Component Image Analysis Using CNN Ensembles for Defect Localization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-based feature recognition systems for components suffer from low reliability, leading to high scrap rates and material waste due to inconsistent detection results and sensitivity issues, especially with defects in components like those manufactured through die-casting.
Innovation Solution
A method utilizing an ensemble of convolutional neural networks (CNNs) with global pooling layers, trained on labeled and unlabeled training images, to enhance feature recognition and localization, including defects and intentional structures, by combining multiple neural networks to improve detection accuracy and visualization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If detection systems are set so sensitively that they fail to miss any defective parts, then detection reliability is improved, but the scrap rate increases due to false positives
Solution Approach 1:
The patent combines multiple neural networks (first neural network for sensitive detection, second neural network for specificity) into an ensemble system. The first neural network operates at high sensitivity to capture all potential defects, while the second neural network filters out false positives. This merging of multiple specialized networks resolves the contradiction by achieving both high detection reliability and low scrap rate through collaborative operation of the ensemble.
Solution Approach 2:
The detection system is segmented into two distinct neural networks with specialized functions: the first neural network is optimized for sensitive detection of all potential defects, while the second neural network is optimized for distinguishing true defects from false positives. This segmentation allows each network to specialize in specific aspects of detection, resolving the contradiction between sensitivity and false positive reduction.
2Productivity
If machine-based feature recognition is used, then productivity is improved, but reliability is insufficient compared to expert inspection
Solution Approach 1:
The patent merges multiple neural networks into an ensemble system that combines the speed of machine processing with the reliability of expert inspection. The first neural network provides rapid initial screening, while the second neural network provides expert-level verification, achieving both high productivity and high reliability through the combined operation of multiple AI models.
Solution Approach 2:
The system creates multiple copies of the inspection process through multiple neural networks, each representing different aspects of expert inspection. The first neural network copies the expert's initial detection capability, while the second neural network copies the expert's verification and discrimination capability. This copying of expert functions at machine speed resolves the contradiction between productivity and reliability.
3Device complexity
If a single neural network is used for feature recognition, then device complexity is reduced, but detection reliability is insufficient
Solution Approach 1:
The patent merges two specialized neural networks into an ensemble system that achieves higher reliability than a single network. The first neural network handles sensitive detection while the second handles specificity, and their combined output provides more reliable detection results than either network could achieve alone, justifying the increased complexity through improved performance.
Solution Approach 2:
Each neural network in the ensemble is optimized for specific local aspects of detection: the first neural network is optimized for sensitive detection of potential defects, while the second neural network is optimized for distinguishing true defects from false positives. This local quality specialization within the ensemble resolves the contradiction by having each component excel at its specific function.
Data Source
Figure 1~2
Figure 3
AI summary
The invention relates to a method (200) for analyzing a component, said method (200) comprising the following steps: - receiving an image of the component; - carrying out (203) a feature recognition on the received image by means of a plurality of neural networks, wherein at least one first neural network of the plurality of neural networks is trained based on a first set of training images, wherein at least one second neural network of the plurality of neural networks is trained based on at least a second set of training images, and wherein the at least one first neural network and the at least one second neural network each have a global pooling layer; and - displaying (205) a result of the feature recognition with respect to a representation of the component.