A deep agent reinforcement learning-based early adaptive multivariate state estimation fault diagnosis method for multi-mode cooling system in data center

Modern cooling systems are multi-mode characteristics for energy conservation, which brings a big challenge for accurate fault diagnosis due to data feature variation in different operation modes. This paper proposes a deep agent reinforcement learning-based early adaptive multivariate state estimation fault diagnosis method to address this issue. Firstly, an enterprise multi-mode data center cooling system is modeled under Beijing's environmental temperature conditions, which can automatically transit between the natural cooling, precooling, and mechanical cooling modes based on outdoor temperature, achieving 39.6 % energy consumption. Then, considering cooling system's multi-mode characteristics, a deep agent reinforcement learning-based early adaptive multivariate state estimation method is proposed, the introduction of deep reinforcement learning improves the fault diagnosis performance of the multi-mode cooling system by adaptively updating the memory matrix, which can be formulated to an agent decision problem in deep reinforcement learning, equivalent to the Markov decision process. Finally, three state-of-the-art comparison methods and five typical cooling system faults are simulated to evaluate the fault diagnosis performance. The proposed method can advance the fault warning time by more than 9 h, and improves the average fault diagnosis rate by 1.38 %, 2.23 %, and 10.38 %, while reduces the fault alarm rate reduced by 0.36 %, 0.46 %, and 8.05 %.