Simulation Tools for Automated and Robust Manufacturing

Lead Research Organisation: University of Sheffield
Department Name: Mathematics and Statistics


The aim of this project is to use statistical methods to develop "green button" manufacturing processes: processes that can be run without a human operator, and can respond to unpredictable variations in the properties of the materials that are being machined. We will be focussing on "high value, low volume" manufacturing: manufacturing relatively small numbers of very expensive components, where it is costly to have to scrap a component because of a fault in the machining process. We will work on a case study: machining the landing gear of an aircraft, which we will use to develop methods that can be applied more generally. The first step will be to build a computer model of the machining process. Given the computer model, we can experiment with different parameters of the machining process such as the speed at which the metal is cut, and the path that the cutting tool takes through the metal. In theory, we could then search for the best choice of parameters, such that the component is machined in the shortest time and is least likely to be defective. However, the properties of the metal to be cut will vary from item to item, so what is best for one item may not be best for another. We can't measure all the relevant properties, so we need to first assess how much variability we are likely to see, and then find parameter settings that best able to handle this variability without producing faulty items.

Once we have determined the best parameter settings, we will then run a small number of machine cutting tests at different choices of machine cutting parameters. During these tests, we will take high quality but expensive measurements, telling us for example, the temperatures and forces exerted on the cutting tools. This information will tell us whether the process is operating satisfactorily, or whether there is a risk of tool damage and possibly a faulty machined component. We will also take lower quality, cheaper, sensor measurements, of the sort that would be available during the manufacturing process in the factory. We will study the relationship between all the variables that we have measured, so that we can construct a simulation model of the entire manufacturing process. (We can also make corrections to the computer model predictions, by inspecting how well the computer model predicts the cutting test outcomes). We can then use the simulation model to explore different strategies for modifying the process mid-production, in response to the cheaper sensor data, to avoid faults (eg "reduce the cutting speed by 10%" if a sensor reports vibration 5% above average"). It will be cheaper and faster to design the automated process using the simulation model, rather than conducting more expensive cutting tests.

The end product will be a manufacturing process that can run efficiently without a human operator, making adjustments as the sensor data are observed, and will be configured in such a way so that it can deal with variability in the properties of the items to be machined. Our aim is to produce statistical methodology for configuring such a process, that can be applied in many different settings.

Planned Impact

This project has been designed around the potential for economic impact, and has an obvious main intended beneficiary: the UK manufacturing sector. We believe that to remain globally competitive, it will be important for manufactures in industries such as aerospace, rail and the energy sector to develop state of the art, fully automated manufacturing facilities. One challenge in developing such a facility is in handling variability in the material properties of items to be machined, exacerbated by difficulties in measuring the properties directly. An impact of this research will be to give manufacturers tools to overcome this challenge. We will develop methodology for quantifying uncertainty about the variable properties, deriving optimal machining settings that account for this variability, and deriving control rules for adjusting the machining settings mid-process, in response to sensor data. Our methodology will be largely simulation-based, relying on as little physical experimentation is possible. This will help manufactures develop their automated processes more quickly, and at less expense.

Other impacts will result from the training of the researchers on this project. They will gain valuable experience of industrial collaboration (for example, with our project partner Messier-Bugatti-Dowty), multidisciplinary research, and giving training to representatives from industry, from helping to prepare and deliver our training courses.


10 25 50
publication icon
Harris K (2016) A Multivariate Control Chart for Autocorrelated Tool Wear Processes in Quality and Reliability Engineering International

Description There have been two methodological themes on this project.

1) We have developed novel methods for inferring the material properties of a metal from sensor data, as the metal is cut during a manufacturing process. This is an alternative approach to standard methods that require destructive testing of the metal. Sensor data involving temperatures and forces is obtained whilst the metal is cut, and then compared with computer model predictions that link the material properties of the metal and the cutting operational parameters to the physical observations obtained from the sensors. By a process of model calibration, input settings corresponding to the material properties are found such that the model outputs are consistent with the sensor data. The calibration methodology makes use of an 'emulator': a fast statistical approximation of the computer model, that allows a more comprehensive exploration of the input space when the model is computationally expensive, and so model runs are limited. Uncertainty in the emulator approximation is accounted for in the calibration methodology. The methodology also makes an allowance for model imperfections: the fact that even at the correct input settings, the model is not expected to reproduce the physical outputs perfectly. Estimates of material properties inferred in this way can then be used to optimise the manufacturing process, to set cutting parameters to optimise tool life and production times. Two journal articles are in preparation: one investigating different calibration methodologies, and one presenting a case study in the machining of a titanium alloy.

2) A second methodological theme is concerned with process monitoring and adjustment. We have developed statistical process control methods for monitoring the condition of a machining tool, in real time, based on sensor measurements of acoustic emission, sound, spindle power and vibration of the tool during a cut. The methodology involves (i) extracting features from the high frequency sensor data that correlate with tool wear, (ii) training a statistical model (a multivariate linear state-space model) on the extracted data when the process is in control (the machine tool condition is satisfactory), (iii) operationalizing the process control monitoring by comparing one step ahead statistical model forecasts with observed sensor measurements, with large discrepancies indicating a process out of control. Following on from this, we have developed for feedback adjustment, where different intervention strategies are investigated to prolong tool life. A journal article has been published, presenting a case-study in the machining of a titanium alloy. A second article is in preparation, presenting the results of simulation studies in feedback adjustment.
Exploitation Route We have developed software that can be used to implement our methods. This will be made available online (for free download) in the Summer of 2020. It is intended that the same framework can be used in other manufacturing settings where there is the combination of computer model predictions and sensor data.
Sectors Aerospace, Defence and Marine,Manufacturing, including Industrial Biotechology