Continuous Low-Power Ammonia Monitoring Using Long Short-Term Memory Neural Networks

Abstract

Accurate and continuous ammonia monitoring is important for laboratory animal studies and many other applications. Existing solutions are often expensive, inaccurate, or unsuitable for long-term monitoring. In this work, we propose a new ammonia monitoring approach that is low-power, automatic, accurate, and wireless. Our system uses metal oxide sensors which change their electrical resistance due to an induced reduction reaction with ammonia at high temperatures. Traditional methods infer the ammonia level by measuring the sensor's electrical resistance after it reaches equilibrium. Such a system consumes a significant amount of energy because reaching equilibrium requires heating the sensor for minutes. Our proposed approach does not wait for equilibrium, but tries to predict the resistance at equilibrium using the sensor's initial resistance response curve in a very short heating pulse (as short as 200ms). The prediction model is built on long short-term memory (LSTM) neural networks. We built 38 prototype sensors and a home-grown gas flow system. In a 3-month in-lab testing period, we conducted extensive experiments and collected 13,770 measurements. Our model accurately predicts the equilibrium state resistance value, with an average error rate of 0.12%. The final average estimation error for the ammonia concentration level is 9.38ppm. Given the ultra low power consumption and accurate measurements, we have partnered with cage vendors and deployed our system at two animal research facilities (NIH and Cornell University) for month-long medical trials.