针对传统卷积神经网络(Convolutional neural network,CNN)应用于焊缝缺陷识别时,池化模型适应差及特征选择能力低、以及易导致过拟合问题,提出了一种基于改进卷积神经网络(Improved pooling model and feature selection of CNN,IPFCNN...针对传统卷积神经网络(Convolutional neural network,CNN)应用于焊缝缺陷识别时,池化模型适应差及特征选择能力低、以及易导致过拟合问题,提出了一种基于改进卷积神经网络(Improved pooling model and feature selection of CNN,IPFCNN)的焊缝缺陷识别方法。结合焊缝缺陷图像本身的特点,对传统平均池化模型做出改进,提出一种综合考虑池化域与其所在区域特征图分布的池化模型;为增强模型特征选择能力,提出将随机森林与卷积神经网络相结合的强化特征选择方法。以某汽轮机制造过程中焊缝缺陷识别案例对所提方法进行了验证和说明,结果表明提出的池化模型在处理不同特征分布的池化域时具有动态自适应性,并通过提高特征选择能力,使得所提方法比传统CNN方法具有更高的缺陷识别率。展开更多
In order to recognize the different operating conditions of a distributed and complex electromechanical system in the process industry,this work proposed a novel method of condition recognition by combining complex ne...In order to recognize the different operating conditions of a distributed and complex electromechanical system in the process industry,this work proposed a novel method of condition recognition by combining complex network theory with phase space reconstruction.First,a condition-space with complete information was reconstructed based on phase space reconstruction,and each condition in the space was transformed into a node of a complex network.Second,the limited penetrable visibility graph method was applied to establish an undirected and un-weighted complex network for the reconstructed condition-space.Finally,the statistical properties of this network were calculated to recognize the different operating conditions.A case study of a real chemical plant was conducted to illustrate the analysis and application processes of the proposed method.The results showed that the method could effectively recognize the different conditions of electromechanical systems.A complex electromechanical system can be studied from the systematic and cyber perspectives,and the relationship between the network structure property and the system condition can also be analyzed by utilizing the proposed method.展开更多
Abnormal conditions are hazardous in complex process systems, and the aim of condition recognition is to detect abnormal conditions and thus avoid severe accidents. The relationship of linkage fluctuation between moni...Abnormal conditions are hazardous in complex process systems, and the aim of condition recognition is to detect abnormal conditions and thus avoid severe accidents. The relationship of linkage fluctuation between monitoring variables can characterize the operation state of the system. In this study,we present a straightforward and fast computational method, the multivariable linkage coarse graining(MLCG) algorithm, which converts the linkage fluctuation relationship of multivariate time series into a directed and weighted complex network. The directed and weighted complex network thus constructed inherits several properties of the series in its structure. Thereby, periodic series convert into regular networks, and random series convert into random networks. Moreover, chaotic time series convert into scale-free networks. It demonstrates that the MLCG algorithm permits us to distinguish, identify, and describe in detail various time series. Finally, we apply the MLCG algorithm to practical observations series, the monitoring time series from a compressor unit, and identify its dynamic characteristics. Empirical results demonstrate that the MLCG algorithm is suitable for analyzing the multivariable linkage fluctuation relationship in complex electromechanical system. This method can be used to detect specific or abnormal operation condition, which is relevant to condition identification and information quality control of complex electromechanical system in the process industry.展开更多
文摘针对传统卷积神经网络(Convolutional neural network,CNN)应用于焊缝缺陷识别时,池化模型适应差及特征选择能力低、以及易导致过拟合问题,提出了一种基于改进卷积神经网络(Improved pooling model and feature selection of CNN,IPFCNN)的焊缝缺陷识别方法。结合焊缝缺陷图像本身的特点,对传统平均池化模型做出改进,提出一种综合考虑池化域与其所在区域特征图分布的池化模型;为增强模型特征选择能力,提出将随机森林与卷积神经网络相结合的强化特征选择方法。以某汽轮机制造过程中焊缝缺陷识别案例对所提方法进行了验证和说明,结果表明提出的池化模型在处理不同特征分布的池化域时具有动态自适应性,并通过提高特征选择能力,使得所提方法比传统CNN方法具有更高的缺陷识别率。
基金supported by the National Natural Science Foundation of China (Grant by No. 51175402)
文摘In order to recognize the different operating conditions of a distributed and complex electromechanical system in the process industry,this work proposed a novel method of condition recognition by combining complex network theory with phase space reconstruction.First,a condition-space with complete information was reconstructed based on phase space reconstruction,and each condition in the space was transformed into a node of a complex network.Second,the limited penetrable visibility graph method was applied to establish an undirected and un-weighted complex network for the reconstructed condition-space.Finally,the statistical properties of this network were calculated to recognize the different operating conditions.A case study of a real chemical plant was conducted to illustrate the analysis and application processes of the proposed method.The results showed that the method could effectively recognize the different conditions of electromechanical systems.A complex electromechanical system can be studied from the systematic and cyber perspectives,and the relationship between the network structure property and the system condition can also be analyzed by utilizing the proposed method.
基金supported by the National Natural Science Foundation of China(Grant No.51375375)
文摘Abnormal conditions are hazardous in complex process systems, and the aim of condition recognition is to detect abnormal conditions and thus avoid severe accidents. The relationship of linkage fluctuation between monitoring variables can characterize the operation state of the system. In this study,we present a straightforward and fast computational method, the multivariable linkage coarse graining(MLCG) algorithm, which converts the linkage fluctuation relationship of multivariate time series into a directed and weighted complex network. The directed and weighted complex network thus constructed inherits several properties of the series in its structure. Thereby, periodic series convert into regular networks, and random series convert into random networks. Moreover, chaotic time series convert into scale-free networks. It demonstrates that the MLCG algorithm permits us to distinguish, identify, and describe in detail various time series. Finally, we apply the MLCG algorithm to practical observations series, the monitoring time series from a compressor unit, and identify its dynamic characteristics. Empirical results demonstrate that the MLCG algorithm is suitable for analyzing the multivariable linkage fluctuation relationship in complex electromechanical system. This method can be used to detect specific or abnormal operation condition, which is relevant to condition identification and information quality control of complex electromechanical system in the process industry.
