TCM in milling processes based on attention mechanism-combined long short-term memory using a sound sensor under different working conditions

Tool condition monitoring (TCM) is essential for the milling process to ensure machining quality, and several deep learning (DL)-based methods have been proposed to obtain good regression accuracy for TCM, such as RNN and LSTM. Unfortunately, the performances of these DL-based methods are not good enough under different working conditions. A novel method combining attention mechanism and long short-term memory (LSTM) is proposed. Firstly, sound time series signal obtained from a machining process is converted into several feature sequences, and these feature sequences are input to the attention mechanism-combined LSTM (AMLSTM) to train the weight of the feature sequences. Finally, the trained AMLSTM model with the optimal weight of the feature can be used to estimate the tool wear value. The application of the proposed method in milling TCM experiments shows that the AM-LSTM-based method is significantly better than SVR-based, RNN-based, and LSTM-based methods under different working conditions. Moreover, skewness and kurtosis are two important features for TCM.