Remaining Useful Life Prediction of Electromechanical Equipment based on Particle Filter and LSTM

The Long Short Term Memory (LSTM) neural network can reach high prediction accuracy when analyzing industrial equipment signals, so it is widely used in Remaining Useful Life (RUL) prediction of industrial equipment. However, there are still several challenges in training LSTM networks, such as prone converging to a local optimal solution, weak generalization ability, and inability to provide uncertainty of estimated RUL, which make it difficult to apply in practice. Aiming at the existing problems, this paper proposes an RUL prediction algorithm based on the model fusion of Particle Filter (PF) and LSTM. The re-sampling process of the PF is improved based on the weight division and the neighboring combination. An LSTM network is deployed as the state transition equation of the PF. The signal noise is extracted and reconstructed based on the wavelet transform to create the particle set. The improved PF algorithm is used to optimize the training of the LSTM to search the global optimal solution. The weight coefficients of the PF are used to generate the CI (CI) of the RUL. The experimental verification on NASA Electromechanical Actuators (EMAs) data set shows that the proposed fusion model reaches higher accuracy and reliability.