Application of long short-term memory recurrent neural networks for localisation of leak source using 3D computational fluid dynamics

Gas leaks represent a major concern in industrial sites due to potential human and economical losses. Prompt identification of leak scenarios favours corrective maintenance avoiding the domino effect. In this paper, long short-term memory recurrent neural networks were trained and tested to CH4 leakage source in a chemical process module. We exploit the benefits of varying the temporal length of input variables, and the datasets were obtained employing 3D-CFD simulations. We consider four leak locations, four wind speeds, and eight wind directions, besides the non-leakage scenario for the same wind speeds and directions. The models were trained using different values of timesteps to evaluate the prediction accuracy for unseen data. Results showed progressive improvement of the performance of the models with greater values of timesteps, and good generalisation with test accuracy over 95.3%, indicating the ability of the model to correctly predict the leakage source using easily monitored variables.(C) 2022 Institution of Chemical Engineers. Published by Elsevier Ltd. All rights reserved.

Automated summary

This paper trained and tested long short-term memory recurrent neural networks on CH4 leakage source in a chemical process module, with models using different values of timesteps to predict the source of leakage. The datasets were obtained through 3D-CFD simulations, considering various leak locations, wind speeds, and wind directions. Results showed improved performance with greater values of timesteps, achieving test accuracy over 95.3% and demonstrating good generalization.