Outlier IC Detection Using XGBoost and Mahalanobis Distance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional outlier integrated circuit (IC) identification methods fail to predict potential outlier ICs, leading to inefficiencies in testing and missed failures in post silicon test flows.
Innovation Solution
A method and system using machine learning frameworks, specifically Extreme Gradient Boosting (XGBoost), to generate a training model from first IC set data, predict data for a second IC set, and apply Mahalanobis distance analysis to identify outlier ICs by creating a bivariate dataset distribution, allowing for accurate prediction and identification of outliers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional outlier IC identification methods are used to analyze measured testing values, then some outlier ICs can be identified, but no prediction function is introduced leading to missed failed ICs and degradation of testing efficiency
Solution Approach 1:
The patent applies preliminary action by training a machine learning model in advance using first measured data from a first IC set before actual testing. The trained model generates predicted data for the second IC set, enabling proactive identification of potential outlier ICs before they are definitively classified, thus improving both accuracy and efficiency
Solution Approach 2:
The patent introduces an intermediary mechanism by using a machine learning prediction function as a mediator between raw measured data and final outlier identification. This prediction function processes measured data to generate predicted data, which then serves as the basis for calculating Mahalanobis distances and identifying outlier ICs, thereby enhancing both precision and productivity
2Productivity
If machine learning frameworks are introduced to predict potential outlier ICs, then testing efficiency is improved, but system complexity increases
Solution Approach 1:
The patent replaces traditional mechanical/conventional statistical outlier detection methods with a machine learning-based prediction system. By substituting conventional analysis with trained ML models that generate predicted data, the system achieves higher efficiency in identifying potential outliers, though this introduces computational complexity that is managed through the structured two-stage approach
Solution Approach 2:
The patent applies parameter changes by transforming the approach from direct outlier detection to a two-stage process: first generating predicted data through ML models, then calculating Mahalanobis distances based on these predictions. This parameter transformation enables more efficient outlier identification while managing system complexity through modular processing steps
Data Source
AI summary
An outlier IC detection method includes acquiring first measured data of a first IC set, training the first measured data for establishing a training model, acquiring second measured data of a second IC set, generating predicted data of the second IC set by using the training model according to the second measured data, generating a bivariate dataset distribution of the second IC set according to the predicted data and the second measured data, acquiring a predetermined Mahalanobis distance on the bivariate dataset distribution of the second IC set, and identifying at least one outlier IC from the second IC set when at least one position of the at least one outlier IC on the bivariate dataset distribution is outside a range of the predetermined Mahalanobis distance.