Prediction of tool-path and cycle time errors of CNC machining using a transformer-based deep learning model

The prediction of tool-path and cycle time errors of computer numerical control (CNC) machining is important in manufacturing field. In this study, these errors are considered as sequential data corresponding to the lines of G-code, i.e., the error prediction is a sequence-to-sequence task, and a Transformer-based deep learning model was constructed to predict the tool-path and cycle time errors simultaneously, taking the positional increments and feed rate as the input. The training datasets were created by running a randomly generated G-code on a real CNC machine tool and subsequently tracking the actual errors, where the motion and feed rate are stochastically defined. The model is validated by conducting the error prediction on six sets of G-code. In addition, the comparison with a Bi-LSTM-based deep learning model is conducted. The results show that the prediction accuracy of the Transformer model can achieve the mean prediction accuracies of ±1 μm and ±1 ms in tool-path and cycle time error predictions, respectively. Furthermore, the Transformer model has a smaller prediction uncertainty than the Bi-LSTM model. Compared with conventional methods only analyzing tool-path errors or cycle-time errors, the utilized model has an advantage in its ability to simultaneously predict tool-path errors and cycle time errors.