基于长短期记忆神经网络的刀具磨损状态监测

Tool wear state monitoring based on long-term and short-term memory neural network

刀具磨损状况的实时检测是目前机床加工状态监测的难点,而对刀具的振动信号分析的常用方法是利用神经网络模型来判断刀具磨损状态。为解决循环神经网络(RNN)模型训练过程中梯度容易消亡的现象,提出基于长短期记忆神经网络的刀具磨损状态在线监测。刀具在进行切削加工时,首先通过加速度传感器采集刀具振动信号,然后对振动信号小波包变换进行分解是让信号通过不同的滤波器进行有条件的选择,由此形成不同的能量值,用作为长短期记忆神经网络的特征输入,从而诊断出刀具磨损状态的3种状态故障;最后利用长短期记忆神经网络模型对处理时间序列的数据有比较好的效果,它可以捕捉长期的依赖关系和非线性动态变化。此外,通过与多层(BP)神经网络和(BP)神经网络故障诊断方法进行比较,结果表明,LSTM网络对刀具磨损状态在线监测更加有效。

The real-time detection of tool wear condition is a difficult point in the current machining condition monitoring of the machine tool.The common method for analyzing the vibration signal of the tool is to use the neural network model to judge the tool wear state.In order to solve the phenomenon that the gradient is easy to die during the training of the cyclic neural network(RNN)model,this paper proposes an online monitoring method for tool wear state based on long-term and short-term memory neural network.When the tool is cutting,the tool vibration signal is first collected by the acceleration sensor,and then the vibration signal wavelet packet transform is decomposed,and the signal is subjected to conditional selection through different filters,and then different energy values are formed.It is used as a feature input of long-term and short-term memory neural network to diagnose three state faults of tool wear state.Finally,the long-term and short-term memory neural network model is used to process time series data,which can capture long-term dependencies.And nonlinear dynamic changes.In addition,the comparison of multi-layer(BP)neural network and(BP)neural network fault diagnosis methods shows that the LSTM network is more effective for on-line monitoring of tool wear status.