Dynamic compaction construction intelligent monitoring and self-adaptive control method based on physical information neural network

Abstract

The invention relates to a dynamic compaction construction intelligent monitoring and self-adaptive control method based on a physical information neural network, and the method comprises the steps: arranging a vibration sensor array in a construction safety region, collecting a compaction vibration signal in real time, and extracting time domain, frequency domain and spatial propagation characteristics; constructing a physical information neural network model, taking a soil elastic wave propagation equation as a physical constraint, taking vibration characteristics and monitoring point space coordinates as input, and predicting three-dimensional distribution of soil physical property parameters; based on the predicted soil body parameters, calculating the distribution and development trend of the bearing capacity of the foundation by adopting a Taisha foundation bearing capacity theory; a multi-target optimization model with the bearing capacity reaching the standard, the minimum energy consumption and the shortest time as targets is established, and follow-up tamping parameters are optimized; and construction is executed according to the optimized parameters, and data and model parameters are continuously updated. Compared with the prior art, the problems that in traditional dynamic compaction construction, the soil body state cannot be observed, and optimization lacks theoretical support are solved, and intelligentization and accurate control over the construction process are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent construction and foundation treatment, and particularly relates to a dynamic compaction construction intelligent monitoring and self-adaptive control method based on a physical information neural network. BACKGROUND

[0002] Dynamic compaction is a widely used foundation treatment technology, which improves the engineering properties of soil foundation through the powerful impact energy generated by the falling of heavy hammer from high altitude. However, the traditional dynamic compaction construction faces the following technical challenges: (1) The internal state of the soil body cannot be observed: The traditional monitoring method is mainly based on ground settlement measurement, which cannot directly observe the density distribution, modulus change and bearing capacity development of the soil body inside. This leads to a lack of scientific basis in the construction process, which seriously affects the quality control of construction. The existing technology CN116399302A proposes a dynamic compaction settlement monitoring method based on binocular vision and neural network, which realizes non-contact real-time measurement, but its technical core is still limited to the observation of ground displacement, and cannot reflect the changes of the internal mechanical state of the soil body, and relies on visible light imaging, and the recognition reliability decreases under dust shielding or complex lighting conditions.

[0003] (2) Difficulty in modeling physical mechanisms: The response mechanism of soil under dynamic compaction is extremely complex, involving nonlinear large deformation, stress wave propagation, pore water pressure change and other physical processes. Traditional empirical formulas are difficult to accurately describe these complex physical phenomena. The existing technology CN112252292A uses artificial neural network to establish a compaction degree calculation model, but its input parameters are concentrated on the control parameters of the road roller and the ground sensor data, and does not involve the deep mechanism of dynamic compaction impact energy transmission and soil body internal wave response. Moreover, the model relies on historical data training, and the generalization ability is limited, which is difficult to adapt to the significant differences in geological conditions in dynamic compaction construction.

[0004] (3) Lack of theoretical basis for construction parameter optimization: Existing technologies mainly rely on engineering experience to determine the number of ramming, energy size and intermittent time, and lack of systematic optimization method. This leads to low energy utilization efficiency and high construction cost.

[0005] (4) Insufficient real-time control capability: Traditional methods cannot adjust construction parameters according to real-time changes in soil state, making it difficult to achieve adaptive control and affecting construction efficiency and quality stability.

[0006] In summary, the existing technologies have not solved the core problems of real-time perception of soil body internal state, coupled modeling of multiple physical processes and dynamic optimization of construction parameters in dynamic compaction construction. Therefore, there is an urgent need for a technology that can deeply integrate soil body internal response monitoring, intelligent modeling and real-time control to realize the precision and intelligentization of dynamic compaction construction. SUMMARY

[0007] The present application aims to overcome the defects of the prior art, and provides a physical information neural network-based intelligent monitoring and adaptive control method for dynamic compaction construction, which solves the technical problems of unobservable soil state, difficult physical mechanism modeling, lack of theoretical basis for parameter optimization, and insufficient real-time control capability in traditional dynamic compaction construction.

[0008] The object of the present application can be achieved by the following technical solutions: The present application provides a physical information neural network-based intelligent monitoring and adaptive control method for dynamic compaction construction, comprising the following steps: S1, arranging a vibration sensor array within a safety range of 20-50 meters from a ramming point, collecting ground vibration signals generated by ramming in real time, and extracting time-domain features, frequency-domain features, and spatial propagation features of the vibration waveform; S2, constructing a physical information neural network model, taking the soil elastic wave propagation equation as a physical constraint, taking the vibration features extracted in S1 and the spatial coordinates of the monitoring points as inputs, and predicting the spatial distribution of soil physical parameters; S3, based on the spatial distribution of soil physical parameters predicted by the physical information neural network model, calculating the three-dimensional distribution and development trend of the foundation bearing capacity using the Terzaghi bearing capacity theory; S4, based on the calculated three-dimensional distribution and development trend of the foundation bearing capacity, establishing a multi-objective optimization model with the goals of meeting the bearing capacity standard, minimizing energy consumption, and minimizing construction time, and optimizing subsequent construction parameters; S5, performing dynamic compaction construction according to the optimized construction parameters, and continuously updating the vibration feature data and the physical information neural network model parameters during the construction process to realize adaptive control of the construction process.

[0009] Further, the time-domain features include peak amplitude , effective value , and duration ; The frequency-domain features include main frequency , frequency band energy , and spectral barycenter ; The spatial propagation features include amplitude decay rate , phase difference , and propagation speed , wherein: , are the vibration amplitudes at distances of 20 meters and 50 meters from the ramming point, , , , , , , is a dimensionless parameter, , respectively represent the radial distance between the sensor i and j and the tamping point.

[0010] Further, in S2, a physical information neural network model is constructed, taking the soil elastic wave propagation equation as a physical constraint, and taking the vibration features extracted in S1 and the spatial coordinates of the monitoring points as inputs, to predict the specific process of the three-dimensional distribution of the soil density field, elastic modulus field and Poisson's ratio field, which includes: Combining the vibration feature vector extracted in S1 and the spatial coordinate vector of the monitoring points into an input vector; Constructing a multi-layer feedforward neural network structure containing an input layer, multiple hidden layers and an output layer, wherein the hidden layers adopt ReLU activation function for nonlinear transformation, and the output layer outputs the three physical parameters of soil density, elastic modulus and Poisson's ratio; In the network training process, the soil elastic wave propagation equation is taken as a physical constraint condition, the residual error between the network output and the physical equation is calculated through automatic differentiation technology, and the residual error is taken as a physical equation residual error loss into the overall loss function, wherein the overall loss function is composed of data fitting loss, physical equation residual error loss, boundary condition loss and initial condition loss, wherein the data fitting loss ensures that the network output is consistent with the measured data, the physical equation residual error loss ensures that the prediction result conforms to the wave propagation physical law, and the boundary condition loss and the initial condition loss ensure the uniqueness and stability of the solution; The gradient-based optimization algorithm is used to minimize the overall loss function, and the network parameters are iteratively updated, and finally a trained physical information neural network model is obtained, which can accurately predict the three-dimensional spatial distribution of the soil physical parameters with vibration features and spatial coordinates as inputs.

[0011] Further, in S2, the input of the physical information neural network model is , wherein: is a vibration feature vector, is a spatial coordinate vector; The output of the physical information neural network model is , wherein: the soil density field distribution, the elastic modulus field distribution, the Poisson's ratio field distribution; The loss function of the physical information neural network model is: , wherein: is the data fitting loss; is the physical equation residual error loss; is a boundary condition loss; is an initial condition loss; , , is a weight coefficient, is the total number of sample data points collected by the vibration sensor, is the true observation value of the vibration feature of the i-th position, is the predicted value of the corresponding vibration feature by the physical information neural network model, represents the residual error obtained by substituting the neural network predicted field quantity u into the soil elastic wave control equation, is the number of spatial points used when calculating the physical residual error.

[0012] Further, in S2, the soil elastic wave propagation equation is a three-dimensional elastic wave propagation equation, and the soil elastic wave propagation equation is specifically: wherein: is a displacement vector, is a soil density field distribution, is time, ∇ is a Hamiltonian operator for calculating spatial gradient and divergence, and is the Lame constant of the elastic properties of the soil, and the relationship with the elastic modulus and Poisson's ratio is: wherein E represents the elastic modulus of the soil, ν represents the Poisson's ratio of the soil.

[0013] Further, in S3, based on the spatial distribution of the soil physical property parameters predicted by the physical information neural network model, the specific process of calculating the three-dimensional distribution and development trend of the foundation bearing capacity using the Terzaghi bearing capacity theory includes: Based on the three-dimensional distribution data of the soil density field, the elastic modulus field, and the Poisson's ratio field predicted by the physical information neural network model in S2, the Lame constant of the soil is calculated according to the elastic modulus and Poisson's ratio; Based on the predicted soil density, the cohesion and internal friction angle of the soil are calculated through an empirical relationship; The calculated cohesion, internal friction angle, as well as the specific gravity of the soil, the foundation burial depth, and the width are used as input parameters, and the Terzaghi bearing capacity theory formula is used to calculate the three-dimensional spatial distribution of the foundation bearing capacity; By analyzing the changes in the spatial distribution of the bearing capacity at different construction stages, the development trend of the foundation bearing capacity is predicted.

[0014] Further, in S3, the Terzaghi bearing capacity theory formula is: wherein: is the bearing capacity of the ground, in kPa, is the cohesion, in kPa, calculated by an empirical relationship , is the reference value of the cohesion, in kPa, is the proportionality coefficient of the cohesion with respect to the density, in Pa / (kg / m 3 ), p is the actual density of the soil, in kg / m 3 , p 0 is the reference density of the soil, in kg / m 3 ; is the unit weight of the soil above the foundation, in kN / m 3 ; is the unit weight of the soil below the foundation, in kN / m 3 ; is the depth of the foundation, in m; is the width of the foundation, in m; , , is the bearing capacity coefficient, dimensionless, calculated based on the internal friction angle, the relationship of which with the density of the soil is: , is the reference value of the friction angle, in rad, is the proportionality coefficient of the friction angle with respect to the density, in rad.

[0015] Further, in S4, based on the three-dimensional distribution and development trend of the calculated bearing capacity of the ground, a multi-objective optimization model is established with the bearing capacity reaching the standard, the minimum energy consumption and the shortest construction time as the objectives, and the specific process of optimizing the subsequent construction parameters includes: based on the development trend of the bearing capacity of the ground predicted by S3 and the current spatial distribution state of the bearing capacity, an optimization problem is constructed which includes multiple conflicting objective functions; the multi-objective optimization model includes: a function aimed at punishing the predicted bearing capacity that does not reach the design target value, a function calculating the total energy consumption of all subsequent ramming, and a function calculating the total time of all subsequent ramming intervals; determine the subsequent ramming times, the energy value of each ramming, and the interval time after each ramming as decision variables, and set the constraint conditions of the decision variables in line with the range of engineering practice; a multi-objective evolutionary algorithm is used to solve the optimization problem, and a set of Pareto optimal solutions that achieve the best balance among multiple objectives is generated through iterative calculation; Finally, a preferred solution is selected from the Pareto optimal solution set, and the corresponding subsequent construction parameter sequence is output, including the specific number of ramming, the energy configuration of each ramming, and the interval time arrangement.

[0016] Further, in S4, the multi-objective optimization model is: Wherein: is the penalty for insufficient bearing capacity, is the preset foundation bearing capacity target value, is the current foundation bearing capacity value based on model prediction; is the total energy consumption, is the energy value applied by the i-th ramming; is the total construction time, is the interval time required after the i-th ramming is completed; is the decision variable, is the total number of ramming; The constraint conditions include: the ramming number range , the single ramming energy kJ; the interval time range h.

[0017] Further, according to the optimized construction parameters, the dynamic compaction construction is performed, and the vibration feature data and the physical information neural network model parameters are continuously updated during the construction process to realize the adaptive control of the construction process. The specific process includes: According to the optimized construction parameter sequence output by S4, subsequent ramming operations are performed, and new ground vibration signals are collected in real time through the vibration sensor array arranged in S1 during each ramming process; The newly collected vibration signals are processed to extract new vibration feature parameters, and the vibration feature data set in S1 is updated and expanded; Using incremental learning, the newly obtained vibration feature data and spatial coordinates are used as training samples to update the parameters of the physical information neural network model constructed in S2, and the prediction accuracy is optimized; Based on the updated physical information neural network model, the bearing capacity calculation and trend prediction of S3 are re-executed, and the subsequent construction parameters are dynamically adjusted through the multi-objective optimization model of S4; Through real-time data collection, continuous model updating, and cyclic parameter optimization, a closed-loop feedback control process is formed to realize intelligent monitoring and adaptive control of the dynamic compaction construction process.

[0018] Compared with the prior art, the present application has the following beneficial effects: Firstly, by arranging a vibration sensor array and extracting multi-dimensional features of the vibration waveform, the energy propagation and soil response signal during ramming are comprehensively perceived, the technical problem of the internal state of the soil being not directly observable in the traditional method is solved, and reliable data foundation is provided for subsequent analysis.

[0019] Secondly, by constructing a physical information neural network model, the physical law of soil elastic wave propagation is embedded into the learning process as a constraint, effectively combining vibration monitoring data and internal physical mechanisms, solving the difficulty of accurately modeling the constitutive relationship of complex soil, and significantly improving the accuracy of soil parameter inversion and state prediction.

[0020] Thirdly, based on the soil parameters predicted by the physical information neural network, the bearing capacity development state of the foundation is scientifically evaluated using the Terzaghi bearing capacity theory, and a multi-objective optimization model is established, which changes the parameter decision-making from relying on experience to an optimization process based on model prediction and theoretical calculation, solving the problem of lack of theoretical basis for parameter optimization.

[0021] Fourthly, by continuously collecting new data and updating the model while executing the optimized parameters, a closed-loop control loop of perception, prediction, optimization, execution and feedback is formed, realizing dynamic adjustment of construction parameters and autonomous optimization of the process, solving the problem of insufficient real-time control capability in traditional dynamic compaction construction, and significantly improving construction quality and efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 The flowchart of the technical scheme of the present application. DETAILED DESCRIPTION

[0023] As a new machine learning method, Physics-Informed Neural Networks (PINNs) can embed physical laws into neural networks to achieve high-precision prediction under small sample conditions. The present application introduces PINNs technology into the field of dynamic compaction construction, providing a new idea for solving the technical problems in the prior art.

[0024] The intelligent monitoring and adaptive control method of dynamic compaction construction based on physical information neural network in the present application arranges a vibration sensor array in the construction safety area, collects ground vibration signals caused by ramming in real time, extracts vibration waveform feature parameters, constructs a PINNs model that integrates soil wave propagation physical equations, takes vibration features and spatial coordinates as input, predicts the three-dimensional distribution of soil density field, modulus field and Poisson's ratio field, calculates the development of foundation bearing capacity based on Terzaghi bearing capacity theory, and determines the optimal construction parameters using a multi-objective optimization algorithm. This method realizes accurate prediction of the internal state of the soil and intelligent control of the construction process.

[0025] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments. In this technical solution, if the component model, material name, connection structure, circuit structure, control method, algorithm and other features are not explicitly stated, they are considered as common technical features disclosed in the prior art.

[0026] Embodiment 1 The technical objectives to be achieved by the technical solution of this embodiment include: (1) constructing a physical mapping relationship between vibration characteristics and the internal state of the soil, and realizing accurate prediction of the soil density distribution and bearing capacity development that cannot be directly observed; (2) developing a PINNs modeling method based on physical constraints to ensure that the prediction results comply with the basic laws of soil mechanics and improve the prediction accuracy under small sample conditions; (3) establishing a multi-objective optimization algorithm for construction parameters to scientifically determine the optimal ramming times, energy allocation and time arrangement, and maximize energy utilization efficiency.

[0027] The present application is realized by the following technical solutions: The present application provides a strong ramming construction intelligent monitoring and self-adaptive control method based on a physical information neural network, referring to Figure 1 , specifically comprising the following steps: S1: A three-dimensional vibration sensor array is arranged within a safe range of 20-50 meters from the ramming point to collect ground vibration signals generated by ramming and extract time domain, frequency domain and spatial propagation characteristics of the vibration waveform.

[0028] In specific implementation, in S1, a three-ring three-dimensional vibration sensor array is arranged within a safe range of 20-50 meters from the ramming point, a GS-20DX three-direction velocity sensor of the GeoSpace company is used, the frequency response range is 4.5Hz-400Hz, and the sensitivity is 28.8 V / (m / s). The sensor array is distributed in three rings, i.e., 25 meters, 40 meters and 55 meters, respectively, 4 measuring points are arranged in each ring, and a spatial layout of 12 monitoring points is formed.

[0029] In specific implementation, in S1, the three-ring sensor arrangement scheme is as follows: the near-field ring ( ) captures high-intensity vibration signals to obtain the dynamic response characteristics of the shallow soil; the middle-field ring ( ) collects medium-intensity vibration signals to reflect the changes in physical parameters of the middle soil layer; and the far-field ring ( m) records the attenuated vibration signals to obtain the propagation characteristics of the deep soil. Each sensor has a buried depth of 0.6m, a 10cm-thick fine sand cushion is laid at the bottom, a high-strength steel protective cover is used, the sealing level is IP68, and it can withstand 100g impact acceleration.

[0030] In specific implementation, in S1, the system adopts a sampling frequency of 1000 Hz to collect the vibration signal, and adopts a 24-bit high-precision A / D converter with a dynamic range greater than 120 dB. The collected original vibration signal is first preprocessed, including removing the direct current component, band-pass filtering to remove high-frequency noise and low-frequency interference, and steps such as abnormal value detection and elimination.

[0031] In specific implementation, in S1, the vibration feature extraction includes three types of key vibration feature parameters: the time domain features include peak amplitude , effective value , and duration ; the frequency domain features include main frequency , frequency band energy , and spectral barycenter ; and the spatial propagation features include amplitude decay rate , phase difference , and propagation speed . Among them , are the vibration amplitudes at distances of , from the ramming point, , are in units of mm / s, , are in units of m, is a dimensionless parameter, , respectively represent the radial distances between the sensors i and j and the ramming point. In specific implementation, by selecting two vibration sensors at different positions (such as distances of = 25 meters and = 40 meters from the ramming point), the actual measurement distance difference and the corresponding amplitude values ( , ) are directly substituted into the formula to calculate the amplitude decay rate.

[0032] S2: Construct a physical information neural network (PINNs) model, take the soil elastic wave propagation equation as the physical constraint, take the vibration feature vector and three-dimensional spatial coordinates as the input, and predict the spatial distribution of the soil physical parameters; In S2, a physical information neural network model is constructed, the network input includes an 8-dimensional vibration feature vector and a 3-dimensional spatial coordinate, and the output is the soil density, elastic modulus and Poisson's ratio of the corresponding position. The network adopts an encoder-decoder structure, the encoder part adopts a three-layer fully connected neural network, the first layer maps the 8-dimensional vibration feature to a 64-dimensional hidden representation, the second layer expands to 128 dimensions and enhances the feature expression ability, the third layer compresses to 64 dimensions, and the fourth layer further compresses to a 32-dimensional encoding vector, and a ReLU activation function is used between layers. The decoder part connects the 32-dimensional encoding vector with the 3-dimensional spatial coordinate to form a 35-dimensional input vector, passes through a 64-dimensional fully connected network, and finally outputs a 3-dimensional soil parameter vector.

[0033] In specific implementation, in S2, the differential equation of elastic wave propagation in the soil is taken as a physical constraint condition: wherein is a displacement vector, and is a Lame constant, is a soil density field distribution, is time, ∇ is a Hamiltonian operator for calculating spatial gradient and divergence. The gradient of the network output parameter with respect to the spatial coordinate is calculated through automatic differentiation technology, and then the wave velocity and vibration attenuation law are calculated. The physical constraint loss function is defined as the mean square error between the predicted vibration attenuation and the actual observed attenuation.

[0034] and The relationship with the elastic modulus and Poisson's ratio is: wherein E represents the elastic modulus of the soil, ν represents the Poisson's ratio of the soil.

[0035] In specific implementation, in S2, the network sets a learnable reference density parameter (1800 kg / m³) and a reference modulus parameter (30 MPa), to ensure that the output parameters are within a reasonable physical range. The Adam optimizer is used for network training, the initial learning rate is set to 0.001, and the training period is set to 3000 rounds. The total loss function is the weighted sum of the data fitting loss and the physical constraint loss: wherein is the observation data fitting loss, is the physical equation residual loss, is the boundary condition loss, is the initial condition loss, and the weight coefficient , , . is the total number of sample data points collected by the vibration sensor, is the true observation value of the vibration feature at the i-th position, is the predicted value of the corresponding vibration feature by the physical information neural network model, represents the residual error obtained by substituting the neural network predicted field u into the elastic wave governing equation of the soil, is the number of spatial points used when calculating the physical residual error.

[0036] S3: Based on the predicted soil physical parameters by PINNs, the three-dimensional distribution and development trend of foundation bearing capacity are calculated using the Terzaghi bearing capacity theory; In S3, the trained PINNs model can predict the density, modulus and Poisson's ratio distribution at any position in the soil. Based on these predicted parameters, the foundation bearing capacity is calculated using the Terzaghi bearing capacity theory: , is the foundation bearing capacity, with the unit of kPa, where is the cohesion, with the unit of kPa, which is calculated by the empirical relationship , is the reference value of cohesion, with the unit of kPa, is the proportional coefficient of cohesion with respect to density, with the unit of Pa / (kg / m 3 ), p is the actual density of the soil, with the unit of kg / m 3 , p 0 is the reference density of the soil, with the unit of kg / m 3 ; is the unit weight of the soil above the foundation, with the unit of kN / m 3 , is the unit weight of the soil below the foundation, with the unit of kN / m 3 ; is the foundation depth, with the unit of m, is the foundation width, with the unit of m; 、 、 is the bearing capacity coefficient, dimensionless, calculated based on the internal friction angle. The relationship between the internal friction angle and the soil density is: . is the reference value of the friction angle, with the unit of rad, is the proportional coefficient of the friction angle with respect to density, with the unit of rad.

[0037] S4: Establish a multi-objective optimization model to optimize the subsequent construction parameters with the goals of bearing capacity compliance, minimum energy, and shortest time; In implementation, in S4, a multi-objective optimization model is established with the goals of bearing capacity compliance, minimum energy consumption, and shortest construction time. wherein: is the penalty for insufficient bearing capacity, is the total energy consumption, is the total construction time. The decision variable is . Optimization constraints include: ramming frequency constraints (3-15 times), single ramming energy constraints (50-500 kJ), and intermittent time constraints (0.5-24 hours).

[0038] In implementation, in S4, the NSGA-II algorithm is used to solve the multi-objective optimization problem, with the following algorithm parameters: population size 100, maximum number of iterations 500, crossover probability 0.9, mutation probability 0.1, and selection strategy tournament selection. A cumulative simulation model of ramming effect is established to predict the cumulative effect of multiple ramming. The bearing capacity increment calculation formula is: wherein is the ramming energy, is the energy conversion efficiency. The consolidation effect coefficient calculation formula is: wherein is the intermittent time.

[0039] S5: Perform construction according to the optimized parameters and continuously update the PINNs model for adaptive control.

[0040] In implementation, in S5, an incremental learning-based model updating mechanism is established, and the vibration feature dataset is automatically updated after each ramming. The PINNs model parameters are updated using incremental learning, the hyperparameters are adjusted based on Bayesian optimization, the prediction uncertainty is evaluated in real time, and the control strategy is adjusted. The system has a learning and updating function: after completing the entire construction of a ramming point, the measured bearing capacity data is fed back to the PINNs model, and the prediction accuracy is continuously optimized.

[0041] The specific process includes: After each ramming operation, the vibration sensor array deployed in S1 collects new ground vibration signals, extracts vibration feature parameters, and adds these new data to the existing vibration feature dataset, achieving dynamic expansion of the dataset; Using incremental learning algorithm, the newly collected vibration feature data and corresponding spatial coordinates are used as new training samples to update the physical information neural network model constructed in S2, improving the model's adaptability to the current construction conditions without losing existing knowledge; Based on the Bayesian optimization framework, the network learning rate and other key hyperparameters are dynamically adjusted according to the model's performance on the new data, optimizing the training process and improving the model's convergence performance; The uncertainty of the physical information neural network model output is evaluated in real time by calculating the posterior distribution of the prediction result, and the conservatism of the control strategy is adjusted according to the degree of uncertainty; After the completion of the entire construction of each tamping point, the measured bearing capacity data is obtained through in-situ testing, and these measured data are fed back to the physical information neural network model as true value labels to further optimize the model parameters and continuously improve the accuracy of soil parameter prediction and bearing capacity calculation.

[0042] In specific implementation, the last step further includes a prediction result verification step: by comparing the soil state predicted by the PINNs with the in-situ test results such as standard penetration test, static sounding test, and load test, the model accuracy is verified. A prediction accuracy evaluation system is established to realize quantitative evaluation and quality control of the prediction results.

[0043] The specific process includes: Based on the soil physical property parameter prediction results and the foundation bearing capacity prediction results obtained from S2 and S3, the measured soil parameters and bearing capacity data at the corresponding positions are obtained through standard penetration test, static sounding test, and load test; The prediction results of the physical information neural network model are compared with the measured data obtained by various in-situ tests, the error between the predicted value and the measured value is calculated, and the model prediction accuracy is verified; A prediction accuracy evaluation system including various quantitative evaluation indexes is established to evaluate the accuracy of the prediction results of soil density, elastic modulus, Poisson's ratio, and bearing capacity; According to the accuracy evaluation results, the reliability of the physical information neural network model is classified, and the comparison results are fed back to the model training process of S2 to provide a basis for model parameter optimization; When the prediction accuracy does not meet the engineering requirements, the model retraining mechanism is triggered, the prediction performance is improved by increasing the training data or adjusting the network structure, and the closed-loop control of the dynamic compaction construction quality is realized.

[0044] Application Example 1 Foundation treatment project of an industrial park 1. Project overview This application example takes a 15000 square meter foundation treatment project of an industrial park as an example. The construction area is a 150m×100m rectangular site, and the design requires that the foundation bearing capacity reaches 200 kPa. The undisturbed soil is soft plastic silty clay, with a natural water content , liquid limit index , internal friction angle , initial bearing capacity Only 85 kPa. The ramming points are arranged in a square grid with a spacing of 6 m x 6 m, and a total of 420 ramming points are set. A 3000 kJ strong ramming machine is used for construction, with a rammer weight of 30 t and a maximum drop distance of 10 m.

[0045] 2. Sensor system layout A representative ramming point P-125 located in the center of the site is selected for detailed monitoring, with coordinates (75 m, 50 m). The sensor layout scheme is as follows: ) 4 measuring points, azimuth angles , , , ; the second ring ( ) 4 measuring points, azimuth angles ; the third ring ( ) 4 measuring points, azimuth angles . A total of 12 three-dimensional vibration sensors, model GS-20DX, frequency response range 4.5 Hz-400 Hz, sensitivity 28.8 V / (m / s). The installation technical requirements are as follows: each sensor is buried 0.6 m deep, a 10 cm thick fine sand cushion is laid at the bottom, a high-strength steel protective cover is used, the sealing level is IP68, and it can withstand 100g impact acceleration.

[0046] 3. Vibration data acquisition and feature extraction First ramming data (energy 3000 kJ): Measuring point T1( ) recorded data: peak amplitude mm / s; effective value mm / s; duration s; main frequency Hz; band energy J·s; spectral centroid Hz; Measuring point T5( ) recorded data: mm / s; mm / s; s; Hz; J·s; Hz; Measuring point T9( ) recorded data: mm / s; mm / s; s; Hz; J·s; Hz; Amplitude attenuation rate: Based on the wave peak arrival time difference calculation, P Wave velocity m / s, S Wave velocity m / s.

[0047] 4. PINNs model construction and training Input feature vector construction is ; network structure is 11→64→128→64→32→3; activation function is ReLU; optimizer is Adam, learning rate is set to 0.001; batch size is set to 32; training rounds are 3000.

[0048] The loss function weight settings are as follows: data fitting loss weight is 1.0; physical constraint loss weight is 1.0; boundary condition loss weight is 0.1; continuity loss weight is 0.1. The training process monitoring is as follows: at the 500th round: ; at the 1000th round: ; at the 1500th round: ; at the 2000th round: ; at the 3000th round: .

[0049] 5. Soil parameter prediction results Based on the trained PINNs model, the distribution of soil parameters around P-125 points is predicted: The prediction results at a depth of 1.0 m are: soil density: kg / m³; elastic modulus: MPa; Poisson's ratio: ; The prediction results at a depth of 2.0 m are: kg / m³; MPa; ; The prediction results at a depth of 3.0 m are: kg / m³; MPa; .

[0050] 6. Bearing capacity calculation Based on the Terzaghi bearing capacity theory, the bearing capacity at each depth is calculated: Internal friction angle calculation: ; Bearing capacity coefficient: , , ; Bearing capacity calculation (foundation width , buried depth ): ; Considering the safety factor of 2.5, the designed bearing capacity is 157.6 kPa, which is less than the target value of 200 kPa, and further ramming is needed.

[0051] 7. Multi-objective optimization calculation Optimization objectives: : insufficient bearing capacity penalty : minimum energy consumption; : minimum construction time. The NSGA-II algorithm parameters are set as follows: population size: 100; maximum number of iterations: 500; crossover probability: 0.9; mutation probability: 0.1. The optimization converges after 472 iterations, and the Pareto optimal solution set is obtained: Scheme A (balanced scheme): subsequent ramming times: 6 times; energy configuration: [2400, 2100, 1800, 1500, 1200, 900] kJ; intermittent time: [2, 4, 8, 12, 16, 24] hours; predicted final bearing capacity: 202.5 kPa; total energy consumption: 10800 kJ; total construction time: 66 hours; Scheme B (fast scheme): subsequent ramming times: 4 times; energy configuration: [2700, 2400, 2100, 1800] kJ; intermittent time: [1, 2, 4, 8] hours; predicted final bearing capacity: 201.2 kPa; total energy consumption: 9000 kJ; total construction time: 15 hours; Scheme C (energy-saving scheme): subsequent ramming times: 8 times; energy configuration: [2100, 1900, 1700, 1500, 1300, 1100, 900, 700] kJ; intermittent time: [3, 6, 9, 12, 15, 18, 21, 24] hours; predicted final bearing capacity: 203.8 kPa; total energy consumption: 11200 kJ; total construction time: 108 hours.

[0052] 8. Construction execution and verification Scheme A is selected for construction, and the actual execution results are as follows: Second ramming (energy 2400 kJ, intermittent 2 hours): actual ramming amount: 11.2 cm; PINNs predicted bearing capacity: 165.3 kPa; Third ramming (energy 2100 kJ, intermittent 4 hours): actual ramming amount: 8.7 cm; PINNs predicted bearing capacity: 178.9 kPa; add new samples and retrain.

[0053] 7th ramming (energy 900 kJ, interval 24 hours): measured ramming amount: 2.1 cm; PINNs predicted bearing capacity: 201.8 kPa.

[0054] Final verification test: Load test: bearing capacity characteristic value: kPa; PINNs predicted value: 201.8 kPa; prediction error: 1.6%.

[0055] The above description of the embodiments is to facilitate the understanding and use of the invention by those of ordinary skill in the art. Those skilled in the art can obviously make various modifications to these embodiments, and apply the general principles described herein to other embodiments without having to go through creative labor. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A method for intelligent monitoring and adaptive control of dynamic compaction construction based on physical information neural networks, characterized in that, Includes the following steps: S1. Deploy a vibration sensor array within a safe range of 20-50 meters from the impact point to collect ground vibration signals generated by the impact in real time, and extract the time domain characteristics, frequency domain characteristics, and spatial propagation characteristics of the vibration waveform; S2. Construct a physical information neural network model, using the soil elastic wave propagation equation as a physical constraint, and taking the vibration characteristics extracted in S1 and the spatial coordinates of the monitoring points as inputs to predict the spatial distribution of soil physical parameters. S3. Based on the spatial distribution of soil physical property parameters predicted by the physical information neural network model, the three-dimensional distribution and development trend of foundation bearing capacity are calculated using Terzaghi bearing capacity theory. S4. Based on the calculated three-dimensional distribution and development trend of foundation bearing capacity, establish a multi-objective optimization model with the goals of achieving bearing capacity standards, minimizing energy consumption, and minimizing construction time, and optimize subsequent construction parameters. S5. Perform dynamic compaction construction based on the optimized construction parameters, and continuously update the vibration characteristic data and physical information neural network model parameters during the construction process to achieve adaptive control of the construction process.

2. The intelligent monitoring and adaptive control method for dynamic compaction construction based on a physical information neural network according to claim 1, characterized in that, The time-domain features include peak amplitude. Valid value Duration ; The frequency domain features include the dominant frequency. Frequency band energy Spectrum centroid ; The spatial propagation characteristics include amplitude attenuation rate. Phase difference speed of transmission ,in: , These are the distances from the impact point. , The vibration amplitude at that location, , The unit is mm / s. , The unit is m. For dimensionless parameters, , These represent the sensors. i and j The radial distance between the impact point and the ramming point.

3. The intelligent monitoring and adaptive control method for dynamic compaction construction based on a physical information neural network according to claim 2, characterized in that, In S2, a physical information neural network model is constructed, using the soil elastic wave propagation equation as a physical constraint and the vibration characteristics extracted in S1 and the spatial coordinates of the monitoring points as input. The specific process of predicting the three-dimensional distribution of the soil density field, elastic modulus field, and Poisson's ratio field includes: The vibration feature vector extracted from S1 and the spatial coordinate vector of the monitoring point are combined into an input vector; A multi-layer feedforward neural network structure is constructed, which includes an input layer, multiple hidden layers, and an output layer. The hidden layers use the ReLU activation function for nonlinear transformation, and the output layer outputs three physical property parameters: soil density, elastic modulus, and Poisson's ratio. During network training, the soil elastic wave propagation equation is used as a physical constraint. The residual between the network output and the physical equation is calculated using automatic differentiation technology. This residual is then incorporated into the overall loss function as the physical equation residual loss. The overall loss function consists of a weighted sum of data fitting loss, physical equation residual loss, boundary condition loss, and initial condition loss. The data fitting loss ensures that the network output is consistent with the measured data, the physical equation residual loss ensures that the prediction results conform to the physical laws of wave propagation, and the boundary condition loss and initial condition loss ensure the uniqueness and stability of the solution. The overall loss function is minimized by a gradient-based optimization algorithm, and the network parameters are iteratively updated to finally obtain a trained physical information neural network model. This model can accurately predict the three-dimensional spatial distribution of soil physical parameters using vibration characteristics and spatial coordinates as input.

4. The intelligent monitoring and adaptive control method for dynamic compaction construction based on a physical information neural network according to claim 3, characterized in that, In S2, the input to the physical information neural network model is ,in: For vibration characteristic vectors, It is a spatial coordinate vector; The output of the physical information neural network model is ,in: Soil density field distribution Elastic modulus field distribution Poisson's ratio field distribution; The loss function of the physical information neural network model is: in: For data fitting loss; This represents the residual loss in the physical equations. For boundary condition loss; Loss due to initial conditions; , , These are the weighting coefficients. This represents the total number of sample data points collected by the vibration sensor. It is the actual observed value of the vibration characteristic at the i-th position. It is the predicted value of the corresponding vibration characteristics by the physical information neural network model. This represents the residual obtained by substituting the field quantity u predicted by the neural network into the soil elastic wave control equation. It is the number of spatial points used when calculating physical residuals.

5. The intelligent monitoring and adaptive control method for dynamic compaction construction based on a physical information neural network according to claim 3, characterized in that, In S2, the soil elastic wave propagation equation is a three-dimensional elastic wave propagation equation, and the soil elastic wave propagation equation is specifically as follows: in: It is a displacement vector. For soil density field distribution, For time, ∇ It is the Hamiltonian operator used to calculate spatial gradients and divergences. and Lamé constant, representing the elastic properties of soil, is related to the elastic modulus and Poisson's ratio as follows: in E This represents the elastic modulus of soil. ν It represents the Poisson's ratio of soil.

6. The intelligent monitoring and adaptive control method for dynamic compaction construction based on a physical information neural network according to claim 1, characterized in that, In S3, the specific process of calculating the three-dimensional distribution and development trend of foundation bearing capacity using Terzaghi's bearing capacity theory, based on the spatial distribution of soil physical property parameters predicted by the physical information neural network model, includes: Based on the three-dimensional distribution data of soil density field, elastic modulus field, and Poisson's ratio field predicted by the physical information neural network model described in S2, the Lamé constant of the soil is calculated according to the elastic modulus and Poisson's ratio. Based on the predicted soil density, the cohesion and internal friction angle of the soil are calculated using empirical relationships. The calculated cohesion, internal friction angle, soil weight, foundation depth and width are used as input parameters, and the three-dimensional spatial distribution of foundation bearing capacity is calculated using Terzaghi's bearing capacity theory formula. By analyzing the changes in the spatial distribution of bearing capacity under different construction stages, the development trend of foundation bearing capacity can be predicted.

7. The intelligent monitoring and adaptive control method for dynamic compaction construction based on a physical information neural network according to claim 6, characterized in that, In S3, the Terzaghi bearing capacity theoretical formula is: in: It refers to the bearing capacity of the foundation, measured in kPa. Cohesion, measured in kPa, is determined through empirical relationships. calculate, This is a reference value for cohesion, in kPa. It is the proportionality coefficient of cohesion as a function of density, with units of Pa / (kg / m³). 3 ), ρ This is the actual density of the soil, expressed in kg / m³. 3 , ρ 0 This is the reference density of the soil, in kg / m³. 3 ; The weight of soil above the foundation is expressed in kN / m³. 3 ; The weight of the soil below the foundation, in kN / m³. 3 ; The foundation burial depth is expressed in meters (m). This is the base width, in meters (m). , , The bearing capacity coefficient is dimensionless and calculated based on the internal friction angle. The relationship between the internal friction angle and soil density is as follows: , This is the reference value for the friction angle, measured in rad. It is the proportionality coefficient of the friction angle as a function of density, and its unit is rad.

8. The intelligent monitoring and adaptive control method for dynamic compaction construction based on a physical information neural network according to claim 1, characterized in that, In S4, based on the calculated three-dimensional distribution and development trend of the foundation bearing capacity, a multi-objective optimization model is established with the goals of achieving bearing capacity standards, minimizing energy consumption, and minimizing construction time. The specific process of optimizing subsequent construction parameters includes: Based on the predicted trend of foundation bearing capacity and the current spatial distribution of bearing capacity in S3, an optimization problem with multiple conflicting objective functions is constructed. The multi-objective optimization model includes: a function designed to penalize the failure of the predicted bearing capacity to reach the design target value; a function to calculate the total energy consumption of all subsequent compaction blows; and a function to calculate the total time of all subsequent compaction intervals. The number of subsequent tamping blows, the energy value of each tamping blow, and the interval time after each tamping blow were determined as decision variables, and constraints that conform to the scope of engineering practice were set for the decision variables. A multi-objective evolutionary algorithm is used to solve this optimization problem, and a set of Pareto optimal solutions that achieve the best balance among multiple objectives is generated through iterative calculation. Finally, an optimal solution is selected from the Pareto optimal solution set, and its corresponding subsequent construction parameter sequence is output, including the specific number of tamping blows, the energy configuration of each tamping blow, and the interval time arrangement.

9. The intelligent monitoring and adaptive control method for dynamic compaction construction based on a physical information neural network according to claim 1, characterized in that, In S4, the multi-objective optimization model is: in: As a penalty for insufficient load-bearing capacity, This is the preset target value for foundation bearing capacity. It is based on the current foundation bearing capacity value predicted by the model; Total energy consumption It is the energy value applied by the i-th impact; Total construction time. It is the interval time required after the i-th tamping is completed; As decision variables, The total number of tamping blows; Constraints include: range of tamping times single impact energy kJ; Intermittent time range h.

10. The intelligent monitoring and adaptive control method for dynamic compaction construction based on a physical information neural network according to claim 1, characterized in that, The specific process of implementing dynamic compaction based on optimized construction parameters, and continuously updating vibration characteristic data and physical information neural network model parameters during construction to achieve adaptive control of the construction process includes: The subsequent compaction operation is carried out according to the optimized construction parameter sequence output by S4, and new ground vibration signals are collected in real time through the vibration sensor array deployed by S1 during each compaction process. The newly acquired vibration signals are processed to extract new vibration feature parameters, and the vibration feature dataset in S1 is updated and expanded. An incremental learning approach is adopted, using newly acquired vibration feature data and spatial coordinates as training samples to update the parameters of the physical information neural network model constructed by S2 and optimize its prediction accuracy. Based on the updated physical information neural network model, the bearing capacity calculation and trend prediction of S3 are re-executed, and the subsequent construction parameters are dynamically adjusted through the multi-objective optimization model of S4. By collecting data in real time, continuously updating the model, and cyclically optimizing parameters, a closed-loop feedback control process is formed, enabling intelligent monitoring and adaptive control of the dynamic compaction construction process.

Images