Abstract: Fault diagnosis of solid rocket motor ground test systems is crucial for test data analysis and status evaluation. To achieve automatic fault diagnosis, a method integrating a Conditional Generative Adversarial Network (CGAN) with a Long Short-Term Memory network and a self-attention mechanism (LSTM-Transformer) was proposed. First, real fault signals were augmented by CGAN. Then, based on the augmented dataset, an LSTM-Transformer model was developed to detect and classify abnormal signals. Experimental results on multiple types of fault signals showed that, while satisfying real-time engineering constraints, the proposed method improved detection accuracy by 8.59% compared with conventional methods. In addition, the model exhibited strong robustness and good generalization capability, indicating promising application prospects in solid rocket motor ground tests.