Predictive maintenance is the most over-claimed capability in fleet software. The concept is sound, some implementations deliver real value, and a large proportion of what is sold under the label is a fault code alert with a dashboard.
Here is what is realistic.
The maintenance strategy ladder
| Strategy | Trigger | Typical cost position |
|---|---|---|
| Reactive | Failure | Highest total cost |
| Preventive | Fixed interval | Lower, some over-servicing |
| Condition-based | Measured condition threshold | Lower still, needs data |
| Predictive | Forecast of impending failure | Lowest in theory, hardest to achieve |
Most fleets should aim for solid preventive maintenance with condition-based triggers layered on. Genuine predictive analytics — forecasting a specific failure with useful lead time — is achievable for some components and speculative for others.
What is genuinely predictable today
Battery failure. Voltage behaviour during cranking and rest is a strong indicator, and battery failure is a common cause of no-starts. This is one of the clearest wins available.
Brake wear. Where pad wear sensors or brake application data exist, remaining life is forecastable with reasonable accuracy.
Tyre issues. Tyre pressure monitoring plus wear tracking predicts both failures and the optimal replacement point.
Fluid degradation. Oil analysis programmes have predicted engine problems reliably for decades — an unglamorous, proven technique that predates the current wave of analytics.
DPF and emissions system problems. Regeneration frequency and duration trends indicate developing issues well before a fault code appears.
Cooling system. Temperature trends under comparable load conditions reveal degradation before failure.
Refrigeration units. Run time, temperature pull-down performance and fuel consumption trends predict failures effectively.
What is not reliably predictable
- Random component failures with no degradation signature
- Damage-induced failures — impacts, kerbing, foreign objects
- Failures caused by poor previous repairs
- Anything on a vehicle where you have no data feed
- Rare failures where there is insufficient historical data to learn from
Vendors claiming to predict "any failure" are describing an ambition. Ask which specific failure modes their model addresses, what data it needs, and what lead time it provides.
Prerequisites
Before considering predictive maintenance:
- Reliable telematics data with the parameters the model needs, at adequate frequency.
- Complete work order history with cause codes — the labels the model learns from.
- Accurate odometer and engine hours.
- A functioning preventive programme. Predictive maintenance layered on a chaotic maintenance operation produces alerts nobody acts on.
- Workshop capacity to respond. A prediction with no capacity to act on it is just an early warning of an unavoidable failure.
- A defined response process per alert type.
Starting sensibly
Phase 1 — Fault code triage. Route diagnostic trouble codes into the maintenance system with a severity mapping and defined actions. Not predictive, but it captures much of the practical value at very low cost.
Phase 2 — Threshold-based condition monitoring. Battery voltage, coolant temperature, DPF regeneration frequency, tyre pressure. Simple rules, immediate benefit.
Phase 3 — Trend analysis. Track parameters over time per vehicle, compare against the vehicle's own baseline and its class peers, and flag deviations. This catches degradation that fixed thresholds miss.
Phase 4 — Statistical or machine-learned prediction, for specific components with sufficient failure history.
Most fleets capture the large majority of the available value in phases 1 and 2, at a small fraction of the cost of phase 4.
Evaluating a predictive maintenance claim
Ask:
- Which specific failure modes does it predict?
- What data does it require, and do our vehicles produce it?
- What lead time does it provide, and is that enough to act?
- What is the false positive rate, and what does a false positive cost us?
- Was the model trained on vehicles like ours, in a duty cycle like ours?
- How does it improve with our data over time?
- What happens when we change vehicle types?
False positive rate is the question that separates serious products from demonstrations. An alerting system that cries wolf is worse than none, because it consumes workshop capacity and trains people to ignore it.
Questions readers send us
Does predictive maintenance reduce costs? Where it works — batteries, brakes, tyres, fluids, refrigeration — it converts unplanned failures into planned work, which is meaningfully cheaper. Fleet-wide claims of dramatic savings should be treated sceptically and tested against your own baseline.
How much data do we need? Enough failure examples for the specific component being predicted, which for rare failures can mean years of fleet-wide history. This is why component-specific approaches with clear degradation signatures work sooner than general failure prediction.
Can we do this without buying a predictive maintenance product? A great deal, yes. Fault code triage, threshold alerts on telematics parameters and oil analysis are all available without a dedicated analytics platform, and they deliver most of the practical benefit.
What is the difference between condition-based and predictive maintenance? Condition-based responds to a measured present state crossing a threshold. Predictive forecasts a future failure from trends. Condition-based is easier, more reliable and sufficient for most fleets.
Should small fleets bother? Small fleets should focus on solid preventive maintenance, accurate mileage and fault code triage. Predictive analytics needs data volume that small fleets rarely have, and the return does not justify the complexity at that scale.
