Machine Learning in the Whole Vehicle Test Community

In brief.

A modern test vehicle logs hundreds of signals across powertrain, chassis, driver inputs and environment, and most test programmes read only a fraction of them. The faults that matter most tend to show in how those systems behave together and drift over a programme, which diagnostics running inside a single controller cannot see.

The paper covers where machine learning earns its place in a test programme, from predictive maintenance and remaining useful life to ADAS feedback loops. Its recommendation for a prototype programme is to start with anomaly detection: it needs no labelled failures, and the faults it confirms become the labels that later predictive models are trained on.

How it runs on ZELLOR.

ZELLOR is developing a hydrogen fuel cell car designed to weigh under 1,000 kg, seat four and travel over 400 miles. Ahead of the demonstrator being built, we apply these techniques to data from the existing test fleet. The models draw on signals from the energy store, fuel cell, inverters, motors, driver inputs and vehicle motion; the anomaly models alone watch 82 of them.

• Isolation ForestTrip against fleet Scores each trip against the rest of the fleet from its summary statistics, then ranks the signals that deviate most.
• LSTM autoencoderWithin a trip · primary detector Reads every monitored signal together in short, overlapping windows, so an anomaly is located to within seconds, even when a window’s averages look normal and only the way the signals move together is wrong.
• Random Forest regressorsExpected behaviour Learn each subsystem’s efficiency map from the fleet’s own telemetry. The gap between measured and predicted current is what becomes a degradation signal as the fleet’s history builds.

What the paper covers.

• What whole vehicle data is
• The problem: rich in data, poor in insight
• Machine learning in vehicle engineering
• Predictive maintenance and remaining useful life estimation
• Anomaly detection, with a comparison of model types
• Autonomous and ADAS feedback loops
• Other applications: energy, warranty and recall, driver-risk scoring
• Case study: ZELLOR
• Conclusions and recommendations

Lessons so far.

• Data quality decides everything. More effort went into aligning, cleaning and validating signals than into the models themselves, and every model failure traced back to the data first.
• False positives are a risk. A detector that flags too much is ignored, so thresholds are tuned conservatively and every flag has to carry the signals behind it.
• Trust is built by communication. Engineers adopted the output because results were shared regularly and the reasoning behind each flag was transparent, not because the model was accurate in isolation.