Gas turbine, as a power source driving the heart of modern industry, is prone to rotor faults such as misalignment and unbalance in long-term high-load operation, which can degrade the performance of the equipment and cause safety accidents. The early fault signals are weak and easily masked by noise, which affects the model identification accuracy. To this end, a hybrid diagnostic method combining empirical modal decomposition (EMD) and wavelet threshold noise reduction (WT) is proposed: firstly, the original signal is decomposed by EMD to obtain a series of IMF components, and the noise in the high-frequency IMF components is removed by WT with a well-set threshold, and then the signal is reconstructed; and then, the fault features are extracted by a multi-scale convolutional network (MSCNN) with the addition of an attention mechanism. Experiments show that the method achieves 94.4% accuracy in the rotor fault test set, effectively improving fault identification accuracy across working conditions.