An Improved Convolutional Neural Network for Weld Defect Recognition

Abstract

Aiming at the problems of poor adaptability of pooling model, low feature selection ability and over-fitting when traditional convolutional neural network (CNN) is applied to weld defect recognition, a new method of weld defect recognition based on improved pooling model and feature selection CNN (IPFCNN) is proposed. According to the characteristics of weld defect image, the average pooling model is improved by taking into account the pooling region and its feature distribution. In order to enhance the feature selection ability of the CNN, an enhanced feature selection method combining random forest and CNN is proposed. A case study of weld defect recognition in the manufacturing process of steam turbine is to illustrate the work. The results show that the proposed method IPFCNN has dynamic adaptability in dealing with pooling region with different feature distributions and improving the feature selection ability, and it has higher defect recognition rate than the traditional CNN method.