DOI:10.3969/j.issn.1672-3872.2023.23.035
作者:寇智铭,和飞翔,李萌瑞(长安大学研究生院,陕西 西安 710064)
摘 要:【目的】针对农机使用环境恶劣、欠缺后期维护等问题,研究农机轴承故障诊断新方法,克服卷积神经网络难以全局提取特征的不足,提高故障诊断准确率以及农机使用的安全性和稳定性。【方法】研究团队引入了二维改进Transformer网络,利用未经优化的短时傅里叶变换生成的二维图像,并结合深度学习模型,进行了农机轴承故障诊断。研究团队还基于凯斯西储大学的轴承故障数据集进行了实验,从而验证该故障诊断方法的有效性。【结果】在训练30轮后即可在故障类型和故障程度的诊断中达到99.9%的准确率,高于常用的卷积神经网络。【结论】该故障诊断方法流程简单,在负载情况下具有可靠性、优越性,减少了对专家系统的依赖,为农机轴承故障诊断提供了更便捷可靠的方法。
关键词:Transformer网络;多头注意力机制;轴承;故障诊断
Fault Diagnosis of Agricultural Machinery Bearing Based on Two-Dimensional Improved Transformer Network
Kou Zhiming, He Feixiang, Li Mengrui (Chang’an University Graduate School, Shaanxi Xi’an 710064)
Abstract: [Objective] To study a new method for fault diagnosis of agricultural machinery bearings in response to the harsh operating environment and lack of post maintenance, in order to overcome the difficulty of global feature extraction by convolutional neural networks, improve the accuracy of fault diagnosis, and enhance the safety and stability of agricultural machinery use. [Method] The research team introduced a two-dimensional improved Transformer network, which used unoptimized short-term Fourier transform to generate two-dimensional images, and combined with deep learning models to diagnose agricultural machinery bearing faults. The research team also conducted experiments based on the bearing fault dataset of Case Western Reserve University to verify the effectiveness of the fault diagnosis method. [Result] After 30 rounds of training, the accuracy in diagnosing fault types and degrees can reach 99.9%, which is higher than commonly used convolutional neural networks. [Conclusion] This fault diagnosis method has a simple process, reliability and superiority under load conditions, reduces the dependence on expert systems, and provides a more convenient and reliable method for agricultural machinery bearing fault diagnosis.
Keywords: Transformer network; multi-head attention mechanism; bearing; fault diagnosis
引文信息:[1]寇智铭,和飞翔,李萌瑞.基于二维改进Transformer网络的农机轴承故障诊断[J].南方农机,2023,54(23):136-138+160.
查看全文请下载PDF文件↓