Transferable Feature Pose Alignment Network for Fault Detection on Unseen Machines

To address the challenge of collecting fault data from high-end equipment for training fault diagnosis models, this paper proposes a transferable Feature Pose Alignment Network (FPAN), which enables the transfer of a trained fault diagnosis model to other machines using only healthy data. The network employs a self-supervised contrastive learning strategy to automatically partition the feature space and generate high-dimensional feature poses for each fault type. Fault types are then identified through a multi-layer attention mechanism and classifier. During feature pose generation and model transfer, a pose alignment strategy based on the Pearson correlation coefficient is established. By aligning the feature poses of new machines with those of old machines, diagnostic knowledge can be shared. Experimental results demonstrate that by aligning the feature poses of normal data from new machines with those from old machines, it is possible to use the fault diagnosis model of the old machine for fault detection in the new machine.