Research On Recognizing Pointer Instrument Of Supervisory Control System Based On Convolutional Neural Network

With the rapid development of machine visual technology,image recognition has been widely adopted for various industrial systems.The pointer instrument has several advantages like operability,water resistance,high ability of anti-freezing,high reliability and favorable price.Therefore,it is applied into different fields,such as electrical power production,oil industry,chemical industry,smelting and nuclear industry.Reading and status indication of such instruments could reflect working state of equipment and environmental data of industrial production.However,the digital communication interface has not been provided for these instruments.Therefore,the data of instrument cannot be automatically collected.As a result,recording data manually is one of problems to achieve automatic industry,besides,safety is another serious problem.Aiming at these problems,consider the demands of the substation’s supervisory control system,this thesis try to investigate the recognizing of pointer instruments in supervisory control system.Because different classes of instruments have different reading recognition models and algorithms,and the corresponding meanings of recognit ion results are also different.Consequently,in the identification system which contains various classes of pointer instruments,class of instruments should also be identified.Based on the theory of Convolutional Neural Network(CNN)and classic CNN models,the thesis gives an improved CNN model to identify the class of instrument.Subsequently,the training dataset and mini-batch gradient descent(MBGD)algorithm with momentum are used to train improved CNN model.Experimental results show that the improved model recognition accuracy is better than those of Alex Net and VGG16 models.Finally,the improved model is effective,because the Top1 accuracy of classification reaches 98.2% on the test dataset.Using digital image processing technology to recognize the reading of pointer instruments is a commonly used technical solution,but it also has shortcomings in practical applications.First of all,process of pre-disposing image is complicated.Secondly,results of recognizing image would be influenced because o f angle of taking photo,lightness and dust of instruments surface.So,this thesis adopts an end to end CNN model to recognize reading of pointer instrument s,which provides another method to recognize pointer instrument’s reading.In this thesis,the improved CNN model is trained with using the augmented pointer instrument’s image samples and experimented.Experiments show that cross-layer feature fusion and shallow-layer enhancement of larger feature maps have brought about an improvement in the precision of instrument reading recognition.Results show that the trained CNN model provides a small error of recognition,with a single image recognition time of about 14.3 ms.Anyway,the quality of instrument’s image is not high,it also shows high robustness and convenience.So,it is a pragmatic method.