OPTICAL transport infrastructures form the physical backbone of modern cloud data centers, 5G/6G front-haul links, and fiber-to-the-premises (FTTP) access networks. Despite exponential growth in line rates, physical-layer fault diagnosis remains predominantly reactive and manual. When an optical link experiences packet loss or intermittent flapping, operators typically dispatch field technicians equipped with legacy 8-bit portable optical time-domain reflectometers (OTDRs) to capture Bellcore/Telcordia SR-4731 (.sor) traces [1], [2]. Because an 8-bit analog-to-digital converter (ADC) provides only 28 = 256 vertical quantization levels, weak Rayleigh backscattered signals returning from 15–25 km fiber spans suffer from severe quantization staircase noise, masking early-stage 0.05 dB microbends and gradual fusion splice degradation.
Furthermore, recent work by the TeCIP Institute at Scuola Superiore Sant'Anna [3] has highlighted the critical distinction between hard failures (such as sudden fiber cuts or Wavelength Selective Switch attenuation spikes exceeding 25 dB) and soft failures (progressive sinusoidal or monotonic optical signal-to-noise ratio (OSNR) decay of 0.5 dB/min). While software-defined networking (SDN) controllers monitor upper-layer bit-error rates (BER), they lack direct visibility into sub-nanosecond analog pulse distortions—such as optical sag, rise/fall time broadening, and I2C transceiver bus desynchronization (NACK stalls)—needed to distinguish a failing laser diode from a mechanical macrobend.
In this work, we bridge the gap between high-precision physical-layer reflectometry and autonomous network orchestration. Our main contributions are threefold:
1) Synchronous Four-Channel 12-Bit Edge Telemetry: We design a hardware-triggered acquisition pipeline sampling at 1.25 GSa/s with 12-bit native vertical resolution (212 = 4,096 levels, a 16× improvement over legacy 8-bit scopes), simultaneously capturing analog InGaAs PIN-TIA waveforms, I2C clock (SCL), SFF-8472 DDM serial data (SDA), and hardware RX_LOS interrupts.
2) Hybrid 1D-CNN/Transformer & Ensemble Diagnostic Engine: We formulate a dual-head neural network that processes raw OTDR/pulse waveforms alongside quantitative DSP metrics to classify four primary fault regimes (Healthy, Fiber Macrobending, Fiber Break/Fresnel Cleave, and Laser/Connector Decay) while regressing exact fault distance to within 1.5 cm RMSE.
3) Sub-50ms Closed-Loop Switch Remediation: We demonstrate an automated remediation controller that translates AI root-cause inferences into deterministic Cisco IOS and RouterOS transceiver reconfigurations in an average of 41.8 ms, preventing link flaps without human intervention.
As an optical probe pulse of peak power P0 and pulse width τp propagates along a single-mode silica fiber (ITU-T G.652) of length L, microscopic refractive index fluctuations frozen into the core during drawing induce continuous Rayleigh scattering. The backscattered optical power PRB(z, λ) returning from distance z at wavelength λ follows the inverse fourth-power law:
where S is the backscatter capture fraction, vg = c / ng is the group velocity in silica (ng ≈ 1.4682), and α(λ) is the total spectral attenuation coefficient (typically 0.33 dB/km at 1310 nm and 0.19 dB/km at 1550 nm). At discrete refractive index discontinuities (e.g., mechanical connectors, air gaps, or fiber breaks), localized Fresnel reflections occur with reflectance coefficient RF:
producing a reflection of approximately −14 dB (4%) at a flat glass-to-air cleave, better than −50 dB for Ultra Physical Contact (UPC) ferrules, and −65 dB for Angled Physical Contact (APC) terminations.
From Channel 1 (CH1) of our 12-bit digitizer terminated at 50 Ω, the ingestion engine computes five real-time morphological features for every captured frame x[n] (N = 25 × 106 points):
1) Peak-to-Peak Voltage (Vpp): Vpp = max(x[n]) − min(x[n]), quantifying instantaneous opto-electric conversion gain and insertion loss.
2) 10%–90% Rise and Fall Times (tr, tf): Measured in picoseconds to capture chromatic dispersion (CD) and modal bandwidth broadening.
3) Optical Sag Percentage (Sopt): Quantifies thermal droop across the flat top of sustained laser pulses:
where Vpeak = max(x[n]) and Vtail is the mean amplitude over the trailing 100 samples of the pulse plateau. An optical sag exceeding 15% strongly indicates laser transmitter aging when accompanied by rising SFF-8472 TX_Bias current.
4) RMS Timing Jitter (Jrms): Computed from threshold-crossing intervals Δtk across median zero-crossings: Jrms = σ(Δtk).
5) SFF-8472 I2C Health Vector: Extracted from CH2 (SCL at 100/400 kHz) and CH3 (SDA at EEPROM addresses 0xA0/0xA2), tracking temperature Tc, supply voltage Vcc, laser bias Ibias, Ptx, Prx, and bus NACK error counters.
| Fault Class / Regime | 12-Bit CH1 Analog Signature | CH2/CH3 SFF-8472 I2C Signature | CH4 RX_LOS | Autonomous Closed-Loop Action (< 50 ms) |
|---|---|---|---|---|
| Class 0: Healthy Link | Vpp ≥ 380 mV, Sopt < 5%, tr < 150 ps | NACK = 0, Ibias ≈ 14.8 mA, Prx > −16 dBm | Deasserted (0) | Maintain baseline profile; log telemetry to time-series DB. |
| Class 1: Fiber Macrobend | Non-reflective step drop (Δα1550 ≫ Δα1310) | NACK = 0, Ibias nominal, Prx drops 2–6 dB | Deasserted (0) | Shift active traffic toward shorter λ / boost dynamic TX power. |
| Class 2: Laser Degradation | Gradual Vpp decay, Sopt > 15%, Jrms > 80 ps | Ibias > 32 mA spike, Tc > 55°C | Intermittent | Pre-emptively reroute trunk to backup SFP port; throttle laser bias. |
| Class 3: I2C / Firmware Desync | Vpp ≈ 384 mV (Optical pulse fully intact) | NACK > 5, SCL clock stretching / frozen DDM | False Positive (1) | Suppress false link-down flap; issue soft I2C transceiver reset via CLI. |
| Class 4: Hard Fiber Break | Sharp Fresnel peak (RF ≈ −14 dB) then noise floor | Prx < −35 dBm | Asserted (1) | Immediate laser safety clamp + protection switchover + exact meter dispatch. |
To simultaneously achieve high fault classification accuracy and centimeter-level spatial localization across noisy field traces, we implement a multi-task deep neural architecture (OpticalFaultClassifier) in PyTorch, paired with an ensemble classifier for tabular DDM fusion.
Given an input reflectometry sequence u ∈ R1×M, two cascading 1D convolutional blocks with kernel size K = 3, batch normalization, ReLU activation, and max-pooling extract shift-invariant local morphological features (such as Fresnel spikes and Rayleigh slope discontinuities):
The resulting feature map H2 ∈ R(M/2)×64 is passed into a two-layer Transformer Encoder with H = 4 self-attention heads (dmodel = 64, dff = 128). Multi-head self-attention computes global spatial dependencies between the launch connector, intermediate splices, and distant attenuation tails:
The encoded representation branches into two parallel Multi-Layer Perceptron (MLP) heads: 1) a Classification Head trained with categorical cross-entropy loss Lcls, and 2) a Localization Regressor Head trained with Huber loss Lloc to predict the exact physical distance ẑfault in meters. The joint training objective is Ltotal = Lcls + β Lloc with β = 0.5.
The proposed system was evaluated on the CableHunt Birmingham Testbed using a 12-bit Rigol DHO804 oscilloscope (1.25 GSa/s, 25 Mpts), a 1.25 Gbps InGaAs PIN-TIA photodetector (1100–1700 nm), a 500 m SMF-28 launch spool, and a Cisco IE-3000-8FM industrial switch, augmented with 5,000 synthetic and real OPM traces derived from the TeCIP optical-failure-dataset [3] and GNPy [4].
| Model Architecture | Precision (%) | Recall (%) | F1-Score (%) | Localization RMSE (m) | Edge Inference Latency (ms) |
|---|---|---|---|---|---|
| Legacy 8-Bit Thresholding (Baseline OTDR) | 74.2% | 68.5% | 71.2% | 1.840 m | 120.0 ms (Manual/Passive) |
| Standalone Random Forest (DDM Features Only) | 88.6% | 86.9% | 87.7% | N/A (No Spatial Output) | 4.2 ms |
| 1D-CNN Only (12-Bit Waveform Input) | 94.8% | 94.1% | 94.4% | 0.062 m (6.2 cm) | 7.8 ms |
| CableHunt Hybrid 1D-CNN + Transformer + DDM Fusion | 98.6% | 98.2% | 98.4% | 0.015 m (1.5 cm) | 12.5 ms |
As shown in Table II and Fig. 2, the complete closed-loop pipeline—comprising hardware RX_LOS edge triggering (3.2 ms), DSP feature extraction (8.4 ms), hybrid neural inference (12.5 ms), and pre-warmed Netmiko/SSH switch reconfiguration (17.7 ms)—executes in 41.8 ms, achieving sub-50ms protection switching and eliminating false-positive link teardowns caused by I2C transceiver bus hangs.
We have demonstrated a 12-bit edge AI reflectometry and autonomous remediation system capable of diagnosing physical optical faults with 98.4% accuracy, 1.5 cm spatial RMSE, and 41.8 ms closed-loop healing latency. Core software drivers are open-sourced at github.com/sassil70/cablehunt (April 2025), while model weights and .sor tokenizers are staged at huggingface.co/sassil/cablehunt-optical-ai for full public release in 2029 following Class 1M laser eye-safety and UK Telecoms Security Act verification.