PREDICTIVE MAINTENANCE AND AI FOR THE MONITORING OF WIND TURBINES
The following case study reports an example of predictive maintenance and AI for the monitoring of wind turbines.
Large production companies have the need to optimize maintenance costs, ensuring a sufficient plant reliability at the same time.
The planned maintenance approach, although often effective in avoiding sudden failures, it is not always cost-effective. Frequently, invasive control interventions or replacements involve parts that are still completely “healthy”, stopping the production without any real need.
The evolution of the maintenance approach is represented by on-condition maintenance and predictive maintenance.
These methods, already validated through specific engineering techniques such as vibration or oil analysis, are strengthened and made increasingly flexible and widely applicable thanks to Machine Learning algorithms.
Through models trained to detect anomalies, Machine Learning is able to intercept potential failures or predict them in advance, before they lead to a production stop or a plant rupture.
The customer who has benefited from the predictive solution implemented by MIPU is a large company operating in the field of energy. It has 7 wind farms with a total capacity of 368 MW gross.
MIPU has developed and implemented a solution to monitor and control the health status of the customer’s wind farm. In fact, through the modelling of different process parameters it was possible to intercept behaviour anomalies in two of the main components of the wind turbine.
Our client was directly involved in the wind turbines maintenance and was very interested in the optimization of their performances. It saw indeed the opportunity to have a solution able to intercept abnormal changes in the wind turbine behaviour.
The purpose is to be capable of anticipating failures, avoiding production downtimes and emergencies, simultaneously avoiding a rigid preventive maintenance that provides a series of complications related to the asset structure.
MIPU designs predictive models for energy monitoring and control and predictive maintenance since 2008, thanks to the development of software tools for the integration and automation of data flows and the industrial training of its engineers.
Thus, we collected the customer’s data and evaluated its potential and quality through a preliminary study of consistency and correlation, with the goal of creating AI models for the monitoring of wind turbines which could be used in all the wind farms.
Plant analysis, in collaboration with the technicians and production managers of the company, has made possible to define the main models to implement for the optimization of control and maintenance. Among these, the model for the health status control of gearbox and slip ring components has been defined.
PREDICTIVE MAINTENANCE AND AI
The algorithm applied to the model is a feedforward artificial neural network, which is structured according to simple computing units called neurons. Once connected to each other they are able to represent and generalize the behaviour of highly complex and non-linear systems.
Artificial Neural Network: calculation tool built by researchers with the aim of replicating the capacity of the human brain to acquire inputs from external environment and process information taking the context into account.
The identified models has been designed according to the parameters available to predict the health status of the two components. The status is represented by a specific parameter that is a symptom of abnormal behaviour when not aligned with the prediction.
In the diagram below the parameters identified for the model are reported in detail:
The models, statistically validated, have been applied on the historical data to verify the sensitivity related to faults or anomalies.
The gearbox model, applied to the historical data that recorded a failure on a component, highlighted a change of behaviour in parallel with the maintenance intervention, anticipated from the out-of-control monitoring tools.
Similarly, the model applied to the slip ring intercepted a second change of behaviour in conjunction with an anomaly intercepted and recorded in the database of the maintenance events.
In both cases the model has clearly intercepted with months in advance the anomalies, which then evolved into a failure.
This system, applied in real time, is therefore able to highlight risky situations that need to be managed, with sufficient anticipation to ensure the organization of the maintenance team.