This study presents an approach for damage quantification of trusses utilizing free vibration signals and Convolutional neural network (CNN) relied on model order reduction (MOR). The input data consists of the values of eigenvectors extracted from several important degrees of freedom (DOFs) instead of all ones, collected from numerical simulations under various random damage scenarios. The output is the truss members’ randomly assumed damage ratios. The Modal strain energy-relied index (MSEI) is applied to eliminate members with a low probability of damage, aiming to reduce the data dimension for CNN. Thereby, its accuracy of predicting the damage detection is improved with the capability of automatically extracting features from CNN, this method significantly reduces the computational cost in training and testing compared to traditional methods. The methodology is validated on a 2D truss model under two damage scenarios programmed in Python. The results are promising for providing the method's potential applications to structural health monitoring (SHM).