Machine failure prediction technique using recurrent neural network long short-term memory-particle swarm optimization algorithm

This paper proposes a hybrid prediction technique based on Recurrent Neural Network Long-Short-Term Memory (RNN-LSTM) with the integration of Particle Swarm Optimization (PSO) algorithm to estimate the Remaining Useful Life (RUL) of machines. LSTM is an improvement of RNN as RNN faces issues with predicting long-term dependencies. Issues such as vanishing and exploding gradients are the results of backpropagating errors, taking place when the network is learning to store and relate information over extended time intervals. RNN-LSTM is a feasible technique for this research due to its effectiveness in resolving sequential long-term dependencies problems. However, the accuracy can still be enhanced to a satisfactory value considering the optimal network topology has not been discovered yet. Accuracy improvement can be achieved by resorting to hyperparameter tuning. A result of proof of concept validates that by increasing the number of epochs, the accuracy of prediction has improved but increases the execution time. To optimize between the accuracy and execution time, a population-inspired Particle Swarm Optimization (PSO) algorithm is employed. PSO will be utilized to select the optimal RNN-LSTM topology specifically the learning rate instead of using manual search. This optimized hybrid prediction technique is useful to be implemented in predictive maintenance to predict machine failure.