행렬 분해 기법이 PCB 오염원 기여도 분석에 미치는 영향: EPA PMF와 NMF 비교

Abstract

The effect of matrix factorization methods on PCB source apportionment was examined using two independent PAS datasets (D1 and D2) from the multi-industrial city of Ulsan to assess reproducibility in relation to rotational ambiguity and user subjectivity. Positive matrix factorization (PMF) and non-negative matrix factorization (NMF) resolved the same number of factors for both datasets (D1: 26 congeners × 19 sites, 6 factors; D2: 34 congeners × 20 sites, 5 factors), indicating consistent resolution of source complexity. When PMF was applied to D1, the resolved source profiles corresponded to legacy Aroclor and Kanechlor technical mixtures and re-volatilization processes. In contrast, when applied to D2, which comprised samples collected primarily from industrial and harbor environments, PMF resolved combustion-related sources with clear factor separation. NMF identified similar source structures but showed greater sensitivity to overlapping congener patterns in the mixed urban-suburban dataset (D1), where improved separation of shared signatures was observed. These results are consistent with findings from previous receptor modeling studies, in which PMF has been widely applied for source apportionment when uncertainty information is available. NMF provides a complementary approach for identifying congener patterns, particularly in complex environments influenced by both legacy sources and combustion emissions. The choice of method influenced the resolution of overlapping signals, while the overall source structure remained consistent across both approaches.