Data driven fault tolerant thermal management of data centers

In this work we address the problem of fault-tolerant operation of data center thermal management systems. A data driven fault tolerant control framework is proposed. First, a data driven dynamic predictive model is trained using subspace identification. This model is utilized in a model predictive controller to predict dynamic behavior of a data center. An economic model predictive control is designed to maintain temperatures within an allowable range while minimizing operational costs of the cooling system. An observer based fault diagnosis method is then developed to detect and isolate sensor faults. When a fault occurs, it is diagnosed and the corresponding measured signal is replaced with the reconstructed signal from a state estimator. The effectiveness of the proposed method is illustrated through simulations on a mechanistic data center model.