Proposals for Availability Prediction Methods for Peripheral Milling Machines

Abstract

Both domestic and global industrial manufacturing companies consistently strive to gain a competitive edge in their respective market segments. In addition to various other strategies, these companies achieve this advantage by harnessing technology and digital services to enhance not only the quality and performance of their manufacturing equipment, but also their availability. These enhancements are particularly evident in industrial processes, such as manufacturing that involves the utilization of milling machines. Rising energy and material costs, as well as overall resource utilization, are major drivers for companies to reduce manufacturing-related expenses by optimizing operational maintenance activities, minimizing unplanned downtimes, and ensuring that manufacturing equipment is available when needed. A vast amount of research in the areas of physics-based simulation models and machine learning (ML) has been performed to enhance the aforementioned measures, yet the combined usage of these approaches in real-life milling applications has remained relatively under-explored. This thesis and the publications referred to provide a thorough analysis of various individual approaches, as well as their hybrid combinations, to improve milling machine availability, while considering a strategic life cycle management standpoint, thus offering tools to improve overall equipment effectiveness (OEE).

First, this thesis focuses on presenting a physics-based simulation model that replicates the peripheral milling machine operational behavior in close approximation in various terms. The physics-based simulation model developed acts as a profound foundation in this thesis for reproducing peripheral milling machine behavior in normal and anomalistic terms and diagnosing abnormal states of operation. The forces affecting the milling spindle are further simulated and connected to a mass-spring-damper system in the simulation environment used to replicate vibrational behavior for dynamic wear prediction. Average cutting force, torque and material removal rate simulations are also time integrated and validated to indicate progressive wear over the temporal operation of the machine. Torque simulations are used to diagnose anomalistic behavior in the milling operation in case of a support roll failure.

Second, to creating a sophisticated physics-based simulation model architecture, this thesis provides a comprehensive analysis of the implementation of various dimensionality reduction methods based on machine data. The Pearson Correlation Coefficient (PCC) and Permutation Feature Importance (PFI) are implemented to explore and guide readers on how to utilize data-driven methods to discover meaningful correlations between machine-related excitations and the wear phenomena of the system. Furthermore, the Support Vector Regression algorithm is implemented and evaluated to predict vibration excitation created by the real acceleration sensor. The Long Short-Term Memory (LSTM) architecture of the Recurrent Neural Network (RNN) is used in regression analysis as well as for classification purposes to showcase its applicability in prediction-making for time- series data.

Finally, in addition to presenting individual tools and their applicability to predict wear-related behavior in peripheral milling machines, this thesis contributes by presenting two novel methods merging physics-based simulation model data and sensor data from real manufacturing processes. The first method, known as the Fused Data Prediction Model (FDPM), is designed to predict the remaining useful life (RUL) of milling machines. The second method, referred to as Simulation- Enhanced Anomaly Diagnostics (SEAD), focuses on generating high-quality synthetic data to enhance the classification capabilities of deep neural networks. They offer validated potential for higher system availability, and thus a way to gain a competitive advantage in the field.