Investigation of artificial neural networks for estimating important component temperatures in electric motors
Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics
Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
Final Report Abstract
The project’s objective was the investigation of artificial neural networks (ANNs) for the data-driven thermal modeling of electric motors subject to a test bench measurement data set, which was to be recorded as well. An algorithm for the offline motor excitation profile generation was developed, which would cover the motor speed and torque operation range as uniformly as possible. Utilizing this algorithm, a 140 hour long sensory data set was recorded at a lab test bench, which, together with a previously collected data set, represents a 185 hour data set and the base for all following investigations. The data was published online without access constraints nor anonymization. On the basis of this data set, several experiments with different ANN architectures and alternative ML methods were conducted. All experiments included a thorough cross-validation according to best practices for a fair comparison. Moreover, comprehensive hyperparameter optimizations were deployed on distributed systems for parallel execution. The superiority of linear regression and ANN-based topologies over alternative ML methods were identified. However, given the empirical results, no unconditional recommendation over classic heat-transferbased lumped-parameter thermal networks (LPTNs) could be expressed yet. Noteworty drawbacks were high amounts of model parameters; a well-conditioned measurement data set; the unadjustable initial estimate value; mandatory additional online moving average calculations; as well as no physically interpretable model states. These are accompanied by the advantage that no expert knowledge is required for the design as thermal model. Surprisingly, in the course of the project work, synergies between ANNs and the LPTN approach have been identified, which were unified into a novel hybrid model type. This data-driven modeling approach was coined ”thermal neural network”(TNN), and it was designed to be applicable to any system for which a LPTN would also be an adequate choice (that is, systems that can be modeled under the assumption of lumped parameters) and for which measurement data is available. Having met these requirements, the TNN can be recommended unconditionally since in contrast to LPTNs neither material nor geometry information about the system is required. Moreover, compared to ML methods that do not inherit differential equations, a higher estimation performance with less model coefficients is achievable, model states can be physically interpreted, setting the initial estimate value is trivial, and no moving averages are required. Behind the principle of the found hybridization procedure, which is leveraged by automatic differentiation frameworks, the project members see a seminal new approach to identification of nonlinear dynamic systems in general. Interesting transfer opportunities to neighboring application and research domains might be revealed during further research. All developed algorithms are published on GitHub as Python code, in order to assist related research.
Publications
- “Deep Residual Convolutional and Recurrent Neural Networks for Temperature Estimation in Permanent Magnet Synchronous Motors”. In: IEEE International Electric Machines Drives Conference. 2019, S. 1439–1446
W. Kirchgässner, O. Wallscheid und J. Böcker
(See online at https://doi.org/10.1109/IEMDC.2019.8785109) - “Empirical Evaluation of Exponentially Weighted Moving Averages for Simple Linear Thermal Modeling of Permanent Magnet Synchronous Machines”. In: Proceedings of the 28th International Symposium on Industrial Electronics. 2019, S. 318–323
W. Kirchgässner, O. Wallscheid und J. Böcker
(See online at https://doi.org/10.1109/ISIE.2019.8781195) - Estimating Electric Motor Temperatures with Deep Residual Machine Learning”. In: IEEE Transactions on Power Electronics 36.7 (2020), S. 7480–7488
W. Kirchgässner, O. Wallscheid und J. Böcker
(See online at https://doi.org/10.1109/TPEL.2020.3045596) - “Data-Driven Permanent Magnet Temperature Estimation in Synchronous Motors with Supervised Machine Learning: A Benchmark”. In: IEEE Transactions on Energy Conversion 36.3 (2021), S. 2059–2067
W. Kirchgässner, O. Wallscheid und J. Böcker
(See online at https://doi.org/10.1109/TEC.2021.3052546) - “Learning Thermal Properties and Temperature Models of Electric Motors with Neural Ordinary Differential Equations”. In: 2022 International Power Electronics Conference (IPEC- Himeji 2022- ECCE Asia). 2022, S. 2746–2753
W. Kirchgässner, O. Wallscheid und J. Böcker
(See online at https://doi.org/10.23919/IPEC-Himeji2022-ECCE53331.2022.9807209) - “Thermal neural networks: Lumped-parameter thermal modeling with state-space machine learning”. In: Engineering Applications of Artificial Intelligence 117 (2023), S. 105537
W. Kirchgässner, O. Wallscheid und J. Böcker
(See online at https://doi.org/10.1016/j.engappai.2022.105537)
