Roadbed pavement test and data acquisition system under action of moving load

Abstract

The invention discloses a roadbed pavement test and data acquisition system under the action of moving load, which comprises the following steps of: acquiring high-precision pavement unevenness data, decomposing features through a space-time convolution variational self-encoding network and inhibiting noise; based on the normal stiffness and the morphological gradient, the density of the sensor is adjusted by adopting dynamic grid division, and multi-modal data is fused by utilizing Bayesian weight. And adaptively extracting signal details by applying multi-scale transformation, and designing a filter to enhance the damage signal and optimize the structure entropy. And constructing a multi-field coupling grid control equation to realize dynamic topological optimization reconstruction. And establishing a damage evolution state space, and training the prediction model through a multi-stage reinforcement learning framework in combination with experience playback. And constructing a digital twin mapping model, integrating Bayesian optimization and reinforcement learning, and establishing a virtual-real feedback channel to realize dynamic decision optimization. According to the invention, innovative improvement on a traditional test method and a data acquisition system is provided, and meanwhile, complex nonlinear dynamic behaviors can be accurately simulated and analyzed.

Description

Technical Field

[0001] The invention relates to the technical field of roadbed and pavement testing, and in particular to a roadbed and pavement testing and data acquisition system under the action of a moving load. Background Art

[0002] With the rapid development of the modern transportation industry, the vehicle loads borne by infrastructure such as roads and railways are becoming increasingly complex and changeable. Especially under high-speed and heavy-load conditions, the roadbed and pavement structures are under complex dynamic loads for a long time, facing severe safety and durability challenges. When different types of vehicles act on the road surface with different speeds, tire contact characteristics and load distribution patterns, they will cause stress redistribution and deformation accumulation of the roadbed and pavement materials, thereby affecting their long-term service performance. Especially in infrastructure such as highways, airport runways and heavy-load railways, the dynamic response caused by the action of moving loads has a crucial impact on the service life of the pavement and driving safety.

[0003] Traditional roadbed and pavement test methods mainly rely on static load tests or simplified dynamic loading tests, which are difficult to truly reproduce the complex working conditions in the actual traffic environment and cannot fully evaluate the dynamic characteristics of the roadbed and pavement under the action of moving loads. In addition, the current data acquisition system still has limitations in spatial resolution, time synchronization, and multimodal fusion capabilities, resulting in insufficient monitoring accuracy of pavement deformation, stress response, and damage evolution. Therefore, it is urgent to develop a test and data acquisition system that can simulate real traffic loads, perceive the state of the roadbed and pavement in real time, and accurately collect multidimensional data to fully reveal the dynamic behavior of the roadbed and pavement under the action of moving loads.

[0004] It is worth noting that the subgrade and pavement may exhibit highly nonlinear dynamic characteristics under the action of moving loads, such as local buckling, overall instability, fatigue cumulative damage, etc. These nonlinear dynamic behaviors are affected by multiple factors such as material properties, structural parameters, environmental factors and load types. Their evolution mechanism is complex and has significant temporal and spatial heterogeneity. At present, related research still lacks systematic theories and methods, and the quantitative analysis and prediction capabilities of nonlinear dynamic behaviors are relatively limited. Therefore, how to accurately analyze and predict the nonlinear dynamic behavior of subgrade and pavement under moving loads to improve the safety and durability of the structure is an important technical problem that needs to be solved urgently. Summary of the invention

[0005] In order to solve the above problems, the present invention provides a roadbed and pavement test and data acquisition system under moving load.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] Subgrade and pavement test and data acquisition system under moving loads, comprising:

[0008] Moving load simulation module: Collect high-precision pavement unevenness data, perform feature decomposition and noise suppression using a spatio-temporal convolutional variational autoencoder network, and optimize the load action path based on adaptive deformation gradient analysis;

[0009] Multi-modal sensing data fusion module: According to the normal stiffness distribution and pavement morphology gradient, adaptively adjust the sensor layout density using adaptive grid division, fuse multi-modal sensor data using Bayesian dynamic weight allocation, and perform noise suppression through neighborhood entropy constraint;

[0010] Nonlinear signal feature enhancement module: Extract signal detail components using multi-scale transformation combined with adaptive scale selection, enhance damage signals and suppress noise through an adaptive filter, and then optimize and enhance the signals based on structural information entropy constraint to improve the identifiability of damage signals;

[0011] Dynamics solving module: Construct an adaptive grid control equation driven by multi-field coupling, adjust the grid density according to the coupling effects of the stress field, velocity field, acceleration field, and damage field, and reconstruct the grid using dynamic topology optimization method;

[0012] Damage prediction module: Establish a state space based on damage evolution, construct a reinforcement learning framework with a multi-stage learning strategy, train the model through experience replay and long-term discount factor, and realize the adaptive optimization of damage prediction strategy;

[0013] Digital twin decision-making module: Construct a digital twin virtual-real mapping model, establish a dynamic feedback channel between the physical system and the virtual simulation system, dynamically adjust and optimize parameters through an intelligent optimization feedback mechanism combined with reinforcement learning and Bayesian optimization, construct an optimization control model based on twin decision-making, and realize dynamic optimization decision-making.

[0014] Furthermore: The moving load simulation module includes:

[0015] Using multi-sensor fusion technology, through the collaborative work of laser point cloud scanning, inertial measurement unit, and ultra-wideband radar, collect pavement height point cloud data, the attitude changes of the vehicle load simulation device, and the non-uniform morphology interference information caused by the change of subgrade materials, and perform fusion processing on the data to establish a continuous, seamless, and high-precision three-dimensional pavement morphology model;

[0016] Perform multi-layer feature decomposition on the collected data using a spatio-temporal convolutional variational autoencoder network, improve data reliability through an adaptive noise suppression mechanism, and thus provide high-precision pavement morphology data input for subsequent adaptive feedback adjustment.

[0017] Furthermore: The multi-modal sensing data fusion module includes:

[0018] Adopt adaptive mesh generation, and dynamically adjust the spatial distribution density of sensors according to the normal stiffness gradient and road surface morphology gradient, optimize the sensor layout, make the sensors in the high stress area more dense, and reduce the sensor density in the low stress area;

[0019] Based on Bayesian dynamic weight assignment, adaptively adjust the data fusion weight according to the signal-to-noise ratio and confidence of different sensors to ensure the accuracy of the fused data;

[0020] Use the neighborhood entropy constraint method to suppress noise from both the time and space domains to improve data quality.

[0021] Furthermore: The non-linear signal feature enhancement module includes:

[0022] Adopt multi-scale transformation with adaptive scale selection, decompose the signal through wavelet transform, calculate the energy distribution of wavelet coefficients, and select the optimal scale to retain micro damage features;

[0023] Signal enhancement based on an adaptive filter, dynamically adjust the filtering parameters through local signal-to-noise ratio calculation and filter gain function, enhance the damage signal and suppress irrelevant noise;

[0024] Use signal optimization based on structural information entropy constraint, calculate the local information entropy of the enhanced signal, and further optimize the signal through the information entropy constraint weight function.

[0025] Furthermore: The dynamics solving module includes:

[0026] Construct an adaptive grid control equation driven by multi-field coupling, comprehensively consider the stress field gradient, velocity field mutation, acceleration field change rate and damage field characteristics to form the driving force for grid adjustment;

[0027] Adjust the grid based on the adaptive density function, dynamically determine the local grid size according to the multi-field coupling control equation, realize grid encryption in the stress concentration and significant damage areas and grid sparsification in the low gradient areas;

[0028] Introduce dynamic topology optimization, adjust the element shape through the grid quality evaluation function, optimize the grid topology structure, and ensure that the overall grid quality is not deteriorated while locally adjusting the grid.

[0029] Furthermore: The damage prediction module includes:

[0030] Construct a state space based on damage evolution, define state variables by combining stress, velocity, acceleration and damage characteristics, construct a damage evolution state vector, and define a damage accumulation objective function to optimize the long-term prediction ability;

[0031] Construct a reinforcement learning framework based on a multi-stage learning strategy, divide the damage evolution stages, including the initial micro-damage stage, the plastic damage development stage, and the fatigue failure stage, design corresponding learning strategies respectively, and dynamically adjust the learning strategies through a multi-stage strategy switching function;

[0032] Adopt experience replay and long-term discount factors for reinforcement learning training, construct an experience pool, update the strategy using the long-term discount factor, and optimize the damage prediction model parameters based on the policy gradient.

[0033] Furthermore: The digital twin decision-making module includes:

[0034] Construct a digital twin virtual-real mapping model, through a high-dimensional state mapping function, dynamically synchronize the stress, displacement, damage, and environmental states of the physical system with the virtual simulation system, and optimize the twin mapping accuracy based on an adaptive filter to ensure the high-dimensional synchronization of the physical system and the virtual system;

[0035] Construct an optimization feedback mechanism, define an optimization objective function, combine the reinforcement learning optimization strategy, dynamically adjust the optimization parameters, and achieve the adaptive adjustment of the parameters through the Bayesian optimization method to ensure the self-adaptability of decision optimization;

[0036] Construct an optimization control model based on twin decision-making, define a twin decision control strategy, and use model predictive control to optimize the control strategy to achieve dynamic optimization control.

[0037] Compared with the prior art, the technical progress achieved by the present invention lies in:

[0038] The present invention adopts an innovative intelligent moving load simulation algorithm based on an adaptive feedback mechanism (AFMLS), which can real-time simulate the spatial variability of vehicle loads and the interaction between the subgrade and tires, and solves the problem of fixed load paths and load distributions in traditional test methods. This advantage enables the system to more accurately reproduce the complex dynamic load conditions in the real traffic environment and provides a more realistic load input for analyzing the nonlinear dynamic behavior of subgrade and pavement.

[0039] The system is based on a multi-modal sensing data fusion algorithm optimized by hierarchical adaptive density, which automatically adjusts the sensor layout density in different regions to cope with the dynamic changes of complex stress fields. Through this technology, the system can efficiently collect sensor data from different parts and different types, improving the quality and accuracy of the data. This is crucial for understanding and predicting the dynamic behavior of subgrade and pavement, especially under complex loads, and can reflect the real-time multi-dimensional states such as stress, strain, displacement, and damage in real time.

[0040] The system adopts a non - linear signal feature enhancement module based on multi - scale transformation, which can extract low - frequency trend information and high - frequency local features under complex loads, enhancing the identifiability of micro - damage signals. This solves the problem that existing methods are prone to losing micro - damage information when dealing with the non - linear behavior of subgrade and pavement, providing more accurate signal processing capabilities for the analysis of non - linear dynamic behaviors.

[0041] By introducing a dynamic solution module based on multi - field coupling adaptive grid adjustment, the present invention can automatically adjust the calculation grid in the stress - concentration area, optimize the calculation efficiency and reduce errors. This is particularly important for complex non - linear dynamic problems, especially in the analysis of phenomena such as instability and buckling of subgrade and pavement, which can effectively reduce calculation errors caused by uneven grid division and improve prediction accuracy.

[0042] The system integrates a damage prediction module based on multi - stage reinforcement learning, which can accurately predict the long - term damage evolution trend of subgrade and pavement under moving loads through the dynamic optimization mechanism of reinforcement learning. This advantage is especially applicable to long - term operating infrastructure. In the case of continuously changing traffic loads, it can provide a scientific basis for pavement maintenance and repair, reducing the risk of sudden damage and destruction.

[0043] By combining a digital twin decision - making module with intelligent optimization feedback, the present invention can adjust and optimize decisions based on real - time data and feedback under different environmental conditions. This innovative feature solves the problem that traditional optimization methods are difficult to adapt to complex engineering environments in real time, providing flexible and efficient decision - making support. Especially in the case of complex traffic flow and load changes, it can optimize the maintenance strategy of subgrade and pavement, improving structural safety and durability.

[0044] In summary, the present invention not only provides an innovative improvement to traditional test methods and data acquisition systems, but also can accurately simulate and analyze complex non - linear dynamic behaviors. Through high - precision load simulation, intelligent data fusion, non - linear feature extraction, efficient solution and long - term damage prediction, the present invention provides strong technical support for improving the safety, reliability and durability of subgrade and pavement structures, solving common problems such as insufficient accuracy, response lag and prediction errors in traditional methods. Brief Description of the Drawings

[0045] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention.

[0046] In the drawings:

[0047] Figure 1 is the system structure diagram of the present invention. Detailed Embodiments

[0048] The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments. The embodiments of the present invention will be described below with reference to the accompanying drawings.

[0049] As Figure 1 shown, the present invention discloses a subgrade and pavement test and data acquisition system under moving loads, including:

[0050] Moving load simulation module: Collect high-precision pavement unevenness data, perform feature decomposition and noise suppression using a spatio-temporal convolutional variational autoencoder network, and then optimize the load action path based on adaptive deformation gradient analysis;

[0051] Multi-modal sensing data fusion module: According to the normal stiffness distribution and pavement morphology gradient, adaptively adjust the sensor layout density using adaptive grid division, fuse multi-modal sensor data using Bayesian dynamic weight allocation, and suppress noise through neighborhood entropy constraint;

[0052] Nonlinear signal feature enhancement module: Extract signal detail components using multi-scale transform combined with adaptive scale selection, enhance damage signals and suppress noise through an adaptive filter, and then optimize and enhance the signals based on structural information entropy constraint to improve the identifiability of damage signals;

[0053] Dynamics solving module: Construct an adaptive grid control equation driven by multi-field coupling, adjust the grid density according to the coupling effects of the stress field, velocity field, acceleration field, and damage field, and reconstruct the grid using a dynamic topology optimization method;

[0054] Damage prediction module: Establish a state space based on damage evolution, construct a reinforcement learning framework with a multi-stage learning strategy, train the model through experience replay and long-term discount factors, and realize the adaptive optimization of the damage prediction strategy;

[0055] Digital twin decision module: Construct a digital twin virtual-real mapping model, establish a dynamic feedback channel between the physical system and the virtual simulation system, dynamically adjust and optimize parameters through an intelligent optimization feedback mechanism combined with reinforcement learning and Bayesian optimization, construct an optimization control model based on twin decision-making, and realize dynamic optimization decision-making.

[0056] Specifically, the moving load simulation module based on the Environment Sensing Neuron (ESN) includes:

[0057] In the moving load simulation module, the Environment Sensing Neuron (ESN) is responsible for real-time collecting the unevenness information on the pavement surface and eliminating random noise through a dynamic filtering mechanism, so as to provide high-precision pavement morphology data input for subsequent adaptive feedback adjustment.

[0058] Module 1.1: High-precision road surface unevenness acquisition based on multi-sensor fusion

[0059] Since traditional single laser or inertial sensors are difficult to meet the requirements of high-frequency micro deformation perception, the present invention uses laser point cloud scanning + inertial measurement unit (IMU) + ultra-wideband radar (UWB-Radar) for data fusion to ensure the optimal combination of spatial accuracy, time synchronization, and noise suppression effect. The core objective of the acquisition is to establish a continuous, seamless, and high-precision three-dimensional road surface morphology model as the input basic data for intelligent load simulation.

[0060] The specific implementation process includes:

[0061] 1. Set N laser scanning sensors (LiDAR) evenly distributed at the bottom of the mobile load simulation system to collect road surface height point cloud data P lidar (x, y, z, t):

[0062] P lidar (x, y, z, t) = f lidar (λ, θ, d)

[0063] where λ is the laser wavelength, θ is the scanning angle, and d is the echo ranging data.

[0064] 2. Use the inertial measurement unit (IMU) to calculate the attitude change of the vehicle-mounted load simulation device, obtain the acceleration a g in the direction of gravity and the angular velocity ω g , and combine the gyroscope data G gyro (t) for tilt compensation:

[0065]

[0066] 3. Combine the ultra-wideband radar (UWB-Radar) for deep structure detection to obtain the non-uniform morphology interference information S uwb (x, y, t) caused by the change of subgrade materials, and correct the laser point cloud data:

[0067] P corr (x, y, t) = P lidar (x, y, t) + S uwb (x, y, t) · w uwb

[0068] where w uwb is the weight coefficient of the radar signal.

[0069] Module 1.2: Multi-layer feature decomposition and noise suppression based on spatio-temporal convolutional variational autoencoder network (SCVAE)

[0070] Since the road surface unevenness data contains short-term scale high-frequency noise (vehicle vibration, environmental interference) and long-term scale terrain trend information (road construction errors, local damage), a single low-pass filter cannot effectively separate the true shape from the noise signal.

[0071] The present invention uses a spatio-temporal convolutional variational autoencoder network (SCVAE, Spatio-Temporal Convolutional Variational Autoencoder) to perform multi-layer feature decomposition on the collected data and improve data reliability through an adaptive noise suppression mechanism.

[0072] The specific implementation process includes:

[0073] 1. Set the input spatio-temporal data matrix:

[0074] X input (x, y, t) = {P corr (x, y, t), R corr (t)}

[0075] Among them, P corr (x, y, t) is the fused road surface point cloud data, and R corr (t) is the gyroscope data after attitude correction.

[0076] 2. Use a spatio-temporal convolutional network (STCN) to extract local road surface shape features:

[0077] H conv = σ(W stcn *X input + b)

[0078] Among them, W stcn is the spatio-temporal convolutional kernel, and σ is the non-linear activation function.

[0079] 3. Perform high-dimensional feature reconstruction through a variational autoencoder (VAE) to generate a low-noise road surface shape estimate:

[0080] Z vae ~q φ (Z|H conv )

[0081] Among them, Z vae is the latent space representation, and q φ is the encoder network of the VAE.

[0082] 4. Optimize using the KL divergence loss:

[0083] L kl = D KL (q φ (Z)||p(Z))

[0084] Among them, p(Z) is the standard normal distribution.

[0085] 5. Calculate the final three-dimensional road surface morphology after denoising:

[0086]

[0087] Module 1.3: Load Path Optimization Based on Adaptive Gradient Deformation Analysis (AGDA)

[0088] During the vehicle driving process, the normal stiffness of the tire-road contact area changes significantly. The present invention adopts Adaptive Gradient Deformation Analysis (AGDA) to dynamically adjust the load application path to ensure that the contact state between the tire and the road surface is closest to the actual working condition.

[0089] The specific implementation process includes:

[0090] 1. Calculate the normal stiffness gradient of the road surface morphology:

[0091]

[0092] 2. Calculate the stress concentration coefficient of the load application area:

[0093] S stress (x, y) = ∫ A G stiff (x′, y′)dx′dy′

[0094] 3. Adopt the gradient descent optimization method to dynamically adjust the load application path to make the tire force uniform:

[0095]

[0096] Among them, α is the optimization step size.

[0097] Final output:

[0098] P filtered (x, y, t): The three-dimensional road surface morphology data after filtering and optimization.

[0099] P opt (x, y): The optimization result of the load application path based on deformation gradient analysis.

[0100] G stiff (x, y): The calculated normal stiffness distribution for subsequent load adjustment.

[0101] This module uses key technologies such as multi-sensor data fusion, high-dimensional feature decomposition, and deformation gradient optimization to obtain high-precision road surface morphology data in real time, optimize the load action path, and provide reliable input data for subsequent adaptive feedback adjustment.

[0102] Specifically, the multi-modal sensing data fusion module based on hierarchical adaptive density optimization includes:

[0103] After the mobile load simulation module completes the real-time acquisition and dynamic filtering of road surface unevenness data based on the Echo State Network (ESN), this embodiment has obtained the high-precision three-dimensional road surface morphology data P filtered (x, y, t) and the corresponding normal stiffness distribution G stiff (x, y), which provide the necessary input information for the multi-modal sensing data fusion module.

[0104] In this module, this embodiment designs a hierarchical adaptive density optimization strategy to dynamically adjust the deployment density of multi-modal sensors according to the stress distribution and morphological change characteristics of different regions, thereby improving the accuracy and robustness of data fusion.

[0105] Module 2.1: Adaptive Grid Partitioning-based Density Optimization (AGP-DO)

[0106] Due to the highly non-uniform stress distribution of the subgrade-pavement system, if a fixed-density sensor deployment method is adopted, it will lead to a waste of sensor resources in low-stress areas and insufficient sensor data in high-stress concentration areas.

[0107] The present invention uses Adaptive Grid Partitioning (AGP) to dynamically adjust the spatial distribution density of sensors according to the normal stiffness gradient G stiff (x, y) and the road surface morphology gradient so that high-stress areas have a denser sensor deployment, while the sensor density in low-stress areas is reduced, thereby improving the measurement accuracy and resource utilization efficiency.

[0108] The specific implementation process includes:

[0109] 1. Calculate the normal stiffness change rate to judge the stress concentration area:

[0110]

[0111] 2. Calculate the road surface morphology gradient modulus to judge the geometric feature mutation area:

[0112]

[0113] 3. Set the sensor density optimization function, where D s (x, y) represents the optimal sensor deployment density at a given coordinate position:

[0114]

[0115] where D min and D max represent the lowest and highest sensor deployment densities respectively.

[0116] 4. Adopt a dynamic grid refinement strategy:

[0117] If D s (x, y) > D threshold , then refine the grid and increase the sensor density in this area.

[0118] If D s (x, y) < D threshold , then merge the grid and reduce the sensor density in this area.

[0119] Module 2.2: Multi-modal Data Fusion Based on Bayesian Dynamic Weight Assignment (BDW-Fusion)

[0120] Since the measurement accuracies and noise characteristics of different sensors (such as strain gauges, accelerometers, fiber Bragg gratings, ultrasonic rangefinders, etc.) are different, how to dynamically adjust the weights of different data sources according to the actual working conditions to ensure the accuracy of data fusion is the core challenge of this module.

[0121] The present invention adopts Bayesian Dynamic Weighting (BDW) to adaptively adjust the data fusion weights of different sensors based on the signal-to-noise ratio (SNR) and confidence of real-time measurement data.

[0122] The specific implementation process includes:

[0123] 1. Set the data sets of different sensors:

[0124] S = {S1, S2,..., S N}

[0125] where S i represents the measurement data of the i-th sensor.

[0126] 2. Calculate the signal-to-noise ratio (SNR) of each sensor:

[0127]

[0128] 3. Calculate the uncertainty of the measurement data:

[0129]

[0130] Among them, λ is the adjustment coefficient, and SNR th is the signal-to-noise ratio threshold.

[0131] 4. Calculate the Bayesian dynamic weight assignment:

[0132]

[0133] 5. Perform weighted fusion to obtain the finally optimized data:

[0134]

[0135] Module 2.3: Noise Suppression Based on Neighborhood Entropy Constraint (NEC-Denoise)

[0136] Since there is inevitably random noise in the sensing data, this embodiment adopts the neighborhood entropy constraint (NEC) method to suppress noise from both the spatial and temporal domains and improve the data quality.

[0137] The specific implementation process includes:

[0138] 1. Calculate the local neighborhood entropy of the data:

[0139]

[0140] Among them, p i is the probability distribution of the data points in this neighborhood.

[0141] 2. Set the noise suppression weight function:

[0142]

[0143] 3. Perform noise suppression processing:

[0144] S denoise (x, y) = W denoise (x, y) · S fused (x, y)

[0145] Final output:

[0146] D s (x, y): Dynamically optimized sensor layout density distribution matrix.

[0147] S fused (x, y, t): Fusion data based on Bayesian dynamic weight assignment.

[0148] Sdenoise (x, y, t): High-precision sensing data optimized by neighborhood entropy constraint denoising.

[0149] This module realizes the dynamic optimization of the sensor layout density through technologies such as adaptive grid division, Bayesian dynamic weight allocation, and neighborhood entropy constraint noise suppression, and improves the accuracy and robustness of data fusion, laying a solid foundation for subsequent nonlinear signal feature enhancement and dynamic solution.

[0150] Specifically, the nonlinear signal feature enhancement module based on multi-scale transform includes:

[0151] After the multi-modal sensing data fusion algorithm based on hierarchical adaptive density optimization in the multi-modal sensing data fusion module is completed, this embodiment has obtained the high-precision sensing data S denoise (x, y, t), which contains the dynamic response information of the subgrade and pavement under the action of moving loads. However, due to the complexity of the subgrade-pavement system under nonlinear dynamics, many subtle damage signals will be masked by background noise or lost due to the limitations of traditional filters. Therefore, the core goal of this module is to enhance the nonlinear signal features using a multi-scale transform domain adaptive filter, thereby improving the identifiability of small damage signals.

[0152] Module 3.1: Multi-scale transform based on adaptive scale selection (MST-ASS)

[0153] Since damage signals usually show local high-frequency abnormal fluctuations, and the low-frequency components of structural responses are mainly affected by the overall stiffness and damping characteristics, a single-scale signal analysis method cannot capture both low-frequency trends and high-frequency local features simultaneously.

[0154] The present invention uses a multi-scale transform (MST, Multi-Scale Transform) to extract the detail components of the signal at different scales and introduces an adaptive scale selection (ASS, Adaptive Scale Selection) mechanism to ensure the selection of the optimal scale, so as to retain small damage features.

[0155] The specific implementation process includes:

[0156] 1. Decompose the signal S using wavelet transform denoise (x, y):

[0157]

[0158] Where:

[0159] W j (x, y, t) is the wavelet detail coefficient (high-frequency component) of the j-th layer, reflecting damage feature information;

[0160] R J (x, y, t) is the wavelet approximation coefficient (low-frequency component) of the Jth layer, representing the overall trend.

[0161] 2. Calculate the energy distribution of wavelet coefficients to screen the optimal scale:

[0162]

[0163] Among them, E j represents the proportion of wavelet energy at the jth layer scale.

[0164] 3. Define the adaptive scale weight function:

[0165]

[0166] Among them, α j represents the importance weight of this scale.

[0167] 4. Select the optimal scale j * :

[0168]

[0169] The detail coefficients corresponding to this scale will be used for subsequent feature enhancement processing.

[0170] Module 3.2: Signal Enhancement Based on Adaptive Filter (AFSE)

[0171] Since the micro-damage signal is often interfered by environmental noise, directly extracting the high-frequency component may lead to misjudgment. Therefore, an adaptive filter (AFSE, Adaptive Filter for Signal Enhancement) is needed to enhance the damage signal and suppress the irrelevant noise.

[0172] This filter dynamically adjusts the filtering parameters through the statistical analysis of local time-frequency features, so as to ensure that the damage signal can be effectively extracted under different working conditions.

[0173] The specific implementation process includes:

[0174] 1. Calculate the local signal-to-noise ratio (L-SNR):

[0175]

[0176] Among them:

[0177] is the local average amplitude;

[0178] σ local(x, y, t) is the local standard deviation.

[0179] 2. Define the filter gain function:

[0180]

[0181] Where:

[0182] β is the adjustment parameter;

[0183] SNR th is the signal-to-noise ratio threshold.

[0184] 3. Enhance the damage signal:

[0185]

[0186] This module can effectively suppress the noise in the low signal-to-noise ratio region and enhance the identifiability of the damage signal.

[0187] Module 3.3: Signal Optimization Based on Structural Information Entropy Constraint (SIE-0)

[0188] Since the damage signal often exhibits local mutation characteristics, the enhanced signal can be further optimized by calculating the structural information entropy of the signal to make it more clearly distinguishable.

[0189] The specific implementation process includes:

[0190] 1. Calculate the local information entropy of the enhanced signal:

[0191]

[0192] where p i is the normalized probability distribution of the signal amplitude.

[0193] 2. Define the information entropy constraint weight function:

[0194]

[0195] 3. Optimize the enhanced signal:

[0196]

[0197] Final output:

[0198] S enhanced (x, y, t): The enhanced signal after multi-scale transformation, adaptive filtering, and structural information entropy optimization. This signal can more clearly reflect the characteristics of micro-damage and suppress background noise.

[0199] Through technologies such as multi-scale transformation, adaptive filtering, and structural information entropy optimization, this module retains both the low-frequency trend information and high-frequency local damage characteristics of the signal, significantly improving the identifiability of micro-damage signals and providing high-quality data input for subsequent non-linear dynamic behavior identification and damage prediction.

[0200] Specifically, the dynamic solution module based on multi-field coupling adaptive grid adjustment includes:

[0201] After the non-linear signal feature enhancement module: the non-linear signal feature enhancement module based on multi-scale transformation, in this embodiment, the dynamic response signal S enhanced (x, y, t) has been obtained, which contains the non-linear dynamic behavior information of the subgrade-pavement system under the action of moving loads. However, traditional dynamic solution methods usually use fixed grid division, which leads to large errors in stress concentration areas and low computational efficiency in low stress gradient areas. Therefore, the core goal of this module is to introduce a multi-field coupling-driven adaptive grid adjustment method so that the grid division can adapt to changes in dynamic characteristics in real time, thereby ensuring the best balance between computational accuracy and efficiency.

[0202] Module 4.1: Construct a multi-field coupling-driven adaptive grid control equation (MAC-CF)

[0203] When the subgrade-pavement system is under the action of moving loads, there will be non-linear coupling effects of stress field, velocity field, acceleration field, and damage field. Therefore, it is difficult to ensure global computational accuracy by relying only on single physical field information to adjust grid division.

[0204] The present invention combines the stress field gradient, velocity field mutation, acceleration field change rate, and damage field characteristics into a grid adjustment driving force to construct a multi-field coupling control equation (MAC-CF, Multi-Field Adaptive Control Function).

[0205] The specific implementation process includes:

[0206] 1. Define multi-field variables:

[0207] Stress field: σ(x, y, t)

[0208] Velocity field: v(x, y, t)

[0209] Acceleration field: a(x, y, t)

[0210] Damage field (calculated by the non-linear signal feature enhancement module of the non-linear signal feature enhancement module): D(x, y, t)

[0211] 2. Calculate the stress gradient field:

[0212]

[0213] Among them, G σ (x, y, t) reflects the severity of local stress changes.

[0214] 3. Calculate the velocity mutation region:

[0215]

[0216] This index is used to identify the region affected by inertial forces under the action of moving loads.

[0217] 4. Calculate the acceleration change rate:

[0218]

[0219] This parameter is used to measure the drastic changes in the acceleration field and reflects the regions of impact loading or nonlinear dynamic phenomena (such as buckling, instability).

[0220] 5. Calculate the damage field gradient:

[0221]

[0222] This item is used to identify potential damage propagation regions and automatically refine the mesh in areas where damage is concentrated.

[0223] 6. Construct the multi-field coupling control equation (MAC-CF):

[0224] G MAC (x, y, t) = w σ G σ (x, y, t) + w v G σ G v (x, y, t) + w a G a (x, y, t) + w D G D (x, y, t)

[0225] Among them:

[0226] w σ , w v , w a , w D are the weight coefficients of different physical fields and can be determined through experimental data or numerical optimization;

[0227] G MAC (x, y, t) serves as the driving force for mesh adaptive adjustment and controls the fineness of mesh generation.

[0228] Module 4.2: Adaptive Density Function Based Mesh Generation (ADF-MG)

[0229] Traditional mesh generation methods usually adopt uniform or single-variable based refinement strategies. However, in this problem, the mesh density in different regions should be dynamically adjusted according to the multi-field governing equation G MAC (x, y, t).

[0230] Therefore, this embodiment uses an Adaptive Density Function (ADF) to determine the local mesh size.

[0231] The specific implementation process includes:

[0232] 1. Define the mesh density function:

[0233]

[0234] Where:

[0235] ρ min and ρ max are the minimum and maximum mesh densities respectively;

[0236] and are the global minimum and maximum multi-field coupling control values respectively;

[0237] This formula ensures that the mesh is automatically refined in areas of stress concentration, significant damage, velocity mutation, and severe acceleration change, while reducing the mesh density in low-gradient areas.

[0238] 2. Calculate the mesh size distribution:

[0239]

[0240] This formula determines the local mesh size h(x, y, t). The smaller the value, the denser the mesh, and the larger the value, the sparser the mesh.

[0241] Module 4.3: Dynamic Topology Optimization Based Mesh Reconstruction (DTO-MG)

[0242] Traditional finite element mesh adjustment methods mainly rely on fixed rules or manual adjustment, and it is difficult to respond in real time to dynamic load and stress distributions.

[0243] The present invention introduces a Dynamic Topology Optimization (DTO) method to ensure that the overall mesh quality is not deteriorated while locally adjusting the mesh.

[0244] The specific implementation process includes:

[0245] 1. Adjust the element shape based on the mesh quality evaluation function:

[0246]

[0247] Where:

[0248] A is the area of the triangular element;

[0249] l i is the side length of the triangular element;

[0250] Q(x, y, t) ranges from 0 to 1, and the closer the value is to 1, the better the mesh quality.

[0251] 2. Mesh reconstruction strategy:

[0252] If Q(x, y, t) < Q th (set threshold), then locally optimize the element shape to reduce distortion;

[0253] If the mesh is over-refined and the calculation time exceeds the limit, then adjust the density function ρ(x, y, t) in combination with error estimation to reduce the computational burden.

[0254] Final output:

[0255] h(x, y, t): The mesh size distribution after adaptive adjustment;

[0256] The optimized mesh topology structure ensures the calculation accuracy while improving the calculation efficiency;

[0257] Minimize the dynamic solution error and provide higher-precision input data for subsequent non-linear behavior modeling.

[0258] This module realizes automatic mesh refinement in regions with high stress gradients, velocity mutations, acceleration changes, and damage concentration characteristics through an adaptive mesh adjustment strategy driven by multi-field coupling. At the same time, it adaptively reduces the mesh density in low-gradient regions, ensuring the best balance between calculation accuracy and efficiency, and providing a solid numerical calculation basis for subsequent non-linear dynamic behavior identification and prediction.

[0259] Specifically, the damage prediction module based on multi-stage reinforcement learning includes:

[0260] In the dynamic solution module: After the dynamic solution module based on multi-field coupling adaptive mesh adjustment, this embodiment has obtained high-precision dynamic stress field, velocity field, acceleration field, and damage field distribution data, and ensured the calculation efficiency through the adaptive mesh optimization strategy. However, traditional damage prediction methods usually rely on static machine learning models or regression analysis based on limited samples, and it is difficult to accurately simulate the damage accumulation and evolution trend under long-term load.

[0261] The core objective of this module is to introduce the multi-stage reinforcement learning (MSRL) method, enabling the damage prediction model to dynamically optimize the learning strategy at different stages, thereby improving the accuracy of long-term prediction.

[0262] Module 5.1: Construct the Damage Evolution State Space (DESS)

[0263] Traditional reinforcement learning methods face the problems of high-dimensional state space sparsity and unstable decision-making paths in complex dynamic systems, making it difficult to effectively learn the long-term damage evolution law.

[0264] The specific implementation process includes: By introducing the Damage Evolution State Space (DESS) and combining stress, velocity, acceleration, and damage characteristics, this module constructs the state representation of reinforcement learning to optimize the long-term prediction ability.

[0265] 1. Define the damage evolution state variables:

[0266] Stress state: S σ (t) = {σ(x, y, t)}

[0267] Velocity state: S v (t) = {v(x, y, t)}

[0268] Acceleration state: S a (t) = {a(x, y, t)}

[0269] Damage state: S D (t) = {D(x, y, t)}

[0270] 2. Construct the damage evolution state vector:

[0271] S(t) = [S σ (t), S v (t), S a (t), S D (t)]

[0272] where S(t) represents the system damage state at the current time step t.

[0273] 3. Define the damage accumulation objective function:

[0274]

[0275] where:

[0276] R(t) represents the reinforcement learning reward function for the degree of damage accumulation;

[0277] α σ ,α v ,α a ,α D is the weight coefficient of each physical field;

[0278] This objective function ensures that the reinforcement learning can accurately capture the damage accumulation effect, thus optimizing the long-term prediction ability.

[0279] Module 5.2: Construct a reinforcement learning framework (MSRL) based on a multi-stage learning strategy

[0280] Traditional reinforcement learning methods adopt a single policy update mechanism, which is difficult to adapt to the evolution characteristics of different damage stages.

[0281] The present invention adopts phased policy optimization to dynamically adjust the learning strategy at different stages, improving the long-term prediction ability of the model.

[0282] The specific implementation process includes:

[0283] 1. Divide the damage evolution stage:

[0284] Initial micro-damage stage (Elastic Deformation Phase): The damage variable D(x, y, t) changes slowly and is mainly controlled by low-frequency loads.

[0285] Plastic damage development stage (Plastic Damage Phase): The damage variable DD(x, y, t) begins to increase rapidly, and the system enters a non-linear damage accumulation process.

[0286] Fatigue failure stage (Fatigue Failure Phase): The damage variable D(x, y, t) approaches the threshold D crit , and the structure enters the critical failure state.

[0287] 2. Construct a multi-stage policy switching function:

[0288]

[0289] Where:

[0290] π1(t) is the learning strategy in the micro-damage stage, mainly optimizing the linear damage prediction model;

[0291] π2(t) is the learning strategy in the plastic damage development stage, using deep reinforcement learning to optimize the damage prediction ability;

[0292] π3(t) is the learning strategy in the fatigue failure stage, optimizing the failure prediction accuracy to ensure accurate identification of the critical damage state.

[0293] Module 5.3: Reinforcement Learning Training Based on Experience Replay and Long-Term Discount Factor

[0294] Due to the time-dependence of damage evolution, it is difficult for traditional reinforcement learning methods to effectively utilize long-term historical data for training.

[0295] The present invention introduces Experience Replay and Long-Term Discount Factor to ensure that the damage prediction model can capture the cumulative damage effect under long-term load action.

[0296] The specific implementation process includes:

[0297] 1. Construct an experience pool:

[0298]

[0299] Where:

[0300] S(t) is the current damage state;

[0301] A(t) is the prediction strategy adopted by the current reinforcement learning model;

[0302] R(t) is the damage accumulation objective function at the current time step;

[0303] S(t + 1) is the damage state at the next time step.

[0304] 2. Update the strategy using the long-term discount factor:

[0305]

[0306] Where:

[0307] γ is the long-term discount factor (usually set to 0.9 - 0.99), ensuring that the model can focus on the long-term damage effect;

[0308] Q(S t , A t ) is the Q-learning value function.

[0309] 3. Optimize the damage prediction model based on policy gradient:

[0310]

[0311] Where:

[0312] θ are the parameters of the reinforcement learning model;

[0313] η is the learning rate;

[0314] This optimization strategy ensures that reinforcement learning can dynamically adjust the prediction strategy and improve the accuracy of damage prediction.

[0315] Final output:

[0316] D pred (x, y, t): Long-term damage prediction result;

[0317] The damage evolution model optimized by reinforcement learning ensures long-term prediction accuracy;

[0318] The adaptive policy optimization mechanism enables the model to automatically adjust the learning strategy at different stages.

[0319] This module realizes the adaptive optimization of the damage prediction strategy through a multi-stage reinforcement learning method, enabling the damage evolution model to dynamically adjust the learning strategy and improve the long-term prediction accuracy, providing more accurate and reliable data support for the nonlinear dynamic analysis, health monitoring, and life prediction of the subgrade-pavement system.

[0320] Specifically, the digital twin decision-making module based on intelligent optimization feedback includes:

[0321] In the damage prediction module: In the damage prediction module based on multi-stage reinforcement learning, a long-term damage prediction model has been established in this embodiment, and the prediction accuracy of damage evolution has been optimized through a multi-stage reinforcement learning strategy. However, traditional optimization methods lack a dynamic feedback mechanism and are difficult to adjust the optimization strategy in real time to adapt to complex engineering environments, resulting in decision-making lags or inadaptability to sudden situations. Therefore, this module introduces an intelligent optimization feedback mechanism (Intelligent Optimization Feedback, IOF) to construct a digital twin (Digital Twin, DT) decision-making framework to achieve dynamic decision-making optimization based on real-time perception, intelligent feedback, and adaptive optimization.

[0322] Module 6.1: Construct a digital twin virtual-real mapping model (Digital Twin Mapping, DTM)

[0323] Traditional digital twin methods are mostly based on static models or predefined parameter optimization and are difficult to adapt to the complex nonlinear characteristics under dynamic loads.

[0324] This module introduces a high-dimensional state mapping function to establish a dynamic feedback channel between the physical system (Physical System, PS) and the virtual simulation system (Virtual System, VS) to ensure the real-time performance and accuracy of the twin model.

[0325] The specific implementation process includes:

[0326] 1. Define the state variables of the physical system:

[0327] Stress state: S σ (t)

[0328] Displacement state: S u (t)

[0329] Damage state: S D (t)

[0330] Environmental state: S E (t)

[0331] 2. Define the digital twin mapping function:

[0332]

[0333] Where:

[0334] S σ (t), S u (t), S D (t), S E (t) are the real-time measurement values of the physical system;

[0335] are the calculated values of the digital twin system;

[0336] This mapping function ensures the high-dimensional synchronization of the physical system and the virtual system.

[0337] 3. Optimize the twin mapping accuracy based on an adaptive filter:

[0338]

[0339] Where:

[0340] λ is the adaptive weight factor to ensure that the twin system can quickly correct errors;

[0341] This optimization strategy can effectively improve the dynamic adaptability of the digital twin model.

[0342] Module 6.2: Build an intelligent optimization feedback mechanism (IOF)

[0343] Traditional optimization methods are based on fixed rules or single-step optimization strategies, lacking real-time feedback capabilities and being difficult to adapt to complex environmental changes.

[0344] This module builds a real-time optimization adjustment strategy based on reinforcement learning through an intelligent optimization feedback mechanism (I0F) to ensure the adaptability of decision-making optimization.

[0345] The specific implementation process includes:

[0346] 1. Define the optimization objective function:

[0347]

[0348] Among them:

[0349] J opt is the optimization objective;

[0350] Θ is the set of optimization parameters;

[0351] ω σ , ω D is the weight factor of each state variable;

[0352] This objective function ensures that the optimization strategy can adjust the damage prediction deviation and stress response error in real time.

[0353] 2. Construct the reinforcement learning optimization strategy:

[0354]

[0355] Among them:

[0356] π IOF (t) is the reinforcement learning strategy of intelligent optimization feedback;

[0357] R IOF (t) is the immediate reward of the optimization strategy;

[0358] This optimization strategy improves the system adaptability by dynamically adjusting the optimization parameters.

[0359] 3. Dynamically adjust the optimization parameters based on Bayesian optimization:

[0360]

[0361] Among them:

[0362] Through the Bayesian optimization method, ensure that the optimization parameters are adaptively adjusted in different damage stages to improve the decision-making accuracy.

[0363] Module 6.3: Construct an optimization control model based on twin decisions

[0364] Traditional decision-making methods adopt an offline optimization strategy and are difficult to achieve real-time optimization.

[0365] This module realizes dynamic optimization control through digital twin + reinforcement learning + intelligent feedback mechanism.

[0366] The specific implementation process includes:

[0367] 1. Define the twin decision control strategy:

[0368]

[0369] Among them:

[0370] u DT (t) is the twin decision control variable;

[0371] K DT is the optimal gain matrix of the twin system;

[0372] This control strategy can effectively reduce the error between the twin system and the physical system.

[0373] 2. Optimization control strategy based on MPC (Model Predictive Contro1):

[0374]

[0375] Where:

[0376] The MPC method is adopted to ensure that the control strategy can be adaptively adjusted at different stages to optimize the overall decision-making effect.

[0377] Final output:

[0378] u DT (t): Control decision optimized based on digital twin;

[0379] Intelligent optimization feedback mechanism (IOF) to ensure that the optimization strategy can be adjusted in real time;

[0380] Optimization control model based on reinforcement learning to improve decision-making accuracy and adaptability.

[0381] This module constructs a high-precision dynamic optimization decision-making system through a digital twin decision-making module based on intelligent optimization feedback, ensuring that the damage prediction model can adjust the optimization strategy in real time to adapt to complex environmental changes and improve the long-term operation safety and durability of the load-subgrade-pavement system.

[0382] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. Subgrade and pavement test and data acquisition system under moving load, characterized in that Including: Mobile load simulation module: Collect high-precision road surface unevenness data, perform feature decomposition and noise suppression using a spatio-temporal convolutional variational autoencoder network, and optimize the load action path based on adaptive deformation gradient analysis; Multi-modal sensing data fusion module: According to the normal stiffness distribution and road surface morphology gradient, adaptively adjust the sensor layout density using adaptive grid division, fuse multi-modal sensor data using Bayesian dynamic weight allocation, and suppress noise through neighborhood entropy constraint; Nonlinear signal feature enhancement module: Extract signal detail components using multi-scale transformation combined with adaptive scale selection, enhance damage signals and suppress noise through an adaptive filter, and then optimize and enhance the signals based on structural information entropy constraint to improve the identifiability of damage signals; Dynamics solution module: Construct an adaptive grid control equation driven by multi-field coupling, adjust the grid density according to the coupling effects of the stress field, velocity field, acceleration field, and damage field, and reconstruct the grid using dynamic topology optimization methods; Damage prediction module: Establish a state space based on damage evolution, construct a reinforcement learning framework with a multi-stage learning strategy, train the model through experience replay and long-term discount factors, and realize the adaptive optimization of the damage prediction strategy; Digital twin decision-making module: Construct a digital twin virtual-real mapping model, establish a dynamic feedback channel between the physical system and the virtual simulation system, dynamically adjust and optimize parameters through an intelligent optimization feedback mechanism combined with reinforcement learning and Bayesian optimization, construct an optimization control model based on twin decision-making, and realize dynamic optimization decision-making.

2. The subgrade and pavement test and data acquisition system under moving load according to claim 1, characterized in that The mobile load simulation module includes: Using multi-sensor fusion technology, through the collaborative work of laser point cloud scanning, inertial measurement unit, and ultra-wideband radar, collect road surface height point cloud data, the attitude changes of the vehicle-borne load simulation device, and the non-uniform morphology interference information caused by the change of subgrade materials, fuse the data, and establish a continuous, seamless, and high-precision three-dimensional road surface morphology model; Use a spatio-temporal convolutional variational autoencoder network to perform multi-layer feature decomposition on the collected data, improve data reliability through an adaptive noise suppression mechanism, and thus provide high-precision road surface morphology data input for subsequent adaptive feedback adjustment.

3. The subgrade and pavement test and data acquisition system under moving load according to claim 2, wherein The multi-modal sensing data fusion module includes: Adopt adaptive grid division, and dynamically adjust the spatial distribution density of sensors according to the normal stiffness gradient and road surface morphology gradient, optimize the sensor layout, make the sensors in the high-stress area more dense, and reduce the sensor density in the low-stress area; Based on Bayesian dynamic weight allocation, adaptively adjust the data fusion weight according to the signal-to-noise ratio and confidence of different sensors to ensure the accuracy of the fused data; Use the neighborhood entropy constraint method to suppress noise from both the spatial and temporal domains to improve data quality.

4. The subgrade and pavement test and data acquisition system under moving load according to claim 3, wherein The nonlinear signal feature enhancement module includes: Adopt multi-scale transformation with adaptive scale selection, decompose the signal through wavelet transform, calculate the energy distribution of wavelet coefficients, and select the optimal scale to retain micro-damage features; Signal enhancement based on an adaptive filter, dynamically adjust the filtering parameters through local signal-to-noise ratio calculation and filter gain function, enhance damage signals and suppress irrelevant noise; Using signal optimization based on structural information entropy constraint, calculate the local information entropy of the enhanced signal, and further optimize the signal through the information entropy constraint weight function.

5. The subgrade and pavement test and data acquisition system under moving load according to claim 4, characterized in that The dynamic solution module includes: Construct an adaptive grid control equation driven by multi-field coupling, and synthesize the stress field gradient, velocity field mutation, acceleration field change rate, and damage field characteristics to form the driving force for grid adjustment; Adjust the grid based on the adaptive density function, dynamically determine the local grid size according to the multi-field coupling control equation, and achieve grid encryption in areas with stress concentration and significant damage as well as grid sparsification in low-gradient areas; Introduce dynamic topology optimization, adjust the element shape through the grid quality evaluation function, and optimize the grid topology structure to ensure that the overall grid quality does not deteriorate while locally adjusting the grid.

6. The subgrade and pavement test and data acquisition system under moving load according to claim 5, characterized in that, The damage prediction module includes: Construct a state space based on damage evolution, define state variables by combining stress, velocity, acceleration, and damage characteristics, construct a damage evolution state vector, and define a damage accumulation objective function to optimize the long-term prediction ability; Construct a reinforcement learning framework based on a multi-stage learning strategy, divide the damage evolution stages, including the initial micro-damage stage, plastic damage development stage, and fatigue failure stage, design corresponding learning strategies respectively, and dynamically adjust the learning strategy through the multi-stage strategy switching function; Adopt experience replay and long-term discount factor for reinforcement learning training, construct an experience pool, update the strategy using the long-term discount factor, and optimize the damage prediction model parameters based on the policy gradient.

7. The subgrade and pavement test and data acquisition system under moving load according to claim 6, characterized in that, The digital twin decision module includes: Construct a digital twin virtual-real mapping model, dynamically synchronize the stress, displacement, damage, and environmental states of the physical system with the virtual simulation system through a high-dimensional state mapping function, and optimize the twin mapping accuracy based on an adaptive filter to ensure the high-dimensional synchronization of the physical system and the virtual system; Construct an optimization feedback mechanism, define an optimization objective function, combine the reinforcement learning optimization strategy, dynamically adjust the optimization parameters, and achieve the adaptive adjustment of the parameters through the Bayesian optimization method to ensure the self-adaptability of decision optimization; Construct an optimization control model based on twin decision-making, define the twin decision control strategy, and adopt model predictive control to optimize the control strategy to achieve dynamic optimization control.

Images