Abstract:

To improve the performance of process monitoring and fault diagnosis system in petrochemical enterprises and meet the requirements of real-time and timeliness of chemical process fault diagnosis, a fault diagnosis model based on mini-1D-CNN driven by process historical data was proposed. First, the data features of different fault types were learned and identified by one-dimensional convolutional kernel, and the dominant features were automatically extracted and classified second, the important feature parameters were selected by stepwise backward regression to optimize the model structure. In this paper, 31 process variables and operating parameters that can be obtained in real time were input into one-dimensional convolutional neural network (1D-CNN) to monitor and diagnose the main faults of TE process. The results show that compared with other fault diagnosis models, the mini-1D-CNN model has a higher fault diagnosis rate(FDR) on test set, which can reach 96.5; at the same time, the mini-1D-CNN model focuses on the important characteristic parameters of TE process fault diagnosis, and has significant advantages in reducing the number of parameters and reducing training and test time.

Key words: minimal-one-dimensional convolutional neural network (mini-1D-CNN ), Tennessee-Eastman(TE) process, fault diagnosis, process monitoring, contribution coefficient