On Pointer Meter Reading Algorithm Based On Deep Learning

There are numerous mechanical pointer meters being used in substations,whose readings need manual collection during regular operation and maintenance.However,this manual data collection is of low accuracy,high labor intensity,and the process being significantly impacted by external factors.Therefore,automatic reading of pointer meters is of great significance to the intelligent operation and maintenance of substations.Detection and reading recognition of pointer meters are studied in this thesis.1.YOLOX algorithm is used to detect pointer meters of multiple types in substations in order to get images of simple gauge.This algorithm forms a high-accuracy,fast-speed,wide-application detector.Experimental results show that the algorithm has an accuracy rate of 99.6% in AP50 with an operating speed of 7.12 milliseconds per image on NVIDIA Ge Force RTX 3090,thus meeting the requirements in accuracy and real-time performance.2.A simple recognition method for single pointer meter based on deep regression is proposed in this thesis to directly obtain the reading of pointer meters located by detection algorithm.This method normalizes meter readings to make data distribution consistent.The backbone network of the model proposed here consists of several residual blocks and the head of the model uses convolutional layers instead of fully connected layers,thus reducing the amount of parameters and improving the inference speed.Experiments show that the algorithm has an accuracy rate of 93.39% under ±5% reading errors,and the accuracy can reach 99.17% under ±10% reading errors.At the same time,it is suitable for pointer meters of multiple types with poor image quality.3.A one-stage,end-to-end,multi-meter reading recognition method based on deep regression is proposed which can locate pointer meters and obtain the readings at the same time just through a simple one-round forward inference.The architecture of the network consists of a backbone for feature extraction and three output branches which are designed respectively to simultaneously locate pointer meters,obtain meter readings and complete over-range classification.Additional masks and indexes are set to make the whole network a fully convolutional neural network,which improves the inference speed greatly.Classification output branch is established to make the training process easier,thus improving the accuracy of the reading regression.Stochastic data augmentation meets different data requirements of the detection and reading recognition tasks.Use of deformable convolution modules enhances the model’s ability of extracting features and transformation of the loss function improves the accuracy of reading recognition.Experimental results show that the algorithm has good performance in detection and reading recognition for pointer meters of different types and styles.The accuracy of key points detection is over 99% when the recall rate is 99.26%.the recognition accuracy is over 92% under ±5% reading error,and the classification accuracy reaches more than 99%.At the same time,the average inference time on a NVIDIA GTX1080 GPU is only 40 ms.