Machine Learning Defect Analysis System with Knowledge Base Segmentation
Solution Overview
Problem
Conventional defect inspection methods using machine learning face accuracy issues due to limited training data, especially when identifying defect causes in industrial products, leading to incorrect apparatus identification and decreased inspection accuracy as the type of inspection targets expands.
Innovation Solution
An analysis system and method that utilizes a machine learning model to identify recognition classes from input data, such as images, and analyzes correspondence relationships with a database to improve detection accuracy and versatility, while minimizing the amount of training data required. This includes using a neural network to convert input data into recognition classes and certainty factors, and weighting these factors based on importance, with database switching or limiting based on image content or context.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning is used to identify defect causes by expanding inspection target types, then versatility is improved, but inspection accuracy decreases due to lack of training data
Solution Approach 1:
The inspection system is segmented into two independent modules: a machine learning model for defect detection and a knowledge base for cause analysis. The ML model handles only defect detection with limited training data, while the knowledge base stores pre-acquired correspondence relationships between defect types and apparatus causes. This segmentation allows the system to maintain high accuracy in defect detection while achieving versatility in cause identification without requiring extensive training data for all inspection target types.
Solution Approach 2:
A knowledge base acts as an intermediary between the machine learning model and the final cause identification. The knowledge base stores pre-acquired correspondence relationships between defect recognition results and apparatus types or manufacturing processes. When the ML model identifies a defect, the system queries the knowledge base to determine the likely cause, eliminating the need for the ML model to be trained on all possible cause scenarios while maintaining accurate cause identification.
2Ease of manufacture
If conventional database methods are used for defect inspection, then implementation is simple, but wrong apparatus is identified due to erroneous identification
Solution Approach 1:
The system incorporates a certainty factor calculation mechanism that provides feedback on the reliability of defect recognition results. When the ML model detects a defect, it calculates a certainty factor based on the confidence of the recognition. The system then queries the knowledge base with this certainty factor and only outputs cause identification results when the certainty exceeds a predetermined threshold, preventing erroneous identification while maintaining simple implementation through rule-based threshold checking.
3Measurement precision
If extensive training data is collected to improve inspection accuracy across all product types, then detection accuracy improves, but the amount of training data required increases significantly
Solution Approach 1:
The system extracts and separates the cause analysis function from the defect detection function. The ML model is trained only on defect detection using limited training data, while the cause analysis knowledge base is populated separately through expert knowledge or historical data. This extraction allows the system to achieve high detection accuracy with minimal training data while maintaining versatile cause identification capabilities across different product types without requiring extensive training data for each.
Data Source
AI summary
Provided is an analysis apparatus that can improve detection accuracy and versatility while suppressing an amount of training data in, for example, identifying a cause of a defect in a product using machine learning. An analysis apparatus includes: an acquisition section that acquires input data; an identification section that identifies any one of predetermined recognition classes by inputting the input data to a machine learning model trained in advance with training data; a database that stores a correspondence relationship between the recognition class and predetermined information; an analysis section that analyzes the correspondence relationship between the recognition class and the predetermined information by using the database; and an output section that outputs an analysis result.