Abstract: Large Language Model (LLM), especially some pre-trained deep neural networks, have strong representation learning capabilities and can automatically learn abstract features and potential error patterns from software code. Large models, with their strong natural language processing capabilities, can not only learn typical software defects but also deeply analyze hidden defects in software and provide risk warnings and improvement suggestions. The application of large models makes testing methods and processes more intelligent and can discover complex defects that are difficult to capture before. Spacecraft software has unique, specialized, complex, high reliability, and high security features, and requires specialized methods to train specialized large models to meet the needs of space embedded software testing. Research shows that a test defect set refined from a large amount of test data, combined with an effective testing method, can train a dedicated large model. Experimental results show that combining the expert's thinking with the large model and guiding the large model to find software defects using the expert's thinking chain can effectively improve code quality and defect detection efficiency, significantly improve the overall professional level of the testing team.