Comparison of CNN and CNN-LSTM Architectures for Tool Wear Estimation

Modern manufacturing needs to guarantee product quality and reduce operating costs. These can be achieved through the use of analytical tools, which depend on the collection of large amounts of data, in this particular case in the form of time series. During the last few years, various conventional and neural network-based methods have shown great promise in problems related to estimating milling cutter wear. Among neural networks, recurrent networks are especially promising due to the memory mechanism they use. In the present work, a comparison is made between a CNN network and a CNN-LSTM network. Both networks extract information directly from the time series of a widely used database. Unlike similar works in the existing literature, two simple preprocessing techniques are used: to remove the tendency of the time series and to equalize the initial values of the tool wear. Additionally, Bayesian optimization of hyperparameters is used. Mean square errors are obtained that are consistently around 10, results equivalent to the state of the art.