Thingworx Analytics Introduction 3/3: Time Series Data TTF Prediction
This is the 3rd article of series: Thingworx Analytics introduction.
1st article introduces Analytics Function and Modules.
2nd article introduces Non Time Series TTF prediction.
Environment: Thingworx Platform 8.5, Analytics Server 8.5, Analytics Extension 8.5.
In TTF prediction, Time Series Data model is better than Non Time Series Data model.
What’s Time Series Data model?
As per Machine Learning concept, the process of training model is to find the best matching parameters, and to use them to combine with features to calculate the result values.
For Non Time Series Data models, it’s not relevant between current values and previous values.
But for Time Series Data, during each time of calculation, it will check not only current values, but also previous values, and the previous values are time relevant.
There’re 2 important terms here: look back size, and data sampling frequency.
Look back size = numbers of data samples need to look back.
Data sampling frequency = the frequency of taking a new data sample.
Look back size * Data sampling frequency = the total numbers will be feed to run calculation, for which I call it Look back area, the data is queried from Value Stream, with maxItems = look back size, and startDate = current date – look back area.
For example, if look back size = 4, data sampling frequency = 1 minute, then look back area = 4 minutes.
Thingworx used Value Stream to log Time Series Data, we can run service of QueryProperyHistory to get historic data:
Some other notes for TTF prediction models:
Always set useRedundancyFilter = true
Always set useGoalHistory = false
Always set lookahead = 1
Training time is much longer than Non Time Series data
After the model is published, test of model requires to input not only current values of all features, but also previous values as defined in look back size, than click Add Row to add the new record.
When creating Analysis Events, please be noted that the Thing Properties of Inputs Mapping should be Logged, because only Logged Properties can be queried with historic values from Value Stream.
And for Results Mapping, if we bind the result to Thing Property, even if that Property is set Logged, somehow the update by Analytics will not be logged into Value Stream. We can create another Property, and bind it with Results Mapping, and sync it with the final Property you’re monitoring by Service or Subscription, and Log it into Value Stream. After that, we can track all of the historic changes with TimeSeriesChart.
Below charts show the comparison between Time Series Data models predicted TTFs and actual TTFs.
We can see that, compare to Non Time Series models, the prediction is much more accurate, and is faster to match the actual TTF curve.
Below table lists out the settings of all models:
Another note, to make sure to write all features into Value Stream in same moment in same record, we should use UpdatePropertyValues instead of SetPropertyValues.