Abstract
Machining is one of the most prevalent manufacturing processes for producing precision engineering components used in aerospace, automotive, and other industries. The demand for higher performance has led to using materials with higher strength and hardness often at high temperatures. These properties can pose significant challenges during machining and cause significant tool wear leading to short tool life and high manufacturing costs. Any damage caused by tool wear during machining can result in high scrap costs or lead to part failure in service. This is particularly important for safety critical components used in aerospace applications such as parts made from Inconel 718. This has led to a common practice in industry to replace tools prematurely to prevent tool wear induced damage to the workpiece. Traditionally, determining when a tool should be replaced is defined by experience and manual measurement or relying on statistical process control leading to high tool consumption.In recent years, research directions have shifted towards implementing sensors and collecting signals for tool condition monitoring. The application of machine learning has led to significant developments in this area for detecting specific features in sensor signals relating to the condition of cutting tools. However, machining is a fast and dynamic process and if a prediction is made in real time, the machining would have moved past this prediction.
The research vision in this thesis is to use deep learning techniques in combination with sensor signals from the machining process to predict what the signal may look like a few time steps ahead and then predict the tool status in the future. This solution involves three distinct phases; signal conditioning followed by a future signal prediction stage, and then finally a classification stage. For the signal conditioning phase, bending moment signals obtained from a sensory tool holder together with continuous wavelet transform was used. These allowed the time and frequency components of the signals to be represented and presented as an image. The future signal prediction stage was based on convolutional long short-term memory networks, which in recent studies have been used for sequence prediction problems in areas such as weather forecasting and video frame prediction. These models used scalogram image sequences to predict future frames. Finally, the classification stage was achieved using convolutional neural networks, a widely used image classification deep learning model, where the future frame predictions can be classified according to the level of cutting tool wear.
The novel contribution of this research is the design, realisation and testing of the tool wear forecasting and prediction model, combining both future prediction and classification deep learning models. The results for each component of this model, and the combination demonstrates that it is possible to predict the future signal and classify the predicted frames. The investigations show that the implementation of this methodology in a real time tool wear condition monitoring system is highly promising.
| Date of Award | 29 Mar 2023 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Alborz Shokrani Chaharsooghi (Supervisor), Evros Loukaides (Supervisor) & Stephen Newman (Supervisor) |