Multi-scale adaptive-routing capsule contrastive network-based intelligent fault diagnosis method for rotating machinery under noisy environment and labels

• Published in: Advanced engineering informatics Vol. 62; p. 102712
• Main Authors: Xiong, Yucheng, Liu, Zhiwen, Tan, Jiyong, Hao, Liang
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.10.2024
• Subjects: Adaptive-routing; Capsule network; Contrastive learning; Intelligent fault diagnosis; Multi-scale residual blocks; Noisy environment and labels; Noisy environment and labels; Multi-scale residual blocks; Adaptive-routing; Intelligent fault diagnosis; Capsule network; Contrastive learning
• ISSN: 1474-0346
• DOI: 10.1016/j.aei.2024.102712

•Multi-scale adaptive-routing capsule contrastive network for rotating machinery intelligent fault diagnosis is proposed.•Multi-scale Residual Blocks is introduced into capsule networks to capture multiscale feature information.•Self-supervised contrastive learning strengthens the model’s representation of raw vibration sample information.•Improved adaptive-routing mechanism is developed to suppress irrelevant features and enhance diagnostic accuracy.•The experimental results show that the proposed model has high robustness under noisy environment and labels. The robustness and accuracy of fault diagnosis are affected by noisy environment and labels from manual annotation errors in vibration signal data collected during the operation of mechanical equipment. Therefore, this study develops a multi-scale adaptive-routing capsule contrastive network-based intelligent fault diagnosis method for rotating machinery under noisy environment and labels. Firstly, multi-scale residual blocks (MSRB) are introduced into the capsule network to utilize receptive fields of different scales, comprehensively capturing multi-scale feature information. Meanwhile, the structure of the capsule network is further enhanced by employing a non-iterative adaptive-routing algorithm, preventing the amplification of irrelevant features due to the sparsification of coupling coefficients and improving the accuracy of the classification model. Secondly, a supervised contrastive learning module is added to enhance the model’s discriminative ability for multi-scale features, improving the model’s representation capabilities. Finally, the proposed method achieves “end-to-end” intelligent recognition of multiple faults by optimizing the model parameters via the established loss function and input data. Experimental results demonstrate that the proposed method significantly reduces the impact of noisy environment and labels, achieving high-precision classification. Compared to other classification methods, the proposed method has a clear advantage in noise resistance and stronger robustness.