Four kinds of predictor
- Anomaly detection: flags what departs from the expected pattern, a vibration for example.
- Forecast: a trend with its band.
- Virtual sensor: estimates a quantity no sensor measures, from those the node reads.
- Probability: the probability that an event defined in Analisi, such as a stop, starts within a horizon. It needs Analisi.
Built from a query, in four steps
You start from a query and answer four questions: «Cosa» to estimate, «Da dove» to take the data from, «Quando» to train and apply, and an optional «Allarme». The probability kind adds the «Evento».
A predictor's versions cannot be edited. The first becomes active, later ones stay candidates until a person promotes or rejects them, and you can go back to one of the last ten retired.
From the cloud to the node on the floor
A node's program can run a predictor as a step, pinned to a version the node may use. The node page says when the pinned version is not the active one, and the predictor lists the nodes that apply it. It needs processing on the node.
A model trained elsewhere can be uploaded as a file: it is tried before it becomes a version.
What a predictor does not do
A predictor estimates a probability or flags a departure. What to do stays with whoever knows the machine, and Muvia states no accuracy. Athena can propose a predictor too, but a person approves it in «Da approvare».