부산대학교 도서관

Condition monitoring (CM) of machinery is important for ensuring the reliability of industrial processes. To adapt to the rareness of data from faulted machinery, semisupervised CM can be implemented by training on only healthy samples. However, the performance of CM can be impaired by the variability of operating data acquired from complex machinery. Additionally, the accuracy of results is limited by the impractical assumption that samples under different health conditions are naturally separable. To address these problems, an adversarial representation learning method is developed herein. The method is trained by reconstructing operating data in both signal and latent spaces, and adversarial evolution is adopted to avoid the convergence at local optima. In this case, data representations of health conditions can be obtained to suppress the volatility of measurements, and redundant information can be reduced by latent codes. Moreover, a strategy of representation embedding is developed to impose constraints on unhealthy data, guaranteeing separable samples under distinct health conditions in the monitoring stage. Furthermore, feature fusion is conducted to avoid missing detailed information on health conditions. The satisfactory performance of the proposed method is demonstrated by experiments in test benches and actual scenarios of wind power generation.