【摘要】 针对传统卷积网络提取数据结构特征与相互关系能力有限，并且图神经网络处理图数据时泛化性能不足的问题，本文提出一种结合shapelets和模式距离的图注意力网络。对原始数据进行节点划分，并对每个节点提取多个shapelets特征序列，突出局部特征，提升数据在低噪声情况下的特征提取能力，将每个节点所取得的shapelets特征序列通过K-最近邻（K-nearest neighbor,KNN）方法构建KNN图模型，进而捕捉节点之间的关系与结构信息，将构建好的图模型输入图注意力网络，对相邻节点自适应分配权重，提高了模型提取故障特征信息的能力，增强了故障分类的性能。为了验证所提方法的有效性，使用多个公共数据集进行了试验验证，准确率均大于99%，并在3种水泥生产设备数据上进行验证，平均准确率达99.5%，结果表明该方法对比现有的传统故障检测模型拥有更好的检测精度和泛化性能。更多还原

【Abstract】 Aiming at the problems of limited ability for traditional convolutional network to extract structural features and interrelationships of data,as well as insufficient generalisation performance of graph neural network when dealing with graph data,this paper proposes a graph attention network based on shapelets and pattern distance. The original data is divided into nodes and multiple shapelets feature sequences are extracted for each node to highlight local features and improve the feature extraction ability of the data under low noise,and the shapelets feature sequences obtained from each node are passed through the K-nearest neighbor method. The KNN graph model is constructed to capture therelationship and structural information between the nodes,and the constructed graph model is input into the graph attention network to adaptively assign weights to the neighbouring nodes,which improves the ability of the model to extract fault feature information and enhances the performance of fault classification. In order to verify the effectiveness of the proposed method,experimental validations are carried out using multiple public datasets, with accuracy greater than 99%. Additionally, validations are performed on data from three types of cement production equipment, achieving an average accuracy of 99.5%. The results show that the proposed method exhibits better detection accuracy and generalization performance compared to existing traditional fault detection models.更多还原