Every year, thousands of freight wagons across Europe are pulled from service for unplanned maintenance — a bearing that overheated, a wheel flat that went undetected, a gearbox defect that progressed silently between inspections. Each unplanned removal costs operators between EUR 2,000 and 10,000 when you factor in traction, crew re-scheduling, and cargo delays. The technology to detect these faults early exists. So why aren’t freight wagons covered in smart sensors?
The answer is deceptively simple: there is no power socket on a freight wagon.
The power problem nobody talks about
Passenger trains have electrical systems that can feed onboard sensors. Freight wagons do not. Most lack electrical couplings between vehicles entirely, and adding wired power would require ATEX-compliant installation — a certification cost that makes no sense for a logistics asset built to move cargo at minimum cost. Even battery-powered solutions fall short: a freight wagon can stay in service for 50 years, and no battery lasts that long without replacement. The recurring cost of accessing remote wagons to swap batteries destroys the business case before it starts.
This is why, despite decades of research into vibration-based condition monitoring, commercial freight fleets still rely primarily on trackside detection systems — hot axle box detectors (HABD) and wheel impact load detectors (WILD) — that can only catch defects when a wagon happens to pass an instrumented checkpoint. By then, the damage is often already done.
A different architecture for a different problem
Within the Chips JU EdgeAI project, demonstrator VCD4.3 takes a fundamentally different approach. Instead of trying to make a conventional sensor node last longer on a battery, the team led by NeuroControls eliminated the two biggest power consumers entirely.
First, the onboard wireless radio was replaced with a passive UHF RFID link. Rather than transmitting data to a carriage-level gateway (which would need its own power supply), the sensor node simply writes its classification results into RFID tag memory. When the wagon passes a trackside reader — the same type of infrastructure already used for wagon identification — the data is harvested using the reader’s own RF energy. No radio power consumed on the wagon at all.
Second, the conventional FFT-based vibration analysis was replaced with a neuromorphic time-domain classifier. Traditional approaches transform raw vibration signals into frequency spectra — a computationally expensive operation that dominates the processor’s power budget. The VCD4.3 classifier works directly on temporal features, inspired by how biological neural systems process sensory information. The result: more than 50% fewer instructions per classification cycle, running on a standard STM32 microcontroller without any specialised hardware accelerator.
Together, these two changes bring the total power budget low enough for a small piezoelectric energy harvester — mounted on the bearing housing where vibration energy is abundant — to sustain the sensor indefinitely. No battery. No wired power. No maintenance access for the life of the wagon.
Not just detecting faults — grading them
There is another challenge specific to freight that most condition monitoring research overlooks. On a passenger train, any detectable vibration anomaly triggers a maintenance action because ride comfort is paramount. On a freight wagon, operators need to know not just whether a defect exists, but how severe it is. Intervening on every incipient fault across a fleet of thousands of wagons would be economically ruinous. The classifier must therefore distinguish between early-stage and advanced degradation — a substantially harder problem than simple fault-present/fault-absent detection.
The VCD4.3 classifier achieves over 93% accuracy on controlled test data across inner race, outer race, and roller element bearing faults at multiple severity levels. On real-world field recordings from 15 wagons under varying speeds and loads, overall accuracy reaches 85%, with roller element defects being the hardest to catch — consistent with published literature, as roller faults produce the weakest vibration impulse of the three bearing fault types.
From sensor to dashboard
Of course, detecting a fault on a wagon is only useful if someone acts on it. GNT, the consortium partner responsible for backend infrastructure, contributed a containerised data pipeline built on Apache Kafka and OpenSearch. Classification results flow from the trackside reader through a secure edge gateway — authenticated via mutual TLS with X.509 certificates — into a Kafka broker that routes data to subscribers. Fleet operators see health status and alerts on Grafana dashboards; maintenance planners receive automated notifications when a defect severity crosses an intervention threshold.
The entire backend runs as Docker containers, deployable on any standard IT infrastructure without vendor lock-in or proprietary middleware — an important consideration for freight rail operators who typically run lean IT departments.
Does it make economic sense?
To answer this question, the team built a large-scale fleet simulation modelling 51 years of operations for a freight fleet growing from 165 to nearly 1,900 carriages. The simulation incorporates maintenance costs, inspection intervals, wayside checkpoint economics, and investment payback over a 10-year loan period.
The conclusion: investment costs are recovered within the loan period, with cumulative savings growing as the monitored fleet expands. At a target sensor unit cost below EUR 25, fleet-wide retrofit becomes viable even for logistics assets with long payback horizons — a market segment that conventional condition monitoring has never been able to address economically.
Complementary, not competing
It is worth emphasising that on-asset sensors are not intended to replace trackside detection systems. HABD and WILD installations remain a valuable last line of defence. But they are reactive by nature — they catch defects that have already reached a failure threshold at a fixed geographic point. On-asset monitoring adds three things that wayside systems cannot provide: predictive lead time (weeks to months of early warning), continuous spatial coverage (no gaps between checkpoints), and alignment with emerging regulatory frameworks for condition-based maintenance — accelerated by incidents such as the 2023 Gotthard base tunnel derailment linked to an axle bearing failure.
Wayside and onboard monitoring are complementary layers. VCD4.3 demonstrates that the onboard layer can now be delivered at a power budget and unit cost that make fleet-wide deployment not just technically feasible but economically essential.
What comes next
The demonstrator has validated the core concept, but open items remain for productisation: improving roller element detection through temperature sensor fusion, optimising RFID read range for trackside deployment at operational distances, and integrating a designed-but-not-yet-implemented cryptographic security architecture — including secure elements, TPM attestation, and blockchain-anchored chain of custody — to protect classification data against tampering.
The path from proof-of-concept to revenue-service deployment is an engineering integration challenge, not a research gap. The hard question — can you do useful AI on a freight wagon with no power supply? — has been answered.
Blog signed by: NeuroControls team