Two stage deep learning for prognostics using multi-loss encoder and convolutional composite features

• Published in: Expert systems with applications Vol. 171; p. 114569
• Main Authors: Pillai, Shanmugasivam, Vadakkepat, Prahlad
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 01.06.2021; Elsevier BV
• Subjects: Algorithms; Coders; Convolution; Deep learning; Encoders-Decoders; Feature extraction; Industrial applications; Machine learning; Multi-loss encoder; Multilayers; Predictive maintenance; Redundancy; Remaining useful life; Representations; Shrink wrapping; Turbofan engines; Deep learning; Multi-loss encoder; Remaining useful life; Predictive maintenance
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2021.114569

•Proposed two-stage deep learning separates feature generation and RUL prediction.•A multi-loss objective function generates relevant information maximizing features.•Generated features capture degradation trend with low-noise and low-redundancy.•Benchmark performance is achieved for C-MAPSS dataset and blade-wear prognostics.•Analysis of data features reveal the benefits of relevant information maximization. Recent advancements in machine learning for remaining useful life (RUL) prediction are driven by the increased availability of data and computing power. Extraction of useful features from raw data leads to better prediction performance. Deep learning based feature generation is superior to labour intensive feature engineering that requires domain expertise. The presence of noise, temporal trends, and irrelevant features in sensor data pose difficulties in training efficient and reliable deep learning models. Proposed Multi-Loss Encoder with Convolutional Composite Features (MLE + CCF), improves feature discovery for deep learning using a two-stage approach, that separates feature generation and RUL prediction. MLE utilizes a multi-loss objective function to train a multi-layer convolutional encoder-decoder network. The relevant information maximizing encoder generates high-dimensional representation characterized by low-noise, low-redundancy, and high-correlation with degradation trend. The second stage implements depthwise separable convolution that learns temporal features from sequential data. The temporal features and encoded representation are concatenated, forming convolutional composite features (CCF). A fully-connected network is trained with CCF for RUL prediction. Experimental results for NASA turbofan engine dataset (C-MAPSS) show performances that are better than benchmark methods. An industrial application demonstrates blade wear prediction in shrink-wrapping equipment using the proposed algorithm. Results and analysis validate the suitability of MLE + CCF for predictive maintenance applications in industries.