Predictive maintenance of wind farms

Reducing losses with machine learning: Predictive maintenance of wind farms


This article was originally published in Bulletin.ch magazine. Proxima Solutions GmbH has the consent to produce the English and Spanish versions and to use the images.

Machine learning is used to identify, segment, and classify temporal signals produced by energy systems in order to detect failure risks. A tool based on this technique has been developed and implemented to optimize the predictive maintenance of a large number of wind power assets.

New decentralized energy sources (wind turbines, photovoltaic panels, batteries) pose many operational challenges [1]. However, some of these challenges can be overcome or simplified thanks to the latest technological advances in data acquisition and processing. This is particularly the case for the maintenance process, where maintenance solutions [2] can be implemented to increase the availability of the installations.

In collaboration with its partners, the BKW Group and Proxima Solutions, CSEM has recently developed a software solution for the predictive maintenance of wind power assets. Based on machine learning, this solution makes it possible to improve the preventive and corrective maintenance process, and thus reduce operating losses linked to wind turbine failures.

Predictive maintenance facing the challenge of uncertainty

For a wide range of systems, functional failures are not spontaneous and can be anticipated. Degradation is often progressive (Figure 1) and manifests itself in the form of anomalies in the measured signals [3]. This observation constitutes the basis for predictive maintenance.

performance-gradual-degradation

With a good set of sensors and a detailed analysis of the data produced, it is thus possible, in theory, to predict the failures of many systems and to send maintenance teams on site in a sufficient time to minimize operational losses. The detectability threshold will obviously depend on the sensors installed. Vibration sensors, for example, are indeed capable of detecting potential failures of rotating parts before temperature sensors. [4]

For complex systems such as wind turbines, however, this theory faces many practical problems. First of all, significant variability of the measured signals (temperature, pressure, vibration) can be observed under similar conditions for wind turbines of the same model, produced by the same manufacturer. Secondly, maintenance operations are themselves at the origin of a major problem. Wind turbines in operation are not isolated system. Numerous interventions - preventive or corrective - take place regularly on site. At the same time, sensors and control systems are updated remotely and regularly. All these disturbances influence the measured values, complicate the implementation of automatic detection solutions and require feedback on the tasks performed in order to exclude false alarms as much as possible.

Targeting alerts

The tool developed during this project aims to solve the above-mentioned problems by relying on three pillars (Figure 2).

three-pillar-solution

The first pillar corresponds to the use of a machine learning algorithm to predict the measurable behavior of each turbine, individually, from its historical data. For this purpose, various data are collected: the data transmitted by the Scada system (temperature, pressure, angle of attack, power) with a resolution of 10 minutes as well as the data from the vibration sensors. The algorithms used, based on neural networks [5], are transferable from one turbine to another as long as they share the same signals. They can also be quickly retrained for a turbine or a group of turbines. A physical approach was used to choose the architecture of these neural networks and to link the external parameters to the measured signals, as well as to take into account the latencies observed in the temperature and pressure signals.

The second pillar corresponds to the comparison of the algorithm's predictions with the on-site observations, that is, determining the model's error. In order to limit false positives, significant work has been done on the processing of this error, including statistical analysis and time-series segmentation. Through this process, each identified deviation is classified according to its severity.

Finally, the last pillar corresponds to the man-machine interface. Even with a very fine sorting, one or several important deviations do not necessarily announce a serious failure. Before presenting the anomalies to the user, the deviations are aggregated by the turbine and a list of potential failures is automatically generated. This list is based on a Failure Mode and Effects Analysis (FMEA) performed with turbine and maintenance experts. The control room analyst can then sort the list, modify the severity scale of the entries, comment on the anomalies and generate intervention requests for the technical teams (Figure 3).

on-site-interventions

In order to keep track of all the information of the maintenance process, each anomaly included in the intervention report must be closed later by the control room manager, specifying the failure found on site by the teams as well as the corrective task performed. This information is saved and then used in a feedback loop to re-train the algorithms, reduce the error rate and increase the accuracy of the diagnosis.

Forty anticipated failures in one year

After a promising test phase, the software tool was put into operation in the BKW control room and is now continuously monitoring more than 200 turbines in Italy, Germany, France, and Switzerland. Since mid-2020, about 40 failures have been anticipated, which has made it possible to avoid production losses and increase the availability of facilities. The failures identified (Figure 4) cover a wide spectrum: minor problems such as contaminated filters, defective fans, insufficient bearing grease, but also medium-sized failures such as defective thermostatic valves in gearboxes, or major failures such as defective generator bearings or large holes or scratches in gearbox transmission elements.

detected-by-algorythms-failures

The involvement of technicians is essential

Various technical and organizational factors are essential for the successful implementation of a predictive maintenance strategy for distributed energy systems. The main technical criterion motivating forward-looking maintenance is the increase in availability achieved through the detection of anomalies in relation to an “until failure” strategy, whether for wind turbines, solar panels, or batteries. Involving maintenance technicians constitutes the main essential element from an organizational point of view. As the continuous improvement of the tool relies on a feedback loop aimed at collecting as much information as possible on the repairs carried out, technicians are an indispensable link in the chain. They must therefore be convinced of the value of this solution and of the importance of their feedback. Thus, the tool will not be perceived as an administrative constraint, but rather as an intervention and decision-making aid in order to guarantee optimal performance of the assets.

References:

  • [1] National Energy Technology Laboratory, «Electric Power System Asset Optimization», NETL Report, US Department of Energy, 2011.
  • [2] Norme NF EN 13306:2001 «Terminologie de la maintenance», 2001.
  • [3] P. Tchakoua, R. Wamkeue, M. Ouhrouche, F. Slaoui-Hasnaoui, T. A. Tameghe, G. Ekemb, «Wind Turbine Condition Monitoring: State-of-the-Art Review, New Trends, and Future Challenges», Energies 2014, 7(4), pp. 2595-2630, 2014.
  • [4] K. Fischer, F. Besnard, L. Bertling, «Reliability-Centered Maintenance for Wind Turbines Based on Statistical Analysis and Practical Experience», IEEE Transactions on Energy Conversion, Vol. 27, no. 1, pp. 184-195, March 2012.
  • [5] I. Goodfellow, Y. Bengio, A. Courville, Deep learning, The MIT Press, 2016, ISBN: 0262035618.

Acknowledgments

The authors would like to thank the following collaborators from the various European subsidiaries of the BKW group for the valuable feedback and sharing their expertise on wind technologies: Arthur Chevalier, Francesco Piersanti, Michele Cacciacarro, Frank Krause, Danilo Grande, Giuseppe Zazzera, Gianpasquale Gambacorta, and Alejandro Sampedro Senen.