Transfer of learning in detecting tightening torque loss in bolted joints

Abstract

Most structural connections are made with bolted joints, aiming to guarantee structural stability, secure connection, and even introduce important global damping mechanisms in the structural dynamics involving nonlinear hysteresis phenomena. However, for the design properties to be maintained, a predetermined tightening torque must be maintained. Unfortunately, the dynamic structural demands to which most systems are subject lead to unavoidable torque losses over time. Such losses motivate the search for viable ways to monitor this torque, maintain the proper functioning of the mechanisms, and prevent future structural failures. The most intuitive way to automatically monitor this variation would be to instrument each joint with a load cell for direct quantification of torque loss. Unfortunately, this option proves to be prohibitive in many applications, mainly due to the high cost and the need for redundancy in the instrumentation used. This motivates the development of indirect techniques that allow evaluating this information from the measurements of the vibration responses at points of the structure that are easier to access. Recent studies with the use of modal parameters proved to be satisfactorily accurate in detecting the variation in tightening torque; however, It still has some limitations, such as the enormous variability of data between nominally identical structures and even between reassembly of the same structure, making it impossible to use the classifier training metrics for similar systems or even for the same system, after eventual maintenance and retightening. Such limitations require that a new model be trained for each structure of interest and each reassembly of the same system. In this sense, this scientific initiation work proposes an approach to the problem via modern Learning Transfer (TA) techniques, making a domain adaptation of the modal parameters. Among the TA methods, domain adaptation with the Transfer Component Analysis (TCA) method is a suitable formulation for this type of application and should be extensively investigated in this IC. The tests of the method to be proposed will use databases from test benches from the laboratory of the advisor's research group involving various vibration levels and different tightening torques for a detailed statistical study of TCA's application to improve the training of detection algorithms. This plan reports the scholarship holder's motivation, objectives, schedule, and description of the activities with the methodology developed in this study. (AU)