改进infoGAN和QPSO-VGG16的小样本条件下电机轴承故障诊断方法

打开文本图片集

DO1：10.15938/j. emc.2025.05.016

中图分类号：TM307;TM343 文献标志码：A 文章编号：1007-449X（2025）05-0167-12

Fault diagnosis method for motor bearings under small sample conditions using improved infoGAN and QPSO-VGG16

LIU Hang， ZHANG Dechun， LIU Zhijian， HE Wei， TAO Yunxu， MENG Xinyu（Department of Electrical Engineering，Kunming Universityof Science and Technology， Kunming 6505Oo，China）

Abstract：In response to scarcity of motor bearing fault data compared to normal data，an improved fault diagnosis method using infoGAN and QPSO-VGG16 was proposed. Initially，the high dimensional fault signalswere transformed into corresponding two-dimensional time-frequency image using the continuous wavelet transform （CWT），creating the raw image dataset. Subsequently，a data augmentation model based on conditional information maximizing generative adversarial nets （cinfoGAN） was established， which generates all categories of fault data within a unified framework.This method achieves enhancements in data augmentation quality and eficiency. Furthermore，a fault diagnosis model based on the VGG16 network was constructed.The VGG16 network was trained alternately on both the raw and augmented image datasets. During the training process， an improved quantum-behaved particle swarm optimization（QPSO） algorithm was employed to jointly optimize the learning rates for both datasets，ensuring that the VGG16 network achieves optimal performance.Numerical experimental results conducted on real motor bearing vibration signals indicate that converting vibration signals into images can fully leverage the feature extraction capability of the VGG16 model for image data. Additionally，data augmentation and alternate training methods can sequentially improve accuracy of fault diagnosis by 2.6% and 4.5% ，respectively.

Keywords： motor bearings； fault diagnosis； continuous wavelet transform； generative adversarial nets ;Visual Geometry Group 16

0引言

电机的故障中约 40% 来源于滚动轴承。（剩余17060字）