Early detection of minor faults is crucial for ensuring production safety and product quality. However, such faults often exhibit weak signals, are easily affected by noise, and are difficult to detect manually, making traditional diagnostic methods less effective. To address this challenge, this paper proposes an intelligent diagnosis method for minor faults in industrial processes based on multi-source feature fusion, namely an attention-fused modeling approach based on graph convolution and gated recurrent units (GCN-GRU-A). Specifically, the method first employs the Gramian angular field (GAF) technique, in which multi-dimensional sensor data at each time step are combined into a one-dimensional vector to generate the corresponding gramian angular summation field (GASF) image, thereby enhancing the spatial structure features of the data. These GAF images are then processed by a graph convolutional network (GCN) to extract deep spatial features. Concurrently, a gated recurrent unit (GRU) network is used to extract temporal features directly from the raw time-series data. The spatial and temporal features obtained from both branches are subsequently fused through a multi-head attention mechanism, which enhances the representation of key fault-related information while suppressing redundant noise. Finally, the fused features are fed into a classifier for accurate identification of minor fault types. The proposed method is validated on the Czochralski (CZ) silicon single crystal growth process. Experimental results demonstrate that the GCN-GRU-A modeling method outperforms conventional single-feature-based models in several key performance metrics, significantly improving detection sensitivity and diagnostic robustness for minor faults.