Research On Fault Diagnosis Of The Key Componets Of Wind Turbine Based On Deep Belief Network

In the past twenty years,the wind power industry in China has developed rapidly.However,under the condition of insufficient operation and maintenance,the main components frequently fail,which challenges the reliability and durability of the system.Generator and gearbox are two key devices of wind turbines,which have a long downtime.If grasping the incipient failure information of generator and gearbox in time,we can take early measures and arrange maintenance plan,which is of great significance for ensuring economic benefits.Aiming at the fault characteristics including asynchronous generator's stator inter-turn short circuit,rotor broken bars and gearbox's uniaxial and biaxial broken tooth,its fault extraction problem is studied respectively.Feature extraction and pattern recognition of generator's stator current signal and multimodal signals composed of vibration signals and current signals are used for fault diagnosis of deep belief network.The incipient fault models of wind turbine are built by Simulink and its correctness is verified.The fault experiments of gear breaking are carried out by building a fan experimental platform.The effectiveness of the proposed methods are verified.The main work of this paper is as follows:Firstly,the physical model of squirrel-cage induction generator was analyzed in this paper.A motor model under normal working condition is established.Through multi-loop theory,the stator winding inter-turn short circuit and rotor bar broken fault are simulated by external resistance mode.Secondly,hierarchical deep belief network is introduced to detect the incipient fault of generator.A deep belief network can directly deal with time domain signal,which do not rely on technology and expert experience.Stator current signal of wind generator is used for feature extraction and classification recognition.The goal is to diagnose faults of different types and different fault levels.Then hierarchical deep belief network is compared with traditional shallow machine learning methods to show its good performance.Thirdly,a method of feature extraction combining multi-channel signal of current signal and vibration signal is studied.Considering the actual condition is very complex,it is difficult to obtain a single signal source characteristics that have high quality.A code and fusion feature extracted method based on multimodal deep belief network is proposed to realize gearbox fault diagnosis.It is compared with single signal source,data level fusion and traditional shallow classification network to prove the effectiveness and advantage of the proposed method.