Abstract: Aiming at the existing tool wear prediction methods which caused the problems of poor prediction accuracy due to lack optimization algorithms and inadequate network structure. A tool wear prediction model with the combination of improved SABO(ISABO) and improved BiLSTM(IBiLSTM) network(ISABO-IBiLSTM model) was proposed. Firstly, the acceleration vibration signal and force signal data were preprocessed by truncation method, Hampel filtering method, and improved complete ensemble empirical mode decomposition with adaptive noise(ICEEMDAN)-improved wavelet thresholding noise reduction method. Then, the time-domain, frequency-domain, and time-frequency-domain features of the preprocessed signal data were extracted, and the features are screened by Spearman and maximum mutual information correlation coefficient to construct the inputs of the model. Finally, the ISABO algorithm was used to perform parameter optimization of the IBiLSTM network, and based on the obtained optimized parameters, the network was trained to achieve wear prediction. The experimental data analysis results show that the proposed ISABO-IBiLSTM model has a prediction accuracy of 98.49% to 98.83% for tool wear, which is significantly improved compared to BiLSTM, IBiLSTM, and improved convolutional neural networks(ICNN)-BiLSTM models.

Key words: tool wear prediction, subtraction-average-based optimizer(SABO) algorithm, bidirectional long-short time memory(BiLSTM) network, signal processing, deep learning

摘要： 针对现有的刀具磨损预测方法因为缺少优化算法及网络结构不完善而导致预测精度不高的问题，提出了一种将改进的减法优化器(SABO)算法和改进的双向长短时记忆(BiLSTM)网络相结合的刀具磨损状态预测模型(ISABO-IBiLSTM模型)。首先，采用截断法、Hampel滤波法、改进的完全自适应噪声集合经验模态分解(ICEEMDAN)-改进的小波阈值降噪法对加速度振动信号与力信号数据进行预处理。然后，提取预处理后的信号数据的时域、频域、时频域特征，并通过斯皮尔曼和最大互信息相关系数筛选特征，构建模型的输入。最后，利用改进的SABO算法对改进后的BiLSTM网络进行参数寻优，基于所得到的优化参数训练网络实现磨损预测。实验数据分析结果表明，所提出的ISABO-IBiLSTM模型对刀具磨损量的预测精度为98.49％～98.83％，较BiLSTM模型、改进的BiLSTM模型、改进的卷积神经网络(ICNN)-BiLSTM模型有了较大的提高。

关键词: 刀具磨损预测, 减法优化器算法, 双向长短时记忆网络, 信号处理, 深度学习