Application of improved least-square generative adversarial networks for rail crack detection by AE technique

• Published in: Neurocomputing (Amsterdam) 2019-03, Vol.332, p.236-248
• Main Authors: Wang, Kangwei, Zhang, Xin, Hao, Qiushi, Wang, Yan, Shen, Yi
• Format: Article
• Language: English
• Subjects: Acoustic emission; Generative adversarial networks; Least-square; Noise suppression; Rail defect detection
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2018.12.057

•An improved LSGANs is proposed to detect rail crack signal under noise interference.•MSE is added to the generator loss as a regularization.•LSGANs is modified into a conditional version to obtain samples’ latent details.•Proposed method is testified to eliminate statistical noise and mechanical noise.•The mechanical noise is acquired from the real operating environment of railway. In order to implement rail crack detection with acoustic emission (AE) technology in the actual application, an important problem to be solved is how to overcome the noise interference of wheel-rail mechanical interaction. In this paper, an improved Least-Square Generative Adversarial Networks (LSGANs) with regularization is proposed for AE signal denoizing. To overcome mode collapse problem and preserve more details for the crack signals, labels of the signals are introduced into the network and Mean Squared Error (MSE) is also added to the generator loss as a regularization. The proposed method is testified in two experiments: Gaussian noise elimination and mechanical noise elimination. The mechanical noise is acquired in actual railway environment. Through the training process of enhancing crack signals submerged by noise, an optimal parametrization for the network was first selected. The trained generator could be seen as an AutoEncoder and automatically obtain the filter architecture for certain noises. The eventual results demonstrated that the improved LSGANs could preserve more details of the crack signals than traditional denoizing method after noise elimination and has a significant robustness. The denoizing ability of the proposed method is verified and it could effectively remove the statistical noise and mechanical noise in rail defect detection.