Artificial intelligence assisted intelligent scheduling method and system for embankment filling and excavation operation

Abstract

The application discloses a kind of based on artificial intelligence auxiliary subgrade filling and digging operation intelligent scheduling method and system, the moisture content distribution data in soil is collected by multispectral imager, and laser point cloud data is obtained by laser sensor, moisture content distribution data and laser point cloud data are fused to generate high-dimensional geological feature tensor;The contribution degree of different sensor data in the real-time state data of the subgrade is dynamically weighted using the attention mechanism;A hybrid model is established, and the cross-modal subgrade data is input into the hybrid model;Based on the geologic stability thermodynamic map, the MOEA / D multi-objective evolutionary algorithm is used to initialize the population, the GWO grey wolf optimization algorithm is used for local search, and the convergence factor is set dynamically adjusted, to obtain the subgrade filling and digging scheduling scheme. It can be adjusted in time according to the actual operation condition, to ensure that the equipment is always in an efficient cooperative operation state, effectively improving the construction efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of engineering construction technology, in particular to a roadbed filling and digging operation intelligent scheduling method and system based on artificial intelligence assistance. BACKGROUND

[0002] In the field of road construction, the traditional roadbed filling and digging operation scheduling method relies on manual experience and simple mechanical equipment, and has many drawbacks. In the data collection and analysis link, the traditional method is difficult to comprehensively and accurately obtain roadbed geological information. For soil moisture content detection, local sampling method is often used, which cannot present the overall moisture content distribution of the entire roadbed area; the collection of topographic data is also limited to single-point measurement by devices such as total station, and the data update is lagging, which is difficult to reflect the real-time changes in the construction process. The operation arrangement between devices often relies on manual coordination, which cannot be dynamically adjusted according to geological conditions and construction progress, resulting in low utilization rate of equipment and low construction efficiency. SUMMARY

[0003] The purpose of the present application is to solve the above problems, and a roadbed filling and digging operation intelligent scheduling method based on artificial intelligence assistance is designed.

[0004] To achieve the above purpose, the technical scheme of the present application is that, further, in the roadbed filling and digging operation intelligent scheduling method based on artificial intelligence assistance, the roadbed filling and digging operation intelligent scheduling method comprises the following steps:

[0005] Collecting soil moisture content distribution data by a multispectral imager, and obtaining laser point cloud data by a laser sensor, fusing the moisture content distribution data and the laser point cloud data to generate a high-dimensional geological feature tensor, and obtaining roadbed real-time state data;

[0006] Using an attention mechanism to dynamically weight the contribution of different sensor data in the roadbed real-time state data, and obtaining cross-modal roadbed data;

[0007] Establishing a Transformer-GNN hybrid model, inputting the cross-modal roadbed data into the Transformer-GNN hybrid model, and generating a geological stability heat map;

[0008] Based on the geological stability heat map, initializing a population using a MOEA / D multi-objective evolutionary algorithm, performing local search using a GWO grey wolf optimization algorithm, setting a dynamically adjusted convergence factor, and obtaining a roadbed filling and digging scheduling scheme.

[0009] Further, in the roadbed filling and digging operation intelligent scheduling method based on artificial intelligence assistance, the collecting soil moisture content distribution data by a multispectral imager, and obtaining laser point cloud data by a laser sensor comprises:

[0010] A calibration model between soil moisture content and spectral reflectance is established by collecting soil samples with different moisture contents in the construction area to obtain spectral data and actual moisture content;

[0011] The image data collected by the multispectral imager is input into the inversion model to calculate the soil moisture content corresponding to each pixel point and generate moisture content distribution data for the entire roadbed area.

[0012] By measuring the time difference from emission to reception of the laser, the distance from the target object to the sensor is calculated, and combined with the scanning angle and rotation speed of the sensor, the three-dimensional coordinates of the target object surface are obtained, and the laser point cloud data is obtained.

[0013] Further, in the above-mentioned intelligent scheduling method for roadbed filling and excavation operation based on artificial intelligence assistance, the fusion of the moisture content distribution data and the laser point cloud data to generate high-dimensional geological feature tensors to obtain real-time roadbed state data includes:

[0014] Convert the moisture content distribution data and the laser point cloud data into a unified global coordinate system.

[0015] Extract moisture features from the moisture content distribution data, including at least mean, variance, and gradient, and extract terrain and topographic features from the point cloud data, including at least elevation, slope, aspect, and curvature.

[0016] Dimensionally extend the moisture features, terrain features, and topographic features to form a multi-dimensional feature vector, and use tensor product operation to fuse the feature vector into high-dimensional geological feature tensor data to obtain real-time roadbed state data.

[0017] Further, in the above-mentioned intelligent scheduling method for roadbed filling and excavation operation based on artificial intelligence assistance, the use of attention mechanism to dynamically weight the contribution of different sensor data in the real-time roadbed state data to obtain cross-modal roadbed data includes:

[0018] For each head in the real-time roadbed state data, calculate the similarity between Query and Key, and use the dot product similarity formula to calculate the matrix to obtain the similarity matrix.

[0019] Perform Softmax normalization on the similarity matrix to obtain an attention weight matrix with weight values ranging from 0 to 1.

[0020] According to the attention weight, the Value matrix is weighted and summed to obtain a cross-modal feature vector for each spatio-temporal grid point.

[0021] Further, in the above-mentioned intelligent scheduling method for roadbed filling and digging operation based on artificial intelligence assistance, the Transformer-GNN hybrid model is established, the cross-modal roadbed data is input into the Transformer-GNN hybrid model, and a geological stability heat map is generated, including:

[0022] The Transformer-GNN hybrid model is established, and the cross-modal roadbed data is input into the Transformer-GNN hybrid model;

[0023] The cross-modal roadbed data is segmented according to time sequence by using the Transformer encoder, the position encoding of the data at each time step is performed, and the time information is integrated into the model. The position encoding is generated by using sine and cosine functions;

[0024] In the multi-head self-attention layer, the self-attention calculation is performed on the cross-modal data at each time step to capture the dependency relationship between different time steps. The FFN layer is used to perform nonlinear transformation on the attention output to extract the time sequence features;

[0025] The construction equipment is regarded as a node in the graph, the node attributes include the type, position and working state of the equipment, the GCN graph convolution network is used for graph convolution operation to obtain the equipment node features.

[0026] Further, in the above-mentioned intelligent scheduling method for roadbed filling and digging operation based on artificial intelligence assistance, the Transformer-GNN hybrid model is established, the cross-modal roadbed data is input into the Transformer-GNN hybrid model, and a geological stability heat map is generated, further including:

[0027] The time sequence features output by the Transformer encoder and the equipment node features output by the GNN are fused, and the geological stability evaluation index is obtained by mapping through the fully connected neural network, including at least soil bearing capacity, settlement risk and sliding possibility.

[0028] Further, in the above-mentioned intelligent scheduling method for roadbed filling and digging operation based on artificial intelligence assistance, the geological stability heat map is used to initialize the population by using the MOEA / D multi-objective evolutionary algorithm, and the local search is performed by using the GWO grey wolf optimization algorithm. A dynamic adjustment convergence factor is set to obtain a roadbed filling and digging scheduling scheme, including:

[0029] A multi-objective optimization model is established, and the objectives include maximizing the filling and digging operation efficiency, minimizing the cost and optimizing the quality;

[0030] The subgrade filling and digging scheduling scheme is expressed as a chromosome, the chromosome code includes allocation of construction equipment, operation path planning and filling and digging volume allocation, and a population is initialized in a random generation manner;

[0031] The GWO improved grey wolf optimization algorithm is used for neighborhood search of each individual, neighborhood operations include equipment reallocation, operation path adjustment and filling and digging volume fine adjustment, and a neighborhood individual is generated each time;

[0032] The objective function values of the neighborhood individuals are calculated, the optimal individual is selected to replace the original individual, and the subgrade filling and digging scheduling scheme is obtained through iteration by dynamically adjusting the convergence factor.

[0033] Further, in an artificial intelligence assisted subgrade filling and digging operation intelligent scheduling system, the subgrade filling and digging operation intelligent scheduling system comprises the following modules:

[0034] A subgrade data acquisition module is configured to acquire water content distribution data in soil through a multispectral imager, acquire laser point cloud data through a laser sensor, fuse the water content distribution data and the laser point cloud data to generate a high-dimensional geological feature tensor, and obtain real-time state data of the subgrade;

[0035] A contribution calculation module is configured to dynamically weight the contribution of different sensor data in the real-time state data of the subgrade by using an attention mechanism to obtain cross-modal subgrade data;

[0036] A stability prediction module is configured to establish a Transformer-GNN hybrid model, input the cross-modal subgrade data into the Transformer-GNN hybrid model, and generate a geological stability heat map;

[0037] A scheduling scheme generation module is configured to initialize a population based on the geological stability heat map by using a MOEA / D multi-objective evolutionary algorithm, perform local search by using a GWO grey wolf optimization algorithm, set a dynamically adjusted convergence factor, and obtain a subgrade filling and digging scheduling scheme.

[0038] Further, in the system for implementing the above-mentioned artificial intelligence assisted subgrade filling and digging operation intelligent scheduling method, the subgrade data acquisition module comprises the following submodules:

[0039] An establishment submodule is configured to establish a calibration model between soil water content and spectral reflectance, acquire spectral data and actual water content by collecting soil samples with different water contents in a construction area;

[0040] A calculation submodule is configured to input image data acquired by the multispectral imager into an inversion model, calculate the soil water content corresponding to each pixel point, and generate water content distribution data of the entire subgrade area;

[0041] The acquisition sub-module is configured to acquire the three-dimensional coordinates of the target object surface by measuring the time difference between the emission and the reception of the laser, and combining the scanning angle and the rotation speed of the sensor, and obtain the laser point cloud data.

[0042] Further, in the system for implementing the above-mentioned intelligent scheduling method for roadbed filling and digging operation based on artificial intelligence assistance, the roadbed data acquisition module further comprises the following sub-modules:

[0043] The conversion sub-module is configured to convert the water content distribution data and the laser point cloud data into a unified global coordinate;

[0044] The extraction sub-module is configured to extract moisture features from the water content distribution data, including at least the mean value, the variance and the gradient of the water content, and extract terrain features and topographic features from the point cloud data, including at least the elevation, the slope, the slope direction and the curvature.

[0045] The fusion sub-module is configured to perform dimension expansion on the moisture features, the terrain features and the topographic features to form a multi-dimensional feature vector, and fuse the feature vector into high-dimensional geological feature tensor data by using tensor product operation, so as to obtain the real-time state data of the roadbed.

[0046] The beneficial effects are as follows: the water content distribution data in the soil is acquired by the multispectral imager, the laser point cloud data is acquired by the laser sensor, the water content distribution data and the laser point cloud data are fused to generate high-dimensional geological feature tensors, and the real-time state data of the roadbed is obtained; the contribution degrees of different sensor data in the real-time state data of the roadbed are dynamically weighted by using the attention mechanism, and the cross-modal roadbed data is obtained; the Transformer-GNN hybrid model is established, the cross-modal roadbed data is input into the Transformer-GNN hybrid model, and the geological stability heat map is generated; based on the geological stability heat map, the MOEA / D multi-objective evolutionary algorithm is used to initialize the population, the GWO grey wolf optimization algorithm is used for local search, and the dynamically adjusted convergence factor is set, so as to obtain the roadbed filling and digging scheduling scheme. 1. The geological stability heat map is generated, so that the construction personnel can intuitively and clearly master the stability of each region of the roadbed, and provide a visual basis for scientific decision-making. 2. The Pareto optimal solution that simultaneously satisfies the maximization of operation efficiency, the minimization of cost and the optimization of quality and other multi-objectives can be quickly found in a complex scheduling scheme space, compared with the traditional single-objective optimization method, the comprehensive performance of the scheduling scheme is significantly improved. 3. The scheduling scheme can be adjusted in a timely manner according to the actual operation situation, so that the equipment is always in a high-efficiency collaborative operation state, the construction efficiency is effectively improved, and the construction cost is reduced. At the same time, by continuously optimizing the construction process parameters, the quality stability of the roadbed filling and digging operation is ensured to meet the standards, and a foundation is laid for the long-term safe use of the road engineering. BRIEF DESCRIPTION OF DRAWINGS

[0047] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings. The drawings are for purposes of illustration only and are not intended to be limiting of the application.

[0048] Figure 1 A first embodiment of a roadbed filling and digging operation intelligent scheduling method based on artificial intelligence assistance according to an embodiment of the present application is shown in the figure.

[0049] Figure 2 A second embodiment of a roadbed filling and digging operation intelligent scheduling method based on artificial intelligence assistance according to an embodiment of the present application is shown in the figure.

[0050] Figure 3 A first embodiment of a roadbed filling and digging operation intelligent scheduling system based on artificial intelligence assistance according to an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0051] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application.

[0052] Those skilled in the art can understand that the singular forms "a", "an" and "the" used herein include plural forms unless specifically stated otherwise. It should be further understood that the use of the term "including" in the specification of the present application means that the features, integers, steps, operations, elements and / or components exist, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0053] The present application will be described in detail below with reference to the accompanying drawings, such as Figure 1 A roadbed filling and digging operation intelligent scheduling method based on artificial intelligence assistance includes the following steps:

[0054] Step 101, acquiring water content distribution data in the soil by a multispectral imager, and obtaining laser point cloud data by a laser sensor, fusing the water content distribution data and the laser point cloud data to generate a high-dimensional geological feature tensor, and obtaining real-time state data of the roadbed;

[0055] Specifically, a calibration model between soil water content and spectral reflectance is established in the present embodiment, soil samples with different water contents are collected in the construction area, and spectral data and actual water content are obtained.

[0056] The image data collected by the multispectral imager is input into the inversion model to calculate the soil moisture content corresponding to each pixel point, and generate the moisture content distribution data of the entire roadbed area;

[0057] By measuring the time difference from emission to reception of the laser, the distance from the target object to the sensor is calculated, and combined with the scanning angle and rotation speed of the sensor, the three-dimensional coordinates of the target object surface are obtained, and the laser point cloud data is obtained.

[0058] Convert the moisture content distribution data and laser point cloud data into a unified global coordinate system;

[0059] Extract moisture features from the moisture content distribution data, including at least mean, variance, and gradient, and extract terrain and topographic features from the point cloud data, including at least elevation, slope, aspect, and curvature;

[0060] Expand the moisture, terrain, and topographic features to form a multi-dimensional feature vector, and use tensor product operation to fuse the feature vector into high-dimensional geological feature tensor data to obtain real-time roadbed state data.

[0061] Step 102, use the attention mechanism to dynamically weight the contribution of different sensor data in the real-time roadbed state data to obtain cross-modal roadbed data;

[0062] Specifically, in the present embodiment, the similarity between Query and Key is calculated for each head in the real-time roadbed state data, and the dot product similarity formula is used to calculate the matrix to obtain the similarity matrix;

[0063] The similarity matrix is normalized by Softmax to obtain an attention weight matrix, and the weight value ranges from 0 to 1;

[0064] According to the attention weight, the Value matrix is weighted and summed to obtain the cross-modal feature vector of each spatio-temporal grid point.

[0065] Step 103, establish a Transformer-GNN hybrid model, input the cross-modal roadbed data into the Transformer-GNN hybrid model, and generate a geological stability heat map;

[0066] Specifically, in the present embodiment, a Transformer-GNN hybrid model is established, and the cross-modal roadbed data is input into the Transformer-GNN hybrid model;

[0067] The cross-modal roadbed data is segmented by time sequence using the Transformer encoder, and the position encoding of each time step data is performed to integrate the time information into the model, and the position encoding is generated using sine and cosine functions;

[0068] In the multi-head self-attention layer, self-attention calculation is performed on the cross-modal data at each time step to capture the dependencies between different time steps. The attention output is nonlinearly transformed using the FFN layer to extract temporal features.

[0069] The construction equipment is considered as nodes in the graph. The node attributes include the type, location, and working status of the equipment. The GCN graph convolutional network is used to perform graph convolution operations to obtain the equipment node features.

[0070] The temporal features output by the Transformer encoder and the device node features output by the GNN are fused and mapped through a fully connected neural network to obtain geological stability assessment indicators, including at least soil bearing capacity, settlement risk, and sliding probability.

[0071] Step 104: Based on the geological stability heat map, initialize the population using the MOEA / D multi-objective evolutionary algorithm, perform local search using the GWO gray wolf optimization algorithm, set a dynamically adjusted convergence factor, and obtain the roadbed filling and cutting scheduling scheme.

[0072] Specifically, in this embodiment, a multi-objective optimization model is established, with objectives including maximizing the efficiency of cut and fill operations, minimizing costs, and optimizing quality.

[0073] The roadbed filling and excavation scheduling scheme is represented as a chromosome. The chromosome encoding includes the allocation of construction equipment, operation path planning and filling and excavation volume allocation. The population is initialized using a random generation method.

[0074] The improved Grey Wolf Optimization Algorithm (GWO) is used to perform neighborhood search for each individual. Neighborhood operations include equipment reallocation, job path adjustment, and fill / cut volume fine-tuning. Each operation generates a neighborhood individual.

[0075] The objective function value of the neighboring individuals is calculated, the optimal individual is selected to replace the original individual, and the roadbed filling and excavation scheduling scheme is obtained through iteration by dynamically adjusting the convergence factor.

[0076] Its beneficial effects are as follows: 1. It generates geological stability heat maps, enabling construction personnel to intuitively and clearly grasp the stability of various areas of the roadbed, providing a visual basis for scientific decision-making. 2. It can quickly find Pareto optimal solutions that simultaneously satisfy multiple objectives such as maximizing operational efficiency, minimizing costs, and optimizing quality within a complex scheduling scheme space, significantly improving the overall performance of the scheduling scheme compared to traditional single-objective optimization methods. 3. It can adjust the scheduling scheme in a timely manner according to actual operational conditions, ensuring that equipment is always in a state of efficient and coordinated operation, effectively improving construction efficiency and reducing construction costs. Simultaneously, by continuously optimizing construction process parameters, it ensures that the quality of roadbed filling and excavation operations consistently meets standards, laying the foundation for the long-term safe use of road projects.

[0077] Please seeFigure 2 In an artificial intelligence assisted roadbed filling and digging operation intelligent scheduling method, the water content distribution data in the soil is collected by a multispectral imager, and the laser point cloud data is obtained by a laser sensor, including the following steps:

[0078] Step 201, a calibration model between soil water content and spectral reflectance is established, spectral data and actual water content are obtained by collecting soil samples with different water contents in the construction area;

[0079] Step 202, input the image data collected by the multispectral imager into the inversion model, calculate the soil water content corresponding to each pixel point, and generate the water content distribution data of the entire roadbed area;

[0080] Step 203, calculate the distance from the target object to the sensor by measuring the time difference from the emission to the reception of the laser, combine the scanning angle and rotation speed of the sensor, obtain the three-dimensional coordinates of the target object surface, and get the laser point cloud data.

[0081] Please refer to Figure 3 In an artificial intelligence assisted roadbed filling and digging operation intelligent scheduling system, the roadbed filling and digging operation intelligent scheduling system includes the following modules:

[0082] The roadbed data acquisition module is used to collect the water content distribution data in the soil by a multispectral imager, and to obtain laser point cloud data by a laser sensor, to fuse the water content distribution data and the laser point cloud data to generate a high-dimensional geological feature tensor, and to obtain real-time state data of the roadbed;

[0083] The contribution calculation module is used to dynamically weight the contribution of different sensor data in the roadbed real-time state data by using the attention mechanism, to obtain cross-modal roadbed data;

[0084] The stability prediction module is used to establish a Transformer-GNN hybrid model, input the cross-modal roadbed data into the Transformer-GNN hybrid model, and generate a geological stability heat map;

[0085] The scheduling scheme generation module is used to initialize the population based on the geological stability heat map, use the MOEA / D multi-objective evolutionary algorithm for local search, set a dynamically adjusted convergence factor, and obtain a roadbed filling and digging scheduling scheme.

[0086] Specifically, the present embodiment can also be implemented in the following way:

[0087] (I) Multispectral imager data acquisition

[0088] Moisture content distribution data extraction: Spectral inversion algorithms, such as partial least squares (PLS) or neural network algorithms, are used to process image data in each band. First, a calibration model between soil moisture content and spectral reflectance is established. This is done by collecting soil samples with different moisture contents in the construction area, obtaining their spectral data and actual moisture content, and training a high-precision inversion model. Then, image data acquired by a multispectral imager is input into the inversion model to calculate the soil moisture content corresponding to each pixel, generating moisture content distribution data for the entire roadbed area with a resolution of 0.5-1 meter.

[0089] (ii) Laser sensor data acquisition

[0090] Point cloud data acquisition: The laser sensor emits laser pulses, and the distance from the target object to the sensor is calculated by measuring the time difference (TOF) or phase difference between emission and reception. Simultaneously, combined with the sensor's scanning angle and rotation speed, the three-dimensional coordinates (X, Y, Z) of the target object's surface are acquired, forming laser point cloud data. The scanning frequency is adjusted according to the operating speed and accuracy requirements of the construction equipment, generally between 50-200Hz, to ensure that the point cloud data can reflect the real-time topographic changes of the roadbed surface.

[0091] (III) Data fusion to generate high-dimensional geological feature tensor data

[0092] Spatiotemporal registration: First, the moisture content distribution data acquired by the multispectral imager and the point cloud data acquired by the laser sensor are transformed into a unified global coordinate system. For sensors on UAVs and construction equipment, their position and attitude information are acquired using the Global Positioning System (GPS) and Inertial Measurement Unit (IMU), and coordinate transformation and attitude correction are performed. For fixed sensors, their position and orientation in the global coordinate system are determined through manual measurement or equipment such as a total station.

[0093] Feature extraction: Moisture-related features, such as mean, variance, and gradient, are extracted from moisture content distribution data to reflect the spatial distribution and variation trends of soil moisture. Topographic features, such as elevation, slope, aspect, and curvature, as well as landform features, such as surface roughness, protrusion height, and depression depth, are extracted from point cloud data. Simultaneously, intensity features of the point cloud can also be extracted to reflect differences in surface material.

[0094] Tensor fusion: The extracted moisture content features and point cloud features are dimensionally expanded to form multidimensional feature vectors. Then, tensor product operations are used to fuse these feature vectors into high-dimensional geological feature tensor data. The tensor dimensions include spatial dimension (X,Y), feature dimension (moisture content, elevation, slope, etc.), and temporal dimension (different acquisition times), thereby constructing a roadbed real-time status data tensor containing spatiotemporal information and multi-source features, with dimensions of N×M×K×T, where N and M are the number of spatial grids, K is the number of features, and T is the length of the time series.

[0095] (I) Construction of Attention Mechanism Model

[0096] A multi-head self-attention model is employed, with a model structure comprising three input matrices: query, key, and value. For high-dimensional geological feature tensor data, it is first flattened into a two-dimensional matrix of dimension (N×M×T)×K, where N×M×T is the number of spatiotemporal grids and K is the number of features. Then, a linear transformation maps the input matrices to the query, key, and value matrices, respectively. The mapping matrix for each head has a dimension of K×d_k, where d_k is the dimension of each head, typically set to 64 or 128.

[0097] (ii) Dynamic weighted calculation of contribution

[0098] Similarity calculation: For each head, calculate the similarity between Query and Key using the dot product similarity formula: Similarity=(Query・Key^T) / sqrt(d_k). The similarity matrix has dimensions of (N×M×T)×(N×M×T), where each element represents the degree of correlation between two spatiotemporal grid points.

[0099] Attention weight generation: The similarity matrix is ​​Softmax normalized to obtain the attention weight matrix. The weight values ​​range from 0 to 1, representing the contribution of each spatiotemporal grid point to the current grid point. For multimodal data, namely water content features and point cloud features, their respective attention weights are calculated, and then fused by weighted averaging or concatenation to obtain cross-modal attention weights.

[0100] Cross-modal data generation: The Value matrix is ​​weighted and summed according to the attention weights to obtain the cross-modal feature vector for each spatiotemporal grid point. After concatenating the outputs of all heads, a linear transformation is performed to restore the original feature dimensions, resulting in the final cross-modal roadbed data. This data dynamically integrates important information from different sensor data, highlighting features that have a significant impact on the current roadbed state.

[0101] (I) Model Architecture Design

[0102] The Transformer-GNN hybrid model consists of two parts: a Transformer encoder and a Graph Neural Network (GNN). The Transformer encoder is used to process time-series data and capture the changing patterns of the roadbed condition over time; the GNN is used to model the spatial topological relationships between construction equipment and analyze the mutual influence between the equipment.

[0103] (ii) Transformer encoder processes timing data

[0104] Input processing: The cross-modal roadbed data is segmented according to time series, with the input data dimension of N×M×K for each time step, where N×M is the number of spatial grids and K is the number of features. Location encoding is performed on the data at each time step to incorporate temporal information into the model. The location encoding is generated using sine and cosine functions, with the formulas: PosEncoding(pos,2i)=sin(pos / 10000^(2i / K)), PosEncoding(pos,2i+1)=cos(pos / 10000^(2i / K)), where pos is the time step position and i is the feature dimension index.

[0105] Encoding Process: The Transformer encoder consists of multiple encoding layers. Each layer includes a multi-head self-attention mechanism and a feedforward neural network (FFN), employing residual connections and layer normalization techniques. In the multi-head self-attention layer, self-attention is calculated for cross-modal data at each time step, capturing the dependencies between different time steps. The FFN layer performs a non-linear transformation on the attention output to further extract temporal features. Through multi-layer encoding, a high-level feature representation containing temporal information is obtained.

[0106] (III) GNN modeling of spatial topological relationships of construction equipment

[0107] Graph Construction: Construction equipment is considered as nodes in the graph. Node attributes include equipment type, location, and operating status (load, speed, fuel consumption, etc.). When the spatial distance between equipment is less than a set threshold (50 meters), an edge connection is established. Edge attributes include distance, communication latency, etc. The graph structure is dynamically updated as equipment moves and its operating status changes.

[0108] Graph convolution operations: Graph convolution networks (GCNs) or graph attention networks (GATs) are used for graph convolution operations. Taking GAT as an example, each node calculates its attention weights with neighboring nodes, aggregates information from neighboring nodes, and updates its own node features. The attention weight calculation considers node attributes and edge attributes, and the formula is: α_ij=Softmax(LeakyReLU(a^T[Wh_i||Wh_j])), where a is the attention parameter, W is the weight matrix, and h_i and h_j are the feature vectors of nodes i and j, respectively. Through multi-layer graph convolution, the spatial topological relationships and collaborative operation patterns between devices are captured.

[0109] (iv) Generating a geological stability heat map

[0110] The temporal features output from the Transformer encoder and the device features output from the GNN are fused and mapped to geological stability assessment indicators, such as soil bearing capacity, settlement risk, and slip probability, through a fully connected neural network. The stability indicators for each spatial grid point are normalized and converted into grayscale or color values ​​ranging from 0 to 255, generating a geological stability heatmap. The heatmap uses different color codes to represent stability levels; for example, red indicates high-risk areas, yellow indicates medium-risk areas, and green indicates low-risk areas. With a resolution of 1×1 meters, it clearly displays the stability status of different areas of the roadbed.

[0111] (a) Initializing the population using the MOEA / D multi-objective evolutionary algorithm

[0112] Objective function definition: A multi-objective optimization model is established, with objectives including maximizing the efficiency of cut and fill operations (shortest completion time, highest equipment utilization), minimizing costs (fuel consumption, equipment wear and tear, labor costs, etc.), and optimizing quality (roadbed smoothness and compaction meeting standards). The specific objective function is as follows:

[0113] Efficiency target: f1 = -1 / T, where T is the time to complete the cut and fill operation;

[0114] Cost target: f2 = Σ(C_fuel + C_wear + C_labor);

[0115] Quality target: f3=Σ(1-|s_i-s_0| / s_max), where s_i is the actual flatness or compaction of the i-th grid point, s_0 is the standard value, and s_max is the maximum allowable deviation.

[0116] Population initialization: The roadbed filling and excavation scheduling scheme is represented as a chromosome. The chromosome encoding includes the allocation of construction equipment (the number and model of excavators, loaders, and bulldozers), operation path planning (the travel route and operation sequence of each piece of equipment), and filling and excavation volume allocation (the filling or excavation volume of each grid point). The population is initialized using a random generation method, and the population size is usually set to 50-100 individuals to ensure population diversity.

[0117] (II) Local Search of the Improved Gray Wolf Optimization Algorithm by GWO

[0118] Improvements to the basic GWO algorithm: Adaptive inertia weights and dynamic learning factors are introduced into the traditional Grey Wolf optimization algorithm to enhance its local search capability and convergence speed. The inertia weight w decreases linearly with the number of iterations, as shown in the formula: w = w_max - (w_max - w_min) * t / T_max, where w_max = 0.9, w_min = 0.4, t is the current iteration number, and T_max is the maximum iteration number. Dynamic learning factors c1 and c2 are adjusted through nonlinear functions to enhance the algorithm's search capability at different iteration stages.

[0119] Local search strategy: Based on the population individuals initialized by MOEA / D, a neighborhood search is performed on each individual. Neighborhood operations include equipment reallocation, job path adjustment, and fine-tuning of cut and fill volumes, generating a certain number of neighborhood individuals (5-10) for each operation. The objective function value of the neighborhood individuals is calculated, and the optimal individual is selected to replace the original individual, achieving local optimization.

[0120] (iii) Dynamically adjust the convergence factor

[0121] The convergence factor 'a' in the GWO algorithm controls the search range. Initially set to 2, it decreases linearly to 0 with each iteration, using the formula: a = 2 - 2*t / T_max. By dynamically adjusting the convergence factor, the algorithm exhibits strong global search capabilities in the early stages and focuses on refined local searches in later stages, improving solution quality and convergence speed. After multiple iterations (typically 100-200), a Pareto-optimal solution set is obtained, from which the solution with the best overall performance is selected as the roadbed cut-and-fill scheduling scheme.

[0122] (I) Implementation of Intelligent Scheduling for Cutting and Filling Operations

[0123] Communication of the plan: The generated roadbed filling and excavation scheduling plan is sent to the vehicle-mounted terminals of the construction equipment via a wireless communication system. The vehicle-mounted terminals display the equipment's work tasks, route planning, and filling and excavation volume requirements. Simultaneously, at the monitoring center on the construction site, the scheduling plan and equipment operating status are displayed in real time on a large screen, facilitating overall monitoring and coordination by management personnel.

[0124] Equipment Collaborative Operation: Construction equipment automatically adjusts its operating parameters according to the scheduling plan, such as the digging depth of excavators, the loading capacity of loaders, and the bulldozing speed of bulldozers. Equipment interacts with each other in real time via a wireless communication network to achieve collaborative operation. For example, an excavator can directly load the excavated soil onto a designated loader, which then transports it to the filling area where a bulldozer levels it, improving operational efficiency.

[0125] (ii) Real-time feedback result collection

[0126] Sensor data acquisition: Various sensors, such as tilt sensors, pressure sensors, and position sensors, are installed on the construction equipment to collect real-time data on the equipment's operating parameters (excavation depth, loading capacity, and travel speed) and the real-time status of the roadbed (smoothness and compaction of the filled and excavated areas). Simultaneously, drones or fixed monitoring equipment are used to collect roadbed moisture content distribution data and point cloud data again, which are then compared with the data from before construction to evaluate the operational effectiveness.

[0127] Feedback information processing: The collected real-time data is transmitted to the central control system for data cleaning and preprocessing to remove noise and outliers. Then, through data analysis and model prediction, the effectiveness of the current scheduling plan is evaluated, and existing problems are identified, such as deviations in cut and fill volumes, low work efficiency, and substandard quality.

[0128] (III) Generating the target roadbed filling and excavation scheduling plan

[0129] Adjustment strategy: Based on real-time feedback results, the original scheduling scheme is adjusted using a feedback control algorithm. If the fill / cut deviation exceeds the allowable range (±5%), the fill / cut volume of each grid point is recalculated, and the equipment's task allocation is adjusted; if the operation efficiency is low, the equipment's operation path and coordination mechanism are optimized; if the quality is substandard, construction process parameters are adjusted, such as the number of compaction passes and compaction speed of the compaction machinery.

[0130] Iterative optimization: The adjusted scheduling scheme is used as new input and re-entered into the data acquisition, model processing, and optimization algorithm flow to form a closed-loop control. Through multiple iterative optimizations, the accuracy and effectiveness of the scheduling scheme are continuously improved, ultimately generating a target roadbed filling and excavation scheduling scheme that meets construction requirements, ensuring the efficient and high-quality completion of roadbed filling and excavation operations.

[0131] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent scheduling of roadbed filling and excavation operations based on artificial intelligence assistance, characterized in that, The intelligent scheduling method for roadbed filling and excavation operations includes the following steps: Soil moisture content distribution data is collected using a multispectral imager, and laser point cloud data is acquired using a laser sensor. The moisture content distribution data and laser point cloud data are then converted into a unified global coordinate system. Moisture features are extracted from the moisture content distribution data, and topographic and geomorphic features are extracted from the point cloud data. The dimensions of the moisture, topographic, and geomorphic features are expanded to form multidimensional feature vectors. Tensor product operations are used to fuse the feature vectors into high-dimensional geological feature tensor data to obtain real-time roadbed status data. The contribution of different sensor data in the real-time status data of the roadbed is dynamically weighted using an attention mechanism. For each header in the real-time status data of the roadbed, the similarity between the query and the key is calculated, and the matrix is ​​calculated using the dot product similarity formula to obtain the similarity matrix. The similarity matrix is ​​then subjected to Softmax normalization to obtain the attention weight matrix, with weight values ​​ranging from 0 to 1. The value matrix is ​​then weighted and summed according to the attention weights to obtain the cross-modal feature vector of each spatiotemporal grid point, i.e., the cross-modal roadbed data. A Transformer-GNN hybrid model is established, and the cross-modal roadbed data is input into the Transformer-GNN hybrid model. The cross-modal roadbed data is segmented into time series using a Transformer encoder, and the data at each time step is encoded with location information, integrating time information into the model. The location encoding is generated using sine and cosine functions. In a multi-head self-attention layer, self-attention calculation is performed on the cross-modal data at each time step to capture the dependencies between different time steps. The attention output is nonlinearly transformed using an FFN layer to extract temporal features. Construction equipment is regarded as nodes in a graph, and node attributes include equipment type, location, and working status. A GCN graph convolutional network is used to perform graph convolution operations to obtain equipment node features. The temporal features output by the Transformer encoder and the equipment node features output by the GNN are fused and mapped through a fully connected neural network to obtain geological stability assessment indicators, including at least soil bearing capacity, settlement risk, and sliding probability, generating a geological stability heatmap. Based on the geological stability heatmap, a multi-objective optimization model is established, with objectives including maximizing the efficiency of cut and fill operations, minimizing costs, and optimizing quality. The roadbed cut and fill scheduling scheme is represented as a chromosome, with chromosome encoding the allocation of construction equipment, operation path planning, and cut and fill volume allocation. The population is initialized using a random generation method. The improved Grey Wolf Optimization Algorithm (GWO) is used to perform a neighborhood search for each individual. Neighborhood operations include equipment reallocation, operation path adjustment, and cut and fill volume fine-tuning, with each operation generating a neighborhood individual. The objective function value of the neighborhood individuals is calculated, and the optimal individual is selected to replace the original individual. By dynamically adjusting the convergence factor, the roadbed cut and fill scheduling scheme is obtained through iteration.

2. The intelligent scheduling method for roadbed filling and excavation operations based on artificial intelligence assistance as described in claim 1, characterized in that, The process of acquiring soil moisture content distribution data using a multispectral imager and obtaining laser point cloud data using a laser sensor includes: A calibration model between soil moisture content and spectral reflectance was established by collecting soil samples with different moisture contents in the construction area to obtain spectral data and actual moisture content. The image data acquired by the multispectral imager is input into the inversion model to calculate the soil moisture content corresponding to each pixel, thereby generating moisture content distribution data for the entire roadbed area. By measuring the time difference between laser emission and reception, the distance from the target object to the sensor is calculated. Combined with the sensor's scanning angle and rotation speed, the three-dimensional coordinates of the target object's surface are obtained, resulting in laser point cloud data.

3. The intelligent scheduling method for roadbed filling and excavation operations based on artificial intelligence assistance as described in claim 1, characterized in that, The moisture characteristics include at least the mean, variance, and gradient of moisture content, while the topographic and geomorphic characteristics include at least elevation, slope, aspect, and curvature.

4. An intelligent scheduling system for roadbed filling and excavation operations based on artificial intelligence assistance, characterized in that, The intelligent scheduling system for roadbed filling and excavation operations includes the following modules: The roadbed data acquisition module is used to acquire soil moisture content distribution data through a multispectral imager and laser point cloud data through a laser sensor, and to convert the moisture content distribution data and laser point cloud data into a unified global coordinate system. Moisture features are extracted from moisture content distribution data, and topographic and geomorphic features are extracted from point cloud data. The moisture, topographic and geomorphic features are expanded in dimension to form multidimensional feature vectors. Tensor product operation is used to fuse the feature vectors into high-dimensional geological feature tensor data to obtain real-time roadbed status data. The contribution calculation module is used to dynamically weight the contribution of different sensor data in the real-time status data of the roadbed using an attention mechanism. For each header in the real-time status data of the roadbed, the similarity between Query and Key is calculated, and the matrix is ​​calculated using the dot product similarity formula to obtain the similarity matrix. The similarity matrix is ​​subjected to Softmax normalization to obtain the attention weight matrix, with weight values ​​ranging from 0 to 1; the Value matrix is ​​weighted and summed according to the attention weights to obtain the cross-modal feature vector of each spatiotemporal grid point, i.e., the cross-modal roadbed data. The stability prediction module is used to establish a Transformer-GNN hybrid model. The cross-modal roadbed data is input into the Transformer-GNN hybrid model. A Transformer encoder is used to segment the cross-modal roadbed data according to time series, and the data at each time step is positionally encoded to incorporate temporal information into the model. The position encoding is generated using sine and cosine functions. In the multi-head self-attention layer, self-attention calculation is performed on the cross-modal data at each time step to capture the dependencies between different time steps. An FFN layer is used to perform a nonlinear transformation on the attention output to extract temporal features. Construction equipment is considered as nodes in a graph, with node attributes including equipment type, location, and working status. A GCN graph convolutional network is used to perform graph convolution operations to obtain equipment node features. The temporal features output by the Transformer encoder and the device node features output by the GNN are fused and mapped through a fully connected neural network to obtain geological stability assessment indicators, including at least soil bearing capacity, settlement risk and sliding probability, and to generate a geological stability heat map. The scheduling scheme generation module is used to establish a multi-objective optimization model based on the geological stability heat map. The objectives include maximizing the efficiency of cut and fill operations, minimizing costs, and optimizing quality. The roadbed cut and fill scheduling scheme is represented as a chromosome. The chromosome encoding includes the allocation of construction equipment, operation path planning, and cut and fill volume allocation. The population is initialized using a random generation method. The improved Grey Wolf Optimization Algorithm (GWO) is used to perform a neighborhood search for each individual. Neighborhood operations include equipment reallocation, operation path adjustment, and cut and fill volume fine-tuning. Each operation generates a neighborhood individual. The objective function value of the neighboring individuals is calculated, the optimal individual is selected to replace the original individual, and the roadbed filling and excavation scheduling scheme is obtained through iteration by dynamically adjusting the convergence factor.

5. The intelligent scheduling system for roadbed filling and excavation operations based on artificial intelligence assistance as described in claim 4, characterized in that, The roadbed data acquisition module includes the following sub-modules: A submodule is established to create a calibration model between soil moisture content and spectral reflectance. This is achieved by collecting soil samples with different moisture contents in the construction area to obtain spectral data and actual moisture content. The calculation submodule is used to input the image data collected by the multispectral imager into the inversion model, calculate the soil moisture content corresponding to each pixel, and generate moisture content distribution data of the entire roadbed area. The acquisition submodule is used to calculate the distance from the target object to the sensor by measuring the time difference between laser emission and reception, and to obtain the three-dimensional coordinates of the target object's surface by combining the sensor's scanning angle and rotation speed, thus obtaining laser point cloud data.

6. The intelligent scheduling system for roadbed filling and excavation operations based on artificial intelligence assistance as described in claim 4, characterized in that, The roadbed data acquisition module also Includes the following sub-modules: The conversion submodule is used to convert the moisture content distribution data and laser point cloud data into a unified global coordinate system. The extraction submodule is used to extract moisture features from moisture content distribution data, including at least the mean, variance, and gradient of moisture content, and to extract topographic and geomorphic features from point cloud data, including at least elevation, slope, aspect, and curvature. The fusion submodule is used to expand the dimensions of the moisture features, terrain features and geomorphic features to form multidimensional feature vectors. The feature vectors are then fused into high-dimensional geological feature tensor data using tensor product operations to obtain real-time roadbed status data.

Images