Kalibre

Exploiting sensor and process characteristics to tackle label scarcity in AIoT sensing

“…This thesis targets AIoT applications where the generation of data is governed by certain first principles and exploits sensor and process characteristics to tackle The study [118] provides an example of this method, in which the cooling and heating units in a data center space are incorporated into the architecture of the neural network for temperature prediction. The method based on learning biases incorporates physical laws into the DNN model's loss function.…”

Physics-informed machine learning for green data center operations

“…A variety of modeling techniques have been proposed for thermal management in data halls. They can be broadly categorized into white-box [10,35,[49][50][51], black-box [52,53] and grey-box [9] methods. The CFD models are representative white-box models, in that they capture the thermodynamic laws followed by the physical processes.…”

“…They can be classified into partial-reserved and full-reserved approaches. The partial-reserved approaches simulate the effects of certain parameters on temperature only at certain discrete points, such as the server inlets/outlets [50] or the cold/hot aisles [35]. To maintain the spatial resolution, the fullreserved reductions are desirable for a complete temperature field approximation.…”

“…The testbed captures a data hall that is equipped with one CRAH unit and two rows of racks hosting 299 servers as shown in Figure 4.14(a). The rated power of each server is 1,000W and the server air flow rates are calibrated by the method in [35].…”

“…(2.1), which can be also incorporated into the loss function for training. However, when the spatial domain is discretized with fine-grained mesh grids, the number of temperature points N can be up to millions [35]. Thus, directly approximating such high dimensional output can cause statistical and computational issues.…”