Development of fatigue crack detection and remaining useful life estimation techniques using deep-learning and nonlinear ultrasonics

Abstract

In this study, to monitor the structural health of structures, noise reduction technique for ultrasonic signals, reference free fatigue crack detection technique using nonlinear ultrasonic and deep learning, and nonlinear ultrasound-based online remaining life estimation (RUL) technique were developed. First, a spectral noise reduction and signal enhancement technique using deep learning was developed. Using deep learning, inherent sequential patterns of nonlinear ultrasonic modulation components were learned and signals were reconstructed. Through this, noise components are removed and important components such as nonlinear modulation components are enhanced. Second, reference free fatigue crack detection technique using nonlinear ultrasonic and deep learning was developed. In order to learn the inherent sequential patterns of ultrasonic signals, a deep learning model was constructed and immediate online training was conducted. Then, multi-step ultrasonic signal prediction was performed through the trained deep learning model, and based on this, the absolute damage index was defined using the rate of change of the nonlinear ultrasonic modulation component. Damage is automatically determined using this absolute damage index without any user-specified threshold value or confidence interval. Thirdly, an online RUL technique was developed through continuous ultrasonic signal measurement. The fatigue index was defined using nonlinear ultrasonic modulation components. Then, the relationship between the defined fatigue index and the loading cycle number was derived as a power function that does not require prior information on the initial crack. Finally, the RUL was estimated by fitting the fatigue index with the derived power function, and the RUL was estimated by continuously updating as the life cycle progressed. The proposed techniques were verified through lug specimens used for structural connection and long-span bridge tests in actual operation.