Semi-Supervised Survival Modelling in Predictive Maintenance

Sammanfattning

The SCANIA Component X dataset is a real-world industrial benchmark that enables the investigation of common modelling challenges in predictive maintenance and reliability estimation. These tasks aim to support timely maintenance decisions by estimating the failure risk or remaining lifetime of industrial components before failure occurs. Two central challenges are irregularly sampled multivariate time series and a limited number of observed failures due to heavy right censoring. These conditions can limit the performance of purely supervised survival models, as such models depend on informative failure events during training. This thesis investigates whether semi-supervised representation learning can improve time-to-failure prediction under these real-world conditions. To address this objective, a controlled experimental design is adopted, where a semi-supervised learning framework is developed and evaluated against established, purely supervised statistical and machine learning survival baselines using three survival evaluation metrics. The study examines the extent to which the proposed framework can extract degradation-related representations that improve time-to-failure prediction, whether contrastive pre-training can learn meaningful temporal structure from censored operational trajectories, and whether this pre-training improves downstream survival modelling under heavy censoring. The results show that the proposed semi-supervised framework outperformed the evaluated baselines across all three metrics. This indicates improved risk ranking and survival probability estimation, suggesting that semi-supervised representation learning can provide useful degradation-related representations for survival-based reliability estimation in heavily censored predictive maintenance data.