基于改进YOLO-tiny的闸板阀开度检测

Opening degree detection of gate valve based on improved YOLO-tiny

闸板阀是煤矿生产过程中控制水、煤流量的常用设备。由于成本、布线等原因,非关键闸板阀的开度检测并未纳入到矿井集中监控系统中。鉴于此,在已有的视频监控系统基础上,针对现有基于图像处理的闸板阀开度检测算法存在需要多模型训练、多步检测、容错率低的问题,提出了一类以改进的YOLO-tiny(包括YOLOv3-tiny和YOLOv4-tiny)为核心的闸板阀开度检测方法。首先,将图像输入到检测网络并使用卷积进行特征提取;其次,为了提高检测模型对多尺度插板的检测精度,从增加网络感受野的角度出发,设计了一类将改进的空间金字塔池化(Spatial Pyramid Pooling,SPP)模块、Sub-stage特征融合和YOLO-tiny相结合的检测器SSA-YOLO(SPP and Sub-stage Aggregated YOLO),对插板及其边框的类别和位置信息进行端到端的预测;最后,检测器输出插板和闸板外框的类别及坐标,并利用它们的位置关系确定闸板阀的开度值。为了更加准确地衡量模型同时检测出插板及其闸板外框的能力,提出使用pairedAP(paired Average Precision)指标对检测模型进行评估。使用3种闸板阀在不同时段的3000张图像和相关监测视频作为数据集对所提方法进行试验,结果表明:2种SSA-YOLO模型在保证实时检测的基础上,其pairedAP指标比对应的YOLO-tiny基准模型分别提高了10.6%和36.2%,并增强了模型的抗干扰能力与泛化性能,即使在闸板开度值连续变化的情形下仍有效。笔者提出的闸板阀开度检测思路能扩展应用于可利用多目标物体之间的空间位置关系来确定特定检测量的问题中。

Gate valve is commonly used for controlling the water flow or coal flow in coal production.Considering the cost,wiring and other factors,the non-critical gate valves have not been remotely monitored by the centralized monitoring system of mine.However,the current-available image processing-based techniques suffer from multi-model training,multi-step detection,and low fault tolerance.To address these concerns,a novel strategy based on the improved YOLO-tiny(including YOLOv3-tiny and YOLOv4-tiny)is proposed.Firstly,the convolution operation is employed to extract the features of input images.Moreover,to enlarge the receptive field and hence improve the detection accuracy for multi-scale plug-in,an effective detector has been developed based on the combination of improved SPP(Spatial Pyramid Pooling)module,Sub-stage feature fusion and YOLO-tiny networks,named SSA-YOLO(SPP and Sub-stage Aggregated YOLO).Finally,the opening degree of gate valve is calculated based on the positional relationship between the plug-in and outer frame of the gate.pairedAP(paired Average Precision)index is proposed to more accurately measure the ability of detector for simultaneously predicting the plug-in and outer frame of the gate.The experimental results on 3000 images and related monitoring videos from three types of gate valves working in different time demonstrate that the proposed SSA-YOLO detectors based on the YOLOv3-tiny and YOLOv4-tiny are able to process images in real time,and achieve significantly higher pairedAP than the corresponding YOLO-tiny backbones,by 10.6%and 36.2%respectively.Meanwhile,the proposed detectors outperform their counterparts in terms of the robustness and generalization ability,and work well even when the opening degree of gate valve continuously changes.The proposed idea of the gate valve opening detection can be applied to measure the specific parameters with spatial position relationship between multiple objects.