基于多注意融合网络的输煤皮带异物识别方法  

Foreign object recognition in coal conveyor belt based on multi-attention fusion network

作　　者：李利[1,2,3] 梁晶 陈旭东 寇发荣 潘红光 LI Li;LIANG Jing;CHEN Xudong;KOU Farong;PAN Hongguang(College of Electrical and Control Engineering,Xi’an University of Science and Technology,Xi’an 710054,China;Xi’an Key Laboratory of Electrical Equipment Status Monitoring and Power Supply Safety,Xi’an 710054,China;College of Mechanical Engineering,Xi’an University of Science and Technology,Xi’an 710054,China)

机构地区：[1]西安科技大学电气与控制工程学院,陕西西安710054 [2]西安市电气设备状态监测与供电安全重点实验室,陕西西安710054 [3]西安科技大学机械工程学院,陕西西安710054

出　　处：《西安科技大学学报》2024年第5期976-984,共9页Journal of Xi’an University of Science and Technology

基　　金：陕西省自然科学基础研究计划项目(2024JC-YBQN-0726);陕西省教育厅科研计划项目(23JK0550);西安市科技计划项目(23DCYJSGG0025-2022)。

摘　　要：为改善现有输煤皮带异物识别算法网络参数量大、识别精度不高的问题,及时避免大块煤和矸石、锚杆等带来的安全隐患,提出了一种基于多注意融合网络的输煤皮带异物识别方法,使用低照度图像处理算法对数据集进行预处理,采用融合局部注意力残差块作为基本特征提取单元,在残差块中融入带有额外偏移量的可变形卷积以增加对不规则特征的描述,用注意力机制对全局特征图做期望最大化处理。结果表明:在Cifar 10数据集和矿用皮带传输异物识别数据集的识别准确率分别为93.7%和84.8%;与ShufflenetV2、MobileNetV2、ResNet 50、ResNet 110、Darknet 53算法相比,识别准确率分别提升了4.7%、3.9%、0.4%、0.5%、1.7%;与识别准确率相近的ResNet 50、ResNet 110算法相比,网络参数量和计算复杂度大大减小。识别方法能够快速识别输煤皮带异物,且具有较高的识别准确率,对保障煤矿运输系统的安全运行具有参考意义。In order to slove the problems of large network parameters and low recognition accuracy of the existing foreign object recognition algorithms for coal conveyor belt,and to avoid safety hazards just in time caused by large blocks of coal,gangue,anchor rods,etc.,a foreign object recognition method for coal conveyor belts based on a multi-attention fusion network is proposed.The low illumination image processing algorithm is adopted to preprocess the dataset.A fused local attention residual block is chosen as the basic feature extraction unit with deformable convolution with additional offsets integrated into the residual block to enhance the description of irregular features.The global feature map is processed with expectation maximization using an attention mechanism.The results show that the recognition accuracy rates on the Cifar10 dataset and the foreign object recognition dataset for mining belt transmission are 93.7%and 84.8%,respectively.Compared with algorithms such as ShufflenetV2,MobileNetV2,ResNet 50,ResNet 110,and Darknet 53,the proposed method increases the recognition accuracy rates by 4.7%,3.9%,0.4%,0.5%and 1.7%,respectively.Compared with algorithms such as ResNet50 and ResNet110 with similar recognition accuracy rates,it reduces the network parameters and computational complexity significantly.This recognition method can quickly identify foreign object in coal conveyor belt and has a high recognition accuracy rate,which has reference significance for ensuring the safe operation of coal mine transportation systems.

关 键 词：异物识别 输煤皮带 Darknet网络 可变形卷积 注意力机制 

分 类 号：TD528[矿业工程—矿山机电]