摘要
在实际工程应用中,进给轴从健康到故障时间跨度长、运行工况复杂、故障数据获取成本高,导致故障数据与健康数据存在严重的不平衡。而传统数据驱动健康监测方法往往需要大量标签数据,且监测结果依赖于标签的数量和准确性。如何在有限数据下,进行健康监测是目前面临的一大挑战。针对这一问题提出了一种基于优化卷积自编码器的机床进给轴健康状态监测方法,首先采用小波包对进给轴振动信号与功率信号进行去噪重构,随后对降噪后的振动信号与功率信号进行时域、频域特征提取形成振动功率多源数据集,之后搭建一种基于卷积自编码器(CAE)与双向长短时记忆网络(BiLSTM)相结合的进给轴健康监测网络,同时在网络中融合残差网络(Res)和注意力模块(SENet)提高模型收敛能力与监测准确性。试验表明所提模型可以仅采用健康数据进行训练,实现进给轴健康状态监测,健康状态监测准确率可达97.7%,优于传统CAE模型。
In practical engineering,the feed axis has a long healthy-to-faulty transition period,complex operating conditions,and high fault data acquisition costs,leading to an imbalance between fault and healthy data.Traditional data-driven methods require extensive labeled data,and their performance depends on label quantity and accuracy.How to monitor health with limited data is a major challenge.To address this,a machine tool feed axis health monitoring method based on an optimized convolutional autoencoder is proposed.First,wavelet packet denoising and reconstruction is applied to vibration and power signals.Time-frequency features are then extracted to form a multi-source dataset.A feed axis health monitoring network is established,combining convolutional autoencoder(CAE)and bidirectional LSTM,with residual and attention mechanisms to improve convergence and accuracy.Experiments show the proposed model can achieve 97.7%health monitoring accuracy using only healthy data,outperforming traditional CAE.
作者
吴楚杰
崔益铭
马骋
王强
赵雷鸣
刘阔
WU Chujie;CUI Yiming;MA Cheng;WANG Qiang;ZHAO Leiming;LIU Kuo(School of Mechanical Engineering,Dalian University of Technology,Dalian 116024,China;Dalian Company,Genertec Machine Tool Engineering Research Institute Co.,Ltd.,Dalian 116620,China;不详)
出处
《组合机床与自动化加工技术》
北大核心
2025年第5期1-6,共6页
Modular Machine Tool & Automatic Manufacturing Technique
基金
国家自然科学基金企业创新发展联合基金重点项目(U22B2085)
通用技术集团科技重大专项项目(GTZD-2022-017)资助。
关键词
残差网络
注意力机制
双向长短期记忆网络
卷积自编码器
进给轴
健康状态监测
residual network
attention mechanism
bidirectional long short-term memory
convolutional autoencoder
feed axis
health status monitoring
