Abstract: Objectives In response to the issue of deteriorated fault diagnosis performance in rotating machinery due to noise interference during actual operation, a novel fault diagnosis approach based on Mel-frequency cepstral coefficients (MFCC) and Parallel Dual-Channel Convolutional Neural Network (PDCNN) is proposed. This approach aims to improve two crucial factors influencing fault diagnosis performance: the quality of fault features extracted from vibration signals and the capability of fault diagnosis models. Methods This method mitigates the impact of noise on the monitored signals by extracting more effective fault features. Building upon this, a novel PDCNN is designed to enhance the fault diagnosis performance in a high-noise environment. The extracted fault features are fed into parallel dual-channel convolutional networks that are designed to accept different modes of features. This further improves the diagnostic performance in the presence of strong noise. Results The feasibility of the fault diagnosis method was evaluated through experiments in different noise environments, where the proposed method achieved a fault diagnosis accuracy of over 98% in high noise environments, demonstrating a performance significantly superior to other methods.Conclusions The experimental results demonstrate that this method exhibits strong noise robustness and excellent diagnostic performance.