| Human society has experienced three industrial revolutions,and in the 21st century it ushered in the fourth industrial revolution of the Internet of things and services in the manufacturing industry,known as industry 4.0.Industry 4.0 three topics are intelligent logistics,intelligence and wisdom factory production,the focus of the wisdom of the factory is the study of intelligent manufacturing systems and processes,as well as the realization of the network distributed manufacturing facilities.At the national strategic level,China and Germany signed the China-Germany cooperation programme of action in 2014,marking the beginning of the three-step of Made in China".Before production after the completion of the warehousing storage,steel coil to coil number and entry system by artificial recognition,in order to unified management in the future,therefore,to replace manual operation number automatic identification coil has become the construction of factory is very important part of wisdom.In this thesis,the traditional convolution neural network itself has carried on the improvement as follows:first of all,to reduce the convolutional neural network,the number of hidden layer reduces the traditional five hidden layer network is 3 layers,after it was found that in terms of the identification results of character images,accuracy without significantly lower,and the training time is greatly reduced;Then,the traditional maximum sampling and mean sampling method are improved,and a random probability matrix is added to carry out random sampling,which can represent the characteristics of the image more authentically.Then,during the training process,the training method can be selected by random selection of training data,which can effectively shorten the training time.In addition add a momentum parameter,accelerate the convergence speed of network,reduce the concussion of network;In addition,the parameters of the network are updated and the target function of gradient descent method is improved,and the efficiency of the algorithm is improved.In order to solve the problem of fitting,the training samples were increased by rotation,translation,scaling and cutting of training images.Then the initial parameters of the convolution neural network are optimized by DBN.In addition,an improved impulse is added to the DBN training to prevent the problems of local optimal solution.In order to improve the network efficiency,the global training method is changed to parallel.Finally,with the improved convolution neural network for 500 steel coil image recognition,the cumulative accuracy reached 95.4%,and recognition time is only 3 minutes,compared with the deep learning algorithm,and the depth of the traditional learning algorithm,effect is obvious. |
