Systems and machines undergo various failure modes that result in their health degradation, so maintenance actions are required to restore them back to a state where they can perform their expected functions. Machine health degradation is inevitable, and so is the maintenance cost associated with it. However, with proper maintenance plans, maintenance costs can be reduced, machine life can be extended, and ultimately we can ensure workplace safety. The field of prognostics is vital to systems health management and proper maintenance planning. A reliable estimation of the remaining useful life (RUL) of machines holds the potential for substantial cost savings. Data-driven methods for predictive maintenance have been recognized as one of the most promising maintenance strategies because of their high efficiency and low cost compared to other strategies.

This work uses a sequential approach through experimentation to investigate the two main machine-learning-based methods for remaining useful life prediction, the similarity-based and direct-approximation methods. Drawing insights from existing works in the literature, the two stages of development of a similarity-based model (SBM) were optimized resulting in the development of improved similarity-based models using supervised and unsupervised machine learning methods for the health index construction. Ultimately, this work proposes a novel remaining useful life estimation model that leverages the concept of Large Language Models (LLMs) for more efficient time series data representation learning and prediction applied to the remaining useful life prediction use case. The experimental results indicate that the proposed Encoder-Transformer architecture outperforms the existing state-of-the-art models.

Other highlights of this work include the bottom-up experimental approach taken to select the best methods and make improvements. The benefits of this approach can be seen from the improved remaining useful life prediction models developed in this work compared to their other counterparts in the literature and the insights this work provides. In this work, ten separate machine-learning models were developed, trained, and tuned for experimentation purposes.

To summarize, three improved RUL prediction models:​ an Encoder-Transformer direct-approximation-based model, an Improved Unsupervised Learning-based Similarity-based model with Principal Component Analysis (PCA), and a Transformer-Assisted Similarity-based models were developed in this work. These models rank first, second, and fourth best amongst the twelve state-of-the-art models they were compared in the literature.