As a critical power component in rotating machinery, transmission bearings are widely used in high-end equipment such as wind power generation, aerospace, and industrial manufacturing. Their health status directly affects the safety and reliability of overall machine operation. Due to complex operating conditions and alternating loads, they are prone to wear and faults. Among them, compound faults of transmission bearings are the main cause of unexpected downtime in rotating machinery. The fault characteristics are coupled and easily masked by the primary fault, greatly increasing the difficulty of diagnosis. Aiming at the limitations of existing compound fault diagnosis methods that mostly rely on single-sensor data, which lack sufficient information representation dimensions and have difficulty fully capturing fault dynamic characteristics, a collaborative diagnosis method for transmission chain compound faults based on zero-sequence current and vibration signal graph-domain fusion is proposed. First, a method based on the symmetrized dot pattern (SDP) is established to achieve spatiotemporal alignment and graph-domain fusion of cross-modal heterogeneous data from zero-sequence current signals and vibration signals. This realizes pixel-level spatial alignment and feature-level information fusion of the two types of signals in the graph domain, effectively avoiding the separation of physical meanings caused by traditional feature concatenation. Second, a collaborative fault feature extraction architecture combining shifted window transformer (SwinT) and convolutional neural network (CNN) is constructed. The SwinT sliding window attention mechanism is designed to extract local detail features, and the CNN global attention module is used to enhance overall spatial feature extraction. Through adaptive average pooling, deep complementary fusion of local-global features is realized. Finally, validation experiments are conducted on a real compound fault dataset. The proposed method achieves a diagnostic accuracy of 98.8%, effectively solving problems of modal aliasing and difficulty in feature extraction under compound faults, providing a new approach for accurate diagnosis of compound faults in rotating machinery.