Multi-spectral radiation thermometry based on mixed kernel support vector regression

• A two-stage feature selection method is proposed to extract the spectral radiation features. • The mixed kernel function makes the SVR model performance more stable. • The Bayesian optimization selects the optimal hyper-parameters for the thermometry model. • The proposed method can calculate the object temperature without the object emissivity. With the development of measurement science and technology, multi-spectral radiation thermometry has been widely used in various fields. In the study of its data processing methods, machine learning technology has gathered wide concern due to the advantage of not affected by the unknown object emissivity. However, the existing machine learning thermometry models need generous data samples to support. It has become a key problem hindering the application of machine learning technology in multi-spectral radiation thermometry. In order to reduce the dependence on training samples while ensuring the high measurement accuracy, this study proposes a multi-spectral radiation thermometry based on mixed kernel support vector regression. In this method, a new method of spectral feature selection is proposed for building appropriate training data set. A mixed kernel function is constructed for support vector regression (SVR) thermometry model to make the nonlinear approximating function between the target spectral radiance and temperature has strong stability. Bayesian optimization is used to adjust the super parameters of the thermometry model, so as to ensure the best performance of model. The temperature measurement experiment results of aviation aluminum alloy show that the performance of proposed thermometry is stable in the measurement range under the small sample training data, and the absolute error is not more than 5K.