SMT Re-Classification Engine for False Defect Reduction

A re-classification engine is introduced as an intermediary system between the optical inspection device and the review station. This engine receives inspection data, applies machine learning models to re-evaluate component positioning, and provides corrected classification results. The intermediary resolves the contradiction by filtering out false defects through intelligent algorithms while maintaining inspection coverage, thereby improving measurement precision and reducing false defect calls simultaneously.

The patent replaces traditional mechanical/optical inspection methods with machine learning-based classification systems. Instead of relying solely on optical measurement thresholds that prone to false calls, the system uses neural networks and deep learning models to analyze component positioning. This substitution of mechanical inspection with intelligent algorithms enables more precise measurement while reducing false defect calls, directly addressing the technical contradiction.

The re-classification engine implements a feedback mechanism that continuously learns from inspection results and refines its classification models. By analyzing confirmed defects and false calls, the system adjusts its decision boundaries and improves its accuracy over time. This feedback loop enables the system to maintain high reliability in defect detection while confidently reducing the number of false positives sent for manual review, thereby improving productivity.

The re-classification engine performs preliminary action by pre-screening and re-evaluating inspection results before they reach the review station. Using machine learning models, it preliminarily filters out false defects and only sends uncertain or confirmed cases for manual inspection. This preliminary processing reduces the workload at the review station, maintains high detection accuracy for real defects, and significantly improves overall processing throughput.

The patent applies segmentation by dividing the classification task into separate models for different component classes. Instead of using a single complex model to handle all classes, the system maintains dedicated models for each component type (e.g., resistors, capacitors, integrated circuits). This segmentation improves classification accuracy for each class while managing complexity through modular architecture, where each model can be independently trained and optimized for its specific component type.

The system implements dynamics by making model selection adaptive based on the specific class being inspected. The re-classification engine dynamically loads and switches between different class-specific models depending on the inspection data being processed. This dynamic model selection allows the system to maintain high accuracy for all classes while managing complexity through on-demand model loading and flexible architecture, rather than maintaining all models simultaneously in memory.