Control rod drive mechanisms (CRDMs), as critical actuators in nuclear power plants, regulate control rod motion by manipulating multiple coil currents through electromagnetic-mechanical coupling. Their operational integrity is vital for ensuring reactor safety. However, existing anomaly detection approaches predominantly model individual coils, neglecting the dynamic interactions among coils and the evolving operational patterns. To address this gap, we propose a multi-coil joint learning (MCJL) model for accurately detecting latent anomalies during control rod operation. In this approach, coil nodes and fully connected edges are defined, and a decay adjacency matrix is introduced to construct a multi-coil interaction graph that captures the dynamic coupling among the lift, moving, and solid coils. A moving graph convolutional network is then employed to efficiently extract local temporal dependencies across coils while jointly reconstructing the current signals of all three coils. The residuals between the reconstructed and actual signals are subsequently calculated, enabling per-cycle anomaly detection using a multi-scale dynamic strategy. Experimental validation using historical data and simulated anomalies from a pressurized water reactor demonstrates that the proposed method effectively captures dynamic inter-coil coupling and temporal patterns, achieving high-precision signal reconstruction. Compared with existing methods, MCJL exhibits superior performance in both reconstruction and anomaly detection. Furthermore, its dynamic thresholding strategy provides flexible decision boundaries and strong fault tolerance.