TY - GEN

T1 - Condition monitoring with wind turbine SCADA data using Neuro-Fuzzy normal behavior models

A1 - Schlechtingen,Meik

A1 - Santos,Ilmar

AU - Schlechtingen,Meik

AU - Santos,Ilmar

PB - American Society of Mechanical Engineers

PY - 2012

Y1 - 2012

N2 - This paper presents the latest research results of a project that focuses on normal behavior models for condition monitoring of wind turbines and their components, via ordinary Supervisory Control And Data Acquisition (SCADA) data. In this machine learning approach Adaptive Neuro-Fuzzy Interference System (ANFIS) models are employed to learn the normal behavior in a training phase, where the component condition can be considered healthy. In the application phase the trained models are applied to predict the target signals, e.g. temperatures, pressures, currents, power output, etc. <br/>The behavior of the prediction error is used as an indicator for normal and abnormal behavior, with respect to the learned behavior. The advantage of this approach is that the prediction error is widely decoupled from the typical fluctuations of the SCADA data caused by the different turbine operational modes. <br/>To classify the component condition Fuzzy Interference System (FIS) structures are used. Based on rules that are established with the prediction error behavior during faults previously experienced and generic rules, the FIS outputs the component condition (green, yellow and red). Furthermore a first diagnosis of the root cause is given. In case of fault patterns earlier unseen the generic rules allow general statements about the signal behavior which highlight the anomaly. <br/>Within the current research project this method is applied to 18 onshore turbines of the 2 MW class operating since April 2009. First results show that the proposed method is well suited to closely monitor a large variety of signals, identify anomalies and correctly classify the component condition. The accuracy of the normal behavior models developed is high and small signal behavior changes become recognizable. The result of the automatic analysis is given in graphical and text format. Within the paper examples of real faults are provided, showing the capabilities of the method proposed. The method can be applied both to existing and new built turbines without the need of any additional hardware installation or manufacturers input.

AB - This paper presents the latest research results of a project that focuses on normal behavior models for condition monitoring of wind turbines and their components, via ordinary Supervisory Control And Data Acquisition (SCADA) data. In this machine learning approach Adaptive Neuro-Fuzzy Interference System (ANFIS) models are employed to learn the normal behavior in a training phase, where the component condition can be considered healthy. In the application phase the trained models are applied to predict the target signals, e.g. temperatures, pressures, currents, power output, etc. <br/>The behavior of the prediction error is used as an indicator for normal and abnormal behavior, with respect to the learned behavior. The advantage of this approach is that the prediction error is widely decoupled from the typical fluctuations of the SCADA data caused by the different turbine operational modes. <br/>To classify the component condition Fuzzy Interference System (FIS) structures are used. Based on rules that are established with the prediction error behavior during faults previously experienced and generic rules, the FIS outputs the component condition (green, yellow and red). Furthermore a first diagnosis of the root cause is given. In case of fault patterns earlier unseen the generic rules allow general statements about the signal behavior which highlight the anomaly. <br/>Within the current research project this method is applied to 18 onshore turbines of the 2 MW class operating since April 2009. First results show that the proposed method is well suited to closely monitor a large variety of signals, identify anomalies and correctly classify the component condition. The accuracy of the normal behavior models developed is high and small signal behavior changes become recognizable. The result of the automatic analysis is given in graphical and text format. Within the paper examples of real faults are provided, showing the capabilities of the method proposed. The method can be applied both to existing and new built turbines without the need of any additional hardware installation or manufacturers input.

BT - Proceedings of ASME Turbo Expo 2012

T2 - Proceedings of ASME Turbo Expo 2012

SP - GT2012-68011

ER -