【作者】 王艳；

【导师】 朱永利；

【作者基本信息】 华北电力大学(北京) ， 电气工程， 2025， 博士

【摘要】 电力变压器作为电力系统的核心设备,其运行状态直接关乎系统的安全与稳定。随着智能电网的快速发展,电力变压器状态感知向多参量、高精度方向发展。变压器的多源监测数据各自蕴含着设备运行状态的关键信息,对其进行深度解析成为设备状态精准评估的关键。深度学习技术凭借其强大的特征提取能力,能够有效地挖掘数据中的隐含特征,为监测数据的诊断提供了有效途径。然而,在实际应用中,数据样本的属性缺失、数量不足、高噪声、模式混叠等都制约着深度学习模型的性能。此外,边缘终端受限于计算资源与实时性要求,如何进行深度模型的轻量化成为模型在边缘侧部署的关键;而数据中心侧的多源异构数据融合诊断,则能够有效突破单一监测数据的局限性,形成“边缘实时预警——中心协同融合诊断”的闭环诊断体系,对提升电力变压器状态感知的可靠性具有重要研究价值。因此,本文围绕电力变压器油中溶解气体与局部放电的深度学习诊断模型、模型的轻量化以及数据中心侧的融合诊断展开研究。针对变压器油中溶解气体分析(Dissolved Gas Analysis,DGA)中数据缺失与气体特征重叠、多因素耦合干扰导致的诊断困难问题,提出了基于降噪自编码器的端到端缺失数据填充方法,构建镜像对称的深度编码器-解码器架构,通过噪声注入与非线性关联重构,实现了 R2=0.923的高精度填充,降低了后续诊断的误差传播风险;提出了多尺度特征融合与注意力机制相结合的DGA深度学习诊断模型,通过构建多尺度特征提取模块,并行捕获油中溶解气体浓度的低阶和高阶交互特征,设计通道注意力与空间注意力的级联,动态强化关键特征、抑制噪声干扰并增强模型的可解释性,实现了端到端的DGA诊断。在轻量化参数量(约126k)下,模型平均诊断准确率达96.4%,CPU推理速度达1.9ms/样本。针对变压器局部放电相位分布图谱(Phase-Resolved Partial Discharge,PRPD)识别中面临的样本数量有限、噪声干扰问题,提出了结合数据增强技术的深度残差收缩网络(Deep Residual Shrinkage Network,DRSN)局部放电识别框架,实现了“数据增强-特征降噪-模式识别”的全链路优化。设计了“规范化-灰度化-归一化-仿射变化”的四阶段PRPD图谱预处理流程,建立了局部放电基准数据集;采用辅助分类边界均衡生成对抗网络构建了高质量的增强数据集,缓解了类间不均衡的问题;通过添加高斯白噪声、周期噪声及脉冲噪声,模拟变压器实际噪声环境,增强了局部放电数据的多样性;提出了 DRSN局部放电识别模型,通过双路特征提取融合局部与全局特征,设计通道域自适应软阈值模块抑制噪声干扰,使非噪声场景下局部放电平均识别准确率达到98.0%以上,不同类型噪声干扰下局部放电平均识别准确率亦达到96.5%以上,显著优于传统的CNN模型;跨场景案例验证表明了模型具有良好的泛化能力。针对深度学习模型参数量大、推理速度慢、难以在边缘设备中部署的问题,提出了输出蒸馏与特征蒸馏协同优化的轻量化DRSN局部放电识别方法。以DRSN识别模型作为教师模型,基于深度可分离卷积、通道压缩与残差收缩结构精简重构轻量化DRSN学生模型结构;设计多层次知识迁移机制,在输出空间引入温度缩放机制传递类别语义关联,在特征空间采用L2损失对齐关键层特征分布,实现残差收缩模块中噪声抑制知识的有效迁移,并结合动态权重多目标损失函数平衡分类准确度与特征保真度。模型在存储空间降至271kB、单样本CPU推理速度提升至19.6ms的情况下,在非噪声数据集、高斯白噪声数据集和混合噪声数据集上的平均识别准确率分别达到97.3%、96.1%和95.0%(较教师模型仅下降0.8%、1.1%和1.5%),为变压器边缘智能终端提供了可行的放电故障检测解决方案。针对电力变压器多源信息融合诊断中的模态异构性、时域异步性、数据冲突性及计算复杂性的问题,设计了面向数据中心侧的多源异构数据分层存储体系,采用ODS-DWD-DWS-ADS四级多粒度数据管理机制,通过原始数据存储、数据清洗重构、多维聚合及应用层服务映射,实现了变压器跨模态数据的整合;提出了基于可信度的分层动态决策融合诊断模型,设计考虑时间敏感性的规则可信度,协同动态迭代推理规则,实现了传感器配置异质、特征空间异构、数据时标异步及诊断证据冲突下多源数据的有效融合;构建基于Storm云平台的动态自适应融合诊断框架,设计任务链式触发的分布式流式拓扑结构,实现了多源诊断的动态自适应触发与融合。实验与案例分析表明,方案实现了变压器故障诊断性能与响应速度的协同优化,为复杂工况下的电力设备状态评估提供了技术方案。更多还原

【Abstract】 As the core equipment of the power system,the operation status of power transformers directly affects the safety and stability of the system.With the rapid development of smart grid,the state sensing of power transformer is developing towards multi parameter and high-precision.The multi-source monitoring data of transformers contain key information about the operation status of the equipment,and deep analysis of these data has become the key to accurate assessment of the equipment status.Deep learning technology,with its powerful feature extraction ability,can effectively mine hidden features in data,providing an effective way for the diagnosis of monitoring data.However,in practical applications,the lack of attributes,insufficient number of samples,high noise,and pattern aliasing all constrain the performance of deep learning models.In addition,edge terminals are limited by computing resources and real-time requirements,and how to lightweight deep models has become the key to deploying models on the edge side;The multi-source heterogeneous data fusion diagnosis on the data center can effectively break through the limitations of single monitoring data and form a closed-loop diagnostic system of "edge real-time warning-center collaborative fusion diagnosis",which has important research value for improving the reliability of power transformer state perception.Therefore,the paper focuses on the deep learning diagnostic models of dissolved gases in oil and partial discharge,the lightweight of models,and the fusion diagnosis at the data center for power transformer.Aiming at the diagnostic difficulties caused by data loss,overlap of gas characteristics,and multi factor coupling interference in dissolved gas analysis(DGA)of transformer oil,an end-to-end missing data filling method based on Denoising Autoencoder is proposed,which achieves high-precision filling with R2=0.923 and reduces the risk of error propagation in subsequent diagnosis by constructing a mirror symmetric deep encoder-decoder architecture,as well as noise injection and nonlinear correlation reconstruction;A DGA deep learning diagnostic model combining multi-scale feature fusion and attention mechanism was proposed,which constructs a multi-scale feature extraction module to capture low-order features and high-order interactive features of dissolved gas concentration in oil in parallel,designs a cascade of channel attention and feature space attention to dynamically enhance key features,suppress noise interference and enhance model interpretability,and achieves end-to-end DGA fault diagnosis.Under the lightweight parameter quantity(about 126k),the model’s average diagnostic accuracy reaches 96.4%,and the inference speed running on CPU reaches 1.9ms/sample.A Deep Residual Shrinkage Network(DRSN)partial discharge recognition framework combined with data augmentation technology is proposed to address the problems of insufficient sample and high noise interference in the recognition of Phase Resolved Partial Discharge(PRPD)in transformers.This framework achieves full chain optimization of "data augmentation-feature denouncing-pattern recognition".Designed a four stage PRPD spectrum preprocessing process of"normalization-grayscale-normalization-affine transformation" and established a benchmark dataset for partial discharge;Using Auxiliary classification boundary balancing generative adversarial network.a high-quality enhanced dataset was constructed to alleviate the problem of class imbalance;By adding Gaussian white noise,periodic noise and impulsive noise to simulate real-world noisy environments.the diversity of partial discharge data was enhanced.A DRSN partial discharge recognition architecture was then designed,which integrates local and global features through dual feature extraction,and suppresses noise interference by designing a channel domain adaptive soft threshold module.This model achieves an average partial discharge recognition accuracy exceeding 98.0%in noise-free scenarios,and above 96.5%under different types of noise interference,significantly better than traditional CNN models.Cross-scenario case studies further demonstrate the model’s strong generalization capability.A lightweight DRSN partial discharge recognition method based on collaborative optimization of output distillation and feature distillation is proposed to address the problems of large parameter count,slow inference speed and difficulty in deploying deep learning models on edge devices.Using DRSN model as the teacher model;Using depthwise separable convolution,channel compression and residual shrinkage structure simplification to reconstruct the lightweight DRSN student model structure.Design a multi-level knowledge transfer mechanism,including introducing a temperature scaling mechanism in the output space to transfer category semantic associations,using L2 loss to align the distribution of key layer features in the feature space to achieve noise suppression feature transfer of the residual shrinkage module,and combining dynamic weight multi-objective loss function to balance classification accuracy and feature fidelity.With its storage footprint remarkably reduced to 271 kB and single-sample CPU inference speed significantly accelerated to 19.6 ms,the model achieves the average recognition accuracies of 97.3%,96.1%and 95.0%on noise-free,Gaussian white noise and mixed-noise datasets,respectively(with only marginal drops of 0.8%,1.1%and 1.5%compared to the teacher model).This provides a feasible discharge fault detection solution for transformer edge intelligence terminals.In order to solve the problems of modal heterogeneity,temporal asynchrony,data conflict and computational complexity in multi-source data fusion diagnosis of power transformers,a multi-source heterogeneous data hierarchical storage system is constructed for the data center,which adopts the ODS-DWD-DWS-ADS four level multi granularity data management mechanism,including raw data storage,data cleaning and reconstruction,multidimensional aggregation,and application layer service mapping,to achieve cross modal data integration;A hierarchical dynamic decision fusion diagnostic model based on credibility is proposed,which designs time sensitive rule credibility and collaborates with dynamic iterative inference rules to achieve effective fusion of multi-source data under heterogeneous sensor configurations,heterogeneous feature spaces,asynchronous data time scales,and diagnostic evidence conflicts;A dynamic adaptive fusion diagnosis framework was constructed based on Storm,which achieved dynamic adaptive triggering and fusion of multi-source diagnosis by designing a distributed streaming topology structure with task chain triggering.Experiments and case studies show that the proposed solution achieves synergistic optimization of transformer fault diagnosis accuracy and response speed,which provides a technical solution for power equipment state evaluation under complex operating conditions.更多还原