<p>Bushings have a wide range of applications in industry. Once the surface of the bushing is defective, it will affect the assembly between bearings resulting in mechanical inefficiency. At present, due to the different target sizes of the different types of defects on the bushing surface, it is difficult to balance inspection accuracy and speed. This paper proposes lightweight You Only Look Once (YOLO) v7 networks to cope with this problem. In this paper, we use a lightweight network, MobileNetv3, as the backbone network, in which a Residual edges CBAM block (RC-block) is designed to retain feature information while focusing on small-scale targets; finally, we use a bi-directional feature pyramid network (BiFPN) to perform feature fusion to further improve the detection accuracy. The experimental results show that the improved model reduces the Mean Average Precision (mAP) by only <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1630_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="35" /> </InlineMediaObject> <EquationSource Format="TEX">$0.7\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>0.7</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> compared with the traditional YOLOv7 model, but the detection speed is increased by <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1630_Article_IEq2.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="43" /> </InlineMediaObject> <EquationSource Format="TEX">$29.4\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>29.4</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>, and the model volume is reduced by <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2025_1630_Article_IEq3.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="43" /> </InlineMediaObject> <EquationSource Format="TEX">$29.9\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>29.9</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>, effectively improves the detection accuracy and speed of all kinds of defects on the surface of the bushings. The improved model was trained under the publicly available dataset NEU-DET, and the results showed the generalisability of the model.</p>

Lightweight YOLOv7 for bushing surface defects detection

期刊论文
发表日期：

2025年2月7日

查看全文

Wenjun Cheng,
Pengfei Zeng,
Yongping Hao

摘要

Bushings have a wide range of applications in industry. Once the surface of the bushing is defective, it will affect the assembly between bearings resulting in mechanical inefficiency. At present, due to the different target sizes of the different types of defects on the bushing surface, it is difficult to balance inspection accuracy and speed. This paper proposes lightweight You Only Look Once (YOLO) v7 networks to cope with this problem. In this paper, we use a lightweight network, MobileNetv3, as the backbone network, in which a Residual edges CBAM block (RC-block) is designed to retain feature information while focusing on small-scale targets; finally, we use a bi-directional feature pyramid network (BiFPN) to perform feature fusion to further improve the detection accuracy. The experimental results show that the improved model reduces the Mean Average Precision (mAP) by only $0.7\%$ $0.7 %$ compared with the traditional YOLOv7 model, but the detection speed is increased by $29.4\%$ $29.4 %$ , and the model volume is reduced by $29.9\%$ $29.9 %$ , effectively improves the detection accuracy and speed of all kinds of defects on the surface of the bushings. The improved model was trained under the publicly available dataset NEU-DET, and the results showed the generalisability of the model.