In this paper we propose a new evaluation method for revealing the inner structure information from the measurements of the thermal transient response of semiconductor devices. In contrast with the conventional methods based on the frequency domain deconvolution, the proposed method casts the problem as a regularized least squares problem for an integral equation directly in the time domain. The regularization is selected based on the knowledge of certain physical quantities involved in the measurement. We observe that the thermal time constant spectrum produced by the proposed method tends to have a better capability of identifying two adjacent parameters. With the benchmark of the true values of the thermal networks, the structure function of the proposed method reduces the relative error by an order of magnitude. A series of simulations based on different parameters and structures verify the viability and performance of the proposed method. A proof-of-concept experiment based on a commercial power semiconductor device also validates the effectiveness of the proposed method.