This paper proposes an ultrasonic multi-feature fusion bolt stress measurement method based on extreme learning machine (ELM) to address the non-linearity and ill-posedness issues in traditional ultrasonic bolt stress measurement. Firstly, based on the theory of acoustic elasticity and scattering theory, ultrasonic feature parameters such as the acoustic time difference of ultrasonic waves and the attenuation coefficient of longitudinal waves in polycrystalline materials within the Rayleigh scattering range are extracted from ultrasonic echo signals. Then, by vector dimension reduction, the acoustic time difference, attenuation coefficient and effective load length are selected as the input feature vector to establish an ELM-based ultrasonic multi-feature fusion bolt stress measurement model. A bolt axial stress ultrasonic measurement experimental platform is set up to measure the stresses of bolts of different materials and specifications. The results are compared with those of traditional ultrasonic measurement methods to verify the limitations of traditional ultrasonic detection methods. The measurement results and precision of ELM are compared with other machine learning methods, including back propagation ( BP) and support vector regression ( SVR). The results show that the method proposed in this paper effectively overcomes the shortcomings of traditional ultrasonic measurement methods, which can measure the stress of bolts of different materials and specifications, and has higher measurement accuracy (with an average relative error of 3. 86% ) and better generalization abili K ty.