An incremental model learning

Incremental learning: combining the best of both worlds

Supervised learning

which consists in injecting a maximum of data as input to the learning process in order to make the model more robust to all the equipment behaviors and to increase its relevance during the recognition of failures.


which consists in learning the nominal behavior of a piece of equipment in order to subsequently detect any behavior that would deviate from this “normality”. This approach makes it possible to initiate a model very quickly, but it will not have the necessary knowledge to recognize failures.

Incremental Model Learning

Amiral Technologies takes advantage of the best of both approaches by working on dynamic learning throughout the life of the equipment. This methodology is particularly relevant in the industrial context where failure occurrences are rare and happen over the long-term.

Incremental Model Learning


Learn the nominal operation of the equipment (create a normality space).


Activate an alert when a probable anomaly occurs, and ask the business expert user to validate or reject this anomaly.


Adjust the normality space accordingly, if a new healthy behavior has not been seen before, to accommodate this new healthy operating point.


Confirm the anomaly, if the behavior is deemed unsound, and ask the user to name the fault in order to enrich the failures dictionary (see below).


Distinguish the fault and identify it automatically in the future if it is likely to occur again.


The failures dictionary

As DiagFit is enriched with anomalies, the signatures of these failures are listed in a “failures dictionary” of the equipment.

When a new anomaly appears, DiagFit is able to calculate the probability of its belonging to one or more categories of known failures. It then gives a confidence score to the user regarding this labeling.

By combining the anomaly detector and this technique, DiagFit is able to perform fault identification (recognizing what has already been seen) in parallel with failure detection on cases never identified before.

Chalk & talk

Thibaut Le Magueresse, Ph.D., Data Scientist
at Amiral Technologies

Labeling anomalies and improving a model

Because equipment lives and evolves as it operates, we at Amiral Technologies have taken the approach of focusing our work around model dynamics and incremental machine learning. ”

Play Video

Request a demo