Unknown Defect Detection via Multi-Model Likelihood Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing information processing methods inaccurately classify product data as non-defective when it actually represents a new or unknown defect type, due to features being more similar to non-defective data than known defective data.
Innovation Solution
Generating non-defective and defective product learning models through machine learning using respective data types, and calculating likelihoods to determine if the target product data satisfies a predetermined requirement for being classified as an unknown defective product.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If determination boundaries are defined between non-defective product data and defective product data for each known defect type, then classification accuracy for known defects is improved, but unknown defect types are misclassified as non-defective products
Solution Approach 1:
The patent segments the classification task into three distinct models: a non-defective product learning model, and defective product learning models for each known defect type. This segmentation allows the system to evaluate target product data against multiple specialized models and determine unknown defects by checking if the data falls outside all known defect boundaries, thereby resolving the contradiction between accurate known defect classification and reliable unknown defect detection
Solution Approach 2:
The patent introduces an intermediary determination process that acts as a mediator between the multiple learning models. This determination unit evaluates the output likelihoods from all learning models and applies logical reasoning to classify target product data, serving as an intermediary that reconciles the conflicting classification results and enables reliable detection of both known and unknown defect types
2Adaptability or versatility
If machine learning is conducted using only known defective product data as teacher data, then the system can accurately identify known defect types, but it cannot detect new or unknown defect types
Solution Approach 1:
The patent performs preliminary action by generating separate learning models for each known defect type before actual classification. This preliminary preparation of multiple specialized models enables the system to later detect unknown defects by recognizing when target data does not fit any of the pre-established known defect patterns, thus achieving both adaptability to known defects and precision in detecting unknown defects
Solution Approach 2:
The patent adds another dimension to the classification space by introducing a fourth category (unknown defects) beyond the traditional binary or multi-class classification. This dimensional expansion allows the system to simultaneously maintain high precision for known defects while gaining adaptability to detect unknown defect types that fall outside established categories
Data Source
AI summary
A learning procedure involves generating a non-defective product learning model by conducting machine learning using non-defective product data as teacher data, and generating a defective product learning model for each defect type by conducting machine learning for each defect type using defective product data as teacher data. A calculating procedure involves calculating the likelihood of a non-defective product from output data calculated using the non-defective product learning model to which target product data is input, and calculating the likelihood of a defective product for each defect type from output data calculated using the defective product learning model to which the target product data is input. A determining procedure involves determining that the target product data is data on a defective product having an unknown defect when the likelihood of a non-defective product and the likelihood of a defective product for each defect type satisfy a predetermined requirement.