Traditionally, in industrial environments maintenance processes have so far been carried out based on rules or advice from the supplier of industrial units. More recently, a conversion to a data driven approach is underway in these ecosystems to take advantage of the enormous amount of data being generated.
In the case of predictive maintenance use there is not much data on industrial units that have suffered failures (thanks to these maintenance policies). For this reason, an unsupervised approach has been chosen, under the premise that in general, the normal behavior of a machine should not vary excessively, and any sufficiently large changes in the sensor data should indicate a deviation to evaluate.
The proposed model compares the different measurements with each other and then infer whether a measurement is an anomaly or not based on how far is this one from the rest of the data. This solution provides Enagás a model for every pump helping their technical team to evaluate the functioning of every pump individually.
Finally, we have developed an API so that the trainings, inferences or updates of the models can be performed easily by business users.