Research On Cross-domain Intelligent Diagnosis Of Wind Turbine Drivetrain Faults

The drivetrain is the core component of the wind turbine to achieve energy conversion,and its operational status significantly impacts on the turbine’s safe and economical operation.Influenced by the intermittent and fluctuating characteristics of wind energy,the wind turbine drives are subjected to loads with strong time-varying and shock,resulting in a high fault incidence.Therefore,the health status monitoring and fault diagnosis of wind turbine drivetrains is conducive to the timely detection of safety hazards and avoiding catastrophic accidents,which is of great practical significance to ensuring the units’ safe operation and improving the economic benefits of wind farms.Data-driven intelligent fault diagnosis is an important part of wind power intelligent operation and maintenance,which aims to mine the manifestations of faults from a large amount of monitoring data and establish mapping relationships with fault patterns to identify equipment health status automatically.The prerequisite for good performance of conventional intelligent diagnosis methods is obtaining a large amount of training data with the same distribution as the data to be diagnosed in advance.However,such a requirement is difficult to be satisfied in the engineering diagnosis tasks of wind turbine drivetrains.To break through the above data dilemma faced by conventional data-driven methods,this thesis focuses on the discussion and methodological exploration of how to achieve cross-domain(cross-distribution)fault diagnosis of wind turbine drivetrains using non-identically distributed data.In this thesis,we first address the diagnosis problem of missing data labels in the target domain and conduct an in-depth study on the methods to achieve cross-domain fault diagnosis using a single source domain and multiple source domains,respectively.On this basis,the constraints on the target data are further relaxed.We study how to improve the generalization performance of the diagnostic model by mining common diagnostic knowledge from multiple source domains in the case of insufficient data in the target domain for the cross-service and cross-machine diagnostic scenarios,respectively.Specifically,the research contents of this thesis are summarized as follows:(1)The necessity and feasibility of cross-domain intelligent diagnosis in wind turbine drivetrain condition monitoring are analyzed and discussed.Combining the data characteristics of the vibration signal model of critical rotating equipment in the drivetrain and the generalization error theory of the conventional data-driven method,the necessity of our research is jointly elaborated in terms of both engineering demand and method demand.For our research’s feasibility,the rationale for cross-domain diagnosis of wind turbine drivetrain is mainly explained from the equipment working principle and fault mechanism perspective.Several feasible implementation methods are also discussed.(2)The problem of single-source cross-domain intelligent diagnosis with missing data labels in the target domain is studied.Existing single-source cross-domain diagnosis methods mainly focus on eliminating feature distribution discrepancies at the global or category level while ignoring the transferability of different features,which limits the improvement of model diagnosis performance to a certain extent.In this thesis,a transferable adaptive channel attention module is proposed.It evaluates the cross-domain transferability of different channel features.Then it enhances or reduces their role in the feature adaptation process according to the evaluation results to achieve optimal feature adaptation and improve diagnostic model generalization to target tasks.Experimental results show that the proposed attention module widely applies to cross-domain diagnostic models with different network structures and can significantly improve their diagnostic performance.(3)The problem of multi-source cross-domain intelligent diagnosis with missing data labels in the target domain is studied.To address the single-source cross-domain diagnosis method,which is prone to suboptimal solutions and the accessibility of wind turbine monitoring data from multiple sources,a cross-domain intelligent fault diagnosis method with multi-source domain factorization is proposed.It focuses on reducing the risk of negative transfer triggered by the complex data distribution characteristics in multi-source scenarios.At the feature space level,by drawing on the idea of shared space component analysis,the potential space of each domain is decomposed into shared and private subspaces to eliminate the interference of domain private features on shared knowledge transfer.Furthermore,a loss based on sample conditional entropy is constructed for guiding the model training,constraining the low-density separation of samples in different categories while reducing the risk of negative transfer at the sample level.The proposed method shows superior diagnostic performance concerning several comparison methods on the cross-working conditions and cross-machine diagnostic tasks of rolling bearings.(4)The problem of real-time fault diagnosis across working conditions with insufficient data in the target domain is studied.The uncertainty of wind turbine operating conditions leads to new conditions that may be unseen and whose monitoring data cannot be obtained in advance for model training.Insufficient target data further increases the difficulty of the fault diagnosis task and makes the methods in the first two chapters no longer applicable.To address this problem,this paper proposes a multi-source augmented generalization network for intelligent diagnosis of unknown working conditions.The main idea is to mine various feature representations with a consistent discriminative structure for fault patterns from multiple source domains to improve the model’s generalization performance.Specifically,a multisource augmentation approach is used to generate inter-domain virtual samples to enrich the training data and enable information to interact across multiple domains,thus supporting the model to achieve diverse domain-invariant features.Furthermore,a sample adaptive filtering strategy is designed to optimize the virtual sample generation process and the original sample training to reinforce the fault pattern discriminative structure of the learned features.The experimental results of both gearbox and rolling bearing cross-working-conditions diagnosis show that the proposed method can maintain a good generalization performance for the data distribution deviation caused by the unknown condition fluctuation and can be effectively applied to the real-time intelligent identification task of wind turbine drivetrains’ operation status.(5)The problem of cross-machine fault diagnosis with insufficient data in the target domain is studied.For a newly put-into-operation wind turbine,the amount and type of fault data are scarce and cannot meet the basic requirements for model training.Therefore,using data from other sources to achieve reliable cross-machine diagnostics is beneficial to ensure the safe operation of new units.This thesis introduces the “domain transferability” concept to guide the model to acquire more generalized diagnostic knowledge from multiple source domains through transferability quantification and weighted training on the domain level.Then,a deep neural network with multiple domain discriminators is designed to address the feature distribution alignment problem when the domain label space is inconsistent,thus removing the assumption of label space consistency from most related research works and improving the method’s applicability.Finally,the discrimination ability of the model is further enhanced by imposing a batch spectral penalty canonical constraint on its training.The proposed method demonstrates better diagnostic performance than diagnostic methods based on traditional supervised learning,deep learning,and deep transfer learning in cross-machine diagnostic tasks of rolling bearings.