Radar collects the data. Intelligence turns it into decisions.
The problem AI solves
A GPR survey vehicle collecting at hi-rail speed generates tens of miles of multi-channel radar data per day — far more than analysts can interpret scan by scan. The FRA’s own evaluation program flagged this directly: data volumes outpace manual processing, and the future of track GPR lies in more sophisticated, more automated algorithms. That future is what KHEERAN builds toward.
Where we apply machine intelligence
Automated fouling classification. Signal-processing features — scattering amplitude envelopes, frequency-spectrum content, time-frequency energy maps — feed classification models that assign fouling categories (clean through highly fouled) consistently along the entire route. The same input always produces the same answer, eliminating analyst-to-analyst variability.
Layer interface tracking. Where clean and fouled ballast grade into each other, there is often no crisp reflection for a human to trace. Pattern-recognition algorithms track textural transitions in the radar data to map clean-ballast thickness continuously, even where traditional layer-picking fails.
Anomaly detection. Models trained on normal substructure signatures flag the exceptions worth an engineer’s attention: ballast pockets, trapped-water reflections, abrupt thickness changes, sections behaving unlike their neighbors. Engineers review the flagged 2 percent, not the unremarkable 98.
Calibration learning. Every ground-truth sample we take improves the mapping between radar response and measured fouling index for your ballast type. Over successive campaigns, classification accuracy for your network compounds.
Network-level trending. Repeat surveys become a time series. Algorithms compare campaigns to estimate fouling rates by segment, forecast when sections will cross maintenance thresholds, and prioritize renewals by predicted condition — the foundation of true condition-based maintenance.
Human in the loop, always
AI accelerates analysis; it doesn’t replace engineering judgment. Every automated classification is spot-checked against calibration samples, every flagged anomaly is reviewed by an experienced analyst, and every report is signed off by an engineer who understands track, not just data. Independent research is clear that radar interpretation has edge cases — unusual particle shapes, abnormally wet or dry conditions — and our workflow is built so the model’s confidence is never mistaken for certainty.
What this means for you
Faster turnaround — automation is how first-pass results arrive within 24 hours. Consistency — the same standards applied to mile 1 and mile 400. Fewer blind spots — algorithms don’t get tired in hour seven of radargram review. Compounding value — each survey makes the next one smarter about your network.