Multimodal Feature Alignment for Early Fault Identification in Smart Manufacturing Systems

Abstract

Unexpected equipment failures remain a major source of production interruptions in modern manufacturing environments. Although machine learning has improved industrial fault diagnosis, existing approaches often struggle to integrate heterogeneous sensor measurements under varying operating conditions. This study introduces a multimodal feature alignment framework that combines vibration signals, temperature measurements, and operational logs to identify emerging equipment abnormalities. A cross-modal attention mechanism captures relationships between sensor streams, while a contrastive learning objective improves the separation of normal and faulty operating states. Experiments are designed using approximately 1.26 million sensor observations collected from 48 industrial machines across six production environments. The proposed framework achieves a fault classification accuracy of 94.3% and a macro-F1 score of 0.918, compared with 87.6% accuracy and 0.854 macro-F1 for conventional deep learning baselines. Furthermore, the framework reduces average fault detection delay by 27.4% and maintains above 90% classification accuracy under moderate sensor noise. These findings suggest that multimodal representation learning could improve the reliability of intelligent manufacturing monitoring systems.