Domain-adaptive shear deformation identification across laminated rubber bearings via transfer learning on electromechanical admittance signals from PZT transducers

In the monitoring of laminated rubber bearing system, the fusion of electromechanical admittance (EMA) data with transfer learning algorithms holds significant promise. However, existing EMA data-fusion approaches are confined to labelled dataset from a single transducer for single-bearing condition classification, which lack the adaptability to transfer classifiers trained on labelled data to other targets with unlabelled datasets. This limitation hinders the practical deployment of the bearing health monitoring. To this end, this article proposed a novel admittance enhanced domain adversarial neural network (AdmiEnDANN) to achieve adaptive deformation detection across different bearings. We introduced Weibull random noise to generate synthetic EMA data, addressing the challenge of limited measured data for model training. Proof-of-concept experiments were conducted on two different-sized rubber bearings under varied shear tests. The labelled EMA data source were from one bearing monitored by three surface-bonded piezoelectric ceramic (PZT) transducers and the unlabelled target data source from three corresponding transducers on another bearing. Experimental results demonstrated that the AdmiEnDANN attained 100.00% classification accuracy for both the task and post-adaptation phases across all three PZT transducer pairs, significantly outperforming traditional DANN models. Promising results provided a data-driven adaptation paradigm for assessing deformation across different rubber bearings with limited monitoring data.