A modular deep learning methodology for multi-fault machine health diagnostics from sparse and imbalanced multimodal data

Data-driven fault detection and diagnostics for machines with multiple concurrent faults pose a complex multilabel classification challenge. Industrial condition-monitoring datasets are typically sparse, imbalanced, and multimodal, requiring meticulous processing, particularly for effective information fusion. This study address key challenges - data quality, uncertainty, and scalability - by proposing a comprehensive end-to-end methodology based on a modular mixture-of-experts (MoE) architecture. The approach encompasses data collection, preprocessing, expert training, dynamic routing, and inference. By incorporating tailored MoE with dynamic gating, the methodology enhances adaptability and efficiency of multimodal fault detection. Its effectiveness is demonstrated through application to a hydrogenerator fleet and validated by industry experts. Additionally, strategies for development under limited computational resources are provided to ensure practical implementation.