Title: Machine oil leakage detection based on an improved YOLOv4-tiny neural network

Authors: Yuehua Li; Sailin Wu; Jiahao Sun; Bin Hu

Addresses: College of Information Science and Technology, Nantong University, Nantong 226019, China ' College of Information Science and Technology, Nantong University, Nantong 226019, China ' College of Information Science and Technology, Nantong University, Nantong 226019, China ' College of Information Science and Technology, Nantong University, Nantong 226019, China

Abstract: Oil leakage is a typical recurrent problem in industrial machine faults. An oil leakage detection network based on the YOLOv4-tiny is proposed. Firstly, we introduce depthwise separable convolution in the feature extraction network to decrease the parameters number. Secondly, MobileNetV3 is used to increase the depth and strengthen feature extraction ability. Finally, a gradient harmonising mechanism in classification loss is adopted to optimise the foreground and background class imbalance. The sufficient experiments show that the improved algorithm has a better result in the data set, which contains 677 oil leakage images. The detection accuracy reaches 89.67%, which is 1.36% higher than the original algorithm. The recall rate reaches 94.04%, 2.3% higher. The model size is reduced to 13.9M, 38.2% lower, and the frame rate is 45.6 fps. The improved model meets real-time requirements, and is lighter and easier to deploy on the edge device.

Keywords: oil leakage; YOLOv4-tiny; deep learning; lightweight.

DOI: 10.1504/IJSNET.2024.136699

International Journal of Sensor Networks, 2024 Vol.44 No.2, pp.124 - 132

Received: 13 Sep 2023
Accepted: 17 Sep 2023
Published online: 16 Feb 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article