Winding deformation diagnosis method, system and equipment based on vector matching and machine learning and medium

Abstract

The invention discloses a winding deformation diagnosis method, system and equipment based on vector matching and machine learning, and a medium, and relates to the technical field of electrical equipment state monitoring and fault diagnosis, and the method comprises the steps: constructing a curve data set of various fault states through frequency response analysis, and carrying out the rational function fitting of a curve through a vector matching technology, the method comprises the following steps: extracting zero-pole feature vectors through iterative optimization, constructing a support vector machine model, automatically optimizing parameters by using a particle swarm algorithm, obtaining a high-precision diagnosis model in combination with cross validation training, and realizing automatic identification and visual display of fault types and degrees through a human-computer interaction interface. According to the method, the machine learning model optimized through the swarm intelligence algorithm is combined, accurate and automatic diagnosis of the transformer winding deformation fault type and degree is achieved, and the comprehensiveness, accuracy and engineering practicability of diagnosis are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrical equipment state monitoring and fault diagnosis, in particular to a winding deformation diagnosis method, system, device and medium based on vector matching and machine learning. BACKGROUND

[0002] Transformer winding deformation diagnosis is one of the core technologies in the field of power equipment state monitoring and fault warning. The frequency response analysis method is currently widely used as the main technical means in the field to reflect mechanical structure changes by scanning the transfer function characteristics of the winding at different frequencies. With the progress of digital signal processing technology, traditional FRA diagnosis has experienced a development process from graphical comparison to numerical analysis: in the early stage, it mainly relied on expert experience to subjectively compare the response curve shape; in the later stage, statistical indicators such as correlation coefficient and Euclidean distance were developed to realize quantitative evaluation; in recent years, with the penetration of artificial intelligence technology, intelligent diagnosis methods based on neural networks and support vector machines have emerged. However, the existing technical solutions are mostly focused on direct analysis of the amplitude-frequency response curve, or extraction of shallow indicators such as resonance point shift and amplitude change through limited feature engineering, forming a technology system from pure manual interpretation to semi-automatic diagnosis.

[0003] However, the existing winding deformation diagnosis technology still has several essential defects: first, in terms of feature representation, the amplitude features or simple numerical indicators relied on by the mainstream methods can only reflect part of the information of the system response, completely ignoring the system dynamic behavior contained in the phase characteristics, resulting in incomplete feature representation; the feature extraction method based on artificial experience is difficult to capture the deep correlation between the winding physical structure and the electromagnetic response, and the traditional statistical indicators are not sensitive enough to distinguish various fault modes; secondly, in terms of diagnosis model construction, the existing intelligent diagnosis methods generally face the problem of relying on expert prior knowledge for model parameter optimization, and the parameter adjustment process lacks self-adaptive ability, affecting the model generalization performance; at the same time, traditional feature extraction and classification models are independent of each other, making it difficult to form an end-to-end optimization system; in addition, the existing technical solutions have limited ability to identify slight winding deformation and composite faults, and the reliability of the diagnosis results is severely dependent on the completeness of the training data. SUMMARY

[0004] In view of the above existing problems, the present application provides a winding deformation diagnosis method, system, device and medium based on vector matching and machine learning, to solve the problems of incomplete feature representation, lack of self-adaptive ability in parameter adjustment process, and difficulty in forming an end-to-end optimization system in the prior art.

[0005] To solve the above technical problems, a winding deformation diagnosis method based on vector matching and machine learning is proposed, which includes, Simulate fault types and different degrees on transformer windings, obtain frequency response data in different states through frequency response analysis technology, and construct a frequency response curve dataset; perform curve fitting processing on the frequency response curve based on a rational function model, determine the optimal fitting order through an iterative optimization process, and extract characteristic parameters representing the winding intrinsic characteristics from the fitted transfer function to form a feature vector; construct a machine learning classification model, and automatically search and optimize key parameters using a swarm intelligence optimization algorithm; select the optimal parameter combination through cross-validation performance evaluation, and train an optimized diagnostic model; input the feature vector into the optimized diagnostic model, use the model to calculate the automatic recognition and classification results of the transformer winding fault type and degree, and display the diagnostic results in a visual form through a human-computer interaction interface.

[0006] As a preferred scheme of the winding deformation diagnosis method based on vector matching and machine learning, the frequency response curve dataset is constructed by setting four fault types and different fault degrees for the transformer winding wound independently, performing frequency response analysis using a frequency response analyzer, obtaining the amplitude and phase angle of the transfer function at different sweep points, displaying the amplitude and phase relationship of the output relative to the input at different frequencies, and calculating the linear amplitude and linear vector of the response.

[0007] As a preferred scheme of the winding deformation diagnosis method based on vector matching and machine learning, the feature vector is formed by performing curve fitting processing on the frequency response curve based on a rational function model to obtain a transfer function representing the dynamic characteristics of the system. The fitting order is adjusted using an iterative optimization process, the fitting quality is evaluated through residual analysis, the system characteristic parameters are extracted from the fitted transfer function, and the feature vector is constructed.

[0008] As a preferred scheme of the winding deformation diagnosis method based on vector matching and machine learning, the training of the optimized diagnostic model includes constructing a machine learning classification model suitable for multi-class fault classification and processing high-dimensional feature input. The model key parameters are searched using a swarm intelligence optimization algorithm, the parameter performance is evaluated using a fitness function, and the parameter combination is iteratively updated. The model performance is evaluated using cross-validation, the optimal parameter combination is selected through multiple training and validation cycles, and the final diagnostic model is trained based on the selected optimal parameter combination and the model parameters are stored.

[0009] ​As a preferred scheme of the winding deformation diagnosis method based on vector matching and machine learning, the rational function model comprises a proportional polynomial model selected as a fitting model, and a transfer function is a quotient of two rational coefficient polynomials. The proportional polynomial model is expressed as: wherein, is a fitted transfer function, is a complex frequency variable, is an order of the model, is a set vector of all polynomial coefficients to be solved, is a coefficient of a numerator polynomial, is a coefficient of a denominator polynomial; initializing zero-pole positions as an iteration starting point, performing parameter estimation through a numerical optimization algorithm, and using a pole relocation iteration technique to convert a nonlinear least square problem into a linear least square problem for solving; The least square solution is expressed as: wherein, is a parameter vector to be solved, i.e., all unknowns to be solved in a current iteration step, is an optimal parameter vector, is a remainder associated with a new pole, is a constant term of a transfer function, is a proportional term of the transfer function, is an observation vector, i.e., a complex frequency response value actually measured at each sampling frequency point, is a conjugate transpose of the matrix A, is an inverse matrix of the matrix . The pole updating formula is expressed as: wherein, is an auxiliary rational function, is an nth remainder obtained from a current solution, is an old pole of a last iteration; The fitting quality is evaluated through residual analysis, a complex residual at each frequency point is calculated, when the residual curve randomly fluctuates around a zero reference line without a random pattern, it is determined that the fitting is effective, and the zero point and pole values are extracted from the converged transfer function to form a feature vector; The complex residual calculation formula is expressed as: wherein, is a frequency point index, is an original complex frequency response value obtained by actual measurement at the kth frequency point, is a predicted frequency response value calculated by using the fitted transfer function model at the kth frequency point, is a complex residual at the kth frequency point.

[0010] As a preferred scheme of the winding deformation diagnosis method based on vector matching and machine learning according to the present application, wherein: the automatic search and optimization of key parameters include initializing a particle swarm, each particle representing a parameter combination of a support vector machine model, the parameters including a penalty coefficient and a kernel function parameter; For each particle, use the leave-one-out method for model training and verification, leave each sample in the data set as a validation set, and use the remaining samples to train the model, and calculate the average classification accuracy as the fitness value; The formula for calculating the average classification accuracy is: wherein, is the accuracy, is the true positive, is the true negative, is the false positive, is the false negative; Update the individual historical best position of each particle and the global best position of the entire population, find the optimal parameter combination through the iterative optimization process, and output the parameter combination corresponding to the global best position; The cross-validation includes solving a convex quadratic programming problem, obtaining a Lagrange multiplier, and determining the support vector and bias term based on the multiplier, and using the decision function to classify the input feature vector, calculating the probability of the current sample belonging to each class, and selecting the class with the highest probability as the final prediction result; The formula for solving the convex quadratic programming problem is: wherein, , and is the Lagrange multiplier, is the number of samples in the current training set, is the feature vector of the i th training sample, is the feature vector of the j th training sample, is the class label of the i th training sample, is the class label of the j th training sample, is the transpose operation, is the constraint condition, and is an index of a training sample.

[0011] As a preferred scheme of the winding deformation diagnosis method based on vector matching and machine learning, the visualization form of the diagnosis result comprises post-processing of the classification result, generating a probability distribution and a confidence score, and providing a graphical operation interface, so that a user can import new test data and a feature vector, call a diagnosis model trained in the background for calculation, output a diagnosis conclusion of a fault type in a text form, and display the diagnosis result through a visual chart, and provide a comparison analysis function with a historical health state curve.

[0012] The preferred technical scheme has the beneficial effects that: a support vector machine classification model is constructed, a particle swarm algorithm is used to automatically search for an optimal parameter combination, and a leave-one-out cross-validation is used for performance evaluation, so that automatic optimization of the diagnosis model is realized, the swarm intelligence optimization algorithm avoids subjectivity and inefficiency of traditional manual parameter adjustment, the model performance is significantly improved through an adaptive search mechanism, the cross-validation ensures unbiasedness of model evaluation, and the obtained diagnosis model has better generalization ability and stability and is suitable for complex and variable working conditions.

[0013] As a preferred scheme of the winding deformation diagnosis system based on vector matching and machine learning, the system comprises a data acquisition module, a data processing and feature extraction module, a model training and optimization module, and a fault diagnosis and human-computer interaction module.

[0014] The data acquisition module is used to perform frequency sweep testing on a transformer winding model through a frequency response analyzer, and acquire original frequency response data.

[0015] The data processing and feature extraction module is used to execute a vector matching algorithm, fit a frequency domain curve into a rational transfer function, automatically determine an optimal fitting order through iterative optimization, extract zero points and pole points representing essential features of a winding physical structure, and form a zero-pole feature vector.

[0016] The model training and optimization module is used to store training data, construct a support vector machine classification model, automatically search for key optimal parameters of a current model through a particle swarm optimization algorithm, and evaluate model performance through a cross-validation method.

[0017] The fault diagnosis and human-computer interaction module is used to receive new test data or a feature vector, call a final diagnosis model for real-time analysis, output a diagnosis conclusion of a fault type and degree, and display the diagnosis result in a visual form through a human-computer interaction interface.

[0018] A computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of the method for winding deformation diagnosis based on vector matching and machine learning when executing the computer program.

[0019] A computer readable storage medium stores a computer program, and the computer program implements the steps of the method for winding deformation diagnosis based on vector matching and machine learning when executed by a processor.

[0020] The present application has the following advantages: the present application provides a high-quality and reproducible data basis for model training by constructing an accurate frequency response curve data set covering various fault types and degrees; and the vector matching technology based on the rational function model is used to extract feature parameters such as zero points and pole points representing the physical nature of the winding from the original response curve, forming a high-dimensional feature vector with strong anti-interference ability and clear physical meaning, which significantly improves the sensitivity of the feature to small deformation; the swarm intelligence optimization algorithm is used to automatically and globally optimize the key parameters of the machine learning classification model, and the cross-validation mechanism is combined to avoid model overfitting, ensuring that the diagnostic model has excellent accuracy, robustness and generalization ability; and the automatic recognition result of the model is intuitively presented in the form of a visual report through the man-machine interaction interface, supporting historical state comparison, reducing the professional threshold of the user, and realizing efficient conversion from complex data to clear decisions, thereby providing technical support for accurate assessment and operation of the transformer winding state. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0022] Figure 1 The overall flowchart of the winding deformation diagnosis method based on vector matching and machine learning provided by an embodiment of the present application.

[0023] Figure 2 The system scheme flowchart of the winding deformation diagnosis system based on vector matching and machine learning provided by an embodiment of the present application. DETAILED DESCRIPTION

[0024] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the present application.

[0025] Embodiment 1, refer to Figure 1 For an embodiment of the present application, a winding deformation diagnosis method based on vector matching and machine learning is provided, comprising: S100: Simulate fault types and different degrees on the transformer winding, obtain frequency response data in different states through frequency response analysis technology, and construct a frequency response curve data set.

[0026] S200: Perform curve fitting processing on the frequency response curve based on the rational function model, determine the optimal fitting order through the iterative optimization process, and extract feature parameters representing the essential characteristics of the winding from the fitted transfer function to form a feature vector.

[0027] S300: Construct a machine learning classification model, and use a swarm intelligence optimization algorithm to automatically search and optimize key parameters, select the optimal parameter combination through cross-validation performance evaluation, and train to obtain an optimized diagnosis model.

[0028] S400: Input the feature vector into the optimized diagnosis model, use the model to calculate the automatic recognition and classification results of the transformer winding fault type and degree, and display the diagnosis results in a visual form through a man-machine interaction interface.

[0029] It should be noted that through data acquisition and data set construction, a standard sample library covering multiple fault modes is established, solving the key problem of difficulty in obtaining fault samples in actual engineering, and through vector matching and feature extraction based on the rational function model, the limitation of traditional methods relying only on amplitude information is broken through, more essential characterization of winding physical structure changes is realized, and through the combination of swarm intelligence optimization and cross-validation model training, automatic and accurate optimization of diagnosis model parameters is realized, improving the generalization ability and stability of the model.

[0030] Embodiment 2, refer to Figure 1 For a second embodiment of the present application, the embodiment provides a winding deformation diagnosis method based on vector matching and machine learning, comprising: In step S100, the construction of the frequency response curve data set comprises steps S101-S103: S101: Set four fault types by using the self-winding transformer winding, including axial displacement (the whole high-voltage winding is moved up), radial deformation (the high-voltage winding is bulged outward), pie space change (the displacement between two middle pies is increased), and pie short circuit (short circuit between two adjacent pies).

[0031] S102: Set several fault degrees respectively, the frequency response analyzer has a sweep frequency range of 1 kHz to 2 MHz, and the amplitude and phase angle of the transfer function at different sweep points are obtained through frequency response analysis, and the amplitude and phase relationship of the output of the display system relative to the input at different frequencies are shown. The transfer function is expressed as: wherein, is a complex transfer function value at an angular frequency ω, is an amplitude response at an angular frequency ω, is a phase response at an angular frequency ω, is an imaginary unit, is a frequency value, is an angular frequency, is a circular constant.

[0032] S103: Calculate the linear amplitude and linear vector of the response according to the known amplitude and phase; The linear amplitude calculation formula is expressed as: The linear vector calculation formula is expressed as: wherein, is a real part, is an imaginary part, is a cosine operation, is a sine operation.

[0033] In the embodiments of the present application, in step S200, the forming a feature vector comprises steps S201-S203: S201: Select a proportional polynomial model, and express the transfer function as the quotient of two rational coefficient polynomials, and the proportional polynomial model is expressed as: wherein, is a fitted transfer function, is the complex frequency variable, is the order of the model, is the set vector of all polynomial coefficients to be solved, is the coefficient of the numerator polynomial, is the coefficient of the denominator polynomial; S202: Randomly initialize the zero-pole, and convert the nonlinear least squares problem into a series of linear problems by the pole relocation iteration technique; is expressed as: wherein, is the coefficient matrix constructed by the current pole position, is the set vector of all polynomial coefficients to be solved, is the observation vector, i.e., the complex frequency response value actually measured at each sampling frequency point; The least squares solution is expressed as: wherein, is the parameter vector to be solved, i.e., all unknowns to be solved in the current iteration step, is the optimal parameter vector, is the remainder related to the new pole, is the constant term of the transfer function, is the proportional term of the transfer function, is the observation vector, i.e., the complex frequency response value actually measured at each sampling frequency point, is the conjugate transpose of matrix A, is the inverse matrix of matrix ; The pole directly corresponds to the natural resonant frequency and damping characteristics of the winding, and the zero point reflects the mutual offset relationship of the internal energy of the system. When the winding causes the physical structure to change due to deformation, the inherent electromagnetic resonance characteristics will be shifted, and this change will directly affect the values of the zero point and the pole point; The pole updating formula is expressed as: wherein, is the auxiliary rational function, is the nth remainder obtained from the current solution, is the old pole of the last iteration.

[0034] S203: Traverse all scanning frequency points, subtract the complex frequency response value obtained by actual measurement from the predicted value calculated by fitting model, obtain a series of complex residuals, evaluate the fitting quality through residual analysis, and extract the zero-pole points from the converged transfer function as the feature vector; The calculation formula of complex residual is: Wherein, is the frequency point index, is the original complex frequency response value obtained by actual measurement at the kth frequency point, is the predicted frequency response value calculated by using the fitted transfer function model at the kth frequency point, is the complex residual at the kth frequency point; When And Random distribution, that is, the residual amplitude tends to zero and the phase is randomly distributed, indicating that the residual curve closely and randomly fluctuates around the zero reference line.

[0035] In an optional embodiment, in step S200, the forming feature vector further includes taking the frequency response data as the output of a dynamic system, constructing a data matrix and a time shift matrix, performing singular value decomposition dimension reduction on the data matrix, solving a low-dimensional linear system matrix, and obtaining the modal parameters (i.e. characteristic frequency and damping ratio) of the system through eigenvalue decomposition of the current matrix as the feature vector.

[0036] In another optional embodiment, in step S200, the forming feature vector can further include constructing a Hankel matrix using the frequency response data, projecting the data to a subspace through orthogonal projection, and performing singular value decomposition on the projected matrix to determine the order of the system, obtaining a state space model through least square calculation, and extracting the eigenvalues of the system matrix from the current model as the feature vector.

[0037] In step S300, the training obtains an optimized diagnostic model, including steps S301-S304: S301: Construct a machine learning classification model suitable for multi-class fault classification, process high-dimensional feature input; Calculate the numerical index value of each fault degree under each fault, and the formula is: wherein, is the reference frequency response curve vector, i.e. the frequency response data measured under healthy state of the transformer, is the to-be-measured frequency response curve vector, i.e. the frequency response data measured under fault or to-be-diagnosed state, is the index of the frequency point, is the total length of the frequency vector, i.e. the total number of samples, is the response value of the reference curve at the u-th frequency point, is the response value of the to-be-measured curve at the u-th frequency point, is the arithmetic mean of all response values of the reference curve V, is the arithmetic mean of all response values of the to-be-measured curve W, is the correlation coefficient, which measures the linear similarity degree in shape between the two curves V and W, is the Euclidean distance, is the sum of squared errors, is the relative root mean square error, is the relative error of the sum of squares, is the standard deviation, is the covariance standard deviation, is the complex correlation coefficient, is the absolute logarithmic spectrum error.

[0038] S302: search the key parameters of the model using swarm intelligence optimization algorithm, evaluate the performance of the parameters through the fitness function, and iteratively update the parameter combination.

[0039] S303: evaluate the performance of the model using cross-validation, and select the optimal parameter combination through multiple training and validation cycles.

[0040] S304: train the final diagnosis model based on the selected optimal parameter combination, and store the model parameters.

[0041] Further, in the embodiments of the present application, in step S302, the automatic search and optimization of the key parameters include steps A1-A3: A1: initialize the particle swarm, and each particle represents a parameter combination of the support vector machine model, including the penalty coefficient and the kernel function parameter.

[0042] A2: for each particle, use the leave-one-out method for model training and validation, and each sample in the data set is left out as a validation set, and the remaining samples are used for training the model, and the average classification accuracy is calculated as the fitness value; The average classification accuracy formula is expressed as: wherein, is the accuracy, is the true positive, is the true negative, is the false positive, is the false negative.

[0043] A3: update the individual historical best position of each particle and the global best position of the whole population, find the optimal parameter combination through an iterative optimization process, and output the parameter combination corresponding to the global best position.

[0044] In an optional embodiment, in step S302, the automatic search and optimization of the key parameters further include encoding the model parameters as chromosomes, initializing a population, selecting by calculating the fitness (classification accuracy) of each chromosome, retaining excellent individuals, performing a crossover operation on the selected individuals to exchange gene fragments, and supplementing a mutation operation to introduce new genes. After multiple iterations, the chromosome with the highest fitness is decoded as the optimal model parameter.

[0045] In another optional embodiment, in step S302, the automatic search and optimization of the key parameters can further include using a grey wolf optimization algorithm, simulating the social hierarchy of a wolf pack, dividing the population individuals into alpha, beta, delta and omega wolves, the alpha, beta and delta wolves representing the current three optimal solutions, guiding the omega wolves (remaining search agents) to surround and approach the prey (optimal solution) position, and simulating the hunting process by adjusting the control parameters, gradually shrinking the surrounding circle, and iteratively updating the positions of all wolves. The position of the alpha wolf is the optimal model parameter.

[0046] Further, in the embodiment of the application, in step S303, the performance evaluation of the model includes steps B1-B3: B1: for a data set containing m samples, sequentially take the i-th sample as the validation set and the remaining m-1 samples as the training set to train the model and make predictions on the current validation sample; B2: repeat m times until each sample has been validated once, and the formula is expressed as: wherein, , and is the Lagrange multiplier, is the number of samples in the current training set, is the feature vector of the i-th training sample, is the feature vector of the j-th training sample, a class label of the i-th training sample, a class label of the j-th training sample, a transpose operation, a constraint condition, and a training sample index.

[0047] B3: Calculate the average classification accuracy of m validation results as an evaluation index of model generalization ability; The formula for calculating the average classification accuracy is: wherein, an accuracy rate, a true positive, a true negative, a false positive, a false negative.

[0048] In an optional embodiment, in step S303, the performance evaluation of the model further includes randomly and uniformly dividing the original data set into k mutually exclusive subsets, selecting one subset as the validation set and the remaining k-1 subsets as the training set each time, repeating k times, changing the validation set each time, and calculating the average performance index of k validation results.

[0049] In another optional embodiment, similar to k-fold cross-validation, the core difference lies in ensuring that the proportion of samples of different classes in each subset is basically consistent with the total proportion of the original data set when dividing the data set.

[0050] In the embodiments of the present application, in step S400, the visualization form displays the diagnostic result, including steps S401-S403: S401: Post-processing the classification result to generate probability distribution and confidence score.

[0051] S402: Provide a graphical operation interface, and the user imports new test data and feature vectors to call the background trained diagnostic model for calculation, and the interface outputs the diagnostic conclusion of the fault type in the form of text.

[0052] S403: Display the diagnostic result through a visual chart and provide a comparison analysis function with the historical health state curve.

[0053] In an alternative embodiment, in step S400, the visualization form of the diagnostic result further comprises that the user accesses the online diagnostic platform through the browser, uploads the frequency response test data file, the platform backend server completes feature extraction and model diagnosis, returns the result to the front end, and the front end displays the diagnostic report and historical record trend chart in the form of a responsive web page, and supports online management and sharing of the diagnostic report by multiple users.

[0054] In another alternative embodiment, in step S400, the visualization form of the diagnostic result can further comprise developing an application on a tablet computer or AR glasses, when the user scans the physical transformer, the device captures the real scene through the camera, and the system superimposes the diagnostic result (such as fault location and severity) on the real transformer image in the form of virtual labels and 3D highlight rendering, to realize immersive and scene-based display of the diagnostic information.

[0055] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application.

[0056] Embodiment 3, refer to Figure 2 As a third embodiment of the present application, the embodiment provides a winding deformation diagnostic system based on vector matching and machine learning, comprising a data acquisition module, a data processing and feature extraction module, a model training and optimization module, and a fault diagnosis and human-computer interaction module.

[0057] The data acquisition module is configured to perform frequency sweep test on the transformer winding model through the frequency response analyzer to obtain the original frequency response data.

[0058] The data processing and feature extraction module is configured to execute a vector matching algorithm, fit the frequency domain curve into a rational transfer function, automatically determine the optimal fitting order through iterative optimization, extract the zero point and pole point representing the essential features of the winding physical structure, and form a zero-pole feature vector.

[0059] The model training and optimization module is configured to store training data, build a support vector machine classification model, automatically search for key optimal parameters of the current model using a particle swarm optimization algorithm, and evaluate the model performance through cross-validation means.

[0060] The fault diagnosis and human-computer interaction module is configured to receive new test data or feature vectors, call the final diagnostic model for real-time analysis, output the diagnostic conclusion of the fault type and degree, and display the diagnostic result in a visual form through the human-computer interaction interface.

[0061] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application, and all the modifications and equivalents should be included in the scope of the claims of the present application.

[0062] Embodiment 4, the fourth embodiment of the present application, which is different from the first three embodiments is: If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the part of the present application which essentially contributes to the prior art or the part of the technical solutions can be embodied in the form of software products, which are stored in a storage medium and include a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0063] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered a list of executable instructions for implementing logic functions, and can be specifically embodied in any computer-readable medium for use by an instruction execution system, apparatus or device, such as a computer-based system, a system including a processor or other system that can fetch instructions from the instruction execution system, apparatus or device and execute the instructions, or in conjunction with these instruction execution systems, apparatus or devices. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transport programs for use by an instruction execution system, apparatus or device or in conjunction with these instruction execution systems, apparatus or devices.

[0064] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can also be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via an optical scanner, then compiled, interpreted, or otherwise processed, and stored in a computer memory in a form that is then employable by a computer. Examples of computer-readable media include but are not limited to prime and non-transitory computer-readable media. Non-transitory computer-readable media specifically include, but are not limited to, magnetic materials, optical media, and solid-state memories. Non- transitory computer-readable media do not include carrier waves.

[0065] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the embodiments described above, various steps or methods can be implemented, for example, by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and in another embodiment, any forms of hardware, or combinations thereof, of the following can be employed: discrete logic circuitry having logic gates for implementing logic functions upon an application of data signals; application specific integrated circuits having appropriate combinational logic gates; programmable gate arrays (PGA), field programmable gate arrays (FPGA), and so forth.

Claims

1. A method for diagnosing winding deformation based on vector matching and machine learning, characterized in that: include, The fault types and different degrees are simulated on the transformer windings, and frequency response data under different states are obtained through frequency response analysis technology to construct a frequency response curve dataset. The frequency response curve is subjected to curve fitting based on a rational function model. The optimal fitting order is determined through an iterative optimization process, and characteristic parameters representing the essential characteristics of the winding are extracted from the fitted transfer function. Forming feature vectors; A machine learning classification model is constructed, and a swarm intelligence optimization algorithm is used to automatically search and optimize key parameters. The optimal parameter combination is selected through cross-validation for performance evaluation, and the optimized diagnostic model is trained. The feature vector is input into the optimized diagnostic model, and the model is used to calculate and output the automatic identification and classification results of transformer winding fault types and degrees. The diagnostic results are then displayed in a visual form through a human-computer interaction interface.

2. The vector matching and machine learning based winding deformation diagnosis method of claim 1, wherein: The construction of the frequency response curve dataset includes setting four fault types using self-wound transformer windings and setting the fault degree for each type, performing frequency response analysis using a frequency response analyzer, obtaining the amplitude and phase angle of the transfer function at different sweep points, displaying the amplitude and phase relationship of the output relative to the input at different frequencies, and calculating the linear amplitude and linear vector of the response.

3. The vector matching and machine learning based winding deformation diagnosis method of claim 2, wherein: The formation of the feature vector includes performing curve fitting processing on the frequency response curve based on a rational function model to obtain a transfer function representing the dynamic characteristics of the system. The fitting order is adjusted using an iterative optimization process, the fitting quality is evaluated through residual analysis, and system characteristic parameters are extracted from the fitted transfer function to form an eigenvector.

4. The vector matching and machine learning based winding deformation diagnosis method of claim 3, wherein: The training to obtain the optimized diagnostic model includes constructing a machine learning classification model suitable for multi-category fault classification and processing high-dimensional feature inputs; The swarm intelligence optimization algorithm is used to search for key parameters of the model, the parameter performance is evaluated through the fitness function, and the parameter combination is iteratively updated. Cross-validation is used to evaluate the model's performance, and the optimal parameter combination is selected through multiple training and validation cycles. The final diagnostic model is trained based on the selected optimal combination of parameters, and the model parameters are stored.

5. The vector matching and machine learning based winding deformation diagnostic method of claim 4, wherein: The rational function model includes selecting a proportional polynomial model as the fitting model, and the transfer function being the quotient of two rational coefficient polynomials. The proportional polynomial model is represented as: wherein, is the transfer function to be fitted, is the complex frequency variable, is the order of the model, is the set vector of all polynomial coefficients to be solved for, is the coefficient of the numerator polynomial, is the coefficient of the denominator polynomial; The zero and pole positions are initialized as the starting point of the iteration. The parameters are estimated by numerical optimization algorithm, and the nonlinear least squares problem is transformed into a linear least squares problem by using the pole relocation iteration technique. The least squares solution is expressed as: wherein, is the parameter vector to be solved, i.e. all unknowns to be solved in the current iteration step, is the optimal parameter vector, is the residual associated with the new pole, is the constant term of the transfer function, is the proportional term of the transfer function, is the observation vector, i.e. the complex frequency response values actually measured at each sampling frequency, is the conjugate transpose of matrix A, is the inverse matrix of matrix . The pole update formula is expressed as: wherein, is an auxiliary rational function, is the nth residual obtained from the current solution, is the old pole of the previous iteration; The fitting quality is evaluated by residual analysis. The complex residuals at each frequency point are calculated. If the residual curve fluctuates randomly around the zero baseline without any regular pattern, the fitting is deemed effective. The zero and pole values ​​are extracted from the convergent transfer function to form an eigenvector. The formula for calculating complex residuals is expressed as: wherein, is a frequency point index, is an original complex frequency response value obtained from actual measurement at the kth frequency point, is a predicted frequency response value calculated at the kth frequency point using the fitted transfer function model, is a complex residual at the kth frequency point.

6. The vector matching and machine learning based winding deformation diagnostic method of claim 5, wherein: The automatic search and optimization of key parameters includes initializing a particle swarm, where each particle represents a parameter combination of the support vector machine model, and the parameters include a penalty coefficient and kernel function parameters. For each particle, the model is trained and verified using the leave-one-out method, each sample in the dataset is left out as a validation set, and the remaining samples are used to train the model, and the average classification accuracy is calculated as the fitness value; The formula for calculating the average classification accuracy is: wherein, is the accuracy, is the true positive, is the true negative, is the false positive, is the false negative; Update the individual historical best position of each particle and the global best position of the whole population, find the optimal parameter combination through the iterative optimization process, and output the parameter combination corresponding to the global best position; The cross-validation includes solving a convex quadratic programming problem, obtaining a Lagrange multiplier, and determining support vectors and bias terms based on the multiplier, and using a decision function to classify input feature vectors, calculate the probability of the current sample belonging to each class, and select the class with the highest probability as the final prediction result; The formula for solving the convex quadratic programming problem is: where, , and are Lagrange multipliers, is the number of samples in the current training set, is the feature vector of the i-th training sample, is the feature vector of the j-th training sample, is the class label of the i-th training sample, is the class label of the j-th training sample, is the transpose operation, is the constraint condition, and are the training sample indices.

7. The vector matching and machine learning based winding deformation diagnostic method as claimed in claim 6, wherein: The visualization form displays the diagnostic results, including post-processing the classification results, generating probability distribution and confidence score, and providing a graphical operation interface, users import new test data and feature vectors, call the trained diagnostic model in the background for calculation, the interface outputs the diagnostic conclusion of the fault type in text form, and displays the diagnostic results through visual charts, and provides a comparison analysis function with historical health state curves.

8. A winding deformation diagnosis system based on vector matching and machine learning, applying the winding deformation diagnosis method based on vector matching and machine learning according to any one of claims 1 to 7, characterized in that, It comprises a data acquisition module, a data processing and feature extraction module, a model training and optimization module, and a fault diagnosis and human-computer interaction module. The data acquisition module is used for performing frequency sweep testing on the transformer winding model through a frequency response analyzer to obtain original frequency response data. The data processing and feature extraction module is used for executing a vector matching algorithm, fitting the frequency domain curve into a rational transfer function, and automatically determining the optimal fitting order through iterative optimization, extracting the zero point and pole point representing the essential characteristics of the winding physical structure, and forming a zero-pole feature vector. The model training and optimization module is used for storing training data, constructing a support vector machine classification model, and automatically searching for key optimal parameters of the current model using a particle swarm optimization algorithm, and evaluating the model performance through cross-validation means. The fault diagnosis and human-computer interaction module is used for receiving new test data or feature vectors, calling the final diagnostic model for real-time analysis, outputting the diagnostic conclusion of the fault type and degree, and displaying the diagnostic results in a visual form through the human-computer interaction interface. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the winding deformation diagnosis method based on vector matching and machine learning in any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the winding deformation diagnosis method based on vector matching and machine learning in any one of claims 1-7.

Images