基于改进Mask Scoring R-CNN的铲齿磨损检测研究

为了实现在电铲工作过程中对铲齿磨损进行实时检测，防止因铲齿磨损而影响电铲开采效率，提出了一种基于改进Mask Scoring R-CNN（region convolutional neural network，区域卷积神经网络）的铲齿实例分割模型。首先，以ResNet-101（residual network， 残差网络）和改进的FPN（feature pyramid networks，特征金字塔网络）作为主干网络，提取高、低特征层的语义信息和细节特征并融合，结合ROI Align层对局部特征层进行裁剪和归一化处理，以完成目标检测与实例分割；然后，基于获取的铲齿分割效果图以及二值化掩码图形信息，计算实例分割后图像中铲齿部分的像素面积，以判断其磨损情况。结果表明，以ResNet-101和改进FPN为主干网络的铲齿实例分割模型在测试集上的平均像素精度为90.76%，平均交并比为83.62%，相比于以ResNet-101和传统FPN为主干网络的实例分割模型分别提升了1.18%和1.21%。在电铲采掘工作现场进行8次铲齿磨损检测实验，检测到的每颗铲齿的磨损程度波动幅度均小于2%，均方差为0.7左右，说明所提出的实例分割模型对铲齿有较好的分割效果和稳定性，基本满足磨损检测要求。研究结果可为铲齿磨损状态的智能化检测提供新思路。

Research on shovel tooth wear detection based on improved Mask Scoring R-CNN

In order to realize the real-time wear detection for shovel teeth during the working process of electric shovel, and prevent the mining efficiency of electric shovel from being affected by the shovel tooth wear, a shovel tooth instance segmentation model based on the improved Mask Scoring R-CNN (regional convolutional neural network) was proposed.Firstly, taking the ResNet-101 (residual network) and improved FPN (feature pyramid networks) as the backbone network, the semantic information and detail features of high and low feature layers were extracted and fused, and then the local feature layer was trimmed and normalized by combining with the ROI Align layer, so as to complete target detection and instance segmentation; then, based on the obtained shovel tooth segmentation effect image and binary mask graphic information, the pixel area of shovel tooth in the image after instance segmentation was calculated to judge its wear condition. The results showed that the mean pixel accuracy of the shovel tooth instance segmentation model with ResNet-101 and improved FPN as the backbone network was 90.76% and its mean intersection over union was 83.62% on the test set, which was 1.18% and 1.21% higher than that of the instance segmentation model with ResNet-101 and traditional FPN as the backbone network, respectively. Eight shovel tooth wear detection experiments were carried out at the electric shovel excavation site, and the fluctuation amplitude of detected wear degree of each tooth was less than 2%, and the mean square error was about 0.7, which indicated that the proposed instance segmentation model had good segmentation effect and stability for the shovel tooth, and basically met the requirements of wear detection. The research results can provide new ideas for the intelligent detection of shovel tooth wear state.