A Real-Time Fault Detection Framework Based on Unsupervised Deep Learning for Prognostics and Health Management of Railway Assets

Fault detection based on deep learning has been intensively investigated in the recent decade due to increasing availability of data and its ability to engineer features with deep neural network architectures. Despite much attention to its application, the major challenge is the lack of available labelled datasets to build the models since maintenance is usually conducted regularly to avoid significant defects. This paper aims to propose a successful real-time fault detection framework based on unsupervised deep learning using only healthy normal data. The approach is based on autoencoder architecture and a one-class support vector machine as a classifier. As a case study, large real-world datasets acquired from railway door systems have been employed. The five different types of deep learning models and a one-class classifier are trained and comprehensively validated based on performance metrics and sensitivity analysis. In addition, two experiments have been carried out to verify the model’s adaptability and robustness to variational time-series data. The result shows a typical autoencoder is the least sensitive to a decision boundary set by the one-class classifier. However, the two experiments show that the fault detection accuracy for a bidirectional long short-term memory-based autoencoder is considerably higher than other autoencoder-based models at 0.970 and 0.966 as F1 score, meaning only this model is adaptable and robust to variational data. The experimental result allows us to obtain the understandability of the deep learning models. Furthermore, the regions of anomalies are localised with unsupervised models, which enables diagnosing the cause of failure.