Leveraging hybrid deep learning and generative modeling to accurately estimate the remaining useful life of mechanical systems

With the advent of Industry 4.0 (I4.0), machine learning (ML) in artificial intelligence (AI), industrial Internet of Things (IIoT), and cyber-physical systems (CPS), the development of data-driven applications like predictive maintenance (PdM) has accelerated. In asset-dependent industries, PdM has reduced operational costs, increased productivity, reduced downtime, and improved safety management. Predictive maintenance solutions also help identify failure sources, eliminating unnecessary maintenance. The concept of prognostics and health management (PHM) is a predictive maintenance approach that has gained recognition as an essential paradigm in smart manufacturing. Its purpose is to provide reliable methods for monitoring the health condition of industrial equipment. To achieve this, efficient and effective methods for monitoring the health of systems are necessary. These methods involve processing and analysing large amounts of equipment data to identify anomalies and provide diagnosis and prognosis. Prognostics is a crucial procedure in the field of PHM that involves predicting future conditions. It primarily focuses on projecting the remaining lifespan of a machine, which is the duration it can continue to work as intended. This estimation is commonly referred to as the remaining useful life (RUL) of the system. The field of prognostic research is still in its early stages, which accounts for the numerous issues that need to be addressed. Despite their potential to reduce costs and increase productivity, prognostics and health management face significant challenges. These include incomplete learning from models, accuracy issues with evolving systems, and the impracticality of human tagging due to the large data volumes generated, particularly in the Industrial Internet of Things (IIoT). A critical aspect of prognostics and health management is the estimation of the remaining useful life (RUL) of equipment. This estimation is complex and requires balancing computational demands with the capabilities of predictive systems. The estimation of remaining useful life (RUL) is a fundamental challenge that must be addressed to successfully adopt predictive maintenance, especially given the early stages of research in the field of prognostics. This thesis will focus on RUL estimation for monitoring equipment using deep learning (DL) methods. This thesis presents the development of hybrid deep learning models, integrating generative modelling techniques to enhance predictive maintenance strategies, specifically in the fields of aero-engine prognostics and Remaining Useful Life (RUL) estimation. It outlines three significant contributions that leverage the capabilities of these hybrid models, supported by a comparative analysis using benchmark Commercial Modular Aero Propulsion System Simulation (C-MAPSS) turbofan engine datasets to validate the advancements in prediction accuracy. Our research is motivated to advance computational efficiency through the design of streamlined pre diction models that not only reduce the complexity of the network but also augment its processing capabilities. To provide a robust hybrid-deep learning framework, focusing on the aerospace sector, thereby contributing to the field through models that offer more precise predictions. The first contribution proposes two hybrid deep learning architectures that leverage the multi-modal and hybrid proficiencies intrinsic to deep neural networks. The objective is to encapsulate critical data and maintain comprehensive information at varying intervals, thereby enhancing the accuracy of RUL estimations. The second contribution introduces another data-driven framework that combines temporal convolution, recur rent skip components, and an attention mechanism to improve the accuracy of RUL estimation. The recurrent skip component finds long-term patterns in time series data, while temporal convolution extracts high-level features from longer sequences. Finding hidden representations and degradation development interactions between features at each window position in the input matrix is what the attention layer does to focus on the most important information for RUL estimation. The third contribution addresses the formidable challenges associated with limited data availability, elevated feature dimensionality, and complex feature interrelationships. To circumnavigate these intricacies, we advocate the application of GANs. GANs can generate synthetic sensor signals. These artificially created signals can effectively supplement the existing data, thus solving the constraints brought about by data scarcity. The data augmentation makes the feature space bigger and helps to understand the original data distribution, providing a new perspective for creating predictive systems for estimating the RUL.

Publisher

University of Galway

Rights

Attribution-NonCommercial-NoDerivatives 4.0 International

cbn

Collections

University of Galway Theses (PhD Theses)