Defective wafer detection is essential to avoid loss of yield due to process abnormalities in semiconductor manufacturing. For most complex processes in semiconductor manufacturing, various sensors are installed on equipment to capture process information and equipment conditions, including pressure, gas flow, temperature, and power. Because defective wafers are rare in current practice, supervised learning methods usually perform poorly as there are not enough defective wafers for fault detection (FD). The existing methods of anomaly detection often rely on linear excursion detection, such as principal component analysis (PCA), k-nearest neighbor (kNN) classifier, or manual inspection of equipment sensor data. However, conventional methods of observing equipment sensor readings directly often cannot identify the critical features or statistics for detection of defective wafers. To bridge the gap between research-based knowledge and semiconductor practice, this paper proposes an anomaly detection method that uses a denoise autoencoder (DAE) to learn a main representation of normal wafers from equipment sensor readings and serve as the one-class classification model. Typically, the maximum reconstruction error (MaxRE) is used as a threshold to differentiate between normal and defective wafers. However, the threshold by MaxRE usually yields a high false positive rate of normal wafers due to the outliers in an imbalanced data set. To resolve this difficulty, the Hampel identifier, a robust method of outlier detection, is adopted to determine a new threshold for detecting defective wafers, called MaxRE without outlier (MaxREwoo). The proposed method is illustrated using an empirical study based on the real data of a wafer fabrication. Based on the experimental results, the proposed DAE shows great promise as a viable solution for on-line FD in semiconductor manufacturing.

晶圆检测不良对于避免由于半导体制造中的工艺异常而导致的成品率损失至关重要。对于半导体制造中最复杂的过程，设备上安装了各种传感器以捕获过程信息和设备状况，包括压力，气体流量，温度和功率。由于有缺陷的晶圆在当前实践中很少见，因此有监督的学习方法通​​常效果不佳，因为没有足够的有缺陷的晶圆用于故障检测（FD）。现有的异常检测方法通常依赖于线性偏移检测，例如主成分分析（PCA），k最近邻（kNN）分类器或手动检查设备传感器数据。然而，直接观察设备传感器读数的常规方法通常不能识别用于检测缺陷晶片的关键特征或统计数据。为了弥合基于研究的知识与半导体实践之间的差距，本文提出了一种异常检测方法，该方法使用降噪自动编码器（DAE）从设备传感器读数中学习正常晶片的主要表示形式，并用作一类分类模型。通常，最大重​​建误差（MaxRE）用作区分正常晶圆和有缺陷晶圆的阈值。但是，由于不平衡数据集中的异常值，MaxRE的阈值通常会产生正常晶圆的高假阳性率。为了解决这一难题，我们使用了Hampel标识符（一种可靠的异常值检测方法），采用来确定检测缺陷晶圆的新阈值，称为没有异常值的MaxRE（MaxREwoo）。使用基于晶圆制造实际数据的实证研究说明了所提出的方法。根据实验结果，提出的DAE有望作为半导体制造中在线FD的可行解决方案。