Analysis of the performance of LSTM-DNN models with the consideration of signal complexity in milling processes

This paper investigates the impacts of signal complexity on the performance of Deep Learning (DL) models in milling processes, with a particular focus on cutting force due to its critical role in system diagnostics and process monitoring. To compare the Cutter-Workpiece-Engagement (CWE) time-dependent and time-independent processes, signal characterisation was firstly conducted using various feature indicators to quantify signal complexity under different conditions. Multiple DL models were then evaluated through controlled cutting trials to assess how variations in signal complexity affect model performance. The signal analysis revealed that the 14 selected indicators, along with the Recurrent Neural Network-Deep Neural Network (RNN-DNN) models, demonstrated monotonicity as the number of embedded 'active frequencies' increased during the "Partial Engagement" or CWE time-dependent process. However, as the depth of the DL models increased, this monotonicity effect diminished. Furthermore, clearer dependencies were observed in the analysis of harmonic prediction performance, particularly with a more pronounced impact on accuracy and uncertainty in the 'Partial Engagement' stage compared to the 'Full Engagement' stage. In conclusion, the complex performance testing of DL models reveals a significant relationship between signal complexity and model performance. This study underscores the importance of incorporating signal complexity analysis as a critical component in applying DL technology within machining processes, as it provides valuable insights into model performance.