Return

Major Breakthrough | AllianStream Photonics × Shanghai Jiao Tong University Publish Research in Light: Zero-Shot Physics-Informed AI Solves Bottlenecks of DAS

News Flash! The full-stack original AI fiber sensing solution jointly developed by Ningbo AllianStream Photonics and Prof. Zuyuan He’s team from Shanghai Jiao Tong University has been officially published in Light: Science & Applications, a top international optics journal under Nature Portfolio. It is a SCI Q1 Top Journal with an Impact Factor of ~23.4, ranking among the world’s top 3 optics journals.

 

Targeting two world-class long-standing bottlenecks of distributed acoustic sensing (DAS) — scarce on-site fault samples and strong environmental noise masking target signals — the team proposes a complete physics-informed neural network paradigm consisting of a PIGN Generative Network and a dedicated denoising network. The whole algorithm training process requires no real on-site fault samples at all. Only relying on physical mechanisms and easily accessible on-site background data, it integrates synthetic event generation, background noise separation and cross-site fault identification. Field tests in coal mines achieve a fault diagnosis accuracy of 91.8%, outperforming traditional data-driven algorithms comprehensively, and providing a disruptive underlying technical route for full-range intelligent monitoring of oil & gas facilities, mines, power grids and submarine cables.

 

Journal Academic Value

Light: Science & Applications is co-published with Nature Portfolio, consistently ranking top 3 globally in optics research. It only accepts original research with paradigm reforming and industrial breakthrough significance. This integrated physics-informed DAS system represents a rarely high-level industry-university joint achievement among domestic fiber acoustic sensing studies, carrying extremely high academic and industrial value.

 

One-Minute Introduction to DAS

DAS relies on the Rayleigh backscattering effect inside optical fibers. A single fiber cable acts as millions of continuous micro-sensors, capturing vibration and acoustic signals over long distances without blind zones. Featuring high voltage resistance and corrosion resistance, and compatible with existing communication cables, DAS is widely adopted for long-term operation and maintenance of mines, long-distance oil & gas pipelines, submarine cables and OPGW power lines, as well as perimeter security systems.

 

Two Global Industrial Pain Points Restricting DAS Commercialization

❶ Extreme Scarcity of Real Fault Samples

Abnormal incidents including pipeline rupture, conveyor roller failure, illegal intrusion and microseisms occur with extremely low probability. Over 95% of field data is meaningless background noise. Manual collection and labeling of fault samples are time-consuming and costly. Traditional deep learning algorithms heavily rely on sufficient annotated fault data; insufficient samples will cause sharp drop of recognition accuracy and hinder large-scale commercial deployment.

❷ Weak Fault Signals Masked by Severe Noise in Complex Scenarios

Mines, factories, offshore sites and wild areas are filled with persistent mechanical vibration, water flow and wind noise. Tiny fault signals can be fully submerged by noise. Conventional denoising algorithms and general AI denoising networks require clean, noise-free reference samples, which are barely obtainable in industrial sites, leading to massive false negatives and false positives and insufficient equipment reliability.

 

Core Original Breakthrough Published in Light: Full-Stack Physics-Informed AI Neural Network System

This research breaks away from the conventional mindset of training AI with massive field measured data, and builds a DAS-dedicated integrated neural network framework. No measured fault data from target sites is needed throughout the training workflow, which includes two core original modules:

1️⃣ PIGN Physics-Informed Generative Network (Original Synthetic Data Module)

Different from traditional GANs which still require measured raw data as input, PIGN constructs spatiotemporal feature constraint functions based on physical dynamic models of events or industry expertise, combined with hardware characteristics of DAS equipment and on-site environmental constraints, to generate massive standard event data autonomously:

  • Accurately replicate time-domain and frequency-domain signals of pedestrian movement, mechanical vibration and various conveyor roller faults;
  • Automatically reproduce inherent distortion features of DAS such as phase wrapping and signal saturation;
  • The time-frequency characteristics of synthetic data match field measurements, and can replace measured fault data as training sets.

2️⃣ DAS Dedicated Background Denoising Network (Noise Separation Module)

Trained with clean fault signals generated by PIGN mixed with easily collected normal on-site background data, no noise-free reference signals are required:

  • Efficiently filter intensive continuous mechanical background noise in mines and factories;
  • Fully retain unique impulse, narrowband and broadband frequency features of faults; the amplitude of fault signals can be increased by up to 6 dB after denoising;
  • Input denoised data into classification networks to drastically reduce missed fault detection and environment-triggered false alarms.

3️⃣ Five-Step Zero-Shot Deployment Workflow

  1. Generate massive synthetic fault data in batches via PIGN guided by physical models or industry experience;
  2. Collect easily accessible fault-free background fiber signals on site;
  3. Train dedicated denoising networks by mixing synthetic fault data and real background data;
  4. Train event classification networks with denoised clean datasets;
  5. The fully trained network can be directly migrated to new operation sites without recollecting fault samples.

 

Figure 1 Schematic diagram of the complete physics-informed DAS neural network architecture

 

Three Levels of Authoritative Verification: Models Trained with Synthetic Data Outperform Traditional Measured-Data-Based Schemes

Validation 1 Identification of Pedestrian and Vibration Events on Public Dataset

Tests are conducted on public spatiotemporal DAS datasets, training mainstream CNN, ResNet and CNN-BiLSTM networks solely with PIGN synthetic data:

  • The recognition accuracy of models trained purely on synthetic data steadily exceeds 70%;
  • After fine-tuning with a small volume of on-site samples, the accuracy of CNN-BiLSTM reaches 86.2%;
  • Under identical training settings, the performance surpasses traditional algorithms trained with a tiny amount of real fault data.

Figure 2 Comparison between PIGN-generated signals and real signals + multi-model confusion matrices for classification

 

Validation 2 Denoising Effect Test on AllianStream Self-Developed Conveyor Testbed

A conveyor test platform is built based on the self-developed ixDAS-4000 Distributed Acoustic Sensing System (Max sensing distance: 60 km, minimum spatial resolution: 3.6 m, self-noise level: 10 pε/√Hz @20 Hz), simulating three typical roller faults: eccentricity, split and crack:

  • Fault features are completely covered by noise in raw signals and become distinct after denoising;
  • The fault recognition accuracy of CNN is increased by 8.4% after training with denoised data;
  • Significant performance improvement is achieved on ResNet, CNN-BiLSTM and other mainstream networks.

Figure 3 Time-frequency curve comparison between raw conveyor signals and denoised signals

 

Validation 3 Cross-Site Industrial Field Test in Underground Coal Mine (Core Engineering Breakthrough)

The complete AI network trained only on the simulation testbed is directly migrated to a 400-meter real coal mine conveyor site, without any mine fault data used during training:

  1. The three-classification accuracy of traditional data-driven models trained with a small amount of mine fault samples is only 71.3%;
  2. The diagnosis accuracy of the proposed physics-informed scheme without mine fault training data reaches 85.6%;
  3. After fine-tuning with a small number of on-site samples, the binary early warning accuracy for faults hits 91.8%;
  4. It comprehensively outperforms traditional time-frequency diagnosis algorithms manually tuned over long-term data (86.6%).

Figure 4 Cross-site migration test flowchart of coal mine + accuracy comparison of various algorithms

 

 Industry-University Collaborative R&D

  • Shanghai Jiao Tong University: Physical mechanism modeling, original architecture design of PIGN and denoising network, multi-scenario dataset verification;
  • Ningbo AllianStream Photonics: Independent R&D and supply of ixDAS series DAS demodulation hardware, construction of conveyor testbed, field measurement in coal mines, modular algorithm optimization and industrial delivery for enterprise clients.

The university-enterprise team has built a complete R&D chain: Physical Modeling → Algorithm Innovation → Hardware Matching → Simulation Test → On-Site Mine Verification, accelerating the transformation of top academic achievements into commercially available industrial intelligent sensing systems.

 

 Application Scenarios

▪ Coal & aggregate conveyor systems: Early warning of roller wear, eccentricity and cracking to reduce downtime losses;

▪ Long-distance oil & gas pipelines: Precise identification of weak vibrations caused by third-party excavation and mechanical impact;

▪ Submarine and terrestrial trunk cables: Long-term monitoring of seabed erosion, geological settlement and illegal construction;

▪ High-voltage OPGW cables: Online early warning of ice coating, tower abnormal vibration and external mechanical damage;

▪ Perimeter security, tunnel structural health monitoring, full-range microseismic sensing.

 

Paper Information

Yangyang Wan, Haotian Wang, Xuhui Yu, Jiageng Chen, Xinyu Fan, Zuyuan He.

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

Light: Science & Applications (2026) 15:281

DOI: 10.1038/s41377-026-02295

Innovation never stops, the future is already here.

AllianStream Photonics, marching toward the vast ocean of fiber sensing together with you!

 

Follow the official WeChat account of AllianStream Photonics to obtain full hardware and AI algorithm industry solutions for DAS.

#DAS #AllianStream Photonics #Shanghai Jiao Tong University #Light #Physics Informed AI #Zero Shot Learning #Convey or Monitoring #Pipeline Security #Industry University Collaboration