What problem does it solve?
A wheel, bearing or brake component on freight rolling stock can start to fail long before anyone sees it. Norfolk Southern says wheel defects are among the most serious mechanical defects in the industry. Railroads monitor the condition of railcars with sensors along the track: BNSF says it has "long used advanced technologies to monitor the condition of our railcars and tracks," and the Association of American Railroads describes trackside sensors that "capture large amounts of data as trains pass by, including wheel profiles, wheel impact loads, bearing acoustics, and component temperatures." But Norfolk Southern says the defects its Wheel Integrity System looks for are "subtle defects difficult for the human eye to identify consistently."
The Association of American Railroads describes Norfolk Southern's Wheel Integrity System as catching wheel defects early, helping prevent failures and derailments.
- The Association of American Railroads says that by analysing large volumes of real time and historical data, AI enabled systems help detect equipment and infrastructure issues early and support predictive maintenance across the US freight rail network.How Freight Rail Uses AI | Safety & Efficiency (2026)
How does it work?
- Sense the wheel or the train. Thermal sensors, acoustic sensors and high resolution cameras mounted at fixed wayside portals capture temperature, sound and images of every wheel, bearing and railcar that passes, without slowing the train.
- Build a picture of healthy equipment. Models trained on large volumes of past readings and images learn what a normal wheel profile, a normal bearing temperature curve or a normal surface looks like for that equipment type.
- Flag the deviations. The system scores each passing wheel or railcar and flags cracks, overheating, unusual wear or other deviations that need a closer look, ranked by urgency.
- Route to the right team. A flagged defect goes to the monitoring desk or the mechanical team nearest the railcar's next stop, with the image or reading attached, so the car can be pulled from service or repaired before its next long run.
- Confirm and feed back. A mechanical inspector confirms the defect on the physical car; confirmed and missed cases both feed back into the model so thresholds improve over time.
- Audience
- Employee facing
- Autonomy
- Assist
- Adoption
- Early adopters
- Channels
- Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has disclosed a measurable outcome yet.
Value drivers: Risk and loss reduction, Lower cost to serve, Speed and cycle time.
Indicative value
A mid sized freight railroad running 1 million loaded railcar trips a year
USD 750,000 to USD 36 million
In service failure and derailment cost avoided per year
How this is calculated
Formula: carTrips * inServiceFailureRate * earlyCatchShare * costPerInServiceFailure. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Loaded railcar trips per year carTrips, railcar trips per year | 1,000,000 | 1,000,000 | The reference railroad. |
| Share of trips with a wheel, bearing or brake related in service failure or setout inServiceFailureRate, fraction of trips | 0.001 | 0.002 | Editorial assumption, replace with your own mechanical setout and derailment data. |
| Share of those failures caught early instead of happening in service earlyCatchShare, fraction of at risk trips | 0.1 | 0.3 | Editorial assumption. BNSF and Norfolk Southern both report scaled, production wheel inspection systems, but neither discloses a share of failures prevented, so this range is not benchmarked to their evidence; replace it with your own catch rate. |
| Cost of an in service wheel, bearing or brake failure (setout, delay, repair, claims) costPerInServiceFailure, USD per event | 15,000 | 60,000 | Editorial assumption; replace with your own cost per mechanical setout or incident, which varies hugely by whether it causes a derailment. |
What it leaves out: Gross avoided cost only, built entirely on editorial assumptions because neither named deployment on this page discloses a catch rate or a percentage outcome. It leaves out the cost of the sensors, cameras and models themselves, false positive handling, and any effect on dwell time or car cycle time.
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
BNSF Railway
United States · Logistics and transportation · 2025
BNSF Railway's condition based maintenance team uses thermal sensors and machine vision at wayside detectors to monitor more than 1.5 million wheels in motion across its network. Machine vision cameras inspect wheel surfaces for cracks and defects, and AI algorithms analyse the resulting wayside detector readings to predict maintenance needs before a breakdown, rather than reacting to a failure in service.
No outcome disclosed.
Norfolk Southern
United States · Logistics and transportation · 2025
Norfolk Southern built a standalone Wheel Integrity System with its own Data Science and AI team, developed with integration support from the Georgia Tech Research Institute. Six synchronized cameras capture about 55 high resolution images per wheel as trains pass at up to 70 mph, and AI algorithms analyse the images to detect subtle defects that are difficult for the human eye to identify consistently. The first site went live near Chicago on November 24, 2025. The new system follows the railroad's existing Digital Train Inspection (DTI) portals, a separate, earlier system that scans entire trains; the DTI portals had already identified and removed from service over 50 wheels with issues since January 2025. Unlike DTI, the new system zeroes in on wheels specifically, and it has pinpointed a vendor wheel casting flaw that triggered an industry recall and seven confirmed defects across North America.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Historical wayside detector readings and images linked to confirmed defects and failures
- Railcar and wheel maintenance history, by equipment type
- A network map of where inspection and repair capacity exists, to route flags usefully
Systems to integrate
- Wayside detector and camera portal network
- Mechanical and maintenance management system for work orders
- Car scheduling or yard system, to know where a flagged car will next be reachable
- Monitoring desk alerting and dashboards
Complexity: High
The sensing hardware (thermal, acoustic, high resolution camera portals) is a capital project in its own right; the AI work is training reliable models per equipment type, keeping false positive rates low enough that mechanical teams trust the flags, and routing a flagged railcar to a stop where it can actually be inspected or repaired.
- 1
Start with one defect type and one corridor
Pick a high value defect type (wheel cracks, bearing overheating) with existing sensor coverage on a busy corridor, so there is enough data and enough traffic to prove the model quickly.
- 2
Get the sensor data trusted before trusting the model
Many first flags turn out to be a dirty lens or a miscalibrated sensor, not a real defect. Build in sensor health checks before tuning detection thresholds.
- 3
Route flags to where action is possible
A flagged car is only useful information if it reaches a point on the network with the inspection or repair capacity to act on it before the car's next long run.
- 4
Keep a person on every removal decision
A model flag advises; a qualified inspector confirms the physical defect and decides whether the car is pulled from service, in line with the railroad's mechanical rules.
- 5
Track false positives as closely as catches
A high false positive rate burns the trust of mechanical teams faster than a missed defect does. Report both the confirmed catch rate and the false positive rate every month.
- 6
Expand by defect type, then by corridor
Reuse the sensing and model pattern for the next defect type or corridor once the first one has a track record of confirmed catches and a manageable false positive rate.
Guardrails
- A qualified mechanical inspector confirms every flagged defect before a car is pulled or repaired
- Flags feed maintenance planning; they do not directly command a train to stop or slow
- Sensor and camera health checks, so faulty hardware is not read as a defective railcar
- Model changes tested against a held out set of confirmed past defects before release
KPIs to instrument
- Confirmed defect catches per period, and the estimated cost of the in service failure avoided
- False positive rate on flagged wheels and railcars
- Time from a flag to a confirmed inspection
- In service wheel, bearing and brake related setouts and derailments, before and after coverage
Human in the loop
Monitoring desk staff and mechanical inspectors review every flagged wheel or railcar and decide whether it is pulled from service or scheduled for repair. Reliability engineers review confirmed catches and any missed failures on a regular cycle to tune detection thresholds.
Common failure modes
- Alert fatigue at the monitoring desk
- Too many low confidence flags and staff start clearing them without a real look. Report the confirmation rate and tune thresholds until flags are worth opening.
- A flag nobody can act on
- A defect is flagged on a railcar that has already left the network segment with inspection capacity. Route flags to where the car will next be reachable, not just to a dashboard.
- Sensor drift read as a fleet problem
- A miscalibrated camera or thermal sensor produces a wave of false flags that looks like a real fleet issue. Monitor sensor health separately from defect detection.
- Treating an advisory flag as a control action
- A flag is wired directly into an automatic stop or slow order without a human check. Keep the model advisory to mechanical teams unless the safety case and conformity assessment for an automated control action have been done separately.
What are the risks and rules?
EU AI Act
Depends on design
A system that flags a wheel or railcar for a qualified inspector to confirm is advisory and usually minimal risk. Under Article 6(1) it is high risk when both conditions hold: the same detection logic is built into a safety component of rolling stock or track equipment (or is itself such a product) covered by Directive (EU) 2016/797 on the interoperability of the rail system, which sits in Annex I Section B, for example if a flag were wired to trigger an automatic stop or speed restriction without a human check, and that directive requires a third party conformity assessment of the product. Under Article 2(2), as amended by Regulation (EU) 2026/1744, a high risk system of that kind is not subject to the full AI Act: only Article 6(1), Article 60a and Articles 102 to 112 apply directly, and Articles 57, 58 and 59 apply only so far as the high risk requirements have been integrated into the interoperability directive. The substantive high risk requirements reach the system through that directive instead, which Article 106 of the AI Act amends to require rail delegated and implementing acts to take those requirements into account.
Rules that apply
Guidance
- Article 2, scope (European Union, Europe). Article 2(2), as amended by Regulation (EU) 2026/1744, says that for high risk AI systems under Article 6(1) related to products covered by the Annex I Section B laws, such as rail interoperability, only Article 6(1), Article 60a and Articles 102 to 112 apply directly; Articles 57, 58 and 59 apply only so far as the high risk requirements have been integrated into that Union harmonisation legislation.
- Annex I, list of Union harmonisation legislation (European Union, Europe). Section B, point 17, lists Directive (EU) 2016/797 on the interoperability of the rail system as the product law behind Article 6(1) for rail equipment.
- Article 6, classification rules for high risk AI systems (European Union, Europe). Sets the two conditions in paragraph 1: (a) the AI system is a safety component of a product, or is itself a product, covered by the Union harmonisation legislation listed in Annex I, such as rail interoperability; and (b) that product or AI system is required to undergo a third party conformity assessment under that legislation with a view to being placed on the market or put into service.
- Article 106, amendment to Directive (EU) 2016/797 (European Union, Europe). Adds a paragraph to Article 5 of the rail interoperability directive requiring that, when the Commission adopts delegated and implementing acts on rail safety components, it takes into account the AI Act's Chapter III, Section 2 requirements for high risk AI systems. This is how the AI Act's substantive requirements reach rail equipment instead of applying directly.
- Digital Omnibus on AI, Regulation (EU) 2026/1744 (European Union, Europe). The regulation of 8 July 2026 that amended Article 2(2) to limit the AI Act's direct application to products covered by Annex I Section B, including rail interoperability, to Article 6(1), Article 60a and Articles 102 to 112, with Articles 57 to 59 applying only so far as integrated into the sectoral law.
Controls to put in place
- Model and sensor inventory with an owner per detection model and the equipment it covers
- Documented separation between advisory detection models and certified train control and protection systems
- Written record of every confirmed catch and every known missed defect
- Cybersecurity controls on data flows from wayside sensors and camera portals, in line with NIS2 where it applies
Frequently asked questions
- What results have freight railroads reported from AI wheel and rolling stock inspection?
- BNSF says its AI algorithms sift through more than 35 million wayside detector readings a day and its machine vision processes over 2 million wheel images daily, across a network that monitors more than 1.5 million wheels in motion. Norfolk Southern built its own, standalone Wheel Integrity System with its in house data science and AI team; it follows the railroad's earlier, separate Digital Train Inspection portals, which had already identified and removed from service over 50 wheels with issues since January 2025. The new system pinpointed a vendor casting flaw, and Norfolk Southern says that finding, coupled with its root cause investigation, triggered an industry recall. Neither railroad has published a percentage reduction in failures or derailments from these systems.
- Does this replace wayside detectors and mechanical inspectors?
- No. It adds AI analysis on top of wayside sensors and camera portals that many railroads already operate, and a qualified mechanical inspector still confirms every flagged defect before a railcar is pulled from service or repaired.
- Is this the same as predictive maintenance for industrial equipment?
- The pattern, anomaly detection on sensor data, is the same, but freight rail's own wayside detector and camera portal network, its safety regime and its car routing constraints (a flagged car has to be reachable at a point with repair capacity) make it different enough in practice to plan separately from an industrial or energy plant's predictive maintenance programme.
- What should stay with a qualified inspector?
- The final call on whether a flagged defect is real and whether a railcar is pulled from service or repaired. Neither BNSF nor Norfolk Southern discloses how its AI flags are wired into day to day operations, so treat the recommendation, not a documented fact about either deployment: keep the AI flags advisory to mechanical and monitoring desk staff, and do not wire them into a certified train control system, unless a separate safety case and conformity assessment for an automated control action have been completed.
How to cite this page
Blits.ai AI Use Case Library, "AI predictive maintenance for freight rail rolling stock", last verified 28 September 2026, https://www.blits.ai/ai-use-cases/freight-rail-rolling-stock-predictive-maintenance. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 28 September 2026: First published