ML Defect Classification With Outlier Filtering for Display Modules
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for identifying manufacturing defects, particularly in display panel modules, are inefficient and prone to errors due to the use of traditional mechanisms and human inspectors, which are time-consuming and prone to labeling inaccuracies, making it difficult to scale and maintain production quality.
Innovation Solution
Employing machine learning models with an outlier filter to identify and remove noisy data during training, using a tuned decision boundary to enhance the accuracy of defect detection models by filtering out unconfident data samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional inspection mechanisms and human inspectors are used to identify manufacturing defects, then the process is simple to implement, but the accuracy and efficiency of defect detection deteriorate due to time-consuming operations and labeling inaccuracies
Solution Approach 1:
The patent replaces traditional mechanical inspection systems (human inspectors and manual examination processes) with an automated machine learning-based inspection system. The ML model automatically analyzes product images, identifies defects, and classifies them, eliminating the need for human inspectors and significantly reducing inspection time while improving accuracy and consistency.
Solution Approach 2:
The patent introduces an automated image analysis system as an intermediary between the manufactured product and the defect identification process. This intermediary system uses machine learning algorithms to process product images, extract defect features, and provide accurate defect classification, serving as a bridge that replaces direct human inspection while maintaining or improving detection quality.
2Productivity
If traditional inspection mechanisms are used, then the system complexity is low, but the scalability and consistency of defect detection deteriorate
Solution Approach 1:
The patent replaces manual inspection processes with an automated machine learning system that processes product images through trained models. This substitution enables high-speed, consistent defect detection across large production volumes, dramatically improving productivity while the system handles the complexity of defect classification internally through algorithmic processing rather than human cognitive effort.
3Measurement precision
If all training data is used to train defect detection models, then the training process is simple, but the model accuracy deteriorates due to noisy labels and unconfident data samples
Solution Approach 1:
The patent extracts and removes noisy, unconfident data samples from the training dataset before model training. By identifying and excluding data samples with unreliable labels or low confidence scores, the system trains the defect detection model only on high-quality, confident data, thereby improving model accuracy and generalization performance while reducing the negative impact of contaminated training data.
Solution Approach 2:
The patent performs preliminary data filtering and quality assessment before the actual model training process. By pre-processing the training data to identify and remove noisy samples, the system prepares a cleaner, higher-quality dataset for training, which leads to better model performance without requiring complex modifications to the training algorithm itself.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Systems and method for classifying manufacturing defects are disclosed. A first machine learning model is trained with a training dataset, and a data sample that satisfies a criterion is identified from the training dataset. A second machine learning model is trained to learn features of the data sample. When an input dataset that includes first and second product data is received, the second machine learning model is invoked for predicting confidence of the first and second product data based on the learned features of the data sample. In response to predicting the confidence of the first and second product data, the first product data is removed from the dataset, and the first machine learning model is invoked for generating a classification based the second product data.