Abstract: An optimal scale selection technique of wavelet multiresolution analysis is proposed, and applied to the identification of time-varying physical parameters. First, time-varying parameters were expressed approximately using wavelet multi-resolution analysis from the perspective of the function space subdivision, and the vibration differential equation can be transformed into a linear regression equation, and the decomposition layers scope was set for every time-varying parameter according to the initial information including the range of frequencies, sampling frequency and the number of linear equations. Then, the optimal decomposition scale was chosen using Akaike information criterion (AIC). In order to enhance the stability of data, the orthogonal least squares algorithm (OLS) was used to estimate the wavelet coefficient instead of the least squares algorithm (LS), and unknown time-varying parameters were reconstructed. Finally, five shear-beam frame models are simulated with two kinds of time-varying parameters cases (abruptly, smoothly). Numerical results show that: in the scope of the decomposition scale preset, identification accuracy increases with decomposition scale when response contains noise, while identification accuracy and the increament of decomposition scale have no obvious connection under the condition that the response data contain noise; appropriate decomposition scale has a great influence on the identification accuracy; and optimal decomposition scale selection can identify the time-varying parameters accurately and improve the computational efficiency and anti-noise ability.