An innovative method for axial pressure evaluation in smart rubber bearing based on bidirectional long-short term memory neural network

Abstract

Rubber bearings are a key component of base-isolated structures. Hence, it is important to effectively monitor the axial pressure in rubber bearings. A smart rubber bearing (SRB) and a wavelet-packet-based axial pressure index were introduced in the previous study. However, the wavelet-packet-based axial pressure index for evaluating the axial pressure in the SRB has notable shortcomings, because the system is susceptible to disturbances such as the change of sensitive subsets, generator-sensor couple and SRB specimen. This paper proposes the discrete wavelet packet transform enhanced bidirectional long-short term memory (DWPT-BiLSTM) method to evaluate the axial pressure state of the SRB. Three full-scale SRBs were tested, and two data augmentation methods were used to increase the dataset capacity and ensure the proper distribution of different axial pressure states in the dataset. By simultaneously considering the validation accuracy and training time, the optimal network structure is one hidden layer with 50 cells, and the optimal learning rate is 0.01. With these hyperparameters, the DWPT-BiLSTM method achieved an accuracy of 99.4% on the training data and an accuracy of 97.3% on the validation data in the evaluation of the axial pressure state of the SRB. The precision and recall of the axial pressure states were above 95% for both the training and validation set.