Noise-Resilient Deep Learning Architecture for Bearing Fault Diagnosis in Integrated IoT Sensing–Computing Environments

Real-time and noise-resilient monitoring in IoT-enabled systems depends on the effective integration of sensing, communication, computing, and control technologies. IoT and edge intelligence have enabled the deployment of real-time fault diagnosis systems for connected industrial and consumer devices. However, the accuracy of such a system is a real challenge due to the presence of the strong and complex background noise that comes along with the data. This work focuses on bearing fault classification under strong noise through a novel enhanced exponential linear unit activation function. An improved end-to-end enhanced convolutional-long short-term memory deep learning model is proposed. The proposed network uses a novel activation function in a convolutional neural network to improve the network’s adaptability against various noise levels. Finally, Case Western Reserve University benchmark data is utilized to assess the model’s performance. Results prove that the proposed model has anti-noise adaptability under strong noises and has a significant advantage in terms of accuracy, precision, and recall over existing methods.