Research On Key Technologies Of Machine Parts Classification Based On Deep Learning

With the new round of scientific and technological revolution,mankind will enter a new era.In the traditional industry,many jobs requiring manual operation are gradually replaced by machines,and the industry will develop towards automation and intelligence.The classification and identification of different mechanical parts to realize automatic sorting conforms to the development of this trend.Aiming at the problem of low accuracy in identifying parts in production line parts sorting system in industrial field,this paper proposes to use convolutional neural network to identify and classify parts.The author mainly conducts related research from two parts,constructs data set and designs network model.In the process of actual data,due to various constraints,the collected data is very limited,there is not enough data quantity will not be able to better trained network model,in order to solve this problem,this paper by the method of multiple data expansion to the original data sets were collected,such as image transformation method,random cutting,gaussian noise,bleaching processing,such as change the contrast,it expanded the amount of data to a certain extent.In order to satisfy the diversity of data sets and train a better network model,the traditional method is too simple.Therefore,this paper also adopts the generative adversarial neural network to expand the data and enrich the diversity of the data set,from the original small sample data set consisting of 400 pictures to 4400 pieces.Followed by the design of network model,this paper selected the VGG network model as the foundation,analyzes the advantages and disadvantages of VGG network model,and then aiming at the shortcomings of the VGG network model based on its was improved,the adjustment parameters,lower class,optimize the loss function and gradient descent method,etc.,design a training network model suitable for small sample data set.This paper builds the Tensorflow learning framework and Pycharm programming environment based on the Linux system of 64-bit Ubuntu 18.04 to complete the network design and experiment.By comparing and analyzing the VGG model and the improved network model,the results show that the improved model is slightly higher in recognition accuracy and speed than the VGG model,indicating that the network model in this paper is feasible.Finally,we compare the method of parts classification based on deep learning proposed in this paper with some traditional methods.The method in this paper has more advantages,mainly in the aspect of feature extraction,which also provides certain reference value for the industry to further move towards automation and intelligence.