Abstract: Since the distribution of air pollution sources is influenced by topography, landform and meteorology, the distribution of air pollution data in space is of arbitrary shapes and densities. To more accurately mine the rule of air pollution, this paper proposed a clustering model based on the DP algorithm for discovering the backbones of the cluster. The model could directly group pollution data without statistical analysis and extract key information from air pollution data by keeping the distribution unchanged, so as to excavate the change law of air pollution more accurately. The proposed clustering model and the k-Means algorithm were compared and analyzed on the three hourly pollutant concentration datasets monitored in January of 2017, 2019 and 2021 in Lanzhou respectively. In these three datasets, our model could more clearly mine the pollution data. The key pollution data accounted for 59.0%, 57.2% and 69.0% respectively in the backbones of pollution cluster, and the primary pollutants causing pollution were NO₂ and particulate matter. To reflect the applicability of the model, we analyzed our model on the pollution data in Lanzhou in January 2021, then found that the variation of air pollution in that month was caused by the joint or alternate action of pollutants NO₂ and PM₁₀, the hourly variation trend of pollution showed a bimodal pattern both on the number of contaminated hours and the occurrence frequency of primary pollutants (NO₂ and PM₁₀), and Chengguan District was the polluted area. The validity of the model was tested using the causes analysis of the above pollution laws, which made the model practical and effective for extracting key air pollution data without changing its complex distribution.