Fault Diagnosis Method Of Wind Turbine Variable Pitch System With Intelligent Optimization Parallel Filtering

Accurate and timely identification of faults in the wind turbine pitch system operation is of great importance to ensure the safe operation of the wind turbine system and reduce or avoid major catastrophic accidents.Modern wind turbines are characterized by large scale,high integration,complex operation mechanism,the uncertainty of parameters and structure,etc.,accompanied by modeling errors,noise disturbance,and high complexity of the algorithm,which makes the requirement of fault diagnosis of wind turbine pitch system highly difficult to increase.How to improve the accuracy and real-time of fault diagnosis of wind turbine pitch system is a key problem in the field of fault diagnosis that needs to be solved urgently.In this research,we take the pitch system as the research object and investigate fault diagnosis accuracy and real-time performance.Innovative research ideas are proposed for the accuracy and real-time performance of multi-model fault diagnosis algorithm and fault diagnosis of hybrid intelligence,which are important for the research of nonlinear filtering algorithms in the field of fault diagnosis.The main research contents of the paper are as follows:(1)Two improved resampling filtering methods are proposed to address the problems of degradation of PF estimation accuracy and computation time consumption caused by particle degradation and sample scarcity,which lead to the poor accuracy of the particle filtering fault diagnosis method.In this thesis,we propose two improved methods based on parallel differential evolution resampling filtering algorithm and parallel multi-prediction resampling filtering using two methods: parallel differential evolution and multi-prediction as the breakthrough point of resampling filtering.In the state estimation method based on parallel differential evolutionary resampling filtering,the segmented prefix and the improved resampling can update the proposed distribution of particle filtering in real-time and adaptively adjust the number of particles to be sampled by particle filtering to a smaller number.In addition,for the problem of inefficient thread execution in the prefix and execution of the particle filtering algorithm execution;the algorithm improves the efficiency of the algorithm thread execution by removing the thread bundle differentiation;based on the parallel multi-prediction resampling filtering method,the multi-prediction method can reduce the number of particles required for the target estimation accuracy and the sequential operation of the resampling part;thus solving the execution efficiency problem of the resampling part and improving the filtering algorithm accuracy.Moreover,the overhead of the multiprediction framework is compensated by a parallel implementation that shifts the workload particle filtering from resampling to the prediction and update steps,and the algorithm mitigates the global sequential operations by adding local parallel computation.Combining multi-prediction with parallel particle filtering reduces the discrepancy between a priori and a posteriori information to some extent,and effectively utilizes the particle set to improve the prediction ability of a posteriori particles.Finally,the improved filtering algorithm is implemented on the CUDA model,and the experimental results show that the proposed method improves the accuracy of the particle filtering algorithm.(2)To address the problems of reduced diagnostic accuracy and estimation precision caused by setting fixed model transfer probability in interactive multi-models;the fault diagnosis method for variable pitch systems using parallel resampling filtering is proposed.The multi-model and model group switching(MGS)are combined into a variable structure multi-model,and combined with resampling particle filtering to achieve multiple fault diagnosis of the pitch system.In the MGS approach,the original model set can be expanded by variable models that are weighted by the model estimation probabilities,and then new models are generated in real-time.Finally,the variable structure multi-models are used to detect and identify faults and determine the extent of faults by maximum likelihood estimation.The results show that the proposed method improves the accuracy of variable pitch fault diagnosis.(3)The resampling filtering method using parallel prefixes and optimization is proposed to address the real-time problem of the algorithm in performing fault diagnosis applications.The process and characteristics of the parallel filtering algorithm for multi-model fault diagnosis are analyzed under the CUDA framework.The resampling filtering algorithm using prefixes and optimization is proposed for the problem of complicated algorithms and insufficient real-time performance of parallel algorithms in the process of multi-model multi-fault diagnosis.The prefix sum is gradually optimized according to interleaved pairing,unfolded statute,unfolded thread,and fully unfolded statute methods to remove the thread differentiation and reduce the lag caused by judgment and branch prediction.The resampling filtering algorithm with improved parallel prefix sum is used for fault diagnosis of wind turbine variable pitch.The experimental results show that the method improves the real-time performance of the resampling filtering algorithm in multi-fault diagnosis applications.(4)For the fault diagnosis problem of hybrid intelligence technology,the pitch system fault diagnosis method of attention mechanism one-dimensional convolutional network is proposed.Firstly,the fault data is obtained from the simulation model of the variable structure multi-model fault diagnosis method.Secondly,the samples are extracted from the generated data,and the one-dimensional convolutional network deep learning method is used,while an attention mechanism is introduced to achieve information screening by adaptively weighting different signal features which will be highlighted and they can be characterized the fault information and suppressed the invalid feature information.This research designs the attention mechanism and global average convergence layer then connects Softmax after the average convergence layer.The proposed method integrates the advantages of model-driven and data-driven methods to achieve data-driven fault diagnosis of pitch systems with pitch systems as the object.All experiments show that the proposed method has better generalization performance under fault multi-mode conditions and has better performance to improve fault diagnosis accuracy.