Heuristic process mining algorithm has a significant advantage in dealing with noise and incomplete logs. However, existing heuristic process mining algorithms cannot handle long-distance dependencies and lenth-2-loop structures correctly in some special situations. Besides, none of them are parallelized. To address the problems, process models are divided into multiple case models according to executed activity set at first. Then the C-nets corresponding to case models are discovered with an improved heuristic process mining algorithm in parallel. After that, these C-nets are integrated to derive the complete process model. Meanwhile, the definition of long- distance dependencies is extended to non-local dependencies between two activity sets in decision points. In addition, a more accurate long- distance dependency metrics and its corresponding mining algorithm are presented. These improvements make the proposed algorithm more accurate and efficient.