1. A Dynamic Adaptive Dy-ASPO for Rolling Bearing Fault Diagnosis
- Author
-
Zhai, Shuo and Gao, Dong
- Abstract
Background: Fault diagnosis methods based on deep learning have achieved remarkable results. However, sample imbalance and constructing an appropriate model are still key issues to be addressed. Purpose: In this work, we present a novel diagnostic model design framework named Dynamic Adaptive Structural Parameter Optimization (Dy-ASPO). The proposed design framework integrates input information and training process information to dynamically and adaptively select the optimal structure for the model. Methods: We adopt a new convolution termed Inception Convolution (IC-Conv), which is based on statistical optimization to select the optimal expansion mode in the supernet in a low-cost manner. Second, our proposal adaptively finds operations that contribute most to the loss in the network and applies Dynamic diverse branch blocks (Dydbb) to enhance their representational capacity. Then, we introduce batch-Attention, which is applied to the batch dimension to implicitly explore sample relationships during training. Thus, the model achieves the collaboration of different samples. Results: Two case studies are constructed to test the performance of the proposed method on different bearing failure data under different scale sample distributions and noise combinations. Comparing the accuracy of various methods, the F1-score, as well as the confusion matrix and T-SNE plot of the output of different datasets, reflects the efficiency and practicality of the proposed method. Conclusion: The experimental results demonstrate that the Dy-ASPO operates the internal parameters and structure of the optimization model in a fast and low-cost manner, and solves the issue of the scarcity of samples in an extremely simple way.
- Published
- 2022
- Full Text
- View/download PDF