Data-Efficient Degradation Progression Modeling in Industrial Compressors via Baseline-Referenced Deep Feature Learning and Unsupervised Clustering

Accurate modeling of degradation progression in rotating machinery remains challenging in real industrial systems, where data are inherently limited and imbalanced because of safety-critical operations and associated risks, and the cost of acquiring fault data is high. These conditions make it difficult for data-driven approaches to reliably capture the evolution of degradation over time. To address this challenge, this study proposes a hybrid framework that models degradation progression as a set of distinct behavioral regimes driven by loss of lubrication. The proposed framework first applies adaptive scaling guided by an α parameter derived from Root Mean Square (RMS) deviation of the vibration signals relative to the baseline condition, aiming to mitigate data leakage during preprocessing while improving robustness to data imbalance. It then performs baseline-referenced deep feature learning using a lightweight Long Short-Term Memory (LSTM) model trained only on baseline data. The trained model is subsequently used to encode the entire dataset into latent representations, which are finally clustered using Mini-Batch K-Means to organize distinct degradation-related behavioral regimes. Results on both real-world and experimental datasets demonstrate that the learned latent representations strongly agree with the degradation regimes associated with baseline characterization and α-guided progression patterns, achieving an Adjusted Rand Index (ARI) of 1.0 across both datasets with respect to the internally defined reference stages.