Optimization of bi-directional gated loop cell based on multi-head attention mechanism for SSD health state classification model
Year: 2025
Type: conference-paper
Abstract: Aiming at the critical role of SSD health state prediction in data reliability assurance, this study proposes a bidirectional gated recurrent unit (BiGRU-MHA) hybrid model incorporating a multi-head attention mechanism, which effectively enhances the accuracy and stability of storage device health classification prediction by innovatively integrating temporal feature extraction and key information focusing capabilities. The model utilizes the bidirectional timing modeling advantage of BiGRU network to capture the forward and backward dependencies of SSD degradation features, and at the same time introduces the multi-head attention mechanism to dynamically assign feature weights to enhance the identification of sensitive indicators of health status. The experimental results show that the proposed model achieves $\mathbf{9 2. 7 0 \%}$ and $\mathbf{9 2. 4 4 \%}$ classification accuracy on the training set and test set, respectively, with a difference of only $0.26 \%$, demonstrating excellent model generalization performance. Further analyzed by the subject work characteristic curve (ROC), the area under the curve (AUC) on the test set reaches 0.94, which confirms that the model has a highly robust binary classification discriminative ability. This study not only provides a new technical path for SSD health prediction but also breaks through the bottleneck of the traditional model in terms of the performance difference between the training-testing set with a generalization error of only $0.26 \%$, which is of great practical value for the preventive maintenance of industrial-grade storage systems. The result can significantly reduce the probability of data loss by warning potential failure risks in advance, while optimizing the maintenance cost, providing verifiable intelligent decision support for building a highly reliable computer storage system, which is widely applicable to the health management of cloud computing data centers and edge storage devices.
Authors: Zhizhao Wen, Ruoxin Zhang, Chao Wang
Institution: Rice University
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4414605449