Robust Evaluation for Multi-Task Predictive Maintenance Models.

Multi-task deep learning models, which simultaneously perform fault classification and remaining useful life (RUL) regression, are increasingly vital in predictive maintenance. However, this research highlights a critical issue: reported performance can be significantly skewed by how sliding-window sequences are partitioned into training and test sets. The study demonstrates that a naive splitting approach can artificially inflate classification accuracy to nearly perfect levels (e.g., 99.9%) or, conversely, reduce it to zero due to poor class representation. To counter this, the researchers introduce a "chunk-based, leakage-audited splitting protocol" and rigorously evaluate models using multiple seeds, ANOVA, and Tukey HSD tests. Using this robust protocol on the NASA C-MAPSS dataset, an attention-enhanced multi-task architecture (AMTLNet) achieved strong classification accuracy and R2 for RUL, matching or outperforming baselines. However, on smaller datasets (Bearing and Hydraulic), multi-task training proved unstable, with classification degrading for one and regression for the other, an asymmetry linked to label provenance. Ablation studies further showed that the multi-head attention branch is crucial for regression stability, while the convolutional branch contributes less despite its parameter count. The paper provides a practical framework for determining when joint training is appropriate under data scarcity and offers a reusable leakage-audit protocol for reliable evaluation.