The Machine Learning Opportunity in Manufacturing, Logistics | Hacker News Reader

The Machine Learning Opportunity in Manufacturing, Logistics

nextplatform.com

29 pointsrbanffy6 comments

I don’t think people realize how big a trend in general manufacturing will be in the future. The life cycle of the world can best be described as new systems / software supporting new infrastructure / hardware.

We’ve had 40 years of massive improvements in software and code. There have absolutely been hardware improvements in that time, but we haven’t seen the levels of improvement we saw from 1920 to 1960 for example. The reason for this is that software or hardware tends to dominate for 40 or so years, followed by a reverse, where the new area takes advantages of the advances laid down in the other. This dynamic tends to play out in other fields of a soft science or liberal arts supporting a hard science.

Interestingly, the 40 year pattern tends to hold for national real estate as well, where prices tend to shift back and forth between coastal areas and the “inner US”.

All of this means that the mid-west, with its incredibly cheap real estate and 100+ year history of manufacturing won’t be the new Silicon Valley ever, but it will be what it once was, a great place to be during a dominant era of American manufacturing supported by advances in software. Probably will also decline again at some point too though.

Also there are plenty of places in Ohio where you can get a 2500 square foot 100k house in a reasonable school district in a town of 200k+. Also means SF and NY real estate likely to revert to what it was like in the 1970’s, a super cheap albeit dangerous place to live.

This article basically says that there's a trend of increasing interest. Well, that's great--I'm sure there's plenty of interest from industrial firms. But there's a lot of grunt work that needs to be done before anyone can hope to make real money off ML in this space... industrial equipment is often ancient and analog, standards are all over the place, and the talent pool isn't there. GE learned this the hard way with Predix.

Yup. This is my day-to-day life. Yes, huge industrial firms might have a ton of data... but it might be spread out over 5000 hard drives from 1995-present. It's extremely hard for an outside AI/ML startup to target that type of space, because they just don't have the data. Or the knowledge that lives in the head of a 30+-year-engineer who knows everything about every piece of equipment that has ever been used on a manufacturing floor.

Do you work directly for a manufacturing company/relevant company?

I wonder if this would be a better opportunity for an outside firm in a longer-term contractual role, so that they both have some autonomy but also access to the data. At least that way they come away with field-knowledge but would obviously be restricted in terms of the proprietary data.

I don't know how real-world unrealistic I am being, but for instance:

- Pitch the potential project, including its mass of limitations, but also its end product / projects given x is possible, y is possible, z is possible, etc

- Present a contract that allows for years, potentially, and enough running capital to hire a staff that can handle modeling the schema, collecting arranging the data, interviewing the experts and quantifying their understanding if necessary.

I mean, if there's a perceived and expected value out of doing it, surely there could be some compromise like start with one sub-process and one type of machine and go from there?

Do you imagine it should need to be holistic?

Or, maybe design a new machine?

For instance, at a major auto sub-manufacturer they are extremely critical and thorough with their quality controls. Currently, this is done 100% manually. Parts (pre- and post-assembly) every N parts or H hours are pulled off the line while the line runs, taken to the lab, welds tested, measurements tested against calibrations, and in some cases chemical tests made to ensure proper galvanization.

That data could be entered initially and measurement data accumulate pretty rapidly.

I'm sure that process could be replaced for the most part using the AI/ML.

I would be keen to have a chat. I come from a software architecture background myself, but we are working on a distributed network of local food manufacturing robotic service locations - http://infinite-food.com/ - and a central industrial facility for ingredient inflow, Q/A, etc. On the latter in particular I'd be keen to pick your brain for insights as there is a whole world of machinery out there, it's not cheap, any insights in to data could be useful and we've already hired ML expertise and begun to apply it to manufacturing logistics. Send me an email if you are happy to have a chat, email in profile.

Yes let's do heuristics instead of first principle modeling and rigorous math programming / control.

Reply on news.ycombinator.com