How To

Health Monitoring - How To

The Health Monitoring section aims to automatically recognize abnormal behavior of machine and line components/stations through software that, by exploiting machine learning algorithms, is able to process the data produced by them and evaluate their conditions.

The trained algorithms follow an "Anomaly Detection" approach: the algorithm, trained to recognize on the basis of "normal" operation data, is able to identify any anomalous patterns, unexpected with respect to normality, even in the absence of direct experience (unsupervised approach).

The following are the steps that operationally lead to the implementation of the Health Monitoring system.

1) Identification of components/stations

Identification of components/stations is possible adopting the following criteria:

• Critical to production

• Subject to breakage and maintenance fairly frequently

• Wear condition difficult to interpret

2) Selection of variables

For each component/station, the most suitable, informative variables should be selected to know the status of the component/station, making use of the manufacturer's know-how:

• Process variables: pressures, currents, positions, temperatures, etc

• Recipe/setpoint variables

• Production environment variables (e.g., ambient temperature)

3) Extraction of "expert-based" features

Mathematical operations to be performed on the variables to obtain meaningful information for assessing the health status of the component, applied during episode evaluation.

These features encode the manufacturer's know-how; they serve as input to the machine learning algorithm, helping the latter to more easily (from a computational and logical point of view) identify the normality of the component's operation.

4) Preliminary data collection

Preliminary data collection for training the Anomaly Detection algorithm, performed autonomously by the system at the EDGE level. The first N. runs of the components in their operational context are performed, then the system performs training of the machine learning models.

At the end of these steps, the training of the Anomaly Detection algorithms is completed and the system is able to evaluate subsequent episodes, providing as output the KPI health status of the component, expressed in percentage form.

Should the health status fall below a certain self-determined threshold (e.g. 50%), the system will register an abnormal episode and notify the appropriate personnel.

This "hybrid" approach, the result of combining machine learning models and "expert-based" features, is called the "grey box approach" and offers the following advantages in summary:

• Encoding the machine/plant manufacturer's know-how through features; the algorithm does not have to "learn" on its own what the manufacturer has learned over decades of operating experience.

• Computational efficiency, with the ability to perform algorithm training directly on low-capacity, low-cost EDGE units.

• Scalability to additional specimens or machine/plant models (and related components/stations) due to the autonomous training process

By "episodes" we mean time sub-windows of data (cycles, minutes of work, periods of time in particular machine states...). The choice of episodes is critical to facilitate learning of the algorithm.