Human auditory system has excellent recognition ability and anti-noise ability in processing non-stationary and nonlinear signals such as sound signals. Based on the homology of vibration and sound signals, a method based on auditory saliency signal data denoising, typical feature extraction and self-organizing feature map (SOM) network was proposed to evaluate the performance of generator set main bearing. Firstly, the human cochlea model was constructed by using a Gammatone filter bank to identify the original vibration signal and eliminate the noise signal. Secondly, the typical features of salient frames and salient channels were obtained by simulating the auditory attention mechanism of human ears, and then the feature space was constructed. Finally, the constructed feature space was divided into training samples and test samples, and the SOM network was used to realize the performance evaluation of the generator set main bearing. The experimental results show that the performance evaluation method proposed can accurately identify the noise signal and construct the feature space, which can effectively evaluate the performance of the generator set main bearing and provide a basis for its condition-based maintenance.