Domain Adaptation Of Population-Based Of Bolted Joint Structures For Loss Detection Of Tightening Torque

Abstract This paper investigates how to improve the performance of a classifier of tightening torque in bolted joints by applying transfer learning. The procedure uses vibration measurements to extract features and to train a classifier using a Gaussian Mixture Model (GMM). The key to enhancing the surrogate model for torque loss detection is considering the bolted joint structures with more qualitative and quantitative knowledge as the source domain, where labels are known and the classifier is trained. After applying a domain adaptation method, it is possible to reuse this trained classifier for a target domain, i.e., a set of different limited data of bolted joint structures with unknown labels. Four different bolted joint structures are analyzed. The new experimental tests adopt a wide range of torque in the bolts to extract the features with the respective labels under safe or unsafe tightening torque. All combinations of possible source or target domains are considered in the application to demonstrate whether the method can aid the detection of the loss of tightening torque, reducing the learning steps and the training sample. A guidance list is discussed based on this population-based SHM of bolted joint structures.