DRCDCT-net: a steel surface defect diagnosis method based on a dual-route cross-domain convolution-transformer network

Abstract

Abstract With the development of industrialization, steel has been widely used in various fields. Current artificial intelligence (AI) methods based on steel surface images can automatically classify defect types on steel surfaces, but they still face challenges when embedded in actual industrial production. For example, the performance of convolutional networks is limited, and some categories of industrial fault data are scarce. In order to alleviate the above problems, this paper proposes a novel network structure, DRCDCT-Net. It is designed as a dual-route structure: a feature attention defect diagnosis module (FAD) and a cross-domain joint learning defect diagnosis module (CJLD). With the Feature Transformer designed as the core, the FAD is mainly responsible for handling defect classification tasks with sufficient samples. It can alleviate the problem of interdependence between features that are difficult for convolutional networks to learn. With the cross-domain joint learning network designed as the core, the CJLD is used to deal with the task of defect classification with extremely scarce samples. It can decouple the domain features of the image, so that the model can use data from different domains to learn the target data. When using the full data of both datasets, the model achieved 99.7 ± 0.2% and 90.0 ± 0.6% precision in Northeastern University (NEU)-CLS and SEVERSTAL, respectively. When using 20 images per class, it achieved 99.5 ± 0.2% and 71.3 ± 0.9% precision in NEU-CLS and SEVERSTAL, respectively. This paper proposes a novel deep learning structure. When faced with sufficient data, the model can take into account both performance and computing resource requirements. When faced with a small amount of sample data, the model can decouple data domain features and use unrelated features to learn the target data. The model proposed is more implementable and builds a bridge for the integration of AI technology and industrial defect real-time monitoring technology.