Optimizing printing systems-of-systems
At their print-on-demand facilities, the big publishers of this world have a variety of different printing and finishing machinery working in close harmony with human operators to ensure print work in all shapes and sizes keeps rolling off the press. “Especially during peak season, typically the end of the year, they want their production to tick like a clock,” says Rob Kersemakers, R&D workflow architect at Canon Production Printing. “They want to get everything out of their systems so that we can all order our books and Christmas cards to have them delivered the next day. At Canon Production Printing, we want to give these customers actionable guidance on how to optimize their complete production lines – what equipment to use, in what order to execute tasks, how many human operators to deploy and where in the chain to deploy them. That’s what we’re looking to achieve in the AIMS project with TNO-ESI.”
The first part of AIMS (Adaptive Intelligent Manufacturing Systems) focused on tightly coupled machinery, ie systems that are physically connected with limited buffering between them. “Think of a printer that automatically hands over the printed sheets to a finisher that combines them into a finished product by cutting, gluing, perforating, stapling, stitching or otherwise binding them,” illustrates Joris Baijens, responsible for third-party finisher integration at Canon Production Printing. “Typically, these finishers are predominantly mechanical, with only a small amount of software. By connecting them to our printer, we’re pulling them into the digital world. This gives us much more flexibility but also a huge integration challenge: if the timing of the two systems is misaligned, their combined performance easily becomes pretty terrible. With the support of TNO-ESI, we’ve developed a methodology to reason about the productivity of such tightly coupled machinery.”
This year, the partners are expanding their focus to loosely coupled systems. “We’re now looking at overall equipment effectiveness and even overall factory effectiveness,” Kersemakers elaborates. “Here, the human aspect is added to the mix. As these loosely coupled systems have no physical connection, the sheets or stacks have to be transported manually from one work cell to the next. A different job order might reduce waiting times down the line. Some print work needs to dry before it’s ready to be used, so the climate on the factory floor needs to be taken into account. There might be service-level agreements in force that command strict delivery times. All these factors now come into the modeling game.”
Timing analysis
The idea behind AIMS is to create models that predict what will happen on the factory floor, explains Marvin Brunner, lead technologist at Canon Production Printing. “Our printer is very flexible in that it doesn’t matter in what order it prints the images. For the post-processing, however, the order is highly relevant. In the first part of AIMS, we modeled the different links in the printing chain and we explored how we can exploit the flexibility in our printer to optimize the total workflow. The models aren’t perfect, but they’re good enough to give us an accurate prediction of how long it takes to make a product from start to finish.”
Jacques Verriet, senior research fellow at TNO-ESI, sets the modeling scene: “We have a production line and the print work goes from an input machine to an output machine, possibly passing some other equipment along the way. Each station performs a number of operations that take time. There’s also some time in between the operations as the machine recovers from one operation and prepares for the next, and there’s transport time between equipment, so the start of the work at the next station may have to be delayed to avoid breaking the flow.”
Using domain-specific languages (DSLs) and tools such as Xtext en Xtend, the partners have modeled the print production line distinguishing four main aspects. “The material model describes all the materials we need to handle, like sheets, stacks of sheets, A4 or A3, as well as the operations we can perform on them. The job model describes all the recipes we need to follow in processing the material to get the work done. The equipment model describes the different material processing stations on the production line, the operations we can perform there, how much time they take and how much time it takes to switch from one operation to the next. The allocation model then describes how the recipes are to be assigned to the various stations,” summarizes Verriet.
From this information, a so-called constraint graph model is automatically generated that specifies all the events that occur while executing the work on the production line, as well as the minimum and maximum time between events. “In the DSLs, we capture all the domain knowledge required for timing analysis,” says Bram van der Sanden, research fellow at TNO-ESI. “Thus, for each operating station, we know what actions are being executed on what type of product, how long they take and what the minimum and maximum durations of task switches are. The resulting constraint graph model serves as input for a scheduler that computes in which order and at what times to execute actions. Using the Bellman-Ford algorithm on the constraint graph model, we arrive at the optimal workflow.”
“With tightly coupled systems, having hardly any buffering imposes severe constraints, but it’s exactly because of this limited freedom why the trick we’ve come up with works so well there,” notes Verriet. “Moving to loosely coupled systems, the big challenge is in the timing variation of the transport from one station to the next. Evidently, humans aren’t as predictable as machines, so we’re going to have to take that variability into account. We’re still exploring the possibilities there. We could stay on the analytic path, we could opt for simulation or we could mix both into a hybrid approach.”
Better insight
“By giving us a better understanding of how our printers cooperate with finishers and other third-party equipment, the AIMS project is bringing us and our partners closer together,” Baijens observes. He sees big value for customers as well. “We have a lot of customers with really big factories, requiring equally big investments. When the need arises to buy new equipment, they all start calculating the total cost of ownership. In doing so, it’s extremely helpful to be able to reason about productivity and performance. Our work with ESI is bringing us closer to our vision where we can analyze their situation and advise them on how to expand their production lines to best fit their needs.” Brunner adds: “When one of our printers is paired with a third-party finisher we haven’t met before, we want to be able to quickly come up with an accurate prediction of their combined performance and assist operators in making the right decisions on planning and scheduling.”
Kersemakers concurs: “I envisage us recommending optimal production routes and providing our customers with actionable guidance for their operators to improve the performance. AIMS will help us give them a better insight into what’s happening on their factory floor so they can get the most out of their production lines and better prepare for those peak loads. That’s becoming increasingly important in these times of narrowing delivery windows and tightening service-level agreements.”
For Brunner, the AIMS project nicely links up with the workflow research he was already conducting at Canon Production Printing. “In our pre-existing efforts, we modeled the equipment from observed behavior: some actions we see being performed in sequence, some in parallel, depending on the material coming in and the status of the machine. One of the insights the collaboration with ESI has brought us is that we can also create this constraint model. And by leaving things out, we have more freedom to explore other analyses, which can lead to interesting alternative optimizations based on more or less the same model.”
Baijens believes that, down the road, the learnings from AIMS could also be very useful for his colleagues in engineering. “We’re aiming to use more models and combine models to get models of models. This will help us be more effective and efficient in engineering new printers.” Brunner: “AIMS gives us an insight into the effects specific machine setups have on the overall equipment effectiveness. That may influence design decisions too.”
“At TNO-ESI, we’ve been investigating performance optimization for quite some time now. Looking at the high-tech industry, I think we’re reaching the limits of what we can squeeze out of a single system,” Van der Sanden points out. “The next step is to view the system in a broader context that also includes the details of the systems it’s cooperating with and try to achieve a higher equipment efficiency at this system-of-systems level. This calls for methodologies that enable us to reason about system-of-systems optimization. The AIMS project is a great stepping stone to see what concepts can be useful there.”
Any production facility
As with all TNO-ESI projects, the results from AIMS have broader applicability. Although the methodology being developed is currently tailored to optimizing the workflow of print service providers, it’s generic enough to be extended to any kind of processing facility. This includes the semicon fabs equipped by ESI partners ASML and Itec, as well as the baggage handling systems built by Vanderlande, for example.
“We’ve already modeled a few other systems,” Verriet concludes. “All fully automated, with limited buffering, so the methodology was easily transferable. Once we finish the extension to loosely coupled systems, really any production facility would work.”
This article was written in close collaboration with TNO-ESI.