基于深度学习的故障电弧检测方法研究

【作者】 杨洋；

【导师】 李平；

【作者基本信息】 西南石油大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 故障电弧是引发电气火灾的主要因素,检测故障电弧对于电气火灾防控具有重要意义。电流信号具有复杂性、随机性和多样化特性,给故障电弧检测带来一定的难度与挑战,比如电流哪方面的特征与故障电弧相关、如何自动获取故障相关的电流特征等。利用深度学习的强表征能力学习电流信号的特征是目前的研究趋势,然而与大多数基于一维信号分析的电弧故障检测方法不同,本文探索了除时间域外的其他高阶特征,并通过构建深度学习模型实现了端对端的故障电弧检测。具体而言,针对故障电弧检测进行了以下几个方面的创新性研究:1、本文利用时频分析方法获取电流信号的二维时频图,采用卷积神经网络模型对直方图均衡化处理后的时频图进行学习。通过在自采的电流数据集上进行对比实验和消融实验,验证了增强时频图像表示方法在故障电弧检测中的有效性,为故障电弧检测任务提供一种新颖的思路。2、离散小波分解是一种常用的时频特征分析方法,然而小波基的选择要求具备先验知识,同时通过预定义小波基的方式对于复杂多样的电流波形不具备自适应性。本文提出一种端对端的深度小波卷积神经网络,自动学习小波基函数,找到最适合电流数据分布的滤波器系数,实现电流特征自适应提取。3、特征融合方法能够丰富电流不同维度的特征,有助于模型达到更准确的检测效果。本文借助特征融合手段实现多特征之间的优势互补,提出一种双通道卷积神经网络模型用于故障电弧检测。通过实验充分说明了双通道的特征融合机制能够更加全面地刻画出电流的关键特征,提升模型性能。4、设计并实现了故障电弧监测系统,该系统能够利用本文提出的故障电弧识别算法进行线路中的故障电弧检测,并展示家用电器的运行状态,保证用电安全。更多还原

【Abstract】 Arc fault is the main factor causing electrical fire,and the detection of arc fault is of great significance for the prevention and control of electrical fires.The complexity,randomness and diversification of current signals bring some difficulties and challenges to arc fault detection,for example,which features of current are related to arc fault and how to automatically obtain the fault-related current features,etc.Utilizing the strong representational ability of deep learning to learn the features of current signals is the current research trend.However,unlike most one-dimensional time series analysis methods based on deep learning,this thesis explores other high-order features in addition to the time domain,and achieves end-to-end arc fault detection by building a deep learning model.Specifically,the following innovative researches are carried out for arc fault detection:1.In this thesis,the time-frequency analysis method is used to obtain the two-dimensional time-frequency image of the current signal,and the Convolutional Neural Network(CNN)is used to learn the time-frequency image after the histogram equalization processing.Through comparative experiments and ablation experiments on the self-collected current dataset,the effectiveness of the enhanced time-frequency image representation method in arc fault detection is verified,which provides a novel idea for the arc fault detection task.2.Discrete wavelet decomposition(DWD)is a common time-frequency feature analysis method.However,the selection of wavelet bases requires prior knowledge,and the pre-defined wavelet bases are not adaptive to complex and diverse current waveforms.In this thesis,an end-to-end deep wavelet convolutional neural network is proposed to automatically learn wavelet functions,find the most suitable filter coefficients for the current data distribution,and realize adaptive extraction of current features.3.Feature fusion method can enrich the features of different dimensions of currents,which is helpful for the model to achieve a more accurate detection result.In this thesis,a dual channel convolutional neural network is proposed for arc fault detection based on feature fusion method to realize complementary advantages among multiple features.The experiments fully demonstrate that the dual channel feature fusion mode can depict the key features of the current and improve the model classification ability.4.An arc fault monitoring system is designed and implemented,which can utilize the arc fault detection algorithm proposed in this paper to detect the arc fault current in the line,and display the running state of household appliances to ensure the safety of using electricity.更多还原