The post-processing of FEM simulations is often a time-consuming and repetitive task, involving lengthy manual operations to detect and quantify deformed regions with respect to criticality, while also dealing with noise. In this paper, we present a novel method for the automatic detection and classification of deformations in a mechanical body, using machine learning within the context of dynamic FEM simulations. Most state-of-the-art methods for deformation and damage assessment of real objects rely on the analysis of 2D images. In contrast, the proposed approach operates directly on the 3D FEM mesh. The 3D information is converted into a graph, initially preserving only nodal curvature change and thereby making the process mesh-independent. Graph Signal Processing (GSP) and standard deep learning are employed to extract meaningful information from the data. Furthermore, rather than using a black-box deep learning model — where tasks are solved by a large, unexplainable neural network — our approach decomposes the problem into a sequence of simpler, well-understood steps, each addressed by tailored algorithms. Description, detection and classification of different deformation classes are demonstrated and validated on synthetic datasets.