摘要
针对部分齿轮的运行环境复杂,导致采集的样本数据不够的问题,提出了一种基于Transformer和卷积神经网络(CNN)的迁移学习齿轮故障诊断方法。首先,采用高斯滤波对原始振动信号进行了预处理,使信号变得平滑,降低了噪声信号的干扰;再将信号处理成带有位置信息的补丁序列以作为Transformer的输入,并增强了Transformer特征提取的能力,提高了诊断精度;然后,将信号输入到CNN继续提取特征信息,在模型中添加了一个残差块以防止网络退化;接着,划分了实验室采集的齿轮数据集和东南大学齿轮箱数据集的源域和目标域,采用了源域数据预训练模型,选择了每种类型的齿轮各100个样本为目标域;最后,以不同数据集为源域共进行了4组10次重复实验,测试了模型的准确率。研究结果表明:以不同数据集为源域的4组10次迁移实验的齿轮故障诊断准确率较高,均在90%以上,最高准确率可达100%;与其他不含Transformer的卷积神经网络、多尺度卷积神经网络和二维卷积神经网络相比,Transformer-CNN的齿轮故障诊断平均准确率更高,其平均准确率可达到99.64%。因此,基于Transformer-CNN的迁移学习方法能在小样本下诊断齿轮的故障。
Aiming at complex operating environment of partial gears which leading to the lack of sample data,a method for diagnosis transfer learning gear fault based on Transformer and convolutional neural network(CNN)was present.First,Gaussian filter was employed to preprocess the original vibration signal.It was able to smooth the signal and reduce interference from noisy signals.Then,the signal as an input signal to the Transformer was transformed into a patch sequence with position information.It enhanced the Transformer s feature extraction capabilities,and it improved model s diagnostic accuracy.Besides,the Transformer output sequence was input into a one-dimensional CNN to keep extracting fault information,and a residual block was added to the model to prevent network degradation.What s more,the gear dataset collected in the laboratory and the gearbox dataset of Southeast University were divided into source and target domains,the model with source domain data was pretrained,and 100 samples of each type of gear were selected as the target domain.Finally,four sets of ten replicates were performed with different datasets as the source and target domains in order to test the accuracy of the model.The experimental results show that the accuracy of gear fault diagnosis with the method of Transformer-CNN transfer learning was more than 90%.Among them,the highest fault diagnosis accuracy can reach 100%.Transformer-CNN also compares the gear fault diagnosis accuracy of other convolutional neural networks,multi-scale convolutional neural networks and two-dimensional convolutional neural networks without Transformer,with an average accuracy of 99.64%,which is higher than that of the above networks.Therefore,the transfer learning method based on Transformer-CNN is able to diagnose gear faults under small samples.
作者
闫绘宇
张超
YAN Huiyu;ZHANG Chao(School of Mechanical Engineering,Inner Mongolia University of Science&Technology,Baotou 014017,China;Inner Mongolia Autonomous Region Key Laboratory of Intelligent Diagnosis and Control for Electromechanical System,Baotou 014010,China)
出处
《机电工程》
CAS
北大核心
2024年第3期409-417,共9页
Journal of Mechanical & Electrical Engineering
基金
国家自然科学基金资助项目(51965052)。
