A Gramian Angular Field for Constructing Graph-Based GNNs and Its Applications in Rolling Bearing Defect Detection

Graph neural networks (GNNs) have unique advantages in terms of processing non-Euclidean spatial data with significant differences and effectively diagnosing faults. However, GNNs typically rely on high-quality graph structures. The current graph construction methods often fail to capture the complexity and global features of vibration signals. To address this shortcoming, a graph construction method called GAFGraph based on a Gramian angular field (GAF) is introduced. GAFGraph offers a more nuanced representation strategy by transforming time series data into a Gramian matrix and then extracting more comprehensive fault node similarity features. It effectively captures the deep internal correlations and long-term dependencies within the input data, enhancing the global nature of the associated graph. Moreover, a feasible graph connectivity evaluation metric is proposed to calculate the relationships among various adjacency parameters, thereby effectively assessing the quality of the constructed graph. In addition, the featurewise linear modulation (FiLM) technique is integrated into a GNN-based fault diagnosis model. This technique dynamically adjusts the feature representations of the target nodes and edges by computing transformations for all incoming information based on the representations of the target nodes, modulating the feature propagation process and enhancing the adaptability and generalization capabilities of the model. Finally, experiments are conducted to validate the GAFGraph and FiLM modules on two datasets with significant differences: one under variable speeds and the other under constant speed conditions.