<p>Accurate diagnosis of bearing faults in gas turbines is essential to ensure the proper functioning of the system. However, the harsh operating conditions of gas turbines result in fault data characterized by high dimensionality, strong coupling and noise. In real-world scenarios, especially when faced with limited fault samples, traditional methods are hampered by these factors while striving for optimal diagnostic accuracy. Furthermore, simply increasing the network depth to improve decoupling capabilities can lead to the disappearance or explosion of gradients in deeper network layers. To overcome these challenges, this study proposes a method that combines MSP data enhancement for preprocessing with pyramidal bottleneck residual networks for fault diagnosis. Specifically, the method uses mix-up and random scaling to augment the data, followed by principal component analysis to extract the principal components. The augmented samples are then fed into pyramidal bottleneck residual networks for training. Primarily, the method enhances defect features through preprocessing, allowing the model to establish smoother decision boundaries and improved generalization within the constraints of limited samples. Subsequently, the specialized structure of the pyramidal bottleneck residual networks enhances decoupling capabilities and reduces computational complexity. The effectiveness and superiority of the proposed method are demonstrated through algorithm validation on the CWRU and XJTU public datasets, together with practical engineering verification using BaiChuan gas turbine bearing monitoring data. The method is shown to have high practical value.</p>