First thank you for an interesting project! I’m exploring the use of Monty for time-series anomaly detection in an industrial setting. I have 3 years of sensor data (1-minute resolution, 50 features), all recorded from an industrial production system.
Together with that I also have access to a SCADA-like interface (see attached image for a similar example found on the internet), which shows the 2D spatial layout of each sensor in the system — pipes, tanks, valves, etc.
Would it make sense to simulate sensorimotor exploration in Monty by:
Treating each feature (sensor) as a “location” in a virtual 2D space,
Using the SCADA layout as a sensor topology map,
And allowing Monty to “move” across sensors (i.e., feature subsets) as if it were navigating the plant?
This way, my hope is that Monty could build object-like representations and learn sensor relationships spatially and temporally, even though the data is historical and fixed. I would simulate a movement policy (e.g., random walk or localized scanning) over the SCADA-defined positions.
Is this aligned with Monty’s intended use? Has anyone tried something similar with non-visual, tabular time series? Any clear obstacles that will inhibit this, or any advice on how to best structure this “virtual movement”?
Hi @brainwaves, and thank you for your reply and for sharing the related thread. I found the discussion on generalizing TBT models very insightful, especially the perspectives on extending it beyond traditional sensory modalities like vision or touch.
That said, I am still curious about how well my idea aligns with Monty’s current capabilities in practice. In particular, I am wondering whether there is any precedent or known limitation around using Monty in a virtual sensory space. My idea is to map features from a time-series dataset to two-dimensional spatial coordinates, such as sensor positions in a SCADA layout, and allow Monty to “navigate” this space to learn both temporal and spatial relationships. This would be done even though the dataset is historical and not interactive.
The overall goal is to apply Monty’s sensorimotor framework to build internal representations of system behavior and to detect anomalies in an industrial process. In essence, I am exploring whether TBT can be used in a simulated context where the structure of the sensor space is known.
If anyone has attempted something similar, perhaps using non-visual data or simulating feature-wise exploration, I would really appreciate hearing about your experience. I would also be interested in any reasons this might not work well, if there are specific limitations I should be aware of.
welcome to the TBP forum! I’m excited to see that you are thinking of using Monty for such a real-world application.
Generally I would say you have the right approach of letting Monty know about the relative location of sensors and measurements and letting it learn a structured model that way. It’s certainly possible for Monty to learn from such an already recorded dataset if there is a way for it to infer structural relationships. I am not an expert with SCADA so my biggest question would be: Does the relative arrangement matter for the measurements? Does it inform the predictions you would make? If so, then I think Monty could be used in this way.
However, there is one big caveat: Monty today can’t model object behaviors yet. So far, it only models static objects. It models them through sensorimotor exploration and recognizes them by moving sensors over them, but the models themselves don’t contain a temporal dimension. So right now, you would not be able to use Monty for an application where you want to model how an object/system changes over time (which I think your application requires).
We are actively working on figuring out how to model object behaviors (in the brain and in Monty) but right now we are still in the conceptual stage and it will still take a while until we will have a functioning implementation of it in Monty.
If you are interested in following along, you can watch our brainstorming sessions on object behaviors (see this playlist https://www.youtube.com/watch?v=_lMffztgr8w&list=PLXpTU6oIscrlBOE69gkxleAeAOS1Q6aZB) on YouTube. We’ve recently made a lot of progress on this and I can let you know once those videos of our latest thinking are posted. However, this is probably just if you are super curious. It won’t help with your application right now. I hope you’ll still stay interested in the project and follow along until this fundamental capability is implemented. Until then, you could have a look at Numenta’s earlier work on HTM for modeling time series and anomaly detection HTM | Numenta
