Vibration Signal Analysis And Fault Diagnosis Of Wind Turbine Gearbox Based On Order Analysis

The working environment of wind farm for the wind turbine is so harsh. Due to the doubleimpact of the environment and its own structure, the external exciting force and vibrationdegrees of freedom is much higher than other heavy machinery. The volume of wind turbine,which increase the probability of failure, will be increased with the addition of unit capacity. Thewind turbine`s running state can directly affect the benefits of the wind farm. Therefore, it is veryimportant to develop a online condition monitoring and fault diagnosis system.The main failures of wind turbine are caused by blades, shaft, gearbox, generator and so on.Gearbox has the highest fault rate and cause the longest shutting down time. In order to improvethe operation efficiency and reduce maintenance costs, this paper focuses on the wind turbinegearbox vibration analysis and fault diagnosis, mainly including the following aspects:First of all, wind turbine often run under the complex conditions, vibration signal exhibits anon-stationary characteristics. The traditional methods are difficult to meet the requirement ofonline condition monitoring. An Even-angle re-sampling order analysis method, which caneliminate the impact of the speed fluctuation and effectively extract the fault feature, is proposedin this paper.Secondly, In the resampling process, two half-band filters and a CIC filter were used tointerpolate values between tacho pulses to obtain the even-angle phasing time scale. Accuracy ofresampling was improved. Compared with the classic calculation order tracking algorithm, theeffectiveness of the proposed algorithm was verified.Finally, a fault diagnosis system based on support vector machine was build，which usefault features extracted by order analysis. Meanwhile, a feature selection and parameteroptimization method of Support Vector Machines(SVM) based on Fruit Fly OptimizationAlgorithm is proposed. Imitating the foraging behavior of fruit flies, the smell concentrationjudgment value is used as the parameter and the feature set is binary-encoded to generate thesubset which can be used to train the model. Then a proper fitness function is constructed tosearch the best parameters and feature subset. The test shows the proposed approach has higherprecision on classification and global search ability compared with other methods.