A Method for Rolling Bearing Fault Diagnosis Based on Dynamic MFCC Features and a Dual-Path Attention-Multi-Scale Aggregation Network (DPA-MSAN) Under Heavy Noise

Abstract

Under severe noise interference in industrial environments, the non-stationary and nonlinear characteristics of rolling bearing vibration signals become more pronounced, making weak fault features easily buried and diagnosis significantly more challenging. To address this issue, this paper proposes a fault diagnosis method that integrates dynamic Mel-frequency cepstral coefficient (MFCC) features with a Dual-Path Attention Multi-Scale Aggregation Network (DPA-MSAN). The method first converts one-dimensional vibration signals into two-dimensional dynamic MFCC feature maps, and then employs multi-scale convolutional branches to jointly capture local transient features and long-term evolution patterns. A dual-path attention (DPA) module further suppresses noise and enhances discriminative information by recalibrating features in both the channel and spatial dimensions. Finally, a multi-scale attention aggregation (MSA) module adaptively fuses features from different branches, strengthening the model’s robustness to complex noise conditions. Experiments on the CWRU and XJTU bearing datasets show that the proposed method maintains excellent diagnostic performance under heavy noise. Even at an extreme SNR of −6 dB, it achieves 91.85% accuracy on the CWRU dataset and 93.02% on the XJTU dataset, demonstrating strong cross-operation generalization and outperforming several advanced models.