QSL-ResNet: Multimodal feature fusion and transfer learning for mechanical fault sound source localization

Traditional sound source localization algorithms suffer from significant performance degradation when applied to a compact microphone array on mobile inspection platforms in large industrial workshops, particularly in near-field scenarios with strong reverberation and multi-source interference. To address this issue, this paper proposes a mechanical fault sound source localization method based on multimodal feature fusion and transfer learning. This method introduces a Quad-stream lightweight ResNet (QSL-ResNet) that extracts and fuses multiple complementary acoustic features in parallel. It also employs a transfer learning strategy “from simulation pre-training to real data fine-tuning” for rapid adaptation using small labeled datasets. In real-world tests, the system was mounted on a quadruped robotic dog and validated using data collected from a faulty bearing test rig in a real workshop setting. Experimental results show that the proposed method achieves high azimuth localization accuracy with small error under various bearing fault scenarios. It effectively improves the localization robustness of small-scale arrays in industrial near-field scenarios, and provides a feasible technical scheme for fault sound source spatial localization of mobile inspection equipment.