On-line Abnormal State Identification Of Pitch System Based On Transitional Mode For Wind Turbine

Pitch system of wind turbine is often switched under different operating modes due to random fluctuation of wind speed and strong coupling between subsystems,which makes it difficult to detect the and locate faults precisely in real time.Therefore,SC AD A system in actual time has a high false alarm rate for pitch system.Aiming at the above problems,this paper proposes an abnormal status monitoring and recognition system online.First,because the contribution rate of characters to state is discrepant under different modes,the paper puts forward reduction model based on the entropy-optimized neighborhood rough set(ENRS)under different operating modes,and a mining strategy of state characteristic parameters of pitch system in all operating modes.Then,the multi-model state monitor based on a small-world particle swarm optimized entropy weighted learning vector quantization(SWPSO-Entropy LVQ)with reduction data set as input sample is constructed to realize locate faults precisely.Finally,train the above models based on actual wing field data.Simulations and test results indicate that the SWPSO-Entropy weighed LVQ based on ENRS can accurately and real-time reflect the abnormal state pattern recognition of the pitch system under transitional operating modes.The specific research is as follows:(1)Research on feature parameter mining method of pitch system based on SCADA data.Combined with the actual data of a certain wind field,under the premise of analyzing the characteristic parameters and fault correlation mode of the electric pitch system,the characteristics of the variable operating modes are analyzed.Considering the SCADA data structure feature and the random fluctuation of wind speed,the characteristic parameter mining model of pitch system based on entropy optimization neighborhood rough set(ENRS)and a mining strategy of state characteristic parameters of pitch system in all operating modes is proposed.The results show that the model can perform real-time extraction of the characteristic parameters of the electric pitch under different operating modes,and the important characteristic parameters can be filtered into the reduction set first.(2)Establish a state recognition model of pitch system based on LVQ(Learning Vector Quantization)network.For solving the problems of the model in the research of pitch system state recognition.First,according to the contribution rate of the characteristic parameters under different operating modes,an adaptive weight distance online adjustment algorithm is designed to solve the defects which magnify impact of small variables in the network distance calculation.Then,use the small world particle swarm optimization algorithm(SWPSO)which is suitable for high-dimensional solution and fast convergence to solve its initial weight sensitivity problem,and further proposes an improved state recognition model of pitch system based on SWPSO-entropy weighted LVQ.(3)Based on the working state of the pitch system detected by SCADA,various states are constructed by multi-model method.That is,a multi-model state monitor based on SWPSO-entropy-weighted LVQ basic model is constructed to solve the precise abnormal state recognition with small difference in features.The test results show that the SWSSO-entropy weighted LVQ multi-model state monitor based on ENRS can effectively identify various working states of the pitch system under variable operating conditions.(4)Apply the method proposed in this paper to the actual electric pitch SCADA data test,and compare it with the state monitor test results based on SWGA-entropy weighted LVQ,PSO-SVM and SWPSO-SVM basic model.A large number of simulation tests for variable and generalized capabilities of various models show that the accuracy of SWPSO-entropy weighted LVQ monitors is higher than other models with the prediction of fault time scale t and the increase of dataset to a certain extent.It reflects higher state recognition sensitivity and generalization ability.