Data-Driven Fault Classification in Large-Scale Industrial Processes Using Reduced Number of Process Variables

In large-scale industrial processes, fault diagnosis is of paramount importance, as faults jeopardize the stability and performance of processes. However, effective fault diagnosis becomes a challenging task, due to having interconnectivity between processes as well as large number of process variables. To enhance fault diagnosis in large-scale industrial processes, while specifically focusing on fault classification, this paper proposes a data-driven method based on a three-stage computational procedure. The computational procedure uses collected data of process variables (i.e., measurements), and achieves the following: (i) time series data clustering using k-means clustering and soft dynamic time warping, (ii) variable selection using LASSO-based or Elastic Net-based regression methods, (iii) data-to-image conversion and fault classification using convolutional neural networks. In addition, a sensitivity metric is defined to select process variables that are deemed more helpful in fault classification, and thereby reducing the size of input data as well as the computational complexity required in fault diagnosis. To illustrate the effectiveness of the proposed computational procedure, two benchmark systems are used, for large-scale as well as small-scale industrial processes.Note to Practitioners-The motivation of this paper stems from the importance of fault diagnosis in industrial processes for practical applications, such as in chemical plants. Data-driven fault diagnosis is studied as it does not rely on exact process models, and therefore, offering practical advantages. However, in large-scale industrial processes, the existence of a large number of process variables as well as the possibility of sensor failure/miscalibration (i.e., leading to unavailable/unreliable measurements) poses limitations. Therefore, in this paper, a computational procedure is proposed to classify faults by only using a few selected process variables. It uses collected process measurement data in industrial control systems (e.g., in historians), and has three stages. The first stage studies the data, to discover and group process variables that have similar time-domain patterns. The second stage selects representative process variables from each group. The third stage uses the data of the representative process variables to classify faults. The effectiveness of the proposed computational procedure is verified, by considering different design approaches as well as operational scenarios, using two benchmark systems/processes.