【摘要】 在刀具磨损监测领域中，传统卷积神经网络难以选择合适的卷积核大小，循环神经网络容易发生梯度消失和梯度爆炸，为克服以上缺点，引入时间卷积网络构建在线监测模型对刀具磨损量进行监测。考虑到原始数据量过大且每次走刀过程所采集数据量不同，对数据进行降采样处理，获得了大小为(7,5000)的网络输入数据。通过一维卷积神经网络和时间卷积块的依次叠加，对数据进行特征提取，使用全连接网络将特征映射到刀具磨损值。最后，使用PHM大赛中铣刀磨损的数据验证了模型的效果。实验结果证明，基于时间卷积网络的刀具磨损在线监测模型具有较强的泛化能力，在验证集上均方误差和平均绝对误差分别仅为65.16与6.21,相较于隐马尔科夫、梯度提升树等模型具有较大的提升。更多还原

【Abstract】 In the field of tool wear monitoring, it is difficult for traditional convolutional neural network to select the appropriate convolutional kernel size, and the recurrent neural network is prone to gradient disappearance and gradient explosion.In order to overcome the above shortcomings, temporal convolutional network was introduced to build an online monitoring model to monitor tool wear.Considering the mass original data and the amount of data collected varies in each tool cutting process, the data was downsampled, and a final input data size of(7,5000) was obtained.Through the superposition of one-dimensional convolutional neural network and temporal convolutional block, the data feature was extracted, and the full connected network was used to map the features to the tool wear value.Finally, the effectiveness of the model was verified using the data of milling cutter wear in the PHM competition.The experimental results indicate that the tool wear online monitoring model based on temporal convolution network has strong generalization ability.The mean square error and the mean absolute error on the validation set is only 65.16 and 6.21 respectively, which suggest that there is a great improvement compared with hidden markov, gradient boosting trees and so on.更多还原