OpenAI API Manufacturing and Industrial Use Cases

API
46

I’m an aerospace engineer at a research institute. I’ve published some work using Whisper and GPT-4 to perform real time monitoring of air traffic control (ATC) audio communications. This could detect dangerous anomalies or monitor ATC performance. It worked correctly on 9 out of 10 examples, using real-world ATC audio. The paper is called: “Aircraft Anomaly Detection using Large Language Models: An Air Traffic Control Application”.

In addition, I’ve linked GPT-4 to aircraft propulsion design software for aircraft conceptual design and sizing. It works well with small projects, but it is context window limited. It needs to be fed a condensed version of the user manual in every chat, driving up token costs, and then all project files need to be included as well. When it works, it works well! It still struggles with spatial reasoning enough that I would not trust it to design an aircraft engine yet.

I’ve also investigated automating satellite component ground testing, but it struggled with multi-step tasks and big-picture thinking. I think the best use case for this application is still regular pair programming with chatGPT or Copilot.

Overall, I haven’t found good ways to enable autonomous behavior, and I keep coming back to “chatbot” as the best solution to many problems. This is unfortunate, as I want to make use of this intelligence, but without strong guardrails and loss of human oversight, I haven’t been able to harness it very well yet. Aerospace engineering tasks require a lot of complex decision making, so designing rigid prompt structures is difficult. Additionally, what I think are Aerospace problems end up being 90% software engineering problems. The nice part is that GPT-4 is pretty solidly at intern level for many tasks. The trick is finding tasks that 1. Require intelligence and 2. Can tolerate failure. I call them intern problems. I’m looking forward to GPT-5 and larger context windows so I can promote it from intern to junior engineer.

5 Likes
47

This is a very interesting topic. One thing I have experimented with is performing analysis on environmental assessment reports that as you know can be massive. I can create a distillation of the key points and risks that can be used to better facilitate discussions around the report. I performed this using a California multi-million dollar environmental assessment report that spaned over 900 pages.

1 Like
48

@logankilpatrick - i am glad to hear this is an area you are exploring. the “built environment” can benefit from this ai boom as much as software.

aside from the obvious customer support chatbot for my industrial additive manufacturing business, i wanted to share that i’ve been working on a 3D gen ai for CAD, which is currently running openai in the backend:

i would be excited to find a way to collaborate more directly with openai.

2 Likes
49

In my day job I do Commercial Building Automation. Not only do I work on analytics and programming of the systems and equipment, but I have to fix controllers and other smart devices.

This means having an extensive knowledge of a huge range of products and areas.

So I built a little tool to help me during the day. When I arrive at a destination, I get a notification on my phone with a link to my previous visit notes. When I’m working on a unit, I take a picture of the unit nameplate and record the info. It starts a ticket and begins populating a report as I take notes.

I can take a picture of an electrical wire diagrams and have the agent help me diagnose based on the reported issue. I take complicated data histographs for unit and building performance, and use GPT for fast analytics that I then use in tandem with a separate analysis of the raw report data.

All this gets put into the report and emailed to me.

My goal is to connect it to CRM and inventory apis and get a full up to date client interaction guide along with my mechanic bot.

Still needs work, but I haven’t seen anyone other than 1 or 2 SaaS in my industry doing this. With a little more time and money I think I could truly disrupt this sector of the service industry.

3 Likes
50

Let’s look at a workflow in a company manufacturing turbine engines. A (traditional) engineer would work with the assistant like this:

E: Create the geometry of a turbine blade with this size and these properties…

A: [Guides the engineer to give all important parameters] and shows the 3D model.

E: Perform the CDF analysis for the blade, in this engine condition…

A: One minute, here is the result: [3D visualization and summary of performance metrics]

E: Perform structural integrity analysis for the blade for this operating environment…

A: Sure, here is the result…

E: Run up to 20 thousands of samples of the above analyses with various parameter configurations within this design space…

A: Sure, here is what I will do… please confirm.

E: Looks good!

A: Ok, it will take a bit of time for me to complete this data generation, please check back with me in a couple hours.

A. Completed. Below is the real time model for you to explore the design space and choose the best design…

E: Looks great. Now, save this workflow as Turbine Blade Design and make it available for everyone to use.

A: Workflow Turbine Blade Design is created with the following details.

Sometime later, another engineer in the enterprise:
E: I need to design the turbine blade.

A: Sure, to design the turbine blade, here are the steps I will do… and here are what I need from you…

The above workflow shares the same requirements as the cases mentioned in the question. Note that up to now, enterprises have been running without GPTs. There are a multitude of mature enterprise software tools that GPTs cannot replace easily. What GPTs should do is to connect with and use existing enterprise tools to intelligently capture the knowledge and automate the workflows. An expert performs a workflow once and GPTs can capture and repeat it, and even more…

I have been building the system like the above, and I can see there is still a lot of room for improvement for GPTs in this space.

2 Likes
51

We are utilizing industrial IoT data and maintenance logs to enhance our predictive maintenance feature by processing it through the OpenAI API.

52

As do I. I’m sorry to say that my experience with this has been extremely frustrating.

More than once I have been in the final stages of something and the api behaviour changes subtly in ways that break what I am working on.

If I could fine tune gpt-4 it would most likely be fine but I have had no response to my application for access or from support.

Good luck.

53

My company, Spiro.ai has built an OpenAI powered CRM for manufacturers.

We use OpenAI to summarize everything that’s happening at a customer – email interactions, transcripts of phone and Teams calls, orders, meetings, etc. We call this an “executive summary” of a company.

We also use OpenAI to draft emails for our customers based on their phone calls and video meetings. And many other things too.

We have over 100 customers and 3,000 users at this point.

2 Likes
54

:scream:

How did you get the customer to agree to that? :laughing:

anyways, welcome to the community!

55

With a shrinking population and a growing avoidance of manufacturing jobs, the industry as a whole is worried about a shortage of workers.

I think it is more efficient to apply the maintenance phase rather than the design phase to develop products.

  • Spare parts inventory management
  • Understanding and creating parts drawings or manuals
  • Image-to-text for maintenance and RCAs
  • Machine diagnostics

All of these are necessary because they are all part of the job of a maintenance person.

1 Like
56

Have spare parts for vehicles as a website service. They would love to be able to have their website look and search for on demand parts from other suppliers that they could then include as it were part of their supply that they can order through and automatically bump the price based off of their cost for their profit to the final appliance so they can service all of the clients needs for vehicles they’re currently working on in one order to simplify the whole process.

1 Like
57

I have developed a chatbot that assist technicians in doing vehicle and equipment repair. It performs diagnostics and does labor and parts cost estimates. It is modeled after the social media personality @SnaponMadness. It is public facing. Google MadnessBot and you can give it a try.

My goal is to develop it into a full multi-model system that creates “how to” videos on-the-fly to assist technicians with the maintenance and repair of various vehicles and equipment.

1 Like
58

Sorry, which specific follow-up that I posted that you were referring to?

59

Hi @VinceZ , I work in a similar space. We build websites primarily for Industrial Distributors. Also do digital marketing, and we specialize in product data for Manufacturers.

Some of our partners have trained LLMs to provide answers to questions about product applicability, provide spec sheets, answer questions about where an order is in transit, estimated arrival time, cross over products. In warehousing there are many applications.

It would be great to discuss with you more, as there arent many in our industry with experience in implementing and would be great to discuss ideas, etc.

-Dan

1 Like
60

Sounds like doing API call for part availability from other suppliers would be more efficient.

61

With over 25 years software background, I’m not sure how LLM would add value to the function of “provide answers to questions about product applicability, provide spec sheets” unless the current “questions” and “answers” data especially the “answers” would not suffice, then, use RAG to beef up “answers”.

Regarding “answer questions about where an order is in transit, estimated arrival time, cross over products.”, since order information must be current, it would be more efficient to simply call up the order database to provide customer questions on their orders. Just a thought.

62

I am trying to find some projects that I can work on for money. Have any people that could use my services? I’m a little behind I need to make up some flow as soon as I can. My last client abused my wallet and left me on empty.

63

Agree, using genAI for this that can be programmatically achievable is not the way…however, one cannot always foresee all the possible use cases, user flows and data representation that would be the most useful for an operation’s team. For that, especially when a small amount of reasoning and data correlation is required (so using tools for the LLMs to be able to request online data), as well as providing tools to the LLM to do something more than just chatting to answer operator’s needs (e.g. show device/machine faceplates, propose specific commands, recover history of certain monitored tags…) can CERTANILY speed up operations. We have been working for a while on AI Operator in SCADA/Automation fields where we operate since many years, and we can already see how disruptive this is for operation or large and complex systems. Just need to focus in what LLMs can do better and/or faster to support humans, and not just create chat bots… cheers

1 Like
64

informative, appreciate your input. Since you mentioned SCADA, have you tried LLM pertaining to MES for any of scheduling, inventory, quality control, and process optimization? If yes, I’d love to learn some details provided it won’t comprise your intellectual property.

1 Like
65

Absolutely, we are one of the world’s most specialized companies in super SCADA systems, particularly WinCC OA. At CERN, we used it to develop the control systems and expert systems for the LHC experiments. Today, we apply it across various industries: transportation, manufacturing, utilities, infrastructure, logistics…you name it! We often develop MES/MOM-like systems directly within SCADA, using an object-oriented development framework that we created inside the SCADA. This allows our SCADA solutions to integrate significant operational functionalities (or full custom MES systems in some cases).

The most exciting aspects of using our frameworks for SCADA development is that while teaching LLMs to engineer or operate SCADA systems remains a difficult challenge today (beyond dump-it-all chatbots), our framework offers a well-structured, object-oriented layer that LLMs can leverage. Our AI Operator will be piloted in a production power utility towards mid-November. It will autonomously monitor processes and assist operators with streamlining, troubleshooting, and recovering devices to improve process efficiency.

The AI Operator communicates with human operators through a dynamic AI-generated UI, where AI can bring up trends, pictograms, suggested commands and much more. LLM agents have access to real-time data, historical data, statistical analysis tools (e.g. forecasting when a signal with reach some value), can set their own schedules, and you can imagine many more options.

I hope to publish more on this soon, as I believe this pilot will be a wake-up call for many operations teams worlwide. Generative AI will only transform the world significantly through its integration into the industrial systems that keep our civilization running; the ones that keep the lights on :)) And this will require substantial efforts within the OT space. So for anyone in this OT industrial realm wanting to accelerate… let me know!

Cheers!

1 Like