Back to the programme printer.gif Print

Delegates are invited to meet and discuss with the poster presenters in this topic directly after the session 'Innovative concepts for drive train components' taking place on Thursday, 13 March 2014 at 11:15-12:45. The meet-the-authors will take place in the poster area.

David Infield University of Strathclyde, United Kingdom
yue wang (1) F P david infield (1)
(1) University of Strathclyde, Glasgow, United Kingdom

Printer friendly version: printer.gif Print

Presenter's biography

Biographies are supplied directly by presenters at EWEA 2014 and are published here unedited

Prof. David Infield has been working in the wind energy field for over 20 years. From 1993 to 2007 he built up CREST (Centre for Renewable Energy Systems Technology) at Loughborough University. In 2007 he moved to Strathclyde as Professor of Renewable Energy Technology. His wind research interests cover resource assessment, flow modelling and condition monitoring. He has published widely, led National and European research teams, and is the founding Editor in Chief of the IET Journal Renewable Power Generation. He coordinates the renewable energy area for the UKERC and sits on the Royal Society Energy Committee.


Multi-machine based wind turbine gearbox condition monitoring using nonlinear state estimation technique


Wind turbine gearbox reliability remains one of the biggest concerns for the wind industry [1]. The total costs involved in hiring the maintenance vessel and associated personnel will be considerable for offshore turbines, plus the revenue lost during the downtime would also be more than for onshore turbines due to the possible delay caused by the severe sea condition. All of these indicate that an effective condition monitoring method for wind turbine gearboxes that is able to identify faults at an early stage would therefore be beneficial for reducing turbine downtime and improving the availability.


Reference [2] presents successful anomaly detections of an individual turbine gearbox using a Nonlinear State Estimation Technique (NSET), where a memory matrix is used to store the representative historical state vectors of the model-correlated variables (with gearbox cooling oil temperature being the condition indicator), based on which the model estimations are calculated and the subsequent anomaly detections are performed. Individual turbine based anomaly detection for a large wind farm will involve extensive data analysis and this could make such an approach impracticable. The development of spatial data-mining techniques that can analysis the SCADA data from a part of a wind farm, or even the complete wind farm with large numbers of identical turbines can dramatically reduce the number of models required and significantly reduce the effort of repeated model construction for each individual turbine. To this end, the NSET method has been extended to multiple turbines in a spatial context, based to an extent on successful application to an individual turbine as in reference [2].

In the multi-turbine model the state vectors in memory matrix include records of the gearbox oil temperature and other associated variables from a group of turbines that are closely located within the wind farm, and thus experience similar wind conditions and operational status. The model outputs are estimations of the relevant variables for all the turbine members within the group, which saves considerable effort as well as computational complexity compared to the repetitive development of individual turbine models. The Welch’s hypothesis test is used for the subsequent anomaly detection.

Main body of abstract

The selection of the turbine group for model construction here relies on the correlation of wind speed values between a selected turbine and those immediately surrounding it. Turbine T16, which is known to have gearbox operational problems, is used as the basis of sub-group selection for the following analyses. The two turbines of interest, T16 and T17, are always included in the sub-group due to their proximity and high correlation with each other.

The spatial model that involves sub-groups of different sizes and also the complete wind farm will be presented later. As a start here, the model candidates are selected based on the wind speed correlations. Turbines T10, T11, T13, T14 and T17 all have nacelle anemometer correlation with T16 in excess of 0.96, so they are selected to be in the sub-group, leading to a sub-group size of 6. It can be seen from the wind farm layout illustrated by Figure that the selected turbines are geographically close to each other since the correlation of wind speed, used as the selection basis, is in general inversely related to the distance between measurement locations.

Three variables including gearbox oil temperature, power output and nacelle temperature are used as model inputs. In contrast to the individual NSET model where the data is filtered only based on the turbine of interest, the data filtration in this case needs to take the selected sub-group into consideration, i.e. the simultaneous measurements for the group of turbines would become disqualified in case any one of them shows an invalid entry for the selected variables. An optimal state number is determined to be around 1000 based on the compromise between model accuracy and computational complexity, resulting in a memory matrix size of 18×1000. The modified NSET model (in which the indicative variable is omitted from the estimation process for the weighting vector, the gearbox cooling oil temperature in this case) is believed to produce more reliable results than the standard model in which the indicative parameter is retained in the weighting calculation. The results are based on the measurements from valid sensors only. Also the state vectors in the memory matrix are normalised in order to get rid of the bias due to the differences in scale of the model inputs, and the variables are assigned with different weights according to their correlations with the variable of interest. Validation and testing results for the modified spatial NSET model are illustrated in Figures , and , which show the excellent model accuracy and detectability for this 6-turbine sub-group.

The effect of sub-group size on the model performance is investigated by varying the sub-group size through modification of the correlation threshold level. The state number remains approximately the same as in the 6-turbine sub-group while the group member increases. Three cases including sub-group size of 10, 15 and 17 are explored and the associated residual statistics are summarised in Table , where the detection ratio (DR) indicates the proportion of the testing instances being detected out of the entire testing population. The statistics in this table present satisfactory model performance in general and it also shows little impact of the sub-group size on the model performance, which implies great potential to extend the spatial model to even larger wind farms so that modelling efforts can be saved without losing the model reliability.

It also can be seen from Table 1 that the spatial model beats the individual turbine model in terms of detection sensitivity but the turbine based model shows higher accuracy.


This paper demonstrates the effectiveness of the spatial NSET model with normalised matrix and weighted parameter, including oil temperature, power output and nacelle temperature as the model inputs. A reasonable state number value of around 1000 is selected for the sub-groups with different sizes based on the compromise between model accuracy and computational complexity. The modified NSET model adds confidence in the model’s reliability by omitting the indicative variable of the relevant turbine from the estimation process for the weighting vector, which in effect results in a model with an auto-sensitivity of zero. The limited impact of sub-group size on model performance indicates the model’s ability to extend and more importantly significant amount of efforts would be saved by wind farm based condition monitoring without losing model effectiveness.

However, the data filtration could lead to failures of anomaly detection due to either predictions based on faulty period of anomalous turbines or inclusion of too many turbines with poor data quality in the sub-group such that the data containing anomalous information are likely to be removed in the filtration process due to invalid entries in other turbine candidates. This indicates that the feasibility of the spatial model implementation largely depends upon the data quality. More case studies need to be done before widely applying the spatial NSET model for entire wind farm condition monitoring.

With regards to the performance comparison between the individual turbine based and the multi-machine NSET model, the individual turbine model has higher accuracy but the spatial model shows more capability for anomaly detection. Therefore no definite conclusions can be come to at this stage.

Learning objectives
In contrast to traditional wind turbine based condition monitoring, in which the implementation procedures need to be repeated for each individual turbine, the proposed multi-machine-based method for estimating for the relevant variable from neighbouring turbines is more effective. This is possible because the turbines experience similar wind profiles and hence share similar operational conditions with each other.

1. C.J. Crabtree, Y. Feng and P.J. Tavner, 2010, Detecting Incipient Wind Turbine Gearbox Failure: a Signal Analysis Method for On-line Condition Monitoring, European Wind Energy Conference, Warsaw, Poland.
2. Y. Wang, D. Infield, Supervisory control and data acquisition data-based non-linear state estimation technique for wind turbine gearbox condition monitoring, IET Renewable Power Generation, Vol. 7, (4), pp. 350-358, 2013.