Hi All,
My Name is Will, I am a final year college student in Milan and the Co-Founder of this project Neutralis. I came across this website super randomly this evening and am fascinated!
I will preface by stating that I am not a programmer nor an AI/ML engineer, just a an economics student with an amazing team of people around me.
This is our first project and for the past year we have meandered through the tests and trial and error required to learn and understand what we need to build but as I’m sure you all know every task is linked to an exponentially increasing number of problems needed to be studied and solved as we trudge towards the market. Since I believe we’re all in the same boat and there are like minded and far more talented people on this page, I wanted to share what were doing and shout into the void to see if we can maybe find some answers to what were looking for!
The project is this, an Intelligent fault detection diagnostics system tailored for industrial scale Heat Pump systems. If you have no bloody idea what that is don’t worry! It’s basically the standard boiler’s successor as an electrical thermal supply system which (at least in the EU) will be replacing all pre-existing systems in the coming years due to legislative changes.
The ecosystem of our product is a model of some format (question to follow) paired with a sensor suite which will be connected to an “OS” for technicians and maintainers with the goal of optimising their post-installation workflow processes.
The software is not obscure, only a little complex w.r.t double format databasing and the presence of multiple user types within an org, but this with time can be organised. The difficulty lies in the model. These systems have datapoints in the 6-7 figures and hundreds of components each requiring enough inference to be able to (with a justifiable accuracy) perform the inference required to pinpoint diagnostics, also including the multitude of ambient/external factors affecting physical systems in real time. This complexity has meant that our ML lead who is finishing his PhD is left scratching his head about what the best approach would be.
Since we would like to have a modular system to allow for any scale our first thoughts lean towards Reinforcement learning. We have a partnership in industry that is allowing us to secure vast stress test datasets from manufacturers of these systems which display the full range of results produced from these systems, but these are only from the Heat pump alone. Therefore, we are also working on gaining access to as many pilot sites as possible to collect data on entire systems so that we can cover all bases. The issue with this is that the time required to have a model viable for launch we fear would be too long and our runway is short.
Option 2 is a digital-twin. If we are able to produce a platform capable enough to align its simulations closely enough to the data collected from the sites, then only a fault library in a relational DB is required to get the desired outcome. However, as specified, to create a modular digital-twin which has internalised all of the external/ambient factors into its computing seems almost impossible. We have been simulating on one produced in Sweden, it doesn’t even come close.
Finally, we have thought instead to look into finding a highly specialised LLM which we could refine well enough to match our use case, for which our understanding is really primitive as we don’t have an AI specialist but the intuition is, produce a technician who is as experienced as humanly possible and so with just a look at the data is able to give you a step by step solution and fix guide.
What do you guys think the best course of action would be?
If you're interested in discussing further reach out at william.taylor@neutralis.it!