Neural compression keeps live fiber sensing streams flowing
STeP-C: Real-Time Neural Compression for Distributed Fiber-Optic Sensing at Scale · arXiv:2611.17126
In plain words
Fiber-optic cables can measure vibrations along many channels, producing a steady flood of detailed data.
This paper introduces STeP-C, a system that compresses those measurements by encoding patterns across nearby channels, predicting what comes next, and storing the remaining differences so the signal can be rebuilt.
In a live one-hour test across 81,000 channels, it processed data faster than the system produced it and dropped no acquisition frames. Reconstruction quality varied across the fiber.
On separate recorded data, the same pretrained detector gave matching top classes on original and reconstructed data for 96.62% of event windows where its confidence was at least 0.6.
The research
Continuous fiber measurements can outpace the systems that store and transmit them, while compressed signals still need to preserve useful waveform detail.
This result changes the expectation that neural compression can be assessed only offline: STeP-C kept pace during a live acquisition test, though reconstruction quality varied across channels.
STeP-C compresses distributed fiber-sensing data by predicting later frames and coding the remaining differences, then reconstructs the signal for downstream use.
In a live test, STeP-C processed one hour of data from 81,000 fiber channels at an average 535 megabytes per second, above the 324 megabytes-per-second acquisition requirement, with no dropped frames.
Reconstruction across the full band reached 31.04 decibels in peak signal-to-noise ratio (PSNR), a measure of reconstruction fidelity where higher values mean a closer reconstruction.
On separate recorded data, the same pretrained detector agreed with its original-input predictions on 96.62% of high-confidence cases. Together, these results support live compression and substantial detector-output consistency under the tested conditions.
The live test covered one hour on one sensing system, and reconstruction quality varied across fiber locations. Detector agreement compares a pretrained detector's outputs, not predictions against independently verified event labels; the study did not test event localization or arrival-time picking.
What the evidence shows
- A compressor shaped for fiber data
STeP-C (Spatio-Temporal Predictive Compression) is a neural compression framework for distributed acoustic sensing (DAS), which records vibration along fiber-optic cables. It encodes patterns across neighboring channels, predicts later frames, and compresses the prediction errors.
This structure addresses the need to reduce continuous, high-volume sensor streams while reconstructing waveforms for physical measurements.
- Coding speed was the bottleneck
A simulated streaming test fed pre-recorded data to the STeP-C-small pipeline at controlled rates, separating processing speed from live acquisition and file handling. The standard arithmetic coder encoded at 13 megabytes per second; the custom coder reached 392 megabytes per second.
This points to entropy coding as the bottleneck in the unoptimized pipeline, though the result comes from a simulated test.
- Prediction reduced bitrate
The comparison shows that temporal prediction reduced the bitrate, measured as average compressed bits per raw data element (BPS), needed for similar reconstruction fidelity. Peak signal-to-noise ratio (PSNR) measures reconstruction fidelity; higher PSNR means a closer reconstruction.
This supports prediction as a source of compression gain on that dataset; the source does not specify whether the two versions shared starting weights.
- Low-bitrate comparison has limits
On NJCFA, STeP-C-small reached 34.51 decibels in peak signal-to-noise ratio (PSNR), a measure of reconstruction fidelity where higher values mean a closer reconstruction, at 0.436 BPS, the average compressed bits per raw data element.
EnCodec was the only other method tested below 0.5 BPS, where its reconstruction quality stayed near 32 decibels PSNR. The other listed methods did not cover that low-bitrate range.
Because the methods' operating points differ, these results do not give an exact bitrate-matched comparison between STeP-C-small and EnCodec.
- Live operation and detector check
The live run shows that the full pipeline kept pace with acquisition across the tested stream.
In a separate DADAS check, the same detector assigned the same top class to original and reconstructed data in 96.62% of event windows where its confidence was at least 0.6; this measures output consistency, not correctness against verified event labels.
Deep
- Original paper authors
- Tingfeng Li, Ming-Fang Huang, Shaobo Han, Yue Tian
- Submitted to arXiv
- License
- arXiv non-exclusive