Industries / Transport and vehicles / Rail
Track and inspection
This is the work of building and keeping the track (the rails, the ties under them, which the British call sleepers, and the crushed stone, or ballast, that holds the ties in place) and of checking it for cracks and for rails that have moved out of line.
5
Milestones achieved, from 1857 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1857
Steel rails outlast iron
Robert Mushet's steel rail was laid early in 1857 at Derby station on the Midland Railway, at a busy spot where iron rails "had to be renewed every six months, and occasionally every three". In 1863, after some 700 trains a day had passed over it, "the rail seemed as perfect as ever".
Credit
Robert Forester Mushet; the Midland Railway
1928
A car finds hidden cracks
A 1911 derailment at Manchester, New York, killed 29 people and was blamed on transverse fissures, cracks that grow inside the rail head where no one can see them. Elmer Sperry's detector car, which found such cracks with an electrical method his team called induction, was accepted by the American Railway Association's Rail Committee on 2 October 1928, and first tested 130 miles of Wabash Railway track for pay that November. Sperry offered ultrasonic rail testing from 1949.
Credit
Elmer A. Sperry and Sperry Rail Service, by William E. Miller's account
1953
Machines pack the stone
Tamping means packing ballast tightly under the ties so the track sits at the right height. Plasser & Theurer of Linz, Austria, lists 1953 as the year of its "first hydraulic track tamping machine", and says that until the middle of the 20th century track maintenance "required heavy manual labour".
Credit
Plasser & Theurer
2000
Hatfield: cracks break a rail
On 17 October 2000 an express traveling at 185 km/h derailed south of Hatfield, England, when a rail broke into pieces under it, killing four people and injuring more than 70. The inquiry found "multiple and pre-existing fatigue cracks", a form of wear called rolling contact fatigue in which repeated wheel loads grow cracks in the rail. British Rail's research center had studied the problem in the 1970s, and Railtrack, which owned the track, was declared bankrupt after the crash.
LessonKemp writes that Railtrack had "divested much of BR's engineering knowledge to contractors", and a known kind of crack still broke a rail.
Credit
the official inquiry, as reviewed by Roger Kemp for the Royal Academy of Engineering
2025
Machines replace one inspection
Automated track inspection measures the shape of the track from instruments on a moving car. In December 2025 the FRA granted the Association of American Railroads (AAR) a five-year waiver: on chosen subdivisions where machines check all main tracks and sidings at least monthly, railroads may cut visual inspections from twice a week to once, and go back to the old schedule if defects exceed set limits. Union Pacific and BNSF began using it in spring 2026, CSX in July and Norfolk Southern in August, by the Alliance for Innovation and Infrastructure's count. The SMART-TD union said the waiver "does nothing to improve safety".
Credit
the FRA; the AAR
Watch for
Fewer visual checks, safer track
The AAR and the large railroads say that monthly machine checks plus one weekly visual inspection keep track safer than two weekly visual inspections alone, and the FRA's administrator said the machines are "designed to enhance already effective visual inspections". The waiver requires railroads to report results, defect rates and derailments on covered track.
Achieved whenThe FRA or an independent party publishes defect and derailment rates from those reports and compares them with track still inspected visually twice a week.
Claimed by
the AAR, in its April 2025 petition, and the railroads using the waiver in 2026
Open
01
Open rail ultrasound benchmark
A public set of ultrasonic scans of rail with flaws of known size, and a baseline detector. Ultrasonic testing sends sound pulses into the rail and records the echoes from cracks; a B-scan is the picture built from those echoes as the probe moves. Use offcuts of scrap rail with reference flaws machined at measured sizes and depths, such as drilled holes and saw cuts, and, where they can be had, pieces of rail taken out of service with real defects.
The briefCredit · unclaimed02
Crack depth from the surface
Head checks are fine cracks on the running corner of the rail head caused by rolling contact fatigue, the kind of wear behind Hatfield. From the surface an inspector sees a crack's length, not its depth, and depth decides whether grinding the rail is enough or the rail must be replaced. Build a public set that pairs close-up photos of head checks with depths measured by cutting the rail and polishing the cut faces, and a model that estimates a crack's depth band from its photo. Use rail already removed from service.
The briefCredit · unclaimed
Open 01 · the brief
Open rail ultrasound benchmark
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of ultrasonic scans of rail with flaws of known size, and a baseline detector. Ultrasonic testing sends sound pulses into the rail and records the echoes from cracks; a B-scan is the picture built from those echoes as the probe moves. Use offcuts of scrap rail with reference flaws machined at measured sizes and depths, such as drilled holes and saw cuts, and, where they can be had, pieces of rail taken out of service with real defects.
Why it is still open
A Beijing Jiaotong University dataset on metro rail shares only processed data, because raw ultrasonic inspection records "contain location-sensitive infrastructure-maintenance information". Published detectors for rail B-scans are trained on railways' in-house data, and this research found no open set of real rail scans.
Built with what exists
Scrap rail offcuts; a machine shop for the reference flaws; a portable ultrasonic flaw detector with standard angle probes, bought used or rented; Python and an open-source object detector.
First step
Buy 2 m of scrap rail, have ten holes of 2 to 8 mm drilled at known depths, and record 100 scans with a rented flaw detector.
Done looks like
At least 20 rail pieces holding at least 200 flaws of measured size and depth, and at least 2,000 scans, under an open license; a published table of the detector's hit rate and false alarms by flaw size and depth, with code public so a second person can rerun it.
Open 02 · the brief
Crack depth from the surface
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Head checks are fine cracks on the running corner of the rail head caused by rolling contact fatigue, the kind of wear behind Hatfield. From the surface an inspector sees a crack's length, not its depth, and depth decides whether grinding the rail is enough or the rail must be replaced. Build a public set that pairs close-up photos of head checks with depths measured by cutting the rail and polishing the cut faces, and a model that estimates a crack's depth band from its photo. Use rail already removed from service.
Why it is still open
Public rail-surface image sets, such as Rail-5k (2021, 1,100 labeled images of 13 defect types), label the kind of defect but not its depth. A 2025 study of head checks on the Belgian network by TU Delft and Infrabel used eddy-current depth measurements, an electromagnetic method, and this research found no public release of those depths.
Built with what exists
Removed rail with visible head checks, from track contractors, scrap dealers or heritage railways; a band saw and a polishing and etching kit, or a university materials lab; a macro lens; dye penetrant to mark the cracks; a vision-language model to draft labels, each checked by a person (see the Note "Three working devices").
First step
Ask a track contractor or heritage railway for five pieces of removed rail with visible head checks, then cut and polish one.
Done looks like
At least 50 rail pieces and 500 cracks, each with a photo and a measured depth, under an open license, and a published table of how often the model puts a crack in the right depth band (under 2 mm, 2 to 5 mm, over 5 mm) on rail pieces it was not trained on.