An Efficient Scan Diagnosis for Intermittent Faults using CNN with Multi-Channel Data

Abstract

Scan chains are essential for enhancing the testability of semiconductor circuits. While scan chains enhance the testing capability, defects can also occur in scan chains due to the hardware overhead caused by scan chains. To prevent a decrease in yield due to such defects, scan chain diagnosis is widely used in semiconductor manufacturing as an important system. Particularly, with the increasing complexity of semiconductor circuits, there is a growing necessity for the diagnosis of intermittent faults. Since the distinct patterns are shown in test results for intermittent faults compared to permanent faults, a decrease in diagnostic accuracy is caused by the occurrence of intermittent faults. To address this problem, this paper introduces a new deep-learning-based scan-chain diagnosis method, designed to diagnose both intermittent and permanent faults. The proposed method introduces two new concepts for diagnosing not only permanent but also intermittent faults. One is a CNN optimized for scan chain diagnosis, and the other is newly optimized input data tailored for this CNN architecture. The proposed CNN architecture is composed of multiple layers centered around the modified inception module adapted for scan chain diagnosis, maintaining spatial and local information while enabling additional feature extraction. Furthermore, the new input data is multi-channel data composed of subset failure vectors (SFVs), integer failure vectors (IFVs), and fan-out vectors, allowing for the maximization of CNN characteristics. The experimental results demonstrate that diagnostic accuracy for intermittent faults is significantly improved by the proposed method compared to previous works.