.png){kind=icon}
.png){kind=icon}
Return
Machine learning based meta-models for sensorless thermal load prediction
.png){kind=icon}
PRE
AI
DOI：10.1109/ITherm51669.2021.9503263.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Lifetime of electronic systems depends various factors determined during product development stage, e.g. design, materials, quality of components and manufacturing. In the field, the load of the system leads to degradation affects the remaining useful life of the product. In this work, we focus on thermal load conditions. The thermal load is induced externally by ambient temperature and internally due to heat generated by operation of components. In most consumer use cases external load is a constant room temperature. In industry, e.g. automotive or aviation, the external load shows strong fluctuation and appears in cycles. The heat loss during operation causes superposition of internal and external thermal loads distributed over the system. Our goal is to create a meta-model that allows the calculation of temperature at various points of interest in the system based on virtual sensors in these points. The model is trained based on experimental data with attached thermal sensors in the points of interest. A temperature chamber provides ambient temperature cycles in a range of -50 to 50 degrees C, resulting in maximum CPU temperatures of 95 degrees C. A program running on the system stresses the CPU of the system to certain load levels. These load levels cause heat loss in the CPU, which is distributed to the system. The results are time series for the CPU load, CPU temperature, ambient temperature and temperatures of the attached thermos couples. We test several mathematical and machine learning approach to obtain the temperature in the points of interest, which are the outputs of the models. CPU load, ambient temperature (and in some cases CPU temperature) serve as inputs. In further research attempts, we run the experiment to failure of components. Based on these data, models can be extended to calculate the remaining useful life of the system.
Keywords:
Meta model
machine learning
PHM