Insight / Blog
Automotive NVH Diagnosis AI : Explainable and OOD-Aware Models
Automotive noise and vibration, together with CAN data, form a demanding signal-analysis domain. Powertrains, motors, and other drive components produce different signal patterns as their condition changes.
Those patterns also vary with driving conditions, measurement environments, and sensor placement. The same anomaly may appear differently from one test to another, making consistent interpretation difficult for researchers and engineers.
Background : Black-Box Predictions and Unseen-Signal Limitations
Company A, an automotive manufacturer, wanted to use its existing automated diagnostic solution more deeply in research workflows.
The existing AI models produced final predictions without showing which signal characteristics influenced the decision. This black-box behavior made it difficult to compare results with engineering knowledge.
The system also lacked a way to flag unfamiliar signals that fell outside the training distribution.
Company A wanted to turn diagnostic AI from a results-only tool into a research tool that could expose evidence and support model improvement.
Approach : Explainability, OOD Detection, and Retraining Comparison
SURROMIND added explainability and out-of-distribution detection to the existing signal-diagnosis workflow so researchers could validate evidence and improve models directly.
- Prediction-evidence visualization: showed which signal features influenced each AI decision.
- Automatic fault classification: classified anomaly types from noise and vibration signals.
- OOD detection: flagged unfamiliar signal patterns outside the training distribution to reduce misclassification risk.
- Data-distribution review: visualized whether data separated clearly by type.
- Retraining comparison: let researchers select important features and compare retrained-model performance with previous results.
Results : 96% Engine-Noise Accuracy and a 28-Point Improvement
| Metric | Previous limitation | Enhanced SURRO MCM workflow |
|---|---|---|
| Diagnostic accuracy | - | 96% engine-noise diagnosis accuracy, 28 percentage points higher than the previous result |
| Evidence review | No visibility into decision drivers | Visualized key frequency regions and signal features that influenced predictions |
| Unfamiliar data | No external-noise or OOD screening | Real-time classification of signal patterns outside the training distribution |
| Model improvement | One-time predictive model | Feature-based retraining and performance comparison using diagnostic history |
Operational Impact : Diagnostic AI as an Evidence-Based Research Tool
Automotive noise and vibration diagnosis must do more than separate normal from abnormal conditions. In a research environment, teams also need to understand why the AI reached its decision.
This project turned an automated diagnostic solution into a more trustworthy research tool. Researchers could review the signal features and frequency regions that influenced a prediction instead of receiving only a final result.
Automatic fault classification also supported more consistent decisions from measurement data, reducing reliance on individual experience alone.
When new data arrived, the system could identify patterns that differed from the training distribution and compare a retrained model with the previous model using selected signal features.
The resulting workflow achieved 96% engine-noise diagnosis accuracy, 28 percentage points above the previous result, while extending the system from one-time prediction to explainable, reviewable model improvement.
