Data Acquisition System for AI-based Odor Classification at the Edge

Air quality in kitchens and adjoining rooms plays an important role for well-being at home. Through the incorporation of olfactory information, the automation of kitchen ventilation systems may be achieved. Considering cost constraints in consumer applications, Volatile Organic Compounds (VOCs) sensors are preferred over expensive gas sensors as a more cost-effective option. Since no comparable dataset for this application existed, a custom dataset was recorded using a specifically developed data logger and labeling tool. A sensor node combines six different sensor types which were used to record cooking situations in eight apartment kitchens with a total of 18 GB of measurement data (VOC, temperature, relative humidity, scattered light). As an initial attempt, a simple Artificial Intelligence (AI) odor classifier was developed. It can distinguish between four different odors by using 1D-Convolutional Neural Networks (CNNs) on time series. The model was quantized to 8 Bit after training and was ported to an Arm Cortex-M4 microcontroller platform. It performs with an F1-score of over 55% (median) in unseen kitchens, requires 183 KB of RAM memory and takes 35 s of inference time.