Answering your questions from Beyond the Line Pt. 3!
Hi everybody! James Newman, Head of Product & Portfolio Marketing at Augury here. On June 18th we shared how an AI-driven factory and Augury’s latest innovations make next-level productivity possible in our special event, Beyond the Line. Catch the full event on-demand here.
There were a few questions we didn’t have time to get to in the Q&A portion of the event that I’d like to to answer here on The Endpoint community. Last week we answered the question, “How do you manage to collect all the operational and process data which are beyond the mechanical dimension such as temp. and vibration?”
Today let’s answer this question:
What are the most promising use cases for Generative AI you see in the near future? Also could you elaborate on some problems that may hinder the use of AI for maintenance at the moment?
For the most part, Generative AI will have a significant impact on internal processes first.
- We are already leveraging Gen AI in our internal process to relabel data based on user and expert feedback, using it to mine our internal data model, using it to accelerate how we rebuild AI models more rapidly, etc.
- Gen AI will also allow us to integrate more rapidly new sensors with high quality data and use those sensors with our existing models and Gen AI capabilities. That isn’t to say that any sensor still works, the sensor will still have to meet minimum data quality standards.
- Lastly, Gen AI allows us (for the first time) to learn from direct interaction with customers and be able to tailor (when required) to specific machines, based on customer interaction and feedback with the system.
For the foreseeable future, these internal use cases are the most likely “trustable” methods for using Gen AI in manufacturing. Use of Gen AI requires the underlying large language model to be very tightly controlled to avoid biases and hallucinations, which will require time to be able to be used in day-to-day operations within manufacturing.
More generically, the limitations to using AI have not changed with the introduction of Gen AI, despite the difference in hype and publicity. The challenge for AI has always been the same, having trustable data sets which can be trained and tested and validated to ensure that the right accuracy, recall and precision can be obtained which makes the result of an algorithm, query or search to be reliable enough to be taken as truth. Generative AI has made that need more visible, but in reality, it has always been and will always be about the quality of the data that is used to train and leverage AI. It’s why Augury made the decision to not be an “AI-platform” but to build purpose-built AI based on specific use cases. That has allowed us to build a data library (now in excess of 400M hours of machine data) for training and re-training models to drive ever increasing accuracy. It’s also why we are able to push the boundaries of what’s possible with vibration analysis, because of the trusted data source underlying the algorithms, we can then explore new use cases and innovative techniques which allow us to apply new methodologies to create trustable outcomes for “non-standard machines” like low RPM or variable speed machines. That underlying data library is also why we can deploy Gen AI internally, as we already know the quality of the library, but also we can use Gen AI capabilities to improve it, based on feedback from our internal experts and customer feedback.
All that to say, regardless of the “type” of AI deployed, the issue comes down to three things: is the data library large enough to build and test the Ai to be trustable? Is the model able to learn from feedback from the field to continually improve the results? And lastly, and most importantly, does the underlying data model consist of high quality data and is the model being managed well enough to ensure that it doesn’t introduce false results and create answers that don’t have a basis in real data? It reinforces the need for purpose-built AI models to drive progress as often as possible, rather than creating multiple models on the fly without the level of control required to ensure accuracy and trustability is maintained.
Have more questions? Drop a comment below and let’s chat.