An enhanced manifold alignment method for fault diagnosis of liquid-cooled plate based on manifold distance features

The application of data-driven fault diagnosis in structural health monitoring (SHM) encounters several challenges. One major obstacle is obtaining a large quantity of accurate fault samples in certain scenarios. Additionally, the complexity of working environments exacerbates the difficulty of diagnostics due to significant distribution disparities among samples from different environments. This paper proposes a manifold alignment method based on manifold distance features (MDFMA), aiming to address the issue of insufficient fault samples during the fault diagnosis process. Initially, a simulated physical model of the monitoring object is established to generate a large number of simulated fault samples. These are supplemented with a small number of actual fault samples and then merged into a training dataset. Subsequently, the training samples undergo reconstruction employing the shortest manifold distance feature. Then the cosine similarity (COS) measuring method is applied to calculate the similarity between simulated and actual samples, thereby facilitating cross-domain feature alignment. Finally, all sample points are projected into the labeled space, and the manifold alignment (MA) method is employed to predict the state of the monitoring subject. The experiments conducted on the liquid-cooled plate confirmed the effectiveness and feasibility of the proposed method. The results demonstrate that this approach exhibits superior robustness and applicability in practical applications compared to other manifold alignment methods.