PMGN-LNCDE: a parallel multi-weight graph network with Linformer-optimized neural controlled differential equations for transformer fault prognosis

Abstract Early fault prognosis in power equipment is essential for ensuring grid stability and preventing cascading failures. Current spatio-temporal models, such as graph neural networks and recurrent neural networks, typically depend on static feature representations and fixed topological structures. These methods often employ high-dimensional parameterizations to capture complex system dynamics, resulting in significant computational overhead. Furthermore, they generally lack adaptive mechanisms to dynamically focus on evolving fault characteristics or sustain robust performance under missing-data conditions. To overcome these limitations, this paper introduces a novel framework integrating a parallel multi-weight graph network with Linformer-optimized neural controlled differential equations (PMGN-LNCDEs). The proposed model utilizes a PMGN to capture local geographical relationships and global semantic dependencies simultaneously. The PMGN is seamlessly embedded within a LNCDE structure, providing dynamic adaptability to complex operational environments in power transformers and effective handling of incomplete data. Experimental evaluations show that the proposed method achieves a sample-level mean squared error of 1.03 × 10 −5 , mean absolute error of 2.23 × 10 −3 , root relative squared error of 1.12%, and relative absolute error of 0.92%. These results demonstrate state-of-the-art prediction accuracy, validating the effectiveness of the proposed framework for vibration-based fault prognosis in power transformers.