DETECTION OF OPERATIONAL FAILURES WITH ARTIFICIAL NEURAL NETWORKS: APPLICATION TO THE TENNESSEE EASTMAN PROCESS
Giovanni Morales Medina, Sebastian Reyes Angarita · Revista Ingenierías Universidad de Medellín · 2024
The purpose of this article is to compare results of fault detection for the Tennessee Eastman (TE) process with the application of artificial neural networks (ANN). The Neuralnet library of the open-source program R, as well as the Keras library of the open-source program Python were used for the training of ANN. The TE process simulation data were down loaded from Harvard University’s server, and subsequently analyzed, defining the trends in the operational variables during the appearance of failures. With the database, the training and validation of different ANN structures were developed, considering the parameters number of hidden neurons, activation function, and number of hidden layers. According to the results, the training and validation of the ANNs with the Neuralnet library yielded a lower performance in fault detection than that obtained with the Keras library. The ANN with the best performance in detecting failures in the TE process was obtained by the application of the Keras library. This ANN considered 52 input variables, 11 neurons in the hidden layer, and one neuron in the output layer, using a logistic function (ANN represented as 52:11:1 logistic) and reporting a prediction efficiency of 92% for the detection of faults with an external test set, which is convenient for future implementation in industrial processes.