Abstract
The vibration signals of gas turbine rolling bearings exhibit significant non-stationarity under complex operating conditions such as frequent start-stop cycles and variable speeds, posing a major challenge for fault diagnosis. To address this issue, this paper proposes a multi-channel variable-speed attention framework (MC-VSAttn). The method first constructs multi-channel inputs to capture rich fault information, then introduces a dynamic time-varying response module to adaptively model non-stationary features, and combines channel and spatial joint attention mechanisms to enhance selective attention to critical information, thereby achieving robust fault identification under complex operating conditions. Compared with existing methods, the proposed framework explicitly models the time-varying characteristics of non-stationary signals and jointly integrates multi-channel fusion with hierarchical attention, enabling more accurate and stable fault diagnosis across variable-speed scenarios. Experimental results based on the variable-speed datasets from Tsinghua University and Huazhong University of Science and Technology show that MC-VSAttn achieves accuracy rates of 99.14% and 98.23%, respectively. Further ablation experiments validate the key role of the dynamic time-varying response module and the joint attention mechanism in performance improvement.