Investigation of artificial neural networks for estimating important component temperatures in electric motors

基于模型的电机重要部件温度估算是近年来的一个重要研究课题。它在运行时用于部件保护，因此是提高现代驱动系统热利用程度的重要基础。通常，基于电机模型的间接方法可以用来检测温度敏感的参数变化，从而观察工作温度。这种模型有一个固有的缺点，即不能估计所谓的热点温度，也不能检测到一些重要的发动机温度(如滚珠轴承)。另一方面，泵浦参数热网络(LPTN)经常被用来直接模拟温度分布。关于LPTN建模深度有很大的范围。对于具有实时能力的模型，特别是抽象的方法被用来估计相关的电机温度，其模型结构尽可能紧凑，从而在计算上高效。这里的问题是，替代的模型方法是否不能更好地适用于复杂系统中对温度估计的经验抽象访问。通过避免基于LPTN的微分方程，可以独立于要建模的零部件的数量来选择模型的自由度数。人工神经网络(KNN)包含了多种不同的黑盒模型，并且已经在语音和图像分析中得到了广泛的应用。本申请完全忽略了这些问题，因此，申请人对KNN的基本适宜性进行了初步调查。这里可以看出，使用长期-短期记忆或门控循环单元的拓扑结构提供了有希望的估计精度。然而，它们仍然低于既定的直接和间接方法。在这次调查中，已经确定了许多开放的研究问题，例如超参数优化问题。这些是所考虑的KNN拓扑(例如，隐藏层的数目或每层的神经元数目)或所使用的训练算法(例如，KNN权重的初始化)的最高配置参数。因此，该项目的目的是系统地研究KNN用于电机温度估计，从而开发出一种通用的方法和过程链，以便项目成果可以直接转移到相关的技术系统，例如电池或电力电子转换器。

The model-based estimation of important component temperatures in electric motors is an important research subject in recent years. It is used for component protection during run-time and thus forms an important basis for increasing the degree of thermal utilization regarding modern drive systems. Typically, indirect methods based on electrical motor models can be used to detect temperature-sensitive parameter changes and thus to observe the operating temperature. This model class has the inherent disadvantage that the so-called hot-spot temperature cannot be estimated, and that some important engine temperatures (e.g., ball bearings) cannot be detected. On the other hand, pumped parameter thermal networks (LPTN) are frequently used to directly model the temperature distribution. There is a wide range with regard to the LPTN modeling depth. For real-time capable models, particularly abstracted approaches are used to estimate the relevant motor temperatures with a model structure that is as compact as possible and thus computationally efficient. The question here is whether alternative model approaches are no better suited for empirical-abstract access to temperature estimation within complex systems. By avoiding LPTN-based differential equations, it is then possible in to select the number of degrees of freedom of the model independently of the number of components to be modeled. As a result, the estimation accuracy against abstract LPTN approaches could be further increased with a reasonable additional calculation effort.Artificial neural networks (KNN) comprise a wide range of different black-box models and already find a wide range of applications, e.g. in speech and image analysis. These have been completely neglected for the present application, so that the applicant has carried out an initial preliminary investigation into the basic suitability of KNN. Here it could be shown that topologies using long-short-term-memories or gated recurrent units provide promising estimation accuracy. However, they are still below the established direct and indirect methods. In this investigation, numerous open research questions have been identified, e.g. the problem of hyper-parameter optimization. These are superordinate configuration parameters of the considered KNN topology (for example, the number of hidden layers or number of neurons per layer) or the training algorithm used (for example, initialization of the KNN weights). This project is therefore aimed at the systematic investigation of KNN for the temperature estimation in electric motors, whereby a general methodological and process chain is developed so that the project results can be directly transferred to related technical systems, e.g. to batteries or power-electronic converters.