基于堆叠降噪自编码器的神经–符号模型及在晶圆表面缺陷识别

深度神经网络是具有复杂结构和多个非线性处理单元的模型, 通过模块化的方式分层从数据提取代表性特征, 已经在晶圆缺陷识别领域得到了较为广泛的应用. 但是, 深度神经网络在应用过程中本身存在“黑箱”和过度依赖数据的问题, 显著地影响深度神经网络在晶圆缺陷识别的工业可应用性. 提出一种基于堆叠降噪自编码器的神经–符号模型. 首先, 根据堆叠降噪自编码器的网络特点采用了一套符号规则系统, 规则形式和组成结构使其可与深度神经网络有效融合. 其次, 根据 网络和符号规则之间的关联性提出完整的知识抽取与插入算法, 实现了深度网络和规则之间的知识转换. 在实际工业晶圆表面图像数据集WM-811K上的试验结果表明, 基于堆叠降噪自编码器的神经–符号模型不仅取得了较好的缺陷探测与识别性能, 而且可有效提取规则并通过规则有效描述深度神经网络内部计算逻辑, 综合性能优于目前经典的深度神经网络.

Application of Neural-symbol Model Based on Stacked Denoising Auto-encoders in Wafer Map Defect Recognition

Deep neural network is a model with complex structure and multiple non-linear processing units. It has achieved great successes in wafer map pattern recognition through deep feature learning. In order to solve the problem of unexplained “black box” and excessive dependence on data in the applications of deep neural networks, this paper proposes a neural-symbol model based on a stacked denoising auto-encoders. Firstly, the symbolic rule system is designed according to the characteristics of stacked denoising auto-encoders. Secondly, according to the inner association between the network and the rules, a knowledge extraction and insertion algorithm is proposed to describe the deep network and improve the performance of the network. The experimental results on the industrial wafer map image set WM-811K show that the neural-symbol model based on stacked denoising auto-encoders not only achieves better defect pattern recognition performance, but also can effectively describe the internal logic of the neural network through rules, and its comprehensive performance is better than that of the current classical classification model.