Quantum-inspired workflow for processing distributed fiber-optic sensor data – Scientific Reports

Turning a single strand of fiber into thousands of virtual microphones is incredible—until the data flood arrives. Distributed Acoustic Sensing (DAS) captures vibration along long fiber runs at high spatial and temporal resolution, making it ideal for monitoring pipelines, wind farms, subsea links, rail, and wellbores. The catch: it can generate terabytes per day, overwhelming storage, bandwidth, and analysis pipelines, especially when decisions must be made in real time. A quantum-inspired approach based on tensor networks offers a way through the bottleneck, shrinking data dramatically while allowing core signal-processing steps to run directly in the compressed space.

Why conventional pipelines struggle

Traditional single-point sensors don’t scale cleanly across vast assets. DAS sidesteps that by turning a cable into a continuous sensor array, but the resulting data volume becomes the new barrier. Keeping everything lossless is often impractical. Even widely used formats and generic compressors struggle to deliver the orders-of-magnitude reduction necessary for fast transfer and low-cost storage. Many practitioners lean on downsampling, filtering, or frequency band extraction (FBE)—useful, but these approaches throw away information and make high-fidelity reconstruction impossible.

From SVD to tensor networks

Low-rank structure is everywhere in signals, and singular value decomposition (SVD) has become a workhorse for denoising and compression. Tensor networks generalize that idea: reshape data into a higher-dimensional array and factor it across dimensions. One of the most practical structures, the Tensor Train, captures correlations with compact “cores” whose sizes are governed by ranks. If a signal’s singular values decay quickly—common for wave-like content—then it packs neatly into a small network without sacrificing accuracy.

Here’s the twist: with tensor networks, you don’t just compress; you compute in the compressed form. Linear operations—filters, transforms, even quantum-analogue gates—can be applied efficiently on the factorized representation. That’s where the “quantum-inspired” angle comes in. The Quantum Fourier Transform (QFT), a close cousin of the classical DFT, can be simulated using tensor networks. For low-rank signals, this can outperform conventional FFT pipelines because it avoids ballooning back to full dense arrays.

A workflow rebuilt for compressed-first processing

This work recasts a complete DAS processing workflow into tensor network form and tests it on a field-scale wellbore experiment aimed at detecting gas flow. The pipeline tackles three big pieces:

• Real-time, on-device compression into a Tensor Train while streaming data in, including multi-threaded stitching to handle long records without stalls.
• A parallelized QFT operating directly on the compressed representation, eliminating back-and-forth decompression.
• Quantum Frequency Band Extraction (QFBE): a tensor-network-native reinterpretation of FBE that isolates target bands in the QFT domain and maps back while preserving compactness.

The results are striking. Compression ratios of 40x–60x were achieved with high fidelity, and the entire process—compress plus QFBE—ran at real-time speeds on a standard laptop. In other words, the cost of “going compressed-first” was comparable to running a traditional FBE alone, but with the bonus of massive storage savings and the ability to continue processing without inflating data back to full size.

How it compares to current practice

Lossless strategies can be convenient and exact, but they rarely deliver the magnitude of reduction needed for continuous DAS monitoring. Lossy approaches—wavelets, zfp, dictionary learning, compressed sensing, and low-rank SVD—push further, but typically require a return to the dense domain for major operations. Tensor networks bridge that gap by combining aggressive compression with in-place computation, preserving a path to reconstruction while enabling transforms and filters to run efficiently on the factorized data.

Contrast that with classic FBE: it compresses by averaging out frequency content. That’s fast and simple but inherently destructive. QFBE keeps the spirit—emphasize the bands you care about, suppress the rest—yet does so on a compact representation and maintains a reversible pathway when ranks remain low. The payoff is better long-term archiving, easier re-analysis, and a cleaner foundation for downstream machine learning.

Why this matters for the edge

DAS is a natural edge workload: sensors stream continuously, links to the cloud are bandwidth-constrained, and many use cases—leak detection, intrusion monitoring, well integrity—demand immediate action. A tensor-network-first workflow makes it feasible to:

• Compress at the source, keeping only compact cores.
• Run core analytics (e.g., QFT, filtering, band extraction) in the compressed space.
• Transmit or archive tiny artifacts instead of raw terabytes.
• Scale up coverage without linear growth in infrastructure costs.

With multi-threaded stitching and parallel QFT, the approach maps well to commodity CPUs and modest GPUs, making it attractive where ruggedized, low-power hardware is a must.

Caveats and opportunities

The magic depends on low-rank structure. If singular values don’t decay quickly—say, in highly chaotic or broadband environments—ranks grow and the advantages shrink. But many industrial signals exhibit exploitable structure, especially after basic preconditioning. That opens the door to adaptive schemes: estimate ranks on the fly, choose between FFT or QFT-in-TN, and fuse with learned priors when appropriate.

Beyond QFBE, the same toolkit invites a rethinking of other DAS staples: matched filtering, beamforming, denoising, and event detection could all be re-implemented to operate natively on tensor trains. There’s also a path to standardization akin to how the world settled on common formats for images and video—only here, the “codec” would double as a compute substrate.

Key takeaways

• DAS generates massive, continuous datasets that strain conventional storage and analytics.
• Tensor networks capture low-rank structure and enable operations directly in compressed form.
• A QFT-based, tensor-network workflow (with QFBE) delivered 40x–60x compression and real-time performance on a laptop for field wellbore data.
• Compared to discard-heavy methods like classic FBE, the compressed-first approach preserves a route to reconstruction and richer re-analysis.
• This quantum-inspired strategy could become a practical edge standard for large-scale infrastructure monitoring.

The bottom line: by treating compression not as a prelude but as the computational substrate, this workflow turns DAS from a data deluge into a streamlined, real-time signal-processing platform—one that’s ready for the edge, scalable across assets, and primed for the next generation of intelligent monitoring.