基于改进ResNet网络的井下钻杆计数方法

Research on underground drill pipe counting method based on improved ResNet network

  • 摘要: 针对现有井下钻杆数量统计方式精度较低、受环境变化影响大等问题,结合卷积神经网络、信号滤波等技术,提出了一种基于改进ResNet网络的井下钻杆计数方法。根据视频图像中卸杆动作与非卸杆动作的差异,采用ResNet-50网络模型对样本集进行分类训练,判断视频中每一帧图像是否包含卸杆动作;结合线性学习率预热和基于Logistic曲线的学习率衰减策略进行学习率更新,以提高模型分类准确率;通过积分法对视频分类置信度进行滤波,并统计置信度曲线下降沿数量,实现钻杆计数。实验结果表明,预热+衰减的学习率更新策略能够有效提高图像分类模型的分类精度,模型分类检测准确率为89%。实际应用结果表明,基于改进ResNet网络的井下钻杆计数方法可以高效识别视频中的卸杆图像,平均钻杆计数精度为97%,满足实际应用需求。

     

    Abstract: In view of problems that the existing underground drill pipe quantity statistics method has low accuracy and is easily affected by environmental changes,an underground drill pipe counting method based on improved ResNet network was proposed combining with convolutional neural network, signal filtering and other technologies. According to the difference between unloading action and non-unloading action in video image, the sample set is classified and trained based on the ResNet-50 network model to determine whether each frame of the video contains the unloading action; linear learning rate preheating and Logistic-based learning rate attenuation strategy are combined to update learning rate and improve the accuracy of model classification; the video classification confidence is filtered through the integration method, and the number of falling edges of the confidence curve is counted to realize drill pipe counting. Experimental results show that the preheating + attenuation learning rate update strategy can effectively improve classification accuracy of the image classification model to 89%. The actual application results show that underground drill pipe counting method based on improved ResNet network can efficiently identify unloaded rod images in the video with an average accuracy of 97%, which meets the actual application requirements.

     

/

返回文章
返回