Control System Question -- Detecting variances in output graphs

I finally figured out where htm could help my control system.
I have a batch-oriented control system. This means that my control system executes a series of jobs rather than a continuous process. The control system has a large number of strip chart readings of various processes vs time for each job. Control systems use these plots for humans to use pattern recognition to monitor the control system. It is amazing that the trained human can detect a problem in an instant from these graphs. The goal is to use htm techniques to replace the human in the loop.

The problem breaks down as follows. There is a curve showing a variable and time. The goal of htm is as follows: At the end of the job use htm techniques to establish a baseline for the curve and detect deviations. The problem with conventional approaches is that baselines need to be maintained. Considering 100 different jobs and 10 variables, that is a 1000 manual baselines which have to be managed using conventional techniques, which is almost impossible.

If this task is a feasible approach for htm, then any papers or videos would be appreciated. Great fans of your work.


Yes it sounds fitting.

1 Like

Thanks for the reference. I read the paper and saw the video. As an aside, coming from a control theory background, a follow up paper or video on how htm was able to predict expected output and how base lines will established would be very useful.

My problem is a batch control system. The expected output is from a combination of prievious runs. If the run has N points, the there would exist N expected outputs.

The question I have is this. Can HTM theory be used in the baseline management. Also, instead of determing a pass or fail but whether there is a new trend developing. The trained human operator can determine this from the time series graphs.

Again thank you for your help

1 Like

Could you describe the input fields a bit more precisely? I think it would help to see some dummy data if possible, just to get a real tangible sense for the structure of the data.

With each new row that streams in what input fields does it contain, what do they each represent, what are their types and ranges and which need to be predicted explicitly. NuPIC’s TemporalAnomaly inf type will do anomaly detection on the system as a whole (all fields combined) but if there are certain fields you’re interested in that needs consideration.