To enhance the audio signal features of plunger pump under strong background noise and improve the accuracy of fault diagnosis, a method for extracting the fault audio features of the plunger pump based on adaptive noise reduction is proposed. By introducing the Gammatone cepstral transform to extract the features, the audio signal of the plunger pump is transformed into the time-frequency domain. Meanwhile, an adaptive noise reduction method is proposed to remove the background noise irrelevant to the fault in the time-frequency signal. Finally, a fault classification comparison experiment is carried out through the Resnet-18 neural network. The results show that the fault diagnosis accuracy of the plunger pump is improved to 96.97% after adaptive noise reduction, which verifies that the proposed feature extraction method can effectively reduce the influence of the background noise of the plunger pump, and improve the fault diagnosis accuracy.