Hybrid Thermal Modeling with LPTN-Informed Neural Network for Multi-Node Temperature Estimation in PMSM

To achieve improved multi-node temperature estimation with limited training data in Permanent Magnet Synchronous Motors (PMSMs), a novel approach of a Lumped-Parameter Thermal Network (LPTN)-informed neural network is proposed in this paper. Firstly, the parameter and model uncertainties of third or higher-order LPTNs with global parameter identification for temperature estimation are systematically stated based on numerical analysis. Then, a two-step parameter identification strategy for a third-order LPTN with simplified thermal transfer paths is proposed to resolve parameter uncertainty. This strategy uses only air-gap structure information to make all parameters converge to their unique solutions without the need for additional geometrical parameters or material features. In response to model uncertainty, an LPTN-informed Long Short-Term Memory (LSTM) framework is designed to compensate for model unaccounted errors and extend temperature estimation nodes that the highly abstract low-order LPTN does not consider.Experimental temperature estimation results validate the effectiveness of the proposed LPTN-informed LSTM framework under a limited 23.8 hours of training data.

为在永磁同步电机（PMSM）中实现有限的训练数据下的多节点温度估计改进，本文提出了一种基于集中参数热网络（LPTN）信息的新型神经网络方法。首先，基于数值分析，系统地阐述了用于温度估计的第三或更高阶LPTN的参数及模型不确定性。随后，针对简化热传递路径的第三阶LPTN，提出了一种两步参数识别策略，以解决参数不确定性问题。该策略仅利用气隙结构信息，使得所有参数均收敛至其唯一解，无需额外的几何参数或材料特性。针对模型不确定性，设计了一种基于LPTN信息的长短期记忆（LSTM）框架，以补偿模型未考虑的误差，并扩展温度估计节点，这些节点为高度抽象的低阶LPTN所忽略。实验温度估计结果验证了所提出的基于LPTN信息的LSTM框架在有限23.8小时训练数据下的有效性。