Skip to content
allanninal.dev/build
Collaborate

Part 3 of 7 · Fuel log auditor series ~5 min read

Why fuel economy only means something against itself

Fuel economy is the obvious metric and it is a difficult one, because almost everything that moves it has nothing to do with the vehicle or the driver.

Key takeaways

  • Load, route, season, traffic and idling dominate. The vehicle is a smaller factor.
  • A fleet league table ranks routes, not vehicles or drivers.
  • Build a baseline per vehicle over at least ten fills, and note it is approximate.
  • Winter economy is worse for everybody; compare to the same season.
  • A trend within one vehicle is the signal worth acting on.

What actually moves it

The typical influence of five factors on vehicle fuel economyA bar chart with five bars showing typical swing in economy as a percentage. Load: eighteen per cent. Route type: twenty-two per cent. Season: twelve per cent. Driving style: ten per cent. Vehicle condition: eight per cent. A note says the thing you want to measure is the smallest bar, hence comparing a vehicle against itself.010203040~18Load~22Route type~12Season~10Driving style~8Vehicle conditionTypical swing in economy, %The thing you want to measure is the smallest bar. Hence: compare against itself.
Fig 1. The approximate size of each influence on fuel economy. Vehicle condition is what the system can usefully detect and it is the smallest effect, which is why the noise has to be removed by comparing like with like.

Route type is the biggest single factor in most small fleets and the one that varies most between vehicles. A van doing urban multi-drop and a van doing motorway runs will differ by twenty per cent or more with nothing wrong with either of them.

Which makes the league table that every fleet system produces — vehicles ranked by miles per gallon — a ranking of routes wearing the costume of a performance measure. It is worse than useless if anybody acts on it.

The per-vehicle baseline

How a per-vehicle fuel economy baseline is builtA vertical chain of five steps entered by a box labelled A vehicle with some history. Step one asks whether there are ten clean fills, matched and with good odometer readings; if not it exits to No baseline yet, reporting the figures without alerts. Step two computes the median economy rather than the mean. Step three computes its spread, asking how variable is normal. Step four adjusts by season if a year of data exists, drawing on a side box noting winter is worse for everybody. Step five produces a baseline with its sample size. A note says a vehicle with a wide normal spread needs a wider band before anything is unusual.AWS ACCOUNTA vehiclewith some historyTen clean fills?matched, odometer goodNo baseline yetreport the figures, no alertsnoMedian economynot meanAnd its spreadhow variable is normal?By seasonif a year existsWinterworse for everybodyA baselinewith its sample sizeA vehicle with a wide normal spread needs a wider band before anything is unusual.
Fig 2. How a vehicle’s own baseline is built. The spread in the third box is as important as the central figure, because a multi-drop van is naturally more variable than a motorway one.
  • Compute
  • App integration
  • Machine learning
  • Management
  • Analytics

Median, and per-fill is noisy

Individual fills are extremely noisy because tanks are never filled to the same level. A fill that stops at the first click and one that is topped to the neck differ by several litres, which on a sixty-litre tank moves the calculated economy substantially.

So the working unit is a rolling window of several fills rather than a single one. That removes most of the fill-level noise and it means a genuine change takes a few weeks to confirm, which is an acceptable trade for a signal that is about slow degradation.

Season

Winter economy is worse for every vehicle: cold engines, heaters, lights, denser air, wetter roads. A ten per cent drop in November is normal and a system that alerts on it will alert on the entire fleet at once, which is how people learn to ignore it.

With a year of data the comparison is against the same months last year. Without one, the comparison is against the rest of the fleet’s movement in the same period, which is the one legitimate use of a cross-vehicle comparison: not levels, but changes.

The one honest cross-vehicle comparison

How comparing changes rather than levels identifies one vehicle with a problemA horizontal row of five boxes. Fleet drops nine per cent in November. Van three drops nine per cent: normal. Van four drops nine per cent: normal. Van five drops twenty-one per cent: not the season. Look at van five, the only real signal. A note says comparing levels ranks routes, and comparing changes finds the vehicle with a problem.CHANGES, NOT LEVELSFleet drops 9%NovemberVan 3 drops 9%normalVan 4 drops 9%normalVan 5 drops 21%not the seasonLook at van 5the only real signalComparing levels ranks routes. Comparing changes finds the vehicle with a problem.
Fig 3. The legitimate use of the fleet as a comparison group. Nobody is compared on how efficient they are, only on how much they moved relative to everybody else.
  • Machine learning
  • Management
  • Analytics

This is the whole method in one picture. The fleet provides the control for whatever is happening in the world that month, each vehicle is measured against its own history, and what emerges is the vehicle that is behaving differently from itself for reasons the season does not explain.

Next: how that becomes a conversation.

All posts