Software crashes represent a critical form of software failure, and Crash Fault Residence Prediction is a key area in software reliability engineering. Crash reports typically contain stack traces, and by determining whether the fault lies within these traces, the prediction task can be framed as a binary classification problem. However, crashes occurring within stack traces are significantly outnumbered by those outside, creating a class imbalance that undermines prediction accuracy. To address this challenge, a conditional variational autoencoder-based data augmentation technique is proposed to generate synthetic instances for the underrepresented class (crashes within stack traces). Experimental findings demonstrate that this approach substantially enhances crash fault residence prediction performance.