Machine Learning-Based Anomaly Detection in Enterprise Hardware Telemetry Streams

Abstract

Enterprise hardware systems continuously generate telemetry streams that can be mined for early signs of faults or abnormal behavior. Machine learning-based anomaly detection offers a proactive approach to identifying performance issues and impending failures from this telemetry data. In this paper, we investigate state-of-the-art ML techniques for anomaly detection in real-world enterprise hardware telemetry, focusing on detection accuracy and false positive rates. We utilize actual telemetry logs (e.g., CPU utilization, memory usage, I/O stats) from data center servers and network devices, applying unsupervised and semi-supervised learning algorithms to flag unusual patterns. Our results demonstrate that a hybrid anomaly detection pipeline can achieve high detection accuracy (over 95% in our tests) while maintaining a low false alarm rate, enabling timely identification of hardware issues with minimal noise. We analyze the trade-offs between different algorithms – including statistical methods, isolation forests, and LSTM-based models – and show how combining detectors improves robustness. The paper is organized into sections covering the introduction of the problem, related work, methodology, results with two illustrative figures and tables, and conclusions.