Physics-Informed Networks for Data Generation and Noise Removal in Distributed Acoustic Sensing

Distributed acoustic sensing, or DAS, is moving closer to becoming an intelligent nervous system for infrastructure, transportation and industrial facilities. A new study published in Light: Science & Applications introduces a physics-informed neural network paradigm designed to recognize acoustic events even when real-world training data are scarce and environmental noise is overwhelming. The approach could help transform ordinary fiber-optic cables into large-scale monitoring systems capable of detecting faults, disturbances and potentially dangerous events over long distances, without depending on vast collections of labeled field measurements.

DAS works by sending laser light through an optical fiber and analyzing tiny changes in the light that are caused by vibrations along the cable. Because the fiber can extend for many kilometers and can be installed beside pipelines, railways, roads, conveyor systems or other structures, it functions as a continuous array of virtual acoustic sensors. The technology can detect vibrations produced by machinery, impacts, ground movement and human activity. Yet the same sensitivity that makes DAS powerful also creates a major obstacle: the system records everything. Wind, vehicles, machinery, construction, electrical equipment and other background sources can obscure the acoustic signatures that operators actually need to identify.

The problem becomes particularly difficult when artificial intelligence is used for event recognition. Conventional neural networks generally require large quantities of representative data, and the most valuable examples are often the hardest to obtain. A pipeline failure, a serious structural fault or an unusual industrial incident may occur only rarely, making it impractical or unsafe to wait for enough real-world examples. Even when data are available, they may be difficult to label accurately. A model trained on one site can also perform poorly at another because the fiber installation, machinery, geology and background noise are different. These limitations can make data-driven DAS systems expensive to develop and slow to deploy.

The research team, led by Professor Zuyuan He of the State Key Laboratory of Photonics and Communications in the Department of Electronic Engineering at Shanghai Jiao Tong University in China, addressed the challenge by embedding physical knowledge into the learning process. Their system combines a physics-informed generative network, a noise-removal network and a classification network. Instead of asking artificial intelligence to learn only from recorded examples, the framework first uses mathematical descriptions and expert knowledge of target events to generate synthetic DAS signals. These generated signals are designed to reproduce important characteristics of real disturbances, including their evolution over time and their distribution along the sensing fiber.

The physics-informed generative network, known as PIGN, is central to the proposed architecture. It does not simply create arbitrary artificial waveforms. It incorporates information about how specific events are expected to appear in a DAS system, while also accounting for system constraints and signal behavior. This allows the network to produce a variety of simulated scenarios without requiring a large archive of actual incidents. Synthetic data can represent changes in event location, intensity, duration and propagation pattern, providing the learning system with a broader range of examples than might be available from field measurements alone. The strategy is especially important for rare events, where collecting real data could take years or involve unacceptable risks.

The generated signals are then used to train a specialized noise-removal network. In practical environments, the desired event is mixed with background signals rather than recorded in isolation. To reflect this condition, the researchers combine generated event data with readily obtainable background measurements. The denoising network learns to separate the event-related component from the surrounding acoustic activity, preserving the temporal and spatial structure that contains diagnostic information. This differs from many conventional filtering techniques, which may suppress noise but also erase weak features or fail when the noise changes unpredictably. By learning the distinction between event patterns and background behavior, the network is intended to operate more effectively in complex environments.

After denoising, a classification network analyzes the extracted signals and determines what type of event has occurred. Crucially, the complete training paradigm is designed not to depend on real-world event data. The generative model creates the event examples, the noise-removal model learns from synthetic events combined with background recordings, and the classifier is trained using the resulting denoised signals. This creates a pipeline that can potentially be adapted to new locations with limited additional data. Instead of rebuilding an entire recognition system every time a DAS cable is moved to a different site, operators could adjust the physical event descriptions and background measurements to match local conditions.

The researchers tested the method in belt conveyor monitoring, an application in which mechanical faults can threaten production, worker safety and equipment reliability. The network achieved a reported fault diagnosis accuracy of 91.8 percent, indicating that the combination of physics-based data generation and learned noise suppression can extract useful information from highly cluttered measurements. The framework also demonstrated event-recognition capability on public DAS datasets, supporting its potential beyond a single industrial experiment. Its performance suggests that artificial data do not necessarily have to be a weakness in machine learning, provided they are generated according to realistic physical principles and processed with models that understand the sensing system.

The significance of the work extends beyond one monitoring task. DAS is already being considered for pipeline surveillance, seismic observation, transportation security and structural health assessment, but deployment often depends on whether reliable training data can be collected at each new site. A physics-informed paradigm could reduce that barrier by combining the scalability of simulation with the adaptability of neural networks. If further validated under diverse field conditions, the method could enable faster installation of intelligent monitoring systems capable of identifying threats in real time. The approach also illustrates a broader shift in scientific artificial intelligence: rather than treating physics and machine learning as competing methods, researchers are using physical laws to guide models where data alone are incomplete. In that combination, optical fibers may become not only passive communication channels but distributed, continuously listening sentinels for the built and natural environments.

Web References: https://doi.org/10.1038/s41377-026-02295-5

References: Wan, Yangyang et al., “Towards a physics-informed network paradigm with data generation and background noise removal for different distributed acoustic sensing applications,” Light: Science & Applications, DOI: 10.1038/s41377-026-02295-5.

Keywords

Distributed acoustic sensing, DAS, physics-informed neural networks, artificial intelligence, fiber-optic sensing, synthetic data, noise removal, event classification, fault diagnosis, optical sensing, infrastructure monitoring, industrial safety