In our first blog post, we described how VCD4.3 eliminates the power barrier that has kept smart sensors off freight wagons for decades. The core idea — replace the radio with passive RFID, replace FFT with time-domain classification, then harvest enough vibration energy to run the whole thing forever — sounds almost too clean on paper. This post is about the engineering reality behind those three claims: what actually happens when you try to run a bearing fault classifier on microwatts.
Why FFT had to go
The conventional approach to vibration-based condition monitoring starts with frequency analysis. You sample the accelerometer at 10–25 kHz, buffer a window of data, compute a Fast Fourier Transform, then look for characteristic frequency peaks — outer bearing, inner bearing, roller bearing defect frequencies. This works reasonably well for passenger rolling stock, where any detectable anomaly triggers intervention because ride comfort is paramount — passengers associate clunking with low quality, and operators can afford the sensor hardware. Freight is a different problem. Operators need to know not just whether a defect exists, but how severe it is. And this is where frequency analysis falls short: five small flat spots spaced around a wheel produce the same defect frequency as one large flat spot. The amplitude of the spectral peak grows with the number of defects, not necessarily with the severity of any single one. Frequency alone cannot reliably distinguish a cluster of minor surface marks from a single deep gouge that will fail catastrophically.
Even setting that limitation aside, FFT is expensive to compute. A 1024-point transform is roughly 5,000 multiply-accumulate operations. That sounds modest until you realise the sensor node must do this continuously, at least once per wheel revolution, across multiple axes. On an STM32L4-class microcontroller, a single FFT cycle consumes around 15–20 µJ. Add windowing, peak detection, and threshold logic, and the full classification pipeline approaches 50 µJ per inference. For a battery-powered node sampling at 5 Hz, that is 250 µJ per second — sustainable on a coin cell for perhaps two to three years, but utterly out of reach for an energy harvester delivering single-digit milliwatts intermittently. So FFT fails freight twice: it cannot grade severity, and it costs too much energy.
The VCD4.3 classifier takes a different path. Instead of transforming the signal into the frequency domain, it works directly on temporal features: peak-to-peak amplitude, crest factor, kurtosis, RMS envelope characteristics, and inter-pulse timing patterns. The computational cost drops to under 2,000 operations per inference, and critically, there is no large buffer to fill. The classifier can operate on short overlapping windows, waking briefly, computing, and sleeping again within milliseconds.
The measured result: more than 50% reduction in instruction count compared to an equivalent FFT-based classifier achieving similar accuracy on the same test dataset. Combined with aggressive clock gating — the STM32 runs at its lowest stable frequency during classification — the total energy per inference drops below 10 µJ. That is the number that unlocks energy harvesting.
Harvesting energy from the defect you are trying to detect
There is an elegant irony in the VCD4.3 power architecture: the very vibration that signals a bearing defect is also the energy source that powers the sensor detecting it. A healthy bearing produces smooth, low-amplitude vibration. A damaged bearing produces sharp impulses — exactly the kind of high-dV/dt mechanical transients that piezoelectric harvesters convert most efficiently.
The harvester is a sealed piezoelectric cantilever (Mide Volture series) bolted rigidly to the bearing housing. There is no compliant mounting, no damping — the harvester is deliberately shock-coupled to capture the instantaneous acceleration spikes from bearing defects and wheel flats. At typical freight speeds of 35–50 km/h on an 810 mm wheel, impact events occur at 3.8–5.5 Hz. Each impact generates a voltage spike that can exceed 50 V open-circuit, though the energy per pulse is small — on the order of tens of microjoules.
A power management IC (LTC3588-1) rectifies the AC output, regulates it to 3.3 V, and charges a small capacitor bank. The deliberate choice here is electrolytic capacitors, not a supercapacitor. A supercapacitor stores more energy but takes 30–45 minutes to cold-start from empty — unacceptable for a system that must begin classifying shortly after a wagon enters service. There is also a regulatory dimension: freight wagons routinely carry hazardous cargo — fuels, chemicals, grain dust — and the energy stored in a supercapacitor can exceed ATEX ignition thresholds. Keeping the storage bank small enough to stay below those thresholds simplifies certification for use in potentially explosive atmospheres. The capacitor bank reaches operating voltage within seconds of the first impact, enabling a fast first classification cycle.
The duty cycle emerges naturally from the energy balance. At 40 km/h with a moderate flat spot, the harvester delivers roughly 200 µW average. The classifier consumes under 10 µJ per inference. Even accounting for MCU wake-up overhead, leakage, and the RFID tag write, the system can sustain one classification every 10–20 seconds — more than sufficient for condition monitoring, where fault progression timescales are measured in days, not seconds.
Writing results without a radio
The third pillar of the power budget is communication — or rather, the elimination of it. A conventional BLE or LoRa transmission costs 10–100 mJ per packet, dwarfing the classification cost by three orders of magnitude. Even the most efficient radio protocols destroy the harvesting budget.
VCD4.3 sidesteps this entirely. The MCU writes its classification result — a compact 23-byte payload containing sensor ID, GPS stamp, fault type, and severity level — into a dual-interface RFID tag via a wired I²C connection. No RF energy is consumed on the wagon. When the wagon passes a trackside reader at a checkpoint, the reader interrogates the tag using its own RF field and the tag backscatters the stored data passively. From the sensor node’s perspective, the “transmission” costs nothing more than an I²C write — roughly 5 µJ.
The reader infrastructure is straightforward. Trackside RFID readers are already part of European rail operations for wagon identification. Adding condition monitoring data to the existing read cycle is a software configuration change, not a new hardware deployment. The short read range — tens of centimetres to around a metre, depending on antenna geometry — also limits the attack surface for RF eavesdropping, complementing the cryptographic payload protection built into the data format.
What the numbers mean for the fleet operator
For a freight operator evaluating condition monitoring, the technical details above collapse into three practical implications. First, the sensor has no consumables. There is no battery to replace, no SIM card to manage, no firmware update channel to maintain. Once installed on the bearing housing, the sensor operates autonomously for the service life of the wagon — potentially 24 years. Second, the infrastructure cost is minimal. Trackside RFID readers already exist across European rail networks for wagon identification. Extending their function to collect condition monitoring data is a configuration task. Third, the classifier runs entirely at the edge. No raw vibration data leaves the wagon. The trackside reader receives only a classification result and severity grade. This eliminates bandwidth costs, reduces data privacy exposure, and makes the system independent of cellular or satellite connectivity — a genuine constraint for freight routes through tunnels, rural areas, and cross-border corridors where roaming agreements do not cover IoT devices.
The fleet simulation built alongside the demonstrator models these economics over 51 years for a growing fleet. At a target sensor unit cost below EUR 25, retrofit achieves payback within the financing period. The deeper finding is that savings compound: as more wagons are monitored, the frequency of unplanned service interruptions drops, and operators can shift from calendar-based to condition-based maintenance intervals — extending asset availability while reducing total maintenance spend.
From demonstrator to deployment
The VCD4.3 demonstrator, developed within the Chips JU EdgeAI project, has validated each subsystem independently. The neuromorphic classifier was trained and tested on field recordings from 15 wagons under five speed conditions, generating over 12,600 classified datasets. The energy harvesting chain has been validated on a bench rig reproducing the mechanical excitation profile of a defective bearing. What remains is closing the loop: running the classifier live on harvested power in a trackside integration test.
Open challenges remain. Temperature sensor fusion would improve roller element detection. The RFID antenna geometry needs optimisation for the specific mounting constraints of European freight bogies. And the cryptographic security layer — designed around secure boot, TPM attestation, and tamper-evident payload signatures — is architecturally complete but awaits hardware integration on the final sensor board.
None of these are research questions. They are engineering tasks with clear specifications and known solutions. The research question — whether useful AI inference is possible on a freight wagon with no external power supply — has been answered. The next step is a pilot deployment on a revenue-service fleet.
Blog signed by: NeuroControls team