(339c) Chemical Process Fault Diagnosis Approach Based on Graph Convolutional Network

Modern chemical processes always form complex systems. Though the process engineering benefits from the development of automatic control technologies, there are always certain risks of process faults, which may cause severe chemical process accidents if process faults cannot be detected and diagnosed at early stages. In several decades, many approaches for fault detection and diagnosis had been proposed, and data-driven methods had drawn much attention recently. Different from model-based methods, data-driven methods have no need for a prior knowledge about the process but large amount of historical process data. Data-driven methods for fault detection and diagnosis can be divided into statistical methods and neural network methods. While statistical methods including PCA, PLS, ICA, FDA etc. have been widely studied, the application of neural networks has been growing rapidly recently with the development of deep learning technologies these years. The common neural network architectures applied in fault detection and diagnosis include auto encoder (AE), convolutional neural network (CNN) and recurrent neural network (RNN), and the stacked layers make the neural network deeper and gain much power for feature extraction, which is essential for fitting the latent relationships between different process variables without a prior knowledge. However, model-based methods can perform better than data-driven methods if detailed mathematical models are constructed based on first principles of the process, which can hardly be available for large-scale processes. So, it may help to improve the data-driven methods if we can find a way to apply the process knowledge to them readily regardless of the complexity of the process.

Motivated by the above, this work proposed a fault diagnosis method based on graph convolutional network, which inherits the benefits of neural network methods and utilizes a prior knowledge of the process. The chemical process is first modeled as a graph using the information about the unit operations and their connections. Then a graph convolutional network model is designed to deal with fault diagnosis end to end. The network is trained using historical process data, which is combined with the graph structure of the process. The proposed method is then tested on the benchmark Tennessee Eastman process and the effects are compared with other data-driven methods proposed for fault diagnosis.