Advancing Bearing Fault Diagnosis Using Deep Transfer Learning for Wireless Sensor Networks

Traditional deep learning (DL) models face key limitations in bearing fault detection within wireless sensor networks (WSNs). They require high computational power and large labeled datasets-resources often unavailable in WSNs due to energy, memory, and processing constraints. The scarcity of some bearing fault classes limits the availability of labeled data, complicating effective model training. Additionally, DL models struggle to generalize across varying operating conditions and sensor types, limiting their robustness. Such constraints highlight the inadequacy of conventional DL in WSN-based fault diagnosis and support the use of advanced DL technique compatible with the resource limitations of WSN platforms. This work presents bearing network BearNet, a novel technique designed to enhance bearing fault diagnosis, that strengthens the functionalities of WSN technology in detecting bearing fault by using the concept of transfer learning (TL) with the pre-trained Yet another audio mobilenet network (YAMNet) neural network. Our method converts sensor data into Mel spectrograms, which serve as critical features for training our neural network model. The application of pre-trained YAMNet, along with our tailored target DL model, allows for efficient and accurate classification of different classes of bearing faults. The proposed architecture addresses the constraints of WSNs, such as limited processing capabilities, by utilizing only the unfrozen and additional layers during validation and testing, rather than the entire YAMNet model, thereby optimizing resource usage. Empirical results conducted on the CWRU and MFPT datasets demonstrate that our BearNet technique achieves high diagnostic accuracy, showing significant improvements of 3.1% and between 0.02-5.26% compared to pure YAMNet and state-of-the-art models, respectively. This validates its effectiveness for practical condition monitoring applications across various industrial settings.