Spike detection

Use case:

Spike detection is useful in identifying rapid and significant changes in a series of process values that could be the sign of e.g. an outage.

This ML Model is pre-trained and can work with process variables directly from ACRON.

Working principle:

The model estimates a baseline across all values in a specified time frame, and tries to find any spikes that are caused by either abnormally large or very rapid changes.

Detailed, technical information about the IID Spike Detection algorithm can be found in the document attached or online.

A tutorial from the original author Microsoft can be found in the document attached or online.

Options and concepts:

Confidence

The confidence parameter controls the sensitivity of the model, a lower confidence score means the model interprets smaller changes as spikes.

Very low values will lead to a lot of false positives.

This option is configurable in the ACRON REST API with the "Confidence" flag. It is recommended to start with a value of 95-99.

P-Value

The probability value is an output of the model that signifies how likely a given value is a spike or outlier. 

The lower the p-value the higher the chance that the associated value is a spike.

This value will be shown as "quality" in the result of the response.

HistoryLength

The HistoryLength parameter defines the number of fixed data points in a sliding window.

The lower the HistoryLength, the faster the model "forgets" previous large spikes.

The standard value of this parameter is "-1" which is equal to 10% of the entire data set.

Any other (positive) values set, refer to fixed number of data points (e.g. "HistoryLength" : 10, means the sliding window consists of 10 fixed data points).