Machine learning method and system for dry-type transformer manufacturing process optimization

Abstract

The present application relates to the field of power equipment manufacturing, in particular to a machine learning method and system for optimizing the manufacturing process of dry-type transformers; the present application collects dry-type transformer production process parameters and quality index data, calculates the correlation coefficient to screen features, and uses recursive feature elimination sorting; based on the process parameters, a quality control model is constructed using support vector regression, and the Bagging method is integrated; the produced transformer is subjected to machine vision quality detection to determine defects; the transformer with quality problems is traced back to the process, and the correlation coefficient and model parameters are adjusted; the present application improves the intelligent level of dry-type transformer manufacturing, realizes online quality prediction and control, improves detection efficiency and accuracy, provides basis for quality problem positioning and improvement, and improves production efficiency and quality level.

Description

Technical Field

[0001] This invention relates to the field of power equipment manufacturing, specifically to a machine learning method and system for optimizing the manufacturing process of dry-type transformers. Background Technology

[0002] Dry-type transformers are crucial electrical equipment in power systems, and their manufacturing quality directly impacts the safe and stable operation of the power grid. Traditional dry-type transformer manufacturing processes rely primarily on manual experience for quality control, resulting in low production efficiency and inconsistent quality.

[0003] The manufacturing process of dry-type transformers is complex, involving multiple stages such as winding, impregnation, and curing. The process parameters at each stage have a significant impact on product quality. Traditional manufacturing methods mainly rely on the experience of operators to set and adjust process parameters, lacking scientific guidance and optimization methods, resulting in unstable product quality.

[0004] The quality inspection of dry-type transformers mainly relies on manual sampling, which has a limited sampling rate, low inspection efficiency, and is prone to missing defective products. Furthermore, manual inspection is susceptible to subjective factors, and inconsistent inspection standards lead to inaccurate quality assessment results.

[0005] When quality problems arise, traditional production models lack effective process traceability and parameter adjustment mechanisms. Often, the cause of the problem needs to be identified manually, and the process adjusted based on experience, which is time-consuming, inefficient, and makes it difficult to guarantee the accuracy and effectiveness of the adjustments.

[0006] With the increasing market demand for dry-type transformers, traditional manufacturing methods are no longer sufficient to meet the requirements of large-volume, high-quality production. There is an urgent need to introduce intelligent manufacturing and management methods to improve production efficiency and product quality consistency.

[0007] To address the aforementioned issues, there is an urgent need for a machine learning method and system for optimizing the manufacturing process of dry-type transformers, in order to improve the production efficiency and quality of dry-type transformers and meet the power system's demand for high-quality dry-type transformers. Summary of the Invention

[0008] To achieve the above objectives, this invention provides a machine learning method and system for optimizing the manufacturing process of dry-type transformers, the specific technical solution of which is as follows:

[0009] Collect process parameters for the production of dry-type transformers and obtain quality indicators for dry-type transformers.

[0010] Calculate the correlation coefficient between process parameters and quality indicators, select process parameters that are strongly correlated with quality indicators as features, and use recursive feature elimination to sort the features;

[0011] Based on the collected process parameters, a support vector regression (SVR) algorithm is used to construct a production quality prediction model for dry-type transformers, and the Bagging learning method is integrated into the production quality control model.

[0012] The dry-type transformers produced are subjected to quality inspection. Image data of the completed dry-type transformers are collected, and the image data is inspected using a machine vision model to determine whether there are quality defects in the dry-type transformers.

[0013] For dry-type transformers with quality problems, process traceability is carried out, and the correlation coefficients of process parameters and quality indicators are adjusted. The parameters in the dry-type transformer production quality prediction model are also adjusted.

[0014] Preferably, the process parameters of the dry-type transformer production process are collected, including: winding tension F, winding speed v, and impregnation time t. imm , Immersion temperature T imm Curing temperature T cure Curing time t cure ;

[0015] The quality indicators of dry-type transformers are collected, including: insulation resistance R. ins Inductance L, transformer ratio N, core loss P iron Copper loss P copper And partial discharge q;

[0016] The collected process parameters and quality index data are cleaned to remove outliers, and missing values ​​are filled in using interpolation. A dataset is constructed based on the preprocessed process parameters and quality index data.

[0017] Let the entire dataset D be represented as: In the dataset, x i =x1,x2,…,x d Let ] be the feature vector of the process parameters of the i-th sample. The feature vector of the process parameters contains d process parameters, x d =[F,v,t imm ,T imm ,T cure ,t cure ];y i Let y be the quality index value corresponding to the i-th sample. i =[y1,y2,…,y d ], y d =[R ins ,L,N,P iron ,P copper ,q], where n is the total number of samples.

[0018] Preferably, the Pearson correlation coefficient between each process parameter and the quality index is calculated, and parameters with a correlation coefficient greater than a preset threshold THp are selected. r The process parameters are used as candidate features;

[0019] The Recursive Feature Emission (RFE) algorithm is used to rank candidate features by importance. The specific steps are as follows:

[0020] Input candidate feature set X = [x1, x2, ..., x d ], where x d Let d represent the d-th feature vector, k be the target number of features (k < 6), and M be the final number of selected features. Set the base model M to evaluate the importance of features.

[0021] Initialize the feature ranking list to store the ranking results of features; record the current feature set S = [1, 2, ..., d], which represents the index value of the features that have not been excluded;

[0022] Repeat the following steps until |S| = k:

[0023] In the current feature subset X S Train the base model M and obtain the model parameters σ;

[0024] Calculate the importance score for each feature: c j =|σ j |,σ j Indicates the model feature weight coefficients;

[0025] Find the feature index j with the lowest importance min =argmin j∈S c j ; will j min Remove from S;

[0026] Output a feature ranking list R. Based on the preset target number of features k, select the top k features from the ranking list R as the final feature subset.

[0027] X k = [X[d-k+1],X[d-k+2],…,X[d]], where X k This represents the k selected feature vectors.

[0028] Preferably, a support vector regression (SVR) algorithm is used to construct a production quality prediction model for dry-type transformers; the process parameter feature values ​​X are selected based on the recursive feature elimination (RFE) algorithm. k and quality indicator data y i , serving as the training set for the dry-type transformer production quality prediction model; the objective function of the SVR algorithm is:

[0029]

[0030] Where w and b are the parameters of the SVR model, C is the penalty coefficient, and ξ i and Let φ(X) be a slack variable, ε be the error tolerance, and φ(X) be the s k ) is the feature mapping function, where i is the parameter and n is the number of samples;

[0031] The Gaussian kernel function RBF is selected as the feature mapping function of SVR, and the hyperparameters A and γ of SVR are optimized by grid search.

[0032] The expression for the Gaussian kernel function RBF is: K(X) g ,X l )=exp(-γ||X g -X l || 2 ); where γ is a hyperparameter of the kernel function, X g and X l Let K(X) be the feature vector of two samples; g ,X l ) represents the Gaussian kernel function, used to calculate the similarity between two samples in a high-dimensional feature space;

[0033] The steps of a grid search are as follows:

[0034] Set the search range and step size for A and γ to generate a grid; for each (A, γ) combination in the grid, perform the following steps: train the model using (A, γ) as hyperparameters for SVR; evaluate the model performance through cross-validation;

[0035] The parameter (A, γ) with the smallest cross-validation evaluation value is selected as the optimal hyperparameter for SVR.

[0036] Preferably, random resampling generates m training subsets, including: setting D as the original training set, and the training subset D obtained by resampling in the f-th round. f for: in, n1 is the number of subset samples randomly drawn from D, and the number of subset samples is equal to the number of original training set samples; the resampling process is repeated m times to obtain m training subsets; For the sample The corresponding label values; using the optimized SVR model parameters on each training subset, train the SVR sub-model, ultimately obtaining m sub-models; for each SVR sub-model, train using the following steps:

[0037] Use the Gaussian kernel function RBF as the feature mapping function for SVR;

[0038] The hyperparameters A and γ of each sub-model are optimized using a grid search.

[0039] Using training subset D i Train the SVR sub-model to obtain the SVR sub-model parameters ω i and b i ;

[0040] The performance of m SVR sub-models is evaluated on the test set, using mean squared error (MSE) and mean absolute error (MAE) as evaluation metrics; the MSE and MAE of each sub-model are calculated, and the weight coefficients of the sub-model are determined based on their reciprocals.

[0041] The prediction results of the m SVR sub-models are weighted and averaged to obtain the final predicted value of the dry-type transformer quality index.

[0042] in, Let α be the predicted value of the u-th sub-model for the g-th test sample. u denoted as the weight coefficient of the u-th sub-model.

[0043] Preferably, image data of the completed dry-type transformer is collected, and the image data of the dry-type transformer includes an external view of the dry-type transformer and an internal structure view of the dry-type transformer.

[0044] The dry-type transformer appearance diagram includes high-resolution images of the transformer's exterior for detecting surface defects, including cracks and dents; the dry-type transformer internal structure diagram includes X-ray images of the transformer's internal structure for detecting internal defects.

[0045] The YOLOv5 object detection algorithm is used to locate defect regions in an image. The steps of the YOLOv5 algorithm are as follows: Divide the input image into E×E grids; predict B bounding boxes for each grid, and each bounding box contains 5 predicted values: x, y, w. b ,h b c; where x, y are the center coordinates of the bounding box, and w b ,h b Let be the width and height of the bounding box, and c be the confidence level; predict P class probabilities for each bounding box, representing the probability that the bounding box belongs to each defect class;

[0046] Based on the confidence level and category probability, some bounding boxes are filtered out to obtain the final defect detection result;

[0047] The detected defect region image patch is input into a pre-trained ResNet-50 classification model to determine the defect type; the ResNet-50 classification model includes an input layer, a convolutional layer, a residual block, a fully connected layer, and an output layer;

[0048] The input layer receives defect image patches. The convolutional layer, consisting of multiple convolutional and pooling layers, extracts hierarchical features from the image. The residual block, consisting of multiple concatenated residual blocks, each containing two convolutional layers and a short-circuit connection, is used to train the deep network. The fully connected layer flattens the features extracted by the convolutional layers and maps them to the probability distribution of defect categories. The output layer uses the Softmax activation function to output the probability of each defect category.

[0049] Based on the defect type and severity, a quality assessment result is given, including pass and fail; let the defect type be q and the severity be s, then the quality assessment function is:

[0050]

[0051] Among them, C pass S is the set of allowed defect types. threshold The severity threshold is Q(q,s) = 1, which indicates that the quality assessment is satisfactory, and Q(q,s) = 0, which indicates that the quality assessment is unsatisfactory.

[0052] Preferably, for dry-type transformers that fail quality inspection, the process parameter data x from their production process is extracted. fail And obtain the corresponding quality index y fail ;

[0053] The process parameters of the non-conforming samples x fail and quality indicators y fail Adding to the original dataset D, we get the updated dataset D'=D∪(x fail ,y fail );

[0054] Based on the updated dataset D', feature selection is performed again, the correlation coefficients between process parameters and quality indicators are recalculated, and the feature ranking is adjusted.

[0055] The SVR model and the Bagging ensemble model are retrained using the updated dataset D' to obtain the updated quality prediction model. The prediction performance of the models before and after the update on the test set is compared using mean squared error (MSE) and mean absolute error (MAE). If the MSE and MAE values ​​of the updated model are both less than those of the unupdated model, then the new model is used to guide the production of dry-type transformers.

[0056] Preferably, the machine learning system for optimizing the manufacturing process of dry-type transformers includes: a data acquisition module, a feature calculation module, a production quality control module, a quality inspection module, and a process traceability adjustment module.

[0057] The data acquisition module is used to collect process parameters of the dry-type transformer production process and to obtain the quality indicators of the dry-type transformer.

[0058] The feature calculation module is used to calculate the correlation coefficient between process parameters and quality indicators, select process parameters that are strongly correlated with quality indicators as features, and sort the features using recursive feature elimination.

[0059] The production quality control module, based on the collected process parameters, uses the Support Vector Regression (SVR) algorithm to construct a production quality prediction model for dry-type transformers, and integrates the Bagging learning method into the production quality control model.

[0060] The quality inspection module is used to inspect the quality of the produced dry-type transformers, collect image data of the completed dry-type transformers, and use a machine vision model to inspect the image data to determine whether the dry-type transformers have quality defects.

[0061] The process traceability adjustment module is used to trace the process of dry-type transformers with quality problems, adjust the correlation coefficients of process parameters and quality indicators, and adjust the parameters in the dry-type transformer production quality prediction model.

[0062] An electronic device includes a processor and a memory, wherein the memory stores a computer program that can be invoked by the processor; the processor executes a machine learning method for optimizing the dry-type transformer manufacturing process by invoking the computer program stored in the memory.

[0063] A computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform a machine learning method for optimizing the dry-type transformer manufacturing process.

[0064] The beneficial effects of this invention are as follows: By collecting process parameters and quality index data of the production process, this invention establishes a data correlation between the production process and product quality, providing a foundation for subsequent data analysis and model building.

[0065] This invention calculates the correlation coefficients between process parameters and quality indicators, enabling a quantitative assessment of the impact of each process parameter on product quality and identifying key process parameters with significant quality influence as features. Using a recursive feature elimination algorithm to sort the features further optimizes the feature subset, removes redundant and noisy features, and improves the model's generalization performance.

[0066] This invention employs the Support Vector Regression (SVR) algorithm to construct a quality control model, which establishes a nonlinear mapping relationship between process parameters and quality indicators, enabling the prediction and control of production quality. SVR exhibits good generalization performance and noise resistance, and can handle small sample sizes and nonlinear problems. Integrating the Bagging learning method into the SVR model further enhances its stability and robustness. By weighted averaging the prediction results of multiple SVR sub-models, the discrepancies and fluctuations of individual models can be reduced, improving the accuracy of quality prediction.

[0067] This invention utilizes machine vision technology to perform quality inspection on manufactured dry-type transformers, enabling automatic defect identification and location. By acquiring images of the transformer's exterior and internal structure, and employing target detection and image classification algorithms, surface and internal defects can be quickly and accurately identified, improving the efficiency and accuracy of quality inspection.

[0068] This invention, through process traceability, can quickly pinpoint key process steps and parameters causing quality problems, providing a basis for root cause analysis and improvement. Simultaneously, by feeding data on non-conforming products back into the original dataset, the correlation coefficients between process parameters and quality indicators are adjusted, and the feature selection results are dynamically updated. This allows for adaptation to changes in the production process and continuous optimization of the feature subset. Attached Figure Description

[0069] Figure 1 Flowchart of the machine learning method for optimizing the manufacturing process of dry-type transformers provided by the present invention;

[0070] Figure 2 The structure diagram of the machine learning system for optimizing the manufacturing process of dry-type transformers provided by this invention. Detailed Implementation

[0071] To better understand the invention, various aspects of the invention will be described in more detail with reference to the accompanying drawings. It should be understood that these detailed descriptions are merely illustrative of exemplary embodiments of the invention and are not intended to limit the scope of the invention in any way. Throughout the specification, the same reference numerals refer to the same elements. The expression "and / or" includes any and all combinations of one or more of the associated listed items.

[0072] In the accompanying drawings, the size, dimensions, and shapes of the elements have been slightly adjusted for ease of illustration. The drawings are for illustrative purposes only and are not strictly to scale. As used herein, the terms “approximately,” “about,” and similar terms are used to indicate approximation rather than degree, and are intended to illustrate inherent deviations in measured or calculated values ​​that will be recognized by one of ordinary skill in the art. Furthermore, the order in which the steps are described in this invention does not necessarily indicate the order in which these steps occur in actual operation, unless otherwise expressly defined or deduced from the context.

[0073] It should also be understood that expressions such as "comprising," "including," "having," "containing," and / or "comprising" are open-ended rather than closed-ended expressions in this specification, indicating the presence of the stated features, elements, and / or components, but not excluding the presence of one or more other features, elements, components, and / or combinations thereof. Furthermore, when expressions such as "at least one of..." appear after a list of listed features, they modify the entire list of features, not just individual elements in the list. Additionally, when describing embodiments of the invention, the word "may" is used to mean "one or more embodiments of the invention." And the term "exemplary" is intended to refer to examples or illustrations.

[0074] Unless otherwise specified, all terms used herein (including engineering and technical terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that, unless expressly stated herein, terms defined in common dictionaries shall be interpreted as having the meaning consistent with their meaning in the context of the relevant art, and not in an idealized or overly formalized sense.

[0075] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other. The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0076] Example 1

[0077] Reference Figure 1 This is the first embodiment of the present invention, which provides a machine learning method for optimizing the manufacturing process of dry-type transformers.

[0078] S1: Collect process parameters for the production of dry-type transformers and obtain quality indicators for dry-type transformers.

[0079] The process parameters for the production of dry-type transformers are collected, including: winding tension f, winding speed v, and impregnation time t. imm , Immersion temperature T imm Curing temperature T cureCuring time t cure .

[0080] The winding tension F controls the wire tension during the winding process to ensure winding quality; the winding speed v controls the operating speed of the winding machine, affecting winding efficiency and quality; the impregnation time t... imm The time the transformer is impregnated in the insulating varnish affects its insulation performance; the impregnation temperature T imm Controlling the temperature of the insulating varnish affects its viscosity and penetration during the impregnation process; the curing temperature T... cure Controlling the curing temperature of the transformer after impregnation affects the curing effect of the insulating varnish; curing time t cure Controlling the curing time of the transformer after impregnation affects the degree of curing of the insulating varnish.

[0081] The quality indicators of dry-type transformers are collected, including: insulation resistance R. ins Inductance L, transformer ratio N, core loss P iron Copper loss P copper And partial discharge q.

[0082] Insulation resistance R ins The transformer's insulation performance is reflected by its main inductance index; inductance L reflects the transformer's inductance value; the transformation ratio N reflects the ratio of the primary and secondary voltages; and core loss P... iron Reflects the transformer's losses under no-load conditions; copper loss P copper It reflects the coil loss under transformer load; partial discharge q reflects the local high-frequency discharge caused by transformer insulation defects.

[0083] The collected process parameters and quality index data are cleaned to remove outliers, and missing values ​​are filled in using interpolation. A dataset is constructed based on the preprocessed process parameters and quality index data.

[0084] Let the entire dataset be represented as: In the dataset, x i =[x1,x2,…,x d Let x be the feature vector of the process parameters of the i-th sample. The feature vector of the process parameters contains d process parameters. d =[F,v,t imm ,T imm ,T cure ,t cure ];y i For the corresponding quality index value, y i =[y1,y2,…,y d ], y d =[R ins ,T,N,P iron ,P copper,q], where n is the total number of samples.

[0085] The data is normalized and scaled to the [0, 1] interval; the normalization formula is: Where x is the original data, x min and x max These are the minimum and maximum values ​​of the data, x. norm This is the normalized data.

[0086] By collecting key process parameters and quality indicators from the dry-type transformer production process, and cleaning and preprocessing the collected data, a high-quality dataset was constructed. This dataset not only includes the key factors affecting the quality of dry-type transformers, but also handles missing and outlier values ​​through interpolation and other methods, laying a solid data foundation for subsequent feature selection and model training, and helping to improve the reliability and effectiveness of optimization methods.

[0087] S2: Calculate the correlation coefficient between process parameters and quality indicators, select process parameters that are strongly correlated with quality indicators as features, and use recursive feature elimination to sort the features.

[0088] Calculate the Pearson correlation coefficient p between each process parameter and quality index. r Select those with a correlation coefficient greater than the preset threshold THp r The process parameters are used as candidate features; the Pearson correlation coefficient formula is:

[0089]

[0090] Where, x z and y z These represent the z-th process parameter and the quality index value, respectively. and These represent the average values ​​of the corresponding process parameters and quality indicators, respectively.

[0091] The Recursive Feature Elimination (RFE) algorithm is used to rank candidate features by importance. The basic idea of ​​the RFE algorithm is to repeatedly build the model and remove the least important features from the current feature set each time until a preset number of features is reached.

[0092] The specific steps of the Recursive Feature Emission (RFE) algorithm are as follows:

[0093] Input candidate feature set X = [x1, x2, ..., x d ], where x d Let d represent the d-th feature vector, k be the target number of features (k < 6), and M be the final number of selected features. Set the base model M to evaluate feature importance; typically, a linear model such as LinearRegression is chosen.

[0094] Initialize the feature ranking list to store the ranking results of features; record the current feature set S = [1, 2, ..., d], which represents the index value of the features that have not been excluded;

[0095] Repeat the following steps until |S| = k:

[0096] In the current feature subset X S Train the base model M and obtain the model parameters σ;

[0097] Calculate the importance score c for each feature. j =|σ j |,σ j The weight coefficient represents the feature weight coefficient. The absolute value of the weight coefficient represents the importance score of the feature. The larger the absolute value of the weight coefficient, the greater the influence of the feature on the model output and the higher its importance.

[0098] Find the feature index j with the lowest importance min =argmin j∈S c j ; will j min Remove from S;

[0099] Output a ranking list R of features, where the features at the end of the list are the least important and the features at the beginning of the list are the most important.

[0100] Based on the preset target number of features k, the top k features of the ranking list R are selected as the final feature subset: X k = [X[d-k+1],X[d-k+2],…,X[d]], where X k This represents the k selected features.

[0101] By repeatedly training the model and evaluating feature importance, the RFE algorithm can gradually select the most relevant and discriminative feature subsets, helping to improve the performance of subsequent quality prediction models.

[0102] A feature selection method combining Pearson correlation coefficient and recursive feature elimination can be introduced to screen key parameters that are closely related to and of high importance to the quality of dry-type transformers from numerous process parameters. Through dimensionality reduction and feature optimization, not only is the complexity of the model reduced and the training efficiency improved, but the generalization ability of the model is also enhanced, enabling it to more accurately predict the quality indicators of dry-type transformers and provide a reliable basis for quality optimization and control.

[0103] S3: Based on the collected process parameters, a support vector regression (SVR) algorithm is used to construct a production quality prediction model for dry-type transformers, and the Bagging learning method is integrated into the production quality control model to improve the robustness of the production quality prediction model for dry-type transformers.

[0104] A support vector regression (SVR) algorithm was used to construct a production quality prediction model for dry-type transformers; the process parameter feature value X was selected based on the recursive feature elimination (RFE) algorithm. k and quality indicator data y i , serving as the training set for the dry-type transformer production quality prediction model; the objective function of the SVR algorithm is:

[0105]

[0106] Where w and b are the parameters of the SVR model, C is the penalty coefficient, and ξ i and Let φ(X) be a slack variable, ε be the error tolerance, and φ(X) be the s k ) is the feature mapping function, where i is the parameter and n is the number of samples.

[0107] The Gaussian kernel function RBF is selected as the feature mapping function of SVR, and the hyperparameters A and γ of SVR are optimized by grid search.

[0108] The expression for the Gaussian kernel function RBF is: K(X) g ,X l )=exp(-γ||X g -X l || 2 ); where γ is a hyperparameter of the kernel function, X g and X l Let K(X) be the feature vector of two samples; g ,X l ) represents the Gaussian kernel function, used to calculate the similarity between two samples in a high-dimensional feature space.

[0109] The basic steps of grid search are as follows:

[0110] Set A and γ as the search range and step size for the grid search, and generate the grid. For each (A,γ) combination in the grid, perform the following steps: train the model using (A,γ) as hyperparameters for SVR; evaluate the model performance through cross-validation.

[0111] The parameter (A, γ) with the smallest cross-validation evaluation value is selected as the optimal hyperparameter for SVR.

[0112] A support vector regression (SVR) algorithm is used to construct a quality prediction model for dry-type transformers, which can accurately predict various quality indicators of dry-type transformers based on key process parameters. Introducing a Gaussian kernel function and grid search effectively handles the nonlinear relationship between process parameters and quality indicators, and optimizes model hyperparameters to further improve the accuracy of quality prediction. Quality prediction based on the SVR model can promptly detect abnormal fluctuations in process parameters, providing a basis for adjusting production processes, thereby achieving quality control and optimization in the dry-type transformer manufacturing process.

[0113] Random resampling generates m training subsets, including: the training subset D obtained by resampling in the f-th round with D as the original training set. f for: in, n1 is the number of subset samples randomly drawn from D, and the number of subset samples is equal to the number of original training set samples; the resampling process is repeated m times to obtain m training subsets; For the sample The corresponding label values; using the optimized SVR model parameters on each training subset, train the SVR sub-model, ultimately obtaining m sub-models; for each SVR sub-model, train using the following steps:

[0114] Use the Gaussian kernel function RBF as the feature mapping function for SVR;

[0115] The hyperparameters A and γ of each sub-model are optimized using a grid search.

[0116] Using training subset D i Train the SVR sub-model to obtain the SVR sub-model parameters ω i and b i ;

[0117] The performance of m SVR sub-models is evaluated on the test set using mean squared error (MSE) and mean absolute error (MAE) as evaluation metrics. The MSE and MAE of each sub-model are calculated, and their reciprocals are used to determine the weight coefficients of the sub-model.

[0118]

[0119] Among them, MSE u Let $\mathbf{u}$ be the mean squared error of the $u$-th sub-model. u Let be the mean absolute error of the u-th sub-model;

[0120] The prediction results of the m SVR sub-models are weighted and averaged to obtain the final predicted value of the dry-type transformer quality index.

[0121] in, Let α be the predicted value of the u-th sub-model for the g-th test sample. u denoted as the weight coefficient of the u-th sub-model.

[0122] The prediction results of the m sub-models are averaged to obtain the final predicted quality index value of the dry-type transformer: f t (x) represents the prediction of the t-th sub-model for the test sample.

[0123] S4: Perform quality inspection on the produced dry-type transformers, collect image data of the completed dry-type transformers, and use machine vision models to inspect the image data to determine whether the dry-type transformers have quality defects.

[0124] Based on the SVR quality prediction model, the Bagging ensemble learning method is introduced. By randomly resampling the training data, multiple SVR sub-models are constructed, and their prediction results are averaged. This effectively reduces the randomness and uncertainty of a single model, improving the stability and robustness of the quality prediction results. The Bagging ensemble model can better adapt to process fluctuations and data noise in the dry-type transformer production process, making quality prediction more reliable. This guides the dynamic optimization of production parameters and ensures the consistency of product quality.

[0125] Image data of the completed dry-type transformer is collected, including an external view and an internal structure view of the dry-type transformer.

[0126] The external view of the dry-type transformer includes high-resolution images of the transformer's exterior for detecting surface defects, including cracks and dents; the internal structure view of the dry-type transformer includes X-ray images of the transformer's internal structure for detecting internal defects, such as insulation failure and short circuits.

[0127] The YOLOv5 object detection algorithm is used to locate defect regions in an image. The basic steps of the YOLOv5 algorithm are as follows: Divide the input image into E×E grids; predict B bounding boxes for each grid, and each bounding box contains 5 predicted values: x, y, w. b ,h b c; where x, y are the center coordinates of the bounding box, and w b ,h b Let be the width and height of the bounding box, and c be the confidence level. For each bounding box, predict P class probabilities, which represent the probability that the bounding box belongs to each defect class.

[0128] Based on the confidence level and class probability, bounding boxes with low confidence and low class probability are filtered out to obtain the final defect detection results.

[0129] The detected defect region image patch is input into a pre-trained ResNet-50 classification model to determine the defect type; the ResNet-50 classification model includes an input layer, a convolutional layer, a residual block, a fully connected layer, and an output layer.

[0130] The input layer receives defect image patches. The convolutional layer, consisting of multiple convolutional and pooling layers stacked together, extracts hierarchical features from the image. The residual block, consisting of multiple concatenated residual blocks, each containing two convolutional layers and a short-circuit connection, is used to train the deep network. The fully connected layer flattens the features extracted by the convolutional layers and maps them to the probability distribution of defect categories. The output layer uses the Softmax activation function to output the probability of each defect category.

[0131] Online quality inspection of completed dry-type transformers is performed using the YOLOv5 target detection algorithm and the ResNet-50 image classification model. This allows for the rapid and accurate identification of external and internal defects in the transformers, achieving automated product quality inspection. Compared to manual inspection, this method is more efficient and accurate, enabling timely detection of quality issues, reducing the outflow of defective products, and improving the product pass rate. By comparing the inspection results with preset quality assessment standards, the system can automatically determine whether the product is qualified, providing a basis for quality grading and improvement.

[0132] Based on the defect type and severity, a quality assessment result is given, including pass and fail; let the defect type be q and the severity be s, then the quality assessment function is:

[0133]

[0134] Among them, C pass S is the set of allowed defect types. threshold The severity threshold is Q(q,s) = 1, which indicates that the quality assessment is satisfactory, and Q(q,s) = 0, which indicates that the quality assessment is unsatisfactory.

[0135] S5: Conduct process tracing for dry-type transformers with quality problems, adjust the correlation coefficients of process parameters and quality indicators, and adjust the parameters in the dry-type transformer production quality prediction model.

[0136] For dry-type transformers that fail quality inspection, extract the process parameter data x from their production process. fail And obtain the corresponding quality index y fail .

[0137] The process parameters of the non-conforming samples x fail and quality indicators y failAdding to the original dataset D, we get the updated dataset D'=D∪(x fail ,y fail ).

[0138] Based on the updated dataset D', feature selection is performed again, the correlation coefficients between process parameters and quality indicators are recalculated, and the feature ranking is adjusted.

[0139] The SVR model and the Bagging ensemble model were retrained using the updated dataset D' to obtain the updated quality prediction model.

[0140] The predictive performance of the models before and after the update on the test set is compared using mean squared error (MSE) and mean absolute error (MAE). If both the MSE and MAE values ​​of the updated model are less than those of the unupdated model, then the new model is used to guide the production of dry-type transformers.

[0141] For dry-type transformers that fail quality inspection, process parameters can be traced back, and the data can be fed back into the original dataset to dynamically update and optimize the model, enabling closed-loop optimization control of the production process. Newly added non-conforming sample data can expand the training set, improving the model's generalization and adaptability; dynamic model updates can adjust the correlation between key process parameters and quality indicators in a timely manner, making quality predictions more accurate and optimization guidance more timely and effective.

[0142] Through the continuous accumulation of production data and the constant self-improvement of models, intelligent optimization of the dry-type transformer manufacturing process can be achieved, thereby continuously improving product quality and production efficiency.

[0143] Example 2

[0144] Reference Figure 2 This is a second embodiment of the present invention, which provides a machine learning system for optimizing the manufacturing process of dry-type transformers.

[0145] The system includes: a data acquisition module, a feature calculation module, a production quality control module, a quality inspection module, and a process traceability adjustment module.

[0146] The data acquisition module is used to collect process parameters during the production of dry-type transformers and to obtain quality indicators of dry-type transformers.

[0147] The feature calculation module is used to calculate the correlation coefficient between process parameters and quality indicators, select process parameters that are strongly correlated with quality indicators as features, and sort the features using recursive feature elimination.

[0148] The production quality control module, based on the collected process parameters, uses the Support Vector Regression (SVR) algorithm to construct a production quality prediction model for dry-type transformers, and integrates the Bagging learning method into the production quality control model.

[0149] The quality inspection module is used to inspect the quality of the produced dry-type transformers, collect image data of the completed dry-type transformers, and use a machine vision model to inspect the image data to determine whether the dry-type transformers have quality defects.

[0150] The process traceability adjustment module is used to trace the process of dry-type transformers with quality problems, adjust the correlation coefficients of process parameters and quality indicators, and adjust the parameters in the dry-type transformer production quality prediction model.

[0151] Example 3

[0152] The present invention also provides an electronic device. This electronic device may include one or more processors and one or more memories. The memories store computer-readable code that, when executed by the one or more processors, can perform the machine learning method for optimizing the dry-type transformer manufacturing process as described above.

[0153] The methods or systems according to embodiments of the present invention can also be implemented using the architecture of the electronic devices of the present invention.

[0154] Electronic devices may include buses, one or more CPUs, read-only memory (ROM), random access memory (RAM), communication ports connected to a network, input / output components, hard disks, etc.

[0155] Storage devices in electronic devices, such as ROM or hard disks, can store the machine learning method for optimizing the dry-type transformer manufacturing process provided by this invention.

[0156] The machine learning method for optimizing the manufacturing process of dry-type transformers includes: collecting process parameters and quality indicators of the dry-type transformer production process; calculating the correlation coefficient between process parameters and quality indicators, selecting process parameters strongly correlated with quality indicators as features, and using recursive feature elimination to rank the features; based on the collected process parameters, constructing a dry-type transformer production quality prediction model using the Support Vector Regression (SVR) algorithm, and integrating the Bagging learning method into the production quality control model; performing quality inspection on the produced dry-type transformers, collecting image data of the completed dry-type transformers, and using a machine vision model to inspect the image data to determine whether the dry-type transformers have quality defects; tracing the process of dry-type transformers with quality problems, adjusting the correlation coefficient between process parameters and quality indicators, and adjusting the parameters in the dry-type transformer production quality prediction model.

[0157] Furthermore, the electronic device may also include a user interface. Of course, the architecture of this invention is merely exemplary; in implementing different devices, one or more components of the electronic device disclosed in this invention may be omitted according to actual needs.

[0158] Example 4

[0159] The present invention also discloses a computer-readable storage medium.

[0160] Computer-readable instructions are stored on a computer-readable storage medium.

[0161] When computer-readable instructions are executed by a processor, a machine learning method for optimizing the manufacturing process of a dry-type transformer according to an embodiment of the present invention, as described with reference to the above figures, can be performed.

[0162] Storage media include, but are not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. Furthermore, according to embodiments of the present invention, the processes described in the above-mentioned flowcharts can be implemented as computer software programs.

[0163] For example, this invention provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be executed by a processor to perform instructions corresponding to the method steps provided by this invention. For example: collecting process parameters of the dry-type transformer production process and obtaining quality indicators of the dry-type transformer; calculating the correlation coefficient between process parameters and quality indicators, selecting process parameters strongly correlated with quality indicators as features, and using recursive feature elimination to rank the features; based on the collected process parameters, using the Support Vector Regression (SVR) algorithm to construct a dry-type transformer production quality prediction model, and integrating the Bagging learning method into the production quality control model; performing quality inspection on the produced dry-type transformers, collecting image data of the completed dry-type transformers, and using a machine vision model to inspect the image data to determine if the dry-type transformers have quality defects; tracing the process of dry-type transformers with quality problems, adjusting the correlation coefficient between process parameters and quality indicators, and adjusting the parameters in the dry-type transformer production quality prediction model.

[0164] When the computer program is executed by the central processing unit (CPU), it performs the functions defined in the method of the present invention. The method, apparatus, and device of the present invention may be implemented in many ways. For example, the method, apparatus, and device of the present invention may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware.

[0165] The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of the present invention are not limited to the order specifically described above, unless otherwise specifically stated.

[0166] Furthermore, in some embodiments, the invention may also be implemented as a program recorded on a recording medium, the program including machine-readable instructions for implementing the method according to the invention. Therefore, the invention also covers recording media storing programs for performing the method according to the invention.

[0167] In addition, the parts of the technical solutions provided in the embodiments of the present invention that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0168] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A machine learning method for optimizing the manufacturing process of dry-type transformers, characterized in that, include: Collect process parameters for the production of dry-type transformers and obtain quality indicators for dry-type transformers. The process parameters include: winding tension. , winding speed Soaking time immersion temperature Curing temperature Curing time The quality indicators include: insulation resistance. ,inductance Transformer ratio Core loss Copper loss and partial discharge ; Calculate the correlation coefficient between process parameters and quality indicators, select process parameters that are strongly correlated with quality indicators as features, and use recursive feature elimination to sort the features; Based on the collected process parameters, a support vector regression (SVR) algorithm is used to construct a production quality prediction model for dry-type transformers, and the Bagging learning method is integrated into the production quality control model. The ensemble Bagging learning method includes: random resampling generation On each training subset, an SVR sub-model is trained using the optimized SVR model parameters to obtain... Sub-models; evaluated on the test set. The performance of each SVR sub-model, using mean squared error. and mean absolute error As an evaluation metric, the calculation of each sub-model is performed. and The weight coefficients of the sub-model are determined based on its reciprocal; The prediction results of each SVR sub-model are weighted and averaged to obtain the final predicted value of the dry-type transformer quality index. The specific method for determining the weight coefficients of the sub-model based on its reciprocal is as follows: Weight coefficients of individual sub-models The calculation formula is: ； in, For the first Mean squared error of each sub-model For the first The mean absolute error of each sub-model; The dry-type transformers produced are subjected to quality inspection. Image data of the completed dry-type transformers are collected, and the image data is inspected using a machine vision model to determine whether there are quality defects in the dry-type transformers. For dry-type transformers with quality problems, trace the process source, adjust the correlation coefficients of process parameters and quality indicators, and adjust the parameters in the dry-type transformer production quality prediction model. The adjustment of parameters in the dry-type transformer production quality prediction model includes: for dry-type transformers that fail quality inspection, extracting process parameter data from their production process. and obtain the corresponding quality indicators. ; Process parameters of non-conforming samples and quality indicators Add to the original dataset In the process, the updated dataset is obtained. Based on the updated dataset Re-select features, recalculate the correlation coefficients between process parameters and quality indicators, and adjust the feature ranking; use the updated dataset. Retrain the SVR model and the Bagging ensemble model to obtain an updated quality prediction model; compare the prediction performance of the models before and after the update on the test set, using mean squared error. and mean absolute error If the mean square error of the updated model and mean absolute error If all values ​​are less than those in the unupdated model, the new model will be used to guide the production of dry-type transformers.

2. The machine learning method for optimizing the manufacturing process of dry-type transformers according to claim 1, characterized in that, The collected process parameters and quality index data are cleaned to remove outliers, and missing values ​​are filled in using interpolation. A dataset is constructed based on the preprocessed process parameters and quality index data. Let the entire dataset Represented as: In the dataset, For the first The feature vector of process parameters for each sample. ; For the first The quality index value corresponding to each sample , where n is the total number of samples.

3. The machine learning method for optimizing the manufacturing process of dry-type transformers according to claim 2, characterized in that, Calculate the Pearson correlation coefficients between each process parameter and the quality index, and select parameters with correlation coefficients greater than a preset threshold. The process parameters are used as candidate features; The Recursive Feature Emission (RFE) algorithm is used to rank candidate features by importance. The specific steps are as follows: Input candidate features Target feature number , , representing the number of features finally selected, and setting the base model. , used to assess feature importance; Initialize the feature ranking list to store the ranking results of the features; record the current feature set. The index value representing the features that were not excluded; Repeat the following steps until... : In the current feature subset On the training base model To obtain model parameters ; Calculate the importance score for each feature. Indicates the model feature weight coefficients; Find the feature index with the lowest importance ;Will from Remove from; Output feature ranking list According to the preset target feature number Select the ranking list The former These features are used as the final feature subset: ,in This indicates a subset of features with k selected features.

4. The machine learning method for optimizing the manufacturing process of dry-type transformers according to claim 3, characterized in that, A production quality prediction model for dry-type transformers was constructed using the Support Vector Regression (SVR) algorithm; process parameter features were selected based on the Recursive Feature Emission (RFE) algorithm. and quality indicator data , serving as the training set for the dry-type transformer production quality prediction model; the objective function of the SVR algorithm is: ； ； in, and For the parameters of the SVR model, The penalty coefficient is... and As slack variables, For error tolerance, For feature mapping function, It is the number of samples; The Gaussian kernel function RBF is selected as the feature mapping function of SVR, and the hyperparameters of SVR are optimized through grid search. and ; The expression for the Gaussian kernel function RBF is: ;in, For the hyperparameters of the kernel function, and These are the feature vectors of two samples; This represents the Gaussian kernel function, used to calculate the similarity between two samples in a high-dimensional feature space. The steps of a grid search are as follows: set up and Given the search range and step size, generate a grid; for each element in the grid... Combine and perform the following steps: using Used as a hyperparameter training model for SVR; model performance is evaluated through cross-validation; Choose the one with the smallest cross-validation evaluation value. As the optimal hyperparameter for SVR.

5. The machine learning method for optimizing the manufacturing process of dry-type transformers according to claim 4, characterized in that, Collect image data of the completed dry-type transformer, including an external view and an internal structure view of the dry-type transformer; The dry-type transformer appearance diagram includes an image of the transformer's exterior for detecting surface defects, including cracks and dents; the dry-type transformer internal structure diagram includes an X-ray image of the transformer's internal structure for detecting internal defects. The YOLOv5 object detection algorithm is used to locate defect regions in an image; the steps of the YOLOv5 algorithm are as follows: the input image is divided into... 1 grid; predict for each grid Each bounding box contains [number] bounding boxes. One predicted value: ;in, The coordinates of the bounding box center are The bounding box width and height, For each bounding box, predict the confidence level. Each category probability represents the probability that the bounding box belongs to each defect category; Based on the confidence level and category probability, some bounding boxes are filtered out to obtain the final defect detection result; The detected defect region image patch is input into a pre-trained ResNet-50 classification model to determine the defect type; the ResNet-50 classification model includes an input layer, a convolutional layer, a residual block, a fully connected layer, and an output layer; The input layer receives defect image patches. The convolutional layer, consisting of multiple convolutional and pooling layers, extracts hierarchical features from the image. The residual block, consisting of multiple concatenated residual blocks, each containing two convolutional layers and a short-circuit connection, is used to train the deep network. The fully connected layer flattens the features extracted by the convolutional layers and maps them to the probability distribution of defect categories. The output layer uses the Softmax activation function to output the probability of each defect category. Based on the defect type and severity, a quality assessment result is given, which includes pass and fail; let the defect type be... The severity is Then the quality assessment function is: ； in, For the set of allowed defect types, The severity threshold, This indicates that the quality assessment is satisfactory. This indicates that the quality assessment is unsatisfactory.

6. A machine learning system for optimizing the manufacturing process of dry-type transformers, used to implement the machine learning method for optimizing the manufacturing process of dry-type transformers as described in any one of claims 1 to 5, characterized in that, include: The system includes a data acquisition module, a feature calculation module, a production quality control module, a quality inspection module, and a process traceability adjustment module. The data acquisition module is used to collect process parameters of the dry-type transformer production process and to obtain the quality indicators of the dry-type transformer. The feature calculation module is used to calculate the correlation coefficient between process parameters and quality indicators, select process parameters that are strongly correlated with quality indicators as features, and sort the features using recursive feature elimination. The production quality control module, based on the collected process parameters, uses the Support Vector Regression (SVR) algorithm to construct a production quality prediction model for dry-type transformers, and integrates the Bagging learning method into the production quality control model. The quality inspection module is used to inspect the quality of the produced dry-type transformers, collect image data of the completed dry-type transformers, and use a machine vision model to inspect the image data to determine whether the dry-type transformers have quality defects. The process traceability adjustment module is used to trace the process of dry-type transformers with quality problems, adjust the correlation coefficients of process parameters and quality indicators, and adjust the parameters in the dry-type transformer production quality prediction model.

7. An electronic device, characterized in that, include: A processor and a memory, wherein the memory stores a computer program that can be called by the processor; the processor executes the machine learning method for optimizing the dry-type transformer manufacturing process according to any one of claims 1 to 5 by calling the computer program stored in the memory.

8. A computer-readable storage medium, characterized in that: The system stores instructions that, when executed on a computer, cause the computer to perform a machine learning method for optimizing the dry-type transformer manufacturing process as described in any one of claims 1 to 5.

Images