Dynamic Anomaly Detection: A Multimodal Spatiotemporal Cross-Fusion Transformer Approach

Accurate fault diagnosis is critical for the reliable and secure operation of modern industrial equipment; however, traditional unimodal data representations often fail to fully capture the complex, multi-dimensional characteristics of mechanical faults. To address this limitation, this paper proposes CFNET, a novel multimodal cross-fusion recognition framework designed for the robust fault diagnosis of rotating machinery, including vane pumps and bearings. The proposed framework employs an innovative three-modal input strategy that systematically extracts deep spatial and temporal features using parallel ResNet50 and stacked GRU networks. To break through the limitations of conventional decision-level fusion, CFNET introduces a dynamic mid-layer cross-fusion mechanism driven by adaptive attention weights, effectively enhancing feature interaction depth while preventing redundancy during back-propagation. Furthermore, handcrafted prior features are integrated via skip connections to supplement the deep learning representations. The aggregated multi-dimensional features are ultimately processed by a Transformer architecture for feature sparsification and high-precision classification. Extensive experiments rigorously validate the superiority of the proposed method. On a custom-built industrial vane pump test rig, CFNET achieved a maximum accuracy improvement of 25% over traditional unimodal methods. Furthermore, it demonstrated state-of-the-art performance on two authoritative benchmark datasets, achieving impressive recognition accuracies of 98.12% and 99.92% on the Paderborn University and Case Western Reserve University bearing datasets, respectively. These results underscore the exceptional robustness, efficiency, and industrial applicability of the CFNET framework.