predicting temperature of something thermally isolated

Question

new to data science/ML but experienced programmer in C/python for engineering/electronics.

I am trying to predict the temperature of a rotor inside an electric motor. I have a lot of data to feed into the model that should correlate to the instantaneous loss in the rotor. I also have data of the instantaneous coolant in the stator of the motor. We do not have a solid working model of the losses or the thermal circuit from the rotor to the coolant so would like to use ML to predict the final rotor temperature.

The issue I am having is that the rotor is thermally isolated and thermally capacitive, and therefore accumulates temperature over time from a known starting temperature.

most of the ML literature I have read about rely on the output to be instantaneously linked to the input dataset and not have this cumulative element to the relationship, can someone suggest an approach or technique to read around that could help with this situation?

Answer 1

It sounds like a nice project!

You should be able to borrow from three related disciplines in order to complete your modeling process: Physics (thermal dynamics), machine learning, and time series.

The final mechanics of your model (pun intended), therefore, can vary considerably based on the approach you take at each step, which should be determined by what data streams you have, what you feel is the correct structure in terms of causality or inter-dependence, and how much you deem each data stream reliable and accurate relative to the physical quantity it represents.

Specifically to your question, my hunch is that the easiest way is not to model the temperature directly, the readings of which are expected to be highly auto-correlated during a normal motor usage session, due to the thermal mass. I would opt for modeling the innovation component in the temperature time series, which (as a hunch/ as a thumb-rule) is the first difference of the temperature time series.

For a more specific answer, I feel I'd need a more detailed description of both your input features and the target variable. For example, you did not mention whether your data includes direct and reliable measurements of your target variable, namely the temperature of the rotor inside the engine.