Defect Probability Estimation Using Neural Network Patch Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing manufacturing processes often generate faulty items, and defects are difficult to detect without prior knowledge of expected variations, necessitating a method for accurate defect detection.
Innovation Solution
A method involving the use of reference patch representations and similarity thresholds to classify manufactured items as good or faulty by comparing evaluated patch representations against reference patch representations, using neural network-generated features and similarity metrics to determine defect probability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional defect detection methods are used without prior knowledge of defects, then the manufacturing process can operate flexibly, but defect detection accuracy deteriorates
Solution Approach 1:
The system performs preliminary actions by collecting training data from the manufacturing process and using neural networks to learn normal variations before actual defect detection occurs. This preliminary learning phase enables the system to subsequently detect defects accurately without requiring prior knowledge of specific defect types.
Solution Approach 2:
The patent introduces an intermediary mechanism - a similarity metric and reference framework - that mediates between the manufactured item features and defect classification. This intermediary enables accurate defect detection by comparing item features against learned references without requiring direct knowledge of defect characteristics.
2Measurement precision
If strict similarity thresholds are used for defect classification, then defect detection accuracy improves, but the number of good items misclassified as faulty increases
Solution Approach 1:
The system dynamically adjusts the similarity threshold parameter based on the distribution of similarity scores in the training data. By changing this parameter adaptively rather than using a fixed threshold, the system achieves accurate defect classification while minimizing false positives of good items.
Solution Approach 2:
The patent applies a similarity threshold that is calibrated to capture the essential distinction between defective and non-defective items without being overly strict. This partial action approach ensures sufficient discrimination power while avoiding excessive rejection of good items.
3Measurement precision
If multiple similarity thresholds are implemented for defect detection, then detection accuracy improves, but system complexity increases
Solution Approach 1:
The patent segments the defect detection process into distinct phases: training phase where the system learns from data, and detection phase where the system applies learned knowledge. This segmentation allows complex multi-threshold logic to be implemented in a structured manner that manages system complexity while maintaining high detection accuracy.
Data Source
AI summary
A method for defect probability estimation based on relationships with concepts, the method may include (a) obtaining an evaluated patch representation, and (b) determining that the evaluates patch representation is not faulty when at least one of the following occurs: (a) for each RPR of a first number (N1) of RPRs, a similarity between the evaluated patch representation is not lower than the first RPR similarity threshold of the RPR; or (b) for each RPR of a second number (N2) of RPRs, a similarity between the evaluated patch representation is lower than the first RPR similarity threshold and not lower than the second RPR similarity threshold of the RPR.