Measurement-Physic-Constrained Neural Network for Multifield Reconstruction of PMC

A permanent magnet coupler (PMC) is a novel noncontact coupling product transmitting torque through complex electromagnetic behaviors. However, its monitoring processes rely on finite single-point sensors, which makes it difficult to characterize the multifield states in the PMC with limited measurement information. To address the above issue, this article proposes a measurement-physic-constrained neural network (MPCNN) for the multifield reconstruction of the PMC, incorporating both the measured MF boundaries and physical knowledge. First, to integrate the limited measurement information, a coefficient decomposition method is proposed to integrate and embed governing equations and various boundary conditions into the loss function of the MPCNN by the multiorder coefficient vectors. Then, the location configuration of the Hall sensor array is optimized to obtain better measurement boundaries. A series of boundary conditions are formulated based on the homogeneous conditions and the MF data measured by the optimized sensors array. The immeasurable fields in the copper layer and the key performances are also estimated at an arbitrary slip and air gap using the MPCNN. Finally, the proposed model is verified and assessed in various aspects. The results show that the proposed MPCNN model is highly compatible with the physical governing equations and measurement boundary conditions.