This paper proposes a novel data preprocessing method that converts numeric data into representative graphs (polygons) expressing all of the relationships between data variables in a systematic way based on Hamiltonian cycles. The advantage of the proposed method is that it has an embedded feature extraction capability in which each generated polygon depicts a class-specific representation in the data, thereby supporting accurate “end-to-end learning” in industrial fault classification applications. Moreover, the generated polygons can play a significant role in the interpretation of trained deep learning fault classifiers. The performance of the proposed method was demonstrated using a benchmark dataset in the process industry. It was also tested successfully to classify challenging faults in major equipment in a thermomechanical pulp mill located in Canada. The results of the proposed method show better performance than other comparable fault classifiers.

本文提出了一种新颖的数据预处理方法，该方法将数字数据转换为代表性图表（多边形），该图表基于哈密顿循环以系统的方式表达数据变量之间的所有关系。所提出的方法的优势在于它具有嵌入式特征提取功能，其中每个生成的多边形在数据中都描述了特定于类别的表示，从而在工业故障分类应用中支持精确的“端到端学习”。此外，生成的多边形在训练有素的深度学习故障分类器的解释中可以发挥重要作用。在过程工业中使用基准数据集演示了该方法的性能。在加拿大的热机械纸浆厂中，它也成功进行了测试，以对主要设备中具有挑战性的故障进行分类。该方法的结果显示出比其他同类故障分类器更好的性能。