An Optimized Boundary to Characterize the Relationship between Training Data Quantity and Neural Network Accuracy in Describing Variable Parameters of PMSMs

In recent years, there is a growing interest in utilizing neural networks (NNs) to model the motor variable parameters, due to their excellent performance in nonlinear fitting. However, it is quite difficult to precisely establish a boundary to characterize the relationship between training data quantity and NN accuracy, since its coefficient depends on the intricate coupling between motor characteristics and trained NNs. As a result, the existing boundary often tends to be conservative, thus leaving room for potential savings regarding redundant training data. Accordingly, this paper proposes a coefficient estimation method, by utilizing prior accuracy tests conducted with a limited quantity of data. Notably, this approach obviates the demand for knowledge pertaining to the structural and material characteristics of motors. The primary objective of this paper is to establish a more specific boundary, thus reducing data acquisition costs. The effectiveness of the proposed method has been validated experimentally.