• Physical modeling using Computational Fluid Dynamic (CFD) simulations and surrogate modeling
• Statistical modeling by machine learning.
• Real-time measurement by an automated data collection framework from wireless sensors.

1- Data Center Wireless Sensor Networks

The main outcome of this research is a reliable, low-power, adaptable, and easy-to-commission Bluetooth Low Energy (BLE) based network architecture for Data Center Wireless Sensor Networks (DCWSNs). Toward this goal:

• A hierarchical network, named Low Energy Monitoring Network (LEMoNet) is designed and evaluated experimentally in a small-scale DC, by simulation in a medium scale DC, and through an analytical model in a large scale DC.
• LeMoNet is further improved by a Software-Defined Networking (SDN) based framework, named SoftBLE.
• Sensor deployment was accelerated by two automated sensor mapping approaches, named thermal piloting and multimodal sensor localization

In this research, a fast, adaptive, and accurate hybrid surrogate model is developed by combining a DDM and physics-based relations to predict real-time temperatures in a DC. Toward this goal:

• An artificial neural network (ANN) in conjunction with a 3D zonal model is employed to predict the spatio-temporal temperature distributions.
• The model is compared with a black-box model based on a nonlinear autoregressive exogenous model (NARX) for interpolative and extrapolative predictions.
• The influence of varying server workload distribution and cooling unit operational conditions on temperature distribution is investigated.

In this research, fault detection and diagnosis (FDD) algorithms and associated control strategies are applied to detect, diagnose, and isolate faults to ensure that the cooling system runs appropriately. Toward this goal:

• A rapid and accurate single and multiple FDD strategy are developed for a DC using data-driven fault classifiers informed by a gray-box temperature prediction model.
• OCSVM and Nonlinear AutoRegressive Exogenous (NARX) techniques are considered for fault detection.
• Two different data-driven classifiers, 2D-CNN and CNN_LSTM, are implemented to predict multiple simultaneous failures while using a few simultaneous failure conditions in training data.