Deep Neural Network Architectures for Prognostics and Health Management of Industrial Systems

This thesis presents novel deep neural network architectures and training techniques to enhance their prediction capabilities in Prognostics and Health Management. These architectural and algorithmic changes are necessary to the standard deep neural network architectures since they were conceptualized to solve problems in different application fields, such as computer vision and natural language processing. Therefore, this thesis presents techniques to effectively model long-term time dependencies, efficient use of a large number of trainable parameters and productive use of unlabelled training data. Each publication proposes a solution to at least one of the above-stated drawbacks of standard deep neural networks. First, the proposed bidirectional LSTM architecture offers a new perspective for time-series analysis, which enhances prediction performance for incipient fault categorization. It is easier to retain long-term data in LSTM cells thanks to the sliding window data function, allowing for considering even longer time series patterns. Second, a novel anomaly detection framework using an auxiliary loss function is proposed to learn a hidden representation which is amenable to the task. Based on the K-means clustering loss, the auxiliary loss function is only computed for a subset of the hidden variables. Third, a generalized dilation layer is proposed for adaptive sampling across the temporal and feature variables for multivariate time-series data analysis. Two novel training approaches are presented to make the generalized dilation training process compatible with gradient-based learning. Moreover, finally, two novel transformer neural network architectures are proposed with a focus on the parameter-sharing phenomena for modelling inter-dependency in the temporal and feature domains. The developed training strategies and algorithms have been tested on benchmark datasets available in the Prognostics and Health Management literature. The benchmark datasets' overall results underline the proposed approaches' superior performance. Lastly, because every strategy suggested in this thesis is very modular, it is possible to combine them in different ways to build dependable and robust architectures for system health monitoring.