The in-depth application of information technology supported by the large-scale internet of things provides the possibility for the discovery of coal mine safety knowledge based on massive data mining. The existing researches on big data of coal mine safety are mostly based on quantitative models, which solve problems with a single perspective and neglect the multi-granularity management needs in coal mine supervision to some extent, so that the risk knowledge of coal mine contained in the data has not been found objectively and comprehensively. From the view of multi-granularity demand in coal mine supervision, with the advantage that cloud models can have a good conversion between quantitative data and qualitative concepts, a multi-granularity representation method for coal mine safety big data based on adaptive hybrid cloud transformation is proposed. This method can effectively meet the variable granularity requirements of coal mine supervision based on macro, micro, different time and space dimensions, and realize the deep mining of the characteristics of coal mine safety big data in different granularity cognitive structure. Through the application in the concept extraction of coal mine data and the comparison with the Gaussian cloud transformation algorithm, the concept coverage of the extraction is more complete and more objective, which verifies the rationality of the proposed method. In the application of coal mine monitoring data prediction, the prediction accuracy is higher than that of the ARIMA algorithm, which verifies the feasibility of the proposed method.