An Innovative Transfer Learning Framework for Rotating Machinery Fault Detection Using Multimodal Diagnosis

Both acoustic and vibration signals generated by rotating machinery contain valuable fault-related information. However, most existing fault diagnosis methods primarily rely on vibration signals, often overlooking the critical fault-related information embedded in acoustic signals. This limitation can lead to missed detections and reduced diagnostic stability. To address these challenges, this paper proposes a transfer learning approach that integrates both acoustic and vibration signals to enhance the precision and robustness of fault diagnosis. Specifically, we introduce a novel fused feature, MAVgram, which combines acoustic and vibration signals to provide a more comprehensive and reliable representation of fault information. This feature is then utilized as input to a Deep Neural Network (DNN)-based classifier, enabling the learning of nuanced fault representations. Furthermore, the backbone undergoes a process of pre-training and fine-tuning to achieve optimal performance in the target task. Experimental results demonstrate the effectiveness of the proposed method, showcasing its effectiveness for improving fault diagnosis in practical industrial applications.