<p>Multi-fidelity surrogate models based on Gaussian processes are widely utilized in engineering and optimization due to their ability to quantify uncertainty. The majority of current multi-fidelity Gaussian process models are constructed using stationary kernel functions, which assume that the fluctuations of the response surface remain consistent across the input space. However, this assumption of stationarity is frequently violated when the response surface exhibits complex variations, leading to an inaccurate representation surface. This paper proposes a non-stationary multi-fidelity surrogate modeling method incorporating sequential sampling, designed to improve the performance of surrogate models for non-stationary systems. In this method, a closed-form distance warping mapping is introduced to transform the non-stationary system into an approximately stationary one by leveraging multi-fidelity samples. Meanwhile, the structure of the distance warping mapping is determined through density clustering on the samples to reduce the complexity of solving the multi-fidelity surrogate model. Through six analytical functions and the breakdown voltage prediction of the SiC-MOSFET device, the superiority of the proposed method is indicated from the aspects of modeling accuracy and suitability for sequential criteria. Experimental results further show that the proposed method improves the performance of sequential sampling to a certain extent due to the correct evaluation of prediction uncertainty.</p>