In pulsed eddy current testing of multi-layer ferromagnetic tubulars, strong interlayer electromagnetic coupling and deep layer diffusion attenuation coexist, making the quantitative decoupling of wall thicknesses in different layers a typical nonlinear ill-posed inverse problem. To address the limitations of traditional data-driven methods, including insufficient physical constraints, strong sample dependence, and limited generalization ability, a pulsed eddy current wall thickness inversion method for multi-layer tubulars based on a physics-guided multi-scale residual network is proposed. First, a semi-analytical forward model for pulsed eddy current testing of multi-layer tubulars is established based on the truncated region eigenfunction expansion method, enabling fast solution and response calculation for finite-length axisymmetric electromagnetic boundary problems. A logarithmic-domain sensitivity matrix and relative contribution indices are then constructed. On this basis, sensitivity-gradient direction and channel-complementarity analyses are introduced to reveal the time-domain evolution law in which the wall-thickness sensitivity of different layers migrates from the inner layer to the outer layer, thereby providing a physical basis for multi time channel feature extraction and joint observation using long coil and short coil information. Next, a deep residual network centered on multi-scale one-dimensional (1D) convolutions is constructed, fusing the complementary information from long and short coils and embedding physical priors at the input stage. Meanwhile, a stage attention aggregation mechanism and a physical consistency constraint are introduced at the feature and loss stages, respectively, guiding the network to learn a wall thickness mapping relationship that complies with the underlying electromagnetic response mechanism. Finally, a hybrid dataset constructed using the semi-analytical forward model and COMSOL simulations is used for training and validation, followed by experimental verification is carried out on three-layer ferromagnetic tubular specimens. The results demonstrate that the proposed method achieves a three-layer average mean absolute error of 0.150 mm and an R2 of 0.995 6 on the simulation test set. Under the measured condition of intact wall thickness, the average mean absolute error is 0.251 2 mm, indicating that the proposed method can effectively characterize the variation law of wall thickness in multi-layer tubulars.