21 November 2023 in IEEE Transactions on Reliability
To address the problem of large parameters in the ResNet model, this article proposes a rolling bearing fault diagnosis method with gram angle field (GAF) and Ghost-ResNet. For addressing the low fault diagnosis accuracy caused by time dimension information loss in image coding, the raw vibration signal is converted into the Gram difference field image. The Ghost convolution module is employed to replace the standard convolution layer in the ResNet model, and a Ghost-ResNet model is proposed for bearing fault diagnosis for the purpose of reducing the parameter sizes and floating point operations. To improve the performance of the Ghost-ResNet model, the RAdam optimization algorithm and Gramian angular difference fields image coding method is used for model training. Experiments with single fault and compound fault are performed for verifying the superiority of the proposed Ghost-ResNet model. The comparison results show that the GAF image coding method is more powerful in representing fault features than the other methods, and the RAdam optimization algorithm has achieved stable and higher diagnostic results with faster convergence speed. Compared with other diagnostic methods, the Ghost-ResNet model has higher diagnostic accuracy, stronger anti-noise performance, and better generalization ability.
5 April 2024 in Applied Mathematical Modelling
16 March 2024 in Reliability Engineering & System Safety
20 March 2024 in Quality and Reliability Engineering International
1 March 2023 in Review of Scientific Instruments
To improve the quality of the non-stationary vibration features and the performance of the variable-speed-condition fault diagnosis, this paper proposed a bearing fault diagnosis approach with Recurrence Plot (RP) coding and a MobileNet-v3 model. 3500 RP images with seven fault modes were obtained with angular domain resampling technology and RP coding and were input into the MobileNet-v3 model for bearing fault diagnosis. Additionally, we performed a bearing vibration experiment to verify the effectiveness of the proposed method. The results show that the RP image coding method with 99.99% test accuracy is superior to the other three image coding methods such as Gramian Angular Difference Fields, Gramian Angular Summation Fields, and Markov Transition Fields with 96.88%, 90.20%, and 72.51%, indicating that the RP image coding method is more suitable for characterizing variable-speed fault features. Compared with four diagnosis methods such as MobileNet-v3 (small), MobileNet-v3 (large), ResNet-18, and DenseNet121, and two state-of-the-art approaches such as Symmetrized Dot Pattern and Deep Convolutional Neural Networks, RP and Convolutional Neural Networks, it is found that the proposed RP+MobileNet-v3 model has the best performance in all aspects with diagnosis accuracy, parameter numbers, and Graphics Processing Unit usage, overcoming the over-fitting phenomenon and increasing the anti-noise performance. It is concluded that the proposed RP+MobileNet-v3 model has a higher diagnostic accuracy with fewer parameters and is a lighter model.
29 January 2025 in IEEE Sensors Journal
Appropriate health indicator (HI) and efficient prediction models are critical factors for accurate remaining useful life (RUL) prediction, particularly, when dealing with fluctuations and redundant information in the HI curve. To address these challenges, this study proposed an HI construction method for better characterization of the degradation behavior based on the entropy weight method and kernel entropy component analysis (EWM-KECA). The HI construction method can eliminate the fluctuations in HI and identify the fault change point position in HI. For RUL estimation, a bearing RUL prediction method was developed by integrating slope-based change point detection with a whale optimization algorithm (WOA)-Attention-bidirectional long short-term memory (BiLSTM) model. By eliminating more than 85% of duplicate data that is not useful for RUL prediction, this approach achieves more accurate RUL predictions while reducing computational resource requirements. The reliability and effectiveness of the proposed method are validated using the bearing degradation dataset. The results from comparative analysis and ablation experiments demonstrate that the proposed method consistently achieves superior performance. Compared with models such as CNN-Attention-BiGRU, WOA-CNN-BiGRU, and WOA-Attention-CNN, the mean absolute error, mean absolute percentage error, and root mean squared error values have been reduced by more than 50%, indicating the proposed RUL prediction methodology represents an advanced and effective approach.