In the mobile industrial Internet of Things, dynamic device topologies and high-precision camouflage attacks present substantial challenges for authentication schemes that rely solely on physical or behavioral layers. These traditional methods exhibit low accuracy and instability in the face of channel variability and sophisticated behavioral mimicry. To overcome these limitations, we propose a novel cross-layer and multi-center fusion authentication framework that integrates both physical and behavioral features. By leveraging a multi-head cross-attention mechanism, our framework dynamically aligns heterogeneous features and captures deep temporal associations between physical signals and behavioral patterns. Furthermore, the multi-center architecture employs Bayesian reliability estimation to dynamically weight inference results from individual nodes, thereby mitigating the influence of low-quality data. This approach offers a robust solution for anti-camouflage authentication of mobile devices in complex industrial environments. Experiments demonstrate that our framework achieves robust authentication accuracy, maintaining stable performance even under severe camouflage attacks and partial data loss conditions.