On-line Process Fault Diagnosis through Integration of Recurrence Plots and Convolutional Neural Network

For industrial manufacturing processes, the occurrence of a small fault may lead to strong consequences on process safety. Although many process monitoring methods have been proposed in recent decades, delays in fault detection, false alarms, and missed detections are still quite common, especially for complex industrial processes. Therefore, there is a need for more efficient and reliable process monitoring methods to be developed. This paper proposes a method for on-line process fault diagnosis through the integration of principal component analysis, recurrence plots, and convolutional neural networks. High dimensional on-line measured process variables are compressed into principal components which are then analyzed using recurrence plots. To carry out online fault diagnosis, sliding windows are used in generating recurrence plots in real time. A convolutional neural network, which is developed from historical process operation data, is then used to classify the resulting recurrence plots into normal and various fault categories. The method has been validated on a simulated continuous stirred tank reactor process and the achieved fault diagnosis accuracy on training data is 99.48% and that on the testing data is 97.73%.