Multi-Modal Data Fusion Approach for Real-Time Monitoring of Industrial Equipment

Real-time monitoring of complex industrial equipment is crucial for ensuring optimal operation and preventing failures. However, traditional approaches that rely on single-source data, such as control system feedback or mechanical vibration signals, often struggle to capture the intricate interplay between process dynamics and equipment health. To address the challenge, this paper proposes a dynamic inner feature fusion network (DiFFNet), which is a multi-modal data feature fusion approach. It uses a unique architecture that combines coarse-grained feature extraction from process data with fine-grained feature extraction from mechanical vibration signals. Spatio-temporal attention mechanisms capture the inherent dependencies within each data mode, while a novel cross-attention mechanism with an integrated autoregressive (AR) module facilitates dynamic information exchange and modeling between the different representations. This comprehensive fusion strategy enables DiFFNet to effectively capture the complex relationships between process variables and equipment vibration patterns. Evaluations of real-world data from a mineral processing plant demonstrate that DiFFNet achieves significantly improved prediction accuracy for the control system feedback index, a key indicator of equipment operating condition, compared to single-source methods.