A ship software performance defect detection method based on pre-trained models
Peng Cui, Ruihu Zhang, Dianwen Liu, Xinda Li · 2025
Shipborne software plays a critical role at military strategic and tactical levels. Its performance directly impacts decisionmaking efficiency in naval warfare, the effectiveness of combat execution, and the overall operational effectiveness of combat forces. Therefore, the performance of shipborne software is a key factor in ensuring maritime power gains an advantage within complex and rapidly changing battlefield environments. Attention to and investment in shipborne software performance are essential in military planning, technological development, and tactical training.However, due to the skillsets and workloads of ship software developers, the utilization of computational resources and user response times often fail to reach optimal levels, resulting in performance defects within the software.Current methods for detecting performance defects in shipborne software predominantly rely on defining and extracting code metrics, combined with machine learning classifiers for defect identification. However, these approaches face challenges such as limited semantic understanding of source code and suboptimal detection effectiveness.To address these issues, this paper proposes a pretrained model-based method for detecting shipborne software performance defects. This method involves source code preprocessing, model fine-tuning and evaluation, and performance defect detection. Leveraging pre-trained models allows for superior extraction and modeling of deep semantic features within source code and facilitates a better understanding of domain-specific knowledge related to performance defects, thereby significantly improving the effectiveness of performance defect detection. The proposed method was experimentally validated on a public dataset. The results demonstrate that our method achieves a 4% improvement in AUC and a 75% improvement in PR-AUC compared to traditional machine learning algorithms.