Author Topic: Using neural networks to get our heads out of the b#)(%* details and into fun!  (Read 1895 times)

0 Members and 1 Guest are viewing this topic.

WILL HATCH

  • Bull Frog
  • Posts: 450
Hey there everyone! Long time to talky talky..
So things took a detour as the oil and gas engineering thing went into decline in Calgary and I started looking at the industry as a whole for a while, then played in politics for the election last year while I've been thinking about how to leverage natural language processing to better organize our group efforts in team projects.

That said, I've read at least fifty papers by now and there's a surprising number of open source tools that are available for us to leverage. We can start by simply feeding complete drawing and documentation packages into a system and have it recognize the features, ESPECIALLY on client specific drawing specifications, or even better have access to a database that drove a drawing and document package.

At the very least I had derived a simple 3 input database for electrical cabling/interconnection design for oil and gas facilities that inferred all the wiring from the process design, the purchasing info, and the object location in the 3d model that could be used for starters. I had walked from this project partially because I didn't have the tools to adequately be able to quickly derive client standards with my current time resources.

I want to put the call out to the community here because ultimately my experience in just the northwestern North American oil and gas industry was just a few years and if it's worth developing a system like this it would be worth targeting as many industries as possible. My interests during university were deeper into hardware than software, lots of high frequency analog electronics, integrated circuit design, power electronics courses in the choosy years at the end. With this came a lot of maths with the imaginary numbers opening little portals into easy ways to describe complex numbers after we got through the mind rape of learning residue theorem.

All that aside, I'm driven by the fact that our engineering industries are slowly being outsourced because after paying the high cost of software and people there seems to be not enough money left for investors, as the trend I've noticed is an increase in outsourcing and a cronic abusive relationship between many firms and their customers creating a bad name for the industry overall. The cost associated with switching firms and reforging a strong relationship is high for many of our clients, as even they are struggling to maintain consistency across all of their design methodologies. None of these are significant problems if we throw some of the more recent advances in neural networking. I think that we could even design a network that can create the dwg files without any other intermediary and recreate the industry essentially overnight. Personally here, I think if I was to work in this direction that I am easily poised to at least create an autocad based app that would wrap up the system I had been developing a couple years back, but before I get started I wanted to see if really seriously looking at this goes with anybody.

Cheers!

Atook

  • Swamp Rat
  • Posts: 1027
  • AKA Tim
I've been curious about using machine learning (via tensorflow) for the same purpose of coming up with design data, but haven't gotten around to coming up with a dataset.

I agree with you, I have access to lots of data sets which could be used for a learning set. My first thought was that I'd need to pick a specific problem, and extract only data related to that problem. The industry I've got the most experience and datasets for is landscape irrigation mostly dealing with hydraulics (no need for machine learning or neural net just yet) and locating sprinklers (the problem I would approach first). My initial thoughts were to choose one discrete part of the overall design problem and solve for that first. After learning to do that, breaking down other parts of the design problem and extending from there. The possibilities are staggering.

I'm not sure what you're asking for, but if you're looking for a partner to explore using existing CAD drawings to create a design application that uses machine learning to create new design drawings, I'm interested.

dgorsman

  • Water Moccasin
  • Posts: 2437
"Are you measuring what you *think* you are measuring?"  One of my favorite phrases, and derives from a story regarding neural net training.  Some researchers were working on a project to recognize camouflaged armored vehicles in photos.  They had a wide set of pictures used during the training cycle of the neural net; they pass in the image, tell the program "vehicle" or "no vehicle", and it did its learning bit.  When it came time to test with other pictures, none of them worked.  They rebuilt the logic, used a different set of images, still nothing.  Finally some bright person picked up on something: with every single image in the learning set, everything with a vehicle was daytime.  The neural net wasn't being taught whether a vehicle was in the image; it was being taught which images were daytime.  Yeah, neural nets can be fun but they also have a fuzziness that isn't going to fly very well with drawing creation - especially when we can just create a mostly-finished template drawing instead.
If you are going to fly by the seat of your pants, expect friction burns.

try {GreatPower;}
   catch (notResponsible)
      {NextTime(PlanAhead);}
   finally
      {MasterBasics;}

Atook

  • Swamp Rat
  • Posts: 1027
  • AKA Tim
"Are you measuring what you *think* you are measuring?"
That's a great question for all kinds of people. :)

Quote
Yeah, neural nets can be fun but they also have a fuzziness that isn't going to fly very well with drawing creation - especially when we can just create a mostly-finished template drawing instead.

Many of the problems solved in the drawings that I send out are fuzzy, meaning there's more than one solution, so a neural net solution may make sense. There isn't an easy way to create a 'mostly finished template' for the irregular shapes and sizes of landscape areas. I certainly think it's something to explore, worst case scenario: I spend time learning about that machine learning fad that never took off. Best case scenario: I figure out how to solve the most time consuming part of my design process with a click of a button, and end up getting a weeks worth of work done in thirty seconds.

Other problems, like load calculations, voltage loss, or hydraulics can be calculated exactly with software and may not be a good candidate for neural nets/machine learning. Those I've got algorithms for, and don't take much time to do.