CDMTNet: a novel transfer learning model for the loosening detection of mechanical structures with threaded fasteners

Abstract

In mechanical systems, various threaded fasteners are widely used to connect separated components. Loosening occurs frequently due to exposure to harsh working environments, especially vibration environments. Loosening detection can eliminate hidden dangers in advance, and its importance cannot be overemphasized. Ultrasonic detection methods based on machine learning have become popular. However, a proposed machine learning model in a published paper is only suitable for a specific mechanical structure with threaded fasteners; the generality of this type of model is poor. In addition, existing models are well-trained based on abundant labeled data. The labeled data are commonly sparse in practical engineering applications, and it is labor-consuming to obtain these data. This paper presents the creation of a generalized detection model suitable for different threaded connection structures based on transfer learning and the exploitation of a small amount of labeled data for model training and accurate loosening detection. A transfer learning network named the cross-domain matching-mix transfer network (CDMTNet) suitable for different mechanical structures was proposed to transfer the knowledge in a source structure to a target structure. The CDMTNet consists of a feature extraction module, a feature disentangling module, and a classification module. The matching-mix method for data augmentation was proposed to finely train the CDMTNet by exploiting only a small amount of labeled data in the target structure. Two different mechanical structures with threaded fasteners (i.e., a lap joint and a globe-cone joint) were tested in the experiment. Their detection accuracies were both greater than 0.8. The results indicated that our method had wide application potential in engineering.