Abstract: To address the pronounced nonlinear characteristics of thermocouple sensors, particularly for the poor linearity exhibited by Type S thermocouples over a wide temperature range of −50~1 768 ℃ in aerospace engine temperature testing, this paper develops a nonlinear compensation model based on hybrid kernel least squares support vector machines to compensate the thermocouple measurement system. Firstly, a hybrid kernel function is constructed by combining radial basis function and polynomial kernels to overcome the inability of single kernel functions to balance local details and global mapping, thereby ensuring linear separability in high-dimensional feature spaces during data processing. Secondly, an improved particle swarm optimization algorithm is proposed for kernel parameter optimization. This enhanced particle swarm optimization incorporates a probability mutation strategy to strengthen its ability to escape local optima, along with adaptive inertia weight and learning factor update strategies designed to adjust the search space dynamically, ensuring both solution accuracy and convergence speed. Experimental results demonstrate that compared with the backpropagation neural network method, the proposed approach reduces the maximum absolute error by 0.260 °C, achieving a post-compensation maximum fitting error of 0.380 °C and an accuracy rate of 98.960%, meeting the stringent requirements for high-precision temperature measurements across broad ranges in practical aerospace engines.