Deep representation clustering of multi-type 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 so reducing manual participation in labeling data as much as possible has been attracted increasing attention. This paper 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 multi-scale features and had better generalization ability. The structure and training method of GAN-SC are studied, and GAN and Spectral Clustering(SC) algorithm are combined for damage diagnosis. The proposed GAN-SC framework harnesses the synergy of GAN’s ability to extract effective multi-scale 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 the Western Reserve University, using cluster evaluation indices (NMI and ARI). The results show that the superior recognition capabilities of GAN-SC emphasizes its potential for real-world applications in structural damage detection by generating virtual labels through spectral clustering. IEEE