Trust Degradation in Multimodal Time-Series Predictive Maintenance Systems

Theodore Tochukwu Omeche, Joseph Azeta

Keywords: Predictive maintenance, time-series analysis, trust degradation, multimodal sensing, edge computing, confidence calibration, temporal reliability

TL;DR: High-performing multimodal predictive maintenance models can maintain accuracy yet hide temporal fluctuations and inter-sensor disagreements, so assessing stability, drift, and trustworthiness is crucial for reliable real-world deployment.

Abstract: Predictive maintenance systems are increasingly deployed on edge platforms to monitor streaming sensor data in real time. While machine learning models often achieve high classification accuracy in offline evaluations, conventional metrics fail to capture the evolution of trust and reliability during continuous deployment. This paper presents a deployment-focused empirical study of trust degradation in a multimodal time-series predictive maintenance system using temperature, vibration, and acoustic sensor streams. We introduce rigorous metrics to quantify temporal stability, confidence drift, inter-modality disagreement, and a composite Trust Degradation Index (TDI) that integrates multiple dimensions of predictive reliability. Longitudinal analyses reveal that, despite stable accuracy, cumulative confidence drift and weighted disagreement indicate silent degradation and latent reliability issues. Visualization of metric evolution over time highlights periods of vulnerability not observable through standard performance measures. These results emphasize the necessity of time-aware evaluation, continuous monitoring, and adaptive strategies to maintain trust in edge-deployed predictive maintenance systems operating under dynamic, real-world conditions.

Submission Number: 18