utilizing the trained machine learning diagnostic model.
9. An electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the transformer winding deformation fault diagnosis method according to any one of claims 1 to 7.
10. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the transformer winding deformation fault diagnosis method according to any one of claims 1 to 7.