Counterfactual Diagnosis of Cooling-System Anomalies in Data Centers

Abstract

Data center cooling systems generate continuous multivariate time series from rack temperature, chilled-water flow, supply-air temperature, fan speed, valve position, power usage, humidity, and workload distribution. These variables have strong causal relationships, and a small abnormal change in one cooling component may trigger delayed effects across racks, air-handling units, and power systems. This study develops a counterfactual dependency modeling method for cooling system anomaly diagnosis in data centers. The proposed method learns causal dependencies among thermal, electrical, and workload-related variables through time-lagged structural modeling. A counterfactual prediction layer estimates how each variable should behave if suspected abnormal drivers were removed. Fine-grained anomaly attribution is then performed by ranking variables according to their counterfactual contribution to system-level deviations. Experiments are conducted on a data center operations dataset containing 14 server rooms, 620 racks, 74 cooling units, and 68 monitoring variables collected every 30 seconds over 210 days. The dataset contains 411 million time-stamped records and 1,960 labeled abnormal episodes, including chilled-water valve failure, fan-speed oscillation, rack hotspot formation, abnormal humidity rise, cooling-loop imbalance, and sensor calibration drift. The proposed method reduces median diagnosis delay from 24.8 minutes to 7.3 minutes compared with a temporal graph autoencoder. The number of unnecessary cooling-unit inspection tickets decreases from 1,420 to 530 during the evaluation period. Causal path analysis identifies 1,310 primary driver variables before downstream temperature deviations become dominant. The median inference latency is 36 ms per monitoring window, supporting online thermal risk diagnosis. These findings indicate that counterfactual dependency modeling can provide interpretable and fine-grained anomaly diagnosis for data center cooling infrastructure.