The present disclosure provides a method for learning emergency injection correction of an injector for preventing misfire. A misfire rate in a multi-stage injection mode is monitored, and it is controlled to be forcibly switched to a single injection mode when the misfire rate by the monitoring is equal to or greater than a specific value, such that injector injection correction learning is performed according to a learning entry condition. Accordingly, it is possible to reduce the deviation between cylinders by the sufficient injection amount deviation correction learning, thereby preventing misfire of the injector.