Causality-Augmented generalization network with cross-domain meta-learning for interlayer slipping recognition in viscoelastic sandwich structures

Abstract

Accurate interlayer slipping recognition in viscoelastic sandwich structures (VSSs) is critical for mechanical equipment’s safety and reliability. However, significant domain shifts exist in VSSs data under variable working conditions, and domain data under certain conditions cannot be directly accessed during training. This renders conventional domain adaptation methods ineffective. To address the problems, we proposed causality-augmented generalization network (CGN) without accessing target domains for VSSs’ slipping recognition. CGN comprises a swin-transformer feature extractor and a capsule network classifier with an FC decoder. The feature extractor aims to fully extract discriminative features of VSSs data and promote their domain invariance across multiple domains. Building on this foundation, the classifier further extracts the underlying causal features associated with the labels and performs slipping recognition, thereby enhancing the model’s generalization and stability across various domains. The decoder serves as a regularizer to assist in learning meaningful representations of input data. Moreover, cross-domain meta-learning strategy is incorporated into the generalized training process to further strengthen the model’s generalization ability. The experiments on VSSs’ cross-domain datasets illustrate that CGN can be trained on some domains and directly tested on multiple unknown domains with desirable results, showing its effective generalization and stability for slipping recognition.