|Parallel Simulation Technology|
PIM/Paracell Application Notes
The process of applying the finish in the production of vehicles is a delicate balance of both art and science. The fit and finish of a vehicle are one of the primary buying critera of the consumer in the marketplace. In the automotive painting process, there are many variables that need to be controlled to provide a quality finish. These variables include the temperature and humidity in the paint booth, the thickness of the finish, even the type of paint being applied. These variables and others must be carefully monitored and controlled in a closed-loop in order to maintain the yield of the paint process and provide a quality product to the marketplace.
In 1990, General Motors agreed to become a beta test site for Parallel Simulation Technology Parallel Inference Machine (PIM) and Paracell Programming Language and environment. They chose paint as an initial application, feeling that this particular area offered a high level of complexity required to evaluate the products, and if solved would provide significant payback. The project was assigned to a GM Paint Process engineer who developed the initial application in a matter of weeks. Once installed, the application ran providing closed loop control for eighty robots and the paint booth air supply houses. Maintenence and enhancement of the application was taken over by a plant floor electrician.
By GM's own reckoning, the application saved over $1.5M in paint during its first year in use. It offered further reductions in cost by reducing reworks, increasing booth yield, and reducing the amount of waste resulting from the process. The year following the installation of the application, the participating plant recieved the J.D. Powers Award for Best Fit & Finsh.
In early 1990, Parallel Simulation Technology PIM and Paracell were ready to emerge as the first commercially available platform for what has come to be known as "autonoumous agents." As the innovator in bringing the programmable logic controller to automotive manufacturing 22 years earlier, Ernie Vahala of GM Truck & Bus was approached to host a beta test of this new control platform. He assigned a young engineer, Gregg Ekberg, to the task of selecting an application for the PIM and seeing that application through to evaluation.
Gregg had an idea for closed loop control in the paint process that he had as yet been unable to implement. There are three parameters in play at the robotic spray nozzle in the paint process; paint flow, fan air, and atomization air. In order to provide the highest quality paint job with the correct amount of material, setpoints for these three parameters must be controlled carefully. They can be influenced by environmental conditions in the paint booth itself, or by the type of paint (manufacturer, color, viscosity, additives) being applied to the vehicle. At the time the beta test was being considered, these parameters were being run open loop. Further, if paint quality seemed off, the common solution was to turn up the flow, resulting in overuse of paint and an increase in the by-products of the process.
Thus, closed loop control of paint flow parameters was selected as the initial application of the PIM. It was also decided that to prove the platforms ability to provide more conventional control, the air supply houses that regulate the environmental conditions in the paint booths would also be part of the test. Inasmuch as monitoring of booth humidity and temperature was required anyway to full test the flow algorithms, control of the enviroment seemed a natural extension to test conventional control.
The site selected was General Motors Ft. Wayne Assembly in Ft. Wayne, Indiana. The Ft. Wayne paint facilities are atypical of other automotive paint shops in that it is a modular shop. A serial queue comes into the paint shop where there are ten painting lines in parallel. Three lines are dedicated to rework and two-tone vehicles, while the other seven paint trucks with uniform colors. Any color can be run in any of the "mods." Each mod consists of eight robots; four painting robots and four clearcoat robots. Between paint and clearcoat is a "flash" booth where trucks stop for a short while for the paint to cure. A flash module follows clearcoat as well. Once trucks have been through the process, they rejoin the serial line for completion of the assembly process.
In beta testing the application, it was important that the system be implemented in such a manner that if the PIM application and contol had to be taken off line, this was easily accomplished. For this reason, PIM generated setpoints were run through an Allen-Bradley PLC-5 sothat taking the application off line and returning to open-loop control was as easy as throwing a switch. The PIM was connected to a VAX via a direct memory interface so that ideal setpoints for particular conditions could be logged on the VAX. A single Macintosh programming workstation was connected to the PIM for application development and to support user SoftScope displays. Once the application was proven, I/O support was added to accommodate ten additional Macintosh workstations. One was placed in each mod to serve as an operator interface to the control application.
Development and Deployment
The application development began at Flavors' offices, and was undertaken by Ekberg. With ladder logic as his primary control language, Gregg was able to quickly familiarize himself with the English-like constructs of Paracell. The challenge in application development on the PIM did not come from the language, but rather from the parallel nature of the machine. For engineers who are used to developing an application in a serial manner where event sequence has critical relevance. programming an agent-based application can be challenging. Ladder logic is a good example of this, where events on a "rung" and the "rungs" themselves are executed in a serial manner, and can be adversely affected by "scan time" variations.
In the earliest weeks of the project, Gregg concentrated on the Paracell code required to implement closed loop control of the flow parameters. This effort was followed by implementation of the air supply houses for the ten mods. Development of this application on the PIM took about a month, and in the course of development a 13:1 reduction in control code was realized. Another two weeks were required to implement the VAX code required to capture and store setpoint information.
Having been developed and tested, the applications were rolled out to the plant floor in Ft. Wayne. Senior paint shop electrician Tom Allyn was assigned to the effort. Tom oversaw integration of the PIM into the existing Allen-Bradley control network, and began assuming day to day responsibilty for the project as it was transitioned from Ekberg to the plant personnel. Though he had no formal programming experience, Tom took to the language easily and in a short time was developing his own SoftScopes and writing Paracell code to examine different aspects of the paint process.
The applications were intially tested on one line. Quality data was collected to determine the correct setpoints for given conditions and a given type of paint. Once determined, it was the PIMs responsibility to control parameters such that these setpoints were kept under control. Film builds were taken to determine the correct amount of paint to apply to assure a quality finish. An objective of this process was to minimize the amount of paint required in order to prevent sagging, "orange peel effect", or uneveness of the coating. GM also sought to minimize the by products of the process in order to control the costs of disposal and minimize the impact on the environment. A futher benefit of the application was obviously cost control. Automotive paint is quite expensive running in excess of $100 per gallon. With PIM control, GM was able to reduce material use while improving quality and protecting the environment.
Once tested, the application was rolled out to the remaining paint lines in the mod. On the flow application alone, this represented 240 closed loop PID functions, each running in parallel and in real-time. Following a year of operation, General Motors performed a beta test evaluation of the installation. It was determined that quality of the product coming out of the booth had been improved, the number of trucks requiring re-work had been reduced, and material savings in the amount of $1.5M per year were realized.
The application was in use in Ft. Wayne for several years following, until such time as the mod was switched over to a different paint application process. During that time, GM continued to use the PIM and Paracell in paint applications. They also developed a simulation of agent-based pull through scheduling based on their experience with the PIM in an attempt to minimize color changes on a line. Rather than assign a truck from the serial queue coming into the booth to a particualar mod, a bidding mechanism was set up whereby each mod would bid on an incoming job. The bid was weighted based on what color a mod was currently painting, and how many jobs it had waiting in its' queue. GM determined that this scheduling algorithm would sigficantly reduce color changes, however the simulation was never deployed as a scheduler.
Following the changeover of the paint shop, the PIM was decomissioned. However, the application remains as the pioneering example of how agents and complexity theory can be applied to meet the challenges of manufacturing and process control. The pull through scheduling paradigm and the concept of machines bidding for work that were conceptualized in Paracell at GM are the subject of numerous schedulers in use today.
Parallel Simulation Technology LLC