Intelligent Bearing Fault Diagnosis With a Lightweight Neural Network
Nguyễn Đức Thuận, Nguyen Thi Hue, Pham Quang Vuong, Hoang Si Hong
Abstract
Bearing fault diagnosis plays a vital role in machine health monitoring. With the support of deep learning methods, bearing fault diagnosis based on neural networks using vibration signals has achieved excellent results. Compact models with high accuracy that can be embedded in handheld devices are in high demand. This research proposes a novel method for bearing fault diagnosis based on vibration signals using a low-computation-cost deep learning model. Generally, the vibration signal collected from the bearing of the electrical motor is transformed into the spectrogram images by Constant-Q nonstationary Gabor transform, in which these images are used as the training data. Subsequently, the proposed method uses a compact convolutional neural network, called MobileNetV3, combined with AutoCompress pruning method to decrease the number of required parameters and multiply-accumulate operations. Experimental results manifest that the nominated method gains an accuracy of up to 99.34%. Compared with other deep learning models with similar accuracy, the proposed model has a 2-times less number of parameters and approximately 8-times less number of multiply-accumulate operations.