Hugging Face's logo

sassil
/
cablehunt-optical-ai

Time Series Forecasting
PyTorch
English
cablehunt_optical_1dcnn_pinn
optical-fiber
otdr
12-bit-adc
1d-cnn
physics-informed-neural-network
telecom
dwdm
time-series
predictive-maintenance
cablehunt
Model card Files
xet
 Community

πŸ€— CableHunt Smart Optical AI (cablehunt-optical-ai)

Physics-Informed 1D-CNN & Spectral Transformer for 12-Bit Optical Reflectometry & 40G Brownfield Remediation

Scheduled Public Weights Release Software Repo Official UK Portal

1. What is the Idea Behind cablehunt-optical-ai?

Every optical network on the planet β€” from London data-centre rings to rural fibre trunks β€” suffers from a blind spot at the physical layer.

Conventional Optical Time-Domain Reflectometers (OTDRs) sample backscattered light using 8-bit Analog-to-Digital Converters ($2^8 = 256$ vertical levels). At 256 levels, a tiny $-0.08\text{ dB}$ splice rot or a slow moisture creep inside an underground duct gets rounded off into coarse staircase noise. Nobody knows the link is dying until the glass snaps on a rainy night and drops traffic.

cablehunt-optical-ai is a purpose-built, physics-informed Smart Optical AI Model created by CableHunt Networks Limited (UK Company No. 16093071) and architected by Eng. Saleem Asseel. Instead of looking at blurry 8-bit traces, it ingests native 12-bit ADC Rayleigh backscatter waveforms ($2^{12} = 4,096$ vertical levels β€” 16Γ— finer resolution) across all six ITU-T optical bands ($1260\text{ nm}$ to $1675\text{ nm}$).

What the Smart Model Does in Real Time:

  1. Sub-Millivolt Fault Classification (<12 ms Inference):** Distinguishes between a degrading fusion splice ($-0.05\text{ dB}$), a physical duct pinch/macrobend, a contaminated ODF SC/APC ferrule (within a $<0.8\text{ m}$ event dead-zone), a ghost reflection, or a clean termination with **>98.4% field accuracy.
  2. Differential Wavelength Physics ($\Delta\alpha_{1625/1550}$): By comparing $1550\text{ nm}$ (C-Band) against $1625\text{ nm}/1650\text{ nm}$ (L/U-Band) mode-field leakage, the model knows immediately whether an attenuation dip is caused by thermal laser drift or someone physically bending the cable in a manhole.
  3. 3-Week Predictive Failure Horizon: Tracks temporal slope derivatives across repeated 12-bit scans to predict exactly when a stressed splice will breach the Forward Error Correction (FEC) cliff β€” alerting engineers 14 to 21 days before an outage.
  4. 1G-to-40G Brownfield Channel Optimiser: Computes the exact chromatic dispersion compensation and ITU-T channel allocation needed to upgrade legacy 1G brownfield G.652 fibres straight to 40G line-rates without civil trenching.
  5. Autonomous Sub-50ms Switch Protection: Generates verified CLI remediation actions via Python Netmiko for Cisco IOS-XR, MikroTik RouterOS v7, and Arista EOS switches in $\approx 38.4\text{ ms}$.

2. Model Architecture Overview 

Raw 12-Bit ADC / Bellcore .SOR Trace (4,096 Vertical Levels, N=1024 Boxcar Averaged)
       β”‚
       β”œβ”€β”€β–Ί [Stage A: Multi-Scale 1D-CNN Feature Extractor]
       β”‚      β€’ Dilated causal convolutions (kernel sizes: 3, 7, 15, 31)
       β”‚      β€’ Captures sharp Fresnel spikes (<0.8m dead-zone) & gradual Rayleigh slopes
       β”‚
       β”œβ”€β”€β–Ί [Stage B: Physics-Informed Spectral Attention (PINN Layer)]
       β”‚      β€’ Enforces ITU-T attenuation constraints: Ξ±(Ξ»), OH⁻ water peak (1383nm),
       β”‚        and macrobend differential loss (Δα = Ξ±_1625nm - Ξ±_1550nm)
       β”‚
       └──► [Stage C: Dual Task Heads]
              β”œβ”€β”€β–Ί 1. Fault Taxonomy & GPS/Distance Localisation (Β±0.42m precision)
              └──► 2. Remaining Useful Life (RUL) Prognostics & Safe Netmiko Action Policy
Specification Value
Model Family CableHunt-Optical-AI-v1 (Hybrid 1D-CNN + Physics-Informed Transformer)
Input Resolution 12-Bit Native Quantisation ($4,096$ vertical levels, $72.2\text{ dB}$ SNR)
Spectral Coverage $1260\text{ nm} – 1675\text{ nm}$ (ITU-T O, E, S, C, L, and U Bands)
Parameter Count $14.8\text{M}$ parameters (FP16 / INT8 quantised for edge FPGA & ARM NPU)
Edge Inference Latency $8.4\text{ ms}$ (CH-12B Handheld Unit) / $3.1\text{ ms}$ (Rack Server GPU)
Classification Precision $98.4%$ across 80 km G.652.D & G.657.A1 Birmingham testbed spools

3. Confirmed Public Release Schedule: 2029

We believe in being 100% transparent with the engineering community about our release timeline:

  • January–April 2025 β€” Initial Research Benchmarks & Core Software Stack Open-Source: Publication of our first 12-bit reflectometry research findings (Jan–Mar 2025) and open-source release of our Python 12-bit DSP acquisition suite, Bellcore .SOR parser, and Netmiko switch drivers on GitHub (sassil70/cablehunt) in April 2025.
  • 2029 β€” Smart Optical AI Model Public Weights Release (This Repository): The complete pre-trained weights (.safetensors), waveform tokenisers, and anonymised 12-bit .SOR benchmark datasets for cablehunt-optical-ai are confirmed for open public release in 2029.

Why 2029 for the AI Model Weights?

Because cablehunt-optical-ai directly governs optical launch power, DWDM channel tuning, and autonomous BGP/MPLS switch failover on live telecommunications infrastructure:

  1. Laser Eye-Safety & Photodiode Protection: We are completing hardware-locked guardrails so no fine-tuned derivative can command optical power levels that violate IEC 60825-1 Class 1M limits or damage remote InGaAs APD receivers.
  2. Adversarial Optical Pulse Hardening: Ensuring the model cannot be spoofed into triggering false core-network reroutes via injected out-of-band optical pulses.
  3. UK Telecoms Security Act (TSA) & Multi-Winter Validation: Completing multi-year endurance validation across UK metro ducts (Birmingham, Slough, and London corridors) to ensure zero false-positive flaps on Critical National Infrastructure (CNI).

4. Engineering Leadership & Direct Contact (@cablehunt.co.uk)

For academic collaboration, early carrier pilot evaluation at our Birmingham 80 km Optical Testbed, or commercial licensing inquiries prior to the 2029 public weight release:

Downloads last month
24
Inference Providers NEW
Time Series Forecasting
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support