TrackSentinel

A visual patrol is only as consistent as the day. The same head checking gets graded differently by two inspectors a week apart, documentation lives in a notebook, and coverage on a short line with no detector car depends on how many miles one crew can walk before dark.

Surface conditions that are cheap to fix when found (shelling, a missing clip, fouled ballast at the tie ends) become the internal defect that only ultrasonic testing can see later. The gap between patrols is where that happens.

Cameras (visible and thermal, 4K recommended) mount on a hi-rail vehicle, a wayside post or a locomotive and capture track imagery continuously, at up to 25 mph on hi-rail and faster on a locomotive. An NVIDIA Jetson processes it on the spot: no cloud upload, detections stay on your network.

Models trained on railroad imagery grade rail surface conditions (head checking, squats, visible cracks, shelling, corrugation), tie condition for wood and concrete (plate cut, rail seat abrasion, splits), fastener anomalies (missing or broken clips and spikes) and ballast surface indicators (fouling, voiding, insufficient ballast at tie ends). Each detection carries a confidence score, a severity class aligned with 49 CFR 213.113 terminology, GPS and a timestamp.

A qualified inspector verifies every detection and can override any classification; their judgement stays authoritative. Verified defects route to Maximo, SAP PM or Infor EAM with coordinates, photos, severity and a recommended action, prioritised by severity, location (mainline or yard) and tonnage per day. After site calibration the true positive rate on anomalies of 5 mm or more is above 90 percent, fastener detection above 95 percent, and the false positive rate under 5 percent.

Proof of concept in 2 to 4 weeks: site survey, baseline corridor scans, calibration to your conditions and accuracy validation against manual assessment. Pilot in 6 to 8 weeks with live inference on a priority corridor, CMMS integration and threshold tuning from inspector feedback. Then multi-corridor rollout with drift monitoring and quarterly accuracy reviews.

The limit is physics. Transverse fissures and other internal defects occur inside the rail steel and are invisible to a camera; they still need ultrasonic testing. TrackSentinel finds the surface conditions that precede them, which is how it pre-positions the UT programme.

Deployment complete in 60 to 90 days. Deployments report detection up 40 percent or more, coverage up 50 to 65 percent and emergency costs down 30 to 45 percent, with the honest note that results vary by railroad size, track class and operating context.