Fire source localization for long short-term memory networks based on Bayesian optimization

Civil aviation aircraft cargo compartment fires often occur in high-altitude, low-temperature and low-pressure environments, posing a significant threat to aircraft safety. A Bayesian optimization (BO) based long short-term memory (LSTM) neural network model (BO-LSTM) is proposed to quickly locate fire source and take specific means for fire suppression. In BO-LSTM model, time series data are fully analyzed to investigate the spatiotemporal correlation between fire characteristics (smoke, temperature, CO concentration) and fire source by using LSTM network. Meanwhile, the Bayesian algorithm is employed to search for the optimal LSTM network hyperparameter combination to improve model robustness and accuracy. In this paper, simulation study was used to validate the proposed BO-LSTM method: 8 widely used aircraft cargo compartment models were built by Pyrosim^® fire simulation software at a scale of 1∶1, and 10 fire source positions were randomly selected in each cargo compartment model to simulate fire data at low-temperature and low-pressure conditions. The experimental results demonstrate that the distance error between predicted fire location and true fire source was less than 0.1 meters, and the predicted two-dimensional fire source localized within the true fire source range. The performance of LSTM neural network has been improved dramatically by the Bayesian optimization method and make it a suitable tool for aircraft cargo compartment fire source localization at the environment of low-temperature and low-pressure.