Estimating Electric Motor Temperatures With Deep Residual Machine Learning
Estimating Electric Motor Temperatures With Deep Residual Machine Learning
复制标题
DOI：
10.1109/tpel.2020.3045596
复制
发表时间：
2021-07-01
影响因子：
6.7
通讯作者：
Boecker, Joachim
中科院分区：
文献类型：
--
作者：
Kirchgaessner, Wilhelm;Wallscheid, Oliver;Boecker, Joachim
大多数牵引驱动应用缺乏精确的温度监测能力，通过昂贵的超大电机设计确保安全运行。经典的热建模需要模型参数选择方面的专业知识，这会受到电机几何形状、冷却动力学和热点定义的影响。此外，他们的主要优势，数据驱动的方法，这是物理可解释性，往往恶化，只要他们的自由度被削减，以满足实时的要求。在这篇文章中，深度递归和卷积神经网络（NN）与剩余连接的经验评估其在预测潜在的高动态温度连续永磁同步电机内部的可行性。在这里，定子齿、绕组和磁轭以及转子永磁体的温度分布被估算，而它们的真实值可作为测试台数据。通过贝叶斯优化和手动合并的目标估计到一个多头架构的自动超参数搜索，精益模型，表现出很强的估计性能，在最小的模型尺寸。已经发现，具有残差连接的深度递归神经网络和卷积神经网络的均方误差和最大绝对偏差性能满足经典的基于神经网络的方法，而不需要领域专业知识，也不需要特定的驱动系统规范来进行拓扑设计。最后，不同的训练集大小和解释模型估计通过预期梯度的学习曲线。
Most traction drive applications lack accurate temperature monitoring capabilities, ensuring safe operation through expensive oversized motor designs. Classic thermal modeling requires expertise in model parameter choice, which is affected by motor geometry, cooling dynamics, and hot spot definition. Moreover, their major advantage over data-driven approaches, which is physical interpretability, tends to deteriorate as soon as their degrees of freedom are curtailed in order to meet the real-time requirement. In this article, deep recurrent and convolutional neural networks (NNs) with residual connections are empirically evaluated for their feasibility on predicting latent high-dynamic temperatures continuously inside permanent magnet synchronous motors. Here, the temperature profile in the stator teeth, winding, and yoke as well as the rotor's permanent magnets are estimated while their ground truth is available as test bench data. With an automated hyperparameter search through Bayesian optimization and a manual merge of target estimators into a multihead architecture, lean models are presented that exhibit a strong estimation performance at minimal model sizes. It has been found that the mean squared error and maximum absolute deviation performances of both, deep recurrent and convolutional NNs with residual connections, meet those of classic thermodynamics-based approaches, without requiring domain expertise nor specific drive train specifications for their topological design. Finally, learning curves for varying training set sizes and interpretations of model estimates through expected gradients are presented.