AI(ML & NLP)-driven mapping of sensor data to standard classes : reducing resource consumption and emissions

Abstract

The increasing prevalence of AI in smart building management necessitates efficient data labelling systems to identify vast amounts of incoming data streams. This work presents the development and implementation of an AI pipeline that integrates Random Forest and BERTje models for accurate and efficient data labelling for buildings’ Heating, Ventilation, and Air Conditioning (HVAC) systems. The pipeline first employs the Random Forest model to classify data points into valid and invalid points, followed by the BERTje model for labelling relevant data points to a standardized labelling convention (BRICK schema). This method achieved an average match accuracy of 95%, surpassing the initial 90% target. A comparative analysis of carbon emissions revealed that the AI pipeline significantly reduces long-term emissions compared to human labelling methods. The findings underscore the potential of using large language models (LLMs) in enhancing the functionality and sustainability of building management systems. Future research is recommended to further optimise the pipeline and explore its applicability in other domains.