Intelligent Temperature Rise And Partial Discharge Monitoring Method Of Electrical Equipment Based On Visible Images

Electrical equipment and transmission lines need to be monitored for abnormal temperature rise and partial discharges.At present,non-contact monitoring methods by images are mainly based on infrared and ultraviolet radiation signals in engineering.Visible light signals and images are only used to assist positioning.The temperature and partial discharge information contained in them have not been fully utilized.On the other hand,it is difficult for traditional modeling methods to establish an effective mapping function due to the nonlinear,non-monotonic and multivariate relationship between visible image features and target state quantities.Accordingly,a new data-driven intelligent monitoring method is proposed and implemented for temperature rise and partial discharge status of electrical equipment by visible images and machine learning technology.The work and achievements are as follows.Aiming at the problem that the visible image of electrical equipment at normal temperature is affected by the change of varying sunlight,which is non-normal incident with multiple incident angles and multiple wavelengths,an intelligent temperature measurement method of metal surfaces by thermal-modulated reflected light is established,which takes the RGB gray level histograms(RGB-GLH)of the visible light image as the normalized comprehensive state quantity.The research compares the accuracy of different machine learning algorithms,and determines the optimal features.The temperature measurement error under laboratory conditions is less than 1.0 ℃.An intelligent data-driven temperature measurement mechanism based on thermal-modulated reflected light is analyzed and proposed.The robustness of the measurement method is analyzed based on it.According to the actual needs of on-site temperature rise fault diagnosis,a temperature difference monitoring method based on images’ differential chromatic features is proposed,which can significantly reduce the interference effect of ambient light changes.This temperature measurement method has been practically applied in the engineering environment.The mean absolute error(MAE)of the temperature and temperature difference monitoring of actual metal devices under sunlight reaches 0.3-0.9 ℃and 0.1-0.5 ℃,respectively.The MAE of the temperature difference monitoring of the busbar clip on the substation site is 0.2-0.7 ℃.To solve the proprietary problems of different on-site environmental monitoring models and the calibration problems of on-site monitoring models,a work plan of updating models on top of base models is designed according to the characteristics of data-driven intelligent modeling.The results show that the MAE of the updated model for testing new samples is 0.9 °C.In view of the multi-spectral lines and multi-time scales of optical radiation in discharges,as well as the complex state quantity problems caused by the spatial non-uniform distribution of streamers and the randomness of discharges,an intelligent monitoring method of discharges is established,which uses the RGB-GLH of visible images as the normalized comprehensive state quantity to automatically divide and identify the discharge states of the surface and corona discharges.The research determines the improved algorithm model and optimized features.The results of state division of discharge image samples using unsupervised learning clustering algorithm have a good correspondence with traditional indicators(current pulse,discharge spectrum).The recognition models of discharge states are trained using supervised learning algorithms.The results show that the recognition accuracy of models based on chromatic features is higher than that of luminance and morphological features.The autoencoder neural network with improved structure and the weighted loss function as the optimization goal can make a biased identification of samples with higher severities of surface discharges.The recognition accuracy for samples of severe and dangerous stages reaches 0.982.Low-dimensional composite features are proposed to characterize the corona discharge state based on the unique "linear" discharge channel spatial structure of corona discharge images.It reduces the amount of data while significantly improving the discharge state recognition accuracy(0.987).The RGB-GLH,as well as the optimal features based on it,proposed in this paper provide a new normalized comprehensive diagnostic index for the diagnosis of complex physical processes with nonlinear,non-monotonicity and multivariate relationships.By choosing the method of regression prediction and classification identification machine learning algorithms,the equipment status and monitoring indicators can be accurately correlated,which is more conducive to the automation and intelligence of status monitoring,and improves the inspection efficiency and equipment cost performance.The above work provides new research and application scenarios for the combination of artificial intelligence modeling and traditional modeling to measure or diagnose physical quantities.It will also play an important role in integrating and sharing visual data on a unified information platform to promote the construction and development of smart grids.