The goal is to parameterize neural networks entirely on empirical data, exploiting their rich expressive power without the need of motor sheet data.
Having test bench data available, estimation accuracy on real-world data is precisely reportable, and has been investigated for the first time in literature by .
The entire process from data acquisition over model training up to temperature monitoring in the field is sketched in Fig. 1.
More specifically, in the field of sequence learning tasks, recurrent neural networks (RNNs) and convolutional neural networks (CNNs) denote the state of the art in classification and estimation performance.
Their topology is depicted in Fig. 2.
They were studied in  and showed superior performance in terms of common regression metrics.
However, they exhibit hundreds of thousands of trainable parameters, which render a lean application on best-cost embedded hardware infeasible so far.
Across the domain of supervised machine learning, linear regression has also been shown to perform competitively  with the advantage of running on very few parameters, albeit not as few as classical physics-based models.
Apart from linear regression and neural networks, data-driven supervised learning algorithms, that stem from the computer science field of pattern recognition from the last decades, were displayed to not achieve sufficient estimation accuracy .
Another shortcoming of the so far investigated approaches is the dependency on exponentially weighted moving averages and standard deviations of definite filter strengths, that are expected to be suboptimal in some operation points.
The project is expected to be concluded with an innovative neural-network-based architecture, that incorporates domain knowledge and combines the benefits of both worlds, data-driven black box models from the machine learning regime and thermodynamic white- or grey-box models, such that a steady estimation on an embedded traction drive control system is feasible in real-time.