In recent years, there has been a rapid increase in the volume of data within the railway data service platform, encompassing multimodal data types such as images, text, and video. To enhance the use of this multimodal data for railway safety purposes, a novel cross-modal intelligent retrieval technique is proposed. This technique involves the utilization of the RegNet network to classify accident and fault images, the Lattice-LSTM model for recognizing textual data as named entities, and the BERT model for achieving mutual retrieval between images and texts by aligning retrieval words with textual semantics. Furthermore, a railway safety service platform has been developed, enabling the realization of cross-modal intelligent retrieval functionality. Through interfacing and visualization, users can conveniently access and retrieve relevant information about railway accidents and failures from the platform.