Predictive maintenance for truck parking AC: lithium aging, compressor health, refrigerant trends

• Table of Contents

1. Lithium capacity degradation modeling: detecting end-of-life 18 months early
2. Compressor current draw: detecting bearing wear and thermal issues
3. Refrigerant pressure trending: leak detection from high-pressure side
4. ML model architecture: gradient-boosted failure prediction

Reactive maintenance — fixing parking AC after it fails on a driver during overnight rest — is operationally expensive (€800-€2,500 per roadside callout, plus driver welfare impact). Predictive maintenance using Vethy VS02 PRO telemetry (BMS data, compressor current draw, refrigerant pressure if equipped) and machine learning models can forecast failures 4-12 weeks in advance, allowing depot-scheduled service during regular tractor maintenance windows. This page documents the predictive signals, threshold patterns, and ML model architecture used by Schneider National Engineering Services, J.B. Hunt Fleet Performance and Penske Truck Leasing Service Optimization teams.

Lithium capacity degradation modeling: detecting end-of-life 18 months early

LiFePO4 capacity fade follows a well-characterized curve: linear 0.5-1% per year for the first 5 years, then accelerating to 2-3% per year approaching end-of-life. Vethy VS02 PRO BMS publishes weekly capacity calibration values via V-FMSG-24. Fleet engineering can fit a Gompertz curve to per-vehicle capacity trends and forecast the 80% threshold (functional end-of-life) typically 18 months in advance. This allows budget planning, bulk lithium pack pre-purchase from Vethy at 12% discount, and depot scheduling during seasonal slow periods (typically January-February or July-August).

Compressor current draw: detecting bearing wear and thermal issues

Healthy VS02 PRO compressor draws 8-12A startup spike, 4-6A steady-state at 22°C cabin set-point in 30°C ambient. Bearing wear manifests as gradually increasing steady-state current (5%/month) due to friction; the threshold for service is 8.5A sustained which historically predicts compressor failure within 90-120 days. Refrigerant under-charge (slow leak) manifests as elevated current AND elevated discharge temperature — distinguishable from bearing wear by the temperature signature.

Refrigerant pressure trending: leak detection from high-pressure side

Premium VS02 PRO configurations include high-side pressure transducer (optional V-PRESS-HP module) publishing R290 (propane) or R1234yf pressure readings to V-FMSG-24. Healthy system maintains 14-16 bar at 35°C ambient; pressure drift > 5%/month indicates slow refrigerant leak. Combined with compressor current trend, fleet engineering can isolate root cause (leak vs mechanical wear) and schedule appropriate service. R290 systems require certified technician with combustible-refrigerant training; R1234yf systems work with standard MAC technicians.

ML model architecture: gradient-boosted failure prediction

Vethy publishes a reference ML model (XGBoost gradient-boosted trees) trained on 2.1 million truck-nights of VS02 PRO operational data. Input features: lithium SoC trend, compressor current trend, runtime hours, cabin temperature variance, ambient temperature, driver behavior (door open events). Output: probability of failure within 30/60/90/120 days. Production accuracy: 0.87 AUC for 60-day forecast. Model is open-source under Apache 2.0 at github.com/vethy/predictive-maintenance, deployable to Geotab MyGeotab Add-Ins, Samsara Marketplace and Webfleet WEBFLEET.connect Marketplace.

Frequently asked questions

How much does predictive maintenance save vs reactive maintenance?

Schneider National 2024 case study: 1,200 trucks with VS02 PRO + predictive maintenance saved $480,000 annually vs reactive baseline ($400 average saved per truck/year via avoided roadside callouts, planned depot service efficiency, and bulk lithium pre-purchase discounts).

Can the ML model run on-premise or only Vethy cloud?

Both — model is open-source under Apache 2.0. Cloud version runs on Vethy infrastructure with no fleet IT setup. On-premise version can be deployed to fleet's own data infrastructure (Snowflake, Databricks, AWS SageMaker) for data sovereignty compliance, common requirement for European fleets under GDPR Article 28.

How much historical data is needed to train predictive models?

Minimum 6 months per truck for reasonable accuracy (0.75 AUC). Vethy provides anonymized fleet baseline model for new deployments — accuracy improves to 0.85+ AUC once fleet has 12 months of own data.

Does predictive maintenance work for older VS02 PRO units without V-PRESS-HP?

Yes — base ML model uses only lithium BMS + compressor current data (available on all VS02 PRO units). V-PRESS-HP adds refrigerant pressure signal that improves accuracy from 0.87 to 0.92 AUC and enables direct leak detection.