Abstract: Industrial safety monitoring, as a core link in intelligent manufacturing, its technical level is directly related to production safety and efficiency. The traditional single-modal perception system has problems such as low reliability and high false alarm rate, making it difficult to meet the precise monitoring requirements in complex industrial environments. Based on this, an industrial safety intelligent perception system based on multi-modal fusion is proposed. By integrating four types of modal data: visual, acoustic, thermodynamic and vibration, and combining them with deep learning algorithms, an industrial safety monitoring system with high robustness is constructed. The system adopts a multi-model data fusion strategy to achieve the fusion of data level, feature level and decision level, and introduces the attention spatio-temporal graph convolutional network ( AST-GCN) realizes abnormal diagnosis and fault early warning. The experimental results show that the anomaly detection accuracy rate of the system in complex industrial environments reaches 98.7% , and the system response time is within 200ms. It can significantly improve the reliability and real-time performance of industrial safety monitoring and provide key technical support for intelligent manufacturing.