针对电力设备运行和维护中所产生的大量碎片化、非系统性以及相关性不足设备缺陷文本,提出了一种电力设备缺陷文本识别模型。使用基于全词掩码的预训练模型(bidirectional encoder representation from transform-ers,BERT)替换基于随...针对电力设备运行和维护中所产生的大量碎片化、非系统性以及相关性不足设备缺陷文本,提出了一种电力设备缺陷文本识别模型。使用基于全词掩码的预训练模型(bidirectional encoder representation from transform-ers,BERT)替换基于随机掩码的BERT模型,提高了对电力词汇的理解力。使用双向长短期记忆网络(bidirection-al long short-term memory,BiLSTM)提高了模型捕获上下文信息的能力,并提高了模型的鲁棒性,引入注意力机制(attention)可以更好地捕获电力设备缺陷实体之间的复杂依赖关系,从而进一步提升模型的表现。实验结果显示,该模型准确率、召回率、F1值分别为96.26%、96.94%、96.60%,在地点、缺陷内容和设备三种实体上的F1值均优于其他模型。展开更多
For modern processes at deep sub-micron technology nodes, yield design, especially the design at the layout stage is an important way to deal with the problem of manufacturability and yield. In order to reduce the yie...For modern processes at deep sub-micron technology nodes, yield design, especially the design at the layout stage is an important way to deal with the problem of manufacturability and yield. In order to reduce the yield loss caused by redundancy material defects, the choice of nets to be optimized at first is an important step in the process of layout optimization. This paper provides a new sensitivity model for a short net, which is net-based and reflects the size of the critical area between a single net and the nets around it. Since this model is based on a single net and includes the information of the surrounding nets, the critical area between the single net and surrounding nets can be reduced at the same time. In this way, the efficiency of layout optimization becomes higher. According to experimental observations~ this sensitivity model can be used to choose the position for optimization. Compared with the chip-area-based and basic- layout-based sensitivity models, our sensitivity model not only has higher efficiency, but also confirms that choosing the net to be optimized at first improves the design.展开更多
This paper proposes an evolutionary optimized recurrent neural network for inspection of open/short defects on thin film transistor (TFT) lines of flat panel displays (FPD). The inspection is performed on digitized wa...This paper proposes an evolutionary optimized recurrent neural network for inspection of open/short defects on thin film transistor (TFT) lines of flat panel displays (FPD). The inspection is performed on digitized waveform data of voltage signals that are captured by a capacitor based non-contact sensor through scanning over TFT lines on the surface of mother glass of FPD. Irregular patterns on the waveform, sudden deep falls (open circuits) or sharp rises (short circuits), are classified and detected by employing the optimized recurrent neural network. The topology parameters of the recurrent neural network are optimized by a multiobjective evolutionary optimization process using a selected training data set. This method is an extension to our previous work, which utilized a feed-forward neural network, to address the drawbacks in it. Experimental results show that this method can detect defects on more realistic and noisy data than both of the previous method and the conventional threshold based method.展开更多
文摘针对电力设备运行和维护中所产生的大量碎片化、非系统性以及相关性不足设备缺陷文本,提出了一种电力设备缺陷文本识别模型。使用基于全词掩码的预训练模型(bidirectional encoder representation from transform-ers,BERT)替换基于随机掩码的BERT模型,提高了对电力词汇的理解力。使用双向长短期记忆网络(bidirection-al long short-term memory,BiLSTM)提高了模型捕获上下文信息的能力,并提高了模型的鲁棒性,引入注意力机制(attention)可以更好地捕获电力设备缺陷实体之间的复杂依赖关系,从而进一步提升模型的表现。实验结果显示,该模型准确率、召回率、F1值分别为96.26%、96.94%、96.60%,在地点、缺陷内容和设备三种实体上的F1值均优于其他模型。
基金Project supported by the National Natural Science Foundation of China (Grant Nos. 61173088 and 61070143)the Programme of Introducing Talents of Discipline to Universities (Grant No. B08038)
文摘For modern processes at deep sub-micron technology nodes, yield design, especially the design at the layout stage is an important way to deal with the problem of manufacturability and yield. In order to reduce the yield loss caused by redundancy material defects, the choice of nets to be optimized at first is an important step in the process of layout optimization. This paper provides a new sensitivity model for a short net, which is net-based and reflects the size of the critical area between a single net and the nets around it. Since this model is based on a single net and includes the information of the surrounding nets, the critical area between the single net and surrounding nets can be reduced at the same time. In this way, the efficiency of layout optimization becomes higher. According to experimental observations~ this sensitivity model can be used to choose the position for optimization. Compared with the chip-area-based and basic- layout-based sensitivity models, our sensitivity model not only has higher efficiency, but also confirms that choosing the net to be optimized at first improves the design.
文摘This paper proposes an evolutionary optimized recurrent neural network for inspection of open/short defects on thin film transistor (TFT) lines of flat panel displays (FPD). The inspection is performed on digitized waveform data of voltage signals that are captured by a capacitor based non-contact sensor through scanning over TFT lines on the surface of mother glass of FPD. Irregular patterns on the waveform, sudden deep falls (open circuits) or sharp rises (short circuits), are classified and detected by employing the optimized recurrent neural network. The topology parameters of the recurrent neural network are optimized by a multiobjective evolutionary optimization process using a selected training data set. This method is an extension to our previous work, which utilized a feed-forward neural network, to address the drawbacks in it. Experimental results show that this method can detect defects on more realistic and noisy data than both of the previous method and the conventional threshold based method.
文摘针对现有基于深度神经网络的代码缺陷检测方法无法分析缺陷特征并输出相关评审建议的问题,提出一种基于大感知域LSTM-Seq2Seq模型的代码缺陷检测方法。首先,使用长短期记忆网络(LSTM,long short-term memory)学习缺陷代码的编码特征,建立缺陷判别模型。其次,针对模型与数据集不匹配的问题,向序列到序列模型(Seq2Seq,sequence to sequence)引入代码段长度系数,提升模型对代码评审任务的适用度;通过建立代码缺陷特征与评审建议特征间的映射关系建立了代码分析模型,实现评审输出功能。最后,利用公开数据集SARD对该方法进行了验证,该方法在准确率、召回率、F1值方面的测试结果分别为92.50%、87.20%、87.60%,典型代码缺陷输出的评审文本与专家评审的文本相似度为85.99%,可有效减少评审过程对专家经验的依赖。
文摘针对基于显性知识的智能制造缺陷检测机制在工程实践中日益凸显的若干缺陷,提出了一种基于机器视觉和深度残差收缩网络(deep residual shrinkage networks,D-RSN)的智能制造缺陷检测方法,并进行了先验环境下的仿真验证。首先利用互补金属氧化物半导体(complementary metal oxide semiconductor,CMOS)相机集群搭建快速机器视觉图像获取装置,形成融合前置训练集和后置测试集的图像特征数据池;然后利用D-RSN对数据池前置训练集进行图像缺陷特征隐性知识学习辨识,构建时间正序下的图像缺陷特征全息感知机制;最后利用深度长短期记忆(deep long short-term memory,D-LSTM)神经网络对数据池后置测试集进行图像缺陷自主检测,借助图像缺陷定位及分类函数输出检测结果。选取某医用外科口罩智能制造生产线为工程实践验证载体,对模型进行了工程应用实践验证,结果表明:所提方法较好地改善了基于显性知识的智能制造缺陷检测机制在工程实践中日益凸显的若干缺陷,可以自主学习辨识图像缺陷特征隐性知识,大幅度提高了智能制造缺陷检测有效率,图像缺陷检测均值有效率达98.37%,符合医用外科口罩智能制造生产线国检要求。
