Discriminative Feature Extraction Through Dictionary Learning for Piezoelectric Admittance-Based Bolt-Loosening Damage Detection

Abstract

Bolt loosening is a major concern in many structural systems, especially in critical applications like pressure vessels. Traditional detection methods, such as acoustic emission, fiber Bragg-grating sensing, and vibration-based techniques, often face limitations in sensitivity, scalability, computational complexity, and implementation for large, multibolt structures. This article introduces a new framework combining piezoelectric active sensing with dictionary learning for bolt-loosening detection. Active interrogation using piezoelectric transducers, known for their sensitivity to structural property changes, is leveraged to collect the piezoelectric admittance of the transducer-structure-coupled system. A dictionary learning framework is established through the sparse representation principle, where the admittance versus frequency data are modeled as a linear combination of a few basis atoms selected from an overcomplete dictionary. Sparse coding algorithms are employed to extract representative features, and subdictionaries trained for different damage scenarios are used to minimize the reconstruction error for identifying bolt-loosening conditions. Experimental case studies on a pressure box emulating space habitat usage demonstrate the effectiveness of the proposed method. Bolt loosening at various locations and different severity levels is successfully detected. Additionally, boundary effects are considered, and for corner bolts with loosening severity ranging from 20% to complete loosening, all five levels are accurately detected at 100%. The learned dictionary enhances the interpretability of the detection process, offering a practical, scalable technique. This approach addresses the limitations of traditional methods by offering a more efficient, data-driven solution well-suited for real-time applications with sparse data.