Abstract Summary
The major challenge in detecting the loss of tightening torque in bolted connections using indirect vibration measurements is the considerable variability and noise presented in the data. This issue is more complex, assuming a heterogeneous population of a set of bolted connections with incomplete data or unlabeled information, i.e., where only some torque conditions are known to train state classification algorithms. Thus, even training algorithms to recognize the loss of tightening torque, generalizations to similar structures are challenging to be produced without retraining these classifiers. This paper illustrates how we can use a population of vibration data measured at bolted joints to avoid this limitation. The key idea is to use the information from a classifier trained with data in a source structure, i.e., with available labeled data, and perform a domain adaptation for a target population of a similar setup with another bolted joint, where we do not have the tagged information previously. Various domain adaptation algorithms can be chosen, but here, a simple transfer component analysis (TCA) was demonstrated to be enough to transfer the information to detect the loss of tightening torque assuming a set of different structures. The method is exemplified in a group of three Orion beams with several tightening torques applied. The features used are the simple resonance frequencies extracted from the transmissibility functions. A Gaussian Mixture Model (GMM) is created to classify the states in a supervised mode using the source data. All detailed steps are carefully discussed with recommendations and advantages of this approach to implement this method.
