Multicase structural damage classification based on semisupervised generative adversarial network

With the rapid development of computer science and the need for structural safety assessment, structural health monitoring (SHM) systems are widely used in structures. SHM systems primarily rely on sensor systems to collect data related to structural safety conditions, which are then analyzed and assessed for performance evaluation. However, structures in real world are often affected by many uncertain factors, making damage detection based on pattern recognition still difficult to apply. In recent years, research on damage recognition based on machine learning has gained considerable attention. One of the research directions is to use machine learning algorithms to extract features from the dynamic response of structures. Aiming at the problem of inaccurate recognition by machine learning in the case of fewer label samples, this paper proposes a structural state classification method based on semisupervised deep learning. The method is verified on the vibration data of a steel truss bridge and a three-story framework structure to realize the classification of structural states under different working conditions. Unlike the general semisupervised learning method, this paper introduces the mean square error (MS) loss function in the loss function of generative adversarial networks (GANs), thereby enhancing the model training effect (mean square error-generative adversarial networks, MS-GAN). The semisupervised learning uses a small amount of supervised information to guide GAN and then sorts and screens unsupervised data through joint probability, which can reduce labeling costs and improve model accuracy. Compared with the general semisupervised GAN, the algorithm proposed in this paper makes full use of some labeled samples to enable the state recognition and classification of semisupervised learning. By properly utilizing labeled data, the accuracy of state recognition is significantly improved. Finally, a range of training tasks are implemented in order to enhance the classification capability of the proposed MS-GAN through the establishment of varying supervised ratios.