Livestock facilities commonly generate NH3, a hazardous substance that may also harm livestock. Therefore, monitoring of NH3 concentrations in livestock facilities is necessary to ensure proper control. However, NH3 is alkaline and toxic, causing corrosion inside detection sensors and making monitoring difficult. This study proposes a virtual sensor concept to complement the durability of NH3 physical sensors. The study also conducts a long-term performance validation of a data-driven NH3 concentration prediction model. Results indicate that the model's prediction performance declines sharply when the data generation pattern inside the livestock facility changes due to changes in outdoor conditions and facility operation. Furthermore, the prediction performance of the model differed depending on the training data period settings when updating the model. Hence, the model needs versioning and update management to respond to the data generation pattern in the livestock facility when operating the NH3 concentration virtual sensor. The virtual sensor is expected to enhance monitoring and reduce sensor management costs in livestock facilities.
Keywords: Data-driven model; Livestock; Model update; NH3; Virtual sensor.
© 2023 The Authors. Published by Elsevier Ltd.