Reconstruction of compressed nonlinear ultrasonic signal for self-healing process of asphalt mixture based on physics-informed neural network

The vast quantity of data generated by ultrasonic monitoring poses challenges for efficient information transfer and significantly prolongs processing times, particularly when data must be exchanged between cloud servers and local endpoints. This situation is in stark contrast to the ongoing drive of the industry towards energy-efficient manufacturing equipment. Data compression, a standard approach for addressing this, is heavily reliant on the quality of the acquired signals, especially when it comes to nonlinear ultrasonic methods, which are sensitive to the characteristics of higher harmonics. Consequently, achieving high-fidelity signal reconstruction from compressed data is a critical challenge. This paper introduces a convolutional autoencoder physics-informed neural network (CAE-PINN) that includes a general loss term along with two terms incorporating physical information. The results demonstrate that this neural network model, which incorporates physical insights, consistently surpasses traditional signal reconstruction techniques in terms of lower reconstruction error and higher signal-to-noise ratio. Moreover, the model trained with numerical simulation datasets effectively guides the reconstruction of experimental signals. CAE-PINN also exhibits robust performance in dealing with noisy signals. Error analysis of nonlinear parameters and ablation experiment reveal that the CAE-PINN is capable of extracting intrinsic physical information from signals, allowing for the accurate recovery of true nonlinear parameters, a feat that surpasses purely data-driven approaches. With this capability, it can be confidently asserted that the CAE-PINN provides an accurate characterization of the self-healing process in asphalt mixture, even under high compression ratios. For further information on the development and replication of the CAE-PINN model, including access to the numerical simulation models and remote-control program code, please refer to the following resources.

https://github.com/Dalongna/Signal_reconstruction.