This article deals with highlighting some pragmatic approaches for determining the key-parameters that one has to monitor in order not to jeopardize the fulfillment of the customer specifications, due to the tolerances, that occur in large-scale mass production. For that, two parametrization models (symmetric vs. asymmetric) and two types of parameter distributions (uniform vs. normal) are compared when considering lamination, stack assembly and material properties tolerances. Results in different scenarios for both the motor (in open-circuit and under load), and the whole drive (with and without position sensor detection errors) are provided and compared with end-of-line measurements (where available) for a 12-slots 10-poles IPM-type PMSM with fractional slot concentrated winding. After including the effect of the temperature in the full asymmetric model, rather good agreements between calculations and measurements are reported for back-EMF and torque constant. Furthermore, it is shown that the interaction between the uncertainties of the motor parameters and the rest of the drive's components (e.g. position sensor) cause large inaccuracies into the system, and they are hardly predictable by taking the individual deviations separately. Finally, towards the end, attempts are made to reduce the computational and the modeling effort using various techniques and ultimately fitted artificial neural networks (ANNs) as surrogates.

中文翻译：

---

通过 IPMSM 的敏感性和统计分析对大规模批量生产制造公差的鲁棒性

本文重点介绍了一些实用的方法，用于确定必须监控的关键参数，以免因大规模批量生产中出现的容差而危及客户规范的实现。为此，在考虑层压、堆叠组件和材料特性公差时，比较了两种参数化模型（对称与非对称）和两种类型的参数分布（均匀与正常）。提供了电机（开路和负载下）和整个驱动器（有和没有位置传感器检测错误）的不同情况下的结果，并与 12-带分数槽集中绕组的 10 极 IPM 型 PMSM。在完全非对称模型中包括温度的影响后，报告了反电动势和扭矩常数的计算和测量之间相当好的一致性。此外，它表明电机参数的不确定性与驱动器的其余组件（例如位置传感器）之间的相互作用会导致系统出现很大的不准确度，并且很难通过单独获取各个偏差来预测它们。最后，最后，尝试使用各种技术和最终拟合的人工神经网络 (ANN) 作为代理来减少计算和建模工作。结果表明，电机参数的不确定性与驱动器的其余部件（例如位置传感器）之间的相互作用会导致系统出现很大的不准确度，并且很难通过单独的偏差进行预测。最后，最后，尝试使用各种技术和最终拟合的人工神经网络 (ANN) 作为代理来减少计算和建模工作。结果表明，电机参数的不确定性与驱动器的其余部件（例如位置传感器）之间的相互作用会导致系统出现很大的不准确度，并且很难通过单独的偏差进行预测。最后，最后，尝试使用各种技术和最终拟合的人工神经网络 (ANN) 作为代理来减少计算和建模工作。