Reliability is an important index to measure the quality of software. Online predicting and improving software reliability are an important research topic. Most existing methods have the following weakness: They can neither predict software reliability on different time intervals nor locate the faulty components that cause the declining of the reliability. This paper proposes a new method to online improve reliability for service composition. The method uses the monitored failure data at ports to predict the reliabilities of service composition on different time intervals. If the predicted reliability is lower than the expected value, it then locates the faulty component that causes the declining of the reliability by using an improved spectrum-fault-localization method. The system is automatically reconfigured by adding a new component or replacing the faulty component to improve the system reliability. An online shop example is used to demonstrate the effectiveness of the proposed method.