(584g) Deep Learning for Pyrolysis Reactor Monitoring: From Thermal Imaging Towards Smart Monitoring System

The pyrolysis reactor is a crucial component of chemical industry, which is used to crack heavier hydrocarbons to lower molecular weight hydrocarbons in a fired furnace. Condition monitoring of pyrolysis reactors is of paramount importance for manufacturing effectively and safely. While in practice, such monitoring is always challenging, particularly the operation condition inside the reactor, due to the high temperature (over 800℃) in the cracking regions where normal sensors can hardly be implemented. Infrared thermography is one of the practical solutions; it provides temperature information inside the fired reactor, where detailed operation conditions such as overheating, tube coking and tube deformation can be observed clearly. To take fully advantage of such information, an automatic monitoring system is built using deep learning techniques, which helps plant operators make correct decisions about the reactor operating.

Recent years, data-driven approaches using machine learning algorithms and statistics have been widely applied in feature learning. As the most advanced branch in machine learning, deep learning (DL) methods bring huge breakthroughs in many areas particularly for image analysis. In our work, based on the infrared thermography, we opted a deep convolutional neural network (DCNN) to build a monitoring system for the surface temperature and shape changing of pyrolysis tubes inside the fired furnace. To recognize the tube regions from the bulk background, a 50-layer ResNet[1] was used to segment tube areas from the raw images. In order to transfer the image classification model into a segmentation model, the strategy of fully convolutional network (FCN)[2] was applied in the ResNet model that tube areas are predicted by the upsampling of multiple-layer feature maps in the ResNet. After the segmentation of detailed tube regions, detailed surface temperature and area of the tube can be identified from the raw photos. Therefore, a monitoring system can be constructed where process operators can easily observe the overheated areas and shape changes through such system. Besides the monitoring efforts by operators, an automatic monitoring system can be further composed by the introduction of multivariate statistics. In this work, a PCA-based monitoring model[3] is introduced to detect faulty conditions from the information provided by six infrared cameras. The monitoring system successfully identified the overheated spots on the tube, as well as abnormal vibration of one infrared camera due to installation problem. This work provides contribution of combining state-of-the-art DL approaches into industrial monitoring systems.

[1] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778.

[2] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit., vol. 07-12-June-2015, pp. 3431–3440, 2015.

[3] L. H. Chiang, E. Russell, and R. D. Braatz, “Fault detection and diagnosis in industrial systems,” Time, vol. 12, no. 10–11, pp. 285–294, 2001.