Research On End-to-end Fault Diagnosis System Of Wind Turbine Gearbox Under Time-varying Speeds

The wind power industry plays a pivotal role in the national energy strategy and is of great significance in maintaining national energy security and promoting emission reduction.With the rapid development of the wind power industry,fault diagnosis technology has proven to be an important technology for ensuring high availability of wind turbines and reducing the occurrence of major accidents.Data-driven intelligent fault diagnosis systems also offer some unique advantages.For example,a large number of data samples can be analyzed in a short period of time.Online intelligent fault diagnosis systems can react promptly to early faults.They do not require a large amount of expert knowledge and avoid human interference.The gearbox is one of the core components of the wind power transmission system and is also a key component that is vulnerable to various types of faults.Therefore,the development of an intelligent fault diagnosis system for wind power gearboxes is an important means of improving the operation and maintenance capabilities of the wind power industry and meeting the industry’s requirements for refinement and efficiency.So far,from the perspective of enhancing and expanding the effectiveness and practicality of wind turbine gearbox fault diagnosis methods,there are still some obvious limitations of such technology.Firstly,the large number of pre-processing algorithms in the front-end tends to reduce the efficiency of fault diagnosis and the robustness of the system.Secondly,traditional feature extraction methods lack attention to the unique structure of wind turbine gearboxes and the vibration energy changes caused by faults.In order to break through these limitations,this paper focuses on the use of an end-to-end system form,incorporating a feature extraction method based on vibration energy variations,using a recurrent neural network(RNN)as the main framework to achieve end-to-end intelligent diagnosis of wind turbine gearboxes.In-service wind turbine gearbox vibration datasets are used to validate the performance of the proposed framework.Specifically,the following research is carried out in this paper:(1)A feature extraction method,Circular pitch cyclic vector(CPCV),is proposed for the unique structure of wind turbine gearboxes.The method captures the regular variation of circular pitch vibration energy caused by faults and attempts to establish a characteristic link between the type of fault and the meshing phase of the planetary structure,thus achieving accurate extraction of wind gearbox faults.The method first resamples the original vibration signal using a resampling algorithm based on the laser speed signal,and then calculates the vibration energy value corresponding to each circular pitch using the root mean square value to obtain CPCV.The statistical information-based feature enhancement algorithm can further improve the variability between different fault features.(2)In order to enhance the processing capability of the time-series data-based fault diagnosis algorithm for bidirectional long-range contextual information,a wind turbine gearbox fault diagnosis method based on CPCV and bidirectional gated recurrent units(Bi GRU)is proposed.Unlike conventional general feature extraction algorithms and one-way timing network models,the CPCV-Bi GRU proposed in this paper is designed with reference to the centrosymmetric structure of planetary drives and uses a bi-directional hidden gated structure to reveal the mapping relationship between CPCV features and planetary gear ring meshing phases,thus enabling the extraction of cyclic energy features from a wider range of time domains and thus improving system diagnosis accuracy.(3)To address the problems of tedious and time-consuming pre-processing and feature extraction processes and low efficiency of model calculation for wind turbine gearbox fault diagnosis,an end-to-end wind turbine gearbox fault diagnosis method based on time-step division(TSD)algorithm and fixed weight integration unit(WFIU)is proposed.In contrast to the traditional numerical interpolation method,the end-to-end diagnostic system in this paper allows the time-varying speed vibration signal to be directly input into the network and uses a time-step division algorithm to estimate the mapping relationship between the gearbox circular pitch and the vibration signal.A fixed weight integration unit is used as the end-to-end interface of the diagnostic network to achieve automatic extraction of energy cycle features of vibration signals in the form of network neurons,thus further enhancing the computational efficiency of the diagnostic model.(4)A network visualization method for RNN models based on sequence Jacobi matrix,sequence Jacobian plot,is proposed,and two typical forms of application are further proposed.The sequence Jacobian matrix is first constructed by back-propagating the output of a given output node at a given time step,and then the information contained in it is integrated and visualized in various forms through slicing or fusion.Through this method,some of the learning characteristics and patterns of vanilla RNN models are explored with the help of simulation experiments.Also,the utilization of input features by the wind turbine gearbox fault diagnosis system at different stages in the execution of the fault diagnosis task is explored using the method proposed in this paper,and some salient features of the decision-making process are summarized.