摘要
针对单传感器状态识别算法存在漏检、误检的缺陷,文章提出一种基于双传声器和深度学习的变压器状态识别算法,即基于一维卷积神经网络和双传声器数据融合算法(1D-CNN based dual microphones fusion algorithm,1D-CNN-DMF)。利用2个传声器分别同时采集变压器声信号,通过一维卷积神经网络对2个传声器采集到的声信号分别进行特征提取,并利用全连接层对特征进行融合,最终通过softmax分类器进行分类。通过采集500 kV变压器的声信号构建数据集进行验证,结果表明1D-CNN-DMF算法可以有效地对变压器不同状态进行分类,分类准确率高于1D-CNN-LSTM、1D-CNN、FFT-BP、SVM和FFT-SAE等算法。最后利用t-SNE可视化工具揭示了1D-CNN-DMF算法的内在机制。
To overcome the shortcomings of single-sensor recognition algorithms with missed and false detections,this paper proposes an algorithm based on multi-sensor data fusion and deep learning,namely,1D-CNN-based dual microphones fusion algorithm(1D-CNN-DMF algorithm for short).The algorithm utilizes dual microphones to simultaneously collect acoustic signals from transformer statuses.A one-dimensional convolutional neural network is used to extract features from the acoustic signals collected by each microphone.The extracted features are then fused using a fully connected layer,and the output features are classified by a softmax classifier.The algorithm is validated using a dataset of acoustic signals collected from a 500kV transformer in four different statuses.The experimental results demonstrate that the proposed algorithm effectively recognizes different transformer statuses,and the accuracy of 1D-CNN-DMF algorithm is higher than 1D-CNN-LSTM,1D-CNN,FFT-BP,SVM and FFT-SAE.Furthermore,the t-SNE visualization tool is used to reveal the inner mechanism of the proposed algorithm.
作者
马裕超
汪欣
钱勇
莫娟
韩利
MA Yuchao;WANG Xin;QIAN Yong;MO Juan;HAN Li(China Electric Power Research Institute Co.,Ltd.,Haidian District,Beijing 100055,China;Shanghai Advanced Research Institute,Chinese Academy of Sciences,Pudong New District,Shanghai 201210,China;Electric Power Research Institute,State Grid Ningxia Electric Power Co.,Ltd.,Yinchuan 750002,Ningxia Hui Autonomous Region,China)
出处
《电力信息与通信技术》
2024年第2期54-60,共7页
Electric Power Information and Communication Technology
基金
国家电网有限公司总部科技项目资助“110kV~750kV变压器出厂噪声指标逆向评测方法研究与应用”(8100-202055154A-0-0-00)。
关键词
深度学习
状态识别
声信号处理
卷积神经网络
deep learning
status recognition
acoustic signal processing
convolutional neural networks