Domain-Adaptive Residual Learning for Rubber Bearing Pressure Identification

Resonance peaks in the electromechanical admittance spectrum shift laterally and vertically when a plate rubber bearing undergoes axial compression and reflect changes in dynamic stiffness and damping that accompany load accumulation and material degradation. Those spectral displacements rarely uniform in practice; peak magnitudes, sub-harmonic features, and local irregularities evolve in ways that complicate straightforward interpretation. In isolation systems designed to moderate seismic forces, such ambiguity carries consequences. Rubber bearings must sustain compressive stresses within prescribed limits, however, construction tolerances, uneven settlement, and service loads introduce deviations that are difficult to quantify once the bearing is embedded within a structure. Base isolation relies on elastomeric components that combine vertical load capacity with horizontal flexibility. Steel-plate-reinforced rubber bearings achieve this through geometric confinement of nearly incompressible rubber layers. Under moderate compression, the constraint imposed by steel shims increases apparent vertical stiffness, while damping characteristics shift with microstructural strain redistribution. Once compressive stresses approach critical thresholds, interfacial damage at the rubber–steel boundary or internal tearing can initiate, altering stiffness in competing directions. Detecting such transitions without dismantling the bearing remains a persistent challenge. The electromechanical admittance method provides one avenue for in-situ interrogation. A surface-bonded PZT patch acts simultaneously as actuator and sensor, coupling electrical excitation with structural impedance. Because admittance depends on the mechanical impedance of the host medium, variations in stress state leave measurable fingerprints in the frequency response. Prior efforts have used peak tracking or statistical indices such as RMSD and MAPD to characterize load evolution but these strategies depend heavily on manual interpretation or fixed reference baselines. They also struggle when signals originate from different transducers whose bonding conditions, local geometry, or manufacturing variability introduce domain shifts in the data. Deep learning has entered this space as a way to bypass manual feature extraction. Convolutional architectures can classify stress states directly from raw spectra. The difficulty lies elsewhere: practical monitoring rarely yields thousands of labeled datasets for every sensor. Bearings in service may contain multiple PZT patches, and labeling each under controlled loading conditions becomes impractical. When a model trained on one transducer encounters signals from another, performance degrades because the statistical distribution of the input space has shifted.

The intellectual tension motivating this study emerges from that mismatch. If stress-induced spectral changes reflect physical mechanisms common to all transducers on the same bearing, then one should not need exhaustive labeled data from every patch. The question is whether a learning framework can extract domain-invariant representations from electromechanical admittance data while preserving sensitivity to pressure and damage states. Addressing that question requires confronting both signal scarcity and cross-domain variability within a unified modeling strategy.

A recent research paper published in Engineering Structures and led by Professor Demi Ai and Dr. Kejun Yang from the Huazhong University of Science and Technology, the researchers developed a residual block–domain adaptation neural network that transfers electromechanical admittance features from one PZT transducer to another for axial pressure and damage classification. They designed a data augmentation and preprocessing scheme that converts one-dimensional spectra into structured two-dimensional representations suitable for convolutional learning. They integrated adversarial domain classification with gradient reversal to enforce feature invariance across sensors.

Briefly, the research team instrumented two plate rubber bearings with pairs of surface-bonded PZT patches positioned diametrically opposite each other. They conducted controlled axial compression tests on one bearing and failure-oriented loading on the other, increasing pressure incrementally until crush failure occurred. During each loading stage, they recorded admittance spectra from 40 Hz to 500 kHz with 801 sampling points under a 1 V excitation. The investigators observed systematic rightward and upward shifts of dominant resonance peaks as compressive load increased within the elastic regime. They interpreted these RU shifts as manifestations of vertical stiffness growth combined with reduced damping, consistent with constrained lateral expansion of nearly incompressible rubber. When stress approached approximately 70 MPa, the researchers detected newly emerged sub-peaks and, in certain patches, abrupt leftward–upward spectral displacements. They attributed these LU shifts to stiffness reduction associated with incipient damage at the rubber–steel interface or internal layer tearing. That interpretation aligns with the mechanical expectation that damage competes with compression-induced stiffening, and the data exhibited precisely that competition in different frequency bands.

The authors computed RMSD and MAPD indices across loading stages and found approximately linear growth with pressure prior to damage onset. After threshold crossing, those metrics displayed inflection behavior or accelerated growth. While these indices captured stage transitions, they did not provide a direct mapping from signal to stress class, nor could they reconcile spectral variability between distinct PZT patches.

To overcome that limitation, the researchers constructed a residual block–domain adaptation neural network. They augmented limited source-domain datasets using a Gaussian-noise-scaled strategy tied to RMSD values, expanding each condition to fifty samples. They preprocessed the data by segmenting frequency bands, computing sub-band RMSD matrices, reshaping them into two-dimensional representations, and applying Z-score normalization and that transformation embedded frequency-domain characteristics into a spatially structured input suitable for convolutional learning. During training, the team fed labeled spectra from one transducer as the source domain and unlabeled spectra from another as the target domain and employed adversarial domain classification with a gradient reversal layer to encourage domain-invariant feature extraction. They found residual convolutional blocks allowed gradients to propagate through deeper layers without degradation, which proved necessary because domain alignment requires delicate feature transformations rather than superficial matching.

Across five independent runs, the researchers reported 100 percent training and testing accuracy for label prediction in both bearings. Domain classification accuracy exceeded 94 percent in cross-domain evaluation, demonstrating that the feature extractor successfully aligned source and target distributions while retaining discriminative structure. Confusion matrices showed correct identification of all stress and damage levels in the target transducers, even though those signals had not been explicitly labeled during training. When the team replaced the residual feature extractor with Transformer, fully connected, or multilayer perceptron variants, domain accuracy declined substantially, indicating that residual mapping facilitated stable adaptation under limited data conditions.  To summarize, Professor Demi Ai and Dr. Kejun Yang successfully built a new framework that enables stress and damage prediction using labeled data from a single transducer while generalizing to unlabeled signals from others. This study reframes pressure and damage monitoring in rubber bearings as a cross-domain learning problem rather than a sensor-specific classification task. That reframing carries practical weight. In isolation systems containing multiple bearings and distributed transducers, calibrating a deep model separately for each sensor is operationally unrealistic. The demonstrated transfer from one PZT patch to another implies that the learned representation captures stress-dependent mechanical behavior rather than patch-dependent idiosyncrasies.e We believe there many implications of the authors’ findings. For instance, electromechanical admittance methods have long promised high sensitivity, but their deployment in field environments has been limited by interpretive complexity and data requirements. Embedding domain adaptation within the modeling architecture, the authors managed to shift the burden from exhaustive labeling toward structural feature alignment. The residual blocks do more than improve gradient flow; they constrain the network to learn deviations from identity mappings, which reduces overfitting when sample sizes remain modest. That architectural choice influences scientific consequence directly: it favors generalizable stress features over sensor-specific artifacts.

The work also contributes to structural health monitoring by illustrating that load-induced spectral migration and damage-induced stiffness reduction share representational structure across sensors. If such invariance persists under environmental variability—temperature, aging, or seismic excitation—then transfer learning frameworks may reduce calibration costs substantially. At the same time, the experiments remain bounded. Only two bearings and two sensor pairs were evaluated. Broader validation across geometries, environmental histories, and multi-domain configurations will determine whether invariance holds at scale. There is also a methodological caution embedded in the findings. Perfect label accuracy within the experimental dataset does not eliminate uncertainty in field deployment. Adversarial alignment depends on sufficient overlap between source and target feature distributions. When operational conditions extend beyond the training manifold—such as long-term material aging or seismic cyclic loading—the learned embedding may require expansion. It is also worth to mention, that work of Professor Demi Ai and Dr. Kejun Yang establishes a pathway for integrating electromechanical impedance sensing with adaptive machine learning in base-isolated structures and shows that quantitative stress and damage classification can emerge from raw admittance data without one-to-one sensor retraining. That progression moves rubber bearing monitoring closer to automated, in-situ evaluation compatible with real infrastructure demands.