The thrust of theIndustry 4.0 is leading more and more production lines to be fully sensorized: the IoT devices with which they are equipped generate Data Lake, large amounts of data that are often historicized and never used again.

How to extrapolate value from data that are in quantities unmanageable by human beings?

The answer is the Machine Learning, which makes it possible to create algorithms that automate the discovery of knowledge contained in data.

Perhaps the best known task for which Machine Learning is applied is thetraining of predictive models: by providing examples derived from the collected data, an algorithm can be trained to make predictions about an objective index, representing the event or business value that one is interested in knowing in advance.

This process enables the creation of a statistical decision support tool to be provided to domain experts.

The predictive models that can be built fall into two macro-families:

  • Models of classification: The objective index is a discrete value; it could be the occurrence or non-occurrence of a particular event to be monitored.
  • Models of regression: the target index is a continuous value, could in this case be a given process variable over time whose trend is to be predicted.

Data Mining and Machine Learning

The Machine Learning however, is not just about creating predictive models: within it is also framed the Data Mining, a set of techniques whose intent is the extrapolation of patterns or rules contained in data.

An example of these approaches is the clustering, which allows examples to be grouped into families, within which similar behaviors are recognized that might be difficult to detect with the human eye.

The two use cases best known where the potential of process data is exploited are certainly:

  • Predictive Maintenance, where the goal is to estimate the remaining life of components in order to take maintenance action on process assets before they can fail. The goal is to preserve Business Continuity, as maintenance operations have significantly less impact on plant availability than downtime due to failure.
  • Predictive Quality, where the goal is to estimate the quality of a semi-finished or finished product, where a check on it might require working under critical conditions (such as high temperatures) or where taking test samples might be unsustainable in cost or time.