| Electrical equipments are important components of substation,power plant,so the daily monitoring of electrical equipment is to promote the development of power system,to ensure its stable operation of the prerequisite.However,the traditional manual patrol has high cost and poor real-time performance,especially in extreme environments such as high temperature and high radiation,so the manual reading may be dangerous.Therefore,there is an urgent need to read the instrument by means of intelligent patrol.This article consider electrical equipment instrumentation pointer and digital meter reading recognition method as the research content,in view of the present reading recognition algorithm of low accuracy,large amount of calculation,multi-type instrument poor compatibility problems are studied,put forward the deep-learning based recognition method for meters of electrical equipment,and through the experiment of verified the effectiveness of the proposed algorithm,this article main research content is as follows:(1)Aiming at the positioning and classification of multi-type instrument targets in natural scenes,a lightweight instrument target detection model(Slim-SE-YOLO)based on attention mechanism is proposed.Combined with the characteristics of multi-scale feature information and the domain knowledge of instrument data,the model pruned the backbone network and detection scale of YOLO-v3,effectively reduced the model parameters and realized the lightweight goal of the model.In addition,the global information of the characteristic spectrum is obtained by the weight of attention,which makes up for the loss of precision caused by pruning.Experimental results show that compared with YOLO-v3,the number of model parameters after improvement decreases by 85% without reducing the detection accuracy.(2)In view of the difficulty of model detection caused by the small volume and dense distribution of instrument details,a pointer instrument reading algorithm based on the detection of key points is proposed.Firstly,a Keypoint detection model(Keypoint-FCN)based on thermal map similarity was established.Utilizing the pixel-level classification ability of the full convolutional neural network and combining with the optimization form of thermal map similarity realized recognition.Then,aiming at the problem of the instrument image tilt and distortion,by determining the key feature points of the instrument,multiple tasks such as image correction,pointer detection and dial projection are completed in parallel.Finally,the reading calculation of the instrument is completed.Taking manual reading as reference value,the accuracy of the proposed algorithm is above 98.3% through many random experiments.(3)In view of the problem that the precision of traditional character recognition method is affected by the character segmentation effect in the process of digital instrument reading,a digital instrument reading algorithm based on depth character recognition model is proposed.Firstly,a context-focused CRNN model(ATT-CRNN)was established based on the cascading structure of CNN and RNN combined with the context vector.Through the end-to-end conversion of image features to serialized features,the steps of pre-segmentation of characters to be recognized were avoided.Secondly,the algorithm uses color feature extraction and image closing operation to complete image segmentation and preprocessing,aiming at the problems of character discontinuity and character spacing caused by digital tube imaging.Then,aiming at the problem of insufficient model training data,we use the method of random combination of single characters and adding noise to achieve the target of data enlargement.Finally,the experimental results show that compared with the character recognition model based on the pre-segmentation method,the model recognition accuracy based on the segmentation free method is improved by more than 20%.Compared with the CRNN model,the recognition accuracy of the improved ATT-CRNN model is improved by 2.1%,reaching 98%. |
PDF Full Text Request