A Coattention Enhanced Multimodal Feature Fusion With Inner Feature for Anomaly Detection

Current anomaly detection methods struggle with the nonlinear, dynamic, and multisource nature of industrial processes. This article proposes a novel end-to-end architecture with multimodal data dynamic inner feature fusion for anomaly detection in high-pressure grinding rolls (HPGRs). We employ a dual spatiotemporal autoencoder (AE) to extract features from both production process data and vibration signals. A coattention mechanism and a coarse-fine-grained feature fusion module enhance the model's ability to capture multimodal feature interactions and recover lost manifold information. An embedded autoregressive network module extracts consistent dynamic feature representations, further improving the AE's ability for interactive multimodal feature fusion. Finally, we propose a new dynamic inner feature fusion anomaly detection method specifically designed for nonlinear dynamic processes with multimodal data. The effectiveness of the proposed method is validated using real-world HPGR production process data.