DGLAConv-TasNet Motor Bearing Fault Sound Source Separation Network

In industrial production, achieving one-to-many audio monitoring and precise separation of overlapping motor sound sources is a crucial step in enhancing the efficiency of equipment fault diagnosis and ensuring the safe operation of production lines. To address the challenging task of separating individual motor sound sources from the audio of multiple motors operating simultaneously in industrial scenarios, we propose a novel end-to-end fully convolutional time-domain source separation network for motor sound source separation under complex interference conditions. This paper constructs a multipurpose acoustic fingerprint bearing fault dataset (HB-bearing fault dataset) tailored for real-world transportation scenarios. The main innovative contributions include: designing an enhancement module based on variational mode decomposition to avoid the issue of traditional de-noising methods treating the overall motor operation sound as noise suppression ; proposing a variable-depth time-domain convolutional network structure to address the mismatch between the original network’s receptive field and the motor pulse width; and designing an attention structure composed of local channel attention and local temporal attention to solve the problem of the network being prone to overfitting and difficult to generalize. Experiments show that the proposed method performs superiorly on the mixed fault task of the HB-bearing fault dataset.