推土机是土方施工中应用广泛的工程机械类型，实现无人推土机自动化施工，为解决建筑行业所面临的施工安全风险高和劳动力短缺等问题带来机遇。非结构化的施工场地有着环境复杂多变的特点，现有基于规则的自动化施工方法在实际工程中存在泛化能力不足的问题。随着人工智能相关技术在无人自主系统中的应用，为实现大规模动态场景下的自主施工带来可能。
环境的不确定性存在多种高维复杂信息，导致多模态数据特征提取困难，数据融合理论和方法有待进一步探索。输入输出的不确定性使得推土机施工路径决策难以建模和优化，需要研究基于深度学习提升非线性模型回归问题泛化能力的方法。本文以无人推土机为对象，研究人工智能融合有经验驾驶员专家知识的建模与分析方法，提出自主施工路径决策模型的有效措施。本文的具体研究内容如下：
（1）推土机施工路径决策感知要素与多模态数据采集。为解决施工决策要素提取困难，数据采集不全面和采集方式孤立的问题，本文基于推土机施工作业特征和动作解耦分析其决策要素，主要包含运距、倒车轨迹以及铲刀三维轨迹。建立推土机施工过程多模态数据采集平台，分别进行山推 SD20，SD90 和 DH17C2U 三种型号推土机的传感器选型和安装，实现推土机施工过程多维异构数据的采集和存储。
（2）推土机运距决策模型研究。为解决现有施工决策方法泛化能力弱的问题，本文结合有经验操作手的施工过程数据，基于观察中模仿的理论和方法实现推土机运距的智能决策。分别以 VGG-16, LeNet-5 和 AlexNet 网络为骨干进行对比测试，结果表明以改进的 Inception-v3 网络为骨干进行图像特征提取，所构建的模型能够得到最小的平均误差，且融合专家知识的决策模型泛化能力强，稳定性高。
（3）推土机倒车轨迹决策模型研究。为解决推土机施工过程倒车轨迹决策存在建模复杂、参数调节难的问题，本文构建端到端深度学习模型，结合注意力机制提升网络对特征向量的提取能力。以 ResNet-50 网络为骨干分别在特征矩阵的水平和垂直方向上实现全局平均池化操作，所构建的模型能够得到最小的均方根误差，满足日常施工的精度要求。最后结合注意力热图，分析施工环境信息中的重点区域感知情况，解释所提出模型做出决策行为的原因。
（4）推土机铲刀三维轨迹决策模型研究。环境的不确定性存在多种高维复杂信息，导致施工决策建模和优化困难。本文结合无人机倾斜摄影技术构建大范围施工场地的三维模型，在实现排土墙边缘检测的基础上，提出将高精度施工地图和群体智能算法相结合的多目标决策优化方法，能够在迭代计算周期内获得最小的适应度函数，实现推土机铲刀三维轨迹决策，以支撑推土机在矿山排土等场景中的高效施工。
最后，本文提出融合数字孪生的无人推土机施工路径决策模型实现方法，在土方项目中对模型进行测试，结果表明本研究所构建模型的平均延迟为 27.03 毫秒，无人推土机每小时的施工区域面积相当于有人驾驶推土机的 94.5%，无人推土机相比于有人驾驶推土机，施工过程的高程差变化减小 60%，在不同施工场景中均能够实现无人推土机施工路径智能决策模型的有效应用。

关键词：推土机；施工自动化；智能决策；卷积神经网络；人工智能

Bulldozers are widely used in earthmoving construction, and the automation of unmanned bulldozers presents opportunities to address challenges in the construction industry, such as high construction safety risks and labor shortages. Unstructured construction sites have complex and dynamic environmental characteristics, posing challenges for rule-based automation methods with limited generalization capabilities in practical engineering. With the application of artificial intelligence (AI) technologies in unmanned autonomous systems, the potential for achieving autonomous construction in large-scale dynamic scenarios has emerged.
The uncertainty of the environment involves a variety of high-dimensional complex information, which makes it difficult to extract multi-modal data features, data fusion theories and methods need to be further explored. The uncertainty of input and output makes it difficult to model and optimize bulldozer construction path decisions. It is necessary to study methods to improve the generalization ability of nonlinear model regression problems based on deep learning. This article takes unmanned bulldozers as the object, studies the modeling and analysis method of artificial intelligence integrating the expert knowledge of experienced drivers, and proposes effective measures for autonomous construction path decision-making models. The specific research contents of this article are as follows:
(1) Perception factors and multi-modal data collection for bulldozer construction path decision-making. In order to solve the problems of difficulty in extracting construction decision-making elements, incomplete data collection and isolated collection methods, this paper analyzes the decision-making elements of bulldozers based on their construction operation characteristics and action decoupling, which mainly include transportation distance, reversing trajectory and three-dimensional blade trajectory. A multi-modal data collection platform for the bulldozer construction process was established to select and install sensors for three types of bulldozers, namely Shantui SD20, SD90 and DH17C2U, to achieve the collection and storage of multi-dimensional heterogeneous data during the bulldozer construction process.
(2) Research on bulldozer transport distance decision-making model. In order to solve the problem of weak generalization ability of existing construction decision-making methods, this paper combines the construction process data of experienced operators and realizes intelligent decision-making of bulldozer travel distance based on the theory and methods imitated in observation. Comparative tests were conducted using VGG-16, LeNet- 5 and AlexNet networks as the backbone. The results showed that using the improved Inception-v3 network as the backbone for image feature extraction, the model built can achieve the smallest average error. The decision-making model that integrates expert knowledge has strong generalization ability and high stability.
(3) Research on bulldozer reversing trajectory decision-making model. In order to solve the problems of complex modeling and difficult parameter adjustment in bulldozer reversing trajectory decision-making during construction, this article builds an end-to-end deep learning model and combines the attention mechanism to improve the network's ability to extract feature vectors. Using the ResNet-50 network as the backbone, the global average pooling operation is implemented in the horizontal and vertical directions of the feature matrix. The constructed model can obtain the smallest root mean square error and meet the accuracy requirements of daily construction. Finally, combined with the attention heat map, the perception of key areas in the construction environment information is analyzed, and the reasons for the decision-making behavior of the proposed model are explained.
(4) Research on the three-dimensional trajectory decision-making model of the bulldozer blade. Environmental uncertainty involves a variety of high-dimensional complex information, which makes construction decision-making modeling and optimization difficult. This paper combines drone oblique photography technology to build a three dimensional model of a large-scale construction site. On the basis of realizing the edge detection of the dump wall, a multi-objective decision-making optimization method that combines high-precision construction maps and swarm intelligence algorithms is proposed, which can iteratively The minimum fitness function is obtained within the calculation cycle to realize the three-dimensional trajectory decision-making of the bulldozer blade to support the efficient construction of the bulldozer in scenarios such as mine dumping.
Finally, this paper proposes an implementation method for the unmanned bulldozer construction path decision-making model that integrates digital twins, and tests the model in earthmoving projects. The results show that the average delay of the model constructed in this study is 27.03 milliseconds, and the construction area of the unmanned bulldozer per hour is Equivalent to 94.5% of that of a manned bulldozer. Compared with a manned bulldozer, the elevation difference during the construction process of an unmanned bulldozer is reduced by 60%. It can effectively apply the intelligent decision-making model of the unmanned bulldozer's construction path in different construction scenarios.