AI-based Gas Turbine Multi-Component Health Prognosis via Recurrent Expansion of Gas Path Parameters

Investigating Gas Turbine (GT) degradation based on Gas Path Parameters (GPPs) is essential for its maintenance and operation. Monitoring GPPs is crucial for early detection of degradation, enabling timely maintenance interventions and preventing potential failures. Integrating physics-based thermodynamic models with representation learning significantly improves predictive maintenance studies and provides a solution to challenges posed by expensive and impractical accelerated aging experiments. However, this integration also presents challenges in managing both data complexity and data drift. In this context, this article aims to address these gaps by extending and improving previous research through (i) exclusive use of GPPs; (ii) the implementation of advanced data preprocessing techniques; and (iii) the use of innovative representation learning strategies. Specifically, it introduces ProgMachina, a tool for data quality analysis in prognosis studies that addresses issues related to data complexity in GPPs. Furthermore, to diversify the feature space and improve the adaptability of representation learning to degradation patterns, this paper proposes using Multiverse Recurrent Expansion with Multiple Repeats (MV-REMR) approach, which is based on a series of Recurrent Neural Networks (RNNs). For evaluation, this study incorporates cross-validation and multiple metrics while comparing against multiple RNNs and state-of-the-art works, demonstrating stable and promising performance, making it a suitable choice for GT prognosis tasks.