| At present,the smart grid puts forward higher requirements for the informatization and automation of power system equipment.Timely and accurately locating and handling faults is an important means to improve the reliability of system power supply.Due to the thermal effect of current or voltage,power equipment is often accompanied by temperature rise when it fails.Infrared thermal imager is the mainstream method for detecting and diagnosing equipment thermal faults,but this method has certain limitations.On the one hand,infrared cameras generally have high prices and low pixels,which makes the spatial positioning of densely installed equipment very poor.On the other hand,infrared equipment has a certain threshold for operation.The emissivity needs to be adjusted according to the measured object.Visible cameras have the advantages of high pixels and good imaging quality,but visible radiation temperature measurement is only suitable for high temperature areas(above 800℃).In order to meet the needs of low-cost and intelligent temperature measurement,this paper proposes a new measurement method for the low temperature field(room temperature to 100℃)of metal materials in power equipment based on image processing and machine learning technology.This method uses only relatively inexpensive visible cameras,and achieves quite high prediction accuracy.The results achieved are as follows:We designed and established a set of experimental platform.A total of 35,000 photos were taken on a metal plate made of brass with a relatively cheap smartphone camera under different temperatures and ambient lighting.All photos form a visible image library suitable for machine learning.The gray frequency distribution of the three primary colors(768 dimensions)in the image was extracted as the input feature vector of the model,and five classical machine learning methods were used for modeling respectively.The average prediction error of the model for training set is 1.02℃.The maximum error for test set is 11.65℃,and the average is 6.77℃.In order to solve the over-fitting,an optimization method using principal component analysis to reduce the dimensions of the original feature vector is proposed.The results show that the model performs best when the features are reduced to 74 dimensions.The average prediction error of the test set is reduced to 3.02℃,and the overfitting value of the model drops from the highest 11.24 to 1.84.When we continue to compress the data dimensions to 31 and16 dimensions,the model complexity decreases excessively,leading to an under-fitting trend.Compared with the 74 dimensions,the prediction error of the test set does not decrease but rises.The absolute error variance increased significantly,indicating that the 74-dimensional feature is ideal and appropriate.Aiming at the problem of increasing error caused by the interference of different ambient lighting on the color information,four optimization methods are proposed.Three of them have positive effects.The addition of the illuminance component to the feature and the global Retinex image enhancement in the RGB space can only reduce the error by 0.1-0.2°C.The image enhancement method based on discrete wavelet transform and I-Retinex can minimize the impact of ambient light disturbance.Under the k-nearest neighbor regression algorithm,the error is reduced to the lowest 0.92℃.The average error of the three algorithms is only 1.36℃,which is 38.4% lower than before optimization.Based on the actual work requirements and the characteristics of the artificial intelligence temperature measurement system,this paper also studied and wrote the "On-site operation specification for live equipment based on thermoreflectance intelligent temperature measurement technology".The specifications include personnel safety,equipment environmental requirements,operation methods,system maintenance,post-inspection and verification,etc.,which provide guidance for the practical application of this new type of temperature measurement technology. |
PDF Full Text Request