Predictive modelling of key performance indicators for a data-driven digital twin of CNC machining processes

Abstract

This thesis introduces a novel Digital Twin architecture for modelling CNC machining processes for better decision-making by predicting key performance indicators (KPIs) with data-driven modelling. The proposed data-driven Digital Twin framework considers human-in-loop decision-makers such as the process planner and the machine operator. The research mainly focuses on methodologies of building data-driven predictive models of KPIs, such as machining quality, energy consumption and machining time, for the proposed Digital Twin architecture. This work primarily uses deep neural networks to prepare respective predictive models at both stages of the CNC machining process from historical data. Three methods of delivering prediction models of KPIs are explained in this work: stage-wise, two-step and hybrid physicsdata-driven methods. In the stage-wise modelling method, the models used in the process planning and machining stages of CNC machining are built using only the data available in each stage of the machining operation. Data-driven models encounter difficulties when an input parameter changes during the machining process. In the five-axis machining processes, the feed rate fluctuates from the planned values due to the delay in synchronising all the axis motors of the five-axis machining process for the shorter cut lengths. A two-step data modelling approach is proposed to account for this. The expected feed rate is predicted first and used as one of the inputs for the energy prediction model. This method improves the accuracy of prediction at the process planning stage. Incorporation of the physics information into the data-driven model results in better accuracy. These methods are referred to as physics-guided data-driven methods. One of the physics-guided approaches is the hybrid physics-data-driven method. This method develops predictive models by featuring physics-based models as inputs to the data-driven models with other feature inputs. Predictive models for energy consumption and surface roughness in a three-axis CNC milling operation are trained with the hybrid physics-data-driven method. Comparison with respective physics-based models and data-driven models shows that the hybrid models perform better than these models. Machining experiments with three-axis and five-axis CNC milling are conducted to generate data for the historical database. The predictive models built using this data form the core of the proposed Digital Twin framework. A graphical user interface has been developed to integrate the predictive models, the CAM planning software and the CNC machine controller for seamless exchange of data between the three. Multiobjective optimisation of the machining parameters and continuous learning are identified as future work for realising the proposed Digital Twin.