Non-Vibration-Based Multimodal Anomaly Fusion with Temporal Persistence Modelling for Degradation Assessment in Industrial Systems

Abstract

Machine failure prediction in industrial systems tends to depend on vibration analysis. Most installed machines, however, do not have vibration or acoustic sensors, leaving alternative prediction methods using proxy sensor readings like temperature, pressure, and environmental data. This work introduces a new framework for multimodal degradation assessment involving anomaly detection and temporal persistence modeling. Through the utilization of time-series sensor data, we extract features representing deviations from an ideal running state and monitor their persistence across time, similar to health decline in living systems. In contrast to conventional methods, which depend on explicit failure signifiers, our method targets long-term anomaly trends and their compounding effects. We compare several machine learning models, such as Random Survival Forest, Isolation Forest, and Recurrent Neural Networks, with results showing that failure prediction accuracy greatly improves when engineered temporal anomaly features are used. The envisioned approach facilitates more powerful predictive maintenance practices applied in actual industry settings, eliminating surprise downtimes and maximizing running efficiency.