Improving Bearing Fault Identification by Using Novel Hybrid Involution-Convolution Feature Extraction With Adversarial Noise Injection in Conditional GANs

• Published in: IEEE access Vol. 11; pp. 118253 - 118267
• Main Authors: Irfan, Muhammad, Mushtaq, Zohaib, Khan, Nabeel Ahmed, Althobiani, Faisal, Mursal, Salim Nasar Faraj, Rahman, Saifur, Magzoub, Muawia Abdelkafi, Latif, Muhammad Armghan, Yousufzai, Imran Khan
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptive systems; channel-agnostic; Conditional generative adversarial networks; Convolution; convolutional neural networks; Datasets; deep learning; Fault detection; Fault diagnosis; Feature extraction; Generative adversarial networks; Generators; Gradient methods; involutional neural networks; mode collapse; Model accuracy; Neural networks; Noise measurement; Robustness; spatial-specific; Training; vanishing gradients
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2023.3326367

Bearing faults are critical in machinery; their early detection is vital to prevent costly breakdowns and ensure operational safety. This study presents a pioneering take on addressing the challenges of imbalanced datasets in bearing fault diagnosis. By mitigating issues such as mode collapse and vanishing gradients, our novel method employs Conditional Generative Adversarial Networks (CGANs) with spectral normalization and adaptive adversarial noise injection to generate high-quality bearing fault samples. This enhances generalization and robustness against noisy data, significantly improving the stability of CGAN training. To extract meaningful features from grayscale bearing fault images, we introduce a novel combination of involution and convolution feature extraction method named I-C FFN. This innovative feature extraction method captures both local and global information, making it capable of handling various types of features, including channel-agnostic, spatial-specific, spatial-agnostic, and channel-specific characteristics. Our proposed oversampling methodology helped enhance the performance of proposed classification scheme as well as of benchmark transfer learning models. Having accuracy values between 99.40% to 99.61% for 0 and 1 HP imbalanced dataset respectively, our models outperformed standard transfer learning methodologies. Furthermore, by the inclusion of our proposed Adaptive Adversarial Class- Conditional GAN (AAC-cGAN), the samples quality and the robustness to noise was significantly increased as demonstrated by the quantitative assessment through various Evaluation Metrics employed in this paper. Lastly, the performance of each combination of both the up-sampled and under-sampled methodologies were assessed through multiple metrics to determine the effectiveness of our proposed approach in addressing class imbalance in bearing fault diagnosis.