Fault Detection of the Wind Turbine Variable Pitch System Based on Large Margin Distribution Machine Optimized by the State Transition Algorithm

Mingzhu Tang,Zijie Kuang,Qi Zhao,Jiahao Hu,Huawei Wu,Shuhao Peng,Yong Chen

doi:10.1155/2020/9718345

Abstract

Aiming at solving the problem that the parameters of a fault detection model are difficult to be optimized, the paper proposes the fault detection of the wind turbine variable pitch system based on large margin distribution machine (LDM) which is optimized by the state transition algorithm (STA). By setting the three parameters of the LDM model as a three-dimensional vector which was searched by STA, by using the accuracy of fault detection model as the fitness function of STA, and by adopting the four state transformation operators of STA to carry out global search in the form of point, line, surface, and sphere in the search space, the global optimal parameters of LDM fault detection model are obtained and used to train the model. Compared with the grid search (GS) method, particle swarm optimization (PSO) algorithm, and genetic algorithm (GA), the proposed model method has lower false positive rate (FPR) and false negative rate (FNR) in the fault detection of wind turbine variable pitch system in a real wind farm.

Highlights

With the rapid development of the wind power industry, the installed capacity and quantity of wind turbines are growing continuously. e World Wind Energy Association predicts that by the end of 2020, the global installed capacity will reach 1.9 × 106 mw [1]
According to the different fault detection of the wind turbine variable pitch system, the sample set in normal operation should be classified as normal, and the sample set in failure should be classified as a fault
Value range e trade-off parameter of margin variance, which is adopted to adjust the weight of margin variance e trade-off parameter of margin mean, which is adopted to adjust the weight of margin mean e loss function parameter, which is adopted to adjust the weight of the loss function in the objective function

Summary

Introduction

With the rapid development of the wind power industry, the installed capacity and quantity of wind turbines are growing continuously. e World Wind Energy Association predicts that by the end of 2020, the global installed capacity will reach 1.9 × 106 mw [1]. SVM is another classic data-driven model based on global optimization It has good performance and can solve the problems of multiclassification recognition and regression prediction [11, 12], which has been widely recognized in the field of wind turbine fault research. E data-driven fault detection model has good practicability in actual wind turbine fault detection and fault diagnosis Most of these models depend on the selection of parameters, so it is necessary to use the parameter optimization algorithm to quickly and accurately find the global optimal model parameters. E STA is a parameter optimization algorithm with four state transition operators, facing the complex fault detection problem; the global optimal value can be quickly and accurately found by the four state transformation operators alternately, which is suitable for detecting the complex fault of the wind turbine variable pitch system. In the fault detection model based on machine learning, parameter optimization is an important process, and how to select appropriate parameters to enable the detection model to meet the fault detection standard of the wind turbine variable pitch system is the key and difficult problem of all machine learning models. erefore, an improved LDM model based on the STA is studied with an aim to effectively finding out the optimal model parameters, making it meet the fault characteristics of the variable pitch system, and improving the accuracy of fault detection

Large Margin Distribution Machine

The State Transition Algorithm

Large Margin Distribution Machine Optimized by the State Transition Algorithm

Conclusion