Defect Classification Using Confidence Filtering for Noisy Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The mobile display industry faces challenges in inspecting surface defects using traditional mechanisms, and machine learning models struggle with accuracy due to erroneous labeling and overfitting, especially with small datasets, leading to poor performance on new data.
Innovation Solution
A two-stage approach is employed to filter out unconfident data samples by training a first machine learning model with clean data and a second model as an outlier filter, using unsupervised or supervised learning to tune decision boundaries and reject noisy data, improving prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional inspection mechanisms are used for surface defects, then the inspection process is simple, but the detection accuracy is insufficient
Solution Approach 1:
The patent replaces traditional mechanical inspection mechanisms with machine learning models that use image data and sensor data to detect defects. The system uses trained ML models to analyze product images and automatically identify surface defects, substituting manual or mechanical inspection with intelligent algorithms that provide higher detection accuracy.
Solution Approach 2:
The patent introduces image processing algorithms and data preprocessing steps as intermediaries between the inspection system and the defect detection process. These intermediaries enhance the quality of input data for machine learning models, improving detection accuracy while maintaining manageable system complexity through modular architecture.
2Reliability
If machine learning models are trained with all available data, then the model covers more data distribution, but erroneous labels cause overfitting and reduce performance on new data
Solution Approach 1:
The patent applies preliminary data filtering and cleaning steps before training the machine learning model. Erroneous labels are identified and removed through validation processes and confidence thresholding during training. This preliminary action ensures that only high-quality data is used for training, preventing overfitting and improving model reliability on new data.
Solution Approach 2:
The patent dynamically adjusts training parameters such as learning rates, batch sizes, and confidence thresholds based on data quality assessment. By changing these parameters adaptively, the system optimizes training performance while filtering out noisy data, thereby improving model generalization without requiring overly complex training procedures.
3Measurement precision
If a single machine learning model is used for defect classification, then the system is simple, but it cannot effectively handle uncertain or noisy data samples
Solution Approach 1:
The patent segments the defect detection system into multiple specialized machine learning models, each trained to handle specific types of data or defect categories. This segmentation allows the system to process different data samples with appropriate models, improving classification accuracy for both clean and noisy data while maintaining manageable complexity through modular model design.
Solution Approach 2:
The patent introduces intermediary components such as data preprocessing modules, feature extraction layers, and post-processing validation steps between the input data and the classification models. These intermediaries enhance the robustness of the system by filtering and preparing data before classification, enabling accurate handling of noisy samples without requiring excessively complex model architectures.
Data Source
AI summary
Systems and method for classifying manufacturing defects are disclosed. In one embodiment, a first data sample satisfying a first criterion is identified from a training dataset, and the first data sample is removed from the training dataset. A filtered training dataset including a second data sample is output. A first machine learning model is trained with the filtered training dataset. A second machine learning model is trained based on at least one of the first data sample or the second data sample. Product data associated with a manufactured product is received, and the second machine learning model is invoked for predicting confidence of the product data. In response to predicting the confidence of the product data, the first machine learning model is invoked for generating a classification based the product data.