Answers are grounded in this paper and cite their sources.
Graph convolutional networks can model relationships among samples in multisensor fault diagnosis. However, their performance depends strongly on graph construction, information propagation, and sensor fusion. Existing methods often rely on a single similarity criterion, suffer from over-smoothing as propagation deepens, and use fixed sensor-aggregation schemes. This study proposes a multisensor information-controlled graph convolutional network (MSICGCN) for supervised fault diagnosis. For each sensor, four distinct candidate graphs are constructed using Euclidean distance, cosine similarity, Mahalanobis distance, and a MST. The corresponding adjacency matrices are integrated through consistency-derived weighting and support-aware weighted sparsification, producing a sensor-specific topology. The resulting graphs are processed by damped graph convolutional layers with initial residual connections and orthogonal feature transformations, which regulate the contribution of propagated information. Sensor-specific representations are subsequently combined using learnable global sensor weights. Across three case studies, MSICGCN achieved classification accuracies of 98.92%, 99.24%, and 99.56%, corresponding to improvements of 1.01, 1.00, and 0.91 percentage points over the strongest comparator, respectively. These results support the combined use of complementary graph structures, controlled propagation, and global sensor weighting for supervised multisensor fault classification.
No takes yet. Share an insight, caveat, or question.
Cao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: