Anomaly Detection in Industrial Air Conditioners in Hangars With Aircraft Spare Parts

Aircraft spare parts must be stored in rooms with cooling systems. The air conditioning outdoor unit faces harsh conditions because it is usually installed without enclosures. This can affect its functionality and expose the spare parts to high temperatures. As a result, the spare parts might lose their calibrations for aircraft use. To solve this problem, we designed an anomaly detection embedded device to recognize air conditioning malfunctions using a machine learning model. First, we took samples of healthy and damaged fans that emulate the functionality of ACs in aircraft hangars. Then, we trained several light ML models with the capability to export their inference into the microcontrollers' memory. Finally, we chose the suitable ML by defining performance metrics, such as memory footprint, execution time, and model accuracy. Our results showed that a neural network model performed better than traditional ML models, such as support vector machines or decision tree algorithms. The neural network model detects 89% of AC failures in real-trial experiments in 2.65 ms, using 254 KB of Flash and 67 KB of RAM.