Abstract
The rapidity of the process in modern manufacturing implies a high speed of continuous monitoring, which often surpasses the potential of manual control. The study presents the adaptive, energy-efficient intelligent platform for monitoring defects with the use of Edge Artificial Intelligence (AI) architecture. In terms of objectives, special attention is paid to solving latent computational problems in real-time processing on the local edge device at "Raduga" LLP. At the beginning, the study performs a comparison of One-to-Many and One-to-One detection architectures. The findings show that the Non-Maximum Suppression (NMS) step decreases the speed of local graphics accelerators; accordingly, the system under development uses NMS-free neural network detectors. In order to standardize the process of defining the data flows and algorithms, the following standards for system-structural modeling are used: IDEF0, IDEF3 and DFD. Besides, the system includes a lightweight You Only Look Once (YOLO) 26 Nano detector and ByteTrack object tracking algorithm. Thanks to this software module, it becomes possible to overcome computational difficulties in post-processing of frames and to perform stable object tracking on a moving conveyor. Finally, operational tests show that the system performs with stable real-time speeds on the NVIDIA GeForce RTX 2050 GPU with FPS above 30 for the conveyor speed of 1.5 m/s..