摘要
针对电机轴承故障诊断过程中现有的特征提取与高效分类问题,提出并验证了一种融合短时傅里叶变换(STFT)、深度残差网络(ResNet18)与K折交叉验证的电机轴承故障诊断方法。该方法首先采用STFT将电机轴承的一维振动信号转换为二维时频图;然后,对转换后的时频图使用ResNet18进行迁移学习训练,通过五折交叉验证策略评估其在电机轴承故障诊断中的表现,在十分类故障诊断训练中取得98.96%的平均准确率;最后,在相同条件下将ResNet18替换为经典的深度学习模型AlexNet重新训练,其在训练集上的平均准确率为97.05%,相较于ResNet18,AlexNet在处理经STFT转换的时频图像时,学习能力明显不足。综合分析,基于STFT、ResNet18和五折交叉验证融合的电机轴承故障诊断方法具有较高诊断精度与鲁棒性。
Aiming at the existing problems of feature extraction and efficient classification in the process of motor bearing fault diagnosis,a motor bearing fault diagnosis method integrating short-time Fourier transform(STFT),deep residual network(ResNet18)and K-fold cross-validation is proposed and validated.The method firstly uses STFT to convert the one-dimensional vibration signals of motor bearings into two-dimensional time-frequency maps.Second,the converted time-frequency maps are trained by migration learning using ResNet18,and its performance in motor bearing fault diagnosis is evaluated by a five-fold cross-validation strategy,which achieves an average accuracy of 98.96%in ten classes of fault diagnosis training.Finally,ResNet18 was retrained by replacing it with the classical deep learning model AlexNet under the same conditions,and its average accuracy on the training set was 97.05%.Compared to ResNet18,AlexNet is significantly less capable of learning when processing time-frequency images transformed by STFT.In the comprehensive analysis,the motor bearing fault diagnosis method based on the fusion of STFT,ResNet18 and five-fold cross-validation has high diagnostic accuracy and robustness.
作者
胡山
朱向华
HU Shan;ZHU Xianghua(School of Automation and Electrical Engineering,Tianjin University of Technology and Education,Tianjin 300222,China;Key Laboratory of Information Sensing&Intelligent Control,Tianjin 300222,China)
出处
《自动化与仪表》
2025年第4期40-45,共6页
Automation & Instrumentation
基金
天津市科技计划项目(23YDTPJC00320)。
关键词
电机轴承故障诊断
深度学习
K折交叉验证
短时傅里叶变换
迁移学习
motor bearing fault diagnosis
deep learning
K-fold cross verificatio
short-time Fourier transform(STFT)
transfer learning
