Enhancing industrial prognostic accuracy in noisy and missing data context: assessing multimodal learning performance

Prognostics and Health Management (PHM) is crucial for the smooth operation of manufacturing systems, and prevention of potentially costly downtime. While data-driven PHM techniques have gained popularity, their effectiveness in industrial application hinges on data quality. Multimodal learning, integrating insights from various data sources, shows promise in handling real world challenges such as noisy or incomplete data. However, existing literature lacks an in-depth exploration of noise and data incompleteness effects on multimodal-based prognostics, particularly in industrial contexts. This study aims to address this gap by analyzing prognostic performance in varying data quality conditions, aiming to guide practitioners in optimizing data strategies. Through a comparative evaluation, involving different multimodal and unimodal learning models under diverse data conditions, this research offers a potential framework, for advancing the prediction of degradation levels in noisy and missing condition-monitoring data scenarios.