New AI Framework Explains Root Causes of Automotive Software Failures in Hardware Testing

Automotive engineers currently spend considerable time manually sifting through massive recordings from hardware-in-the-loop (HiL) validation tests—sessions where actual car software runs against simulated hardware. Pinpointing why something failed, rather than just that it failed, remains a manual and labor-intensive task.

A new approach combines knowledge graphs with large language models to turn raw sensor data into actionable diagnostic insights. Rather than training a neural network to classify faults, the framework first converts time-series recordings into evidence snapshots, links that evidence to a map of how sensors and subsystems interact, and then retrieves similar past cases before asking an LLM to reason through the diagnosis.

How the System Works#

The method operates in stages. First, raw multivariate recordings are condensed into compact diagnostic summaries. These summaries are then enriched with sensor-to-location relationships and information about how faults propagate through the vehicle—knowledge typically held by engineers in design documents or past incident reports.

Next, the system retrieves historically similar cases from a database of prior faults. This retrieval step gives the language model concrete analogues to reason from. Finally, an LLM acts as a reasoning and explanation layer, generating both a ranked list of likely fault locations and a narrative explanation of the suspected cause—not merely a classification label.

Results and Real-World Testing#

The framework was evaluated on two automotive test scenarios: an advanced spark-ignition gasoline engine and an electric vehicle powertrain. On the gasoline engine, the best configuration achieved 90% accuracy in pinpointing the fault location on a per-recording basis. On the EV system, accuracy climbed to 94%. When aggregating results across multiple recordings in the same test file, the system achieved perfect localization on the tested subset.

Why This Matters#

Deep learning systems typically struggle with generalization: retraining is often needed when new faults, vehicle platforms, or test conditions arise. By anchoring the analysis in knowledge graphs and past cases rather than training end-to-end neural networks, this approach generalizes more readily to new scenarios. Equally important, engineers get interpretable explanations—not a black-box prediction—making the tool suitable for actual engineering diagnosis and decision-making.