摘要
为了解决传统的传送带托辊异常检测方法效率低、实时性差等问题,提出一种基于红外图像识别的托辊异常检测模型。通过现场采集并使用标签平滑和Mosaic数据增强处理对托辊红外图像数据集进行扩充,降低模型的训练成本。在特征提取模块提出使用GhostNet骨干特征提取网络,能够有效地降低特征提取所需成本。在特征融合模块,提出使用SPP-Net模块优化PaNet特征融合网络,增加模型的感受野。通过深度可分离卷积块简化模型结构,降低模型的计算量和参数量,并通过LeakyReLU激活函数提高模型的学习能力。试验结果表明:该检测模型能够有效识别托辊异常。在实际检测中,该方法在托辊检测中平均准确率达到94.9%,检测速度达到39.2 FPS,为矿山传送带托辊的准确高效巡检提供了保障。
With the development of intelligent construction of mine,the detection of mine equipment is becoming more and more intelligent.The ore conveyor belt is one of the important production equipment in the mine,and the abnormal detection of the conveyor roller is one of the important contents of mine safety.At present,most of the ore conveyor roller inspection is manual inspection,and depends on the personal experience of the inspector to judge the working status of the roller,which will lead to problems such as the damage of the roller is not found in time.Therefore,it is urgent to study a more objective,intelligent and efficient method for abnormal detection of rollers.In order to solve the problems of low efficiency and poor real-time performance of the traditional ore conveyor roller anomaly detection method,an optimization model of ore conveyor roller anomaly detection based on infrared image recognition was proposed.The infrared image of the roller was collected on the spot and the infrared image data set of the roller was expanded by using label smoothing and Mosaic data enhancement processing to prevent the overfitting of the detection model and reduce the training cost of the model.In the feature extraction module,it was proposed to use GhostNet backbone feature extraction network,which can effectively reduce the image redundancy produced by feature extraction,accelerate the learning speed of the model,and further optimize the backbone feature extraction network through LeakyReLU activation function to improve the learning ability of the model.In the feature fusion module,multi-dimensional feature fusion was realized through the feature pyramid structure and the bottom-up feature fusion layer,and the SPP-Net module was used to optimize the PaNet feature fusion network to increase the effective receptive field of the model.And through the depth separable convolution block to simplify the model structure,reduce the amount of calculation and the number of parameters of the model.The experimental results show that,compared with the mainstream detection model,the detection model can more effectively identify rollers and distinguish between normal and abnormal rollers.In the actual detection,the detection accuracy of the idlers is 96.2%,the recall rate is 95.9%,and the average detection accuracy is 94.9%,in which the accuracy of abnormal rollers is 99.6%,the accuracy of normal rollers is 90.2%,the detection speed is 39.2 FPS,and the number of model parameters is only 1.1×107.The method provides a guarantee for accurate and efficient inspection of mine conveyor rollers.
作者
阮顺领
阮炎康
卢才武
顾清华
RUAN Shunling;RUAN Yankang;LU Caiwu;GU Qinghua(School of Resource Engineering,Xi’an University of Architecture and Technology,Xi’an 710055,Shaanxi,China;Xi’an Key Laboratory of Intelligent Industry Perception Computing and Decision Making,Xi’an 710055,Shaanxi,China)
出处
《黄金科学技术》
CSCD
2023年第1期123-132,共10页
Gold Science and Technology
基金
国家自然科学基金项目“地下金属矿山岩体破坏多源异质流数据智能融合与态势评估研究”(编号:51974223)
陕西省自然科学基金项目“多模态融合学习下尾矿坝安全态势感知与协同预警研究”(编号:2022JM-201)联合资助。
关键词
机器视觉
红外图像识别
深度学习网络
网络结构优化
托辊检测
异常检测
machine vision
infrared image recognition
deep learning network
network structure optimization
roller detection
anomaly detection
