Online Identification of Nonlinear Flux Linkages Using Neural Networks for Highly Utilized PMSMs

Dynamic model-based control and efficient operation of highly utilized permanent magnet synchronous motors (PMSMs) require detailed knowledge of the magnetic behavior of the machine. Due to sparse yet efficient usage of iron and magnet material, the flux linkages of such PMSMs are highly nonlinear. To model the nonlinearities of the machine exactly usually huge lookup tables (LUTs) containing the flux linkages are used, which need to be obtained by finite element analysis (FEA) or by extensive measurement.As a lean and adaptive alternative, this paper proposes a data-driven method which uses a multilayer perceptron (MLP) neural network model, mapping current to flux linkage, which is trained from measurement and control values processed by a model-based open loop current controller. As no large lookup tables are required, this implementation is especially memory efficient.The proposed identification method is validated by simulation and dynamic tests on a test bench.

I. Introduction

Reduction of iron material in PMSMs decreases cost, size and weight. However, this also results in motors featuring intense nonlinear magnetic properties and strong cross-saturation. Highly dynamic model-based current control of these motors requires precise electromagnetic models of the machine which need either parametrization from FEA or extensive measurements. To achieve accurate machine modelling taking into account manufacturing tolerances, degradation, parameter drift and ambient conditions, online identification methods are desirable. Beyond precise model based control, online identified and tracked models also enable condition monitoring applications as well as self-commissioning of the electric drive system.