In view of the low accuracy, high false alarm rate and lack of real-time risk prediction capabilities caused by the current pipeline leakage monitoring system's reliance on single sensor data, this study proposes an intelligent monitoring system based on a multimodal large model to build a pipeline safety assurance system with real-time monitoring and advance prediction functions. First, a distributed sensor array and visual acquisition device are deployed to build a multi-source data acquisition network, and wavelet transform denoising and spatiotemporal alignment algorithms are used for data preprocessing. Then, a multimodal large model based on the Transformer-GCN hybrid architecture is constructed. Through self-supervised learning, 100,000 h of pipeline operation data are pre-trained, and a dynamic attention mechanism is introduced to realize multimodal feature fusion. Finally, combined with transfer learning, fine-tuning is performed on a database containing 2000 leakage cases to develop a dual-function module with real-time leakage detection and risk prediction for the next 30 min. The system achieves a detection accuracy of 98.7% and a false alarm rate of 1.2% on the test set. The risk prediction module achieves an F1 score of 93.5% for leakage events within 30 min. During the field deployment verification, it successfully warns of 5 potential leakage events, with an average advance warning time of 22 min. This multimodal large model system can provide an innovative solution for the security protection of industrial infrastructure.