TRODEON · Research

Research & Evidence

History-informed features improved repair-event ranking over our initial snapshot baseline in independent offline experiments. A separate calibration experiment improved probability scoring. These results describe research on a public historical maintenance dataset—not validation of the Trodeon application or a customer fleet.

Independent offline research performed in Arena. Reused-test development evidence—not fresh blind confirmation, named-fault diagnosis or operational readiness.

01 / Measurement

What we measured

Source: SCANIA Component X, Version 3. The research target was recorded repair/replacement within 48 anonymized dataset time units. Those units are not real days, and the recorded event is not confirmed physical root cause or proof of successful repair. This binary research target differs from the official five-class challenge task.

All three comparisons below use the same reused published test cohort: 5,045 vehicles and 142 labelled events. The ranking comparison selects 505 records in each method; it is not a recommended operating workload.

Offline / historical test-data example

This supplied Signal Room display uses historical U.S. published-source values. It is not live vehicle data and does not verify live Trodeon operation.

Signal Room display with a truck illustration and historical U.S. published data values.
The image is part of the offline research context, not a live vehicle feed or evidence of verified live operation.

02 / Same-cohort results

Results on the same reused test cohort

Results on the same reused test cohort
MethodAverage precisionROC AUCBrier scoreEvents among top 505
Initial snapshot baseline0.0560540.6861410.03381436
History-informed features0.0994970.7409990.02973449
History plus calibration0.0994970.7409990.02667149

Initial snapshot baseline

Average precision
0.056054
ROC AUC
0.686141
Brier score
0.033814
Events among top 505
36

History-informed features

Average precision
0.099497
ROC AUC
0.740999
Brier score
0.029734
Events among top 505
49

History plus calibration

Average precision
0.099497
ROC AUC
0.740999
Brier score
0.026671
Events among top 505
49

Higher average precision and ROC AUC indicate better ranking; lower Brier score indicates better probability scoring. AUC is not accuracy.

History features captured 13 additional labelled events within the same 505-record ranking budget compared with the snapshot baseline. That does not establish avoided breakdowns, correct repairs or customer savings.

The calibration map preserved ranking, so average precision, ROC AUC and the top-505 event count did not change. The training-prevalence constant had a test Brier score of 0.027355: uncalibrated history scores were worse than this constant, while calibrated history scores were better.

03 / Method

How the experiments were conducted

The offline research retained the publisher’s training, validation and test vehicle splits and checked them for overlap. Eleven publisher files were checked against published sizes and SHA-256 values. Model preprocessing was fitted on training data only.

The initial snapshot model used histogram gradient boosting. The history experiment retained the learner and weighting approach while adding 241 past-only history features. Future and same-time readouts were excluded from prior history. The initial baseline predictions and metrics were reproduced before the history comparison.

Calibration used a positive-slope sigmoid applied to fixed history-model scores. Of 5,046 development vehicles, 2,523 were assigned to calibration fitting and 2,523 to threshold selection using label-independent assignment. The calibration map was not fitted on test outcomes.

These are summaries of the Arena research methods, not claims that Polsia independently reproduced the experiments.

04 / Limitations

What remains uncertain

  • The test cohort was examined repeatedly across research stages. These results are reused-test benchmark iteration, not fresh blind or external validation. The initial benchmark also inspected two test-label rows during schema review before its original protocol.
  • Calibrated mean test probability was 1.93%, compared with observed event prevalence of 2.81%. Individual probabilities remain imperfect.
  • Maximum-F1 threshold behavior was unstable in later development-only checks. Those repeated checks reused the same development cohort; they were not independent new datasets.
  • Ranking budgets and decision thresholds are different. At its validation-selected maximum-F1 threshold, the history model recorded 42 true positives, 285 false positives, 100 missed events and 4,618 true negatives on this test cohort—not the top-505 ranking result.
  • Non-event labels do not establish vehicle health. No safe-to-operate, diagnostic, repair, dispatch or automatic work-order decision is supported.
  • U.S. Class 8 applicability, real-day warning time, named-fault diagnosis, successful repair outcomes, avoided breakdowns, customer economics and Trodeon application/fleet performance have not been established.

05 / Attribution

Source and attribution

Dataset: SCANIA Component X, Version 3.

DOI: https://doi.org/10.5878/bnh5-ka77

Official catalogue: https://researchdata.se/en/catalogue/dataset/2024-34/3

License: Creative Commons Attribution 4.0.

License link: https://creativecommons.org/licenses/by/4.0/

Dataset creators credited by the publisher: Tony Lindgren, Olof Steinert, Oskar Andersson Reyna, Zahra Kharazian and Sindri Magnússon.

Trodeon’s offline research transformed the source data into features and a binary repair-event target for the comparisons described here. These analyses are separate from the source publication. Source attribution does not imply a partnership, customer relationship or endorsement.

06 / Before additional signal work

Signal-source approval checklist

Review and document each item for the exact source and use being considered. No new source, account, fleet, vehicle, device, field set, time window, cadence, retention period, driver-linked scope, or additional use is identified or confirmed here; those details must remain unconfirmed until they are supplied and reviewed.

  1. Vehicle type and target population

    Identify the vehicle and engine type, and the specific target fleet or vehicle population. Record any limits on which vehicles or operating contexts are included.

  2. Source provenance and authority

    Identify the source or provider and product or API version. Review its authoritative documentation and schema, field meanings, units, and time semantics. Document where the data came from, the applicable rights, and who has authority to grant access.

  3. Exact data scope

    Specify the account, fleet, vehicles, and devices; each field; the historical window; collection cadence; retention period; and whether any driver-linked data is included. Leave every unspecified value unconfirmed.

  4. Permitted uses and explicit exclusions

    State each intended use and explicitly list excluded uses. Do not infer permission for a different purpose, secondary use, or operational action from the existence of a source or access to its data.

Owner approval is required before work proceeds

The owner must explicitly approve the identified source, defined data scope, and permitted use before any additional source, scope expansion, ingestion, or integration work proceeds. Reviewing or completing this checklist is not approval. Approval for one source, scope, or use does not cover a different source, scope, or use.

This public page is informational only: it has no approval control and does not record or imply owner approval. The offline dataset attribution above applies only to the research described on this page; it does not approve another source or use.

Publication scope

Aggregate results and methods summary only

This page presents aggregate results and a methods summary only. It contains no vehicle-level records, identifiers or predictions and is not a complete executable reproduction package.

Back to the TRODEON homepage