Unsupervised Remaining Useful Life Estimation of Tool-Holder Bearings from Sparse-Autoencoder Reconstruction Error and Exponential Degradation Modeling

This paper presents a hybrid approach for tool-holder bearing prognosis, combining a Sparse Autoencoder with an Exponential Degradation Model. The objective is to enable early detection of wear and degradation, allowing for timely maintenance interventions and minimizing unplanned downtime. The Sparse Autoencoder is employed as an unsupervised anomaly detection tool, trained solely on defect-free data to learn normal bearing behavior and identify deviations through reconstruction errors. These errors serve as a health indicator, capturing early signs of degradation. To estimate the Remaining Useful Life, the Exponential Degradation Model is applied to the health indicator, fitting an exponential curve to predict the degradation rate and forecast the time remaining until failure. The experimental study involved two tool-holders operating under identical conditions. Tool-holder 1 was used for SAE development and for the offline calibration of the health-indicator thresholds, whereas the complete fixed configuration was subsequently applied to tool-holder 2 without SAE retraining or threshold recalibration. The results obtained on tool-holder 2 provide a preliminary within-domain evaluation of the proposed framework on an unseen component. A comparative analysis with three baseline models—Principal Component Analysis (PCA), a Naive Constant Model (NCM), and a supervised Long Short-Term Memory (LSTM) network—further contextualized the results. The Sparse Autoencoder consistently outperformed PCA and NCM in terms of accuracy and robustness, while requiring only a short defect-free baseline for training. The LSTM achieved the lowest Root Mean Square Error values overall, but its dependence on full life-cycle data and higher computational demands limit its practicality in industrial scenarios. Overall, the main contribution of this study lies in demonstrating that a simple, interpretable, and data-efficient hybrid framework can achieve reliable Remaining Useful Life prediction without relying on large supervised datasets, thus bridging the gap between deep learning performance and industrial applicability.