Towards practical data-driven predictive maintenance: a robust and generalizable deep learning approach

• Main Author: Mohamed Ragab Mohamad Adam
• Other Authors: Kwoh Chee Keong
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; Engineering::Mechanical engineering::Motors, engines and turbines
• Online Access: https://hdl.handle.net/10356/163792
• Institution: Nanyang Technological University

Predictive maintenance (PdM) is a prevailing maintenance strategy that aims to minimize downtime, reduce maintenance costs, and increase machine reliability. With the huge data availability in the era of Industry 4.0, data-driven predictive maintenance leverages Artificial Intelligence (AI) and historical data to forecast potential machine failure, enabling accurate in-advance maintenance schedules based on the machine condition. Recently, deep learning has shown widely acclaimed potential for data-driven predictive maintenance tasks. Nevertheless, deep learning only performs well under two main assumptions: (1) the availability of vast amounts of labeled data; (2) the training phase and deployment phase share the same data distribution or working condition. These constraints can significantly hinder the deep learning applicability to real-world environments. On the one hand, data labeling and annotation can be very laborious tasks, especially for sensor data that encompass complex underlying patterns. Furthermore, running to failure for vast and sophisticated machines can be costly and cause catastrophic consequences. This imposes an additional challenge on obtaining labeled data that identify the health status of the machine. On the other hand, under the premise of the real-world manufacturing environment, variation in working environments is inevitable. This can yield different sensor readings across the working environments even for the same industrial asset. As a result, the data generated from each working condition may have different statistics (i.e., distributions). Under those circumstances, a model that performs well under one working condition performs poorly when tested on data from different working conditions. Therefore, there is a need for more practical deep learning approaches that can handle the aforementioned limitations. This thesis aims to develop robust and generalizable deep learning approaches that can work under real-world environments i.e., variability of working conditions and lack of labeled data. The focus of this thesis is on two main tasks of predictive maintenance: fault diagnosis and fault prognosis (i.e., Remaining Useful Life Estimation (RUL)) tasks. The thesis' contributions are structured according to the predictive maintenance tasks. Part I includes two pieces of work to address the real-world challenges of the remaining useful life estimation task. First, an attention-based sequence-to-sequence model is proposed to exploit historical information better, improving robustness and generalization across different industrial assets. Second, a novel contrastive adversarial domain adaptation (CADA) approach is proposed to address the lack of labeled data coupled with the variability of working conditions. The proposed approach aims to transfer the knowledge learned from one labeled operating condition (source domain) to another operating condition (target domain) with no available labels while preserving the target-specific information during the adaptation process. Part II focuses on the fault diagnosis task and includes three pieces of work that have been developed to address the limitations of deep learning in real-world environments. The thesis first proposes a novel Self-supervised Autoregressive Domain Adaptation (SLARDA) to address the lack of labeled data while explicitly considering the dynamics of time series data during both feature learning and domain alignment. Subsequently, towards a more scalable domain adaptation approach, a novel adversarial multiple domain adaptation (AMDA) method for single-source multiple targets (1SmT) is proposed to handle multiple unlabeled working conditions concurrently. Last, unlike previous domain adaptation approaches that assume access to unlabeled data from the target working condition, a novel contrastive domain generalization (CDG) is designed to generalize to unseen working conditions with no prior data available. The main goal of this thesis is to pave the way for deep learning in real-world environments, towards practical data-driven predictive maintenance. Despite the significant thesis's contributions, there remain some obstacles that hinder the applicability of data-driven predictive maintenance. These obstacles have been regarded in the future works of the presented thesis, which can be briefed as follows: considering the class-label shifts across different working environments; (2) quantifying the uncertainty of the proposed data-driven approaches; (3) realizing explainable deep learning approach that can provide reasons behind the decisions and give descriptions to the end-users.