Enhanced reliability prediction of rolling bearings through a hybrid gamma process and BOA-TCN-BiLSTM-attention mechanism

To address the issue of the lack of expertise in data-driven reliability assessment and the difficulty in accurately characterizing the degradation state of bearings, this manuscript proposes a novel scientific machine learning method. Based on the hybrid Gamma process and BOA-TCN-BiLSTM-Attention mechanism, by combining data-driven and model-driven, enhances the ability to predict the reliability of rolling bearings. Firstly, the gradual reliability mathematical model of rolling bearing is established based on Gamma degradation model. Secondly, combined with the gradual reliability, the dimensionality of the high dimensional set consisting of the time domain and frequency domain features of the bearing vibration signal is reduced by using the Pearson correlation coefficient and the improved parrot optimizer random forest (IPO-RF). Thirdly, the composite mechanism combining Bayesian optimization algorithm, time convolutional network, bidirectional long short-term memory (BiLSTM) network and attention mechanism is developed to achieve accurate prediction of bearing reliability. Finally, a series of experiments on Xi’an Jiaotong University data set and PHM2012 data set verify the excellent performance of the scientific machine learning method.