Predicting Car-Engine Manufacturing Quality with Multi-Sensor Data of Manufacturing Assembly Process.

Area of Science:

• Manufacturing Engineering
• Artificial Intelligence
• Data Science

Background:

• Car engine quality control faces challenges with high-dimensional, noisy, and imbalanced multi-sensor data.
• Existing methods struggle to effectively process complex manufacturing data for diagnostics and prediction.

Purpose of the Study:

• To develop an edge-deployable framework for enhanced diagnostic and predictive capabilities in car engine quality control.
• To address data challenges including high dimensionality, noise, and class imbalance in manufacturing sensor data.

Main Methods:

• Utilized a Sparse Autoencoder (SAE) to reduce dimensionality and filter noise from over 12,000 manufacturing parameters.
• Employed a Class-Specific Weighted Ensemble (CSWE) to mitigate class imbalance in defect classification.
• Implemented an Adaptive Regime-Switching Regression (ARSR) for unsupervised transient performance tracking and prediction.

Main Results:

• The framework achieved an ultra-low inference latency of 80±3 ms.
• Defect interception recall improved by 7.72% due to effective handling of class imbalance.
• Relative prediction error was reduced by 12% through dynamic weighting of expert models.
• Practically reduced the engine rework rate by 7.2% on a physical H4 engine assembly line.

Conclusions:

• The proposed framework effectively addresses the challenges of multi-sensor data in car engine quality control.
• The edge-deployable solution offers significant improvements in diagnostic accuracy, predictive performance, and operational efficiency.
• Validated across diverse datasets and a physical assembly line, demonstrating practical applicability and substantial rework rate reduction.