阅读说明：本技术 一种基于机器学习的定子线棒绝缘老化状态复合预测方法 (Machine learning-based stator bar insulation aging state composite prediction method ) 是由 钱翰宁 孙兆龙 黄垂兵 刘振田 庄哲鑫 周炜昶 卯寅浩 郑伟 丁安敏 于 2021-09-27 设计创作，主要内容包括：本发明公开了一种基于机器学习的定子线棒绝缘老化状态复合预测方法,包括：图像数据采集：包括采集日常预测数据及采集节点预测数据；采集日常预测数据只采集定子线棒的宏观老化图像数据；采集节点预测数据只采集定子线棒微观老化图像数据；图像数据预处理；预测情况分类；宏观及微观老化图像数据特征提取；数据分类及通过训练集构建模型：通过验证集,得到训练好的BP神经网络模型,通过测试集验证最终的准确率后,得到最终可用于预测老化状态的模型,具有可快速、准确得到预测结果的优点。(The invention discloses a machine learning-based stator bar insulation aging state composite prediction method, which comprises the following steps: image data acquisition: acquiring daily prediction data and acquiring node prediction data; acquiring daily prediction data and only acquiring macroscopic aging image data of the stator bar; collecting node prediction data and only collecting stator bar microscopic aging image data; preprocessing image data; classifying the prediction situation; extracting macro and micro aging image data features; classifying data and constructing a model through a training set: and obtaining a trained BP neural network model through the verification set, obtaining a model which can be finally used for predicting the aging state after verifying the final accuracy through the test set, and having the advantage of quickly and accurately obtaining a prediction result.)

1. A stator bar insulation aging state composite prediction method based on machine learning is characterized by comprising the following steps:

step 1: image data acquisition: acquiring daily prediction data and acquiring node prediction data; collecting daily prediction data and only collecting stator bar macroscopic aging image data; collecting node prediction data and only collecting stator bar microscopic aging image data;

step 2, preprocessing image data: preprocessing the image data acquired in the step 1;

step 3, predicting condition classification: for the daily prediction situation, entering the step 4; for the node prediction situation, entering step 7;

and 4, extracting the data features of the macroscopic aging image: performing feature extraction on the stator bar macroscopic ageing image data preprocessed in the step 2 to obtain feature data of a macroscopic ageing image;

step 5, data classification and model construction through a training set: dividing the characteristic data obtained in the step 4 into a training set, a verification set and a test set, inputting the characteristics of the training set into a BP neural network, and constructing a BP neural network model;

step 6, obtaining a trained BP neural network model through a verification set, and obtaining a model which can be finally used for predicting an aging state after verifying the final accuracy rate through a test set;

and 7, extracting the data characteristics of the microscopic aging image: performing characteristic extraction on the stator bar microscopic aging image data preprocessed in the step 2 to obtain characteristic data of the microscopic aging image;

step 8, data classification and model construction through a training set: dividing the characteristic data obtained in the step 7 into a training set, a verification set and a test set, inputting the characteristics of the training set into a BP neural network, and constructing a BP neural network model;

and 9, obtaining a trained BP neural network model through the verification set, and obtaining a model which can be finally used for predicting the aging state after the final accuracy is verified through the test set.

2. The machine learning based stator bar insulation aging state composite prediction method according to claim 1, characterized in that in the step 1, the method for collecting stator bar macroscopic aging image data is as follows: a camera device is arranged in the motor, the surface of the stator bar is photographed at fixed time intervals through the camera device, image data in different aging stages are obtained, and the image data are transmitted to an external data storage device.

3. The method for compositely predicting the insulation aging state of the stator bar based on machine learning as claimed in claim 1, wherein in the step 1, the method for collecting the microscopic aging image data of the stator bar is as follows: stripping the insulation sample from the new stator bar to obtain single-layer or multi-layer stacked insulation sheets; cutting the insulating sheet into a plurality of sheet samples, placing the sheet samples in a ventilated vessel, carrying out thermo-oxidative aging, and equally dividing the sheet samples into a plurality of parts according to the number; taking out a sheet sample at intervals until all the sheet samples are taken out, obtaining aging samples at different aging stages, observing diffraction spots of the aging samples at different aging stages by using two-dimensional small-angle X-ray scattering equipment, and obtaining diffraction spot maps of the aging samples; and observing the surface morphology of the aged sample in different aging states on a nanometer scale and the interface state between different components by using an atomic force microscope to obtain the micro-topography map of the aged sample.

4. The machine learning based stator bar insulation aging state composite prediction method of claim 1, characterized in that the image data preprocessing method in the step 2 is picture enhancement or morphological erosion and expansion processing.

5. The method for compositely predicting the insulation aging state of the stator bar based on machine learning as claimed in claim 1, wherein in the step 4, the method for performing feature extraction on the preprocessed macroscopic aging image data of the stator bar comprises the following steps: and extracting by adopting a feature extraction network, wherein the feature extraction network is formed by removing the conv5 module and all the layers after the conv5 module from the residual error network of 50 layers.

6. The method for compositely predicting the insulation aging state of the stator bar based on machine learning as claimed in claim 1, wherein the method for performing feature extraction on the stator bar microscopic aging image data preprocessed in the step 2 in the step 7 is as follows: and extracting directional gradient histogram characteristics of a diffraction spot map of the preprocessed stator bar microscopic aging image data and local binary pattern characteristics of a microscopic morphology map.

7. The machine-learning based stator bar insulation aging state composite prediction method of claim 6, wherein the extraction process of the histogram of oriented gradient features comprises: and (3) segmenting the diffraction spot spectrum of the aged sample, calculating the direction and gradient of pixels in each segmentation block, accumulating the gradients according to the direction of pixel change to obtain a directional gradient map in one segmentation block, and counting the directional gradients in all the segmentation blocks to obtain the directional gradient histogram characteristics of the image.

8. The machine learning based stator bar insulation aging state composite prediction method of claim 6, wherein the local binary pattern feature extraction process comprises: the original local binary pattern operator is defined as that in a 3 x 3 window, the central pixel of the window is used as a threshold value, the gray values of 8 adjacent pixels are compared with the central pixel, if the values of the surrounding pixels are greater than the value of the central pixel, the position of the pixel is marked as 1, otherwise, the position of the pixel is marked as 0, 8 points in the 3 x 3 neighborhood are compared to generate 8-bit binary numbers, and the 8-bit binary numbers are converted into decimal numbers to obtain the local binary pattern characteristic of the central pixel of the window.

9. The machine learning based stator bar insulation aging state composite prediction method of claim 1, characterized in that the step 6 or step 9 specific method is: by setting different model parameters of the BP neural network, finding out the model parameter with the highest accuracy rate for predicting the aging state of the stator bar in a verification set, and taking the model parameter as the final model parameter to obtain a trained BP neural network model; inputting the test set into the trained BP neural network model, predicting the aging degree of the images in the test set, and predicting the aging state of the stator bar by using the model after verifying the accuracy of the test set.

Technical Field

The invention relates to the field of prediction of aging degree of stator bar materials, in particular to a stator bar insulation state prediction method based on deep learning.

Background

The insulation of the stator winding of the generator is an important component of the generator, and the influence of environmental stress on the insulation is increased along with the continuous improvement of the installed capacity of a large motor. The life of the generator is mainly determined by the insulation life, and the insulation can be affected by various factors to cause failure.

For example, in the design and manufacture of a motor, insulation failure can be caused by no strengthening treatment at a position where electric field concentration is easy to occur, or insulation of a stator bar is impacted by operation overvoltage due to frequent starting, so that the insulation aging of the motor is accelerated, or the insulation aging is caused by heating of components due to the loss of the insulation.

The service life of the generator is reduced due to the aging of the insulation of the stator bar, so that the research on the service life prediction of the insulation of the generator is of great significance to the improvement of the operation stability of the generator. The service life prediction is divided into early stage prediction and later stage prediction, wherein the early stage prediction refers to predicting the service life of insulation at the initial stage of design according to the design structure of the generator and by combining the insulation performance of an insulation material; the later prediction is the prediction of the residual life of the common generator insulation, in the operation process, the insulation performance characterization data is obtained, the residual service life of the insulation is predicted, and the generator fault caused by insulation failure is prevented in advance.

The traditional method for predicting the insulation aging degree of the stator bar is mainly a formula method, and mainly comprises a partial discharge parameter prediction method, a D-image method and other non-electrical parameter life evaluation methods by establishing the relationship between various electrical parameters such as dielectric loss factors, partial discharge amount and the like and the residual life. The methods have great difference with the evaluation result and are not accurate enough. At present, some methods for obtaining the microscopic pattern of the insulating sheet on the winding bar to perform characteristic value extraction are used for state prediction, but the method is complex in operation and data sample obtaining, needs to disassemble a motor, is difficult to obtain a prediction result quickly, and is not suitable for daily prediction.

Disclosure of Invention

Aiming at the defects or improvement requirements of the prior art, the invention aims to solve the problems that the traditional stator bar insulation aging degree prediction method is inaccurate, complex to operate, difficult to obtain a prediction result quickly and not suitable for daily prediction.

The present invention provides: a stator bar insulation aging state composite prediction method based on machine learning comprises the following steps:

step 1: image data acquisition: acquiring daily prediction data and acquiring node prediction data; collecting daily prediction data and only collecting stator bar macroscopic aging image data; collecting node prediction data, namely only collecting stator bar microscopic aging image data;

step 2, preprocessing image data: preprocessing the image data acquired in the step 1;

step 3, predicting condition classification: for the daily prediction situation, entering the step 4; for the node prediction situation, entering step 7;

and 4, extracting the data features of the macroscopic aging image: performing feature extraction on the stator bar macroscopic ageing image data preprocessed in the step 2 to obtain feature data of a macroscopic ageing image;

step 5, data classification and model construction through a training set: dividing the characteristic data obtained in the step 4 into a training set, a verification set and a test set, inputting the characteristics of the training set into a BP neural network, and constructing a BP neural network model;

step 6, obtaining a trained BP neural network model through a verification set, and obtaining a model which can be finally used for predicting an aging state after verifying the final accuracy rate through a test set;

and 7, extracting the data characteristics of the microscopic aging image: performing characteristic extraction on the stator bar microscopic aging image data preprocessed in the step 2 to obtain characteristic data of the microscopic aging image;

step 8, data classification and model construction through a training set: dividing the characteristic data obtained in the step 7 into a training set, a verification set and a test set, inputting the characteristics of the training set into a BP neural network, and constructing a BP neural network model;

and 9, obtaining a trained BP neural network model through the verification set, and obtaining a model which can be finally used for predicting the aging state after the final accuracy is verified through the test set.

Further, in the step 1, the method for acquiring the macroscopic aging image data of the stator bar comprises the following steps: a camera device is arranged in the motor, the surface of the stator bar is photographed at fixed time intervals through the camera device, image data in different aging stages are obtained, and the image data are transmitted to an external data storage device.

Further, in the step 1, the method for acquiring the microscopic aging image data of the stator bar comprises the following steps: stripping the insulation sample from the new stator bar to obtain single-layer or multi-layer stacked insulation sheets; cutting the insulating sheet into a plurality of sheet samples, placing the sheet samples in a ventilated vessel, carrying out thermo-oxidative aging, and equally dividing the sheet samples into a plurality of parts according to the number; taking out a sheet sample at intervals until all the sheet samples are taken out, obtaining aging samples at different aging stages, observing diffraction spots of the aging samples at different aging stages by using two-dimensional small-angle X-ray scattering equipment, and obtaining diffraction spot maps of the aging samples; and observing the surface morphology of the aged sample in different aging states on a nanometer scale and the interface state between different components by using an atomic force microscope to obtain the micro-topography map of the aged sample.

Further, the image data in step 2 is preprocessed by image enhancement or morphological erosion and expansion.

Further, in the step 4, the method for extracting the features of the preprocessed stator bar macroscopic aging image data includes: and extracting by adopting a feature extraction network, wherein the feature extraction network is formed by removing the conv5 module and all the layers after the conv5 module from the residual error network of 50 layers.

Further, the method for extracting the characteristics of the stator bar microscopic aging image data preprocessed in the step 2 in the step 7 comprises the following steps: and extracting directional gradient histogram characteristics of a diffraction spot map of the preprocessed stator bar microscopic aging image data and local binary pattern characteristics of a microscopic morphology map.

Further, the process of extracting the histogram of directional gradients includes: and (3) segmenting the diffraction spot spectrum of the aged sample, calculating the direction and gradient of pixels in each segmentation block, accumulating the gradients according to the direction of pixel change to obtain a directional gradient map in one segmentation block, and counting the directional gradients in all the segmentation blocks to obtain the directional gradient histogram characteristics of the image.

Further, the local binary pattern feature extraction process includes: the original local binary pattern operator is defined as that in a 3 x 3 window, the central pixel of the window is used as a threshold value, the gray values of 8 adjacent pixels are compared with the central pixel, if the values of the surrounding pixels are greater than the value of the central pixel, the position of the pixel is marked as 1, otherwise, the position of the pixel is marked as 0, 8 points in the 3 x 3 neighborhood are compared to generate 8-bit binary numbers, and the 8-bit binary numbers are converted into decimal numbers to obtain the local binary pattern characteristic of the central pixel of the window.

Further, the specific method in step 6 or step 9 is as follows: by setting different model parameters of the BP neural network, finding out the model parameter with the highest accuracy rate for predicting the aging state of the stator bar in a verification set, and taking the model parameter as the final model parameter to obtain a trained BP neural network model; inputting the test set into the trained BP neural network model, predicting the aging degree of the images in the test set, and predicting the aging state of the stator bar by using the model after verifying the accuracy of the test set.

In general, by the above technical solution of the present invention, compared with the prior art, the following beneficial effects can be obtained:

according to the invention, the stator bar macroscopic aging image data and the microscopic aging image data are respectively collected, and the daily prediction model and the node prediction model are respectively established, so that the daily rapid prediction and the further accurate prediction of major nodes (the major nodes are major overhaul periods or other time periods needing to be checked) can be satisfied, and the defects of inaccuracy, complex operation and difficulty in rapidly obtaining prediction results of a conventional prediction mode are avoided through the composite cross use of the daily prediction model and the node prediction model.

Drawings

FIG. 1 is a flow chart of the preferred embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

Referring to fig. 1, an embodiment of the present invention provides a stator bar insulation aging state composite prediction method based on machine learning, including the following steps:

step 1: image data acquisition: acquiring daily prediction data and acquiring node prediction data; collecting daily prediction data and only collecting stator bar macroscopic aging image data; collecting node prediction data, namely only collecting stator bar microscopic aging image data;

specifically, the method for acquiring the macroscopic aging image data of the stator bar comprises the following steps: a camera device is arranged in the motor, the surface of the stator bar is photographed at fixed time intervals through the built-in camera device, image data in different aging stages (aging images usually show the characteristics of cracking of the surface of the stator bar, color change, powder appearance and the like) are obtained, and the image data are transmitted to an external data storage device.

The method for acquiring the microscopic aging image data of the stator bar comprises the following steps: stripping the insulation sample from the new stator bar (i.e. the new stator bar not yet used) to obtain a single-layer or multi-layer stack of insulation sheets; cutting an insulating sheet into a plurality of sheet samples, placing the sheet samples in a ventilated vessel (the ventilation is used for reducing the influence of oxygen in the air on thermal aging and simulating the actual environment of a stator bar), thermally oxidizing the samples, and equally dividing the samples into a plurality of parts according to the number; taking out a sheet sample at intervals until all the sheet samples are taken out, obtaining aging samples at different aging stages, observing diffraction spots of the aging samples at different aging stages by using two-dimensional small-angle X-ray scattering equipment, and obtaining diffraction spot maps of the aging samples; and observing the surface morphology of the aged sample in different aging states on a nanometer scale and the interface state between different components by using an atomic force microscope to obtain the micro-topography map of the aged sample.

Step 2, preprocessing image data: the image data acquired in step 1 is preprocessed, for example, by image enhancement or morphological erosion and dilation.

Step 3, predicting condition classification: for the daily prediction situation, entering the step 4; for the node prediction situation, entering step 7;

and 4, extracting the data features of the macroscopic aging image: and (3) performing feature extraction on the stator bar macroscopic aging image data preprocessed in the step (2), wherein the extraction method comprises the following steps: extracting by adopting a feature extraction network, wherein the feature extraction network is formed by removing a conv5 module and all the layers behind the conv5 module from a residual error network of 50 layers; obtaining characteristic data of the macroscopic aging image;

step 5, data classification and model construction through a training set: dividing the characteristic data obtained in the step 4 into a training set, a verification set and a test set, inputting the characteristics of the training set into a BP neural network, and constructing a BP neural network model;

step 6, obtaining a trained BP neural network model through a verification set, and obtaining a model which can be finally used for predicting an aging state after verifying the final accuracy rate through a test set; the specific method comprises the following steps: by setting different model parameters of the BP neural network, finding out the model parameter with the highest accuracy rate for predicting the aging state of the stator bar in a verification set, and taking the model parameter as the final model parameter to obtain a trained BP neural network model; inputting the test set into the trained BP neural network model, predicting the aging degree of the images in the test set, and predicting the aging state of the stator bar by using the model after verifying the accuracy of the test set.

And 7, extracting the data characteristics of the microscopic aging image: performing characteristic extraction on the stator bar microscopic aging image data preprocessed in the step 2 to obtain characteristic data of the microscopic aging image;

the specific method for extracting the characteristics of the stator bar microscopic aging image data preprocessed in the step 2 comprises the following steps: and extracting directional gradient histogram characteristics of a diffraction spot map of the preprocessed stator bar microscopic aging image data and local binary pattern characteristics of a microscopic morphology map.

The extraction process of the histogram features of the directional gradients comprises the following steps: and (3) segmenting the diffraction spot spectrum of the aged sample, calculating the direction and gradient of pixels in each segmentation block, accumulating the gradients according to the direction of pixel change to obtain a directional gradient map in one segmentation block, and counting the directional gradients in all the segmentation blocks to obtain the directional gradient histogram characteristics of the image.

The local binary pattern feature extraction process comprises the following steps: the original local binary pattern operator is defined as that in a 3 x 3 window, the central pixel of the window is used as a threshold value, the gray values of 8 adjacent pixels are compared with the central pixel, if the values of the surrounding pixels are greater than the value of the central pixel, the position of the pixel is marked as 1, otherwise, the position of the pixel is marked as 0, 8 points in the 3 x 3 neighborhood are compared to generate 8-bit binary numbers, and the 8-bit binary numbers are converted into decimal numbers to obtain the local binary pattern characteristic of the central pixel of the window.

Step 8, data classification and model construction through a training set: dividing the characteristic data obtained in the step 7 into a training set, a verification set and a test set, inputting the characteristics of the training set into a BP neural network, and constructing a BP neural network model;

and 9, obtaining a trained BP neural network model through the verification set, and obtaining a model which can be finally used for predicting the aging state after the final accuracy is verified through the test set: the specific method comprises the following steps: by setting different model parameters of the BP neural network, finding out the model parameter with the highest accuracy rate for predicting the aging state of the stator bar in a verification set, and taking the model parameter as the final model parameter to obtain a trained BP neural network model; inputting the test set into the trained BP neural network model, predicting the aging degree of the images in the test set, and after verifying the accuracy of the test set, predicting the aging state (such as residual life) of the stator bar by using the model.

In the embodiment, the BP neural network is adopted as a prediction algorithm model, and the BP neural network is mature in both network theory and performance. Its outstanding advantage has:

1) non-linear mapping capability: the BP neural network essentially realizes a mapping function from input to output, and mathematical theory proves that the neural network with three layers can approximate any nonlinear continuous function with any precision. This makes it particularly suitable for solving problems with complex internal mechanisms, i.e. the BP neural network has strong nonlinear mapping capability.

2) Self-learning and self-adaptive capacity: when the BP neural network is trained, reasonable rules between output and output data can be automatically extracted through learning, and learning contents are self-adaptively memorized in the weight of the BP neural network. I.e., the B neural network has a high degree of self-learning and self-adaptation capabilities.

3) Generalization ability: the generalization capability refers to that when designing a pattern classifier, the network is considered to ensure that objects to be classified are correctly classified, and whether the network can correctly classify unseen patterns or noise-polluted patterns after training is also considered. I.e., the neural network, has the ability to apply learning outcomes to new knowledge.

4) Fault tolerance capability: the BP neural network does not have a great influence on the global training result after local or partial neurons of the BP neural network are damaged, that is, the system can still work normally even if the system is locally damaged. Namely, the BP neural network has certain fault-tolerant capability.

In the actual prediction work, other prediction algorithm models, such as random forests, logistic regression, k-means clustering and other algorithms, can also be adopted.

In addition, in the actual prediction work, daily prediction and node prediction can be combined and verified mutually, so that the prediction accuracy is improved;

it will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.