Abstract: Cutting tool condition is an important consideration in machining.Aiming at the on-line prediction of milling cutter wear,a deep learning based prediction model of milling cutter wear was established.The vibration sensor was used to collect the vibration signal and extract the feature vectors in time domain,frequency domain and time-frequency domain by wavelet denoising,fast Fourier transform,wavelet packet decomposition and other technologies.The correlation analysis was applied to select qualified feature vector as the noise samples of Stacking Sparse Denoising Auto-Encoder(SSDAE),then the selected feature vectors were treated with monotonous non-decreasing and smoothing by feature post-processing,and were used as the noise-free samples of SSDAE to train the network.Subsequently,the feature vectors after dimension reduction of SSDAE were taken as the input of Back Propagation Neural Network (BPNN),and the actual wear of milling cutter corresponding to these features was taken as the label to conduct fitting training for the Network.The trained model was verified experimentally,and the comparison between the normal test data set and the test data set with artificial noise input proved that the model established by this method had not only higher prediction accuracy but also higher robustness.

Key words: milling cutter wear, stacking sparse denoising auto-encoder, feature post-processing, robustness, back propagation neural network

摘要： 刀具磨损状态是机械加工过程中需考虑的重要因素之一。针对铣刀磨损的在线预测问题,建立了一种基于深度学习的铣刀磨损预测模型。首先,将采集到的铣刀切削时的振动信号进行小波去噪后,利用快速傅里叶变换和小波包分解等技术提取时域、频域及时频域上的特征参数,并根据相关性分析从中筛选出合格的特征参数合并为特征向量,以此作为堆叠稀疏去噪自动编码网络(SSDAE)的含噪样本。其次,利用特征后处理的方式对已经筛选出的特征参数进行单调不递减及平滑处理,并将其作为SSDAE的无噪样本来训练该网络。然后,将经过SSDAE降维后的特征向量作为多隐层反向传播神经网络(BPNN)的输入,以这些特征对应的实际铣刀的磨损量作为标签对该网络进行拟合训练。最后,对训练好的模型进行实验验证,通过测试数据集和人为加入噪声的测试数据集的对比,结果显示所提模型不仅具有较高的预测精度,还具有较高的鲁棒性。

关键词: 铣刀磨损, 堆叠稀疏去噪自动编码网络, 特征后处理, 鲁棒性, 反向传播神经网络