Software changes constantly in its lifecycle to adapt to the changing requirements and environments. In order to predict whether each change will introduce any bugs timely, various bug prediction methods for software source code changes have been proposed by researchers. However, there are three deficiencies in existing methods: 1) The prediction granularities are limited at the coarse-grained levels (i.e. transaction or file levels); 2) As vector space model is used to represent software changes, abundant information in software repositories, such as program structure, natural language semantic and history information, can not be mined sufficiently; 3) Only short-time prediction is explored without considering the concept drift caused by new requirements, team restructuring or other external factors during the long time software evolution process. In order to overcome the shortcomings of existing methods, a bug prediction method for source code changes is proposed. It makes prediction for fine-grained (i.e. statement level) changes, which reduces the quality assurance cost effectively. By in-depth mining software repositories with static program analysis and natural language semantic inference technologies, feature sets of changes are constructed in four aspects (i.e. context, content, time, and developer) and key factors that lead to bug injection are revealed. Characteristics of concept drift in software evolution process are analyzed by using matrix of feature entropy difference, and an algorithm of adaptive window with concept reviewing is proposed to achieve stability of long-time prediction. Experiments on six famous open source projects demonstrate effectiveness of the proposed method.