Backfill soil rolling control system and rolling method based on artificial intelligence

Abstract

The invention discloses a backfill soil rolling control system based on artificial intelligence and a rolling method, and relates to the technical field of intelligent construction of water conservancy projects. The system comprises an edge computing node, a 5G communication module, a positioning module, a sensor module and a road roller group, and the edge computing node receives the positioning data and the soil material parameter data in real time, generates a digital geotechnical parameter map, predicts the compactness through an LSTM model, generates a regulation and control instruction, and realizes low-delay closed-loop control. The method comprises the steps of rolling test and model initial building, digital map generation and model optimization, compaction degree prediction and evaluation, error analysis and coupling relation modeling, parameter dynamic adjustment, intelligent rolling and quality closed-loop control and the like. The rolling parameters can be accurately regulated and controlled in real time, the one-time compaction qualification rate is increased to 95% or above, and the construction cost and the construction period risk are remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction technology for water conservancy projects, and in particular to an artificial intelligence-based backfill compaction control system and compaction method, which is applicable to the collaborative operation of unmanned road roller groups in large-scale earthwork projects of water conservancy projects. Background Technology

[0002] Traditional earth-rock dam compaction construction methods rely on manual experience to judge compaction parameters (such as the number of compaction passes, travel speed, and excitation force). However, in actual construction, the characteristics of soil materials (such as clay), moisture content, and loose thickness have significant spatial variability, which can easily lead to local under-compaction (under-compaction) or over-compaction (over-compaction). This results in poor compaction uniformity, low first-pass yield (usually below 85%), and a large number of compaction passes, which in turn leads to project delays and increased costs (rework costs account for 15% to 20% of the total construction cost). It may also cause dam leakage later on.

[0003] To address the aforementioned issues, existing technologies have gradually introduced intelligent control methods. For example, patent application CN119466049A (application number 2024113762651) proposes a non-destructive testing method for compaction degree based on intelligent sensors, using a vibration wheel accelerometer to collect CMV (Compaction Meter Value) values ​​to assess compaction degree; patent application CN115952437A utilizes a one-dimensional convolutional neural network to predict compaction degree. Existing technology (CN120556338A) further discloses an intelligent control method for road rollers based on the real-time status of road compaction quality. This method uses a host computer (in-situ testing vehicle) to detect compaction degree, strength, and modulus in real time, while the slave computer (road roller) optimizes parameters such as vibration frequency, amplitude, and travel speed based on an LSTM model, achieving coordinated control of "compacting, inspecting, and adjusting simultaneously." However, these technologies still suffer from insufficient multi-source data fusion capabilities, lagging dynamic optimization of compaction parameters, and bottlenecks in real-time control. Summary of the Invention

[0004] The purpose of this invention is to provide a technical solution to the real-time control of compaction quality, thereby overcoming the shortcomings of insufficient precision of compaction parameters in existing technologies, reducing the number of compaction passes, improving the first-pass compaction qualification rate, and reducing construction costs and schedule risks.

[0005] To achieve the above objectives, the artificial intelligence-based backfill compaction control system of the present invention includes an edge computing node, a 5G communication module, a positioning module, a sensor module, and a group of road rollers. The edge computing nodes are connected to the 5G communication module, the positioning module, and the sensor module, respectively. The 5G communication module is connected to the grader and the road roller group for communication. The edge computing node receives centimeter-level positioning data from the positioning module and soil parameter data collected by the sensor module in real time. The edge computing node performs spatial gridding processing on the received soil parameter data and positioning data to generate a digital geotechnical parameter map containing the spatial distribution of soil type, moisture content and loose paving thickness. It generates control commands for the roller's travel speed and excitation force through a preset compaction degree prediction model, and sends the control commands to the roller group through the 5G communication module, realizing low-latency closed-loop control of data acquisition, prediction calculation and command issuance. It also includes a parameter storage module, which is configured to store the optimal moisture content of the soil determined by indoor compaction tests and the standard value of the loose paving thickness preset before construction. The edge computing node is connected to the parameter storage module and is used to call the optimal moisture content and the preset loose-lay thickness standard value for deviation judgment.

[0006] The sensor module is positioned 5–20 cm behind the grader scraper. The sensor module's detection surface is 5-20cm away from the soil surface and is level with the soil surface after it has been leveled by the scraper. That is, the horizontal projection of the detection surface completely overlaps with the soil surface and retains a vertical gap of 5-20cm.

[0007] The compaction degree prediction model is an LSTM model (specifically, a temporal prediction network that integrates convolutional feature extraction). Its front end integrates a lightweight convolutional layer, taking historical data sequences of continuous compaction units as input. These historical data sequences include one-hot encoding of soil type, moisture content, loose-lay thickness, number of compaction passes, travel speed, and excitation force. The LSTM model employs a 5-layer structure, specifically including an input layer, 3 hidden layers, and an output layer. The hidden layers consist of one lightweight convolutional layer (using 32 3×3 convolutional kernels), one pooling layer (compressing feature dimensions through 2×2 max pooling), and one fully connected layer (containing 128 neurons). A Dropout regularization mechanism (with a dropout rate set to 0.3) is introduced after the fully connected layer.

[0008] It also includes a parameter control module, which is configured to adjust the moisture content according to the priority level of "abnormal moisture content > abnormal thickness" when the moisture content deviation is greater than the optimal moisture content stored in the parameter storage module ±1% or the loose-lay thickness deviation is greater than the preset loose-lay thickness standard value stored in the parameter storage module ±5%. Specifically, a four-level early warning mechanism is used to quantify and bind the deviation range and control amplitude. When the moisture content deviation is 1% to 2%, the speed is reduced by 10%; when the deviation is 2% to 3%, the excitation force is adjusted by 15%; when the deviation is 3% to 4%, the speed is reduced by 10% and the excitation force is increased by 15%; when the deviation is >4%, the soil replacement process is started. Loose-laid thickness deviation: When the deviation is 5% to 10%, reduce the speed by 10%; when the deviation is 10% to 15%, adjust the excitation force by 15%; when the deviation is 15% to 20%, reduce the speed by 10% and the excitation force by 15%; when the deviation is >20%, start the re-leveling process.

[0009] This invention also discloses a corresponding compaction method, which employs the aforementioned AI-based backfill compaction control system and proceeds according to the following steps: S1: Compaction test and initial LSTM model establishment; The optimal moisture content and maximum dry density of the soil were determined by indoor compaction tests to control the target compaction degree. Field compaction tests were conducted under different soil types, moisture contents and loose thicknesses to collect data on the number of compaction passes, travel speed, excitation force and measured compaction degree. Based on the data, a preliminary LSTM model was trained to establish the mapping relationship between "input parameters and compaction degree". S2: Digital geotechnical parameter map generation and model optimization; After paving is completed, a grader scans the backfill soil layer at a preset speed using a positioning module and a sensor module. The soil parameter data obtained from the scan is then processed into a spatial grid to obtain a 2m×2m grid of digital geotechnical parameters. The map contains spatial distribution data of soil type, moisture content, and loose paving thickness. During compaction, the compaction degree and rolling parameters of each unit node are recorded in real time and uploaded to the edge computing node via a 5G communication module to dynamically optimize the LSTM model weights. S3: Compaction degree prediction and dispersion assessment; The edge computing nodes use LSTM models to predict the number of local compaction passes required to achieve the target compaction degree based on digital geotechnical parameter map data, and take the average number of passes for the entire layer as the control value; then predict the compaction degree of each plane coordinate node again, generate a compaction degree cloud map, calculate the average value and standard deviation of the predicted compaction degree, and evaluate the degree of dispersion of compaction quality. S4: Error analysis and coupling modeling; If the deviation between the measured compaction degree and the predicted value exceeds the threshold, an error analysis is initiated, and a coupled mapping matrix with moisture content and loose-lay thickness as the main control variables is established to clarify the control priority: abnormal moisture content > abnormal loose-lay thickness. S5: Dynamic adjustment of roller group parameters; Based on the prediction results and the coupling mapping matrix, the edge computing nodes dynamically generate control commands for the travel speed and excitation force of the roller group: in areas with low compaction, the parameters are adjusted in conjunction with the direction of moisture content and thickness deviation; in areas with moderate compaction, the current parameters are maintained; in areas with high compaction, the travel speed is increased; among them, the deviation range and the control amplitude are quantitatively bound according to a four-level early warning mechanism, the control strategy prioritizes ensuring the stability of the average number of compaction passes, and the speed and excitation force are adjusted in conjunction according to the coupling relationship.

[0010] S6: Intelligent compaction and quality closed-loop control; Edge computing nodes send control commands to the roller group via 5G communication modules. The roller group constructs according to the control commands and monitors the compaction degree in real time. After reaching the preset number of rolling passes, the overall pass rate is evaluated. If the pass rate is met, the machine stops. If the pass rate is not met, rolling is automatically added, and the LSTM model is updated through the edge computing nodes. After the construction is completed, a compaction degree distribution map and acceptance report are generated and uploaded to the BIM system for archiving.

[0011] In step S2, the scanning speed of the grader is 4 km / h, and the walking distance is 2 meters. The digital geotechnical parameter map is generated by acquiring data from a multispectral infrared sensor, a soil identification infrared sensor, and a layer thickness ultrasonic sensor installed 5-20 cm behind the grader's scraper. The sensor detection surface is horizontally flush with the surface of the soil layer after it has been leveled by the scraper, and this leveled surface is used as the detection reference surface. There is a 5-20 cm vertical gap between the sensor detection surface and the soil surface.

[0012] In step S3, the input to the LSTM model is a historical data sequence of five consecutive compaction units. The hidden layers are, in order, a lightweight convolutional layer (using 32 3×3 convolutional kernels), a pooling layer (compressing the feature dimension through 2×2 max pooling), and a fully connected layer (containing 128 neurons). A dropout mechanism is introduced to prevent overfitting, and the compaction degree prediction error is set to <3% as the convergence target. The model is trained with a time step of 24~48 (including the two extreme values), and the parameters are optimized through early stopping and cross-validation.

[0013] In step S5, when generating the control command, if the moisture content and the loose thickness are both abnormal, the moisture content is controlled first. Specifically, when the moisture content deviation is 1% to 2%, the speed is reduced by 10%, and when the deviation is 2% to 3%, the excitation force is adjusted by 15%.

[0014] I. Hardware System.

[0015] The hardware system mainly includes a grader for inspection, a group of road rollers, a positioning module, and a communication module. The grader for inspection, serving as a mobile inspection platform, is modified as follows: a Beidou RTK high-precision positioning module (centimeter-level positioning) and a 5G communication module are integrated on the roof; a multispectral infrared moisture content sensor (wavelength 3-12μm), a soil material identification infrared sensor (spectral resolution ≤10nm), and a layer thickness ultrasonic sensor are installed 5-20cm behind the scraper (5-20cm from the soil surface, scraper width 2m). The sensor detection surfaces are directly aligned with the soil surface after leveling (maintaining a 5-20cm vertical interval). Through these sensors, the spatial distribution data of the backfill soil layer's moisture content, loose-lay thickness, and soil type are collected in real time, generating a 2m×2m grid of digital geotechnical parameter maps.

[0016] Edge computing node installation location: Independently deployed in a fixed control center or dedicated cabinet at the construction site.

[0017] Roller group: Each roller is equipped with a Beidou RTK high-precision positioning module and a 5G communication module to receive control commands issued by edge computing nodes; a speed controller and a vibration force controller are installed to adjust the travel speed and vibration force; a triaxial accelerometer is installed on the vibrating drum to collect vibration signals in real time and detect the compaction degree by calculating the CMV value.

[0018] Positioning and communication module: It adopts Beidou RTK technology to achieve centimeter-level positioning; and builds a low-latency transmission network (transmission latency <10ms) through a 5G communication module to ensure real-time interaction of sensor data, positioning information and control commands.

[0019] II. Software System.

[0020] The software system's functions are integrated and implemented by edge computing nodes, including data acquisition, prediction, correction, control command generation, and monitoring. Data acquisition function: Real-time acquisition of soil type (one-hot code), moisture content, loose thickness and compaction degree data of each unit node (2m×2m grid), and transmission to edge computing node through 5G communication module; Prediction Function: A multivariate dynamic prediction model is constructed based on the LSTM model. The input is a historical data sequence of 5 consecutive compaction units (including soil type, moisture content, loose thickness, number of compaction passes, travel speed, and excitation force). The hidden layers are, in order, one lightweight convolutional layer (using 32 3×3 convolutional kernels), one pooling layer (compressing feature dimensions through 2×2 max pooling), and one fully connected layer (containing 128 neurons). The output is the predicted compaction degree (error <3%). Correction function: Based on the deviation between the measured compaction degree and the predicted value, dynamically adjust the weights of the LSTM model and optimize the parameter control strategy; Control command generation function: Based on the prediction results and coupling mapping matrix, control commands for the travel speed and excitation force of the road roller group are generated and controlled in a graded manner according to the priority of "abnormal moisture content > abnormal thickness" (e.g., if the moisture content deviation is 1% to 2%, the speed is reduced by 10%), and the commands are sent to the road roller group through the 5G communication module. Monitoring functions: Real-time visualization of compaction quality (compaction degree cloud map, pass rate), equipment status (speed, excitation force) and construction progress, supporting BIM platform integration and digital twin simulation.

[0021] The present invention has the following advantages: 1. The 2D map is sent from the grader and received from the road roller based on the 5G cluster collaborative control protocol. It can locate data such as the type of backfill soil, the thickness of the loose layer, and the moisture content. It can also propose a differentiated compaction scheme for the corresponding coordinate unmanned road roller cluster in a timely and accurate manner, realize the real-time and precise control of compaction parameters, and achieve unmanned operation around the clock.

[0022] 2. The sensor module and the scraper are designed in a coordinated spatial configuration, using the smooth surface leveled by the scraper as the detection reference surface to eliminate measurement errors caused by soil surface roughness. This achieves simultaneous integration of the leveling and detection processes, avoiding the redundancy of manual pre-processing of the soil surface required by traditional independent detection equipment, thus improving detection efficiency and accuracy.

[0023] 3. Specific measures to prioritize the control of moisture content: Ensure that the main control variable (moisture content) is controlled in a priority manner, avoid conflicts in parameter adjustment, and improve the first-pass compaction qualification rate to over 95%.

[0024] The Level 4 early warning mechanism has the following technical advantages: ① Achieve quantitative binding between deviation and control to improve the accuracy of parameter adjustment. By corresponding the deviation range of moisture content / loose layer thickness (e.g., moisture content 1%~2%, 2%~3%) with specific control ranges (speed reduced by 10%, excitation force adjusted by 15%), the blindness of traditional experience-based control is avoided, making parameter adjustment quantifiable and traceable, and ensuring that the compaction deviation is controlled within ±1%.

[0025] ② Clarify the priority of multi-variable conflicts to resolve parameter control coupling issues. Based on soil mechanics principles, set the priority of "abnormal moisture content > abnormal thickness". When both deviate from the threshold simultaneously (e.g., moisture content deviation of 2% and thickness deviation of 8%), prioritize the response of moisture content control to avoid compaction quality fluctuations caused by multi-parameter adjustment conflicts and improve decision-making efficiency.

[0026] ③ Dynamic graded response to adapt to complex construction conditions. The four-level early warning system covers the entire range of scenarios from minor deviations (early warning 1) to severe anomalies (early warning 4). It adopts step-by-step measures for different degrees of deviation (from fine-tuning of speed / vibration force to soil replacement / re-leveling) to achieve dynamic adaptation of "small deviations require small adjustments, and large deviations require strong interventions", reducing the risk of over-compaction or under-compaction.

[0027] ④ Pre-construction quality risk control reduces rework costs. In extreme cases where the deviation is >4% (moisture content) or >20% (thickness), radical measures such as "soil replacement" and "re-leveling" are triggered to prevent the accumulation and expansion of local defects, eliminate potential quality problems during the construction stage, reduce the rework rate in the later stage, and reduce rework costs.

[0028] 4. Supports unmanned collaborative operations and adapts to intelligent construction needs. The control logic of edge computing nodes can be directly embedded to achieve a fully automated closed loop of "deviation detection - instruction generation - cluster execution" without manual intervention, significantly improving the first-pass yield, reducing redundant compaction, and saving costs and time. Attached Figure Description

[0029] Figure 1 This is a flowchart of the closed-loop control based on edge computing and 5G.

[0030] Figure 2 This is a schematic diagram of the LSTM model structure.

[0031] Figure 3 This is a closed-loop diagram of the entire compaction process based on LSTM and edge computing. Detailed Implementation

[0032] like Figure 1 , Figure 2 and Figure 3 As shown, the artificial intelligence-based backfill soil compaction control system of the present invention includes an edge computing node, a 5G communication module, a positioning module, a sensor module, and a group of road rollers. The edge computing nodes are connected to a 5G communication module, a positioning module, and a sensor module, respectively, and the 5G module is connected to the road roller group for communication. The edge computing node receives centimeter-level positioning data from the positioning module and soil parameter data collected by the sensor module in real time. The edge computing node performs spatial gridding processing on the received soil parameter data and positioning data to generate a digital geotechnical parameter map containing the spatial distribution of soil type, moisture content and loose thickness. It generates control commands for the roller travel speed and excitation force through a preset compaction degree prediction model (i.e., LSTM model) and sends the control commands to the roller group through the 5G communication module, realizing low-latency closed-loop control of data acquisition, prediction calculation and command issuance. It also includes a parameter storage module, which is configured to store the optimal moisture content of the soil determined by indoor compaction tests and the standard value of the loose paving thickness preset before construction. The edge computing node is connected to the parameter storage module and is used to call the optimal moisture content and the preset loose-lay thickness standard value for deviation judgment.

[0033] The positioning module is a Beidou RTK positioning system; the sensor module includes a multispectral infrared sensor (wavelength 3-12μm) and a soil identification infrared sensor (spectral resolution ≤10nm).

[0034] Multispectral infrared sensor models are selected based on the requirement of "stability error <2% in ambient temperatures ranging from -20℃ to 60℃". A spectral database of infrared sensors for soil identification is established. The sensor's detection surface must be aligned with the soil surface. The sensor's detection surface is not the sensor itself; the sensor should be 5–20 cm above its detection surface.

[0035] The sensor module is positioned 5–20 cm behind the grader's scraper. The sensor module's detection surface is 5–20 cm from the soil surface and is horizontally aligned with the leveled soil surface, meaning the horizontal projection of the detection surface completely overlaps with the soil surface while maintaining a 5–20 cm vertical gap. The core purpose is to ensure the sensor's detection direction is perpendicular to the leveled soil surface, avoiding measurement errors caused by angular deviations. The vertical distance between the sensor module's detection surface and the soil surface remains 5–20 cm. The scraper is 2 m wide.

[0036] The sensor module and the scraper are designed in a coordinated spatial configuration, using the smooth surface leveled by the scraper as the detection reference plane to eliminate measurement errors caused by soil surface roughness. This achieves simultaneous integration of the leveling and detection processes, avoiding the redundancy of manual pre-processing of the soil surface required by traditional independent detection equipment, thus improving detection efficiency and accuracy.

[0037] The 5-20cm range is a design parameter determined based on experiments on the surface stability of the soil layer after leveling with a scraper. If the distance is too close, it is easily affected by scraper vibration, while if it is too far, the surface will regenerate roughness. The design concept is to balance vibration interference and surface roughness. Without this design concept, the 5-20cm design parameter would not exist.

[0038] The compaction degree prediction model is an LSTM model, which takes a historical data sequence of five consecutive compaction units as input. The historical data sequence includes soil type one-hot encoding, moisture content, loose layer thickness, number of compaction passes, travel speed, and excitation force; loose layer thickness is the thickness of the soil layer before compaction. The hidden layers consist of one lightweight convolutional layer (using 32 3×3 convolutional kernels), one pooling layer (compressing the feature dimension through 2×2 max pooling), and one fully connected layer (containing 128 neurons), with a dropout rate of 0.3.

[0039] The dropout ratio of 0.3 is designed for small sample data (≤500 groups) in engineering projects to prevent overfitting.

[0040] A compaction unit is the smallest data acquisition and processing unit in backfill compaction construction, based on spatial gridding. Its spatial scale is 2m×2m grid, corresponding to a grid cell in a digital geotechnical parameter map. This unit contains soil parameters (soil type, moisture content, loose thickness) and construction parameters (number of compaction passes, travel speed, excitation force) within a specific area, and is the basic input unit for LSTM models to predict compaction degree.

[0041] The key features of a compaction unit include: 1. Spatial scale: 2m×2m grid, matching the width of the grader scraper (2m) to ensure that the sensor detection range covers the entire scraper-leveled area.

[0042] 2. Data composition: Each compaction unit contains parameters such as soil type (one-hot encoding), moisture content, loose thickness, number of compaction passes, travel speed, and excitation force, forming the input sequence of the LSTM model.

[0043] 3. Temporal correlation: The LSTM model takes "historical data sequences of 5 consecutive compaction units" as input. By capturing the spatiotemporal dynamic correlation of the compaction process, prediction errors are reduced, and the shortcomings of traditional static models (such as CNN and linear regression) that cannot reflect the influence of preceding units on subsequent units are addressed.

[0044] The dropout ratio of 0.3 is a design parameter. It was determined through comparative experiments that the model achieved the highest accuracy on the validation set when the ratio was 0.3 within the range of 0.1-0.3. The time step of 5 is also a design parameter. Based on the stress transfer and attenuation characteristics of soil, the influence weight of the first 5 elements on the current element is greater than 90%.

[0045] The specific hyperparameters for training the LSTM model are as follows: initial learning rate of 0.001 (using cosine annealing decay strategy), optimizer Adam (β1=0.9, β2=0.999), 300 training iterations, and batch size of 32.

[0046] The training dataset is derived from 500 sets of field data from the construction of a dam in a water conservancy project (containing 5 types of clay with a moisture content ranging from 12% to 25%).

[0047] Model evaluation metrics: Root mean square error (RMSE) and coefficient of determination (R²) are used, where RMSE < 0.03, R² > 0.95, and prediction accuracy error < 3% on the validation set.

[0048] It also includes a parameter control module, which is configured to adjust the moisture content according to the priority level of "abnormal moisture content > abnormal thickness" when the moisture content deviation is greater than the optimal moisture content stored in the parameter storage module ±1% or the loose-lay thickness deviation is greater than the preset loose-lay thickness standard value stored in the parameter storage module ±5%. Specifically, a four-level early warning mechanism is used to quantify and bind the deviation range and control amplitude. When the moisture content deviation is 1% to 2%, the speed is reduced by 10%; when the deviation is 2% to 3%, the excitation force is adjusted by 15%; when the deviation is 3% to 4%, the speed is reduced by 10% and the excitation force is increased by 15%; when the deviation is >4%, the soil replacement process is started. Loose-laid thickness deviation: When the deviation is 5% to 10%, reduce the speed by 10%; when the deviation is 10% to 15%, adjust the excitation force by 15%; when the deviation is 15% to 20%, reduce the speed by 10% and the excitation force by 15%; when the deviation is >20%, start the re-leveling process.

[0049] This invention is based on soil mechanics principles to design priority. Moisture content is more sensitive to compaction than thickness (a 1% deviation in moisture content leads to a 2% to 3% deviation in compaction, while a 5% deviation in thickness leads to a 1% to 1.5% deviation in compaction).

[0050] The specific configuration of the parameter control module resolves parameter adjustment conflicts caused by multivariate interference, avoids the blindness of traditional experience-based control, and improves the compaction qualification rate by 10% to 15%.

[0051] This invention also provides a corresponding compaction method, which employs the aforementioned AI-based backfill compaction control system and proceeds according to the following steps: S1: Compaction test and initial LSTM model establishment; The optimal moisture content and maximum dry density of the soil were determined by indoor compaction tests to control the target compaction degree. Field compaction tests were conducted under different soil types, moisture contents and loose thicknesses to collect data on the number of compaction passes, travel speed, excitation force and measured compaction degree. Based on the data, a preliminary LSTM model was trained to establish the mapping relationship between "input parameters and compaction degree". S2: Digital geotechnical parameter map generation and model optimization; After paving is completed, a grader scans the backfill soil layer at a preset speed using a positioning module and a sensor module. The soil parameter data obtained from the scan is then processed into a spatial grid to obtain a 2m×2m grid of digital geotechnical parameters. The map contains spatial distribution data of soil type, moisture content, and loose paving thickness. During compaction, the compaction degree and rolling parameters of each unit node are recorded in real time and uploaded to the edge computing node via a 5G communication module to dynamically optimize the LSTM model weights. In step S2, the generation of the digital geotechnical parameter map is performed by the edge computing node. Specifically, the edge computing node receives soil parameter data (soil type, moisture content, loose thickness, etc.) collected by the sensor module and centimeter-level positioning data from the positioning module, performs spatial gridding processing according to a 2m×2m grid, and generates a digital geotechnical parameter map.

[0052] S3: Compaction degree prediction and dispersion assessment; The edge computing nodes use LSTM models to predict the number of local compaction passes required to achieve the target compaction degree based on digital geotechnical parameter map data, and take the average number of passes for the entire layer as the control value; then predict the compaction degree of each plane coordinate node again, generate a compaction degree cloud map, calculate the average value and standard deviation of the predicted compaction degree, and evaluate the degree of dispersion of compaction quality. S4: Error analysis and coupling modeling; If the measured compaction degree deviates from the predicted value by more than a threshold, an error analysis is initiated, and a coupled mapping matrix with moisture content and loose layer thickness as the main control variables is established (this matrix is ​​generated by clustering historical deviation data and reflects the combined influence weight of moisture content and thickness on compaction degree). The control priority is clarified as follows: abnormal moisture content (deviation greater than the optimum moisture content ±1%) > abnormal loose layer thickness (deviation greater than the preset thickness ±5%). S5: Dynamic adjustment of roller group parameters; Based on the prediction results and the coupling mapping matrix, the edge computing nodes dynamically generate control commands for the travel speed and excitation force of the roller group: in areas with low compaction, the parameters are adjusted in conjunction with the direction of moisture content and thickness deviation (e.g., the speed is reduced by 10% when the moisture content deviation is 1% to 2%, and the excitation force is adjusted by 15% when the deviation is 2% to 3%); in areas with moderate compaction, the current parameters are maintained; in areas with high compaction, the travel speed is increased; among these, the deviation range and the control amplitude are quantitatively bound according to a four-level early warning mechanism, the control strategy prioritizes ensuring the stability of the average number of compaction passes, and the speed and excitation force are adjusted in conjunction according to the coupling relationship.

[0053] S6: Intelligent compaction and quality closed-loop control; Edge computing nodes send control commands to the roller group via 5G communication modules. The roller group constructs according to the control commands and monitors the compaction degree in real time. After reaching the preset number of rolling passes, the overall pass rate is evaluated. If the pass rate is met, the machine stops. If the pass rate is not met, rolling is automatically added, and the LSTM model is updated through the edge computing nodes. After the construction is completed, a compaction degree distribution map and acceptance report are generated and uploaded to the BIM system for archiving.

[0054] In step S2, the scanning speed of the grader is 4 km / h, or matched to the sensor sampling frequency, ensuring that one point is collected every 2 meters. The grader travels at 2-meter intervals (the scraper width is 2m), resulting in a 2m×2m grid of digital geotechnical parameters. This digital geotechnical parameters map is generated by a multispectral infrared sensor, a soil identification infrared sensor, and a layer thickness ultrasonic sensor installed 5–20cm behind the grader scraper. The sensor detection surface is horizontally aligned with the soil surface after scraping, and this leveled surface serves as the detection reference surface. There is a 5–20cm vertical gap between the sensor detection surface and the soil surface. "Horizontally aligned" means that the horizontal projection of the detection surface completely overlaps with the soil surface (ensuring the detection area corresponds to the scraper leveling range).

[0055] In step S3, the input to the LSTM model is a historical data sequence of five consecutive compaction units (2m×2m grid). The hidden layer adopts a three-layer structure, consisting of one lightweight convolutional layer (using 32 3×3 convolutional kernels), one pooling layer (compressing feature dimensions through 2×2 max pooling), and one fully connected layer (containing 128 neurons). A Dropout mechanism (ratio 0.3) is introduced to prevent overfitting, and the compaction degree prediction error is set to <3% as the convergence target. The model is trained with a time step of 24~48 (including the two extreme values), and the parameters are optimized through early stopping and cross-validation.

[0056] The training dataset contains 300-500 sets of field test data, and the validation set has a Nash efficiency coefficient (NSE) ≥ 0.95.

[0057] In step S5, when generating the control command, if the moisture content and the loose thickness are both abnormal, the moisture content is controlled first. Specifically, when the moisture content deviation is 1% to 2%, the speed is reduced by 10%, and when the deviation is 2% to 3%, the excitation force is adjusted by 15%.

[0058] Specific measures to prioritize moisture content control ensure that the main control variable (moisture content) is controlled first, avoids parameter adjustment conflicts, and increases the first-pass compaction qualification rate to over 95%.

[0059] The deviation calculation benchmark is the "optimal moisture content" and "preset loose thickness" (e.g., 35cm) determined by the indoor test in step S1.

[0060] The four-level early warning mechanism corresponds to the priority mechanism of "abnormal moisture content > abnormal thickness", quantifies the mapping relationship between the deviation range and the control amplitude, and ensures the feasibility of parameter adjustment.

[0061] Hardware configuration of edge computing nodes: configured with an 8-core CPU, a 512-core GPU and ≥32GB of memory, supporting ≤20ms local data processing latency; for example, using an NVIDIA Jetson AGXXavier processor (8-core ARMv8.2 CPU, 512-core Volta GPU), 32GB of memory, running Ubuntu 20.04 system and TensorRT 8.2 inference engine.

[0062] Table 1. Comparison of Edge Computing and V2X Communication Performance:

[0063] The fundamental difference between edge computing and V2X lies in the fact that edge computing is a "distributed real-time control hub," while V2X is an "information transmission tool." The former can independently complete closed-loop control, while the latter needs to be used as an auxiliary means in conjunction with other computing units. This invention achieves lower latency through edge computing, matching the dynamic real-time requirements of compaction operations.

[0064] • Sensor fault diagnosis: By “cross-validation of adjacent sensor data” (such as triggering an alarm when the moisture content detection deviation between the multispectral infrared sensor and the soil identification sensor is >5%), redundant sensor channels are automatically enabled.

[0065] The design of the three hidden layer structures of the LSTM model in this invention is explained as follows: 1. Number of hidden layers (3 layers).

[0066] The number of hidden layers in an LSTM model directly determines the depth of feature extraction from time-series data. In this invention, the input to the LSTM model is a "historical data sequence of five consecutive compaction units" (containing six parameters such as soil type, moisture content, and loose paving thickness), and its core requirement is to capture the spatiotemporal dynamic correlation of multiple coupled parameters.

[0067] 2-layer LSTM: It can extract low-order nonlinear features (such as the interaction between "moisture content + number of compaction passes"), but it is difficult to model the high-order coupling relationship between "soil type - loose thickness - excitation force" (the patent has 6 input parameters, requiring multi-level feature fusion). 3-layer LSTM: Through the hierarchical structure of "extracting single-parameter temporal features at the bottom layer → fusing multi-parameter correlations in the middle layer (such as the synergistic effect of moisture content and excitation force) → outputting the predicted compaction value at the top layer", it can completely cover the spatiotemporal dynamics of the rolling process (the weight of the first 5 units on the current unit is >90%, and the 3-layer structure is sufficient to model the weight decay characteristics). ≥4 layers of LSTM: In engineering small sample scenarios (500 training data sets), it will lead to redundant model parameters (when each layer has ≥128 neurons, the total number of parameters exceeds 100,000), causing overfitting (the accuracy of the validation set drops by more than 15%, which conflicts with the design goal of "Dropout ratio of 0.3 to suppress overfitting".

[0068] 2. Number of neurons per layer (128).

[0069] The number of neurons needs to match the dimensionality of the input features and the required output accuracy to avoid the "curse of dimensionality" or "underfitting". Input dimensions: Each compaction unit contains 6 parameters (soil one-hot encoding, moisture content, loose paving thickness, number of compaction passes, travel speed, and excitation force). Five consecutive units form a 5×6=30-dimensional input tensor. 128 neurons can provide sufficient nonlinear mapping space (the number of neurons is usually 3-5 times the input dimension, 30×4≈120, and 128 is an engineering rounding value). Output accuracy: This invention requires a compaction degree prediction error of <3%, necessitating a sufficient number of neurons to improve model fitting ability. Theoretically, the number of neurons is negatively correlated with the model fitting error (within a certain range); when the number of neurons is <64, the root mean square error (RMSE) of the validation set is >0.05 (corresponding to a compaction degree error >5%); when the number of neurons is 128, the RMSE is <0.03 (meeting the convergence target of <3% error); further increasing to 256, the RMSE does not decrease significantly (<0.029), but the parameter computation doubles (edge ​​computing node processing latency needs to be <20ms; 128 neurons can balance accuracy and computational efficiency).

[0070] Quantitative design explanation for "speed reduction by 10%" and "excitation force adjustment by 15%" in the four-level early warning mechanism: Core logic: Based on the theories of "sensitivity relationship between moisture content and compaction degree" and "adjustment efficiency of compaction parameters and compaction effect" in soil mechanics, a quantitative mapping between deviation and control range is established.

[0071] 1. Theoretical correlation between moisture content deviation and control range.

[0072] According to the soil mechanics compaction principle, moisture content is the main controlling variable affecting compaction degree (the parameter control module clearly states that "abnormal moisture content > abnormal thickness" has a higher priority), and its deviation on compaction degree conforms to the "nonlinear decay law near the optimum moisture content": Theoretical formula: Compaction degree = K0 × [1 - a(⊿ω)] 2 ], where K0 is the target compaction degree at the optimum moisture content, Δω is the moisture content deviation (%), and α is the sensitivity coefficient (a≈0.02 for clay soils and a≈0.015 for sandy soils).

[0073] When the deviation is 1%~2%, reduce the speed by 10%. When Δω = 1%, the degree of compaction K ≈ K0 × [1 - 0.02 × (1)] 2 =0.98K0 (2% loss). According to the compaction theory, the degree of compaction is inversely proportional to the travel speed (K∝1 / v). Reducing the speed by 10% (v'=0.9v) can increase the degree of compaction to K'≈K0×0.98 / 0.9≈1.09K0. After compensating for the 2% loss, the target degree of compaction is still met (error <1%).

[0074] When the deviation is 2%~3%, adjust the excitation force by 15%. When Δω = 2%, the degree of compaction K ≈ K0 × [1 - 0.02 × (2)] 2 = 0.92K0 (8% loss). At this point, adjusting the speed alone cannot fully compensate (a 10% reduction in speed only increases the compaction by 11%, 0.92 × 1.11 ≈ 1.02K0, risk of overshoot). Adjustment through the excitation force is necessary (compaction degree is proportional to the square of the excitation force, K∝F). 2 The excitation force is increased by 15% (F'=1.15F), and the compaction degree is increased to K'≈0.92×(1.15). 2 ≈0.92×1.32=1.21K0, combined with speed fine-tuning (such as synchronously reducing by 5%), can be precisely controlled to the target compaction degree.

[0075] 2. Theoretical relationship between the deviation of the loose layer thickness and the adjustment range.

[0076] Deviation in the loose-lay thickness affects the compaction effect by altering the compaction effort of the soil layer; its sensitivity is lower than that of moisture content ("abnormal moisture content > abnormal thickness" priority). The design concept is as follows: Theoretical formula: Compaction degree K=K0×[1-β(⊿h / h0)], where ⊿h is the thickness deviation, h0 is the preset thickness, and β is the thickness sensitivity coefficient (β≈0.01, which is only 1 / 2 of the moisture content sensitivity coefficient).

[0077] When the deviation is 5%~10%, reduce the speed by 10%: When the thickness deviation is 5%, the compaction loss is about 0.05×0.01×100%=0.5%, which can be compensated by reducing the speed by 10% (same logic as the moisture content deviation of 1%); when the deviation is 10%~15%, the compaction loss is 1%~1.5%, and the excitation force needs to be adjusted by 15% (refer to the excitation force adjustment efficiency of moisture content deviation of 2%~3%).

[0078] Experimental subjects. A typical backfill soil compaction engineering scenario was selected (such as the construction of a water conservancy project dam, with an area of ​​1000㎡, the soil type being clay, and a target compaction degree of 93%).

[0079] Control group: Using existing technology (traditional manual control, relying on operator experience to adjust compaction parameters).

[0080] Experimental group: The invention "Artificial Intelligence-Based Backfill Soil Compaction Control System" (system + method) was adopted.

[0081] Control variables: The soil type, optimum moisture content (12.5%), loose thickness (35cm), and target compaction degree (93%) were completely consistent in both groups of experiments, with only the control method being changed (manual / existing technology vs. this invention).

[0082] Sample size: Each experiment was repeated 3 times (parallel experiments) to avoid single-experiment error.

[0083] Experimental methods.

[0084] Control group (traditional manual control): Based on experience, the operator determines the number of compaction passes (preset to 8 passes), travel speed (2-4 km / h), and excitation force (300-500 kN). After compaction, 30 test points are randomly selected using the ring cutter method to determine the compaction degree and calculate the first-pass yield (the percentage of test points with a compaction degree ≥90%).

[0085] Experimental group (this invention): Perform steps S1-S6 according to claim 5: S1: The optimum moisture content was determined to be 12.5% ​​and the maximum dry density was 1.8 g / cm³ through indoor compaction tests; S2: Generate a digital geotechnical parameter map with a 2m×2m grid; S3-S5: LSTM model predicts compaction degree, and a four-level early warning mechanism dynamically adjusts parameters (e.g., when the moisture content deviation is 1%, the adjustment rate is reduced by 10%). S6: After compaction, 30 test points are randomly selected, and the pass rate is calculated once, as shown in Table 1: Table 1: Comparison of Pass Rates between Experimental and Control Groups

[0086] The experimental group had an average first-time compaction pass rate of 95.5%, which was significantly higher than the control group's 84.3%.

[0087] The above embodiments are only used to illustrate and not limit the technical solutions of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention without departing from the spirit and scope of the present invention. Any modifications or partial substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An artificial intelligence-based backfill soil compaction control system, characterized in that: This includes edge computing nodes, 5G communication modules, positioning modules, sensor modules, and a group of road rollers; The edge computing nodes are connected to the 5G communication module, the positioning module, and the sensor module, respectively. The 5G communication module is connected to the grader and the road roller group for communication. The edge computing node receives centimeter-level positioning data from the positioning module and soil parameter data collected by the sensor module in real time. The edge computing node performs spatial gridding processing on the received soil parameter data and positioning data to generate a digital geotechnical parameter map containing the spatial distribution of soil type, moisture content and loose thickness. The system generates control commands for the roller's travel speed and excitation force by using a preset compaction degree prediction model, and sends these commands to the roller group via a 5G communication module, thus achieving low-latency closed-loop control of data acquisition, prediction calculation, and command issuance. It also includes a parameter storage module, which is configured to store the optimal moisture content of the soil determined by indoor compaction tests and the standard value of the loose paving thickness preset before construction. The edge computing node is connected to the parameter storage module and is used to call the optimal moisture content and the preset loose-lay thickness standard value for deviation judgment.

2. The artificial intelligence-based backfill compaction control system according to claim 1, characterized in that: The sensor module is positioned 5–20 cm behind the grader scraper. The sensor module's detection surface is 5-20cm away from the soil surface and is level with the soil surface after it has been leveled by the scraper. That is, the horizontal projection of the detection surface completely overlaps with the soil surface and retains a vertical gap of 5-20cm.

3. The artificial intelligence-based backfill compaction control system according to claim 1, characterized in that: The compaction degree prediction model is an LSTM model with a lightweight convolutional layer integrated at the front end. It takes historical data sequences of continuous compaction units as input, including one-hot encoding of soil type, moisture content, loose thickness, number of compaction passes, travel speed, and excitation force. The LSTM model adopts a 5-layer structure, specifically including an input layer, 3 hidden layers, and an output layer. The hidden layers consist of one lightweight convolutional layer, one pooling layer, and one fully connected layer. A Dropout regularization mechanism is introduced after the fully connected layer to mitigate the risk of overfitting on medium-sized datasets. The lightweight convolutional layer uses 32 3×3 convolutional kernels to control the parameter size while ensuring basic feature extraction capabilities; the pooling layer compresses the feature dimension through 2×2 max pooling to reduce redundant information; the fully connected layer contains 128 neurons, which is suitable for feature fitting requirements of medium-sized data; the dropout regularization mechanism has a dropout rate of 0.

3.

4. The artificial intelligence-based backfill compaction control system according to claim 1, characterized in that: It also includes a parameter control module, which is configured to adjust the moisture content according to the priority level of "abnormal moisture content > abnormal thickness" when the moisture content deviation is greater than the optimal moisture content stored in the parameter storage module ±1% or the loose-lay thickness deviation is greater than the preset loose-lay thickness standard value stored in the parameter storage module ±5%. Specifically, a four-level early warning mechanism is used to quantify and bind the deviation range and control amplitude. When the moisture content deviation is 1% to 2%, the speed is reduced by 10%; when the deviation is 2% to 3%, the excitation force is adjusted by 15%; when the deviation is 3% to 4%, the speed is reduced by 10% and the excitation force is increased by 15%; when the deviation is >4%, the soil replacement process is started. Loose-laid thickness deviation: When the deviation is 5% to 10%, reduce the speed by 10%; when the deviation is 10% to 15%, adjust the excitation force by 15%; when the deviation is 15% to 20%, reduce the speed by 10% and the excitation force by 15%; when the deviation is >20%, start the re-leveling process.

5. The compaction method, using the AI-based backfill compaction control system of claim 4, is carried out according to the following steps: S1: Compaction test and initial LSTM model establishment; The optimal moisture content and maximum dry density of the soil were determined by indoor compaction tests to control the target compaction degree. Field compaction tests were conducted under different soil types, moisture contents and loose thicknesses to collect data on the number of compaction passes, travel speed, excitation force and measured compaction degree. Based on the data, a preliminary LSTM model was trained to establish the mapping relationship between "input parameters and compaction degree". S2: Digital geotechnical parameter map generation and model optimization; After paving is completed, a grader scans the backfill soil layer at a preset speed using a positioning module and a sensor module. The soil parameter data obtained from the scan is then processed into a spatial grid to obtain a 2m×2m grid of digital geotechnical parameters. The map contains spatial distribution data of soil type, moisture content, and loose paving thickness. During compaction, the compaction degree and rolling parameters of each unit node are recorded in real time and uploaded to the edge computing node via a 5G communication module to dynamically optimize the LSTM model weights. S3: Compaction degree prediction and dispersion assessment; The edge computing nodes use LSTM models to predict the number of local compaction passes required to achieve the target compaction degree based on digital geotechnical parameter map data, and take the average number of passes for the entire layer as the control value; then predict the compaction degree of each plane coordinate node again, generate a compaction degree cloud map, calculate the average value and standard deviation of the predicted compaction degree, and evaluate the degree of dispersion of compaction quality. S4: Error analysis and coupling modeling; If the deviation between the measured compaction degree and the predicted value exceeds the threshold, an error analysis is initiated, and a coupled mapping matrix with moisture content and loose-lay thickness as the main control variables is established to clarify the control priority: abnormal moisture content > abnormal loose-lay thickness. S5: Dynamic adjustment of parameters for road roller groups; Based on the prediction results and the coupling mapping matrix, the edge computing nodes dynamically generate control commands for the travel speed and excitation force of the roller group: in areas with low compaction, the parameters are adjusted in conjunction with the direction of moisture content and thickness deviation; in areas with moderate compaction, the current parameters are maintained; in areas with high compaction, the travel speed is increased; among them, the deviation range and the control amplitude are quantitatively bound according to a four-level early warning mechanism, the control strategy prioritizes ensuring the stability of the average number of rolling passes, and the speed and excitation force are adjusted in conjunction according to the coupling relationship; S6: Intelligent compaction and quality closed-loop control; Edge computing nodes send control commands to the roller group via 5G communication modules. The roller group constructs according to the control commands and monitors the compaction degree in real time. After reaching the preset number of rolling passes, the overall pass rate is evaluated. If the pass rate is met, the machine stops. If the pass rate is not met, rolling is automatically added, and the LSTM model is updated through the edge computing nodes. After the construction is completed, a compaction degree distribution map and acceptance report are generated and uploaded to the BIM system for archiving.

6. The compaction method according to claim 5, characterized in that: In step S2, the scanning speed of the grader is 4 km / h, and the walking distance is 2 meters. The digital geotechnical parameter map is generated by acquiring data from a multispectral infrared sensor, a soil identification infrared sensor, and a layer thickness ultrasonic sensor installed 5-20 cm behind the grader's scraper. The sensor detection surface is horizontally flush with the surface of the soil layer after it has been leveled by the scraper, and this leveled surface is used as the detection reference surface. There is a 5-20 cm vertical gap between the sensor detection surface and the soil surface.

7. The compaction method according to claim 5, characterized in that: In step S3, the input of the LSTM model is a historical data sequence of 5 consecutive compaction units. The hidden layers are a lightweight convolutional layer, a pooling layer and a fully connected layer in sequence. The Dropout mechanism is introduced to prevent overfitting, and the compaction degree prediction error is <3% as the convergence target. The lightweight convolutional layer uses 32 3×3 convolutional kernels; the pooling layer compresses the feature dimension using 2×2 max pooling. The fully connected layer contains 128 neurons; The model was trained using a time step of 24 to 48, and the parameters were optimized using early stopping and cross-validation.

8. The compaction method according to claim 5, characterized in that: In step S5, when generating the control command, if the moisture content and the loose thickness are both abnormal, the moisture content is controlled first. Specifically, when the moisture content deviation is 1% to 2%, the speed is reduced by 10%, and when the deviation is 2% to 3%, the excitation force is adjusted by 15%.

9. The backfill soil compaction control method based on artificial intelligence according to claim 6, characterized in that, In step S6, the edge computing node has a processing delay of less than 20ms and the 5G communication module has a transmission delay of less than 10ms, thus achieving real-time closed-loop control.

Images