Research On Meter Identification Algorithm Of Traction Substation In Low Light Mobile Environment

In order to ensure the reliability and economy of electrified railway operation,the national railway vigorously promotes the unattended system of traction power supply,replacing the manual inspection with high labor intensity,scattered detection quality,and many subjective factors by intelligent inspection of robots to accelerate the traction substation Unattended process.As the "main force" of routine inspections in substations,inspection robots can perform meter identification tasks with the help of machine vision technology.However,the systematic and mechanical errors of mobile robots often make the target meter not in the center of the image,so it is difficult to locate the dial in complex environments.At the same time,the inspection environment such as haze,rainy days and night also makes the detected dial have some low light images with insufficient light,which seriously affects the accuracy of subsequent meter identification.In addition,special meters such as the lightning arrester leakage current meter,main transformer oil temperature meter,main transformer oil pillow oil level meter and other special meters existing in the substation need to design appropriate meter identification algorithms.Therefore,the main work of this paper is as follows:First of all,due to the low detection efficiency,weak generalization ability and the need to input clear images to be detected by traditional methods,this paper proposes an improved YOLOv3 deep learning detection model to complete the detection of the dial area of the substation.This improved algorithm uses Soft-NMS to replace YOLOv3 original NMS to be more suitable for dense and target occlusion application scenarios.The output of the multi-scale detection network is Gaussion modeled and the standard deviation of the model is used to estimate the position confidence of the detection frame,so that the detection frame is closer to the real object.then the YOLOv3 model was improved by combining the pre-training data set and the self-built substation meteration data set to adapt to the actual inspection environment of the traction substation.Compared with the traditional method,theimproved YOLOv3 model in this paper does not require image preprocessing and has strong anti-interference.It can detect and extract the target meter panel area in a low-light mobile environment,which greatly reduces the calculation complexity of subsequent meter identification algorithms.Then,for the part of the low-light dial image detected,this paper uses fractional-order differential to reconstruct the lighting prior and reflection prior of the low-light image,and uses the camera response function closer to the real camera exposure to adjust the brightness information of the illuminated image,obtain a low-light enhancement algorithm based on fractional-order improved Retinex.In order to extract more dark content in the low-light area of the image,this paper chooses to perform secondary enhancement processing and use Fast ABF algorithm to filter out environmental noise.Based on this,a low-light enhancement fusion algorithm is proposed,which is based on the original image,the first low light enhancement image,the Fast ABF enhanced image,and the second low-light enhanced image are respectively subjected to Laplacian pyramid decomposition and multiplied by the corresponding Gaussian weight matrix to fuse to obtain the final enhanced image.The experiment proves that the low-light enhancement fusion algorithm can better save the image details while extracting more dark content in the low-light area of the image,and achieve a good balance between enhancing brightness and avoiding overexposure,which provides a clear table image for the subsequent meter recognition algorithm.Finally,this paper combined perspective transformation and SIFT feature point registration method to correct partially deformed dial images.Aiming at the three main meters outside the substation,this paper designs the lightning arrester leakage ammeter recognition algorithm based on dynamic threshold segmentation,the main transformer oil temperature meter recognition algorithm based on Hough rectangle detection and the main transformer oil pillow oil level gauge recognition algorithm based on FCM and Otsu combined segmentation.In addition,this article compiles and debugs the above three types of meter identification algorithms in the environment of Microsoft Visual Studio 2015 and Open CV3.4.3,and designs the meter identification software of the traction substation.Experiments show that thereadings obtained by the three meter recognition algorithms designed in this paper are basically the same as the values read manually,and the average recognition accuracy has reached the requirements of practical applications.