SAGECN: A Spectral-and-GraphSAGE Enhanced Convolutional Network for Robotic Arm Fault Diagnosis

Modern industrial robotic systems often operate under dynamic working conditions, which can cause faults of motors in robotic arms and further lead to system reliability. However, existing fault diagnosis approaches typically rely on static signal features in combination with fixed thresholds, which limits their ability to capture temporal dynamics and structural dependencies. These methods treat sensor channels independently and lack mechanisms to model the spatiotemporal correlations within sensorial time series, which are critical for detecting subtle and localized fault patterns. To address these challenges mentioned above, this paper proposes a Spectral-and-GraphSAGE Enhanced Convolutional Network (SAGECN). This paper introduces a hybrid graph neural architecture that embeds GraphSAGE-based neighborhood aggregation into a Graph Convolutional Networks (GCN) backbone, enabling joint modeling of global topological structures and localized temporal perturbations. Sensor time-series data are first transformed into structured path graphs using a sliding window mechanism, preserving temporal continuity and directional progression. Then, a multi-stage Graph Neural Network encoder is constructed by stacking GCN and Graph Sample and Aggregate (GraphSAGE) layers, enabling the extraction of both global and local structural features. Finally, top-k pooling is applied for hierarchical node selection and graph-level representation learning. Experimental results on a public dataset from fault-injected motors in robotic arms demonstrate that the proposed model achieves superior performance in fault detection accuracy and generalization compared with existing graph-based models.