Multiple domain dynamic feature adaption transfer learning method for stranded wires health monitoring under variable vibration working conditions using laser-generated ultrasonic guided wave

Abstract

Laser-generated ultrasonic guided wave is proved as a promising way to the structural health monitoring (SHM) of Stranded wires. However, variable vibrations would greatly affect the damage identification precision of stranded wires for the consequent distribution changes to collected signals. Aiming at this problem, a robust domain adversarial adaptation capsule network (DAACN) is proposed to reduce the influence of vibration and achieve reliable health state identification. Firstly, the UGW signals are preprocessed by normalization, phase alignment and bandpass filtering. Afterward, these preprocessed signals are divided into source and target signals and mapped to a high-dimensional feature domain space by deep feature extractor. Furthermore, an adversarial learning model, which contains a global domain discriminator and a local domain discriminator, is designed to help align the marginal distribution and conditional distribution of the two domain features. Then, a damage decoupled capsule network is built to obtain the relationship between the source features and damage labels. Finally, by joint adversarial training among the damage decoupled capsule network, global and local discriminator, the final model is obtained to identify the real structural state of test signals. Experimental results show that the proposed DAACN method can effectively align the damage feature under different vibration working conditions. Identification precision of more than 99% in all conditions also proves the superiority of the proposed method.