Temporal Variational Auto-Encoders for Semi-Supervised Remaining Useful Life and Fault Diagnosis

Abstract

Deep Learning has seen an incredible popularity surge in recent years mostly due to the state-of-the-art results obtained by neural networks. Nevertheless, within the Prognostics and Health Management community, even though its application in research endeavors is extensive, its use in practice still has many challenges to overcome. A major one is the difficulty to generate or obtain labeled data (e.g., health states) for fault scenarios, mostly because generating such data often involves costly experiments in which real components must be forced to fail. However, labeled data is what allows the supervised learning of neural networks, which is by far the most used and proven approach for training deep learning models. A method to tackle this problem is the use of semi-supervised training schemes, in which the model learns useful information from both labeled and unlabeled data. In this paper, a family of temporal aware Variational Auto-Encoders is proposed in conjunction with a novel semi-supervised training scheme. These two elements are used to encode both supervised and unsupervised information into the health state or RUL (Remaining Useful Life) model, effectively requiring significantly less labeled data to learn how to perform both fault classification and RUL evaluation when compared to the exact same structural model using a traditional semi-supervised learning scheme. The proposed approaches are validated using the well-known CMAPSS dataset in a RUL case study as well as a fault diagnostics case study using a proprietary dataset regarding the health state of a pumping system used in offshore oil production. Results indicate that the proposed approaches deliver better performance (in terms of balance accuracy for the classification case study and root mean square error for the regression case study) consistently through almost all levels of labels availability.