Predictive maintenance schemes are becoming a standard of maintenance in the industry. From a theoretical point of view, it is easy to understand their strengths, oriented towards intervention on equipments at the optimal time.
However, when evolving from the theoretical image to the actual implantation, the need to manage the details appears, since in reality it will never be so easy to detect the beginning of the degradation. The success of the project will depend largely on the care of the details.
This article presents those elements that, in the authors’s opinion, must be considered from the beginning of a migration process to predictive schemes.
We will present the concepts of false positive and false negative as the elements to avoid, we will mention the impossibility of eliminating both simultaneously, so the migration process should aim to achieve the most appropriate balance according to the criticality of each team.
We will refer to the need for sufficient sensorization of the installation and the importance of an adequate configuration and training process for the predictive models (this does not mean over sensoring, but reaching the best cost vs. risk balance). This training should contemplate the variability of the boundary conditions but also the different possible states of the plant and its operating strategies.
Thanks to this, we will be able to subtract, from the evolution of each variable, the part corresponding to the variability of the environment to be able to locate the moment in which the degradation processes become perceptible.
We will insist on the importance of looking for alerts that provide value to the operation through the selection, by experts in the process rather than by statistical criteria, of the maneuver bands that we will consider acceptable before activating an alert to maintenance personnel.
Finally, we will explain the importance of the explanability of the algorithm as the key to interpreting the results. Interpretation that will require support from specialized personnel through internal training or through the provision of monitoring as a service. Predictive monitoring as a service until alerts are activated, is the optimal option in the understanding of the authors, who represent Uniper and Tecnatom, and who at the time of writing this article are monitoring more than 35 GW of various generation technologies.
Finally, we will present three success stories of predictive maintenance applied to the electricity generation industry, insisting on its ability to detect degradation in its initial stages and appreciably before the activation of alarms in the control room. Not only that, but we will also mention its ability to detect unexpected operation in equipment due to sudden failures.