A Physics-Specific Change Point Detection Method Using Torque Signals in Pipe Tightening Processes

Abstract

Change point detection in torque signals has widely been adopted for quality inspection during pipe tightening processes. Previous studies on the change point detection in this process generally focus on directly detecting the change points throughout torques without considering the underlying mechanism that generates various quasi-periodic nonlinear profiles, thereby introducing a series of false change points and increasing the risk of releasing defective pipes. To overcome this problem, we propose a novel change-point detection approach by fully considering the profile generating mechanism, and introduce a similarity-weighted matrix with an integration of dynamic time warping and tightening physics. Thus, the probability of false detection of change points is reduced. A weighted regression model is developed to determine the authentic change points by introducing the tightening process constraints. The performance of the proposed approach is demonstrated by both numerical and real case studies, and results show that the proposed method achieves a more effective detection power than the other existing methods in the pipe tightening processes. Note to Practitioners —This paper was motivated by the real industrial needs of change point detection in torque signals during pipe tightening processes. Identifying the change points that substantially captures the tightening process conditions, and thus assuring the quality of pipe connections, is of great practical interest. A key challenge to this problem is that multiple uncertainties exist in the pipe tightening process, thereby leading to various nonlinear profiles with different lengths in torque signal. This fact introduces a number of fake change points and consequently mixes with the real change points regarding the process conditions. To address this issue, we propose to generate a “physical basis function” to adaptively assign the weights to those nonlinear profiles that may produce the fake points, by using dynamic time warping. To better use this method for change point detection, two factors should be noted: 1) the generated basis function should fully comply with the process physics and 2) this method can be widely used in torque signals with different lengths and multiple nonlinear profiles.