Application of physical model-based machine learning to the temperature prediction of electronic device in oil-gas exploration logging

Abstract

Logging tools used for underground oil and gas resource exploration often operate at temperatures above 200 °C, and their internal electronics are highly susceptible to burnout. In order to prevent over-temperature failure of the electronics due to prolonged operation, we avert the risk of over-temperature failure by predicting the real-time temperature with the physical model-based machine learning method. First, based on the transient heat transfer equation of the logging tools, the temperature of the electronic device is determined by the combination of its own historical temperature, the historical temperatures of the surrounding devices, and the physical parameters. Therefore, the impact of surrounding devices on the temperature of the electronic device is considered. Furthermore, we obtain the new ERD (Ensemble Recurrent Neural Network and Deep Neural Network) model with physically meaningful linear assumption, which benefits the reduction of complicated formula calculations. Compared with other machine learning methods, prediction error of ERD model can be reduced by 2 °C, and ERD model also shows good prediction when predicting temperature over a longer period of time. The real-time prediction will expand the application of oil and gas resource exploration techniques in deeper, hotter wells.