ASC-PIT: Affine–Spectral Coordinated Physics-Informed Transformer for offline PMSM commissioning
Year: 2026
Type: article
Abstract: Accurate maximum torque per ampere (MTPA) mapping for interior permanent magnet synchronous machines (IPMSMs) is fundamentally constrained by the trade-off between electromagnetic fidelity and data efficiency under sparse commissioning measurements. While physics-informed neural networks (PINNs) embed d – q frame residuals to enforce physical consistency, they systematically under-represent the geometric structure of the operating manifold, especially when ground-truth target labels are scarce. This work proposes ASC-PIT, an Affine–Spectral Coordinated Physics-Informed Transformer, which integrates a dual-stream manifold encoder for explicit cross-motor geometric alignment, a Mixture-of-Experts feed-forward sublayer for regime-adaptive capacity allocation, and a hybrid physics–data head with soft d – q regularization. Evaluated on two public IPMSMs (5.2 kW, 2.2 kW) under a few-shot transfer protocol, the Transformer-based pipeline reduces torque mean absolute error by approximately one order of magnitude relative to a matched MLP-PINN baseline. This accuracy gain is predominantly driven by the PIT-STN-style encoder with sparse target adaptation, rather than manifold encoding and MoE routing alone. Under multi-seed evaluation, ASC-PIT achieves statistically equivalent performance to PIT-STN (paired p = 0.224 , 0.089). The architectural redesign adds value through interpretable manifold alignment and quantifiable expert routing, enabling offline IPMSM commissioning with inference relying solely on electrical angle, currents, and voltage-derived signals.
Source: Measurement
Authors: Gelin Wang, Junjie Niu, Xiao Ying, Xinmin Chen
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,760
OpenAlex ID: W7214426235