Hybrid MB/ML PCB Defect Detection via Autoencoder Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated optical inspection systems for printed circuit boards (PCBs) are inefficient, detecting only about 10% of defects early enough for remediation and produce high false positive rates, making them labor-intensive and unreliable.
Innovation Solution
A hybrid model-based and machine learning-based PCB inspection system using an autoencoder that combines model data and training data to recognize features of interest, reducing the need for extensive training data and minimizing false positives by generating a reconstruction error value to detect defects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If automated optical inspection is used to detect PCB defects, then defect detection capability is improved, but false positive rate increases and labor-intensive manual inspection is required
Solution Approach 1:
The inspection system segments the defect detection process into two distinct modules: a model-based inspection module that uses PCB design data to detect structural defects, and a machine learning-based module that uses training images to detect appearance defects. This segmentation allows each module to specialize in specific defect types, improving overall detection accuracy while reducing false positives by avoiding the need for a single complex system to handle all defect types.
Solution Approach 2:
The patent introduces an image processing module as an intermediary between the imaging device and the defect detection modules. This intermediary processes the captured PCB images to enhance quality, normalize lighting variations, and prepare the images for both model-based and machine learning-based analysis. This intermediary layer improves measurement precision by ensuring that input images are optimized for defect detection while reducing false positives caused by imaging artifacts.
2Reliability
If automated optical inspection is used to detect PCB defects, then defect detection capability is improved, but labor-intensive manual inspection following detection is required
Solution Approach 1:
The system segments defect detection into model-based inspection for structural defects and machine learning-based inspection for appearance defects, allowing automated detection of both defect types without requiring manual verification. This dual-module approach enables the system to handle diverse defect types autonomously, reducing labor requirements while maintaining high reliability.
Solution Approach 2:
The patent employs different parameter sets and detection thresholds for the model-based and machine learning-based modules. The model-based module uses parameters derived from PCB design data, while the machine learning module uses parameters learned from training images. This parameter differentiation allows the system to optimize detection sensitivity for each defect type, improving reliability while maintaining automated operation with minimal manual intervention.
3Measurement precision
If model-based inspection alone is used, then structural defect detection is improved, but appearance defects and variations are not detected
Solution Approach 1:
The inspection system is segmented into two specialized modules: the model-based module focuses on structural defect detection using PCB design data, while the machine learning-based module focuses on appearance defect detection using training images. This segmentation allows each module to excel at its specific function, with the model-based module providing high precision for structural defects and the machine learning module providing versatility for appearance defects and variations.
Solution Approach 2:
The patent merges the outputs of the model-based inspection module and the machine learning-based inspection module to produce a comprehensive defect detection result. The model-based module detects structural defects with high precision, while the machine learning module detects appearance defects and variations, and their combined results provide both measurement precision and adaptability. This merging allows the system to leverage the strengths of both approaches without compromising either structural detection accuracy or appearance defect versatility.
4Adaptability or versatility
If machine learning-based inspection alone is used, then appearance defect detection is improved, but extensive training data is required
Solution Approach 1:
The system segments the inspection task so that the machine learning module handles only appearance defect detection, while the model-based module handles structural defect detection. This segmentation reduces the training data requirements for the machine learning module because it only needs to learn appearance characteristics rather than all defect types. The model-based module supplements the inspection by detecting structural defects using PCB design data, reducing the overall data requirements while maintaining versatility.
Solution Approach 2:
The image processing module serves as an intermediary that prepares images for both model-based and machine learning-based inspection. By optimizing image quality and normalizing variations before inspection, this intermediary reduces the amount of training data needed for the machine learning module to achieve effective appearance defect detection, while still maintaining adaptability to various defect types.
5Measurement precision
If extensive training data is used for machine learning, then detection accuracy is improved, but system complexity and training time increase
Solution Approach 1:
The inspection system is segmented into model-based and machine learning-based modules, with the machine learning module trained only on appearance defect data. This segmentation reduces the overall training data requirements and simplifies the machine learning model, thereby reducing system complexity and training time while maintaining detection accuracy for appearance defects. The model-based module handles structural defects without requiring training data, further reducing overall system complexity.
Solution Approach 2:
The patent applies partial action by using the machine learning module only for appearance defect detection rather than all defect types. This partial application of machine learning reduces the training data requirements and model complexity compared to a full machine learning approach, while still achieving high detection accuracy for appearance defects. The model-based module handles the remaining structural defect detection, allowing the system to achieve high overall accuracy with reduced complexity.
Data Source
AI summary
Hybrid MB/ML techniques for automated printed circuit board (PCB) defect detection. In one example, a PCB inspection system implements a hybrid solution using model based (MB) and machine learning (ML) technologies to detect possible defects in a PCB via an automated image capture device and processing methodology. The processing methodology fuses features from MB and ML at the latent representation. An autoencoder can be used to learn the fused data by training a neural network to produce a reconstructed image that can be compared to an original image to generate a reconstruction error value. The system produces an output indicating detection of a defect in one or more features of interest based on the reconstruction error value transgressing a threshold value.