A lightweight and efficient deep learning model for wear particle classification in lubricating oil

Purpose This study aims to develop a lightweight and efficient classification model for wear particle analysis in lubricating oil, a crucial task in machinery condition monitoring and fault diagnosis. A key challenge is the severe class imbalance among particle categories, where conventional models often fail to recognize minority classes. Design/methodology/approach A lightweight classification model, termed MobileNeSt, is proposed by integrating depthwise separable convolution, Split Attention and the H-Swish activation function within an efficient convolutional architecture. To alleviate class imbalance, a weighted loss function is incorporated to enhance the recognition capability for underrepresented wear particle categories. Findings Experimental results demonstrate the superior performance of MobileNeSt compared with representative models such as ResNet, DenseNet and MobileNet, achieving notable gains for minority classes (Fatigue and Sliding). Ablation studies further confirm the effectiveness of its components and validate its efficiency. MobileNeSt achieved 93.54% accuracy with only 10.6 M parameters, 0.68 GFLOPs and an inference speed of 255.52 FPS, striking an excellent balance between accuracy and efficiency. Originality/value This study proposes a lightweight wear particle classification framework that effectively mitigates class imbalance while satisfying practical requirements for lightweight design and fast inference.