Qunbo Wang

Railway scene understanding is critical for ensuring train operational safety and advancing intelligent railway systems. Existing railway vision methods mainly focus on perception and classification, while lacking regulation-guided semantic reasoning capabilities in complex environments. To address these limitations, this paper proposes a retrieval-augmented generation (RAG)-enhanced vision–language framework for railway signal cognition and safety reasoning. The proposed method integrates railway signal perception, regulatory knowledge retrieval, and multi-modal reasoning to improve factual consistency, reasoning reliability, and operational interpretability. In addition, a dedicated railway signal dataset comprising 500 standardized railway scene images with structured QA annotations is constructed to support regulation-oriented multi-modal recognition evaluation. Experimental results show that the proposed framework improves reasoning accuracy from 28.40% to 67.20% with an average end-to-end inference latency of 11.31 s per sample, and the inference speed can be further improved by adjusting experimental configurations to trade off between efficiency and accuracy, demonstrating the potential of RAG-enhanced architectures as a foundational step toward reliable multi-modal cognition in intelligent railway systems.

Although many researchers use pre-trained models to better solve downstream tasks, further exploration of more effective pre-training methods remains necessary, especially for multi-modal pre-training where high-quality training data is more difficult to obtain. This work aims to improve the knowledge-learning performance in multi-modal pre-training. Some researchers focus on injecting entity knowledge into language pre-trained models based on masked entity model (MEM) training, which masks entities randomly and lets the model recover. These methods cannot guarantee good performance due to the lack of consideration of which entities are more valuable for learning. Moreover, in multi-modal training data, some entities may be unrelated to visual content. In this work, for the vision-language pre-trained model, we propose a Masked Entity Model pre-training method based on Active learning (ActiveMEM). It is designed to actively mask important and informative entities—those that are both informative and uncertain—for the model to recover, thereby encouraging it to extract more valuable knowledge from the data. The proposed method is evaluated using three pre-training datasets and four downstream datasets, and the experimental results demonstrate the effectiveness of our method.