Multi scale parallel frequency attention network for bearing fault diagnosis under severe noise

Vibration signals collected from harsh environments often suffer from severe noise interference, which can submerge critical fault characteristics. To address this challenge, a frequency attention network based on multi-scale parallel ResNet (MSPResNet-FA) is proposed. The architecture integrates a Fast Fourier Transform Convolution (FFT-Conv) module to extract harmonic frequency impulse feature, while parallel multi-scale kernels in the convolutional layer are employed to extract various scale local frequency features. Furthermore, a Convolutional Block Attention Module (CBAM) is utilized to enable the network to focus on and extract important features from the frequency domain especially the multi-scale harmonic frequency impulse feature. The effectiveness of MSPResNet-FA was validated using four benchmark datasets: Politecnico di Torino (PDT), Xi’an Jiao Tong University (XJTU), Paderborn University (PU), and the Society for Machinery Failure Prevention Technology (MFPT). Experimental results indicate that MSPResNet-FA outperforms current state-of-the-art methods, notably achieving accuracy gains of over 5% on the XJTU, MFPT, and PDT datasets. Ablation studies confirm that the FFT component is foundational to this high performance, contributing an improvement of over 10%, while the CBAM refinement layer provides an additional 1% enhancement.