Bearing Fault Diagnosis Based on Convolutional Neural Network using Estimated Motor Current Signals and their Spectral Portrait

Abstract

Induction motor bearing fault diagnosis stands as a crucial aspect of rotating machinery maintenance. Numerous studies have delved into employing current signals and machine learning methods for this purpose. However, the effectiveness of these approaches relied heavily on manually selecting features for training. Moreover, traditional machine learning techniques struggle with large volumes of computational data. To address these limitations, researchers have turned to deep learning architectures such as Convolutional Neural Networks, ResNet, and AlexNet, either individually or in combination with traditional machine learning methods, for bearing fault diagnosis. Published convolutional neural network-based works usually use basic CNN networks. The experimental data are time or frequency domain data, and the fault classification accuracy is high only with noise-free signals. This paper proposes a novel approach aimed to enhance the accuracy of bearing fault identification by leveraging a CNN model trained on both the estimated motor current signals and their corresponding Fast Fourier Transform values. Comparative analysis against existing methodologies including machine learning and single-input convolutional neural networks or multiinput convolutional neural networks demonstrates that the proposed method achieves impressive results. The bearing fault accuracy reaches up to 99.88% for noisefree signals and 99.14% for signals with added noise at a Signal-to-Noise Ratio of -10 dB.