Particle Learning and Gated Recurrent Neural Network for Online Tool Wear Diagnosis and Prognosis

摘要

Automated tool condition monitoring is critical in intelligent manufacturing to improve both productivity and sustainability of manufacturing operations. Estimation of tool wear in real-time for critical machining operations can improve part quality and reduce scrap rates. The motivation of this work is to study two approaches, which aim to provide an online diagnosis and prognosis of machine tool wear conditions using indirect measurements. This work covers four aspects within the approach: 1) Diagnosis of the tool wear itself; 2) Prognosis estimates of the tool wear ahead in time; 3) Mode of collecting indirect measurements from the process, and finally, 4) A method by which we continuously update in real-time tool wear estimates during a machining operation.;The first proposed approach is a probabilistic method based on Particle Learning by building a linear system transition function whose parameters are updated by online in-process observations of the machining process. By applying Particle Learning (PL), the method helps to avoid developing the closed form formulation for a specific tool wear model. It increases the robustness of the algorithm and reduces the time complexity of the computation. Our first approach assumes linearity and a Markovian process, which may not always hold for broader applications. Our second approach is based on Recurrent Neural Networks (RNN) for the online diagnosis and prognosis for cutting tool wear. It avoids the need to build an analytical model for specific tool wear model, and aims to capture the long term dependencies.;Capturing both long-term and short term memories through Gated Recurrent Units distinguishes our work from other RNNs developed by the community. Without increasing the complexity of the Neural Networks, our approach can realize multi-step ahead tool wear prediction and forecasting Remaining Useful Life (RUL). Both methods were tested experimentally to validate the diagnosis (online estimation), arbitrary multiple-step ahead prediction and Remaining Useful Life capability of our approach.