Smart manufacturing towards cyber-physical resilience in 3D printing process monitoring and anomaly detection

Abstract The digital threads of additive manufacturing (AM), originating from smart manufacturing, leverage Cyber-Physical Systems (CPS) that integrate interconnected cyber and physical domains, wherein the cyber domain encompasses product design processes (e.g., CAD modeling, .STL file creation, G-code generation, and cloud-based data sharing), while the physical domain employs the G-code to drive 3D printing and enable post-process monitoring. G-code modifications are a common vector for malicious cyberattacks on AM systems, enabling alterations to part designs and print parameters that disrupt the printing process and degrade the mechanical properties and functionality of mission-critical parts. To detect cyber-attacks in AM systems, this study employs National Institute of Standards and Technology (NIST) cybersecurity frameworks to systematically detect cyber-attacks in AM systems and uses a fused filament fabrication (FFF) 3D printer testbed to simulate raster angle-based alteration classification (RAAC) for sensitivity analysis. This work focuses on G-code alteration detection using Euclidean distances along the x, y, and z axes, and employs layer-by-layer in-situ video monitoring of the FFF process via a digital microscope camera mounted on the extruder head to identify fabrication defects and potential cyber-physical attacks. Thus, layer-wise image frames are processed using adaptive region-of-interests (ROIs) to adjust spatiotemporal image fames’ raster angle and RGB-to-grayscale conversion to extract rasterized surface texture features (STFs), which capture surface patterns at varying raster angles and are refined using principal component analysis (PCA) for dimensionality reduction and feature extraction. This study evaluates a multiclass support vector machine (SVM) to detect RAAC attacks on AM systems. The model’s effectiveness is measured by using confusion matrices to analyze classification accuracy. The RAAC approach proved more effective than benchmark methods like Independent Component Analysis (ICA) and Convolutional Neural Networks (CNNs).