Internal Rail Defect Detection Methods, Ranked by Maturity

Internal rail defect detection methods split into two groups. In revenue service, contact ultrasonic testing (UT) does the work, with induction as a complement and hand probes for verification. Phased array, EMAT, laser, air-coupled and guided-wave systems are still pilots or research. In a 1998 benchmark, six UT test vehicles detected 72% of 49 known flaws.

This page ranks each technology by how far it has got: from routine use on US track to laboratory work. For each one it gives what the method finds, its known limits and the published evidence, mostly FRA reports, TTCI (now MxV Rail) digests and journal papers. It covers internal defects only: transverse defects, detail fractures, split heads, bolt hole cracks and weld defects.

Three related pages go deeper on questions this one only touches. For which method suits which defect, see the defect-to-method matrix on detecting rail defects with NDT methods. For what an ultrasonic rail test covers and reports, see ultrasonic rail testing capabilities. For the parts of a test system, see the ultrasonic rail flaw detection system anatomy.

What counts as an internal rail defect detection method?

An internal rail defect detection method is one that can find flaws inside the rail head, web or base, not just on the running surface. In the United States the legal test is 49 CFR 213.237, which requires "internal rail inspections" on Class 3 to 5 track and defines what a usable inspection is.

The rule never names a technology. It sets two performance conditions instead. Paragraph (f) says: "Inspection equipment shall be capable of detecting defects between joint bars, in the area enclosed by joint bars." Paragraph (j)(4) defines the result the equipment must deliver:

"Valid search means a continuous inspection for internal rail defects where the equipment performs as intended and equipment responses are interpreted by a qualified operator as defined in § 213.238." — 49 CFR 213.237(j)(4)

That definition is the real filter for this page. A technology that can't search continuously, can't reach inside a joint, or produces responses no qualified operator interprets, can't yet carry a 213.237 inspection on its own, however good its laboratory results. The intervals themselves (370 days or 30 MGT on Class 4 and 5 track, for example) are explained on how often rail must be tested and in our 49 CFR 213.237 guide.

Why some familiar methods are left out. Eddy current, ACFM, magnetic flux leakage and machine vision mainly measure the surface or the first few millimeters of the head. They help with rolling contact fatigue and grinding decisions, but they are not internal methods. Radiography can image the inside of a weld, but in track it is a niche tool: it needs radiation controls and exposure time per joint and can't scan miles of rail. Ground-penetrating radar looks at ballast and subgrade, not rail steel. The original version of this article listed all of these as internal methods; it was wrong to.

Key takeaway: a method counts here only if it can find flaws inside the rail, and it counts for compliance only if it can deliver a 213.237 valid search.

How are internal rail defect detection methods ranked by maturity?

Only contact UT is in routine use as the primary method on US track; induction is a complement, hand UT is the verification tool, and everything else is a pilot or research. FRA said so directly in the preamble to its 2014 rail integrity rule:

"The main technologies utilized for non-destructive testing on U.S. railroads are the ultrasonic and induction methods. Ultrasonic technology is the primary technology used, and induction technology is currently used as a complementary system." — FRA, Track Safety Standards; Improving Rail Integrity, 79 FR 4234 (Jan. 24, 2014)

A 2008 journal review reached the same split from the research side. Papaelias, Roberts and Davis found that in-service rail inspection had "extensively used" ultrasonic transducers and magnetic induction sensors, while EMAT, laser ultrasonics, guided waves and ACFM "are also under development" (Proc. IMechE Part F 222(4), 2008).

This page uses four maturity levels. They are our editorial judgment from the published sources, not an FRA or AREMA classification:

Key takeaway: if a proposal presents any method below the first four rows as a replacement for contact UT, ask for its benchmark data first.

How does contact UT with roller search units and sliders work?

Contact UT sends 2.25 to 5 MHz sound into the rail from probes that ride on the head, and maps the echoes from internal flaws, bolt holes and the base. It is the deployed baseline every other method on this page is measured against.

The 2014 FRA preamble gives the clearest public description. The frequency range "normally utilized during current flaw detection operations is 2.25 MHz … to 5.0 MHz", and the transducers sit "at various angles with respect to the rail surface". The sound "normally scan[s] the entire rail head and web, as well as the portion of the base directly beneath the web", and "Interpretation of the reflected signal is the responsibility of the test system operator."

Roller search units (RSUs) are fluid-filled wheels with transducers inside. TTCI describes conventional detector cars as using "piezoelectric transducers housed in roller search units that run atop the railhead", with fixed-angle probes (TD-16-030, 2016). A 1999 Colorado State report for the Mountain-Plains Consortium (MPC 99-106) described three to six transducers per wheel and up to 12 per rail. Typical angles are 0° for horizontal flaws and loss of bottom, about 37° to 45° for the web and bolt holes, and 70° for transverse defects in the head, plus side-looking probes. No public source names a standard North American channel set, so treat any list as typical.

Sliders (sleds) put the probe in a shoe that slides on the head under a water film. One NDT maker's application note says sliders give finer scan pitch but use far more water, and that the choice between wheels and sliders "mainly depends on tradition of a railway service company". That is a vendor view; no independent comparison was found. The system anatomy page covers both in more detail.

Known limits, from the sources:

• Base flanges. Head-contact UT reaches the base only under the web. Thermite and other field welds need a hand scan of the base.
• Surface masking. FRA's 2014 preamble lists shells, engine burns, spalling, flaking, corrugation and head checking as conditions that "negatively influence the detection of rail flaws". See shelling and spalling for how a shell blocks the beam.
• Worn rail. TD-16-030 notes that on worn rail fixed-angle probes can inspect "the wrong volume of material".
• Lost coupling. A stretch with no back-wall echo is untested rail, not sound rail, and falls under 213.237(g) and (h).

Test speed is its own subject, covered on our high-speed ultrasonic rail inspection page.

Key takeaway: contact UT is the only method with a published North American benchmark, and that benchmark also documents its blind spots.

What does the published benchmark evidence show for contact UT?

The main public benchmark is TTCI's 1998 evaluation of six hi-rail UT systems over 49 known flaws, which found 72% overall, 75% of transverse defects and 33% of bolt hole cracks. It was published as Technology Digest TD-99-006 (Garcia and Reiff, February 1999).

The tests ran from February to May 1998 on the Rail Defect Test Facility (RDTF) at FRA's Transportation Technology Center near Pueblo, Colorado. The flaws had been found in revenue service, verified by TTCI and placed in the test track. Three North American contractors supplied the cars and crews. Hits were hand-mapped afterward to confirm them.

TD-99-006 also compared the results with a "Recommended Minimum Performance Guideline for Rail Testing", which it says was added to the AREMA Manual for Railway Engineering in 1992. The comparison below is as TTCI printed it for mainline track over 3 MGT a year. Check the current MRE edition before quoting AREMA's figures as current.

MPC 99-106 (Jeffrey and Peterson, 1999) analyzed the same runs. It counted 19 of the 49 flaws as "frequently missed", meaning missed three or more times. Most were small detail fractures, many of 3% to 8% of head area, and most sat under "no surface anomalies". Only four of the 19 were under shells, spalls or flaking. The authors found strong evidence that detection depended on defect size. They ran no statistics on orientation or surface condition because the samples were too small.

TTCI's later summary is blunter. In 2016 it wrote that conventional UT "has proven effective at finding most rail flaws (about 90 percent), but inherent limitations leave some flaws (about 10 percent) undetected during periodic inspection" (TD-16-030). The digest gives no test basis for the 90%, so treat it as TTCI's estimate, not a measured rate.

Key takeaway: the benchmark is real, but it is from 1998, uses one test track and has very few flaws in some rows.

Worked example: how much a benchmark percentage tells you

A detection percentage from a small benchmark carries a wide uncertainty band, so two numbers that look different may not be. This example uses TD-99-006's shell result to show why, without inventing any new rate.

Inputs from TD-99-006: seven TDs under shells and 33 TDs with no surface anomaly, each run by six systems. That gives 7 × 6 = 42 detection opportunities under shells and 33 × 6 = 198 without. The reported rates (about 62% and 77%) imply about 26 hits of 42 and 152 hits of 198.

Step 1: compute a 95% interval for each. Using the Wilson score interval for a proportion:

Step 2: read the result. The two TD intervals overlap between 70.4% and 75.0%. The direction (shells hurt detection) agrees with FRA's 2014 description of masking, but the benchmark alone can't pin the size of the gap.

Step 3: allow for repeated flaws. The 42 opportunities are only seven physical defects seen six times. Repeat looks at the same defect are not independent, so the true uncertainty is wider than the table shows.

What to take from it. When a vendor quotes a detection rate for any method on this page, ask how many distinct flaws were tested, what sizes, at what detection threshold, and with what false-call rate. The USAF reliability handbook MIL-HDBK-1823A makes the same point: POD comparisons between methods only mean something at equal false-alarm rates.

Is induction testing still used for internal rail defects?

Yes, as a complement: FRA described induction in 2014 as "a complementary system" to ultrasonic testing, and it is the oldest detector-car method still cited. It is not the same thing as magnetic flux leakage, and it doesn't replace UT.

FRA's description of the principle is plain: "Induction testing requires the introduction of a high-level, direct current into the top of the rail and establishing a magnetic field around the rail head. An induction sensor unit is then passed through the magnetic field." A flaw distorts the current flow, and the search unit picks up the resulting field distortion.

Induction came first. The first commercial detector car test, in November 1928, used an induction method; ultrasonic cars followed, with the first for the New York City Transit Authority in 1959 (MPC 99-106). Our page on how detector-car testing began gives that history. In 1999 MPC 99-106 reported that most major contractors relied only on ultrasonic cars, while some ran larger rail-bound units carrying both ultrasonic and induction systems.

What the evidence says about pairing. A 2003 US patent on processing rail test data (US 6,594,591 B2, filed by a rail testing contractor) states: "Not all rail defects are detectable by either the magnetic induction technique or the ultrasonic technique." It adds that combining the two "greatly reduces the number of 'false calls'". That is the patent holder's statement, not an independent study.

Known limits:

• Induction senses discontinuities in the railhead. The same patent notes that induction data carry no depth information.
• It needs a high direct current in the rail, which is why the units described in the literature are rail-bound cars, not light hi-rail trucks.
• No public benchmark of induction alone was found. TD-99-006 tested ultrasonic systems only.

Whether a given test car still carries induction in 2026 is a question to put to the contractor; the 2014 preamble is the latest public FRA statement found.

Key takeaway: induction is a head-only cross-check on UT, not a stand-alone internal search.

Where does hand UT fit among internal rail defect detection methods?

Hand UT is the deployed verification method: it confirms and sizes what a test car flags, covers rail the car couldn't test, and scans weld bases the car can't reach. It is slow, so it is never the primary search on long lines.

Hand UT does four jobs in practice:

1. Verification. In stop-and-verify testing the operator checks each suspect with a hand probe before marking a defect. For owners using continuous testing, 49 CFR 213.240 requires suspect locations to be field-verified within set windows. The trade-offs are compared on continuous vs stop-and-verify rail testing.
2. No-test follow-up. Where a valid search couldn't be made, 213.237(h)(1) lets the owner conduct one before the interval expires. On a short stretch that is often a hand test.
3. Weld bases. Field welds need a base scan that head-contact cars don't give.
4. Short or awkward track. Walking carts and single-probe units cover yards, turnouts and transit track; see our ultrasonic rail detection cart page.

The best recent evidence is FRA's own. DOT/FRA/ORD-26/07, Application of MAPOD to Handheld Ultrasonic Rail Flaw Inspection (Bruzek, Einbinder and Tunna, ENSCO, July 2026), applied model-assisted probability of detection to handheld rail UT. It used 512 simulated responses for flaws from 1% to 76% of head area. Its Table 24 shows how much the detection threshold matters at a 25% head-area flaw:

Worked reading. Raising the threshold from 0.20 to 0.30 drops the lower 90% bound from 95.1% to 91.4%, a fall of 3.7 points. Lowering it the other way cuts misses but, in the report's words, "too low a detection threshold can result in an unacceptable number of false positives". Those are simulated results for one setup, not a field detection rate.

Known limits. ORD-26/07 found that, with a small exception, no two operators produced statistically similar calibration results, and that North American railroads have no standardized calibration procedure for handheld rail UT. Flat probes also coupled poorly on heavy gage-face wear. Our page on false calls in rail UT covers the human-factor side.

Key takeaway: hand UT is only as consistent as its calibration, and FRA's 2026 study found that calibration varies by operator.

What has phased array shown in rail?

Phased array UT (PAUT) on a test vehicle performed as well as or better than a conventional UT truck on every defect type tested, but only as a 2016 prototype; it remains a pilot. It steers the beam electronically instead of using fixed angles.

TTCI built a PAUT roller search unit that fits in place of a conventional one and tested it over known defects on the RDTF. TD-16-050 (Witte and Poudel, November 2016) reports that "The prototype performed equivalent or better than a conventional UT inspection truck on every type of rail defect tested", and that it detected "transverse defects under shells, defects in welds, and web defects at speeds up to 20 mph". It can "inspect to the base of the rail under the web".

The same digest states the limit just as clearly: PAUT "is still a head contacting inspection technique and so suffers the limitation of being unable to inspect the rail base flanges". TTCI said comparison in the service environment was future work. No published revenue-service POD for vehicle PAUT was found.

A second FRA project used PAUT for sizing rather than searching. FRA RR 20-16 (TTCI, September 2020) located and sized head defects for rail head repair welds, which need defect position to within ±1 in. That is a verification use. For PAUT basics see our phased array ultrasonic testing page.

Key takeaway: PAUT fixes the worn-rail and shell problems on a test track, but not the base-flange gap.

Is EMAT ready for rail flaw detection?

No: electromagnetic acoustic transducers (EMATs) have been studied for rail since at least 1980 and are still at the research-to-pilot stage. They generate sound inside the steel with a magnet and a coil, so they need no couplant.

FRA/ORD-80/45 (Alers and colleagues, Rockwell International, September 1980) found that "virtually all of the critical rail flaws can be detected in the laboratory by EMAT's", with sensitivity comparable to piezoelectric probes (TRID record). The words that matter are "in the laboratory".

The obstacle is signal strength. ASTM's EMAT guide, E1774-17(2022), describes "very low efficiency", with insertion losses of 40 dB or more, and notes that corroded surfaces can make EMAT unsuitable. The EU RIFLEX project (2014–2016) validated a flexible EMAT on a rotating rig at up to 80 km/h and called it a prototype "ready for further development" (CORDIS report). Its higher speed and coverage figures were projections.

Key takeaway: EMAT removes the water but not the need for a tiny, constant gap on a rough rail head.

How far have laser and air-coupled UT come?

Non-contact laser and air-coupled UT reached instrumented tests on a test track, where performance was good at walking speeds and fell sharply above about 8 km/h. No routine revenue-service system was found.

The main program ran at the University of California, San Diego, with FRA funding. FRA's DOT/FRA/ORD-12/02, On-line High-Speed Rail Defect Detection: Part II (March 2012), reports seven field tests of a non-contact prototype on FRA's hy-rail vehicle (ROSA P record). Its abstract claims blind-test reliability "exceeding industry average" for those tests.

The later air-coupled field paper is more cautious. Mariani, Nguyen, Zhu and Lanza di Scalea reported "good detection performance at the speeds of 1.6 and 8 km/h, with a much poorer performance at the speeds of 16 and 24 km/h" in two tests at the RDTF (J. Transp. Eng. Part A 143(5), 2017). A laser-air hybrid was reported earlier by Kenderian and colleagues (Insight 47, 2005). Our advanced nondestructive evaluation page reviews these programs further.

Key takeaway: the best published non-contact UT result is at 8 km/h (5 mph), slower than a normal test car.

Can guided and passive waves find internal defects?

Guided waves can detect complete breaks and larger defects over long distances, and a passive version has run at up to 80 mph on a test track, but neither finds small internal defects reliably enough to replace a search. Guided wave and long-range UT are one method under two names.

Active guided waves use low-frequency (kHz) sound that travels along the rail. In pulse-echo, Loveday, Long and Ramatlo detected and located an artificial defect 370 m from the transducers on operational heavy-haul track (Structural Health Monitoring 19(6), 2020). The artificial defect was a glued-on mass, not a crack. In finite-element and laboratory work, Gharaibeh and colleagues reported transverse defects "as small as 2 mm and 5 mm in the head and foot" (Proc. IMechE Part F 225(3), 2011). That is laboratory scale. Range and mode questions are worked through on long-range ultrasonic rail inspection.

Passive systems use no transmitter. They listen to the guided waves that the train's own wheels excite in the rail. Datta and Lanza di Scalea reported test-track runs showing "possible test speeds as high as 80 mph" and evaluated detection with ROC curves against speed, sensor position and signal quality (Experimental Mechanics 63, 2023). The paper's claim that existing methods are limited to about 30 mph is the authors'.

Known limits. Rail is a multi-mode waveguide, so signals are hard to interpret. Small flaws give small reflections, and no published study shows these systems finding joint-area defects or reporting untested rail the way 213.237(g) requires.

Key takeaway: guided waves are the strongest candidate for monitoring between tests, not a substitute for the test.

What evidence should move a method up the maturity ladder?

A method moves from pilot to deployed when it has a known-defect benchmark with enough flaws, a revenue-service trial, and a clear route to a 213.237 valid search. An owner evaluating a new technology can use these checks:

Under 213.237(i) the operator must also have "demonstrated proficiency in the rail flaw detection process for each type of equipment the operator is assigned". A new technology therefore brings a training and proficiency task, not just new hardware.

What are the limits of internal rail defect detection methods?

No method on this page finds every internal defect, and the public evidence is thinner than vendor brochures suggest. These limits apply across the ladder:

• The base flanges. Head-contact UT, induction and vehicle PAUT all miss most of the base outside the web. TD-16-050 says inspection technologies for the base flange area "are needed in order to detect all critical rail defects".
• Surface damage. Shells and heavy RCF block or scatter sound for every contact UT variant, and FRA lists them as detection problems.
• Small detail fractures. The 1998 benchmark's frequently missed flaws were mostly small DFs with no visible surface sign.
• Old benchmarks. The only public multi-system benchmark of contact UT is from 1998. Equipment has changed since, but no newer public multi-system result was found.
• No speed-versus-POD study. No public FRA or TTCI study measures contact UT detection as a function of test speed (search, Oct 2026).
• Rates don't transfer. A detection rate belongs to its system, threshold, operators, rail condition and flaw set. Don't apply TD-99-006, ORD-26/07 or a pilot's result to a different setup.
• Detection is not disposition. Finding a flaw starts the 49 CFR 213.113 remedial action; none of these methods decides it.

How EMA Industries applies this

EMA Industries provides ultrasonic rail flaw testing for transit agencies, short lines and industrial track owners, and rail weld inspection, from Southern California, mobilizing nationwide. The deployed methods on this page, contact UT for the search and hand UT for verification, no-test follow-up and weld bases, are what a 213.237 program depends on today; pilot technologies are worth tracking against the evidence checklist above. If you are scoping a 213.237 program or evaluating a new technology's evidence, see our ultrasonic examination services or contact EMA Industries.

Frequently asked questions

What is the most common method for detecting internal rail defects?

Contact ultrasonic testing is the primary method on US railroads. FRA's 2014 rail integrity rule calls ultrasonic technology "the primary technology used", with induction "used as a complementary system". Probes in roller search units or sliders ride on the rail head and send 2.25 to 5 MHz sound into the head and web.

Does FRA require ultrasonic testing for internal rail inspection?

FRA doesn't name a method. 49 CFR 213.237 requires internal rail inspections with equipment that can detect defects between joint bars, and a valid search interpreted by a qualified operator under 213.238. Ultrasonic testing is the industry practice that meets those conditions.

How accurate is ultrasonic rail flaw detection?

It depends on flaw size, type, rail surface and system setup. In TTCI's 1998 benchmark, six UT systems found 72% of 49 known flaws: 96% of transverse defects of 21% head area or larger, but only 66% of smaller ones and 33% of bolt hole cracks. Those figures belong to that test and shouldn't be applied to a different system.

Is phased array replacing conventional rail ultrasonic testing?

Not yet. TTCI's vehicle PAUT prototype performed as well as or better than a conventional UT truck on a test track in 2016, at up to 20 mph. It still couldn't inspect the base flanges, and no revenue-service detection data has been published. In rail, PAUT is a pilot for searching and a tool for sizing.

Can guided-wave monitoring replace periodic rail testing?

Not under current evidence. Guided waves can detect breaks and larger defects at long range, and a passive system has run at up to 80 mph on a test track. No published study shows them finding small internal flaws, defects inside joints or untested rail the way a 213.237 valid search must.

Sources

• 49 CFR 213.237, Inspection of rail, eCFR (text checked on Cornell LII, Oct 2026)
• 49 CFR 213.238, Qualified operator, eCFR
• 49 CFR 213.240, Continuous rail testing, eCFR
• 49 CFR 213.241, Inspection records, eCFR
• 49 CFR 213.113, Defective rails, eCFR
• FRA, Track Safety Standards; Improving Rail Integrity, final rule, 79 FR 4234, Jan. 24, 2014
• Garcia and Reiff, TTCI TD-99-006, Benchmarking Results of North American Rail-Flaw Detection Technologies, Feb. 1999
• Jeffrey and Peterson, Assessment of Rail Flaw Inspection Data, MPC 99-106, Mountain-Plains Consortium, Aug. 1999
• Witte and Poudel, TTCI TD-16-030, High-Speed Rail Flaw Detection Using Phased Array Ultrasonics, July 2016
• Witte and Poudel, TTCI TD-16-050, Characterizing Rail Flaws Using Phased Array Ultrasound, Nov. 2016
• Bruzek, Einbinder and Tunna, Application of MAPOD to Handheld Ultrasonic Rail Flaw Inspection, DOT/FRA/ORD-26/07, FRA, July 2026
• TTCI, Locating and Sizing Rail Head Defects for Rail Head Repair Welds Using Phased Array Ultrasonics, RR 20-16, FRA, Sept. 2020
• Alers et al., Development of a Prototype EMAT System for Inspection of Rails, FRA/ORD-80/45, 1980 (TRID record)
• RIFLEX project final report (EMAT rail head inspection), CORDIS, 2016
• ASTM E1774-17(2022), Standard Guide for Electromagnetic Acoustic Transducers, store page
• Coccia et al., On-line High-Speed Rail Defect Detection: Part II, DOT/FRA/ORD-12/02, FRA, 2012
• Mariani, Nguyen, Zhu and Lanza di Scalea, field test performance of a noncontact ultrasonic rail inspection system, J. Transp. Eng. Part A 143(5), 2017
• Kenderian, Cerniglia, Djordjevic and Garcia, laser-air hybrid ultrasonic rail inspection, Insight 47, 2005
• Loveday, Long and Ramatlo, guided-wave monitoring of operational rail track, Structural Health Monitoring 19(6), 2020
• Gharaibeh, Sanderson, Mudge, Ennaceur and Balachandran, long-range ultrasonic testing of rail, Proc. IMechE Part F 225(3), 2011
• Datta and Lanza di Scalea, defect detection performance of a high-speed ultrasonic rail inspection system during field tests, Experimental Mechanics 63, 2023
• Papaelias, Roberts and Davis, A review on non-destructive evaluation of rails, Proc. IMechE Part F 222(4), 2008
• Clark, Boyle and LaMacchia, US 6,594,591 B2, Method and system for processing rail inspection test data, 2003
• US Air Force, MIL-HDBK-1823A, Nondestructive Evaluation System Reliability Assessment, 2009