Dynamic failure analysis of process systems using neural networks

Abstract

Complex and non-linear relationships exist among process variables in a process operation. Owing to these complex and non-linear relationships potential accident modelling using an analytical technique is proving to be not very effective. The artificial neural network (ANN) is a powerful computational tool that assists in modelling complex and nonlinear relationships. This relationship has good potential to be generalized and used for subsequent failure analysis. This paper integrates ANNs with probabilistic analysis to model a process accident. A multi-layer perceptron (MLP) is used to define the relationship among process variables. The defined relationship is used to model a process accident considering logical and casual dependence of the variables. The predicted accident probability is subsequently used to estimate the likelihoods of failure to the process unit. A backward propagation technique is used to dynamically update the variable states and the failure probabilities accordingly. Integrating ANN with a probabilistic approach provides an efficient and effective way to estimate process accident probability as a function of time and thus the risk can be easily predicted upon quantifying the damage. The updating mechanism of the approach makes the model adaptive and captures evolving process conditions. The proposed integrated approach is applied to the Tennessee process system as a case study.