Detection And Classification Of Surface Defects In Arc Deposition Based On Deep Learning And SVM

Arc Deposition Manufacturing(ADM), which not only shares the inherent advantage of Additive Manufacturing in making complex shape components, but also has the ability of forming components efficiently with a low cost, has got wide attention form researchers in areas of 3D metal printing. ADM is a complicated processing for its principle of layer manufacturing, which means that the surface quality of every single layer has a fundamental connection with the forming precision and mechanical performance of the final deposition components. To manufacture a ADM component of good quality, detecting the surface defects that may appear in the arc deposition process efficiently and accurately would be a significant work. In consideration of this, a detecting and classification system of the surface defect s has been designed to meet the requirement of surface quality monitoring, which is based on the deep learning algorithm and support vector machine(SVM).An industry CCD camera of 1.3 billion pixels has been used to acquire the surface image of the arc deposition layer, however, to use it as the input image immediately would make the deep learning network complex and hard to train. A region of interest(ROI) with a pixel size of 85*109 could be an efficient way to cut down the input image size. It is necessary to preprocess the ROI with graying, histogram equalization and Gaussian filter before being sent into the convolutional neural network(CNN) for the purpose of abstraction and feature extraction. With the feature extracted by CNN, SVM is trained to achieve a better classification.This system is built on the platform of Visual Studio 2010, consisting of six modules, which are training module, testing module, loading and saving module of both CNN and SVM. There are 680 samples of five classes, that is normal, welding pores, welding hump, welding depressions and welding undercut, and a quarter of them are used for testing with the rest of them for training. In the final experiment of the system, a n accuracy of 99% in the training stage and 95.29% in the testing stage could be reached with a speed of 237 samples per second. It is concluded that the detecting system can work well on the deposition layer defect detection and has practical engineering value.