Deep Learning OOD Detection for Semiconductor Defect Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning performance in semiconductor device manufacturing is degraded by out-of-distribution (OOD) data, necessitating efficient detection and restoration of performance.
Innovation Solution
A deep learning device and system that includes processing circuitry to determine image abnormalities using an anomaly detection model, perform PCA on merged vectors, and utilize a probability approximation model to detect OOD data, storing results in a database and extracting target labeling when the OOD rate exceeds a threshold, thereby enhancing performance through active learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning is used for quality control in semiconductor manufacturing, then quality prediction and defect detection capability is improved, but performance degradation occurs due to out-of-distribution data
Solution Approach 1:
The system performs preliminary detection of OOD data before it significantly degrades model performance. By continuously monitoring input data distribution and comparing it with the training data distribution, the system identifies OOD samples early and prevents them from corrupting the model's decision boundaries, thus maintaining both detection accuracy and performance stability
Solution Approach 2:
The system implements a feedback mechanism where detection results are continuously fed back to adjust the monitoring thresholds and retraining schedules. When OOD data is detected, the system triggers retraining with corrected data, and the improved model performance feeds back into enhanced detection capability, creating a self-improving cycle that maintains reliability while preserving measurement precision
2Reliability
If OOD data detection and restoration mechanisms are added, then model performance stability is improved, but system complexity increases
Solution Approach 1:
The system employs multi-functional components that serve multiple purposes. The feature extractor used for normal defect detection is also utilized for OOD detection by comparing feature distributions. The same neural network architecture performs both classification and anomaly detection through different output heads, reducing the need for separate dedicated structures and thereby limiting complexity increase while improving reliability
Solution Approach 2:
The system introduces an intermediary feature space that mediates between raw input data and final classification decisions. By detecting OOD conditions in this intermediate feature representation rather than directly in the raw data or final outputs, the system adds a relatively simple monitoring layer that provides robust performance stability without significantly increasing overall system complexity
3Reliability
If continuous monitoring and retraining are performed, then deep learning performance is maintained, but computational time and processing speed increase
Solution Approach 1:
Instead of continuous retraining, the system employs periodic retraining triggered by OOD detection thresholds. Normal operation proceeds with fast inference, and retraining is performed only when OOD data exceeds predetermined thresholds, creating a periodic cycle of fast detection followed by slower retraining. This maintains performance reliability while minimizing time loss by avoiding unnecessary continuous retraining operations
Solution Approach 2:
The system applies different processing intensities to different data regions. For in-distribution data, fast standard inference is used. For suspected OOD data, more intensive feature analysis and comparison are performed locally. This localized quality adjustment maintains overall performance while reducing average processing time by applying heavy computation only where necessary
Data Source
AI summary
A deep learning device and system including the same is provided. The deep learning device comprising processing circuitry configured to determine whether a received image is abnormal using an anomaly detection model; merge at least some vectors extracted from the anomaly detection model; input, to a probability approximation model, principal components generated by a principal component analysis (PCA) to detect whether out of distribution (OOD) occurs in data of the received image; store a result of the determinations; and extract at least some the data in which the OOD occurs, as target labeling, using a target labeling extraction model when a rate of the data in which the OOD occurs is greater than or equal to a threshold value, wherein the anomaly detection model determines whether the received image is abnormal using the target labeling.