Deep Learning-Based Predictive Maintenance Framework for Marine Propulsion Systems Using Multi-Sensor Data Fusion

Abstract

In this study, multi-sensor data fusion is being used for a deep learning-based predictive maintenance framework of marine propulsion systems. So as to capture complementary fault characteristics under simulated marine operation conditions, vibration, acoustic and thermal signals were considered. Signal synchronization, noise modeling, denoising and feature level fusion based methodology has been developed. This multi-domain signal analysis revealed significant overlap among individual sensing modalities. That limits the standalone diagnostic capability, to make it more robust and to evaluate classification performance across six fault conditions, deep learning models are trained for single modality and fused feature inputs as well. Results shows that all the three models fall in range of 47-59% accuracy whereas the fusion model achieved 90%. Which in turn confirmed using repeated validation as the fusion model achieved 88.52% ± 4.01% showing reduced variability. After which ablation study is followed which demonstrated a progressive performance improved from single to dual and full fusion configurations. These findings undermines the need of multi sensor integration for reliable fault diagnosis and a scalable framework for future predictive maintenance.