Comparative Analysis of Machine Learning Techniques for Temperature Compensation in Microwave Sensors

The planar nature of microwave sensors leaves them vulnerable to ambient temperature changes with potential impact on the perception of the material under test. A temperature compensation technique is required to consider its direct effect on the dielectric property of materials. In this article, machine learning algorithms are employed in two configurations of classifier and regressor on frequency response of a split-ring resonator operating at 1.19 GHz. A wide range of dielectric constant is covered with concentrations of [0:20%:100%]-methanol/acetone in water with a temperature cycle of 25 °C–50 °C. This broad variety of cases captures the complicacy of entangled trends that are recognized using classifiers regardless of the measurement temperature. In the next step, the ambient temperature is extracted from the same measured data. This is accomplished by cascading the classifier with a regressor pool that contains trained parameters for individual classes. Highly accurate classification of material types followed by their corresponding temperature (using linear regression with $R^{2}=97$ % averaged over tenfold cross validation and a mean absolute error of 0.58 °C) leads to investigating limit of detection in the proposed scheme. This step, through testing of [0:1%:5%] methanol in water, identified multilayer perceptron (MLP) and support vector machine (SVM) as the best performing algorithms. Final hyperparameter optimization yields parameters for these two models that provide accuracy of 0.97 and 0.99, respectively.