Deep Representation Clustering of Multitype Damage Features Based on Unsupervised Generative Adversarial Network

Damage identification based on deep learning has become a hot topic recently. Damage identification and classification methods based on neural networks are much concerned, and therefore, reducing manual participation in labeling data as much as possible has attracted increasing attention. This article presents the work on developing a damage detection method by using limited information features to improve the performance of clustering in unsupervised learning. In order to improve the accuracy of unsupervised clustering algorithm, a damage classification method is proposed by using measured data based on deep learning network. A generative adversarial network (GAN) is introduced into the unsupervised clustering process, which is able to extract effective multiscale features and has better generalization ability. The structure and training method of GAN-spectral clustering (SC) are studied, and the GAN and SC algorithm are combined for damage diagnosis. The proposed GAN-SC framework harnesses the synergy of GAN's ability to extract effective multiscale features and SC's potential to generate virtual labels, improving generalization capabilities. Some signal preprocessing methods are used to reduce the noise of the original data while retaining the high features of the fault data as much as possible. The proposed method is verified by a numerical bridge dataset and a popular experiment dataset from Case Western Reserve University (CWRU), using cluster evaluation indices [normalized mutual information (NMI) and adjusted Rand index (ARI)]. The results show that the superior recognition capabilities of GAN-SC emphasize its potential for real-world applications in structural damage detection by generating virtual labels through SC.