Cyclic Code Word Embedding and Chi-Squared Test-Based Attention Mechanism for Deep Learning-Based Server Fault Detection

• Published in: International Wireless Communications and Mobile Computing Conference (Online) pp. 959 - 964
• Main Authors: Xiong, Yiyang, Dong, Shilei
• Format: Conference Proceeding
• Language: English
• Published: IEEE 19.06.2023
• Subjects: Analytical models; Attention Mechanism; Chi-Squared Test; Codes; Cyclic Code; Data models; Deep Learning; Fault diagnosis; Feature extraction; Machine learning algorithms; Wireless communication
• ISSN: 2376-6506
• DOI: 10.1109/IWCMC58020.2023.10182514

Server log data offers a comprehensive record of system operations, making the analysis of this data via algorithms for autonomous fault diagnosis a critical research area. At present, two primary methods of automatic fault diagnosis prevail: traditional machine learning algorithms, such as random forests, which can efficiently and accurately deliver fault diagnosis results after manual feature extraction; and deep neural network algorithms, which can execute end-to-end fault diagnosis with large data sets and yield superior outcomes compared to the former method. However, both techniques exhibit a significant shortcoming: log files constitute time-series data, but machine learning algorithms struggle to effectively capture sequence information. Conversely, deep learning algorithms based on recurrent neural networks can grasp sequence information but lack a thorough comprehension of the system, rendering their implementation relatively demanding, as it necessitates a considerable volume of labeled data. In this paper, we put forth a groundbreaking solution that amalgamates the strengths of both algorithms. We employ the chi-square test algorithm to extract features via machine learning algorithms and subsequently in-corporate an attention mechanism layer into the neural network. This layer modifies the attention mechanism parameters using the chi-square test output. Moreover, we integrate cyclic codes into the neural network's word embedding, a technique that has proven highly effective in communication channel coding. We introduce the utilization of Hamming distance to quantify the disparities among various data, thereby facilitating rapid data comprehension by the neural network. Our proposed algorithm has been corroborated using a publicly accessible dataset and demonstrated a 2 percent enhancement in the F1 score.