Physics first, then judgement
Detecting a failing bearing or a damaged gear tooth from vibration is textbook condition monitoring, with decades of literature and an ISO standard for severity. The hard part is doing it on a rock truck, which is never at steady state. This page is how.
How RockLogik is trained, and how it works
Why a haul truck defeats ordinary vibration analysis, how the sensors make the signature hold still, and how the system learns a fleet's normal before it judges anything against it.
Read the transcript Hide the transcript
Every defect in a bearing or a gearbox has a frequency, and every one of them is a multiple of shaft speed. On a rock truck, shaft speed changes every few seconds. Loading, hauling, dumping, returning, queueing. Five regimes, five different signatures from the same healthy gearbox. Average them together and a fault disappears into the noise. RockLogik fits six IVAN sensors flush along the drivetrain, and one Firefly hub in the cab. Firefly listens to the truck's own bus for gear, speed, load and grade. It only listens; it never transmits. When the truck enters a known operating window, loaded on grade in third, say, it tells the sensors to capture. Each sensor records a few seconds at full bandwidth. Its magnetometer reads the rotating shaft through the housing, so the vibration is resampled against shaft angle, not time. Bearing and gear frequencies land at fixed orders, whatever the speed. The sensor computes the diagnostic features on board: bearing defect orders, gear mesh and sidebands, the envelope spectrum where early faults live. Kilobytes leave the sensor, not megabytes. Then the system learns. For eight to twelve weeks it records what normal looks like for every truck, at every location, in every regime. Nothing is judged yet. Your work orders and teardown reports go into the same timeline. Before it goes live, it runs in shadow: it logs what it would have said, and your planners check that against what actually happened. Watch levels are set with your team, on evidence. A false alarm costs more trust than a miss. Then it goes live. Every capture is scored against its own baseline. A defect order climbing out of it raises a watch on a named component, with the trend and a recommended action. In the cab and on the dashboard, every component carries a symbol and a word, never a colour alone. Most watches resolve at the next scheduled service. Critical is rare by design. A watch becomes a notification in your maintenance system. What the mechanic finds comes back and sharpens the model. Each inspection makes the next call better. RockLogik. Physics on the sensor, judgement in the cloud, and a breakdown that becomes a scheduled shop visit instead.
A signature that will not hold still
Every defect frequency in a gearbox or a bearing is a multiple of shaft speed. On a haul truck the shaft speed changes every few seconds.
Loaded uphill, empty downhill, gear changes, dumping, idling at the loader. A plain spectrum of haul-cycle data smears every defect frequency across the band and into the noise floor. This is why generic telematics show nothing about the drivetrain until it fails: the analysis assumes a steady machine, and a rock truck is not one.
Fixed-plant monitoring solved this decades ago with two ideas, and RockLogik brings both to the truck: only compare like with like, and measure in orders of shaft speed rather than in hertz.
What a rock truck does in ten minutes
- Loading idle, shock from each bucket, no drivetrain load
- Loaded haul low gear, full torque, grade, the regime that matters
- Dump hoist pressure, body up, frame flex
- Empty return high gear, light load, higher speed, braking
- Queue idle again
Five regimes, five different vibration signatures from the same healthy gearbox. A single averaged spectrum is meaningless.
How RockLogik makes the signature hold still
Four techniques, each standard on its own. Together they make a haul truck look like fixed plant to the analysis.
Regime segmentation from the truck's own bus
Firefly listens to the J1939 bus for gear, ground speed, engine load and grade, and defines a handful of operating windows: loaded on grade in third, say. Sensors capture only inside a window, so every measurement is comparable with the last one in the same window. It also means the sensors sleep most of the time, which is where the battery life comes from.
Order tracking with a magnetometer tachometer
A rotating steel shaft carries residual magnetisation, and the magnetometer in IVAN reads that flux leakage through the housing as a once-per-revolution signal. The vibration is resampled against shaft angle rather than time, so bearing and gear frequencies land at fixed orders whatever the speed. No encoder, no hole in the housing.
Envelope analysis for early faults
A defect impact does not show up at its own frequency first; it rings the structure at high frequency. Demodulating that ringing, the envelope spectrum, reveals the defect order while the fault is still small. This is why the accelerometer is wideband: a sensor that stops at 3 kHz sees late-stage faults only.
Physics-informed features, not a black box
For every location we compute the known diagnostic quantities: bearing outer-race, inner-race, roller and cage orders from the bearing geometry; gear-mesh order and its sidebands from the tooth counts; band RMS, kurtosis and crest factor; ISO 20816 severity. A rising trend at a specific defect order is interpretable, and a maintenance planner can act on it.
The sensor, in section
What each part is for, and why it sits where it does.
Accelerometer on the base, not the board
The vibration sensor is bonded to a rigid pedestal that is one piece with the base pad, so what it measures is the component, not the circuit board. The rest of the electronics float in potting compound above it.
Sleep, wake, capture, compute, send
The sensor spends most of its life asleep at microamps. On a regime trigger from the hub it records a few seconds at full rate, computes the feature set on the Cortex-M4F, and sends a few kilobytes over Bluetooth. Raw bursts go up on request for model improvement.
Built for the outside of a differential
Surface temperatures near a transmission run high, trucks are pressure-washed, and everything gets hit by rocks. So the housing is a single aluminium billet, the electronics are potted solid, the one joint carries an O-ring to IP67, there is no cable or connector to work loose, and the primary cells are rated to −40 °C with no charging circuit to fail in the cold. It is built to outlast the component it watches.
The diagnostic feature set
Every feature below is computed per truck, per sensor location and per operating regime, and trended against a baseline learned from that same combination.
Bearing defect orders
Outer race, inner race, roller and cage frequencies for each bearing, tracked as orders of shaft speed so they hold still while the truck does not.
Gear mesh and sidebands
Mesh order amplitude and the modulation sidebands that appear when a tooth is damaged or a shaft is bent.
Envelope spectrum
Demodulates the high-frequency ringing a defect impact excites, which is where early-stage faults live.
Band energy and shape
RMS, kurtosis and crest factor per band, plus ISO 20816 severity trending.
Temperature against load
Component temperature compared with the fleet at the same gear, speed and grade.
Bus diagnostics
Diagnostic trouble codes and pressure or temperature warnings from the J1939 bus, correlated with what the sensors see.
Normal is learned. Deviation is the signal.
Drivetrain failures are rare, which is good for you and hard for a classifier. So the system is built the other way round.
In the first weeks on a site the system learns what healthy looks like for each truck, at each location, in each regime. There is a great deal of that data. From then on, every capture is scored against its own baseline, and a defect order climbing out of it is what raises a watch.
Known failure signatures from the condition-monitoring literature are matched on top of that. And every work order, oil analysis and teardown report on the fleet is pulled into the same timeline, so each inspection your team performs makes the models sharper. That exchange is written into the programme from day one.
Watch
A trend has left its baseline on a named component. Inspect at the next scheduled service. Most items resolve here.
Advisory
Something worth knowing that is not yet a trend: a temperature above the fleet norm for the same load, a bus code correlated with a signature.
Critical
A signature consistent with imminent failure. Bring the truck in. Rare by design, and set with your team.
Thresholds are chosen with your maintenance team, deliberately. In a maintenance organisation a false alarm costs more trust than a miss, and an unnecessary teardown is expensive for the person who authorised it. The system runs in shadow mode first so the operating point is agreed on evidence.
Want the engineering detail?
The signal-processing design, the sensor qualification plan and the analysis pipeline are documented. Reliability engineers and vibration analysts are welcome to go through them with us.