Compensation of accuracy by increased data “thickness” for high timeliness in fault diagnosis of chemical process

Abstract

BackgroundFault diagnosis rate (FDR) and fault diagnosis timeliness (FDT) are critical to a fault diagnosis model applied to industrial processes. Generally, high FDR requires process data of longer time, which means a low timeliness; while, the time-length reduction of data will result in a decreased FDR as learning features from less data is more difficult. MethodsTo address this problem, an idea is proposed that the time-length reduction of process data could be compensated by enhancing the data “thickness”, i.e., the shortage of feature information of data could be compensated by investigating more patterns from multiple different aspects of the data. Based on this, the multiple pattern representation-convolutional neural network (MPR-CNN) is proposed, in which multiple pattern representation algorithms are used first to pre-extract features from data. The resulting three-dimensional matrix contains richer features, which could be learned more efficiently by the CNN. Significant FindingsThe diagnosis results from a continuous and a batch chemical process show that MPR-CNN significantly improves the fault diagnosis performance, which achieved 96.1% of FDR in the Tennessee Eastman process and 98.1% of FDR in a semi-batch crystallization process, and outperforms other CNN structure-based models.