Appearance Inspection Feedback Learning for Defect Reannotation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional appearance inspections in nuclear power plants face challenges due to poor visibility caused by radiation, leading to increased inspector burden and inefficiencies in defect detection using machine learning, with issues like long learning times and risk of over-learning.
Innovation Solution
An appearance inspection system that utilizes a trained model generated through progressive learning, where annotation information is associated with input images to improve the model, allowing for relearning when inference results differ from user determinations, thereby reducing inspector burden and enhancing defect detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by stationary object
If all obtained data is used as teacher data for progressive learning, then the learning process can be continuous, but the learning time increases and there is a risk of over-learning which may lead to a decrease in defect detection accuracy
Solution Approach 1:
The system implements feedback by comparing the trained model's inference results with user determination results. When discrepancies are found, the corresponding input images are selected as teacher data for relearning. This feedback mechanism ensures that only data that can improve the model is used for progressive learning, avoiding over-learning while maintaining continuous learning capability.
Solution Approach 2:
Instead of using all obtained data for progressive learning, the system selectively uses only the necessary portion of data - specifically, input images where the trained model's inference results differ from user determination results. This partial action approach reduces learning time and prevents over-learning while still achieving continuous improvement of the model.
2Ease of operation
If machine learning is used to detect and identify defects, then inspector burden is reduced, but long learning times are required to train the model and identify defect patterns
Solution Approach 1:
The system uses feedback from user determination results to guide the progressive learning process. By comparing model predictions with user judgments and selectively relearning from discrepant cases, the model rapidly adapts to specific defect patterns without requiring extensive initial training time, thus reducing both inspector burden and learning time.
Solution Approach 2:
The trained model performs self-improvement through progressive learning automatically. The system autonomously selects teacher data based on inference result discrepancies and performs relearning without requiring manual intervention for data selection or model retraining, significantly reducing inspector burden while maintaining efficient learning time.
3Reliability
If the trained model relearns from all input images, then the model may improve comprehensively, but the learning efficiency decreases and time is wasted on already well-performing cases
Solution Approach 1:
The system uses feedback from comparing inference results with user determination results to identify specific cases where the model needs improvement. By selecting only input images with discrepancies as teacher data, the system targets relearning efforts where they are most needed, maintaining high model accuracy while maximizing learning efficiency.
Solution Approach 2:
The progressive learning process applies local quality improvement by focusing relearning efforts on specific problematic cases rather than uniformly processing all data. The system identifies and addresses local weaknesses in the model's performance through targeted relearning from discrepant cases, improving overall reliability without sacrificing productivity.
Data Source
AI summary
An appearance inspection apparatus includes: an annotation unit that associates annotation information indicating whether a defect of a structure is normal with an input image to be used for an appearance inspection of the structure, the information being determined by a user based on the input image; a learning unit that generates a trained model by deep learning using the information and the input image associated with the information with respect to an initially-set trained model; and a defect inference unit that outputs a result obtained by inferring the defect based on the input image using the trained model generated by the learning unit, in which the annotation unit associates the information with the input image for which the result is different from a result of the determination made by the user for the input image to cause the learning unit to relearn the trained model.