基于深度残差收缩网络的刀具故障自感知系统研究
Research on tool fault self-sensing system based on deep residual shrinkage network
李嘉豪 1张兵 2朱建阳1
作者信息
- 1. 武汉科技大学 机械自动化学院,湖北 武汉 430081
- 2. 东莞理工学院 机械工程学院,广东 东莞 523808
- 折叠
摘要
传统机器学习在特征提取方面需要手动进行特征提取工程,存在局限性和维度灾难问题,同时处理非线性关系能力不足且精度有限.针对这些问题,基于深度残差收缩网络(DRSN),提出一种融合了注意力机制(MCA-a)及空间细化特征(SRU)和通道细化特征(CRU)的改进深度残差收缩网络(DRSN-IAM),有效解决传统神经网络由于网络过深而导致的梯度爆炸、梯度消失和特征提取能力不足等问题.使用公开数据集(即美国纽约预测与健康管理学会(PHM)2010 年高速数控机床刀具健康预测竞赛的开放数据)进行验证,结果表明所提出的算法模型分类精度达到 99.2%,比现有的DRSN错误率降低了 40.1%.
Abstract
With the application of next-generation artificial intelligence technologies,CNC machine tools are encountering new development opportunities,moving towards high precision,high efficiency,automation,and intelligence.Traditional machine learning faces challenges in feature extraction due to manual feature engineering requirements,limitations,the curse of dimensionality,insufficient handling of non-linear relationships,and low accuracy.To address these issues,proposes an improved Deep Residual Shrinkage Network(DRSN)fused with Multidimensional Cooperative Attention-alter(MCA-a),Spatial Recalibration Unit(SRU),and Channel Recalibration Unit(CRU),termed Deep Residual Shrinkage Improvement of Attention Mechanism(DRSN-IAM).This effectively tackles problems such as gradient explosion and vanishing gradients caused by deep networks in traditional neural networks.By optimizing the model,it reduces issues such as insufficient feature extraction capabilities due to the soft threshold function.Validation using the publicly available dataset,specifically the open data from the 2010 PHM High-Speed CNC Machine Tool Health Prediction Competition by the American Society of Mechanical Engineers(ASME),achieves a classification accuracy of 99.2%.Compared to classical neural networks,the proposed improved Deep Residual Shrinkage Network(DRSN-IAM)demonstrates superior classification performance.
关键词
刀具故障诊断/人工智能/深度残差收缩网络/注意力机制Key words
tool fault diagnosis/artificial intelligence/deep residual shrinkage network/attention mechanism引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据