Design of a Real-Time Abnormal Detection System for Rotating Machinery Based on YOLOv8

Jianli Chen^1*Jie Tong²Jiang Su²

• ¹Guangdong Institute of intelligent science and Technology, Zhuhai, China
• ²College of Robotics, Guangdong Polytechnic of Science and Technology, Guangdong, China

To address the low accuracy and poor real-time performance of detecting minor abnormalities such as cracks, oil leaks, and loose bolts in industrial rotating machinery under dynamic vibration conditions, this paper introduces a lightweight detection system based on adaptive feature enhancement using YOLOv8 (You Only Look Once version 8). Firstly, a temporal motion compensation module based on the optical flow method is introduced to estimate and align the vibration displacement between consecutive frames. Secondly, a lightweight YOLOv8 network is designed, and depthwise separable convolution is used instead of standard convolution. An adaptive spatial-channel attention module is embedded in the Neck layer to enhance the expression ability of small abnormal features. Finally, a weighted fusion strategy is introduced to improve the detection rate of small targets in complex backgrounds. The model is deployed on the Jetson AGX Xavier edge platform, using FP16 (half-precision floating-point)/INT8 (8-bit integer) quantization and asynchronous pipelined inference to ensure real-time processing capabilities at the edge. Experimental results show that this method achieves an average accuracy of 97.8% (mAP@0.5) and 86.6% (mAP@0.5:0.95), respectively, with an average inference speed of 29.5 FPS (Frames Per Second). This method achieves an industrial-grade balance between detection accuracy, real-time performance, and deployment stability, demonstrating strong value for field applications.