Indicator Fault Detection Method Based on Periodic Self Discovery and Historical Anomaly Filtering

Abstract

Data centers’ information systems typically encompass a variety of operational objects including applications, systems, networks, and devices, which generate a large volume of indicator data during operation. The traditional threshold-based indicator fault detection method has struggled to adapt to the massive and heterogeneous nature of distributed architectures, resulting in numerous false positives and negatives. Existing research has limitations in terms of the accuracy of fitting indicator operating characteristics, as well as performance. To address the problem of fault detection for indicator data, this article proposes a method based on periodic self-discovery and historical anomaly filtering. It achieves periodic self-discovery based on Fourier transformation, significantly reducing the training cost such as model parameter tuning. Additionally, it introduces a quadratic fuzzy filter based on periodicity to effectively solve possible local misclassification issues while generating a baseband that is more suitable for indicator operating characteristics, improving detection accuracy. Through experimental validation, the fault detection method proposed in this article significantly improves the accuracy of indicator for operational objects. Compared to directly using ARIMA and SARIMA models, this method has improved the MSE by up to 44% and 36%, respectively. This method has also been practically applied in multiple scenarios in a bank’s production environment, increasing alert accuracy by 20% and achieving a ratio of 10:1 for alert convergence. Compared to traditional methods, it significantly enhances the ability of fault detection for indicators.

Keywords

• Fault detection
• periodic self discovery
• historical anomaly filtering
• dynamic baseline
• indicator