为解决智能化盾构施工中，施工隧道内所采集的渣土车原始点云数据存在畸变，从而影响渣土体积测量准确性的问题，基于点云数据处理与深度学习的方法，建立基于深度卷积自编码网络的渣土车厢自适应去畸变模型。首先，采用包围盒滤波和反距离加权插值对渣土车厢点云数据进行滤波与补全操作，接着使用真实渣土车厢尺寸构建网络的理想输出，并通过灰度化将点云数据转换为伪特征图，构建网络的数据集；
然后，以传统卷积自编码网络为基础构建渣土车厢自校正网络，网络设计融合堆叠式卷积层和栈式自编码的处理方法，增加网络层数以获得更优的特征表达； 最后，使用盾构掘进现场数据进行试验。结果表明：
本文提出的渣土车厢点云数据自校正方法在保证时间效率的前提下，渣土车厢数据的峰值信噪比（PSNR）达到29.73，结构相似性（SSIM）达到0.86，均优于传统自编码网络与几何约束矫正的方法。证明了本文方法的正确性与有效性，能够提高隧道内所获取渣土车厢点云数据的可用性，同时为点云数据去畸变技术在三维重构和工程领域的应用提供了理论依据。
In intelligent shield tunneling, the accuracy of dump
volume measurements is compromised by distortions in the raw point cloud data
of the dump carriages. To address this issue, a model that utilizes a deep
convolutional selfcoding network for selfadaptive distortionremoval in dump
carriages is developed. This model combines point cloud processing with deeplearning techniques. Initially, surround box filtering and antidistanceweighted interpolation are applied
for the data refinement. Subsequently, the actual size of the slag soil
carriage is utilized to establish the ideal output of the network, transforming
the point cloud into pseudosymbols using grayscale
values to create a network dataset. Furthermore, the dump carriage selfcoding network based on a traditional convolutional selfcoding framework is constructed. This involves an integrated network
design with stacked convolutional layers and autoencoders to enhance the
network depth to achieve superior feature representation. Finally, the method
is tested using shield tunneling data. The experimental results indicate that
the proposed method significantly outperforms traditional selfcoding networks and geometric constraint correction methods in terms
of time efficiency. With a peak signaltonoise ratio of 29.73 and a structural similarity index of 0.86, the
proposed approach demonstrates its superiority and validity. This advancement
not only enhances the utility of dump carriage point cloud data, but also lays
a theoretical foundation for applying point cloud distortionremoval technology in 3D reconstruction and engineering projects.
