Research On Multi Side Joint Inspection Method Of Industrial Parts Based On Model Fusion

With the development of China’s manufacturing industry and the goal of building an industrial and manufacturing power,China’s industrial production mode is changing from small-scale,multi labor,low efficiency to large-scale,automatic,safe and reliable production mode.The scale of production and the types of business lines are increasing day by day.At the same time,new industrial standards and new names are constantly put forward in industrial production Words: industry 4.0,synchronous engineering,big data technology,etc.The purpose of intelligent warehousing is to effectively solve the problem of rapid flow of materials and products in industrial production and inventory.Intelligent storage is different from the old storage method,which requires a lot of manpower and energy to obtain the information of materials in the storage.Rapid access and identification of a large number of industrial materials and parts,and maintenance of inventory information are the key and difficult points of warehouse management technology.In order to solve the problems of type identification and quantity management of such industrial parts and materials,this paper proposes a method of multi side joint identification of industrial parts based on multi model fusion,which is mainly divided into two stages: the first stage is to use neural network to extract multiple side features of the same part to form the first stage of category prediction,and the prediction of single model has preference for specific types and does not need to be improved It can achieve good prediction effect for all kinds,so the multi model method is used for prediction.Using a variety of ways to fuse and combine the first stage of multiple side prediction results into new features,and then send them to the second stage of BP neural network for prediction to produce the final results.In the data set of self built 100 parts,the average accuracy rate is improved from 88.5% to 98%,which has a good effect.The main contents of this paper are as follows:1.This paper improves the existing multi-channel acquisition system.Because there are many kinds of industrial materials and parts,individual parts have small differences,and have multi side similarity features,it is difficult to extract the key features for recognition.In view of the characteristics of industrial production materials and parts,and the shortcomings of slow manual image acquisition,this paper improves the existing multi-channel visual acquisition system to make the existing acquisition system easier to use and save time and labor costs,and constructs the industrial parts data set as the experimental object.2.A series of standard processes of preprocessing methods are designed to reduce the error of deep learning model caused by image quality problems.Through Gaussian blur,binarization,edge detection and clipping,the main part of the part can be found and the influence of the background can be removed,which can effectively reduce the complexity of image operation and extract more effective features.The method of data enhancement is used to expand the part data set,including the whole image is shifted to four directions,the mirror operation on the plane is applied to the image,and the clockwise and counterclockwise rotation operation is applied to increase the number and size of the training set.3.Based on the idea of model fusion,this paper proposes two models to solve the problem that the prediction accuracy of the current single depth learning model is not high enough,and it can also solve the problem of highly similar parts classification.The densenet network with the highest classification accuracy is used as the bottom feature extractor in the two models to carry out the first level prediction.Multiple parts data are transmitted to the model to obtain the prediction results.The prediction results are reconstructed through a variety of ways,and the reconstructed feature results are input into the BP neural network model of the second level to obtain the final results.Experiments show that the accuracy of the multi-stage prediction model based on multi feature fusion is greatly improved in the part data set.