Deep Learning-based Recognition Algorithm For Industry Meters And Its Applications

Industrial meters are widely used in military,aerospace,industrial and other fields,especially in harsh environments with high temperature,high pressure and high radiation,such as substations.Classified by value displaying method,industrial meters can be divided into pointer meters and digital meters.However,a large number of industrial meters nowadays cannot exclude human operation while recording.The shortcomings of manual recording,such as high work intensity,high labor cost,poor immediacy,low efficiency,and high error rate,can no longer meet the needs of modern industrial production and development.In addition,monitoring various meters in a repetitive scheme for long time can cause visual fatigue,which will inevitably lead to misjudgment.Therefore,it is urgent to suggest an intelligent reading recognition algorithm for industrial meters.Images of industrial meters can be highly affected by natural environments such as uneven illumination in each image,large range variation of illumination in different images,complex backgrounds,image blur,tilting of pointer meters,scale change,and cumbersome image preprocessing.Looking forward to resolve the above problems,this paper proposes a reading recognition algorithm for industrial meters based on deep learning.In order to handle the challenge posed by natural environmental impact on industrial meters reading recognition,this paper introduces several machine learning and deep learning algorithms,such as Mask-RCNN instance segmentation network,principal component analysis(PCA),fully convolutional networks(FCN),and K-Means convergence.The algorithm in this paper has greatly improved the positioning accuracy and reading recognition accuracy of the industrial meters.The main contents of this thesis are as follows:(1)A reading recognition algorithm for pointer meters based on Mask-RCNN was proposed.The Mask-RCNN instance segmentation network mainly performs tasks including industrial meters classification,industrial meters positioning,instrument dial segmentation,and pointer segmentation.In order to improve the instance segmentation accuracy of the instrument dial and pointer,this paper proposes a method to improve the performance of Mask-RCNN based on bilinear difference.In order to solve the problem that the meter image may be distorted and tilted,this paper proposes a perspective transformation correction algorithm based on irregular ellipses,which further improves the accuracy of the reading recognition of pointer meters.Experimental results show that the improved Mask-RCNN's target positioning accuracy is improved by 2% ~ 3%,and the instance segmentation accuracy is improved by 1% ~ 2%.In addition,taking the manual reading value as the true value in the experiment,the average relative error of the pointer meter reading was reduced to 2.35%,and the average citation error was only 0.1708%.(2)A reading recognition algorithm for digital meters based on full convolutional network was proposed.This algorithm uses the concept of semantic segmentation in a full convolutional network to realize the positioning and recognition of digital instrument dial,digital display areas,and digital characters.This algorithm solves the problem that the traditional digital meter reading recognition process is easily affected by the character area positioning and character segmentation effect,and greatly improves the accuracy and reliability of digital meter reading recognition.(3)An algorithm for simultaneous reading of pointer and digital meters was proposed.A traversal enumeration multi-instrument mechanism is added,combining the above two algorithms,to realize the simultaneous reading recognition of pointer and digital meters.Finally,the algorithm in this paper was successfully applied to the substation inspection robot project.The experimental results show that the algorithm in this paper has higher accuracy of reading,stronger universality and generalization ability than the traditional algorithm.