Manufacturing · After-sales service
Which trucks get an air-system check, from data they already log
A workshop score built on Scania’s published data from 76,000 heavy trucks in everyday use, with the cut-off set from what a check and a breakdown cost.
Executive summary
A score from readings trucks already log sends 4.8% of trucks for an air-system check and catches 365 of 375 failures, at under half the cost of the best single-sensor rule.
Situation
Checking every truck’s air system wastes technician hours; checking none leaves brake and gear-change failures on the road. The workshop needs to know which trucks to inspect.
Approach
LightGBM on 170 on-board readings from 60,000 Scania trucks, with the cut-off set from what a check and a missed failure cost, chosen on training data and applied once to 16,000 test trucks.
Result
Cost 8,960 (95% CI 6,170 to 12,201), matching the best published result on this data (9,920). The score stays the cheapest option from 10:1 to 200:1 cost ratios.
The business problem
A truck dealer’s service network or a large fleet workshop sees trucks come in for service with faults that are not yet obvious. The air pressure system supplies the brakes and gear changes, and a failing component that leaves the workshop unchecked can mean a roadside breakdown, a tow and a missed delivery. Manufacturers already use connected-truck data to tailor service plans, and the question for the workshop is which trucks to inspect, from counters the truck already logs.
I set the cut-off from what an inspection costs against a breakdown and measured how many failures a score catches for the checks a workshop can afford.
- Team
- A service manager, 4 to 15 heavy-vehicle technicians, a parts desk
- Fleet
- 300 to 20,000 trucks reporting to the service network
- Systems
- Telematics or on-board counters read at service, a dealer management system, workshop job cards
- Constraint
- Heavy Vehicle National Law brake standards; a technician makes the call, the score sets the order
What was built
A score for each truck from its on-board counters and histograms, read when the truck is connected at service or through telematics. Trucks above a cut-off get an air-system inspection added to the job card. The cut-off is set from what a check and a missed failure cost, not from accuracy.
Results
Scania published this data with a cost: 10 for an unnecessary check by a mechanic, 500 for a missed failing truck. The test set holds 16,000 trucks, 375 of them with the failing component.
| Approach | Checks | Failures caught | Missed | Unneeded checks | Total cost (95% CI) |
|---|---|---|---|---|---|
| Check every truck | 16,000 | 375 of 375 | 0 | 15,625 | 156,250 (155,900–156,620) |
| Check no trucks | 0 | 0 of 375 | 375 | 0 | 187,500 (168,988–205,000) |
| Single-sensor rule | 1,421 | 357 of 375 | 18 | 1,064 | 19,640 (15,950–24,233) |
| Logistic regression | 1,017 | 362 of 375 | 13 | 655 | 13,050 (9,730–17,040) |
| Gradient boosting | 761 | 365 of 375 | 10 | 396 | 8,960 (6,170–12,201) |
The cut-off was chosen on training data only and applied to the test set once. The best published entry in the 2016 industrial challenge on this data scored 9,920 (542 unneeded checks, 9 missed). This pipeline scores 8,960 (396 unneeded, 10 missed); the 95% interval (6,170 to 12,201) overlaps it, so the fair reading is that it matches the best known result. Against logistic regression the saving is 1,319 to 7,050 (95% CI).
Where the failures sit
What the score reads
Sensor names are anonymised by Scania, so the score cannot say which physical reading matters. One counter, ck_000, carries a third of the model’s gain and is also the best single-sensor rule. In a live service network the names are known, and the reasons can go on the job card.
Data and method
- Source
- APS Failure at Scania Trucks, Scania CV AB (2016), UCI ML Repository 421, GPL-3
- Size
- 76,000 trucks in everyday use; 170 anonymised readings, 7 of them histograms with 10 bins
- Split
- Published split: 60,000 train (1,000 failures), 16,000 test (375 failures)
- Audit
- 8.3% of readings missing; 99% of trucks have at least one gap; 8 readings over half missing; 1 constant reading
Four approaches beyond check-all and check-none: the best single reading and threshold found on training data, logistic regression on log-scaled readings with a missing count, and LightGBM (385 rounds). Cut-offs come from five-fold out-of-fold predictions on the training set. Cost intervals from 1,000 bootstrap redraws of the test trucks.
Every truck in the data had a fault of some kind. The negative class is trucks whose fault was not in the air-system component. The score separates the two; it was not tested on healthy trucks.
Assumptions and limitations
- Costs of 10 and 500 are Scania’s published figures; an Australian workshop should use its own labour rate and breakdown cost. The table below shows the score stays the cheapest option from 10:1 to 200:1
- One snapshot per truck, no dates; a live version needs readings over time and checks that it holds as trucks age
- Readings are anonymised, so the reasons cannot be written in workshop language from this data
- The score does not replace brake inspections required under heavy-vehicle law
If the costs are different
| Cost of a miss against a check | Model | Check every truck | Check none | Checks | Failures caught |
|---|---|---|---|---|---|
| 10:1 | 5,100 | 156,250 | 37,500 | 500 | 340 of 375 |
| 20:1 | 6,180 | 156,250 | 75,000 | 636 | 358 of 375 |
| 50:1 | 8,960 | 156,250 | 187,500 | 761 | 365 of 375 |
| 100:1 | 10,190 | 156,250 | 375,000 | 1,192 | 373 of 375 |
| 200:1 | 12,190 | 156,250 | 750,000 | 1,192 | 373 of 375 |
Cut-off re-chosen on training data at each ratio, then applied to the test set.
What a service network would do next
- Weeks 1 to 3: export 18 months of on-board readings and air-system job outcomes from the dealer management system
- Weeks 4 to 5: refit with named readings, set the cut-off from the workshop’s own costs, agree the job-card wording
- Next three months: add flagged inspections at half the branches and compare roadside breakdowns
- Then: extend to other systems with the same counters
Technical depth
Missing values are left as missing for gradient boosting, which learns its own split direction for them; logistic regression gets median fill, log scaling and a missing count. The positive class is weighted five times in boosting, then the cut-off is set on expected cost, which matters more than the weight. The single-sensor rule is searched over every reading and 100 thresholds per reading.