<p>Vibration-based assessment of conventional concrete and steel beams is well established, but analogous tools for 3D-printed plastic and composite beams remain limited. In particular, natural frequency shifts alone may be weak indicators of damage in layered materials with distributed micro-defects. This study proposes an integrated analytical–experimental–machine-learning framework in which a representative power spectrum serves as a damage-sensitive vibration parameter for 3D-printed PLA beams. Free and forced vibration responses of simply supported beams with systematically machined notches are measured at multiple sensor locations, and the acceleration time histories are transformed into normalized one-sided power spectra. These spectra are interpreted using an energy-based stiffness-degradation model and are simultaneously used as input to a compact convolutional neural network–multilayer perceptron (CNN–MLP) classifier that maps spectral signatures to discrete defect scenarios. The results show that, although global natural frequencies change only modestly across defect configurations, the spectral shape encoded in the proposed parameter is highly sensitive to the presence, number, and depth of notches. The CNN–MLP model reliably discriminates between intact and damaged beams, separates multiple defect patterns under both free and forced vibration conditions, and captures the progressive reduction in effective bending stiffness predicted by the analytical model. Overall, the framework demonstrates that full-spectrum vibration features, when coupled with lightweight neural networks and physically grounded stiffness models, provide a more informative and practically deployable basis for quality control and damage monitoring in 3D-printed and other layered beam-like structures than traditional frequency-only approaches.</p>