Abstract: It is importance to detect accurately and effectively the residual stress existed in entire lifecycle of mechanical components. Due to the insufficiency of existed methods, a tangential stress detection method was proposed based on ultrasonic amplitude energy attenuation in the time domain to precisely characterize residual stress in components. First, using the attenuation theory of ultrasonic scattering in loaded polycrystalline media, the attenuation coefficient of ultrasonic waves within the Rayleigh scattering range of polycrystalline materials was derived, and a corresponding method was introduced to calculate amplitude energy attenuation and residual stress in the time domain. A tangential residual stress loading system was then established to minimize noise in the collected ultrasonic signals. Finally, linear regression and neural network regression models were built and compared with the traditional cross-correlation algorithm for residual stress detection. Experimental results show that the proposed ultrasonic time-domain amplitude energy attenuation method can achieve higher detection accuracy than the acoustic time difference algorithm, and the best performance can be exhibited with the neural network regression model. Its average error can reach to 4.77 MPa, and get a 52.2% reduction compared to the acoustic time difference method, making it highly applicable in engineering.