基于改进深度学习的风机油污识别
2023,31(5):174-179
摘要：针对风机设备油液渗漏影响风机正常运行亟需解决的对风机设备油污的识别问题，提出了一种基于改进深度学习的风机油污检测方法。基于深度学习在目标检测中的应用特点，对目标检测网络YOLOv5n（You Only Look Once v5n）进行改进，将原网络中的非极大抑制（Non Maximum Suppression，NMS）替换为Soft-NMS，降低了网络的误检率，添加CA (Coordinate Attention)注意力机制，增强了模型对目标的定位能力，改进原网络损失函数为α-IoU（Alpha- Intersection over Union）损失函数，提高了边界框检测的准确度。实验结果表明:模型平均精度提升了8.1%，查全率提高了19.1%，网络推理速度提高了28.6%。改进后的模型能准确检测风机油污，有效解决了风机实际运行中油液渗漏所带来的问题。
关键词：深度学习;风机油污;注意力机制;损失函数;非极大值抑制
Fan oil contamination identification based on improved deep learning
Abstract：Aiming at the identification problem of oil pollution of fan equipment that needs to be solved urgently when the oil leakage of fan equipment affects the normal operation of fan equipment, a method of oil pollution detection of fan equipment based on improved deep learning is proposed. Based on the application characteristics of deep learning in object detection, the object detection network YOLOv5n (You Only Look Once v5n) is improved, the non maximum suppression (NMS) in the original network is replaced by Soft-NMS, the false detection rate of the network is reduced, the CA (Coordinate Attention) attention mechanism is added, and the positioning ability of the model to target is enhanced. Improved the original network loss function to the α-IoU (Alpha-Intersection over Union) loss function, improving the accuracy of bounding box detection. Experimental results show that the average accuracy of the model is improved by 8.1%, the totality rate is increased by 19.1%, and the network inference speed is increased by 28.6%. The improved model can accurately detect the oil pollution of the fan, and effectively solve the problem caused by oil leakage in the actual operation of the fan.
Key words：deep learning; fan oil pollution; attention mechanisms; loss function; non-maximum suppression
收稿日期：2022-09-22
基金项目：江苏省高等学校自然科学基金面上项目(21KJB120005)
