Physics-guided neural network for structural health monitoring with lamb waves through boundary reflection elimination

Lamb-wave-based structural health monitoring (SHM) technology for damage location in plate-like structures relies on the postprocessing of captured signals after interacting with damage. Traditional methods typically leverage the time of flight (ToF) of scattered waves from damage. However, these methods are prone to reflected waves from structural boundaries which mix with scattered waves from damage. This is a vital problem faced by most ToF-based detection methods, which seriously narrows the inspection area. To tackle this problem, a machine learning framework, consisting of a multiscale spatiotemporal (MSST) fusion network, is proposed to facilitate the accurate extraction of the ToF of scattered waves through eliminating the influence of boundary reflections. Experiments are conducted with the time-domain Lamb wave signals recorded by a tactically designed piezoelectric sensor array on a 2-mm-thick Al-6061 plate. A pair of circle magnets is attached onto the plate as the wave reflectors. Through step-by-step moving of the magnets in the predefined grids, the corresponding Lamb wave signals are measured to construct a database. An MSST is subsequently designed to minimize the error between estimated and theoretical ToFs, with wavelet coefficients of the signals and transducer position as inputs. The model is trained with the Adam algorithm where 80% of samples in the database are used for training and the rest for evaluation. The final validations are conducted with the scatters off the predefined grids. Results demonstrate that the designed neural network architecture can effectively eliminate boundary reflections and enable precise ToF extraction of the scattered waves from damage. This allows the enlargement of the detection area and presents a promising and useful tool for enhancing the detection performance of existing SHM methods in complex structures.