本文设计了一种冷再生沥青路面材料配比优化系统，采用深度学习技术实现配比参数与路面性能的非线性映射。系统由数据采集、预处理、模型构建 / 训练、配比预测优化和结果验证5个模块组成，引入残差网络与双向长短期记忆网络，以捕捉材料性能演化规律，并采用量子遗传算法生成帕累托最优解集。工程试验表明，系统生成的配比方案动态模量均大于3000MPa，劈裂强度大于0.8MPa，冻融劈裂强度比高于75%，疲劳寿命超106次循环，单位成本低于300元 /m3，验证了系统的实用价值。

　　This article designs a cold recycled asphalt pavement material proportioning optimization system, which uses deep learning technology to achieve nonlinear mapping between proportioning parameters and pavement performance. The system consists of five modules: data collection, preprocessing, model construction/training, ratio prediction optimization, and result verification. Residual networks and bidirectional long short-term memory networks are introduced to capture the evolution of material properties, and quantum genetic algorithms are used to generate Pareto optimal solution sets. Engineering tests have shown that the dynamic modulus of the ratio scheme generated by the system is greater than 3000 MPa, the splitting strength is greater than 0.8 MPa, the freeze-thaw splitting strength ratio is higher than 75%, the fatigue life exceeds 106 cycles, and the unit cost is less than 300 yuan/m3, verifying the practical value of the system.

　　关键词：冷再生沥青 ；深度学习 ；配比优化

　　Keywords: cold recycled asphalt; Deep learning; Ratio optimization

　　随着我国对交通基础设施建设的重视和投入力度不断加大，全国各地沥青路面的养护修缮需求日益增长[1]。冷再生沥青路面是绿色低碳技术的典型代表，其材料配比的科学性和精准性直接决定了路面性能和使用寿命。目前，人工智能技术蓬勃发展，深度学习具有强大的非线性拟合能力和数据挖掘潜力，为解决冷再生沥青路面材料配比优化这个复杂工程难题提供了新思路。

　　With the increasing emphasis and investment in transportation infrastructure construction in China, the demand for maintenance and repair of asphalt pavements across the country is growing day by day. Cold recycled asphalt pavement is a typical representative of green and low-carbon technology, and the scientific and precise material ratio directly determines the performance and service life of the pavement. At present, artificial intelligence technology is flourishing, and deep learning has strong nonlinear fitting ability and data mining potential, providing new ideas for solving the complex engineering problem of optimizing the proportion of cold recycled asphalt pavement materials.

　　1相关理论基础

　　1. Relevant theoretical basis

　　1.1 冷再生沥青路面材料配比影响因素冷再生沥青路面材料的基本组成包括旧沥青混合料（Reclaimed Asphalt Pavement， RAP）、新集料和沥青乳化剂（Asphalt Emulsion，AE）等。RAP 为主要原料，其粒径分布、针入度等物理性质直接影响再生混合料的路用性能。合理选配新集料能够调节混合料级配，弥补 RAP 缺陷。沥青乳化剂为粘结剂，其残留沥青含量、乳化剂类型等参数与混合料的离析抗裂性密切相关。此外，温度、湿度等外界因素也会影响材料组分间的兼容性和化学反应动力学，进而改变路面力学性能。例如，当环境温度从 10℃升至40℃时，再生混合料的动态模量会下降 20%~30%。

　　1.1 Factors affecting the proportion of cold recycled asphalt pavement materials The basic components of cold recycled asphalt pavement materials include recycled asphalt pavement (RAP), new aggregates, and asphalt emulsifiers (AE). RAP is the main raw material, and its physical properties such as particle size distribution and penetration directly affect the road performance of recycled mixtures. Reasonable selection of new aggregates can adjust the gradation of the mixture and compensate for RAP defects. Asphalt emulsifier is a binder, and its residual asphalt content, emulsifier type, and other parameters are closely related to the segregation and crack resistance of the mixture. In addition, external factors such as temperature and humidity can also affect the compatibility and chemical reaction kinetics between material components, thereby altering the mechanical properties of the road surface. For example, when the ambient temperature rises from 10 ℃ to 40 ℃, the dynamic modulus of the recycled mixture will decrease by 20% to 30%.

　　1.2 深度学习用于材料配比的优势深度学习具有强大的非线性拟合能力和数据挖掘潜力，为解决冷再生沥青路面材料配比这个复杂工程难题提供了新思路。传统的配比优化方法通常基于人工经验和有限的试验数据，难以准确描述材料组分与路面性能间的内在联系。而深度学习模型可以从海量历史配比数据中自主学习，提取隐藏在高维参数空间中的关键特征，建立配比参数与目标性能指标间的非线性映射关系。

　　1.2 Advantages of Deep Learning in Material Proportioning Deep learning has strong nonlinear fitting ability and data mining potential, providing new ideas for solving the complex engineering problem of cold recycled asphalt pavement material proportioning. Traditional proportioning optimization methods are usually based on manual experience and limited experimental data, making it difficult to accurately describe the intrinsic relationship between material composition and pavement performance. Deep learning models can autonomously learn from massive historical matching data, extract key features hidden in high-dimensional parameter spaces, and establish nonlinear mapping relationships between matching parameters and target performance indicators.

　　例如，应用卷积神经网络处理骨料图像，即可定量评估骨料形状、粒径分布等微观形貌特征对路面强度和水稳定性的影响规律，大幅提升了配比方案的针对性和可靠性。此外，生成对抗网络在创新选材方面也具有卓越潜力，可根据设计需求自主优化母体沥青的级别和掺量，拓展了再生路面的适用范围。2基于深度学习的冷再生沥青路面材料配比优化系统设计2.1 系统整体架构设计本文提出了一种基于深度学习的冷再生沥青路面材料配比优化系统，其整体架构如图 1 所示。该系统由数据采集、预处理、深度学习模型构建与训练、材料配比预测优化以及结果展示验证 5 个模块构成。

　　For example, by applying convolutional neural networks to process aggregate images, the influence of micro morphological features such as aggregate shape and particle size distribution on pavement strength and water stability can be quantitatively evaluated, greatly improving the pertinence and reliability of the proportioning scheme. In addition, generative adversarial networks have excellent potential in innovative material selection, as they can independently optimize the grade and dosage of the parent asphalt according to design requirements, expanding the applicability of recycled road surfaces. 2. Design of a deep learning based cold recycled asphalt pavement material ratio optimization system 2.1 Overall system architecture design This paper proposes a deep learning based cold recycled asphalt pavement material ratio optimization system, and its overall architecture is shown in Figure 1. The system consists of five modules: data collection, preprocessing, deep learning model construction and training, material ratio prediction and optimization, and result display and verification.

　　其中，数据采集模块用于收集并标注原材料性质、配比参数和力学性能等试验数据，采用 JSON 格式存储，并由 FTP协议上传至数据库服务器。预处理模块能够对异常数据进行清洗和归一化处理，生成格式统一的训练数据集。在模型训练阶段，系统调用 TensorFlow 框架，构建包括 CNN、LSTM 等在内的混合神经网络，并使用 Adam 优化器和交叉熵损失函数对模型进行训练和调优，最终得到配比优化任务的预测模型。

　　Among them, the data collection module is used to collect and annotate experimental data such as raw material properties, proportioning parameters, and mechanical properties. It is stored in JSON format and uploaded to the database server via FTP protocol. The preprocessing module can clean and normalize abnormal data, generating a uniformly formatted training dataset. In the model training phase, the system calls the TensorFlow framework to build a hybrid neural network including CNN, LSTM, etc., and uses the Adam optimizer and cross entropy loss function to train and optimize the model, ultimately obtaining a predictive model for the proportioning optimization task.

　　预测模块用于接收用户输入的目标性能指标，结合已训练的深度学习模型，搜索满足约束条件的最优配比方案。优化结果经过可视化处理后，通过 Web 接口返回用户，并与试验结果进行比较，以动态调整模型参数，持续提升系统性能。2.2 配比优化过程

　　The prediction module is used to receive target performance indicators input by the user, combined with a trained deep learning model, to search for the optimal matching solution that meets the constraints. After the optimization results are visualized, they are returned to the user through a web interface and compared with the experimental results to dynamically adjust the model parameters and continuously improve system performance. 2.2 Proportion optimization process

　　2.2.1 建立参数标准化矩阵在冷再生沥青路面材料配比优化系统中，建立参数标准化矩阵是生成智能配比方案的关键一环。系统采用特征工程手段，提取旧沥青混合料（RAP）的针入度、软化点和黏度等物理性质参数，并结合 RAP 的粒径分布、矿物成分等形态学特征，构建原料性质描述矩阵 XRAP。同理，新集料和沥青乳化剂的性质参数也被编码为矩阵XAgg和XAE。不同性质参数的量纲差异显著，为了消除量纲影响，系统会采用极差归一化算法对各参数进行无量纲化处理 [4]。极差归一化计算过程如公式（1）所示。值；max（x）和 min（x）分别为该参数的最大值和最小值。考虑环境因素的动态变化对材料性能的影响，系统引入了滑动窗口机制，通过在归一化后的矩阵中加入温度和湿度等环境参数，动态补偿环境因素的耦合效应。滑动窗口的尺寸 w 根据环境因素的变化周期进行自适应调整，不同环境因素下的滑动窗口尺寸见表 1。通过融合材料性质和环境因素，系统最终建立了覆盖全工况的多维参数描述矩阵 X，如公式（2）所示。基于标准化参数矩阵 X，系统构建了一个可迭代更新的参数知识库。当新的原料组分或环境条件出现时，相应的性质参数将被自动编码并进入知识库，实现了参数库的动态扩展。同时，历史配比方案的性能反馈数据也会被记录下来，用于对深度学习预测模型进行动态校准，提高模型的鲁棒性和适应性。这种参数标准化机制不仅规范了材料配比优化问题的数学描述，而且为后续的性能预测和配比优化提供了统一的输入接口，是实现全流程闭环优化的基础。

　　2.2.1 Establishing Parameter Standardization Matrix In the optimization system of cold recycled asphalt pavement material ratio, establishing parameter standardization matrix is a key step in generating intelligent ratio schemes. The system adopts feature engineering methods to extract physical property parameters such as penetration, softening point, and viscosity of old asphalt mixture (RAP), and combines morphological features such as RAP particle size distribution and mineral composition to construct a raw material property description matrix XRAP. Similarly, the property parameters of new aggregates and asphalt emulsifiers are also encoded as matrices XAgg and XAE. The dimensional differences of parameters with different properties are significant. In order to eliminate the influence of dimensionality, the system will use range normalization algorithm to dimensionless each parameter [4]. The process of range normalization calculation is shown in formula (1). Value; Max (x) and min (x) are the maximum and minimum values of the parameter, respectively. Considering the impact of dynamic changes in environmental factors on material properties, the system introduces a sliding window mechanism to dynamically compensate for the coupling effects of environmental factors by adding temperature and humidity parameters to the normalized matrix. The size w of the sliding window is adaptively adjusted according to the changing cycle of environmental factors. The sliding window sizes under different environmental factors are shown in Table 1. By integrating material properties and environmental factors, the system ultimately established a multidimensional parameter description matrix X that covers all operating conditions, as shown in formula (2). Based on the standardized parameter matrix X, the system has constructed an iteratively updated parameter knowledge base. When new raw material components or environmental conditions arise, the corresponding property parameters will be automatically encoded and entered into the knowledge base, achieving dynamic expansion of the parameter library. At the same time, the performance feedback data of historical matching schemes will also be recorded for dynamic calibration of deep learning prediction models, improving the robustness and adaptability of the models. This parameter standardization mechanism not only standardizes the mathematical description of material ratio optimization problems, but also provides a unified input interface for subsequent performance prediction and ratio optimization, which is the foundation for achieving closed-loop optimization of the entire process.

　　2.2.2 计算性能边界约束获得标准化的材料配比参数矩阵后，系统需要进一步计算配比方案的性能边界约束，以保证生成的配比满足路面工程的性能要求。本文采用深度残差网络（Deep Residual Network，即 ResNet）构建非线性映射模型，预测再生路面的动态模量衰减梯度和疲劳裂纹萌生的临界应变阈值。ResNet 通过引入恒等映射（Identity Mapping）的捷径连接（Shortcut Connection），有效解决了深层网络的退化问题，提高了模型的收敛速度和泛化能力。ResNet 的前向传播如公式（3）所示。

　　After obtaining the standardized material ratio parameter matrix, the system needs to further calculate the performance boundary constraints of the ratio scheme to ensure that the generated ratio meets the performance requirements of pavement engineering. This article uses Deep Residual Network (ResNet) to construct a nonlinear mapping model to predict the dynamic modulus attenuation gradient and critical strain threshold for fatigue crack initiation of regenerated pavement. ResNet effectively solves the degradation problem of deep networks by introducing identity mapping and shortcut connections, improving the convergence speed and generalization ability of the model. The forward propagation of ResNet is shown in formula (3).

　　同 时， 为了捕捉历史配比方案的时序演化规律，系统还集成了双向长短期记忆网络（Bidirectional Long Short-Term Memory，BiLSTM），通过前向和后向 2 个方向的信息传递，量化路面强度与耐久性间的平衡关系。基于 ResNet 和 BiLSTM预测所得性能指标边界域，系统可以快速判断新生成的配比方案是否满足工程规范要求，进而在解空间中搜索最优配比组合。2.2.3 生成帕累托最优解集获得配比方案的性能约束边界后，系统需要在约束域内搜索最优配比组合，生成兼顾路面性能和成本效益的帕累托前沿解集。本文采用量子遗传算法（Quantum Genetic Algorithm，QGA）进行配比方案多目标优化。与经典遗传算法相比，QGA利用量子比特（Qubit）编码表示个体，通过量子旋转门操作进行基因概率搜索，大幅提高了算法的收敛速度和全局寻优能力。

　　At the same time, in order to capture the temporal evolution of historical proportioning schemes, the system also integrates a Bidirectional Long Short Term Memory (BiLSTM) network, which quantifies the balance between road strength and durability through information transmission in both forward and backward directions. Based on ResNet and BiLSTM prediction, the performance index boundary domain can be used to quickly determine whether the newly generated proportioning scheme meets the engineering specification requirements, and then search for the optimal proportioning combination in the solution space. After generating the Pareto optimal solution set and obtaining the performance constraint boundary of the proportioning scheme, the system needs to search for the optimal proportioning combination within the constraint domain to generate a Pareto frontier solution set that balances road performance and cost-effectiveness. This article uses Quantum Genetic Algorithm (QGA) for multi-objective optimization of proportioning schemes. Compared with classical genetic algorithms, QGA uses quantum bit encoding to represent individuals and performs gene probability search through quantum rotation gate operations, greatly improving the convergence speed and global optimization ability of the algorithm.

　　QGA 的量子编码形式如公式（4）所示。在优化过程中，系统随机生成一组量子个体，并将其解码为初始配比方案，进而进行适应度评估、量子旋转门更新和量子测量等一系列遗传操作，不断更新种群，逐步逼近帕累托最优解集 [5]。为了平衡配比方案的局部搜索和全局探索能力，系统还引入了种群分化机制，根据个体的适应度水平动态调整搜索半径，平衡收敛速度和解的多样性。优化流程如图 2 所示。

　　The quantum encoding form of QGA is shown in formula (4). In the optimization process, the system randomly generates a set of quantum individuals and decodes them into an initial matching scheme, which then performs a series of genetic operations such as fitness evaluation, quantum rotation gate update, and quantum measurement, continuously updating the population and gradually approaching the Pareto optimal solution set [5]. In order to balance the local search and global exploration capabilities of the proportioning scheme, the system also introduces a population differentiation mechanism, dynamically adjusting the search radius based on individual fitness levels, balancing convergence speed and diversity of solutions. The optimization process is shown in Figure 2.

　　最终，QGA 输出的帕累托前沿解集为决策者提供了丰富的优选方案，既满足了路面性能要求，又兼顾了材料成本和环保效益。

　　In the end, the Pareto front solution set output by QGA provides decision-makers with a rich selection of optimal solutions, which not only meet the requirements of road performance, but also take into account material costs and environmental benefits.

　　3配比优化系统的工程应用

　　Engineering Application of the 3-Ratio Optimization System

　　3.1 试验场景选取与试验方案本试验选取某城市郊区的一段二级公路作为试验场景。该路段年均交通量约为 5000 辆/日，气候条件为亚热带季风气候，年均气温为 10℃ ~28℃，湿度变化范围为 60%~90%，具有典型的冷再生沥青路面施工环境。试验目标是评估系统生成的配比方案在实际工程中的性能表现，包括动态模量、水稳定性、抗裂性和成本效益等指标。试验变量设置包括旧沥青混合料（RAP）掺量（40%~70%）、新集料级配比例（0.5 ∶ 1~1.5 ∶ 1）、沥青乳化剂类型（慢裂型和中裂型）以及环境温度（10℃、20℃和 30℃）。在试验实施阶段，首先，利用“数据采集模块”将现场原材料的关键性质参数（例如 RAP 的针入度、软化点和粒径分布等）输入系统，结合实时采集的环境温 / 湿度信息，系统会自动构建标准化参数矩阵，并上传至中央数据库。其次，系统调用已训练完成的 ResNet-BiLSTM 深度学习混合模型，根据当前输入参数预测关键路面性能指标（例如动态模量、疲劳寿命等），并根据工程目标设定性能约束条件。模型预测完成后，系统启动“配比优化模块”，调用量子遗传算法（QGA），在满足性能边界的前提下搜索最优材料配比组合。在优化过程中，系统会动态调整量子旋转角度，以加快收敛速度，并利用种群分化机制控制局部最优陷阱。优化完成后，系统输出一组帕累托最优配比解集，并通过 Web 平台，以图表和表格形式展示各方案的性能与成本权衡关系。项目技术人员可以根据目标权重（例如耐久性或成本控制）选择具体的配比方案。试验操作步骤如下 ：

　　3.1 Selection of test scenario and test scheme This test selects a section of second-class highway in the suburb of a city as the test scenario. The average annual traffic volume of this section is about 5000 vehicles per day, and the climate conditions are subtropical monsoon climate. The average annual temperature is 10 ℃~28 ℃, and the humidity range is 60%~90%. It has a typical cold recycled asphalt pavement construction environment. The experimental objective is to evaluate the performance of the system generated proportioning scheme in practical engineering, including indicators such as dynamic modulus, water stability, crack resistance, and cost-effectiveness. The experimental variable settings include the content of old asphalt mixture (RAP) (40%~70%), the gradation ratio of new aggregate (0.5:1~1.5:1), the type of asphalt emulsifier (slow cracking and medium cracking), and the ambient temperature (10 ℃, 20 ℃, and 30 ℃). In the experimental implementation phase, firstly, the "data acquisition module" is used to input the key property parameters of the on-site raw materials (such as RAP penetration, softening point, and particle size distribution) into the system. Combined with real-time collected environmental temperature/humidity information, the system will automatically construct a standardized parameter matrix and upload it to the central database. Secondly, the system calls the trained ResNet BiLSTM deep learning hybrid model to predict key pavement performance indicators (such as dynamic modulus, fatigue life, etc.) based on current input parameters, and sets performance constraints based on engineering objectives. After the model prediction is completed, the system starts the "ratio optimization module" and calls the quantum genetic algorithm (QGA) to search for the optimal material ratio combination while satisfying the performance boundary. During the optimization process, the system dynamically adjusts the quantum rotation angle to accelerate convergence speed and utilizes population differentiation mechanisms to control local optimal traps. After optimization is completed, the system outputs a set of Pareto optimal allocation solutions and displays the performance and cost trade-off relationship of each solution in the form of charts and tables through the web platform. Project technicians can choose specific proportioning schemes based on target weights, such as durability or cost control. The experimental operation steps are as follows:

　　1）在实验室制备不同配比的冷再生混合料样本。

　　1) Prepare cold recycled mixture samples with different ratios in the laboratory.

　　2）在标准试验条件下测量其力学性能，并记录试验数据。

　　2) Measure its mechanical properties under standard test conditions and record the test data.

　　3）将实测数据输入系统进行模型校准。系统根据反馈数据对基础深度学习模型进行微调，自动更新参数权重，并将试验结果写入知识库，提升后续预测的准确性与适应性，保证模型能够持续学习和演化。控制条件包括恒温 / 恒湿养护环境（温度为 20℃ ±2℃，湿度为 70%±5%），加载频率为 10Hz 的动态模量测试，并采用 Marshall 试验测定水稳定性。采用回归分析和误差评估进行数据分析，评价指标包括动态模量（≥ 3000MPa）、劈裂强度（≥ 0.8MPa）、冻融劈裂强度比（≥ 75%）、疲劳寿命（≥ 106 次循环）和单位成本（≤ 300 元 /m）。试验共采集了 200 组数据，以保证试验的准确性和可重复性。

　　3) Input the measured data into the system for model calibration. The system fine tunes the basic deep learning model based on feedback data, automatically updates parameter weights, and writes experimental results into the knowledge base to improve the accuracy and adaptability of subsequent predictions, ensuring that the model can continue to learn and evolve. The control conditions include a constant temperature/humidity curing environment (temperature of 20 ℃± 2 ℃, humidity of 70% ± 5%), dynamic modulus testing with a loading frequency of 10Hz, and Marshall test to determine water stability. Regression analysis and error evaluation were used for data analysis, with evaluation indicators including dynamic modulus (≥ 3000MPa), splitting strength (≥ 0.8MPa), freeze-thaw splitting strength ratio (≥ 75%), fatigue life (≥ 106 cycles), and unit cost (≤ 300 yuan/m). A total of 200 sets of data were collected for the experiment to ensure its accuracy and reproducibility.

　　3.2 结果讨论实施上述试验方案共获得 200 组试验数据，包括动态模量、劈裂强度、冻融劈裂强度比、疲劳寿命和单位成本等关键指标，见表 2。

　　3.2 Results Discussion: A total of 200 sets of experimental data were obtained through the implementation of the above experimental plan, including key indicators such as dynamic modulus, splitting strength, freeze-thaw splitting strength ratio, fatigue life, and unit cost, as shown in Table 2.

　　试验结果表明，系统生成配比方案的多个性能指标均达到预设目标值，验证了系统的实际应用价值。在系统优化运行过程中，模型预测出的动态模量、疲劳寿命结果与实测值的平均误差分别为 3.2% 和 5.7%，验证了深度学习模型的预测精度。量子遗传算法在第 28 代种群迭代过程中即达到稳定最优解，整体优化时间不超过 12min，计算效率优异。同时，通过用户交互界面，技术人员可以实时调整性能指标权重（例如增加水稳定性权重），系统能够迅速重构目标函数并生成新的优化解集，体现了其高度的定制化能力。

　　The experimental results show that multiple performance indicators of the system generated proportioning scheme have reached the preset target values, verifying the practical application value of the system. During the system optimization process, the average errors between the predicted dynamic modulus and fatigue life results of the model and the measured values were 3.2% and 5.7%, respectively, which verified the prediction accuracy of the deep learning model. The quantum genetic algorithm achieved a stable optimal solution during the 28th generation population iteration process, with an overall optimization time of no more than 12 minutes and excellent computational efficiency. At the same time, through the user interaction interface, technicians can adjust the weight of performance indicators in real time (such as increasing the weight of water stability), and the system can quickly reconstruct the objective function and generate new optimization solution sets, reflecting its high degree of customization ability.

　　从表 2 数据可以看出，所有配比方案的动态模量均超过3000MPa，满足预设目标值 ；劈裂强度均大于 0.8MPa，最高为 0.86MPa，表明系统生成的配比方案具有较高的抗压性能 ；冻融劈裂强度比均高于 75%，体现了良好的水稳定性 ；疲劳寿命均超过 106 次循环，具有较强的抗疲劳能力；单位成本均低于 300 元 /m，符合经济性要求。

　　From the data in Table 2, it can be seen that the dynamic modulus of all proportioning schemes exceeds 3000MPa, meeting the preset target value; The splitting strength is all greater than 0.8 MPa, with the highest being 0.86 MPa, indicating that the proportion scheme generated by the system has high compressive performance; The freeze-thaw splitting strength ratio is higher than 75%, indicating good water stability; The fatigue life exceeds 106 cycles, indicating strong resistance to fatigue; The unit cost is all below 300 yuan/m, which meets the economic requirements.

　　在进一步分析中，系统利用深度模型的敏感性分析功能，识别出“新集料级配比例”和“乳化剂类型”对疲劳寿命、劈裂强度影响最大，提示配比优化应优先调控这 2 个变量。系统还根据回归残差自动提示某些试验样本可能存在测量误差，辅助人工复查，提高了试验的质量控制水平。

　　In further analysis, the system utilizes the sensitivity analysis function of the deep model to identify that the "new aggregate gradation ratio" and "emulsifier type" have the greatest impact on fatigue life and splitting strength, suggesting that the optimization of mix proportion should prioritize the control of these two variables. The system also automatically prompts for possible measurement errors in certain experimental samples based on regression residuals, assisting manual review and improving the quality control level of the experiment.

　　4 结语

　　4 Conclusion

　　本文构建了冷再生沥青路面材料配比优化系统，采用深度学习技术，解决了传统经验配比的局限性问题。系统通过参数标准化矩阵建立、性能边界约束计算以及帕累托最优解生成，实现了配比方案的智能化生成与动态优化。工程应用试验表明，该系统能够有效平衡路面性能和成本效益，具有良好的实用价值。

　　This article constructs a cold recycled asphalt pavement material proportioning optimization system, which uses deep learning technology to solve the limitations of traditional empirical proportioning. The system achieves intelligent generation and dynamic optimization of proportioning schemes through parameter standardization matrix establishment, performance boundary constraint calculation, and Pareto optimal solution generation. Engineering application tests have shown that the system can effectively balance road performance and cost-effectiveness, and has good practical value.