AI Assisted Fan Failure Prediction using Workload Fingerprinting

Server fans are one of the most critical parts of the data center server’s thermal system solution and a significant failure cause of many electronic systems. A fan inside the computer chassis is used to draw cooler air from outside and expel exhausted air from inside to adjust a server’s thermal range of electronic components. Fan failure can reduce cooling efficiency and result in unexpected throttling, overheating, and thermal shutdown. This paper demonstrates an AI-assisted methodology that predicts fan failure by developing a probabilistic model that correlates fan speed to workload fingerprint representing resource consumption patterns in the spatio-temporal domain. We use training data from many workloads that execute on many machines of similar configurations and record resource utilization statistics along with environmental parameters with a configurable sliding window. The algorithm evaluates the phase distribution using CPU utilization features and corresponding fan speeds to characterize the expected behavior of phases within individual phase boundaries. This allows us to explain abnormal fan speeds due to CPU’s ambient temperature, voltage, and peak hour workload demands. The paper also presents a novel simulator that helps generate artificial datasets covering various edge cases to evaluate the prediction algorithm’s performance.