Towards a physics-informed network paradigm with data generation and background noise removal for different distributed acoustic sensing applications.

Distributed acoustic sensing (DAS) has attracted considerable attention across various fields, and artificial intelligence (AI) technology plays a vital role in DAS applications for event recognition and denoising. Existing AI models require real-world data (RWD), whether labeled or not, for training, which is contradictory to the reality of limited available event data in practical scenarios. Here, a physics-informed DAS neural network paradigm is proposed, which eliminates the need for real-world event data during training. By physically modeling the target events along with real-world and DAS system constraints, physical functions are derived to train a generative network for the synthesis of DAS event data. A DAS noise-removal network is then trained using the generated data to effectively eliminate background noise in DAS measurements. The effectiveness of the proposed paradigm is demonstrated in two applications: event identification based on a public DAS spatiotemporal dataset, and belt conveyor fault monitoring based on DAS time-frequency data. In both cases, the paradigm achieves comparable or superior performance to data-driven networks trained with RWD. Owing to the incorporation of physical information and the ability to remove background noise, the proposed approach shows strong generalization capability across different sites within the same application. Notably, a fault diagnosis accuracy of 91.8% is achieved in a real belt conveyor field using networks transferred from a simulation test site, without using any fault event data from the target field during training. The proposed paradigm offers a potential solution to the critical challenges of limited data availability and intense noise in practical DAS applications.