Research on Edge Deployment of Anomaly Detection Algorithm Based on Domestic AI Chip

With the rapid development of artificial intelligence and edge computing technologies, their integrated applications in safety monitoring have become a research hotspot. This study focuses on concrete crack detection. Three pretrained lightweight convolutional neural networks—ResNet18, MobileNetV3-Large, and GoogLeNet (Inception V1)—are fine-tuned via a transfer learning strategy and trained on the SDNET2018 concrete crack dataset. To address the resource constraints of deploying deep learning models on edge devices, the RKNN toolkit is employed for architectural optimization and deployment. Experimental results demonstrate that MobileNetV3-Large achieves the best overall performance after deployment. Its precision, recall, and F1-score reach 90.68%, 95.43%, and 92.99%, respectively. The model maintains high accuracy while occupying only 11.36 MiB of memory and achieving an inference speed of 6.527 ms. The practicality of MobileNetV3-Large for edge-based concrete crack detection is validated, and the advantages of the domestic RK3588 chip in low-power, high-performance AI deployment are highlighted. The proposed framework provides a reliable edge solution for safety monitoring of concrete structures such as bridges, dams, and buildings.