A winding method and device for transformer processing
Through multi-sensor data acquisition and machine learning algorithms, real-time abnormality detection and adaptive control of the transformer winding process are realized, which solves the problems of low efficiency and unstable quality of traditional winding methods, and improves winding efficiency and product quality.
Patent Information
- Application Number
- CN202411124407.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-16
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-08-16
AI Technical Summary
Traditional transformer winding methods are inefficient, difficult to ensure consistency and traceability of winding quality, and difficult to adapt to the complex and changeable winding environment and the special needs of different types of transformers, resulting in difficult to detect and deal with abnormal states in a timely manner, affecting production. Efficiency and product quality.
Multi-sensor data acquisition is adopted to perform time-frequency domain analysis and feature extraction, and form a fusion feature space through multi-scale decomposition and data fusion, combine machine learning algorithms to perform abnormal detection and defect recognition, real-time feedback and adaptive control, and optimize winding process parameters. .
The efficiency and accuracy of the winding for transformer processing is improved, the stability and controllability of the production process is enhanced, and human intervention is reduced through automated closed-loop processes, and product quality and production efficiency are continuously improved.
Smart Images

Figure CN118658727B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of winding control technology, and in particular to a winding method and device for transformer processing. Background Art
[0002] Transformers are key equipment in power systems, and their performance and reliability directly affect the stable operation of the entire power grid. In the transformer manufacturing process, winding is a crucial process, which not only determines the electrical performance of the transformer, but also affects the life and reliability of the transformer. Traditional transformer winding methods mainly rely on manual operation and experience judgment, which is not only inefficient, but also difficult to ensure the consistency and traceability of winding quality. With the continuous improvement of power system requirements for transformer performance and the widespread application of large-capacity and ultra-high voltage transformers, traditional winding methods have been unable to meet the needs of modern transformer manufacturing.
[0003] In recent years, with the rapid development of sensor technology, data analysis and artificial intelligence technology, the concept of intelligent manufacturing has received widespread attention in the field of transformer production. However, current research mainly focuses on the analysis of single sensor data and simple control strategy adjustment, lacking the comprehensive utilization and in-depth mining of multi-source heterogeneous data. In addition, the existing winding control methods are often based on fixed control strategies, which are difficult to adapt to the complex and changeable winding environment and the special needs of different types of transformers. These problems make it difficult to detect and handle abnormal conditions in the winding process in a timely manner, affecting the production efficiency and product quality of the transformer. Summary of the invention
[0004] The present application provides a winding method and device for transformer processing, which are used to improve the efficiency and accuracy of winding for transformer processing.
[0005] In a first aspect, the present application provides a winding method for transformer processing, the winding method for transformer processing comprising:
[0006] Performing multi-sensor data acquisition on the winding process of the transformer to obtain an original sensor data set, wherein the original sensor data set includes tension data, speed data, vibration data and temperature data;
[0007] Performing time-frequency domain analysis and feature extraction on the original sensor data set to obtain a multi-dimensional feature parameter set;
[0008] Performing multi-scale decomposition and data fusion on the multi-dimensional feature parameter set to obtain a fused feature space;
[0009] Performing anomaly detection and defect recognition on the fused feature space to obtain anomaly marks and defect type information;
[0010] Performing real-time feedback and adaptive control according to the abnormal mark and the defect type information to obtain a target winding control strategy;
[0011] Quality evaluation and process optimization are performed on the target winding control strategy and the fused feature space to obtain an optimized winding process parameter set.
[0012] In a second aspect, the present application provides a winding device for transformer processing, the winding device for transformer processing comprising:
[0013] An acquisition module is used to perform multi-sensor data acquisition on the winding process of the transformer to obtain an original sensor data set, wherein the original sensor data set includes tension data, speed data, vibration data and temperature data;
[0014] An analysis module, used for performing time-frequency domain analysis and feature extraction on the original sensor data set to obtain a multi-dimensional feature parameter set;
[0015] A decomposition module, used for performing multi-scale decomposition and data fusion on the multi-dimensional feature parameter set to obtain a fused feature space;
[0016] An identification module is used to perform anomaly detection and defect identification on the fused feature space to obtain anomaly marks and defect type information;
[0017] A control module, used for performing real-time feedback and adaptive control according to the abnormal mark and the defect type information to obtain a target winding control strategy;
[0018] The optimization module is used to perform quality evaluation and process optimization on the target winding control strategy and the fused feature space to obtain an optimized winding process parameter set.
[0019] The third aspect of the present application provides a winding device for transformer processing, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the winding device for transformer processing executes the above-mentioned winding method for transformer processing.
[0020] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned winding method for transformer processing.
[0021] In the technical solution provided by this application, by simultaneously collecting tension, speed, vibration and temperature data, the key parameters of the winding process are fully covered, and a variety of advanced signal processing methods such as fast Fourier transform, wavelet transform, Hilbert-Huang transform, etc. are used to extract data features from different angles. Through empirical mode decomposition, nonlinear mapping, autoencoder network and other technologies, the effective fusion of multi-scale and multi-modal data is achieved, and the robustness and information richness of feature representation are improved. Combined with machine learning algorithms, a multi-level anomaly detection and defect recognition system is constructed, which can accurately identify various types of winding defects. Fuzzy rule reasoning and dynamic programming methods are used to achieve real-time adjustment and optimization of the winding process, and improve the stability and controllability of the production process. Systematically optimizing the winding process parameters through simulation experiments helps to continuously improve product quality and production efficiency. A closed-loop automated process is formed from data acquisition to process optimization, which greatly reduces human intervention and improves the efficiency and accuracy of winding for transformer processing. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.
[0023] Figure 1 A schematic diagram of an embodiment of a winding method for transformer processing in an embodiment of the present application;
[0024] Figure 2 This is a schematic diagram of an embodiment of a winding device for transformer processing in an embodiment of the present application. DETAILED DESCRIPTION
[0025] The embodiment of the present application provides a winding method and device for transformer processing. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0026] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the transformer winding method in the embodiment of the present application includes:
[0027] Step S101, performing multi-sensor data acquisition on the winding process of the transformer to obtain an original sensor data set, where the original sensor data set includes tension data, speed data, vibration data and temperature data;
[0028] It is understandable that the execution subject of the present application may be a winding device for transformer processing, or a terminal or a server, which is not limited here. The present application embodiment is described by taking a server as the execution subject as an example.
[0029] Specifically, the winding process of the transformer is monitored, and the tension sensor of the winding machine is sampled at high frequency to obtain the original tension time series data. The original tension time series data is segmented by sliding windows, and statistical features are extracted to obtain the tension feature vector. These statistical features include mean, standard deviation, peak value, etc. The speed sensor of the winding machine is monitored in real time to obtain the original speed time series data. The original speed time series data is subjected to Fourier transform and spectrum analysis to obtain the speed frequency domain features. These frequency domain features can reveal the periodicity and mutation points in the speed change. At the same time, the vibration sensor array around the winding machine is synchronously collected to obtain a multi-channel vibration signal matrix. Due to the multi-dimensional and complex characteristics of the vibration signal, principal component analysis and singular value decomposition are performed on the multi-channel vibration signal matrix to extract a representative vibration feature set. Principal component analysis can reduce the dimension of the data while retaining most of the information, while singular value decomposition can identify the main patterns in the signal. The temperature sensor in the winding process is sampled regularly to obtain the temperature time series data. In order to analyze the temperature change trend over time, the temperature time series data is subjected to time series decomposition and trend analysis to extract the temperature change characteristics. These features help to understand the fluctuation and overall change trend of temperature at different time points. The tension feature vector, velocity frequency domain features, vibration feature set, and temperature change features are synchronized and aligned in time to obtain a multimodal sensor feature matrix. The time synchronization and alignment process ensures that different sensor data are compared and analyzed under the same time reference. Data normalization and outlier detection are performed on the multimodal sensor feature matrix to eliminate the dimensional differences between different sensor data and remove outliers to obtain the original sensor data set. The original sensor data set includes tension data, velocity data, vibration data, and temperature data.
[0030] Step S102: performing time-frequency domain analysis and feature extraction on the original sensor data set to obtain a multi-dimensional feature parameter set;
[0031] Specifically, the tension data is subjected to a fast Fourier transform to obtain a spectrum of the tension signal. By analyzing the spectrum, the main frequency components in the tension signal are found. Peak detection and harmonic analysis are performed on the spectrum of the tension signal to extract the tension frequency domain feature vector. Peak detection can identify the main frequency peaks in the spectrum, while harmonic analysis can help understand the relationship and mutual influence between these frequency components. At the same time, wavelet transform is applied to the speed data to obtain the time-frequency distribution diagram of the speed signal. Wavelet transform is a time-frequency analysis tool that can provide local information of the signal in time and frequency at the same time. By calculating the energy distribution of the time-frequency distribution diagram of the speed signal, the energy distribution of the signal at different times and frequencies is determined. Feature point extraction is performed on the time-frequency distribution diagram to obtain the speed time-frequency feature matrix. Feature point extraction can identify important change points and characteristic modes in the signal. The vibration data is subjected to a Hilbert-Huang transform to obtain the intrinsic mode functions of the vibration signal, which represent different modal components in the signal. The modal energy ratio calculation and instantaneous frequency estimation are performed on the intrinsic mode functions to obtain the vibration modal feature set. Modal energy ratio calculation can evaluate the energy share of each mode in the entire signal, while instantaneous frequency estimation provides information on the change of modal frequency over time. Autoregressive integrated moving average analysis is performed on temperature data to establish a time series model of temperature signals and extract its model parameters. Model parameters describe the dynamic change characteristics of temperature signals. Residual analysis and prediction error calculation are performed on the time series model parameters of temperature signals to obtain temperature dynamic feature vectors. Residual analysis can reveal the fitting error of the model, while prediction error provides the prediction accuracy of the model in practical applications. The tension frequency domain feature vector, velocity time-frequency feature matrix, vibration modal feature set and temperature dynamic feature vector are feature concatenated to form a comprehensive multimodal feature tensor. Feature data from different sensors are integrated for unified analysis and processing. Tension decomposition is performed on multimodal feature tensors to obtain a multidimensional feature parameter set. Tensor decomposition is a multidimensional data processing technology that decomposes complex multimodal feature tensors into several low-dimensional subspaces to extract the main modes and features in the data.
[0032] Step S103, performing multi-scale decomposition and data fusion on the multi-dimensional feature parameter set to obtain a fused feature space;
[0033] Specifically, empirical mode decomposition is performed on a multidimensional feature parameter set to obtain multiple intrinsic mode functions, and Hilbert spectrum analysis is performed on multiple intrinsic mode functions to obtain a time-frequency energy distribution matrix. The matrix provides the energy distribution of the multidimensional feature parameters at different times and frequencies. Singular value decomposition is performed on the time-frequency energy distribution matrix to extract the main feature modes and corresponding eigenvalues, and the most representative components in the data are identified through these feature modes and eigenvalues. The main feature modes and eigenvalues are sorted by importance and threshold screened to obtain a feature subspace. The feature subspace contains the most representative feature information. The feature subspace is analyzed, and a nonlinear mapping of the feature representation is obtained through a nonlinear mapping method. The feature representation after nonlinear mapping is locally linearly embedded to obtain a low-dimensional manifold structure. The low-dimensional manifold structure can retain the main features of the data while reducing the dimension of the data. The low-dimensional manifold structure is trained by an autoencoder network to obtain encoder and decoder model parameters. The autoencoder network can effectively compress and reconstruct data through learning to ensure the retention of important features. The encoder and decoder model parameters are used to reconstruct the low-dimensional manifold structure and perform error analysis to obtain the reconstruction error matrix. The reconstruction error matrix can reflect the error distribution in the data reconstruction process. Fuzzy c-means clustering analysis is performed on the reconstruction error matrix to obtain the cluster center of feature fusion. The cluster center represents the gathering point of different features in the data. Adaptive weight calculation and weighted fusion are performed on the cluster center of feature fusion and the low-dimensional manifold structure to obtain the final fusion feature space. Through adaptive weight calculation, the weight of the feature is dynamically adjusted according to its importance, making the fusion feature space more representative and robust.
[0034] Step S104: performing anomaly detection and defect recognition on the fused feature space to obtain anomaly marks and defect type information;
[0035] Specifically, an isolation forest analysis is performed on the fusion feature space. This method is an unsupervised learning algorithm based on random forests and can effectively identify anomalies. Through the isolation forest analysis, the anomaly score of each sample is obtained. The higher the anomaly score, the more likely the sample is an anomaly. The sample anomaly score is thresholded, and the sample is marked as normal or abnormal according to the preset threshold to obtain the initial label. The local anomaly factor is calculated for the initially labeled sample. The local anomaly factor is a density-based anomaly detection method. The degree of abnormality is evaluated by calculating the density difference of the sample in the local neighborhood to obtain the local density anomaly. The local density anomaly is adaptively clustered. Cluster analysis can cluster samples with similar anomalies to obtain abnormal sample clusters. Support vector data description is calculated for the abnormal sample cluster to obtain an abnormal boundary hypersphere. Support vector data description is a method for describing the boundary of data distribution. By constructing a hypersphere to surround normal data, abnormal samples beyond the boundary are identified. According to the abnormal boundary hypersphere, new samples are discriminated to obtain abnormal labels. Multi-class support vector machine training is performed on the abnormally labeled samples to obtain a defect classifier model. Multi-class support vector machine is a supervised learning method for classification. Through training, multiple classification hyperplanes can be constructed to classify abnormal samples according to different defect types. According to the defect classifier model, the abnormal samples are classified and predicted to obtain the preliminary defect type. In order to improve the accuracy of defect type identification, Bayesian network inference is performed on the preliminary defect type. Bayesian network is a probabilistic graph model that calculates the probability distribution of each defect type by combining prior knowledge and data. According to the probability distribution of defect type, the maximum a posteriori probability is estimated to determine the most likely defect type, and finally obtain the abnormal label and defect type information.
[0036] Step S105: Perform real-time feedback and adaptive control according to the abnormal mark and defect type information to obtain a target winding control strategy;
[0037] Specifically, the abnormality mark and defect type information are weighted to obtain the abnormality severity score. The score is calculated based on the severity of each abnormality and its potential impact on the winding quality. According to the abnormality severity score, the sensitivity analysis of the current winding control strategy is performed to identify the control parameters that have the most significant impact on the winding process and obtain the strategy adjustment priority. Fuzzy rule reasoning is performed on the strategy adjustment priority, and the adjustment priority is converted into a preliminary control decision using the preset fuzzy rules. These fuzzy rules are based on experience and expert knowledge and can effectively handle the complex uncertainties in the winding process. Based on the preliminary control decision, similarity matching is performed to find the historical control scheme that is closest to the current abnormal situation and obtain a set of candidate control schemes. The candidate control scheme set is searched and an optimization algorithm such as a genetic algorithm or a particle swarm algorithm is used to find the optimal control sequence, which can minimize the impact of the abnormality and improve the winding quality. The optimal control sequence is dynamically programmed to decompose the overall control target into several staged control targets. Dynamic programming can effectively handle multi-stage decision problems and ensure that each step of control is optimal. Model predictive control is performed on the staged control objectives, and the future winding process is simulated using the predictive model to obtain the predictive control output. Model predictive control can foresee possible problems in advance and provide corresponding control adjustment plans. According to the predictive control output, the corresponding control stability boundary is generated to ensure that the control plan can operate stably and effectively in actual applications. According to the control stability boundary, the current winding control strategy is constrained and optimized. Constrained optimization is to adjust the control parameters on the premise of meeting all process and safety constraints to obtain the final target winding control strategy.
[0038] Step S106: perform quality assessment and process optimization on the target winding control strategy and the fused feature space to obtain an optimized winding process parameter set.
[0039] Specifically, simulation experiments are conducted on the target winding control strategy. Various situations that may be encountered in the actual production process are simulated through simulation experiments to obtain simulation result data sets. Statistical analysis and performance index calculation are performed on the simulation result data sets to extract multiple key strategy evaluation indicators, including winding accuracy, speed stability, tension consistency and temperature control effect. Multi-objective genetic analysis is performed on the strategy evaluation indicators and fusion feature space. Multi-objective genetic algorithm is an optimization tool that can optimize multiple objective functions at the same time to obtain the Pareto optimal solution set. The Pareto optimal solution set represents a set of solutions that achieve the best balance between different objectives, and these solutions cannot be dominated by other solutions on any objective. The Pareto optimal solution set is analyzed, and the key process parameter combination is identified using the decision tree analysis method. Decision tree analysis can effectively reveal the influence of each parameter on the process results and help understand which parameter combinations are most important for optimizing the process. Orthogonal experiments are performed on the key process parameter combinations. Orthogonal experimental design is an efficient experimental design method that systematically adjusts and combines different process parameters to comprehensively evaluate the influence of each parameter on the process effect and obtain an optimized process parameter scheme. Generate the global optimal parameter combination according to the process parameter optimization plan. The global optimal parameter combination is the result of further optimization based on comprehensive consideration of all key evaluation indicators and the Pareto optimal solution set and orthogonal test. Perform process verification and adjustment on the global optimal parameter combination. Apply the optimal parameter combination to actual production, observe its effect and make necessary adjustments to ensure that the expected effect and stability can be achieved in actual operation. Through repeated verification and adjustment, the optimized winding process parameter set is obtained.
[0040] In the embodiment of the present application, by simultaneously collecting tension, speed, vibration and temperature data, the key parameters of the winding process are fully covered, and a variety of advanced signal processing methods such as fast Fourier transform, wavelet transform, Hilbert-Huang transform, etc. are used to extract data features from different angles. Through empirical mode decomposition, nonlinear mapping, autoencoder network and other technologies, the effective fusion of multi-scale and multi-modal data is achieved, and the robustness and information richness of feature representation are improved. Combined with machine learning algorithms, a multi-level anomaly detection and defect recognition system is constructed, which can accurately identify various types of winding defects. Fuzzy rule reasoning and dynamic programming methods are used to achieve real-time adjustment and optimization of the winding process, and improve the stability and controllability of the production process. Systematically optimizing the winding process parameters through simulation experiments helps to continuously improve product quality and production efficiency. A closed-loop automated process is formed from data acquisition to process optimization, which greatly reduces human intervention and improves the efficiency and accuracy of winding for transformer processing.
[0041] In a specific embodiment, the process of executing step S101 may specifically include the following steps:
[0042] (1) Monitor the winding process of the transformer and perform high-frequency sampling on the tension sensor of the winding machine to obtain the original tension time series data;
[0043] (2) Perform sliding window segmentation and statistical feature extraction on the original tension time series data to obtain the tension feature vector;
[0044] (3) The speed sensor of the winding machine is monitored in real time to obtain the original speed time series data, and the original speed time series data is subjected to Fourier transform and spectrum analysis to obtain the speed frequency domain characteristics;
[0045] (4) Synchronously collect the vibration sensor array around the winding machine to obtain a multi-channel vibration signal matrix, and perform principal component analysis and singular value decomposition on the multi-channel vibration signal matrix to obtain a vibration feature set;
[0046] (5) Perform regular sampling of the temperature sensor during the winding process to obtain temperature time series data, and perform time series decomposition and trend analysis on the temperature time series data to obtain temperature change characteristics;
[0047] (6) Time synchronization and alignment of the tension feature vector, velocity frequency domain features, vibration feature set, and temperature change features to obtain a multimodal sensor feature matrix;
[0048] (7) Perform data standardization and outlier detection on the multimodal sensor feature matrix to obtain the original sensor data set, which includes tension data, velocity data, vibration data, and temperature data.
[0049] Specifically, monitor the winding process of the transformer and perform high-frequency sampling on the tension sensor of the winding machine to obtain the original tension time series data. Through high-frequency sampling, the subtle dynamics of tension changes can be captured to ensure the accuracy and integrity of the data. Set the sampling frequency to The collected tension time series data is ,in Indicates a time point. The original tension time series data is segmented by sliding windows and statistical features are extracted. The time series data is divided into windows of fixed length, and the data in each window is analyzed independently. The length of the window is , the sliding step length is Statistical feature extraction is to calculate the statistical features within each window, such as the mean , Standard Deviation , Peak Etc. Through these statistical features, the tension feature vector is obtained , which contains the key characteristic information of the tension signal. At the same time, the speed sensor of the winding machine is monitored in real time to obtain the original speed time series data The original velocity time series data is Fourier transformed and converted to the frequency domain for spectrum analysis. The Fourier transform formula is as follows:
[0050] ;
[0051] in, represents the representation of the speed signal in the frequency domain, is the frequency. Spectrum analysis to obtain the velocity frequency domain characteristics , such as the main frequency components and their amplitudes. The vibration sensor array around the winding machine is synchronously collected to obtain a multi-channel vibration signal matrix , each column represents a vibration signal of a sensor. The multi-channel vibration signal matrix is subjected to principal component analysis and singular value decomposition. Principal component analysis is used for dimensionality reduction and extraction of main features. The formula is as follows:
[0052] ;
[0053] in, is the eigenvector matrix. The singular value decomposition formula is:
[0054] ;
[0055] in, and is the matrix obtained by singular value decomposition, is a diagonal singular value matrix. Through principal component analysis and singular value decomposition, the vibration feature set is obtained During the winding process, the temperature sensor is sampled regularly to obtain temperature time series data. . Perform time series decomposition and trend analysis on temperature time series data. Time series decomposition can be divided into trend , Seasonality and residual Three parts, the formula is as follows
[0056] ;
[0057] Through trend analysis, the temperature change characteristics are obtained . The tension feature vector, velocity frequency domain features, vibration feature set and temperature change features are synchronized and aligned in time to obtain a multimodal sensor feature matrix. The sensor data are aligned according to a unified time axis to ensure that the data of each feature at the same time point can match. Data standardization and outlier detection are performed on the multimodal sensor feature matrix. The purpose of standardization is to eliminate the dimensional differences of different sensor data and make them on the same scale. Outlier detection is to identify and process abnormal points in the data to ensure the reliability of the data. The original sensor data set is obtained, which includes tension data, velocity data, vibration data and temperature data.
[0058] In a specific embodiment, the process of executing step S102 may specifically include the following steps:
[0059] (1) Perform fast Fourier transform on the tension data to obtain the spectrum of the tension signal, and perform peak detection and harmonic analysis on the spectrum of the tension signal to obtain the tension frequency domain feature vector;
[0060] (2) Apply wavelet transform to the speed data to obtain the time-frequency distribution diagram of the speed signal, and perform energy distribution calculation and feature point extraction on the time-frequency distribution diagram of the speed signal to obtain the speed time-frequency feature matrix;
[0061] (3) Performing Hilbert-Huang transform on the vibration data to obtain the intrinsic mode function of the vibration signal, and performing modal energy ratio calculation and instantaneous frequency estimation on the intrinsic mode function of the vibration signal to obtain a vibration mode feature set;
[0062] (4) Perform autoregressive integrated moving average analysis on the temperature data to obtain the time series model parameters of the temperature signal, and perform residual analysis and prediction error calculation on the time series model parameters of the temperature signal to obtain the temperature dynamic feature vector;
[0063] (5) Perform feature concatenation on the tension frequency domain feature vector, velocity time-frequency feature matrix, vibration modal feature set, and temperature dynamic feature vector to obtain a multi-modal feature tensor;
[0064] (6) Perform tensor decomposition on the multimodal feature tensor to obtain a multi-dimensional feature parameter set.
[0065] Specifically, the tension data is subjected to a fast Fourier transform to convert the tension signal in the time domain into the frequency domain in order to analyze its frequency components. Assume that the tension time series data is ,in is the time variable. The formula for fast Fourier transform is as follows:
[0066] ;
[0067] in, is the frequency domain signal, is the signal length, is an imaginary unit. Through fast Fourier transform, the spectrum of the tension signal is obtained, which represents the amplitude of the signal at different frequencies. Peak detection and harmonic analysis are performed on the spectrum to extract the main frequency components and their corresponding harmonics to form the tension frequency domain feature vector For the speed data, wavelet transform is applied to obtain the time-frequency distribution diagram of the speed signal. Assume that the speed time series data is , the wavelet transform formula is as follows:
[0068] ;
[0069] in, are the wavelet coefficients, is the scale parameter, is the translation parameter, is the mother wavelet function, is its complex conjugate. Through wavelet transform, the time-frequency distribution diagram of the speed signal is obtained, revealing the energy distribution of the speed signal at different times and frequencies. The energy distribution calculation and feature point extraction of the time-frequency distribution diagram are performed to form the speed time-frequency feature matrix The vibration data is processed by Hilbert-Huang transform. The vibration signal is decomposed into several intrinsic mode functions by empirical mode decomposition. Assuming that the vibration signal is , after decomposition IMF is IM For each IM Perform Hilbert transform to get the instantaneous frequency and modal energy The specific formula is as follows:
[0070] ;
[0071] Through the above process, the vibration mode feature set is obtained , including features such as modal energy ratio and instantaneous frequency. For temperature data, the autoregressive integrated moving average model is used for analysis. Assume that the temperature time series data is , the form of the autoregressive integrated moving average model is:
[0072] ;
[0073] in, is the autoregressive coefficient, is the sliding mean coefficient, is the error term. Through the autoregressive integral sliding average model, the time series model parameters of the temperature signal are extracted, and the model residual is analyzed to calculate the prediction error and form the temperature dynamic feature vector . Perform feature concatenation on various features. Assume that the tension frequency domain feature vector is , the velocity time-frequency feature matrix is , the vibration mode feature set is , the temperature dynamic eigenvector is . Through feature cascading, a multimodal feature tensor is formed , expressed as:
[0074] ;
[0075] For multimodal feature tensors Perform tensor decomposition to extract its multi-dimensional feature parameter set. Tensor decomposition can be performed using CP decomposition, and its formula is:
[0076] ;
[0077] in, is a factor matrix, represents the outer product operation, is the decomposition rank. Through tensor decomposition, we get a multidimensional feature parameter set ,These parameter sets contain the key features of multimodal sensor data.
[0078] In a specific embodiment, the process of executing step S103 may specifically include the following steps:
[0079] (1) Performing empirical mode decomposition on the multidimensional feature parameter set to obtain multiple intrinsic mode functions, and performing Hilbert spectrum analysis on the multiple intrinsic mode functions to obtain the time-frequency energy distribution matrix;
[0080] (2) Perform singular value decomposition on the time-frequency energy distribution matrix to obtain the main eigenmodes and corresponding eigenvalues;
[0081] (3) Sort the main feature patterns and feature values by importance and perform threshold screening to obtain the feature subspace, and analyze the feature subspace to obtain the feature representation after nonlinear mapping;
[0082] (4) Perform local linear embedding on the feature representation after nonlinear mapping to obtain a low-dimensional manifold structure, and perform autoencoder network training on the low-dimensional manifold structure to obtain encoder and decoder model parameters;
[0083] (5) Reconstruct the low-dimensional manifold structure and perform error analysis based on the encoder and decoder model parameters to obtain the reconstruction error matrix;
[0084] (6) Perform fuzzy c-means clustering analysis on the reconstruction error matrix to obtain the cluster center of feature fusion, and perform adaptive weight calculation and weighted fusion on the cluster center of feature fusion and the low-dimensional manifold structure to obtain the fused feature space.
[0085] Specifically, empirical mode decomposition is performed on the multidimensional feature parameter set to decompose the complex signal into several intrinsic mode functions. Assume that the multidimensional feature parameter set is ,in is the time variable. Through empirical mode decomposition, Can be decomposed into multiple IMs , each IM Represents an intrinsic mode in the signal. The formula is as follows:
[0086] ;
[0087] in, is the number of modes, is the residual term. For each IM Perform Hilbert spectrum analysis to obtain its instantaneous frequency and modal energy By calculating the Hilbert transform of each mode , and obtain the time-frequency energy distribution matrix ,in For time, is the frequency, and the matrix elements represent the energy distribution at different times and frequencies. Perform singular value decomposition to extract the main characteristic modes and corresponding eigenvalues. The formula for singular value decomposition is as follows:
[0088] ;
[0089] in, and are the left and right singular matrices, is a diagonal singular value matrix. Singular values Indicates the matrix The energy size of each characteristic pattern. Sort the importance of the main characteristic patterns and eigenvalues, retain the patterns with greater energy contribution, and set a threshold for screening to obtain the characteristic subspace. Perform nonlinear mapping on the characteristic subspace. Use kernel methods or other nonlinear techniques to analyze the characteristic subspace and obtain the characteristic representation after nonlinear mapping. . Nonlinear mapping can reveal the complex structure and relationship in the data, making the feature representation more effective. The feature representation after nonlinear mapping is locally linearly embedded to obtain a low-dimensional manifold structure. Local linear embedding is a dimensionality reduction method that embeds high-dimensional data into a low-dimensional space by retaining the linear relationship of data points in the local neighborhood. The formula is as follows:
[0090] ;
[0091] in, is a weight matrix, which represents the linear reconstruction relationship of each data point in its neighborhood. Through local linear embedding, a low-dimensional manifold structure is obtained. The autoencoder network is trained on the low-dimensional manifold structure to obtain the encoder and decoder model parameters. The autoencoder is a neural network that can learn the compression representation and reconstruction representation of data through training. Assume that the encoder function is , the decoder function is , the goal of the autoencoder is to minimize the reconstruction error Through training, the model parameters of the encoder and decoder are obtained. According to the model parameters of the encoder and decoder, the low-dimensional manifold structure Perform reconstruction and error analysis and calculate the reconstruction error matrix , the formula is as follows:
[0092] ;
[0093] Reconstruction error matrix Represents the difference between the original data and the reconstructed data. The reconstruction error matrix Perform fuzzy c-means clustering analysis to obtain the cluster centers of feature fusion. Fuzzy c-means clustering is a soft clustering method that assigns data points to multiple cluster centers by minimizing the weighted square error. The formula is as follows:
[0094] ;
[0095] in, It is a data point Belongs to the cluster center The membership degree of is the cluster center, is the fuzzy factor. By minimizing the objective function , get the cluster center of feature fusion. Perform adaptive weight calculation and weighted fusion on the cluster center of feature fusion and low-dimensional manifold structure to get the final fusion feature space. Adaptive weight calculation dynamically adjusts the weight of each feature according to the similarity between the cluster center and the manifold structure, making the fusion feature space more reasonable and effective.
[0096] In a specific embodiment, the process of executing step S104 may specifically include the following steps:
[0097] (1) Perform isolation forest analysis on the fused feature space to obtain the sample anomaly score, and then perform threshold processing on the sample anomaly score to obtain the initial label;
[0098] (2) Calculate the local anomaly factor of the initially marked samples to obtain the local density anomaly, and adaptively cluster the local density anomaly to obtain the abnormal sample cluster;
[0099] (3) Perform support vector data description calculation on the abnormal sample cluster to obtain the abnormal boundary hypersphere, and then judge the new sample based on the abnormal boundary hypersphere to obtain the abnormal label;
[0100] (4) Perform multi-class support vector machine training on the abnormally labeled samples to obtain a defect classifier model. Based on the defect classifier model, classify and predict the abnormal samples to obtain the preliminary defect type.
[0101] (5) Perform Bayesian network inference on the preliminary defect type to obtain the defect type probability distribution, and perform maximum a posteriori probability estimation based on the defect type probability distribution to obtain the abnormal label and defect type information.
[0102] Specifically, an isolation forest analysis is performed on the fused feature space. Isolation forest is an unsupervised learning algorithm based on decision trees, which is used to detect outliers. By constructing multiple random trees to isolate data points, isolation forest can effectively identify abnormal samples. Assume that the fused feature space is ,in Indicates The output of the isolation forest is the anomaly score for each sample. , the higher the score, the more likely the sample is an outlier. Threshold the sample anomaly score and set a threshold , the score is greater than The samples are marked as abnormal and the initial label is obtained ,in Indicates The label of the sample, 0 means normal, 1 means abnormal. The local anomaly factor is calculated for the initially labeled sample. The local anomaly factor is used to evaluate the degree of abnormality of the sample in the local neighborhood. The degree of abnormality is determined by comparing the sample density with the density of its neighborhood. Assume that the sample The local density is , the sample density in the neighborhood is , then the local anomaly factor The calculation formula is as follows:
[0103] ;
[0104] The larger the LOF value, the more likely the sample is an outlier. By calculating the local density anomaly, the abnormal samples are further identified. The local density anomaly is adaptively clustered, and a clustering algorithm such as DBSCAN (density-based clustering method) is used to cluster the abnormal samples together to obtain abnormal sample clusters. Support vector data description is calculated for the abnormal sample clusters. Support vector data description is a method for describing the boundaries of data distribution. It constructs a minimum hypersphere to surround the normal data, thereby identifying abnormal samples beyond the boundaries. Assume that the center of the hypersphere is , the radius is , the goal of the support vector data description is to minimize the following objective function:
[0105] ;
[0106] in, is the slack variable, is the penalty parameter. By optimizing the objective function, the abnormal boundary hypersphere is obtained. According to the abnormal boundary hypersphere, the new sample is judged to determine whether it is an abnormal sample and obtain the abnormal label. The multi-class support vector machine is trained on the abnormally labeled samples to obtain the defect classifier model. The multi-class support vector machine is a supervised learning method for classification. By constructing multiple classification hyperplanes, the abnormal samples are classified according to different defect types. Assume that the training data is ,in Representation sample The goal of the multi-class support vector machine is to minimize the following objective function:
[0107] ;
[0108] in, is the weight vector of the classification hyperplane, is the bias term. By optimizing the objective function, the defect classifier model is obtained. According to the defect classifier model, the abnormal samples are classified and predicted to obtain the preliminary defect type. In order to improve the accuracy of defect type identification, the preliminary defect type is inferred by Bayesian network. Bayesian network is a probabilistic graph model that calculates the probability distribution of each defect type by combining prior knowledge and data. Assume that the defect type is , the observed data is , the goal of the Bayesian network is to calculate the posterior probability , the formula is as follows:
[0109] ;
[0110] Through Bayesian network reasoning, the probability distribution of defect types is obtained. According to the probability distribution of defect types, the maximum a posteriori probability estimation is performed to determine the most likely defect type and obtain the final abnormality mark and defect type information.
[0111] In a specific embodiment, the process of executing step S105 may specifically include the following steps:
[0112] (1) Assign weights to the abnormality mark and defect type information to obtain the abnormality severity score. Based on the abnormality severity score, perform sensitivity analysis on the current winding control strategy to obtain the strategy adjustment priority.
[0113] (2) Perform fuzzy rule reasoning on the strategy adjustment priority to obtain a preliminary control decision, and perform similarity matching based on the preliminary control decision to obtain a set of candidate control schemes;
[0114] (3) Search the candidate control scheme set to obtain the optimal control sequence, and perform dynamic programming decomposition on the optimal control sequence to obtain the staged control target;
[0115] (4) Perform model predictive control on the staged control objectives to obtain predictive control outputs, and generate corresponding control stability boundaries based on the predictive control outputs;
[0116] (5) According to the control stability boundary, the current winding control strategy is constrained and optimized to obtain the target winding control strategy.
[0117] Specifically, the anomaly mark and defect type information are weighted. The anomaly severity score is achieved by assigning corresponding weights to each anomaly and defect type. Assume that the anomaly mark is , the defect type is , then the abnormality severity score The calculation formula is:
[0118] ;
[0119] in, Is the defect type The weight of is the indicator function, when belong If the type is 1, otherwise it is 0. According to the abnormality severity score, a sensitivity analysis is performed on the current winding control strategy to determine which control parameters have the greatest impact on the abnormality severity, and then determine the priority of strategy adjustment. Assume that the parameter set of the current control strategy is , the results of the sensitivity analysis are that each parameter The degree of impact on the anomaly severity score. Perform fuzzy rule reasoning on the strategy adjustment priority. Fuzzy rule reasoning is a method of making decisions using fuzzy logic. Define fuzzy rules based on parameter sensitivity and anomaly severity scores, such as "If the anomaly severity is high and the parameter sensitivity is high, the parameter needs to be adjusted immediately." Obtain a preliminary control decision through fuzzy rule reasoning. Based on the preliminary control decision, perform similarity matching to find control strategies similar to the current situation from historical data to form a set of candidate control schemes. Search the set of candidate control schemes and use an optimization algorithm (such as a genetic algorithm or a particle swarm algorithm) to find the optimal control sequence in the candidate scheme set. The optimal control sequence refers to the strategy that can minimize the anomaly severity score among multiple control strategies. Perform dynamic programming decomposition on the optimal control sequence. Dynamic programming is a technique that decomposes a problem into multiple sub-problems and solves them step by step. Assume that the optimal control sequence can be decomposed into multiple stage control objectives Each stage control target corresponds to a control parameter adjustment within a specific time period. Model predictive control is performed on the stage control target. Model predictive control is a method that uses system models to predict future behavior and optimize control. Through model predictive control, the predicted control output of each stage control target is obtained. Assume the system state is , the control input is , the prediction model is , the goal of model predictive control is to minimize the following cost function:
[0120] ;
[0121] in, is the reference state, is the control input. By optimizing this cost function, the predicted control output is obtained . The corresponding control stability boundary is generated according to the predicted control output. The control stability boundary is used to ensure that the control strategy can operate stably and effectively in practical applications. The stability boundary can be defined by constraints such as the upper and lower limits of the control parameters and the control change rate. According to the control stability boundary, the current winding control strategy is constrained and optimized. Constrained optimization is to adjust the control parameters under the premise of satisfying all process and safety constraints to obtain the target winding control strategy. Assume that the constraints are , the goal of constrained optimization is to minimize the following objective function:
[0122] ;
[0123] in, is the optimal control parameter, is the weight parameter.
[0124] In a specific embodiment, the process of executing step S106 may specifically include the following steps:
[0125] (1) Conduct simulation experiments on the target winding control strategy to obtain a simulation result data set, perform statistical analysis and performance index calculation on the simulation result data set, and obtain strategy evaluation indicators;
[0126] (2) Perform multi-objective genetic analysis on the strategy evaluation index and fusion feature space to obtain the Pareto optimal solution set;
[0127] (3) Conduct decision tree analysis on the Pareto optimal solution set to obtain key process parameter combinations, and conduct orthogonal experiments on the key process parameter combinations to obtain process parameter optimization solutions;
[0128] (4) Generate a global optimal parameter combination based on the process parameter optimization plan, and perform process verification and adjustment on the global optimal parameter combination to obtain an optimized winding process parameter set.
[0129] Specifically, a simulation experiment is conducted on the target winding control strategy, and the actual winding process is simulated in a virtual environment to verify the effectiveness of the control strategy. Through the simulation experiment, a large amount of data is collected to form a simulation result data set. Assume that a winding control strategy is designed and run in a simulation experiment. The collected data include parameters such as tension, speed, and temperature. Statistical analysis and performance index calculation are performed on the simulation result data set. Statistical analysis is to extract useful information from a large amount of data, and performance index calculation is to evaluate the effectiveness of the control strategy. Common performance indicators include tension stability, speed uniformity, temperature fluctuation, etc. Similarly, statistical analysis is performed on data such as speed and temperature to obtain corresponding performance indicators. Multi-objective genetic analysis is performed on the strategy evaluation index and the fusion feature space. The multi-objective genetic algorithm is an evolutionary algorithm used to optimize multi-objective problems. By simulating the natural selection process, the multi-objective genetic algorithm can find a set of solutions that balance each objective to form a Pareto optimal solution set. Assume that the performance indicator is , the fusion feature space is , the goal of multi-objective optimization is to minimize the following objective function:
[0130] ;
[0131] in, and is the weight parameter. Through the selection, crossover and mutation operations of the genetic algorithm, a set of Pareto optimal solutions can be obtained. Perform decision tree analysis on the Pareto optimal solution set. Decision tree analysis is a supervised learning method for classification and regression. It identifies the features that have the greatest impact on the target variable by recursively splitting the data. Perform decision tree analysis on the Pareto optimal solution set to obtain key process parameter combinations. Perform orthogonal experiments on key process parameter combinations. Orthogonal experimental design is an efficient experimental design method. By systematically arranging and combining various parameters, the impact of each parameter on the process results can be comprehensively evaluated to obtain an optimized process parameter solution. Assume that the key process parameters are , through orthogonal experiments, a series of experimental schemes are formed, the effect of each scheme is calculated, and an optimal scheme is obtained. The global optimal parameter combination is generated according to the process parameter optimization scheme. The global optimal parameter combination is obtained through optimization and adjustment based on the orthogonal test results. Assume that the process parameter optimization scheme is , the global optimal parameter combination is , where each parameter Yes The optimization result is obtained. The global optimal parameter combination is verified and adjusted. Process verification is to apply the global optimal parameter combination in the actual production environment, observe its effect and make necessary adjustments. Through repeated verification and adjustment, the optimized winding process parameter set is finally obtained.
[0132] The above describes the winding method for transformer processing in the embodiment of the present application. The following describes the winding device for transformer processing in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the winding device for transformer processing includes:
[0133] The acquisition module 201 is used to perform multi-sensor data acquisition on the winding process of the transformer to obtain an original sensor data set, wherein the original sensor data set includes tension data, speed data, vibration data and temperature data;
[0134] An analysis module 202 is used to perform time-frequency domain analysis and feature extraction on the original sensor data set to obtain a multi-dimensional feature parameter set;
[0135] A decomposition module 203 is used to perform multi-scale decomposition and data fusion on the multi-dimensional feature parameter set to obtain a fused feature space;
[0136] The recognition module 204 is used to perform anomaly detection and defect recognition on the fused feature space to obtain anomaly marks and defect type information;
[0137] A control module 205, configured to perform real-time feedback and adaptive control according to the abnormal mark and the defect type information to obtain a target winding control strategy;
[0138] The optimization module 206 is used to perform quality evaluation and process optimization on the target winding control strategy and the fused feature space to obtain an optimized winding process parameter set.
[0139] Through the collaboration of the above components, by collecting tension, speed, vibration and temperature data at the same time, the key parameters of the winding process are fully covered. A variety of advanced signal processing methods such as fast Fourier transform, wavelet transform, Hilbert-Huang transform, etc. are used to extract data features from different angles. Through empirical mode decomposition, nonlinear mapping, autoencoder network and other technologies, the effective fusion of multi-scale and multi-modal data is achieved, and the robustness and information richness of feature representation are improved. Combined with machine learning algorithms, a multi-level anomaly detection and defect recognition system is constructed, which can accurately identify various types of winding defects. Fuzzy rule reasoning and dynamic programming methods are used to achieve real-time adjustment and optimization of the winding process, improving the stability and controllability of the production process. Systematically optimizing winding process parameters through simulation experiments helps to continuously improve product quality and production efficiency. A closed-loop automated process is formed from data acquisition to process optimization, which greatly reduces human intervention and improves the efficiency and accuracy of winding for transformer processing.
[0140] The present application also provides a winding device for transformer processing, which includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor executes the steps of the winding method for transformer processing in the above-mentioned embodiments.
[0141] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the winding method for transformer processing.
[0142] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.
[0143] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.
[0144] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A winding method for transformer processing, characterized in that: The method comprises: The winding process of the transformer is subjected to multi-sensor data acquisition to obtain an original sensor data set, wherein the original sensor data set includes tension data, speed data, vibration data and temperature data; specifically, the winding process of the transformer is monitored, and the tension sensor of the winding machine is sampled at high frequency to obtain original tension time series data; the original tension time series data is subjected to sliding window segmentation and statistical feature extraction to obtain tension feature vectors; the speed sensor of the winding machine is monitored in real time to obtain original speed time series data, and the original speed time series data is subjected to Fourier transform and spectrum analysis to obtain speed frequency domain features; the vibration sensor array around the winding machine is synchronously collected to obtain to a multi-channel vibration signal matrix, and perform principal component analysis and singular value decomposition on the multi-channel vibration signal matrix to obtain a vibration feature set; perform timed sampling on the temperature sensor in the winding process to obtain temperature time series data, and perform time series decomposition and trend analysis on the temperature time series data to obtain temperature change characteristics; perform time synchronization and alignment on the tension feature vector, the speed frequency domain feature, the vibration feature set and the temperature change feature to obtain a multi-modal sensor feature matrix; perform data standardization and outlier detection on the multi-modal sensor feature matrix to obtain an original sensor data set, wherein the original sensor data set includes tension data, speed data, vibration data and temperature data; The original sensor data set is subjected to time-frequency domain analysis and feature extraction to obtain a multidimensional feature parameter set; specifically, the method comprises: performing fast Fourier transform on the tension data to obtain a spectrum of the tension signal, performing peak detection and harmonic analysis on the spectrum of the tension signal to obtain a tension frequency domain feature vector; applying wavelet transform on the speed data to obtain a time-frequency distribution diagram of the speed signal, performing energy distribution calculation and feature point extraction on the time-frequency distribution diagram of the speed signal to obtain a speed time-frequency feature matrix; performing Hilbert-Huang transform on the vibration data to obtain an intrinsic mode function of the vibration signal, and performing peak detection and harmonic analysis on the spectrum of the tension signal to obtain a tension frequency domain feature vector; applying wavelet transform on the speed data to obtain a time-frequency distribution diagram of the speed signal, performing energy distribution calculation and feature point extraction on the time-frequency distribution diagram of the speed signal to obtain a speed time-frequency feature matrix; performing Hilbert-Huang transform on the vibration data to obtain an intrinsic mode function of the vibration signal, and performing peak detection and harmonic analysis on the spectrum of the tension signal to obtain a tension frequency domain feature vector; performing ... and performing energy distribution calculation and feature point extraction on the time-frequency distribution diagram of the speed signal to obtain a speed time-frequency feature matrix; performing wavelet transform on the vibration data to obtain an intrinsic mode function of the vibration signal, and performing peak detection and harmonic analysis on the spectrum of the tension signal to obtain a tension frequency domain feature vector; performing wavelet transform on the speed data to obtain a time-frequency distribution diagram of the speed signal, and performing energy distribution calculation and feature point extraction on the time-frequency distribution diagram of the speed signal to obtain a speed time-frequency feature matrix; performing wavelet The inherent modal function of the vibration signal is subjected to modal energy ratio calculation and instantaneous frequency estimation to obtain a vibration modal feature set; the temperature data is subjected to autoregressive integral sliding average analysis to obtain time series model parameters of the temperature signal, and residual analysis and prediction error calculation are performed on the time series model parameters of the temperature signal to obtain a temperature dynamic feature vector; the tension frequency domain feature vector, the velocity time-frequency feature matrix, the vibration modal feature set and the temperature dynamic feature vector are subjected to feature cascading to obtain a multi-modal feature tensor; the multi-modal feature tensor is subjected to tensor decomposition to obtain a multi-dimensional feature parameter set; The multidimensional feature parameter set is subjected to multi-scale decomposition and data fusion to obtain a fused feature space; specifically, the method includes: performing empirical mode decomposition on the multidimensional feature parameter set to obtain a plurality of intrinsic mode functions, and performing Hilbert spectrum analysis on the plurality of intrinsic mode functions to obtain a time-frequency energy distribution matrix; performing singular value decomposition on the time-frequency energy distribution matrix to obtain main feature modes and corresponding eigenvalues; performing importance sorting and threshold screening on the main feature modes and the eigenvalues to obtain a feature subspace, and analyzing the feature subspace to obtain a feature representation after nonlinear mapping; performing local linear embedding on the feature representation after nonlinear mapping to obtain a low-dimensional manifold structure, and performing autoencoder network training on the low-dimensional manifold structure to obtain encoder and decoder model parameters; performing reconstruction and error analysis on the low-dimensional manifold structure according to the encoder and decoder model parameters to obtain a reconstruction error matrix; performing fuzzy c-means clustering analysis on the reconstruction error matrix to obtain a clustering center of feature fusion, and performing adaptive weight calculation and weighted fusion on the clustering center of feature fusion and the low-dimensional manifold structure to obtain a fused feature space; Anomaly detection and defect identification are performed on the fused feature space to obtain anomaly labels and defect type information; specifically, the method includes: performing isolation forest analysis on the fused feature space to obtain sample anomaly scores, and performing thresholding processing on the sample anomaly scores to obtain initial labels; performing local anomaly factor calculation on the initially labeled samples to obtain local density anomaly degrees, and adaptively clustering the local density anomaly degrees to obtain abnormal sample clusters; performing support vector data description calculation on the abnormal sample clusters to obtain an abnormal boundary hypersphere, and discriminating new samples based on the abnormal boundary hypersphere to obtain abnormal labels; performing multi-class support vector machine training on the abnormally labeled samples to obtain a defect classifier model, and classifying and predicting the abnormal samples based on the defect classifier model to obtain preliminary defect types; performing Bayesian network inference on the preliminary defect types to obtain defect type probability distribution, and performing maximum a posteriori probability estimation based on the defect type probability distribution to obtain anomaly labels and defect type information; According to the abnormal mark and the defect type information, real-time feedback and adaptive control are performed to obtain a target winding control strategy; specifically, the following steps are performed: weighting the abnormal mark and the defect type information to obtain an abnormal severity score, and based on the abnormal severity score, sensitivity analysis is performed on the current winding control strategy to obtain a strategy adjustment priority; fuzzy rule reasoning is performed on the strategy adjustment priority to obtain a preliminary control decision, and similarity matching is performed based on the preliminary control decision to obtain a candidate control scheme set; the candidate control scheme set is searched to obtain an optimal control sequence, and the optimal control sequence is dynamically programmed to obtain a staged control target; model predictive control is performed on the staged control target to obtain a predictive control output, and a corresponding control stability boundary is generated based on the predictive control output; based on the control stability boundary, constrained optimization is performed on the current winding control strategy to obtain a target winding control strategy; The target winding control strategy and the fused feature space are subjected to quality assessment and process optimization to obtain an optimized winding process parameter set; specifically comprising: performing simulation experiments on the target winding control strategy to obtain a simulation result data set, and performing statistical analysis and performance index calculation on the simulation result data set to obtain a strategy evaluation index; performing multi-objective genetic analysis on the strategy evaluation index and the fused feature space to obtain a Pareto optimal solution set; performing decision tree analysis on the Pareto optimal solution set to obtain a key process parameter combination, and performing orthogonal experiments on the key process parameter combination to obtain a process parameter optimization scheme; generating a global optimal parameter combination according to the process parameter optimization scheme, and performing process verification and adjustment on the global optimal parameter combination to obtain an optimized winding process parameter set.
2. A winding device for transformer processing, characterized in that: Used to perform the transformer winding method according to claim 1, the device comprises: The acquisition module is used to perform multi-sensor data acquisition on the winding process of the transformer to obtain an original sensor data set, wherein the original sensor data set includes tension data, speed data, vibration data and temperature data; specifically, the module includes: monitoring the winding process of the transformer, and performing high-frequency sampling on the tension sensor of the winding machine to obtain original tension time series data; performing sliding window segmentation and statistical feature extraction on the original tension time series data to obtain a tension feature vector; performing real-time monitoring on the speed sensor of the winding machine to obtain original speed time series data, and performing Fourier transform and spectrum analysis on the original speed time series data to obtain speed frequency domain features; synchronously sampling the vibration sensor array around the winding machine; A multi-channel vibration signal matrix is obtained, and principal component analysis and singular value decomposition are performed on the multi-channel vibration signal matrix to obtain a vibration feature set; the temperature sensor in the winding process is sampled regularly to obtain temperature time series data, and time series decomposition and trend analysis are performed on the temperature time series data to obtain temperature change characteristics; the tension feature vector, the speed frequency domain feature, the vibration feature set and the temperature change feature are time synchronized and aligned to obtain a multi-modal sensor feature matrix; data standardization and outlier detection are performed on the multi-modal sensor feature matrix to obtain an original sensor data set, and the original sensor data set includes tension data, speed data, vibration data and temperature data; The analysis module is used to perform time-frequency domain analysis and feature extraction on the original sensor data set to obtain a multidimensional feature parameter set; specifically, the module includes: performing fast Fourier transform on the tension data to obtain a frequency spectrum of the tension signal, performing peak detection and harmonic analysis on the frequency spectrum of the tension signal to obtain a tension frequency domain feature vector; applying wavelet transform on the speed data to obtain a time-frequency distribution diagram of the speed signal, performing energy distribution calculation and feature point extraction on the time-frequency distribution diagram of the speed signal to obtain a speed time-frequency feature matrix; performing Hilbert-Huang transform on the vibration data to obtain an intrinsic mode function of the vibration signal, The modal energy ratio and instantaneous frequency of the inherent modal function of the vibration signal are calculated to obtain a vibration modal feature set; the temperature data is analyzed by autoregressive integral sliding average to obtain the time series model parameters of the temperature signal, and the time series model parameters of the temperature signal are analyzed by residual analysis and prediction error calculation to obtain a temperature dynamic feature vector; the tension frequency domain feature vector, the velocity time-frequency feature matrix, the vibration modal feature set and the temperature dynamic feature vector are feature concatenated to obtain a multi-modal feature tensor; the multi-modal feature tensor is decomposed into a tensor to obtain a multi-dimensional feature parameter set; The decomposition module is used to perform multi-scale decomposition and data fusion on the multidimensional feature parameter set to obtain a fused feature space; specifically comprising: performing empirical mode decomposition on the multidimensional feature parameter set to obtain multiple intrinsic mode functions, and performing Hilbert spectrum analysis on the multiple intrinsic mode functions to obtain a time-frequency energy distribution matrix; performing singular value decomposition on the time-frequency energy distribution matrix to obtain main feature modes and corresponding eigenvalues; performing importance sorting and threshold screening on the main feature modes and the eigenvalues to obtain a feature subspace, and analyzing the feature subspace to obtain a nonlinear mapping The method comprises the following steps: performing local linear embedding on the feature representation after nonlinear mapping to obtain a low-dimensional manifold structure, and performing autoencoder network training on the low-dimensional manifold structure to obtain encoder and decoder model parameters; reconstructing and error analyzing the low-dimensional manifold structure according to the encoder and decoder model parameters to obtain a reconstruction error matrix; performing fuzzy c-means clustering analysis on the reconstruction error matrix to obtain a clustering center of feature fusion, and performing adaptive weight calculation and weighted fusion on the clustering center of feature fusion and the low-dimensional manifold structure to obtain a fused feature space; The identification module is used to perform anomaly detection and defect identification on the fused feature space to obtain anomaly labels and defect type information; specifically comprising: performing isolation forest analysis on the fused feature space to obtain sample anomaly scores, and performing thresholding processing on the sample anomaly scores to obtain initial labels; performing local anomaly factor calculation on the initially labeled samples to obtain local density anomaly degrees, and adaptively clustering the local density anomaly degrees to obtain abnormal sample clusters; performing support vector data description calculation on the abnormal sample clusters to obtain an abnormal boundary hypersphere, and discriminating new samples based on the abnormal boundary hypersphere to obtain abnormal labels; performing multi-class support vector machine training on the abnormally labeled samples to obtain a defect classifier model, and classifying and predicting the abnormal samples based on the defect classifier model to obtain preliminary defect types; performing Bayesian network inference on the preliminary defect types to obtain defect type probability distribution, and performing maximum a posteriori probability estimation based on the defect type probability distribution to obtain anomaly labels and defect type information; A control module is used to perform real-time feedback and adaptive control according to the abnormal mark and the defect type information to obtain a target winding control strategy; specifically comprising: weighting the abnormal mark and the defect type information to obtain an abnormal severity score, and performing sensitivity analysis on the current winding control strategy according to the abnormal severity score to obtain a strategy adjustment priority; performing fuzzy rule reasoning on the strategy adjustment priority to obtain a preliminary control decision, and performing similarity matching based on the preliminary control decision to obtain a candidate control scheme set; searching the candidate control scheme set to obtain an optimal control sequence, and performing dynamic programming decomposition on the optimal control sequence to obtain a staged control target; performing model predictive control on the staged control target to obtain a predictive control output, and generating a corresponding control stability boundary according to the predictive control output; performing constrained optimization on the current winding control strategy according to the control stability boundary to obtain a target winding control strategy; The optimization module is used to perform quality assessment and process optimization on the target winding control strategy and the fused feature space to obtain an optimized winding process parameter set; specifically comprising: performing simulation experiments on the target winding control strategy to obtain a simulation result data set, and performing statistical analysis and performance index calculation on the simulation result data set to obtain a strategy evaluation index; performing multi-objective genetic analysis on the strategy evaluation index and the fused feature space to obtain a Pareto optimal solution set; performing decision tree analysis on the Pareto optimal solution set to obtain a key process parameter combination, and performing orthogonal experiments on the key process parameter combination to obtain a process parameter optimization scheme; generating a global optimal parameter combination according to the process parameter optimization scheme, and performing process verification and adjustment on the global optimal parameter combination to obtain an optimized winding process parameter set.
3. A winding device for transformer processing, characterized in that: The transformer processing winding device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory to enable the transformer winding equipment to execute the transformer winding method according to claim 1.
4. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the transformer winding method according to claim 1 is implemented.
Citation Information
Patent Citations
Transformer winding tension intelligent control system and control method thereof
CN114864269A
Automatic winding machine monitoring method and system
CN116631769A
Plastic master batch quality inspection method and system based on machine vision
CN117095005A
Intelligent sensing and fault diagnosis method for transformer winding insulator processing equipment
CN118396942A
Failure diagnosis method for power transformer winding based on gsmallat-NIN-CNN network
US20210382120A1
