At its most basic, a closed-loop control system needs input from sensors, some kind of interpretation of that input (for example to determine whether a process parameter is in tolerance), and the ability to alter the original parameters of the process to change the end result: to course-correct.
Closed-loop control systems prevent processes from operating outside their production parameters automatically, reducing the potential for scrap and rework, if the process is well-understood and the technological levers can be put into place.
Doing just that was demonstrated recently at the University of Sheffield AMRC in the context of metalcutting, in particular in a legacy CNC (computer numerically controlled) large-format milling machine. In five-axis machine tool operations, a control system determines position and velocity of a cutting tool mounted in a spindle in the X, Y and Z linear axes, plus rotary drives A and B (rotation around X and Y), as well as spindle rotation speed – all closed-loop controlled, according to Robert Ward, industrial research fellow at the University of Sheffield. The tool cuts a metal billet mounted in the machine according to a program. The way the tool pushes into the material may produce efficient or inefficient cuts, and can stress the tool so much that it bends or breaks. A tool break may damage the part, causing delays due to rework, scrappage and resetting the machine.
To avoid these risks, machinists deliberately create cutting programs with a safety margin. Because of that safety margin, these result in less-than-optimal performance. It would be more efficient if the machine itself could sense problems in their very early onset, and course-correct, or even better before they occur.
In fact, some already do. Ward refers to adaptive feed control from machine tool controller manufacturer Heidenhain, which automatically slows the spindle movement in response to the load felt by the tool, which is similar to Siemens’s Omative Adatptive Control and Monitoring for its Sinumerik CNC controller. But they are rare. Such systems have not been taken up partly because contractors to the aerospace sector, at least, are bound by the sector’s very tight process controls.
HOW TO DO IT
Large machine tools are extremely expensive, so are long-lived. While modern controls are compatible with big data exporting processes, older controls have been far less amenable, until now. With the project, Ward’s team was able to tap into the Starrag machining centre’s Siemens 840D Powerline controller, a product no longer supported by the manufacturer, at speeds fast enough to use a closed-loop control to alter the metalcutting process.
Here is how. The team used Siemens hardware, its Axis Data Stream monitoring system, to extract encoder positions at 250Hz directly from the numerical controller kernel of the controller (up against the maximum clock speed of the controller’s processor); this was far faster than PLC data. That data was extracted via Profibus protocol to a comms processor, in the form of a card loaded into the researchers’ high-end Dell laptop.
Inside the laptop resided synchronised metalcutting simulations, digital twins, the fruit of previous research programmes.
Feeding the results of the digital twin predictions back into the machine tool were so-called ‘analogue modules’, which plugged into the numerical control terminal back panel of the machine. (Actually, the signal flow was more complicated than that, since both source – the laptop – and destination – the CNC controller – were digital. From the laptop, the signal progressed through a National Instruments digital-analogue converter, to the analogue modules, through their own analogue/digital converter, and into the CNC).
The set-up exploited a special function of the Siemens controller called Synchronous Actions, a way of programming in G-code, the CNC machine language. Adds the researcher: “You can almost write control loops within the part program itself without having to do any form of in-depth programming behind the scenes, and it operates at the interpolated clock speed as well.”
THE CHATTERING CASES
One problem in metalcutting is ‘chatter’, in which the tool starts vibrating uncontrollably when its movement matches an aspect of the material properties – it hits a resonant frequency, like the ringing of a bell. According to Ward, many control systems have an anti-resonance control that will increase spindle speed or lay off based on spindle power data. But, he adds, that means that something has already happened.
His team had a different aim, he reports. “What we were looking at is, can you predict that in advance, and you can change it before you even get there? And that’s where things like digital twins and these kind of forecasting models in the near term will really come into their own. And that worked quite nicely actually. There were some challenges with it, for sure. But it did work: it predicted ahead that it was going to go into chatter; it then changed spindle speed, and it didn’t go into chatter.”
FINDINGS AND FUTURE STEPS
The point of the project was not to produce an embedded control system, or a finished product, he cautions – but rather as primary research to develop the field, which he calls ‘rapid control prototyping’.
In doing the work, he reported that one of the biggest challenges was computational bottlenecks; although computer processing power was fine for relatively simple machining toolpaths, it struggled when doing complex five-axis machining. “If you’re operating very quick as well, if you’re roughing [out], perhaps, you need to be calculating fast, because otherwise your calculation is here and your tool is over here.”
Although the project has now ended, the digital machining team at the AMRC is developing hybrid models as part of the high value manufacturing catapult research streams; this is where they are integrating AI into the system. That will be a significant upgrade, Ward argues. “What we did was purely mechanistic modelling; there was no learning as such within our models. Now the way that it’s going is how to combine real data with measurements we are getting from machine tools to update those models. How do we fit that within the digital twin framework; that’s where it’s going. So you can look at digital twins plus machine learning and AI. But the trick is, how do we do that quickly, in real time. That’s the challenge.”