Detection of Rupture Damage Degree in Laminated Rubber Bearings Using a Piezoelectric-Based Active Sensing Method and Hybrid Machine Learning Algorithms

Abstract

Laminated rubber bearings may exhibit rupture damage due to factors such as temperature variations and seismic activity, which can reduce their isolation performance. Current detection methods, including human-vision inspection and computer-vision inspection, have certain limitations in accurately assessing the degree of rupture damage. This study attempts to combine the piezoelectric-based active sensing method with a machine learning algorithm to detect rupture damage in laminated rubber bearings. A series of laminated rubber bearings with varying degrees of rupture damage were fabricated, and 1440 sets of detection signals were obtained through experiments using the active sensing method. This study proposes a hybrid machine learning algorithm that integrates a one-dimensional convolutional neural network (1DCNN), long short–term memory (LSTM) network, Bayesian optimization (BO) algorithm, and extreme gradient boosting (XGB) algorithm. The algorithm involves using the 1DCNN and LSTM algorithms to extract the deep features from the wavelet packet energy spectra of the detection signals, and then employing the XGB algorithm optimized by the BO algorithm to construct the prediction model. The research results indicate that the proposed 1DCNN–LSTM–BO–XGB model achieved an accuracy value of 98.6% on the test set, outperforming the 1DCNN–LSTM (91.7%), 1DCNN (88.9%), LSTM (25.0%), XGB (90.3%), and SVM (66.7%) algorithms. Therefore, the combination of the active sensing method and machine learning algorithm shows promising application prospects in detecting the degree of rupture damage in laminated rubber bearings.