Automatic identification of factor profiles can be achieved by improved machine learning model

Abstract

The identification of factor profiles is a pivotal step in the source apportionment model. Currently, this process heavily relies on human experience, resulting in high subjectivity in the results and requiring a time-consuming procedure. In this study, a pseudo label extra trees classifier model was proposed to facilitate the automated identification of factor profiles. The source profiles serve as domain knowledge to train the model, as they accurately reflect the distinctive characteristics of emission sources. The findings indicate that the recognition rate of seven factors is 94.3%, significantly outperforming four factors (25%), five factors (30%), six factors (60%). Significantly, the model demonstrates its proficiency in determining the optimal number of factors. And the factor profiles identified using this approach demonstrate complete concurrence with manual recognition. For offline datasets, the model is also proficient at identifying factor profiles and exhibits excellent generalization. This approach facilitates the identification of emission sources in intricate environments and advances the model's capacity to automatically discern source categories by utilizing domain knowledge characteristics.