Design of a real-time abnormal detection system for rotating machinery based on YOLOv8

To address the issues of low detection accuracy and poor real-time performance in existing methods for detecting minor abnormalities such as cracks, oil leaks, and loose bolts in rotating industrial machinery under dynamic vibration conditions, this paper proposes a lightweight detection system based on YOLOv8 (You Only Look Once version 8) with adaptive feature enhancement. First, this paper employs a temporal motion compensation module based on optical flow to estimate and correct the vibration displacement between adjacent frames. Second, this paper designs a lightweight YOLOv8 network, using depthwise separable convolution instead of traditional convolution. Finally, this paper employs a weighted fusion strategy to improve the accuracy of small object detection in complex backgrounds. This model is deployed on the Jetson AGX Xavier edge computing platform, utilizing FP16 (half-precision floating-point) / INT8 (8-bit integer) quantization and asynchronous pipeline inference to ensure real-time processing capabilities on edge devices. The experimental results show that the method achieves an average detection accuracy of 97.8% (mAP@0.5) and 86.6% (mAP@0.5:0.95), with an average inference speed of 29.5 FPS (frames per second). This demonstrates that the method has reached industrial-grade performance in terms of detection accuracy, real-time performance, and deployment stability, making it highly valuable for practical applications.