This paper presents a data-driven framework for interpreting high-dimensional multivariate fire sensor data. The framework utilizes principal component analysis (PCA), data clustering, and sparse partial least squares (S-PLS) to assess the integrity of field measurements and then elucidate statistical correlations and trends using gas species and mass loss data obtained in an under-ventilated compartment fire experiment. PCA was able to reinforce the identification of faulty or incorrect measurements in sensors and reveal the temporal trajectory of the fire. Clustering methods implemented allowed the categorization of time clusters that correspond to physical phenomena (ignition, growth, fully-developed, and decay phases) while also providing practical insights into the chemical evolution and fire dynamics within complex experimental datasets. This framework provides a powerful and interpretable approach for linking sensor measurements to key fire behaviour indicators, enabling improved sensor selection and experimental design. • Propose general data science framework for characterizing multi-dimensional fire data. • Systematically identify stages of under-ventilated compartment fire. • Identify chemical species that significantly correlate with mass loss rate. • Detect faulty sensor measurements.
Mazzadi et al. (Sun,) studied this question.