Fault Detection Using the k-Nearest Neighbor Rule for Semiconductor Manufacturing Processes

TLDR

Effective fault detection reduces scrap, boosts equipment uptime, and cuts test wafer usage, yet semiconductor processes’ nonlinearity, multimodal batch trajectories, and variable step durations challenge traditional univariate and PCA‑based statistical process control methods. The paper develops a fault detection method based on the k‑nearest neighbor rule to address these unique characteristics of semiconductor processes. The FD‑kNN method uses only normal operation data, applies a nonlinear k‑nearest neighbor classifier that handles nonlinearity, bases decisions on small local neighborhoods to accommodate multimodal batches, and performs automatic preprocessing for online detection. The method’s effectiveness is shown through simulated examples and an industrial case study.

Abstract

It has been recognized that effective fault detection techniques can help semiconductor manufacturers reduce scrap, increase equipment uptime, and reduce the usage of test wafers. Traditional univariate statistical process control charts have long been used for fault detection. Recently, multivariate statistical fault detection methods such as principal component analysis (PCA)-based methods have drawn increasing interest in the semiconductor manufacturing industry. However, the unique characteristics of the semiconductor processes, such as nonlinearity in most batch processes, multimodal batch trajectories due to product mix, and process steps with variable durations, have posed some difficulties to the PCA-based methods. To explicitly account for these unique characteristics, a fault detection method using the k-nearest neighbor rule (FD-kNN) is developed in this paper. Because in fault detection faults are usually not identified and characterized beforehand, in this paper the traditional kNN algorithm is adapted such that only normal operation data is needed. Because the developed method makes use of the kNN rule, which is a nonlinear classifier, it naturally handles possible nonlinearity in the data. Also, because the FD-kNN method makes decisions based on small local neighborhoods of similar batches, it is well suited for multimodal cases. Another feature of the proposed FD-kNN method, which is essential for online fault detection, is that the data preprocessing is performed automatically without human intervention. These capabilities of the developed FD-kNN method are demonstrated by simulated illustrative examples as well as an industrial example.