【作者】 李莉；

【导师】 杨明顺； 王荪馨； 侯增刚；

【作者基本信息】 西安理工大学 ， 机械工程（专业学位）， 2024， 硕士

【摘要】 风能作为世界范围内最重要的能源之一,它储量富足、可再生且低碳无污染,在解决世界能源短缺问题、推动我国能源转型过程中承担着重要角色。我国的风力发电机组累计和新增容量在全球范围内持续保持领先地位。随着风电装机量的不断提升,风机故障诊断也逐步成为业界关注的重点。齿轮箱作为风电机组的关键组成部分,其运行状况严重影响机组的稳定性。因此,实时监测风电机组齿轮箱的工作状况,并对其进行故障诊断,对确保风机的安全运行、减少运维开支以及提高发电效率具有重要的意义。分析齿轮箱故障诊断的基本流程,构建了“云-雾-边”协同工作的齿轮箱故障诊断框架,对数据采集与传输、边缘端的数据预处理、云端样本量充足时高性能的齿轮箱故障诊断模型构建、雾端的小样本故障诊断模型构建和信息可视化过程中涉及的关键技术进行分析。通过云雾边协同来减少网络延迟和带宽资源的浪费,充分挖掘雾节点的计算能力,减轻云端的处理负担,突显边缘节点在故障诊断中实时响应的优势。采用基于随机森林、奇异值分解、孤立森林、Z-score标准化的数据预处理方法,处理原始数据传输过程中的缺失、异常、噪声等现象。采用随机森林算法进行数据插补以处理缺失值,使用孤立森林算法识别异常值,利用奇异值分解技术对数据进行降噪处理,采用Z-score标准化方法统一数据量纲。通过边缘端的数据预处理方法,提高数据挖掘效率,在一定程度上增强了故障诊断模型的性能。构建一个融合卷积神经网络、双向门控循环单元和注意力机制的齿轮箱故障诊断模型。考虑齿轮箱监测数据具有规模大、时序性和多特征等特性,利用卷积神经网络提取齿轮箱振动数据的重要特征,利用门控循环网络捕捉数据的时序特性,利用注意力机制来动态调整特征的权重分布,从而强化重要特征对模型的影响,同时采用蝴蝶优化算法对模型的超参数进行优化,提升模型的训练效率。通过实例分析验证该方法的可行性。构建了齿轮箱深度迁移故障诊断模型。考虑齿轮箱更新换代或者工况变化,缺乏足够的样本数据对模型性能的影响,以样本量充足的齿轮箱数据集为源域样本,使用这些样本进行模型训练和特征提取,将稀疏样本作为目标域输入模型中提取目标域特征;采用最大均值差异方法衡量源域和目标域之间的差异,对模型进行微调提升模型的诊断精度;应用经过迁移学习优化的模型对齿轮进行故障诊断,通过实例验证模型的有效性。基于云雾边协同框架,设计并开发了齿轮箱故障诊断原型系统,实现齿轮箱状态监测、数据管理、故障诊断与实时预警功能,为齿轮箱的故障诊断和预测提供更为精确和可靠的数据支撑及技术手段。更多还原

【Abstract】 Wind energy,a crucial global energy source,is characterized by its abundance,renewability,and low-carbon,pollution-free nature.It plays an important role in solving the world’s energy shortage problem and promoting China’s energy transformation.China’s wind turbine cumulative and new capacity in the world continues to maintain a leading position.As the installed capacity of wind power continues to rise,wind turbine troubleshooting has gradually become the focus of industry attention.As a key component of wind turbines,the gearbox’s operating condition seriously affects the stability of the unit.Therefore,real-time monitoring and troubleshooting of wind turbine gearboxes is crucial for guaranteeing the secure operation of wind turbines,reducing operational and maintenance costs,and enhancing power generation efficiency.The basic process of gearbox fault diagnosis is analyzed,and a "cloud-fog-edge"collaborative gearbox fault diagnosis framework is constructed,and the key technologies involved in data acquisition and transmission,data preprocessing at the edge,high-performance gearbox fault diagnosis model when the sample size is sufficient at the cloud end,and smallsample fault diagnosis method and information visualization at the fog end are analyzed.Key technologies are analyzed.Through cloud-fog edge collaboration to reduce network delay and bandwidth resource waste,while fully exploiting the computational capacity of the fog nodes to reduce the processing burden of the cloud,highlighting the advantages of edge nodes in real-time response in fault diagnosis.Data preprocessing methods based on Random Forest,Singular Value Decomposition,Isolated Forest,and Z-score standardization are used to deal with missing,abnormal,and noisy phenomena in the transmission of raw data.Random forest algorithm was used to interpolate missing data,isolated forest algorithm was used to detect outliers,singular value decomposition technique is used to reduce noise in the data,and Z-score standardization method is used to unify the data scale.By employing various edge based data preprocessing techniques,the effectiveness of data extraction is enhanced,and the performance of model is enhanced.A model for gearbox fault diagnosis is developed,incorporating a convolutional neural network,bidirectional gated recurrent units,and an attention mechanism.Considering that the gearbox monitoring data has the characteristics of large scale,temporal sequence and more features,a convolutional neural network is applied to mine the essential features from gearbox vibration data,a bidirectional gated recurrent unit is utilized to extract the timing characteristics of the data,an attention mechanism is applied to adjust the feature weight distribution dynamically,thereby enhancing the influence of the important features on the model.At the same time,a butterfly optimization algorithm is utilized to optimize the hyperparameters of the model,enhancing its training efficiency.The method’s viability is demonstrated through analytical case studies.A gearbox deep migration fault diagnosis model is constructed.Considering the impact of gearbox upgrading or working condition changes and lack of sufficient sample data on model performance,the gearbox data set with sufficient sample size was taken as source domain samples to conduct model training and feature extraction,and sparse samples were taken as target domain input models to extract the features.Employing the maximum mean discrepancy technique allows for the assessment of discrepancies between the source and target domains.The model undergoes refinement to boost its diagnostic accuracy.The model optimized by transfer learning is utilized to the gear fault diagnosis,with its efficacy being demonstrated through an example.Based on the cloud-fog-edge collaboration framework,the prototype system for gearbox fault diagnosis is engineered and developed.This system aims to achieve the objectives of monitoring the condition of gearboxes,managing data,diagnosing faults,and providing real-time early warnings.It is intended to offer more precise and dependable data support and technological methods for gearbox fault diagnosis and prediction.更多还原