AI Accelerator Neural Networks
Neural Networks “solve the unsolvable” problems in manufacturing:
Sequencing of Jobs to reduce Setup Time by 75%*: Consider a machine in one of the Pull Groups that is about to complete its current batch of products. In the “Aerospace” factory, Fig 6-2, (below) from page 122 of the book “Lean Six Sigma in the Age of Artificial Intelligence” we note that there are about 50 jobs waiting at that Pull Group. Which of the 50 jobs should we assign to this machine? Most companies would select the next job up on the customer schedule by date. This generally means that a completely new setup process will be undertaken which uses few if any of the tools and methods of the job about to be completed. One might instead pick the job which is the Nearest Neighbor in setup tooling and methods to the current job, so long as AI Accelerator WIP Control verifies that all jobs of the Pull Group will still be on time to >95%. This will then reduce the setup time waste on this job. However, what we really would like to do is to minimize the total setup waste on the next, for example, four jobs that this machine processes which is always significantly lower than the “Nearest Neighbor” approach. To do this we would have to examine1 50!/(4!46!)=230,000 sequences! This will find the sequence of four jobs with lowest possible setup waste while still delivering on-time with >95% probability. Given all the possible combinations that go into 50 jobs, finding the minimum total setup time sequence is clearly beyond human capacity. However, a trained Neural Network can accomplish this feat in a few seconds2! In larger factories, there may be 200 or more jobs at a Pull Group, which only increases the likelihood of near perfect setup matches and further reduces total setup waste. Before we release a sequence of four jobs, we first use AI Accelerator WIP Control to verify that all four jobs will be shipped on time with >95% probability.
Scrap after Setup: In this case a process that is under-control after setup is completed suddenly generates scrap. How did it happen? How can it be prevented? Everyone in the factory will have an opinion, but the human mind can only consider a few possibilities. Most processes are well instrumented and provide data on scores or even hundreds of process parameters. Buried in this “Big Data” is an often counter-intuitive answer which is dependent on scores of parameters. The Neural Network will search for common patterns in the data of scrap events versus a prior run that was scrap free. When we speak of “Big Data” this also includes variation in the process parameters of material from outside suppliers.
*U.S. Patent Pending
“Lean Sigma in the Age of Artificial Intelligence” Page 122