Neural Network AOI Inspection With Three-Way Operator Review
Solution Overview
Problem
Existing automated optical inspection (AOI) systems in manufacturing have high false positive and false negative rates, requiring significant manual intervention and labor to correct, especially during new product introductions, due to insufficiently tuned classification rules and the need for extensive human supervision.
Innovation Solution
A Vision Analytics and Validation (VAV) system utilizing a convolutional neural network (CNN) with a three-way classifier that classifies components as 'good', 'bad', or 'do not know', allowing for rapid retraining based on operator feedback to adapt to new conditions and reduce false positives and negatives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional AOI classification rules are used, then inspection coverage is achieved, but false positive and false negative rates remain high
Solution Approach 1:
The system implements a feedback mechanism where operator corrections of false positives and false negatives are fed back into the training dataset. The CNN model is periodically retrained with this corrected data, allowing it to learn from errors and continuously improve classification accuracy, thereby reducing both false positives and false negatives over time.
Solution Approach 2:
The system enables self-improvement by automatically retraining the CNN model using feedback from operator corrections. This self-service capability allows the inspection system to autonomously enhance its classification rules without requiring manual reprogramming, continuously adapting to reduce errors and minimize manual review requirements.
2Reliability
If extensive manual tuning of classification rules is performed, then false positive rate decreases, but labor cost and time increase
Solution Approach 1:
The system replaces manual mechanical tuning of classification rules with an automated machine learning approach. The CNN model automatically learns optimal classification parameters from training data, and the system automates the retraining process using feedback from operator corrections, eliminating the need for manual rule adjustment while maintaining low false positive rates.
Solution Approach 2:
The system dynamically changes classification parameters through automated retraining of the CNN model. Instead of manual parameter adjustment, the model's internal parameters (weights and biases) are automatically updated based on feedback from operator corrections, allowing the system to adapt classification thresholds and rules efficiently without human intervention.
3Productivity
If AOI classification rules are not fine-tuned, then production speed is maintained, but all boards fail inspection requiring manual review
Solution Approach 1:
The system performs preliminary training of the CNN model using a training dataset before production inspection begins. This preliminary action establishes baseline classification capabilities that can operate at high speed during production, reducing the need for manual review while maintaining throughput. The model is prepared in advance to handle production volume requirements.
Solution Approach 2:
The system maintains continuous operation during production inspection, with the CNN model processing boards at full production speed. The feedback collection and model retraining occur continuously in the background or during idle periods, ensuring that production throughput is not interrupted while still improving classification accuracy over time.
4Ease of manufacture
If traditional AOI systems are used during new product introduction, then initial inspection is possible, but 100% of boards fail due to insufficiently tuned rules
Solution Approach 1:
The system performs preliminary training of the CNN model using a training dataset specific to the new product before production begins. This preliminary action allows the model to learn product-specific characteristics and establish accurate classification rules in advance, enabling immediate accurate inspection without requiring extensive manual rule tuning during new product introduction.
Solution Approach 2:
During new product introduction, the system collects feedback from operator corrections and uses it to rapidly retrain the CNN model. This feedback mechanism allows the system to quickly adapt to the specific product variations and characteristics, improving inspection accuracy iteratively without delaying production or requiring extensive manual setup time.
Data Source
AI summary
A vision analytics and validation (VAV) system for providing an improved inspection of robotic assembly, the VAV system comprising a trained neural network three-way classifier, to classify each component as good, bad, or do not know, and an operator station configured to enable an operator to review an output of the trained neural network, and to determine whether a board including one or more “bad” or a “do not know” classified components passes review and is classified as good, or fails review and is classified as bad. In one embodiment, a retraining trigger to utilize the output of the operator station to train the trained neural network, based on the determination received from the operator station.