An Integrated Multi-Task Learning Model for Predictive Maintenance (PdM) in Industrial Equipment

Abstract

Abstract

Predictive maintenance (PdM) is increasingly essential in modern manufacturing environments, where unexpected equipment failures can cause production loss, quality deviation, schedule disruption, and elevated operational risk. While many PdM studies focus on a single objective such as Remaining Useful Life (RUL) estimation, practical maintenance decisions often require complementary outputs, including near-term risk indication and interpretable health-stage information. To address this gap, this thesis presents an integrated multi-task learning framework that jointly performs RUL prediction, imminent-failure detection, and health-stage classification from multivariate sensor time-series data. To ensure reproducibility, the study specifies an end-to-end pipeline including leakage-safe unit-level splitting, consistent preprocessing, fixed-length window construction, and unified label generation. Inputs are transformed into 30-cycle windows with a stride of 1 cycle, and labels are assigned to the most recent time point of each window for real-time inference alignment. RUL targets are capped at 125 cycles, imminent failure is defined as RUL ≤ 30, and health-stage classification uses three stages: normal (RUL > 80), degradation (31–80), and imminent (≤ 30). Experiments on the NASA C-MAPSS turbofan engine dataset (FD002 and FD003) show that the proposed shared-encoder CNN–LSTM model learns a degradation-relevant representation that supports multiple PdM objectives within a single network. Quantitative evaluation demonstrates comparable or improved task-level performance versus single-task baselines, and time-axis case analysis further indicates that operational usability is strengthened when outputs evolve coherently as failure approaches (declining RUL, increasing risk, and consistent stage transitions). The thesis concludes with limitations and deployment considerations, including domain-dependent variability, class imbalance in near-failure regions, and the need for calibration and governance of threshold-based actions, and suggests future work on consistency-aware objectives and deployment-oriented decision policies.

Keywords: predictive maintenance, remaining useful life, imminent detection, health-stage classification, multi-task learning, time-series deep learning, CNN–LSTM, output coherence