Structured Denoising Autoencoder for Fault Detection and Analysis

This paper proposes a new fault detection and analysis approach which can leverage incomplete prior information. Conventional data-driven approaches suffer from the problem of overfitting and result in high rates of false positives, and model-driven approaches suffer from a lack of specific information about complex systems. We overcome these problems by modifying the denoising autoencoder (DA), a data-driven method, to form a new approach, called the structured denoising autoencoder (StrDA), which can utilize incomplete prior information. The StrDA does not require specific information and can perform well without overfitting. In particular, an empirical analysis with synthetic data revealed that the StrDA performs better than the DA even when there is partially incorrect or abstract information. An evaluation using real data from moving cars also showed that the StrDA with incomplete knowledge outperformed conventional methods. Surprisingly, the StrDA results were better even though the parameters of the conventional methods were tuned using faulty data, which are normally unknown. In addition, the StrDA fault analysis was able to extract the true causes of the faulty data; the other methods were unable to do this. Thus, only our proposed method can explain why the faults occurred.