Many researches of data mining have paid close attention to multi-scale theory. However the study of multi-scale data mining still comes short on universal theories and approaches. To overcome this limitation, this paper conducts a study of universal multi-scale data mining on theoretical and methodological aspect. First, the paper lays out the definition of data-scale-partition and data-scale based on concept hierarchy, and characterizes the relationship of upper-layer and lower-layer datasets between multi-scale datasets. Next, it illustrates the definition and essence of multi-scale data mining, and presents the classification of multi-scale data mining methods. Finally, it introduces the algorithm framework and its theoretical basis of multi-scale data mining, and proposes an algorithm named MSARMA (multi-scale association rules mining algorithm) to realize the transition of knowledge in multi-scale data expressions. Experiments are carried out to test MSARMA with the help of IBM T10I4D100K dataset and demographic dataset from H province, and the results indicate that MSARMA is effective and feasible with better coverage rate, better accuracy and lower average support error.