Hi All, I am working on the following PoC for industrial application in mining:
n-number of sensors streaming data into the model hourly:
datetime, sensor_id, value
2016-07-07 19:00:00, 1, 2
2016-07-07 19:00:00, 2, 5
2016-07-07 20:00:00, 1, 15
The downstream system expects to receive 23 predictions for each sensor.
Nupic is used for something similar in one_hot gym tutorial; however, it does so for one sensor only. How can a model be designed and trained to consume data for multiple sensors?
This is not possible through the OPF, but you could set up your own Network with multiple classifiers to extract multiple predictions as described in this issue: https://github.com/numenta/nupic/issues/1712
Also if the language is not a concern you could use HTM.Java? As originally designed, HTM.Java classifies every field of the input and we working now to be able to specify which ones to classify (so as not to wind up classifying every one), so as a byproduct of that design oversight, you can do it with HTM.Java’s Network API. And soon will be able to limit it to the desireds only, but for now it does every one.
Thank you, then I will try to spawn a model per sensor. Another question, if you don’t mind, in one hot gym, when we run model.run with 1 step prediction, it predicts the value for the next hour? Do I have to rerun the whole sequence to get to the last prediction?
It depends on the interval. If you are sending data in at 15-minute intervals and using a 1-step ahead prediction, you’ll get a prediction that’s 15 minutes ahead. If you used a 4-step prediction it would be 60 minutes.