From Edge Screening to Cloud Adjudication: Intelligent Multimodal Anomaly Detection in Large-Scale Sensor-Cloud Networks

Most multimodal anomaly detection methods in sensor-cloud networks rely on static fusion or cloud-centric processing, and their performance degrades under semantic conflicts or resource constraints. To address these challenges, we propose the Edge-Cloud Collaborative Mamba-based Multimodal Anomaly Detection (EC-MAD) framework, which couples lightweight Mamba-based encoders at the edge with large-scale pretrained models in the cloud. First, edge-hosted Mamba encoders independently learn image and text features, preserving unimodal specificity while maintaining low latency. Second, a cloud-hosted shared space aligns modalities and captures their interactions, leveraging pretrained models such as BERT, ClipCap, and VGG19. Third, a reliability-aware dynamic fusion mechanism adaptively reweights modalities according to calibrated predictive confidence and cross-modal agreement, suppressing unreliable or adversarial signals and allowing a single modality to dominate when required. This design enables efficient edge processing and high-accuracy cloud inference, making the framework both resource-friendly and robust for multimodal anomaly detection in large-scale sensor-cloud networks. We validate EC-MAD through comparative and ablation experiments on two large-scale public multimodal datasets. The results indicate superior or competitive accuracy and F1-scores relative to strong single-modal and multimodal baselines.