Edge-AI implementation of lightweight convolutional models for resource-constrained additive manufacturing systems

Real-time defect detection in Additive Manufacturing (AM) demands robust quality assurance, yet traditional deep convolutional networks such as VGG19 and ResNet50 have prohibitive memory and computational requirements for resource-constrained edge devices. This paper introduces a lightweight Convolutional Neural Network (CNN) architecture optimised through pruning and quantisation for accurate, real-time defect detection on edge devices in AM settings. The CNN was trained and validated on four datasets — Fused Deposition Modelling (FDM), fabric, surface, and casting defects (22,700 images total) — and deployed on an NVIDIA Jetson Nano. Performance was evaluated using accuracy, F1 score, inference latency, and memory footprint against MobileNetV2, ShuffleNetV2, and VGG19. The lightweight CNN achieved 97.8% accuracy with an F1 score of 0.96, matching VGG19 detection performance while requiring only 12 MB of memory and 15 ms inference time — representing 78% memory and 67% latency reductions over VGG19 and fourfold efficiency gains over MobileNetV2. Ablation experiments confirmed that pruning and quantisation were essential to achieving high accuracy with minimal resource consumption. These results demonstrate that purpose-built lightweight CNNs can deliver robust, continuous quality control in resource-limited AM environments, advancing Industry 4.0 adoption by eliminating cloud dependency and mitigating latency, bandwidth, and privacy constraints.