Towards Similar User Utterance Augmentation for Out-of-Domain Detection