摘要
针对传统智能监测方法中信号特征提取繁琐,以及铣削过程中刀具磨损对工件质量和生产效率的影响等问题,提出一种基于改进GRU神经网络的刀具磨损状态预测方法(BiGRU-1DCNN-CBAM)。利用统计量、时域分析、频域分析和小波变换等方法提取刀具信号的均值、峰度和功率谱密度等24个特征量,将多模态时间序列数据转换为刀具特征的时间序列图像;引入卷积神经网络挖掘信号数据的深层特征,结合卷积注意力机制模块增强模型对振动和切削力信号特征图的捕捉能力,将特征层展平拼接后输入BiGRU捕获长时依赖,通过全连接层回归刀具磨损量,实现数控加工过程中刀具磨损状态的剩余寿命预测。实验结果表明,该模型在PHM2010数据集上的均方根误差(RMSE)和平均绝对误差(MAE)分别为2.17μm和1.29μm,相比Bayesian-MCMC-Prognostics,SBiLSTM,RIME-CNN-SVM,MobilenetV3,TDConvLSTM,ISABO-IBiLSTM,IWOA-IECA-BiLSTM,LSTM-CNN-CBAM等模型,提升分别超过40.5%和52.1%,相比同类模型耗时至少减少2.8%。本文提出的模型能够有效表征刀具磨损程度,减小预测误差,取得了较好的预测效果。
To address the challenges of cumbersome signal feature extraction in traditional intelligent monitoring methods and the adverse effects of tool wear on workpiece quality and production efficiency during milling,a tool wear state prediction method based on an improved GRU neural network(BiGRU-1DCNN-CBAM)is proposed.Using statistical methods,time‑domain analysis,frequency‑domain analysis,and wavelet transform,24 feature parameters of the tool signals-such as mean,kurtosis,and power spectral density-are extracted,and the resulting multimodal time‑series data are converted into time‑series images of tool features.A convolutional neural network(CNN)is then introduced to mine deep features of the signal data,and a convolutional block attention module(CBAM)is integrated to enhance the capability of the model to capture feature maps of vibration and cutting force signals.After flattening and concatenating the feature layers,the fused features are fed into a bidirectional GRU(BiGRU)to capture longterm dependencies,and the tool wear amount is predicted through a fully connected layer,thereby enabling remaining useful life prediction of the tool wear state during CNC machining.Experimental results on the PHM2010 dataset show that the RMSE and MAE of the proposed model are 2.17μm and 1.29μm,respectively.Compared with Bayesian-MCMC-Prognostics,SBiLSTM,RIME-CNN-SVM,MobileNetV3,TDConvLSTM,ISABO-IBiLSTM,IWOA-IECA-BiLSTM,and LSTM-CNN-CBAM models,the prediction accuracy in terms of RMSE and MAE is improved by more than 40.5%and 52.1%,respectively,while the time consumption is reduced by at least 2.8%relative to similar models.These results demonstrate that the proposed model can effectively characterize tool wear,reduce prediction errors,and achieve superior prediction performance.
作者
巢渊
张俊杰
谈齐峰
张宜珺
戴国洪
夏志杰
张志胜
CHAO Yuan;ZHANG Junjie;TAN Qifeng;ZHANG Yijun;DAI Guohong;XIA Zhijie;ZHANG Zhisheng(School of Mechanical Engineering,Jiangsu University of Technology,Changzhou 213001,China;School of Mechanical Engineering,Southeast University,Nanjing 211189,China;Jiangsu Nangao High-end CNC Machine Tools and Complete Equipment Manufacturing Industry Innovation Center,Nanjing 211189,China)
出处
《光学精密工程》
北大核心
2025年第23期3765-3783,共19页
Optics and Precision Engineering
基金
国家自然科学基金资助项目(No.51905235)
江苏省自然科学基金资助项目(No.BK20191037)
江苏高校“青蓝工程”项目(No.KYQ23002)。
