Remaining Useful Life Prediction of Milling Tool Based on Improved PSO-MultiAM-BiLSTM

Abstract

To improve the accuracy of remaining useful life (RUL) prediction for milling tools, this study proposes an enhanced PSO-MultiAM-BiLSTM model integrating particle swarm optimization (PSO), multi-head attention mechanism (MultiAM), and bidirectional long short-term memory (BiLSTM). The model captures key information in input sequences, alleviating early feature attenuation in BiLSTM from “chain propagation.” A logarithmic decreasing strategy adjusts PSO inertia weights, balancing global and local searches while optimizing BiLSTM parameters. Validated on the PHM2010 dataset, the model attains an average coefficient of determination of 0.97, with average root-mean-square error and mean absolute error of 0.062 and 0.045, improving prediction accuracy by 9.64% and 4.06% over MultiAM-BiLSTM and PSO-AM-BiLSTM, respectively. Such a result attests to the effective extraction of degradation features of tools and provides a valuable reference for predicting the RUL of milling tools.