Research brief · May 2026

Rail AI at SBB

What Switzerland's federal railway, and its peers, publicly do with machine learning. Compiled from primary sources for a master's thesis application.

SBB · DB · SNCF · Network Rail · NS
11 SBB programs · 4 peer operators · 8 problem areas

01 · SBB programs

Computer vision

Center of Competence in Machine Perception

SBB's named group for vision-based infrastructure inspection. Co-led by Dr. Ilir Fetai (Albanian, see section 02) and Dr. Andre Roger.

  • Camera-equipped inspection trains feed defect-detection models
  • Stated open problem: safety-critical model reliability under distribution shift, weather, lighting, track conditions
  • Stack: LatticeFlow (ETH spin-off, model robustness), Siemens, ETH Zurich
Multi-agent RL

Flatland Challenge

Multi-agent reinforcement learning for train rescheduling on a 2D rail grid. SBB ran the original 2019 edition, then co-organized the NeurIPS 2020 edition jointly with Deutsche Bahn and SNCF.

  • ~700 teams in round 2
  • Open-source simulator and follow-on environments
  • The canonical academic-industry RL benchmark in rail
Predictive maintenance

Wheel Load Checkpoints

Wayside strain-gauge sensors capture 8 measurements per wheel pass. Cyclic-shift-invariant neural nets exploit the periodic sensor geometry to detect:

  • Flat spots
  • Shelling
  • Non-roundness
Physics-informed ML

Bridge lifecycle ML with ETH Zurich

Active PhD: Sophia Kuhn in Walter Kaufmann's bridge group, co-supervised by Fernando Pérez-Cruz (ETH CS).

  • ML models for reinforced-concrete rigid-frame railway bridges
  • Direct SBB collaboration with structural engineer Marius Weber
  • Surrogate modeling of scientifically constrained systems
Geospatial CV

Vegetation control from orthophotos

Helicopter-captured orthophotos plus Picterra's geospatial platform detect neophyte plants along tracks.

  • 6 detector classes
  • 60-90% per-class accuracy
  • Operational pipeline, not a research demo
XR + AI

Fleet maintenance digitization

SBB and BCG: extended reality and AI applied to fleet maintenance workflows. Public framing is reliability and cost reduction.

Systems

Smart Rail 4.0

Industry-wide architecture program (SBB, BLS, SOB, RhB, VöV) covering traffic management, electronic interlocking, localization, and automatic train operation.

  • GoA 2 ATO trial on Lausanne-Villeneuve, 2018
  • Remote-control loco trials with Alstom, Feb-Mar 2024
Open source

github.com/SchweizerischeBundesbahnen

110+ public repos. ML-relevant subset:

  • sbb-ml-models, mobile object detection for train components (lighting, side windows, loudspeakers). Archived.
  • sbb-ml-ios, Core ML + Combine inference
  • sbb-ml-android, TFLite / YOLO inference
  • SBB-IIoT-to-SAP-PdMS, edge preprocessing for SAP Predictive Maintenance Services
University ties

Thesis and research collaborations

ZHAW · Timetable planning

SBB-funded research on automated timetable generation.

EPFL + ETH · CCMH

Co-Creating Mobility Hubs. SBB-funded, jointly run on station and mobility-hub transformation.

HSLU · NLP on customer feedback

Dominik Finzer's 2023 thesis on topic modeling of free-text customer comments. Now at SBB Customer Experience & Analytics. A documented thesis-collaboration template.

StationRank · academic

Markov-chain network-risk model built on SBB's published daily train-traffic streams.

02 · Direct outreach · Ilir Fetai

Strongest angle

Why this is the lead lane

Fetai is Albanian, teaches at the main Albanian-language university in the region, and leads the SBB group most aligned with an AI thesis. A direct, Albanian-language outreach from Roni lands in a category that no German-speaking applicant can replicate.

Role at SBB
Head of the Center of Competence on Machine Perception since November 2019. Also Head of AI in Railinspect, covering all AI for railway infrastructure monitoring and maintenance. At SBB since 2016.
PhD
University of Basel, 2016, under Prof. Heiko Schuldt (DBIS group). Thesis: Cost- and Workload-based Data Management in the Cloud. Three protocols: C3 (adaptive consistency), Cumulus (adaptive data partitioning), QuAD (quorum-based replication).
Academic ties
Professor at South East European University (SEEU, Tetovo) teaching Service-Oriented Architectures, Distributed Computing, Big Data Technologies. Lecturer at Fernfachhochschule Schweiz since 2011, leads the Web Technologies department.
Trajectory
Distributed systems → enterprise AI for safety-critical applications. Not a lifelong ML researcher; he came to ML through an applied infrastructure problem.

Notable publications

  • Jüngling, Fetai, Rogger, Morandi, Peraic (2022). On the Track to Application Architectures in Public Transport Service Companies. Applied Sciences 12(12), 6073. The most directly relevant paper: ML models, frameworks, and AI services as building blocks in IT architectures of transport companies, framed around the gap from POC to operational deployment.
  • Fetai, Stiemer, Schuldt (2017). QuAD: A quorum protocol for Adaptive Data Management in the Cloud. IEEE Big Data.
  • Stiemer, Fetai, Schuldt (2015). Comparison of Eager and Quorum-based Replication in a Cloud Environment. IEEE Big Data.
  • Fetai, Murezzan, Schuldt (2015). Workload-Driven Adaptive Data Partitioning and Distribution: The Cumulus Approach. IEEE Big Data.
  • Speaker at AMLD EPFL 2022, framing himself as working on enterprise-readiness of AI solutions.

Outreach playbook

Channel. Email direct, not via SBB careers form. SBB email pattern: firstname.lastname@sbb.ch. Cross-reference via LinkedIn before sending.

Language. Open in Albanian. One short paragraph, then switch to English for the technical pitch. This signals the connection without making him do extra work.

Subject line angle. Reference the 2022 Applied Sciences paper or the AMLD talk by name. Proves you read.

Pitch hooks. Two of these, not all. (1) Bioinformatics ML transferring to safety-critical inspection-CV under distribution shift. (2) The POC-to-production gap his Applied Sciences paper named, framed as the kind of thesis problem you want. (3) His own trajectory from distributed systems to ML mirrors your willingness to cross domains from bioinformatics.

Ask. Concrete: a 20-minute call to discuss whether a thesis fits, and what supervision route works (SBB-direct vs co-supervised with a university PI).

03 · Peer operators

Germany

Deutsche Bahn / DB Systel

Strategic AI bets: predictive maintenance and automated dispatching.

  • AIM, acoustic infrastructure monitoring for escalators, mic plus structure-borne sensors
  • Vsion.ai, image and video analysis of pantographs and platforms
  • S-Bahn wheelset RUL forecasting, elevator RUL dashboards
  • Public talk: Reinforcement Learning of Train Dispatching at Deutsche Bahn (Tobias Keller)
  • Konux partnership on smart point-machine sensors, claimed 25% maintenance cost reduction
France

SNCF

  • DataLab Réseau, 350+ datasets, internal data-rooms model
  • Decision Science group, Miramas marshalling yard: weeks of OR collapsed to minutes
  • Physical plus data hybrid models for track geometry and rail fatigue
  • Predictive maintenance via Trealis since 2013
  • Internal SNCF Group GPT launched Jan 2024
  • AI and Sustainable Mobility Chair with École Polytechnique
  • Quantum optimization on the public roadmap
UK

Network Rail

  • Intelligent Infrastructure, 30k+ sensors feeding the Insight platform on Azure (Data Factory, Databricks, Azure ML)
  • UK patent GB2620615 granted Oct 2024: prediction of cyclic top events
  • Industry-wide AI in Rail Action Plan, launched April 2026 by GBRX, with an AI Incubator Accelerator
  • Operator work: LNER delay-risk ML, South Eastern fault-detection cameras, Lewisham CCTV trespass detection, Waterloo LiDAR crowd monitoring
Netherlands

NS

  • SLT HealthMeter, per-unit failure probability from diagnostic messages, maintenance history, timetable performance
  • Advanced Analytics stack: Python, PyTorch, YOLO, segmentation on Azure (implies CV for asset inspection)
  • Co-organizes academic delay-prediction research, XGBoost with network-topology features on Dutch rail
Coverage gaps

Not surveyed

Within the research budget no primary sources at comparable depth were found for:

  • JR East / JR Central
  • Trenitalia / FS Italiane
  • China Railway / CRRC

04 · Problem areas in rail ML

Predictive maintenance from sensors

Vibration, acoustic, strain. DB Konux switches · SBB WLC · NS SLT HealthMeter.

CV for track and asset inspection

Inspection trains, drones, on-board cameras. SBB + LatticeFlow + Siemens · SBB Picterra · South Eastern UK.

Multi-agent RL for dispatching

The Flatland line. SBB + DB + SNCF, NeurIPS 2020.

Timetable optimization, OR + ML

SBB-funded ZHAW automated timetabling. SNCF marshalling-yard solver.

Demand and delay forecasting

NS XGBoost with network features. LNER delay-risk model.

NLP on operational text

SBB / HSLU customer-comment topic modeling. SNCF Group GPT.

Structural ML for infrastructure

ETH-SBB bridge surrogates. SNCF physics-plus-data rail-fatigue models.

Energy-aware trajectory optimization

ETH IVT theses on energy-efficient driving and railway power peaks.

05 · Application angles

  1. Inspection-train CV under distribution shift. Center of Competence in Machine Perception (Fetai, Roger). The SBB-LatticeFlow-Siemens-ETH stack explicitly frames distribution shift as the open safety-critical problem.
  2. Multi-agent RL for rescheduling. Flatland and the NeurIPS 2020 SBB-DB-SNCF edition. Ask where production traffic management sits today on the RL ↔ OR spectrum.
  3. Physics-informed ML for infrastructure. ETH-SBB bridge work (Kuhn, Kaufmann, Pérez-Cruz). Surrogate modeling of scientifically constrained systems, a natural bridge from bioinformatics ML.
  4. Applied NLP on operational text. HSLU customer-comment topic modeling is a documented template. Adjacent targets: incident reports, driver feedback, maintenance logs.
  5. On-device CV for rolling stock. The sbb-ml-models mobile pipeline is archived; ask whether on-device CV is dormant or has migrated into the inspection-train pipeline.

06 · Open questions

  1. Is the Center of Competence in Machine Perception the right home for an AI thesis, or do theses typically sit in BI Data Science, Customer Experience Analytics, or an embedded Infrastructure team?
  2. Does SBB still co-supervise theses externally with ETH, EPFL, ZHAW, HSLU? Apply direct, or via a university PI?
  3. Status of the Flatland line internally: rescheduling RL still active, moved to OR-heavy hybrids, on hold post-2021?
  4. Active needs around physics-informed or scientifically-constrained ML, where surrogate-model intuition transfers?
  5. Inspection-train CV bottleneck today: label scarcity, distribution shift, or downstream safety certification? Each implies a different thesis.