Physics-Motivated Domain Adaptation for Smart Touch-Based Bolted Flange Looseness Detection Using Stress Wave Sensing, Frequency Response Functions, and Machine Learning

Abstract

Bolt looseness in flanged connections can compromise the structural integrity and safety of pipeline systems, particularly in subsea environments. Stress wave-based structural health monitoring combined with machine learning has shown promise for detecting bolt preload loss, but existing approaches typically assume training and testing data originate from the same structure. Deploying developed model across flanges of different geometries induces domain shift that substantially degrades model performance. This paper proposes a physics-motivated domain adaptation framework based on frequency response function (FRF) signal transformation for bolt looseness monitoring across flanges of different diameters. Rather than performing feature-level distribution alignment, the method performs signallevel adaptation by modeling each flange as a linear time-invariant system per bolt preload state and estimating its FRF via the H1 spectral estimator. A transformation based on the ratio of source and target FRFs maps stress wave signals into the target flange, after which Mel-frequency cepstral coefficient (MFCC) features are extracted and classified using a support vector machine (SVM). Critically, this framework substantially reduces the data collection burden on new structures: since large-diameter flanges predominate in field deployments but large-scale data collection there is often impractical, the method requires only a small target-structure calibration set to estimate its FRF, which is then used to transform signals from a data-rich source structure rather than collecting extensive labeled target data. The framework is evaluated across three cross-flange transfer scenarios: transfer between 6-inch and 9-inch diameter flanges in air, transfer between 6-inch and 9-inch diameter flanges underwater, and transfer between 4.25-inch and 5- inch diameter flanges underwater. Using only twelve calibration signals from each target structure, the method achieves 96.0%, 92.2%, and 91.9% 3-class classification accuracy across these three scenarios respectively (93.4% average), substantially outperforming no adaptation (32.3% average), a Domain-Adversarial Neural Network baseline (38.7% average), and a calibration-only classifier (71.3% average).