摘要
针对供电企业火灾隐患检测中存在的特征提取不精准、多源数据融合困难等问题,提出一种改进的反向传播(Back Propagation,BP)神经网络方法。该方法融合自适应学习率与Dropout正则化技术,并优化了适用于电力场景的灰度分布相关性、圆形度等静态特征提取算法,有效实现了静态与动态特征的高效融合。结合专家规则库,构建了分级预警与闭环处置机制。实验结果表明:所提方法在电弧场景下的辨识准确率达到97%,蒸汽场景下稳定在91%以上;特征轮廓清晰度较基线方法提升了43%;在7维融合特征中,能量特征与灰度相关性特征的贡献权重分别达0.32与0.28;模型经40次迭代后准确率稳定于97.5%,验证了其在多模态火灾隐患辨识中的有效性与鲁棒性。
To address the problems of inaccurate feature extraction and difficulties in multi-source data fusion in fire hazard detection in power supply enterprises,an improved back propagation(BP)neural network method is proposed.This method incorporates an adaptive learning rate and dropout regularization techniques,and optimizes static feature extraction algorithms,such as gray-level distribution correlation and circularity,for power scenarios,effectively achieving efficient fusion of static and dynamic features.A hierarchical early warning and closed-loop response mechanism is constructed by combining an expert rule base.Experimental results show that the proposed method achieves a recognition accuracy of 97%in the electric arc scenario and remains stable above 91%in the steam scenario;the feature contour clarity is improved by 43%compared to the baseline method;in the 7-dimensional fusion features,the contribution weights of energy features and gray-level correlation features reach 0.32 and 0.28,respectively;after 40 iterations,the accuracy of the model stabilizes at 97.5%,verifying its effectiveness and robustness in multimodal fire hazard identification.
作者
李晨光
刘洋
刘平
梁建磊
王鹏
LI Chengguang;LIU Yang;LIU Ping;LIANG Jianlei;WANG Peng(State Grid Inner Mongolia Eastern Electric Power Co.,Ltd.,Hohhot 010000,China;Comprehensive Service Branch of State Grid Inner Mongolia Eastern Power Co.,Ltd.,Hohhot 010000,China;Electric Power Science Research Institute of State Grid Inner Mongolia Eastern Electric Power Co.,Ltd.,Hohhot 010000,China)
出处
《国外电子测量技术》
2025年第11期314-320,共7页
Foreign Electronic Measurement Technology
