<p>Planetary gearboxes, as critical components within rotating machinery transmission systems, often operate under extreme conditions with variable speeds and high loads. Over time, key parts may develop anomalies, leading to failures. Early-stage faults in planetary gearboxes are subtle and challenging to classify and diagnose accurately due to strong noise and random impacts in industrial settings. Recent studies have utilized signals from multiple sensors for fault diagnosis through multi-source information fusion. However, the credibility of classification results under variable speed conditions close to real equipment operation remains a complex issue. To address these challenges, a multi-channel-SCConv residual block (MSResBlock) has been developed for the diagnosis of planetary gearbox faults. This approach enables synchronous feature extraction from various signals to accurately identify fault types under complex variable speed conditions. Initially, a novel multi-channel-SCConv enhanced residual network was proposed, enhancing the network’s ability to integrate spatio-temporal features of different scales from various signals through MSResBlock. A planetary gearbox test rig simulating a production environment was designed, and several experiments were conducted to verify the accuracy and stability of the proposed network. The results demonstrate that the network achieves high classification accuracy and stability on test rig datasets under two types of operating conditions. Results indicated that the network achieved a classification accuracy of 100 % under constant speed conditions and 99 % under variable speeds. In noise resistance tests, it consistently outperformed all comparative methods across various noise intensities, demonstrating excellent noise robustness.</p>