Motor Temperature Virtual Sensing for a Small Scale Propulsion System

Electric propulsion systems are increasingly adopted in aerospace applications such as drones, Unmanned Aerial Vehicles (UAVs), and emerging Urban Air Mobility (UAM). They offer advantages in efficiency, reduced emissions, and mechanical simplicity compared to combustion-based propulsion, but also introduce challenges. Among the most critical is thermal management: motors can overheat under high loads or prolonged operation, leading to reduced performance, reliability issues, and safety risks. Accurate monitoring and prediction of motor temperatures is therefore essential. In testing, fiber optic temperature sensors are often used because they provide high accuracy and immunity to electromagnetic interference. However, they are costly and fragile: bending their cables can distort wavelengths, making them unsuitable for widespread or long-term use. If motor temperatures could instead be predicted from accessible electrical measurements such as currents and voltages, reliance on these delicate sensors could be reduced. This would enable safer and more economical testing while still detecting overheating risks. This thesis addresses this challenge by developing a machine learning system for virtual engine temperature sensing via standard electrical signals, which could also be used for conditional monitoring. A dedicated test bench was designed to collect a dataset including motor currents, resistance, and speeds, alongside reference data from fiber optic sensors. Several Neural Network (NN) architectures were investigated, based on Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs), and more. The sequential nature of temperature evolution motivated RNNs, while CNNs were considered for feature extraction. Within the recurrent family, Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) layers were compared with lightweight counterparts minimal LSTM (minLSTM) and minimal GRU (minGRU). The results showed that the considered architectures are suitable for virtual measurement of engine-related temperatures. For predictive purposes, however, the amount and variability of the acquired data proved insufficient to describe the full thermal dynamics. Consequently, the networks often converged on predicting average values rather than tracking actual fluctuations, limiting their usefulness for motor temperature condition monitoring. This outcome emphasizes that richer datasets with greater operating diversity and longer recording times are essential for developing condition monitoring systems. In summary, this thesis demonstrates that neural network architectures can support condition monitoring of electric propulsion systems using readily available electrical measurements. By reducing reliance on fiber optic sensors, the approach contributes to safer and more economical testing while laying the groundwork for predictive monitoring in UAVs, robotics, and electric vehicles.