2 minutes read
A robotics company was falling behind on completing their machines. These were large works, consisting of up to, and sometimes more, than 20 000 individual parts. While putting together subassemblies, some of the more complex parts that required both normal and specialty fasteners would take weeks to complete, since parts needed to be sourced. The parts containers just for fasteners was truly vast, requiring thousands of bins spread across dozens of dollies throughout the workplace. Parts were deemed to be low only when visually confirmed by an assembler, who would then fill out a request form. The quantity requested was generally just enough to get the project done, but in base cases, special parts would be over ordered, not to be used for years.
After back to back delays on projects completion, which threatened long standing international contracts, Automation Toronto was contacted to find a solution.
We started our solution from the engineering team, where interestingly, one of our proposed optimizations would actually be counterintuitive: There was no real need to use the best part for the job. The designers had been using fasteners only as large as needed to withstand the stresses and without realizing it, they were using non standard sizes, vastly increasing the cost and specialization. With the suggestion to always use the next standard size up if possible, we consolidated a large swath of the resources.
We then made a custom ingest server for all their past parts orders, passing the parsed data to pytorch. The data we were interested in getting was what parts are needed, how many of each and when’s the earliest they would be used.
Machine learning allowed for patterns to emerge that a person couldn’t see, (barring John Forbes Nash. Certain sectors work in a cyclical pattern, which meant those industries would look for new machinery at particular times in the year. The system output a standard file on general parts requirements, which was optimized further on every subsequent parts list fed to it.
Delays in completion due to parts shortages were reduced by over 70%, while over buying was reduced by 40%. We further suggested the engineering team output a parts manifest, one that could be used to optimize the system to its maximum potential.