In today's manufacturing industry it is of great importance to know the condition of a tool as it wears out. This knowledge allows us to reduce possible errors and failures in the manufacturing process. This work focuses on classifying the level of wear of a machine tool, allowing us to know its current state using measurements of forces and accelerations. First, tool wear levels are obtained by unsupervised learning (through the Gaussian mixture model). Then a convolutional neural network trained directly using the measured time series predicts the level of tool wear. After careful selection of the optimal window length, optimization algorithm, and number of epochs, wear levels were predicted with accuracy greater than 85%.