Detection method of drilling rig for ground surface settlement based on artificial intelligence assistance

Abstract

The invention relates to the technical field of civil engineering, and discloses an artificial intelligence assistance-based drilling rig detection method for ground surface settlement, which systematically solves the disjunction problem between a geomechanical mechanism and data-driven prediction by deeply coupling drilling data acquisition, physical constraint embedding and an artificial intelligence model. Compared with the problem of lack of physical constraint embedding of a deep learning settlement prediction model constructed based on drilling data in the prior art, the method has the advantages that the time sequence requirement of the Biot consolidation theory is strictly followed in the data acquisition link, and the dynamic characteristics of the geomechanical process naturally carried by the original data are ensured. Besides, various physical equation constraint terms are embedded in a loss function, a complete reasoning chain is generated in combination with a layer-by-layer correlation propagation algorithm, and causal chain visualization from stratum parameter identification to settlement volume deduction is realized, so that the physical interpretability and engineering practicability of the model are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of civil engineering technology, specifically to a method for detecting surface subsidence using a drilling device based on artificial intelligence assistance. Background Technology

[0002] In the construction of underground engineering projects such as subway tunnels and deep foundation pits, surface settlement monitoring is a core technical means to prevent engineering accidents. With the expansion of the application of deep learning algorithms in the engineering field, settlement prediction technology based on artificial intelligence has gradually become a research direction. Drilling equipment, as the main equipment for geological structure detection, soil physical parameter acquisition, and groundwater condition investigation, directly affects the reliability of prediction models due to the quality of its data acquisition. However, the current drilling data processing technology is not sufficiently integrated with geomechanical theory, resulting in technical problems such as weak interpretability of physical mechanisms in artificial intelligence models.

[0003] In existing technologies, deep learning settlement prediction models built based on drilling data have the following shortcomings: First, although neural network models can achieve data fitting, the prediction process lacks a verification mechanism that correlates with basic geomechanical equations such as the Terzaghi effective stress principle and Biot consolidation theory, and some prediction results may violate physical constraints such as mass conservation or seepage direction; Second, the model output lacks a complete reasoning path from formation parameter identification to settlement calculation, making it difficult for engineers to determine whether the prediction results meet the relevant technical specifications; Third, the raw data collected by existing drilling equipment has not undergone structured processing based on physical constraints and cannot be directly applied to the physical information neural network architecture, limiting the depth of integration between data-driven methods and theoretical calculation methods.

[0004] The aforementioned technical deficiencies have the following impacts on engineering practice: Firstly, construction managers are unable to obtain the mechanical causes of the early warning results, affecting the efficiency of risk response measures; some projects have experienced deviations between predicted trends and actual monitoring data. Secondly, when settlement calculations conforming to engineering technical standards are required, the output of purely data-driven models is difficult to meet professional review requirements. Furthermore, if the model's predicted values ​​exceed the physical boundaries of soil compressibility, technicians must invest additional time in manual verification, reducing the application value of intelligent prediction technology. The root cause lies in the lack of a data organization method adapted to geomechanical equations during the drilling data acquisition phase, resulting in a lack of a data foundation for subsequent model construction with embedded physical constraints.

[0005] Therefore, we propose an AI-assisted drilling method for detecting surface subsidence to address the aforementioned problems. Summary of the Invention

[0006] The purpose of this invention is to provide an artificial intelligence-assisted method for detecting surface subsidence using drilling devices, in order to solve the problem mentioned in the background art where construction managers are unable to know the mechanical causes of early warning results, which affects the efficiency of risk response measures and where some engineering projects have experienced deviations between predicted trends and actual monitoring data.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting surface subsidence using a drilling device based on artificial intelligence assistance, the specific steps of which are as follows: S1. During the drilling process, the drilling device simultaneously collects parameters such as formation depth, soil density, water content, and penetration resistance. At the formation interface, it obtains the total stress and pore water pressure data required for effective stress calculation through pore water pressure sensors and permeability coefficient testing modules, and records the pressure dissipation curve according to the time intervals required by Biot consolidation theory. S2. Physical constraint labeling of the collected data, calculation of seepage velocity and labeling of seepage direction according to Darcy's law, labeling of stress-strain relationship according to Duncan-Chang constitutive model, labeling of soil stability level using Mohr-Coulomb criterion, and establishing mapping relationship between data labels and geomechanical equations; S3. Construct a physical information neural network architecture, embed the mass conservation equation, energy conservation equation and Biot consolidation equation constraint terms into the loss function, set physical constraint weight coefficients, so that network training can minimize data fitting error and physical equation residuals at the same time. S4. Input the ideal formation model with known analytical solution to verify the conformity between the predicted settlement curve and the Terzaghi consolidation theory, and check the consistency of the superposition of layered settlement and the physical correctness of the direction of the hydraulic head gradient of the confined aquifer. S5. The decision path is traced in reverse by using the layer-by-layer correlation propagation algorithm to extract key input features and intermediate physical parameters, and generate a complete reasoning chain including pore water pressure identification, consolidation degree calculation and settlement inference. S6. When new operating condition data is detected, analyze the physical causes of the prediction deviation, use Bayesian optimization to adjust the physical constraint weights or constitutive model parameters, retrain and verify the convergence of the equation residuals. S7. Based on the inference chain, automatically generate technical documents that meet the requirements of the specifications, including stratigraphic profiles, calculation process of the layered summation method, consolidation degree time history curves, and safety factor verification sheets.

[0008] Preferably, step S1 is performed in the following manner: S1.1 During the drilling process, data is collected by a sensor array integrated on the outer wall of the drill pipe. The sensor array includes a depth encoder, a gamma ray density probe, a capacitive moisture content sensor, and a dynamic penetration resistance meter. The depth encoder records the drilling depth h at a resolution of 1 mm. The density probe measures the soil density ρ and converts it to unit weight γ. The moisture content sensor measures the mass moisture content w. The penetration resistance meter records the standard penetration test blow count N. A set of data is collected every 10 cm ± 2 cm of drilling, with a collection time interval of no more than 5 seconds. Each set of data includes depth coordinates, soil physical and mechanical parameters, and a timestamp, forming a basic parameter dataset for the stratigraphic profile. S1.2. When a formation interface is detected, the test is initiated. The interface determination criteria are that the difference in N value between adjacent measuring points is greater than 5 blows or the density change rate exceeds 15%. The pore water pressure sensor penetrates the soil to a depth of not less than 20cm and is left to stand for 3 minutes before measuring the initial pore water pressure u0 with an accuracy of not less than 0.5kPa. Based on the data in S1-1, the total stress σ is calculated using the layered summation method to obtain a pair of dual-parameter data. The permeability coefficient testing module measures the permeability coefficient k. The pore water pressure u is recorded at time points of 0.5 minutes, 1 minute, 2 minutes, 4 minutes, 9 minutes, 16 minutes, 25 minutes, and 36 minutes. t The process continues until the dissipation rate reaches 90% or more, or the pressure change rate at three consecutive measuring points is less than 2%, or the monitoring lasts for 60 minutes, to form a time-series pressure dissipation curve dataset.

[0009] Preferably, step S2 is performed in the following manner: S2.1. Based on the data in S1, physical parameters are calculated. According to Darcy's law, the seepage velocity v is calculated using the permeability coefficient k and the hydraulic gradient i. The hydraulic gradient i is calculated by the ratio of the difference in pore water pressure between adjacent measuring points to the vertical distance. The unit weight of water γw is taken as 10 kN / m³. When i is greater than 0, it is marked as upward seepage; when i is less than 0, it is marked as downward seepage. According to the Duncan-Chang model, the tangent modulus Et is calculated using the modulus parameters K and n, and the stress-strain relationship type is marked. The estimated range of K is 100 to 400, and the range of n is 0.4 to 0.8. The safety factor Fs is calculated using the Mohr-Coulomb criterion using the cohesion c, the internal friction angle φ, and the effective stress σ'. When Fs is greater than or equal to 1.5, it is marked as high stability; when Fs is greater than or equal to 1.2 and less than 1.5, it is marked as medium stability; when Fs is less than 1.2, it is marked as low stability. S2.2 Establish a database mapping data labels to geomechanical equations. Map seepage parameters to the boundary conditions of the mass conservation equation, constraining the consistency of the seepage direction with the hydraulic gradient and the conservation of mass flux. Map stress-strain parameters to the Duncan-Chang incremental equation, constraining the stress path not to exceed the failure envelope. Map stability parameters to the Mohr-Coulomb yield criterion and force balance equation, constraining the yield function value to be less than or equal to zero and the resultant force to be zero. Generate a composite label vector for each set of data. The composite label vector contains the equation type code, parameter list, constraint expression, and penalty weight. The equation type code is 1 for mass conservation equation, 2 for constitutive relation equation, and 3 for force balance equation. The penalty weight ranges from 0.1 to 10, forming a structured training dataset containing original measurements, physical parameters, equation types, and constraint conditions.

[0010] Preferably, step S3 is implemented in the following manner: S3.1 Construct a multi-layer feedforward neural network. The input layer receives features from the S2 dataset, including seven parameters: formation depth h, soil weight γ, water content w, permeability coefficient k, pore water pressure u, effective stress σ', and time t. The hidden layer is a fully connected structure of 3 to 5 layers, with 50 to 200 neurons per layer, using the tanh activation function. The output layer includes total settlement s, settlement distribution s(z), and degree of consolidation U(t). Parallel physical constraint calculation branch modules are set up. The partial derivatives are calculated using automatic differentiation technology and substituted into the continuity equation to calculate the mass conservation residual R_conservation, where the water density ρw is taken as 1000 kg / m³. The energy residual R_energy is substituted into the energy conservation equation, and the consolidation residual R_consolidation is substituted into the Biot consolidation equation. Three types of residual values ​​are output. S3.2 Construct a composite loss function comprising a data fitting loss term Ldata and physical constraint loss terms Lconservation, Lenergy, and Lconsolidation. The data loss is calculated using the average of the squared differences between predicted and measured values, while the physical loss is calculated using the average of the squared residuals. The number of sampling points is 2 to 5 times that of the training samples. Initial weights are λconservation = 0.5, λenergy = 0.3, and λconsolidation = 1.0. When the mean residual Ri is greater than a threshold, the coefficients are adjusted according to the weight update formula, with an adjustment rate α ranging from 0.1 to 0.5. The thresholds are 0.01, 0.05, and 0.01, respectively. The Adam optimizer is used with a learning rate of 0.001 to 0.01, batch size of 64 to 256, and iterations of 5000 to 20000. Termination conditions include a loss change rate less than 10 for 50 consecutive epochs. -4 The residual R is conserved < 0.01, R energy < 0.05, and R consolidation < 0.01. The validation set RMSE is less than 10% of the measured mean.

[0011] Preferably, step S4 is performed in the following manner: S4.1 Construct a single-layer homogeneous saturated clay model with a soil layer thickness H of 10m, a compression modulus Es of 5MPa, and a permeability coefficient k of 1×10⁻⁶. -8 With a speed of m / s, Poisson's ratio μ of 0.35, and an applied load q of 100 kPa, the boundary condition is that the top surface is permeable and the bottom surface is impermeable. The parameters are input into the PINN model trained on S3 to obtain the predicted degree of consolidation time history curve. The theoretical analytical solution is calculated based on Terzaghi consolidation theory. The time factor is calculated using the consolidation coefficient, time, and drainage distance. The consolidation coefficient is calculated using the permeability coefficient, water unit weight, and volumetric compressibility coefficient. The water unit weight is taken as 10 kN / m³. 3 The relative error was less than 5% within the consolidation degree range of 0% to 90%. A three-layer foundation model was constructed, consisting of a 5m thick silty clay layer with a compression modulus of 8MPa, an 8m thick silty clay layer with a compression modulus of 3MPa, and a 6m thick silty sand layer with a compression modulus of 15MPa. The settlement of each layer and the total settlement were predicted, and the superposition relative deviation was verified to be less than 3%. S4.2 Construct a model containing a confined aquifer, with an upper impermeable clay layer 5m thick and a permeability coefficient of 1×10⁻⁶. -9 m / s, with a pressure sand layer thickness of 10m and a permeability coefficient of 1×10⁻⁶. -4 m / s, bottom layer is impermeable bedrock, pore water pressure of top plate is 150 kPa and bottom plate is 200 kPa. Input PINN to predict the internal pressure distribution and seepage velocity distribution of the confined layer. Calculate the hydraulic gradient based on the predicted pressure distribution. The hydraulic gradient is calculated by dividing the partial derivative of pore pressure with depth by the unit weight of water, where the unit weight of water is taken as 10 kN / m. 3 The experiment verified that a hydraulic gradient greater than 0 indicates upward seepage. The theoretical seepage velocity was calculated using Darcy's law by multiplying the permeability coefficient by the hydraulic gradient. The relative error between the predicted velocity and the theoretical value was less than 10%. The dot product of the velocity vector and the head gradient vector was verified to be greater than 0. A control volume was selected to calculate the inflow and outflow mass fluxes. The mass fluxes were calculated by multiplying the water density, flow velocity, and cross-sectional area. The water density was taken as 1000 kg / m³. 3 The relative deviation in flux is less than 2%.

[0012] Preferably, step S5 is performed in the following manner: S5.1. The PINN model trained by S3 is back-analyzed using a layer-by-layer correlation propagation algorithm. The correlation score is propagated back from the output layer to the input layer. The initial correlation score in the output layer is set to be equal to the predicted settlement value. The correlation score in each hidden layer is allocated to the node of the previous layer according to the connection weight ratio. The contribution value of each input feature is calculated. The top five key input features with the highest contribution are identified, including formation depth h, pore water pressure u, permeability coefficient k, effective stress σ', and water content w. The physical parameters activated in the intermediate hidden layers are extracted, including the degree of consolidation U, soil compression Δs, and seepage velocity v. The physical parameters are obtained by substituting the activated values ​​of the hidden layers into the constraint terms of the S3 physical equation and solving in reverse. The numerical range, activation intensity, and influence weight of each physical parameter are recorded. S5.2. Based on the key features extracted in S5.1 and the physical parameters of the intermediate layer, a reasoning chain is constructed according to the geomechanical causal relationship. The first step is pore water pressure identification. When the ratio of pore pressure u to hydrostatic pressure is greater than 1.2, it is identified as a confined aquifer. The water unit weight γw is taken as 10 kN / m³. The second step is consolidation degree calculation. The consolidation degree U is calculated based on the effective stress change Δσ' and the pore pressure dissipation rate combined with the residual value of the Biot consolidation equation. The consolidation degree is obtained by dividing the difference between the initial pore pressure and the current pore pressure by the initial pore pressure. When U is greater than 60%, it is identified as the main stage of consolidation. The third step is settlement estimation. The final settlement s is estimated based on the consolidation degree U, soil layer thickness h, and volume compressibility coefficient mv combined with the tangent modulus Et. When s exceeds the design allowable value, it is identified as a high-risk area. The thickness is 30 mm for subway tunnels and 50 mm for building foundations. A reasoning chain report containing physical parameter identification results, calculation process, and risk conclusions is generated. The source of parameters and theoretical basis are marked in the report.

[0013] Preferably, step S6 is performed in the following manner: S6.1 When new working condition data is detected, it is input into the PINN model trained in S3 to obtain the predicted settlement s. The actual settlement s is collected and measured. The prediction deviation Δs is calculated. When the absolute value of the deviation is greater than 15% of the measured value, the physical cause analysis is initiated. The analysis includes three aspects. First, the soil parameters of the new working condition are compared with the statistical range of the S2 training set to identify parameters that exceed the range and the degree of deviation is recorded. Second, the residuals R conservation, R energy, and R consolidation of the three types of physical equations of the new working condition are calculated. The convergence criteria of 0.01, 0.05, and 0.01 are compared to identify equations that exceed the standard and the excess multiple is recorded. Third, the stress-strain relationship data is extracted and the deviation from the theoretical curve of the Duncan-Chang model is calculated. When the deviation is greater than 20% or the stress ratio is greater than 0.85, it is determined that the constitutive parameters need to be corrected and the strategy for adjusting the physical constraint weights or constitutive model parameters is determined. S6.2. Based on the analysis results of S6-1, Bayesian optimization method is used to adjust the parameters. The objective function is to minimize the weighted sum of prediction bias and physical residual. The optimization variables include the weight coefficient λ conservation, λ energy, λ consolidation, and constitutive parameters K, n, and Rf. The search range is weight 0.1 to 10, K 50 to 500, n 0.2 to 1.0, and Rf 0.7 to 0.95. A probabilistic surrogate model is constructed using a Gaussian process. The next evaluation point is selected by the expected improvement EI criterion. Iterative execution is performed to evaluate the objective function value, update the surrogate model, calculate the EI value, and select the parameter combination with the maximum EI. The termination condition is that the improvement amount is less than 0.01 for 5 consecutive times, or after 50 iterations, or the objective function is less than 0.05. After obtaining the optimal parameters, the model is retrained using the S3 architecture. The training set is the original data from S2 combined with the new working condition data. The verification is that the mean of the three types of residuals is less than 0.01, 0.05, and 0.01, and the test relative error is less than 10%. After the conditions are met, the parameters of the S3 model are updated.

[0014] Preferably, step S7 is implemented in the following manner: S7.1. Based on the S5 inference chain report, extract technical parameters to generate calculation documents. The calculation documents include a stratigraphic profile, a settlement calculation report using the layered summation method, a consolidation degree time history curve, and a safety factor verification report. The stratigraphic profile extracts the depth of the S1-1 stratum and the soil type of S2-1 to draw a vertical profile and label six types of parameters, including thickness h, unit weight γ, compression modulus Es, permeability coefficient k, internal friction angle φ, and cohesion c. The layered summation method calculation report is prepared according to GB50007 format and includes four parts: self-weight stress, additional stress, compression amount, and total settlement. The consolidation degree curve extracts the data from the nine time nodes predicted by S3 to draw a Ut relationship curve and overlay the Terzaghi theoretical curve, labeling the t90 time point. The safety factor verification report is prepared according to GB50007 and GB50157 to calculate the anti-sliding safety factor Fssliding and the anti-uplift safety factor Fsuplift, and to determine whether Fssliding is greater than or equal to 1.2 and whether Fsuplift is greater than or equal to 1.5. S7.2. Conduct a compliance review of the technical documents generated in S7-1. The review includes three aspects: 1) Compliance review of the calculation method, verifying whether the layered summation method complies with the standard; 2) Checking the influence of groundwater level on the calculated buoyant unit weight γ' (taking 10 kN / m³); 3) Checking the stress diffusion angle range of 20 to 30 degrees; 4) Checking the layered calculation when the difference in compression modulus Es is greater than 2 times; 5) Reviewing the rationality of parameter values, verifying the empirical range of Es for cohesive soil (3 to 15 MPa, φ for 10 to 30 degrees) and Es for sandy soil (10 to 30 MPa, φ for 25 to 40 degrees), verifying that the parameter consistency deviation is less than 5%; 6) Checking that Fs slip is not less than 1.2 and Fs uplift is not less than 1.5; 7) Compliance review of the conclusion statement, verifying the final settlement amount, settlement rate, and settlement stabilization time t90. After the review is passed, output the technical documents in PDF and Word formats, along with the S5-2 inference chain report, indicating the generation time, project name, coordinates, signature, and settlement calculation safety verification conclusion according to the project overview and geological conditions.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. By deeply coupling drilling data acquisition, physical constraint embedding, and artificial intelligence models, this application systematically solves the disconnect between geomechanical mechanisms and data-driven prediction. Compared to existing technologies where deep learning settlement prediction models based on drilling data lack physical constraint embedding, this application ensures that the raw data naturally carries the dynamic characteristics of the geomechanical process by strictly adhering to the time series requirements of Biot's consolidation theory during the data acquisition stage. Furthermore, by embedding multiple physical equation constraint terms into the loss function and combining them with a layer-by-layer correlation propagation algorithm to generate a complete inference chain, the causal chain from formation parameter identification to settlement extrapolation is visualized, thereby significantly improving the model's physical interpretability and engineering applicability.

[0016] 2. By quantifying the difference in N-values ​​between adjacent measuring points to be greater than 5 blows or the density change rate to exceed 15% as the criteria for determining the formation interface, objective and accurate formation interface identification was achieved. The pore water pressure sensor was penetrated to a depth of not less than 20 cm and allowed to stand for 3 minutes to ensure a stable measurement environment and improve the measurement accuracy of the initial pore water pressure u0. The permeability coefficient testing module recorded the pore water pressure u at specific time points. t These time points strictly match the timescale requirements of Biot's consolidation theory, ensuring the physical validity of the pressure dissipation curves. The resulting time-series pressure dissipation curve dataset allows the original data to directly meet the requirements for verifying the physical equations, ensuring the feasibility of embedding physical constraints in subsequent neural network training. Attached Figure Description

[0017] Figure 1 This is a diagram illustrating the method steps of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Please refer to Figure 1 A method for detecting surface subsidence using a drilling device based on artificial intelligence assistance, the specific steps of which are as follows: S1. During the drilling process, the drilling device simultaneously collects parameters such as formation depth, soil density, water content, and penetration resistance. At the formation interface, it obtains the total stress and pore water pressure data required for effective stress calculation through pore water pressure sensors and permeability coefficient testing modules, and records the pressure dissipation curve according to the time intervals required by Biot consolidation theory. S2. Physical constraint labeling of the collected data, calculation of seepage velocity and labeling of seepage direction according to Darcy's law, labeling of stress-strain relationship according to Duncan-Chang constitutive model, labeling of soil stability level using Mohr-Coulomb criterion, and establishing mapping relationship between data labels and geomechanical equations; S3. Construct a physical information neural network architecture, embed the mass conservation equation, energy conservation equation and Biot consolidation equation constraint terms into the loss function, set physical constraint weight coefficients, so that network training can minimize data fitting error and physical equation residuals at the same time. S4. Input the ideal formation model with known analytical solution to verify the conformity between the predicted settlement curve and the Terzaghi consolidation theory, and check the consistency of the superposition of layered settlement and the physical correctness of the direction of the hydraulic head gradient of the confined aquifer. S5. The decision path is traced in reverse by using the layer-by-layer correlation propagation algorithm to extract key input features and intermediate physical parameters, and generate a complete reasoning chain including pore water pressure identification, consolidation degree calculation and settlement inference. S6. When new operating condition data is detected, analyze the physical causes of the prediction deviation, use Bayesian optimization to adjust the physical constraint weights or constitutive model parameters, retrain and verify the convergence of the equation residuals. S7. Based on the inference chain, automatically generate technical documents that meet the requirements of the specifications, including stratigraphic profiles, calculation process of the layered summation method, consolidation degree time history curves, and safety factor verification sheets.

[0020] In this embodiment, the drilling device achieves simultaneous acquisition of multiple parameters during the drilling process through a multi-sensor array integrated on the drill rod. Specifically, an ultrasonic ranging sensor is used to monitor the drilling depth in real time, a gamma density sensor is used to measure the soil density, a resistive sensor is used to obtain the soil moisture content, and the penetration resistance is measured by a strain gauge force sensor installed near the drill bit. Furthermore, pore water pressure sensors are deployed at different depths in the borehole, automatically recording the dissipation process of pore water pressure at preset time intervals (e.g., every 5 minutes). This time interval is determined based on the calculation requirements of the consolidation time factor in Biot's consolidation theory, thereby ensuring that the collected time-series data accurately reflects the consolidation characteristics of the soil. In addition, the permeability coefficients of soil layers at different depths are obtained through field or laboratory permeability tests, providing necessary soil parameters for subsequent physical constraint labeling.

[0021] The collected raw data needs to undergo physical constraint annotation processing to transform discrete measurements into structured labels that conform to geomechanical principles. First, based on Darcy's law, the seepage velocity is calculated using the ratio of the pore water pressure difference between adjacent measuring points to the vertical distance, and the seepage direction is marked according to the pressure gradient direction to ensure the seepage field satisfies the continuity equation. Second, combined with stress path test data, the stress-strain relationship is established using the Duncan-Chang hyperbolic model or other applicable constitutive models (such as the modified Cambridge model), parameterizing the measured load-deformation curves into model parameters and annotating them as structured features. Third, based on the Mohr-Coulomb failure criterion or other applicable strength criteria (such as the Tresca criterion), the safety factor for each soil layer is calculated, and the soil stability is classified into stable, critically stable, and unstable levels for annotation. Through the above annotation process, an explicit mapping relationship is established between the measured data and the geomechanical equations, allowing the data labels to naturally carry physical constraint information.

[0022] Based on the labeled dataset, a neural network architecture incorporating physical constraints is constructed. The loss function of this network consists of a data fitting term and a physical constraint term. The data fitting term uses mean squared error to measure the deviation between the predicted and measured values, while the physical constraint term includes the following: a residual term from the mass conservation equation, used to constrain volume changes during soil deformation; a residual term from the energy conservation equation, ensuring system energy balance; and a residual term from the Biot consolidation equation, forcing the model output to satisfy the coupling relationship between pore water pressure dissipation and effective stress growth. Each physical constraint term is weighted and summed using adjustable weight coefficients. These weight coefficients can be fixed values ​​(e.g., empirically set between 0.1 and 1.0) or dynamically adjusted using a strategy (e.g., adaptively adjusted based on the relative magnitude of the residuals of each constraint term). The network training process employs the Adam optimizer or other suitable gradient descent algorithms to simultaneously minimize both the data fitting error and the physical equation residuals, thereby ensuring that the prediction results conform to both the characteristics of the measured data and the fundamental laws of geomechanics.

[0023] To verify the physical rationality of the model, the following verification methods were adopted: First, an ideal geological model with analytical solutions (such as a one-dimensional consolidation problem of a single-layer homogeneous soil) was input, and the settlement time history curve predicted by the model was compared with the analytical solution of Terzaghi's one-dimensional consolidation theory to check the degree of agreement between the two. Second, for multi-layer geological models, it was verified whether the superposition result of the settlement of each soil layer was consistent with the calculation result of the layered summation method. Third, it was checked whether the direction of the hydraulic head gradient of the confined aquifer was consistent with the actual groundwater flow field to eliminate abnormal predictions that violate physical laws. Through the above verification steps, possible physical inconsistencies in the model can be effectively identified and corrected.

[0024] The trained neural network is backpropagated using either Layer-wise Relevance Propagation (LRP) or SHAP (SHapley Additive exPlanations) algorithms to trace the decision path from input features to the final prediction result. Specifically, the contribution of each input parameter (such as soil layer thickness, compression modulus, and initial void ratio) to the prediction result, as well as the physical quantities characterized by the intermediate hidden layers (such as effective stress and degree of consolidation), are extracted to generate a complete inference chain. This chain includes key steps such as pore water pressure identification, degree of consolidation calculation, and settlement extrapolation for each layer. By visualizing the inference chain, technicians can trace the mechanical basis of each prediction result and clarify the role mechanism of each physical parameter in the settlement evolution process.

[0025] When the model is applied to new working conditions, if the predicted results deviate from the measured data, the physical causes of the deviation are analyzed (such as the compressibility parameter of a certain soil layer deviating from the actual value, or the permeability coefficient being estimated inaccurately). Bayesian optimization or grid search methods are then used to adjust the corresponding physical constraint weight coefficients or constitutive model parameters. The network is retrained, and the convergence of the physical equation residuals is verified. This dynamic optimization mechanism enables the model to adaptively adjust according to actual working conditions, maintaining predictive accuracy and physical rationality under complex geological conditions.

[0026] Based on the inference chain and prediction results output by the model, the system automatically generates technical documents that comply with current engineering technical specifications. These documents include: a stratigraphic profile, indicating the thickness, physical and mechanical parameters, and groundwater level information of each soil layer; the calculation process using the layered summation method, detailing the formulas and numerical results for calculating the compression of each soil layer; a consolidation time-history curve, showing the consolidation process of the soil at different times; and a safety factor verification report, calculating the foundation bearing capacity and stability safety factor according to relevant specifications. The automatically generated documents are formatted correctly and are complete in content, allowing direct use for engineering design review and acceptance, thus solving the problem that purely data-driven model outputs often fail to meet professional review requirements.

[0027] This technical solution, by strictly adhering to the temporal requirements of geomechanics theory during data acquisition, embedding physical law constraints during data annotation, integrating multiple physical equations during model construction, conducting theoretical analysis and comparison during verification, and visualizing the inference path and optimizing dynamic parameters during application, forms a complete physical constraint system from data acquisition to engineering document output. Compared to existing deep learning settlement prediction models based on drilling data, this solution systematically solves the following problems: First, through physical constraint annotation and loss function design, it ensures that the model prediction results naturally satisfy basic physical laws such as mass conservation, energy conservation, and consolidation theory, avoiding the risk that purely data-driven models may output results that violate physical laws; second, through inference chain extraction and visualization, it achieves transparency of the causal relationship from formation parameter identification to settlement calculation, significantly improving the model's interpretability; third, through theoretical verification and dynamic optimization mechanisms, it enhances the model's adaptability and prediction reliability under complex geological conditions; and fourth, by automatically generating standardized engineering documents, it improves the engineering practicality of the technical solution and meets the stringent requirements of professional review.

[0028] Example 2: Please refer to Figure 1 The specific method for step S1 is as follows: S1.1 During the drilling process, data is collected by a sensor array integrated on the outer wall of the drill pipe. The sensor array includes a depth encoder, a gamma ray density probe, a capacitive moisture content sensor, and a dynamic penetration resistance meter. The depth encoder records the drilling depth h at a resolution of 1 mm. The density probe measures the soil density ρ and converts it to unit weight γ. The moisture content sensor measures the mass moisture content w. The penetration resistance meter records the standard penetration test blow count N. A set of data is collected every 10 cm ± 2 cm of drilling, with a collection time interval of no more than 5 seconds. Each set of data includes depth coordinates, soil physical and mechanical parameters, and a timestamp, forming a basic parameter dataset for the stratigraphic profile. S1.2. When a formation interface is detected, the test is initiated. The interface determination criteria are that the difference in N value between adjacent measuring points is greater than 5 blows or the density change rate exceeds 15%. The pore water pressure sensor penetrates the soil to a depth of not less than 20cm and is left to stand for 3 minutes before measuring the initial pore water pressure u0 with an accuracy of not less than 0.5kPa. Based on the data in S1-1, the total stress σ is calculated using the layered summation method to obtain a pair of dual-parameter data. The permeability coefficient testing module measures the permeability coefficient k. The pore water pressure u is recorded at time points of 0.5 minutes, 1 minute, 2 minutes, 4 minutes, 9 minutes, 16 minutes, 25 minutes, and 36 minutes. t The process continues until the dissipation rate reaches 90% or more, or the pressure change rate at three consecutive measuring points is less than 2%, or the monitoring lasts for 60 minutes, to form a time-series pressure dissipation curve dataset.

[0029] In this embodiment, the sensor array integrates and installs various sensors with different functions on the outer wall of the drill pipe to achieve synchronous measurement of formation parameters during drilling. Specifically, the sensor array includes, but is not limited to, the following devices: a depth encoder, a gamma-ray density probe, a capacitive water content sensor, and a dynamic penetration resistance gauge.

[0030] The depth encoder uses photoelectric or magnetic encoding to record the drill pipe's advance depth in real time with a resolution of 1mm, providing a high-precision spatial positioning reference for strata stratification and ensuring accurate correspondence between the data at each measuring point and the actual stratum depth. The gamma-ray density probe utilizes the scattering characteristics of gamma rays in soil to determine the soil's wet density and converts it to the soil's natural unit weight based on the moisture content data, directly obtaining the basic physical parameters required for settlement calculations. The capacitive moisture content sensor determines the soil's mass moisture content by measuring changes in the soil's dielectric constant, accurately reflecting the soil's moisture state and providing a basis for saturation calculations and effective stress analysis. The dynamic penetration resistance gauge records the penetration resistance or equivalent standard penetration test blow count (N value) during drilling, used to assess the soil's mechanical strength and deformation characteristics.

[0031] To ensure the continuity and temporal consistency of the collected data, the drilling rig collects data according to the following rules: a complete set of multi-parameter data is collected every 10cm of drill pipe advance (allowable deviation ±2cm), and the time interval between two adjacent sets of data collection is no more than 5 seconds. This dual spatial and temporal constraint mechanism effectively avoids data sparsity or redundancy caused by uneven collection intervals, ensuring the continuity and uniformity of the data sequence.

[0032] The determination of stratigraphic interfaces adopts a quantitative standard: when the difference in the standard penetration test blow count (N value) between adjacent measuring points is greater than 5 blows, or the soil density change rate exceeds 15%, a stratigraphic interface is determined to exist at that location. This quantitative determination condition is based on the significant differences in the physical and mechanical properties of the soil layers, achieving objective and accurate identification of stratigraphic interfaces and avoiding the subjective bias that may arise from traditional judgments based on experience.

[0033] The installation and measurement of the pore water pressure sensor should follow these operating procedures: the sensor probe should penetrate the soil to a depth of not less than 20 cm. After penetration, allow it to stand for 3 minutes until the pore water pressure reading stabilizes, then record the initial pore water pressure. This set time is intended to eliminate the influence of penetration disturbance on the measurement results, ensuring that the measurement accuracy of the initial pore water pressure reaches ±0.5 kPa, thus providing reliable initial conditions for subsequent effective stress calculation.

[0034] The permeability coefficient testing module records the dissipation process of pore water pressure over time according to a preset time series. Specifically, the selection of time points is determined based on the time factor in Biot's consolidation theory. Typical time series include the first time point at 15 seconds, the second at 30 seconds, the third at 1 minute, the fourth at 2 minutes, the fifth at 5 minutes, the sixth at 10 minutes, and the seventh at 30 minutes, etc. The time interval can be adjusted appropriately according to the different permeability of the soil layer. By collecting the pore water pressure values ​​at each moment through the above time series, a complete pressure dissipation curve is formed. The time scale of this curve strictly matches the time dimension requirements of Biot's consolidation theory, ensuring the physical validity of the data.

[0035] The consolidation coefficient of a soil layer can be calculated using the pressure dissipation curve. Specifically, the consolidation coefficient equals the product of the time factor (theoretically approximately 0.197) at which the degree of consolidation reaches 50% and the square of the drainage distance, divided by the time required for the measured pore water pressure to dissipate to 50% of the initial excess pore water pressure. Furthermore, by combining this with the soil layer's compressibility coefficient, the permeability coefficient can be calculated, which is equal to the product of the consolidation coefficient, compressibility coefficient, and the unit weight of water. Through this calculation process, the field-measured pressure dissipation curve is transformed into key physical parameters such as the soil layer's consolidation coefficient and permeability coefficient, providing a quantitative basis for subsequent physical constraint labeling.

[0036] The aforementioned multi-parameter synchronous acquisition scheme solves the problems of data temporal misalignment and unclear spatial correspondence caused by traditional discrete measurement methods. The high-precision depth benchmark provided by the depth encoder ensures accurate matching between the data from each measuring point and the actual stratum depth; the combined use of the density probe and water content sensor enables simultaneous determination of soil weight and water content, avoiding parameter mismatch caused by separate measurements; and the N-value data obtained by the penetration resistance gauge provides a quantitative basis for assessing the mechanical strength of the soil layer.

[0037] By collecting data at 10cm intervals and strictly controlling the time intervals, the continuity and uniformity of the data sequence are guaranteed through dual spatial and temporal constraints, thus solving the data quality problem caused by uneven collection intervals. The quantitative determination criteria for stratigraphic boundaries enable objective and accurate stratigraphic division, laying the foundation for establishing precise stratigraphic profile models.

[0038] The standardized operating procedure for pore water pressure sensors, including penetration depth control, settling time, and time series design, ensures high-precision measurement of initial pore water pressure and the pressure dissipation process. Time nodes set according to Biot's consolidation theory ensure that the acquired pressure dissipation curve data naturally meet the time scale requirements of consolidation theory, guaranteeing the physical validity of the data. The resulting time-series dataset contains complete information on formation depth, soil physical properties, mechanical strength parameters, and consolidation characteristics. This data directly meets the requirements for verifying physical equations and provides a reliable data foundation for embedding physical constraints in subsequent training of the physical information neural network.

[0039] Compared to existing technologies that lack clear data acquisition standards, have incomplete measurement parameters, and suffer from discontinuous time-series data, this solution significantly improves the integrity, accuracy, and physical consistency of drilling data through systematic sensor configuration, standardized acquisition rules, and strict quality control measures. This provides solid data support for building a highly reliable foundation settlement prediction model and enhances the model's physical interpretability and engineering applicability.

[0040] Example 3: Please refer to Figure 1 The specific method for step S2 is as follows: S2.1. Based on the data in S1, physical parameters are calculated. According to Darcy's law, the seepage velocity v is calculated using the permeability coefficient k and the hydraulic gradient i. The hydraulic gradient i is calculated by the ratio of the difference in pore water pressure between adjacent measuring points to the vertical distance. The unit weight of water γw is taken as 10 kN / m³. When i is greater than 0, it is marked as upward seepage; when i is less than 0, it is marked as downward seepage. According to the Duncan-Chang model, the tangent modulus Et is calculated using the modulus parameters K and n, and the stress-strain relationship type is marked. The estimated range of K is 100 to 400, and the range of n is 0.4 to 0.8. The safety factor Fs is calculated using the Mohr-Coulomb criterion using the cohesion c, the internal friction angle φ, and the effective stress σ'. When Fs is greater than or equal to 1.5, it is marked as high stability; when Fs is greater than or equal to 1.2 and less than 1.5, it is marked as medium stability; when Fs is less than 1.2, it is marked as low stability. S2.2 Establish a database mapping data labels to geomechanical equations. Map seepage parameters to the boundary conditions of the mass conservation equation, constraining the consistency of the seepage direction with the hydraulic gradient and the conservation of mass flux. Map stress-strain parameters to the Duncan-Chang incremental equation, constraining the stress path not to exceed the failure envelope. Map stability parameters to the Mohr-Coulomb yield criterion and force balance equation, constraining the yield function value to be less than or equal to zero and the resultant force to be zero. Generate a composite label vector for each set of data. The composite label vector contains the equation type code, parameter list, constraint expression, and penalty weight. The equation type code is 1 for mass conservation equation, 2 for constitutive relation equation, and 3 for force balance equation. The penalty weight ranges from 0.1 to 10, forming a structured training dataset containing original measurements, physical parameters, equation types, and constraint conditions.

[0041] In this embodiment, physical constraint annotation transforms the collected discrete measurement data into structured labels constrained by geomechanical laws. First, seepage parameters are labeled based on Darcy's law. The hydraulic gradient is calculated using the ratio of the pore water pressure difference between adjacent measuring points to the vertical distance. The seepage velocity is equal to the product of the permeability coefficient and the hydraulic gradient, and the seepage direction is determined by the sign of the hydraulic gradient. This annotation method ensures that the seepage velocity vector and the hydraulic gradient vector are always aligned, avoiding errors in neural network learning that violate mass conservation. Second, stress-strain relationships are labeled based on the Duncan-Chang hyperbolic constitutive model. The modulus parameter K ranges from 100 to 400, and the exponent parameter n ranges from 0.4 to 0.8. By limiting the parameter range, the tangent modulus is ensured to always be positive and decrease with increasing stress, keeping the labeled stress-strain relationship within the elastic deformation range of the soil and effectively preventing the stress path from exceeding the failure envelope. Furthermore, the Mohr-Coulomb failure criterion is used to label the stability. The safety factor is equal to the shear strength divided by the actual shear stress and is converted into a discrete level: a safety factor greater than or equal to 1.5 is labeled as high stability, between 1.2 and 1.5 is labeled as medium stability, and less than 1.2 is labeled as low stability, so that the model can clearly identify the critical boundary of the soil stability state.

[0042] To establish an explicit mapping between data labels and geomechanical equations, this scheme designs a composite label vector structure. This vector is a multi-dimensional array containing three dimensions: basic physical parameters, equation type codes, and penalty weights. The basic physical parameter dimension stores values ​​such as seepage velocity, seepage direction, tangent modulus, and safety factor after physical constraint annotation. The equation type code dimension uses integer identifiers: code 1 represents the mass conservation equation, constraining the continuity of fluid mass during seepage; code 2 represents the constitutive relation equation, constraining the stress-strain relationship to conform to the Duncan-Chang model; and code 3 represents the force balance equation, constraining the stress distribution within the soil to satisfy static equilibrium. The penalty weight dimension stores the weight coefficients of each physical equation constraint term in the loss function, ranging from 0.1 to 10, differentiated according to the importance of different physical equations. For example, in soft soil foundation settlement prediction, the penalty weight of the Biot consolidation equation is set to 5 to 10 to strengthen the constraint of the consolidation process, while the penalty weight of the mass conservation equation in sandy soil foundations is appropriately increased to ensure rationality in seepage scenarios.

[0043] Each sample in the generated training dataset contains input features and a composite label vector. The input features are the original formation parameters, and the composite label vector is the annotation result after physical constraint processing, along with its associated equation type and weight information. This structured dataset enables the training of the physical information neural network to achieve accurate embedding of physical constraints: when constructing the loss function, the system automatically matches the corresponding physical equation constraint expressions according to the equation type encoding, calculates the residuals of the mass conservation equation, constitutive relation equation, and force balance equation separately, and adds them to the loss function; by using penalty weights to weightedly sum the residuals of various physical equations, the constraint strength is dynamically adjusted. During training, the neural network learns to simultaneously satisfy multiple physical equation constraints during prediction, ensuring that the predicted seepage direction is consistent with the hydraulic gradient, the stress-strain curve does not exceed the failure envelope, and the stratified settlement conforms to the principle of mass conservation. The above technical solutions systematically solve the problems of traditional data annotation lacking physical constraints, which leads to neural networks learning non-physical mapping relationships and prediction results that may violate the basic laws of geomechanics. They fundamentally ensure the physical rationality and engineering reliability of the model prediction results, realize the deep integration of data-driven methods and geomechanical theory, and significantly improve the prediction accuracy and engineering application value of the model under complex geological conditions.

[0044] Example 4: Please refer to Figure 1 The specific method for step S3 is as follows: S3.1 Construct a multi-layer feedforward neural network. The input layer receives features from the S2 dataset, including seven parameters: formation depth h, soil weight γ, water content w, permeability coefficient k, pore water pressure u, effective stress σ', and time t. The hidden layer is a fully connected structure of 3 to 5 layers, with 50 to 200 neurons per layer, using the tanh activation function. The output layer includes total settlement s, settlement distribution s(z), and degree of consolidation U(t). Parallel physical constraint calculation branch modules are set up. The partial derivatives are calculated using automatic differentiation technology and substituted into the continuity equation to calculate the mass conservation residual R_conservation, where the water density ρw is taken as 1000 kg / m³. The energy residual R_energy is substituted into the energy conservation equation, and the consolidation residual R_consolidation is substituted into the Biot consolidation equation. Three types of residual values ​​are output. S3.2 Construct a composite loss function comprising a data fitting loss term Ldata and physical constraint loss terms Lconservation, Lenergy, and Lconsolidation. The data loss is calculated using the average of the squared differences between predicted and measured values, while the physical loss is calculated using the average of the squared residuals. The number of sampling points is 2 to 5 times that of the training samples. Initial weights are λconservation = 0.5, λenergy = 0.3, and λconsolidation = 1.0. When the mean residual Ri is greater than a threshold, the coefficients are adjusted according to the weight update formula, with an adjustment rate α ranging from 0.1 to 0.5. The thresholds are 0.01, 0.05, and 0.01, respectively. The Adam optimizer is used with a learning rate of 0.001 to 0.01, batch size of 64 to 256, and iterations of 5000 to 20000. Termination conditions include a loss change rate less than 10 for 50 consecutive epochs. -4 The residual R is conserved < 0.01, R energy < 0.05, and R consolidation < 0.01. The validation set RMSE is less than 10% of the measured mean.

[0045] In this embodiment: This technical solution uses a multi-layer feedforward neural network as its basic architecture, which can be implemented based on deep learning frameworks such as TensorFlow or PyTorch. The network consists of an input layer, three to five hidden layers, and an output layer. Each hidden layer contains 64 to 256 neurons, using ReLU or Tanh activation functions, and overfitting is prevented through Dropout and L2 regularization. A physical constraint calculation branch module is added to the standard architecture. This module extracts intermediate variables and prediction results from the main network, specifically calculating the residual terms of mass conservation, energy conservation, and Biot's consolidation equation, enabling real-time evaluation of the physical rationality of the model output. The input layer receives seven key parameters: soil layer thickness, compression modulus, initial void ratio, permeability coefficient, load magnitude, loading time, and drainage condition type. These parameters are directly related to the fundamental equations of geomechanics, ensuring the inherent consistency between the model input and physical laws. The output layer is designed with three nodes: total settlement, settlement distribution function, and average degree of consolidation, directly corresponding to engineering evaluation requirements and facilitating subsequent technical documentation generation and safety assessment.

[0046] The model training employs a composite loss function, consisting of a data fitting term and a physical constraint term. The data fitting term uses mean squared error to measure the deviation between predicted and measured values. The physical constraint term includes three residual sub-terms: mass conservation, energy conservation, and Biot's consolidation equation. The mass conservation residual assesses the consistency of soil volume changes before and after deformation; the energy conservation residual assesses the balance between work done by external forces and the increase in strain energy; and the Biot consolidation equation residual assesses the consistency between the increase in effective stress and the dissipation of pore water pressure. The composite loss function is constructed using a weighted summation. The weight of the data fitting term is set to 1.0, the weight of the mass conservation residual is 0.3 to 0.5, the weight of the energy conservation residual is 0.2 to 0.4, and the weight of the Biot consolidation equation residual is 0.4 to 0.6. A dynamic weight adjustment strategy is adopted: when the mean residual of a physical constraint term exceeds a preset threshold for five consecutive rounds, the weight of that term is increased to strengthen the constraint; when the residual falls below the threshold, the weight is appropriately reduced to avoid over-constraint. The model was trained using the Adam optimizer with an initial learning rate of 0.001 to 0.01 that gradually decreased. The training was terminated when the composite loss function did not decrease significantly for ten consecutive rounds and the residuals of each physical constraint term were below the engineering allowable error.

[0047] Through the above architectural design, a deep integration of data-driven learning and physical constraints is achieved. The input layer parameters comprehensively cover the physical process of soil consolidation and settlement, the hidden layer configuration provides sufficient nonlinear expressive power while effectively preventing overfitting, and the output layer indicators directly correspond to engineering requirements. The composite loss function incorporates data fitting accuracy and consistency with physical laws into a unified optimization objective, and the real-time evaluation mechanism of the physical constraint calculation branch ensures that the prediction results conform to the laws of mass conservation, energy conservation, and consolidation theory. The dynamic weight adjustment strategy enables the model to adaptively optimize according to the training process, effectively solving the problem of persistently high residuals that may be caused by fixed weights. Strict training termination conditions and data augmentation techniques ensure that the model achieves engineering-acceptable accuracy and possesses good generalization ability. The final output settlement prediction results conform to Terzaghi consolidation theory and the principle of mass conservation; the prediction curves highly match the theoretical analytical solutions; and the superimposed results of layered settlement meet the requirements of the layered summation method. Compared to existing purely data-driven models, this approach systematically addresses issues such as weak physical interpretability, prediction results that may violate physical laws, and difficulty in meeting engineering specifications by embedding physical constraint calculation branches, introducing multiple physical equation residual terms, adopting dynamic weight adjustment, and employing a rigorous physical verification process. This significantly improves the practicality and reliability of artificial intelligence models in the field of foundation settlement prediction.

[0048] Example 5: Please refer to Figure 1 The specific method for step S4 is as follows: S4.1 Construct a single-layer homogeneous saturated clay model with a soil layer thickness H of 10m, a compression modulus Es of 5MPa, and a permeability coefficient k of 1×10⁻⁶. -8With a speed of m / s, Poisson's ratio μ of 0.35, and an applied load q of 100 kPa, the boundary condition is that the top surface is permeable and the bottom surface is impermeable. The parameters are input into the PINN model trained on S3 to obtain the predicted degree of consolidation time history curve. The theoretical analytical solution is calculated based on Terzaghi consolidation theory. The time factor is calculated using the consolidation coefficient, time, and drainage distance. The consolidation coefficient is calculated using the permeability coefficient, water unit weight, and volumetric compressibility coefficient. The water unit weight is taken as 10 kN / m³. 3 The relative error was less than 5% within the consolidation degree range of 0% to 90%. A three-layer foundation model was constructed, consisting of a 5m thick silty clay layer with a compression modulus of 8MPa, an 8m thick silty clay layer with a compression modulus of 3MPa, and a 6m thick silty sand layer with a compression modulus of 15MPa. The settlement of each layer and the total settlement were predicted, and the superposition relative deviation was verified to be less than 3%. S4.2 Construct a model containing a confined aquifer, with an upper impermeable clay layer 5m thick and a permeability coefficient of 1×10⁻⁶. -9 m / s, with a pressure sand layer thickness of 10m and a permeability coefficient of 1×10⁻⁶. -4 m / s, bottom layer is impermeable bedrock, pore water pressure of top plate is 150 kPa and bottom plate is 200 kPa. Input PINN to predict the internal pressure distribution and seepage velocity distribution of the confined layer. Calculate the hydraulic gradient based on the predicted pressure distribution. The hydraulic gradient is calculated by dividing the partial derivative of pore pressure with depth by the unit weight of water, where the unit weight of water is taken as 10 kN / m. 3 The experiment verified that a hydraulic gradient greater than 0 indicates upward seepage. The theoretical seepage velocity was calculated using Darcy's law by multiplying the permeability coefficient by the hydraulic gradient. The relative error between the predicted velocity and the theoretical value was less than 10%. The dot product of the velocity vector and the head gradient vector was verified to be greater than 0. A control volume was selected to calculate the inflow and outflow mass fluxes. The mass fluxes were calculated by multiplying the water density, flow velocity, and cross-sectional area. The water density was taken as 1000 kg / m³. 3 The relative deviation in flux is less than 2%.

[0049] In this embodiment: This technical solution constructs a standardized theoretical verification system to systematically verify the consistency between the prediction results of the physical information neural network and the basic principles of geomechanics. First, a single-layer homogeneous saturated clay model is used as an idealized verification benchmark. This model specifies specific parameters such as a soil layer thickness of 10 meters, a compression modulus of 5 MPa, an initial void ratio of 0.8, and a permeability coefficient of 1 micrometer per second, giving the verification conditions clear physical meaning. These parameters are input into the trained physical information neural network model to obtain the predicted consolidation time history curve, which is then compared with the analytical solution of Terzaghi's one-dimensional consolidation theory. The root mean square error between the predicted and theoretical curves is calculated to evaluate the model's simulation accuracy of the basic consolidation process. Second, a three-layer foundation model is constructed, including an overlying clay layer, an intermediate sand layer, and an underlying clay layer. The superposition of settlement amounts from each layer is verified to be consistent with the results calculated using the layered summation method, thus verifying the model's applicability under complex geological conditions. Furthermore, for the geological model containing confined aquifers, the pore water pressure distribution and seepage velocity distribution within the confined aquifer are predicted. The hydraulic gradient is obtained by calculating the ratio of the hydraulic head difference between adjacent measuring points to the vertical distance, verifying the consistency between the sign of the hydraulic gradient and the actual seepage direction. The theoretical seepage velocity is calculated based on Darcy's law and compared with the model's predicted value. Finally, it is verified whether the dot product of the seepage velocity vector and the hydraulic head gradient vector is greater than zero, and the difference in fluid mass flowing into and out of the control volume per unit time is calculated to verify whether the model output satisfies the principle of mass conservation. Through these multi-dimensional physical verifications, it is ensured that the model's prediction results not only fit the measured data but also conform to the fundamental laws of geomechanics, effectively addressing the technical defect that the prediction results might violate physical equations.

[0050] This technical solution employs a layer-by-layer correlation propagation algorithm to reverse-engineer the trained physical information neural network, extracting the complete decision-making path from input features to prediction results. First, starting from the output layer, the initial correlation score is set to the predicted settlement value, directly anchoring the analysis to the core engineering objective. Then, according to the connection weights between layers of the neural network, the correlation score is propagated back to the input layer layer by layer, quantifying the contribution of each neuron node to the final prediction result through a weighted allocation mechanism. By ranking the contributions of all input features, the key parameters most significantly affecting settlement prediction are identified, including formation depth, pore water pressure, permeability coefficient, effective stress, and water content. These parameters are essential variables determining soil settlement behavior in geomechanics. Furthermore, the activation values ​​of the hidden layers are substituted into physical constraints such as the mass conservation equation, energy conservation equation, and Biot's consolidation equation. Through reverse solving, intermediate physical parameters such as degree of consolidation, soil compressibility, and seepage velocity are extracted, establishing an explicit mapping relationship between the internal representation of the neural network and geomechanical theory. Based on the extracted key input features and intermediate layer physical parameters, a complete reasoning chain is constructed according to the causal relationships in geomechanics, including three stages: pore water pressure identification, consolidation degree calculation, and settlement extrapolation. In the pore water pressure identification stage, the location and head distribution of the confined aquifer are determined based on hydrogeological principles. In the consolidation degree calculation stage, the core formulas of Biot's consolidation theory are applied to ensure the calculation process is traceable and consistent with the theoretical analytical solution. In the settlement extrapolation stage, the layered summation method and constitutive model are integrated, and a risk assessment is conducted according to engineering technical specifications.

[0051] Based on the results extracted from the above reasoning path, the system automatically generates a complete reasoning chain report, including the physical parameter identification process, calculation formula derivation, and risk assessment conclusions. This report details the source, value, and contribution to settlement prediction of each key parameter, and marks the geomechanical theories and engineering specifications upon which each calculation step is based, enabling technicians to clearly trace the mechanical basis and calculation path of each prediction result. The reasoning chain report is compiled according to the engineering design document format, including a stratigraphic profile diagram, a layered settlement calculation table, a consolidation degree time history curve, and a safety factor verification sheet, ensuring that the output document complies with current engineering technical standards and review requirements, and can be directly used for engineering design review and construction guidance. Through this technical solution, the problem of insufficient interpretability of artificial intelligence prediction results is systematically solved, achieving transparency and traceability in the settlement prediction process. It transforms the purely data-driven black-box model into a white-box model with clear physical logic, significantly improving the engineering credibility and decision support value of the prediction results, and meeting the stringent requirements of professional technicians for the physical rationality and standard compliance of the prediction model.

[0052] Example 6: Please refer to Figure 1 The specific method for step S5 is as follows: S5.1. The PINN model trained by S3 is back-analyzed using a layer-by-layer correlation propagation algorithm. The correlation score is propagated back from the output layer to the input layer. The initial correlation score in the output layer is set to be equal to the predicted settlement value. The correlation score in each hidden layer is allocated to the node of the previous layer according to the connection weight ratio. The contribution value of each input feature is calculated. The top five key input features with the highest contribution are identified, including formation depth h, pore water pressure u, permeability coefficient k, effective stress σ', and water content w. The physical parameters activated in the intermediate hidden layers are extracted, including the degree of consolidation U, soil compression Δs, and seepage velocity v. The physical parameters are obtained by substituting the activated values ​​of the hidden layers into the constraint terms of the S3 physical equation and solving in reverse. The numerical range, activation intensity, and influence weight of each physical parameter are recorded. S5.2. Based on the key features extracted in S5.1 and the physical parameters of the intermediate layer, a reasoning chain is constructed according to the geomechanical causal relationship. The first step is pore water pressure identification. When the ratio of pore pressure u to hydrostatic pressure is greater than 1.2, it is identified as a confined aquifer. The water unit weight γw is taken as 10 kN / m³. The second step is consolidation degree calculation. The consolidation degree U is calculated based on the effective stress change Δσ' and the pore pressure dissipation rate combined with the residual value of the Biot consolidation equation. The consolidation degree is obtained by dividing the difference between the initial pore pressure and the current pore pressure by the initial pore pressure. When U is greater than 60%, it is identified as the main stage of consolidation. The third step is settlement estimation. The final settlement s is estimated based on the consolidation degree U, soil layer thickness h, and volume compressibility coefficient mv combined with the tangent modulus Et. When s exceeds the design allowable value, it is identified as a high-risk area. The thickness is 30 mm for subway tunnels and 50 mm for building foundations. A reasoning chain report containing physical parameter identification results, calculation process, and risk conclusions is generated. The source of parameters and theoretical basis are marked in the report.

[0053] In this embodiment, the technical solution employs a layer-by-layer correlation propagation algorithm to back-analyze the trained physical information neural network model, extracting the complete decision path from input features to prediction results. Starting from the output layer, the algorithm sets the initial correlation score as the predicted settlement value, directly anchoring the analysis process to the core engineering objective. Then, according to the connection weights between layers of the neural network, the correlation score is propagated back to the input layer layer by layer, quantifying the contribution of each neuron node to the final prediction result through a weight ratio allocation mechanism. Specifically, the correlation score received by a neuron in a hidden layer is equal to the proportion of its output connection weight to the sum of all output connection weights, multiplied by the correlation score passed from the previous layer. Through layer-by-layer calculation, the correlation scores of all input features are finally obtained. The contributions of the input features are sorted in descending order to identify the key parameters that have the most significant impact on settlement prediction, including formation depth, pore water pressure, permeability coefficient, effective stress, and water content. These parameters are essential variables in geomechanics that determine soil settlement behavior. Furthermore, the activation values ​​of each neuron in the hidden layer are substituted into physical constraints such as the mass conservation equation, energy conservation equation, and Biot's consolidation equation. Intermediate physical parameters such as degree of consolidation, soil compressibility, and seepage velocity are extracted through inverse solving. This inverse solving process is based on the mathematical form of the physical equations. For example, the degree of consolidation is calculated by combining the effective stress value output from the hidden layer with the relationship between effective stress and pore water pressure in Biot's consolidation equation. This establishes an explicit mapping relationship between the internal representation of the neural network and geomechanical theory, giving the model's prediction process a clear physical meaning.

[0054] Based on extracted key input features and intermediate layer physical parameters, a complete reasoning chain is constructed according to the causal logic of geomechanics, including three stages: pore water pressure identification, consolidation degree calculation, and settlement extrapolation. In the pore water pressure identification stage, the system determines the groundwater type and distribution characteristics based on hydrogeological principles. It identifies confined aquifers by analyzing their position relative to impermeable layers in the stratigraphic profile, and determines the confined hydraulic head and hydraulic gradient direction based on measured or predicted pore water pressure distribution, providing initial conditions for subsequent consolidation analysis. In the consolidation degree calculation stage, the core formula of Biot's consolidation theory is applied. The degree of consolidation is equal to the ratio of pore water pressure dissipation at a given moment to the initial excess pore water pressure. By inputting stratigraphic depth, permeability coefficient, compressibility coefficient, and time parameters, the system calculates the degree of consolidation at different depths and times. The theoretical formulas and parameter sources used are clearly marked during the calculation process to ensure that each calculation step is traceable and consistent with the theoretical analytical solution. In the settlement estimation stage, the layered summation method and the Duncan-Chang constitutive model are integrated for settlement calculation. The foundation soil layers are divided into several calculation layers according to differences in compressibility. The compression caused by additional stress is calculated for each layer, and the sum of the compression values ​​of all layers is the total foundation settlement. The Duncan-Chang model is used to describe the stress-strain relationship of the soil in the compression calculation. The tangent modulus is determined based on the model parameters and stress level, and then the compressive strain and compression value are calculated. At the same time, a risk assessment is conducted according to engineering technical specifications. When the predicted settlement exceeds the allowable value of the specifications or the settlement rate is too fast, a high-risk level is marked and engineering treatment suggestions are proposed.

[0055] Based on the above reasoning process, the system automatically generates a complete reasoning chain report containing physical parameter identification results, detailed calculation process explanations, and risk assessment conclusions. The report records in detail the value, source, and contribution to settlement prediction of each key parameter, and marks the geomechanical theoretical formulas and engineering specification clauses used in each calculation step, enabling technicians to clearly trace the mechanical basis and calculation path of each prediction result. The reasoning chain report is compiled according to the engineering technical document format, including a parameter identification table, a consolidation degree calculation table, a layered settlement calculation table, and risk assessment conclusions, ensuring that the output document complies with current engineering technical standards and review requirements. The report provides formula annotations and theoretical explanations for key calculation steps; for example, the consolidation degree calculation section explicitly references the time factor formula of Biot's consolidation theory, and the settlement calculation section details the calculation formula of the layered summation method and the parameter values ​​for each soil layer. The above technical solutions systematically solve problems such as insufficient interpretability of artificial intelligence prediction results, lack of physical basis in the prediction process, and difficulty in meeting the verification requirements of engineering specifications. They realize the transparency and traceability of the settlement prediction process, transforming the purely data-driven black box model into a white box model with clear physical logic. This significantly improves the engineering credibility and decision support value of the prediction results, providing reliable technical support for engineering design and construction management.

[0056] Example 7: Please refer to Figure 1 The specific method for step S6 is as follows: S6.1 When new working condition data is detected, it is input into the PINN model trained in S3 to obtain the predicted settlement s. The actual settlement s is collected and measured. The prediction deviation Δs is calculated. When the absolute value of the deviation is greater than 15% of the measured value, the physical cause analysis is initiated. The analysis includes three aspects. First, the soil parameters of the new working condition are compared with the statistical range of the S2 training set to identify parameters that exceed the range and the degree of deviation is recorded. Second, the residuals R conservation, R energy, and R consolidation of the three types of physical equations of the new working condition are calculated. The convergence criteria of 0.01, 0.05, and 0.01 are compared to identify equations that exceed the standard and the excess multiple is recorded. Third, the stress-strain relationship data is extracted and the deviation from the theoretical curve of the Duncan-Chang model is calculated. When the deviation is greater than 20% or the stress ratio is greater than 0.85, it is determined that the constitutive parameters need to be corrected and the strategy for adjusting the physical constraint weights or constitutive model parameters is determined. S6.2. Based on the analysis results of S6-1, Bayesian optimization method is used to adjust the parameters. The objective function is to minimize the weighted sum of prediction bias and physical residual. The optimization variables include the weight coefficient λ conservation, λ energy, λ consolidation, and constitutive parameters K, n, and Rf. The search range is weight 0.1 to 10, K 50 to 500, n 0.2 to 1.0, and Rf 0.7 to 0.95. A probabilistic surrogate model is constructed using a Gaussian process. The next evaluation point is selected by the expected improvement EI criterion. Iterative execution is performed to evaluate the objective function value, update the surrogate model, calculate the EI value, and select the parameter combination with the maximum EI. The termination condition is that the improvement amount is less than 0.01 for 5 consecutive times, or after 50 iterations, or the objective function is less than 0.05. After obtaining the optimal parameters, the model is retrained using the S3 architecture. The training set is the original data from S2 combined with the new working condition data. The verification is that the mean of the three types of residuals is less than 0.01, 0.05, and 0.01, and the test relative error is less than 10%. After the conditions are met, the parameters of the S3 model are updated.

[0057] In this embodiment: when the physical information neural network model is applied to a new working condition, if the relative deviation between the predicted settlement and the measured value exceeds 15%, the physical cause analysis mechanism is triggered. This threshold setting avoids overreaction to minor fluctuations, allowing the system to focus on diagnosing and handling significant errors. The physical cause analysis includes a three-level systematic diagnostic process: First, parameter comparison analysis is performed, comparing key parameters such as soil layer thickness, compression modulus, initial void ratio, and permeability coefficient in the new working condition with the statistical range of corresponding parameters in the training set. The standard deviation multiple of each parameter from the mean of the training set is calculated to quickly identify abnormal parameters and their degree of deviation. For example, if the permeability coefficient of a soil layer is 0.1 micrometers per second, while the permeability coefficient range in the training set is 1 to 10 micrometers per second, this parameter is determined to be abnormal and marked as a key adjustment target. Second, physical equation residual verification is performed. After inputting the new working condition data into the model, the residuals of the mass conservation equation, energy conservation equation, and Biot consolidation equation output from the physical constraint calculation branch are extracted and compared with preset convergence criteria. If the mass conservation residual exceeds 1%, the energy conservation residual exceeds 2%, or the Biot consolidation equation residual exceeds 5% of the measured settlement, the model is deemed to violate the corresponding physical constraints under this working condition, and the corresponding physical constraint weight coefficients need to be adjusted. Next, a constitutive model suitability test is performed. The stress-strain relationship data predicted by the model is extracted and compared with the theoretical curves of the Duncan-Chang hyperbolic model or the modified Cambridge model. The root mean square deviation between the predicted and theoretical curves is calculated. When the deviation exceeds 10%, the existing constitutive model parameters are deemed unsuitable for the soil properties under this working condition, and the constitutive model parameters need to be recalibrated or a different constitutive model type needs to be used.

[0058] Based on the physical causal analysis results, Bayesian optimization is used to adaptively adjust the model parameters. Bayesian optimization, a global optimization algorithm based on a probabilistic surrogate model, uses a Gaussian process to model the objective function and combines it with the expected improvement criterion for efficient parameter search, finding the optimal parameter combination under reasonable computational cost. Specifically, the optimization variables are determined based on the physical causal analysis results: if the physical equation residuals exceed the limit, the mass conservation weight, energy conservation weight, and Biot consolidation equation weight are used as optimization variables; if the constitutive model deviation exceeds the limit, the initial tangent modulus, failure ratio, and exponential parameters of the Duncan-Chang model are used as optimization variables. The search range of the optimization variables is set based on engineering experience and literature data; for example, the mass conservation weight ranges from 0.2 to 0.8, and the constitutive model parameter range fluctuates by 50% based on typical values ​​for the soil type. The objective function is designed as a weighted sum of the data fitting error term and the physical constraint residual term. The data fitting error is the square of the difference between the predicted and measured settlement, and the physical constraint residual is the sum of the squares of the residuals of each physical equation. The weighting coefficients are set according to the importance of the project. The Bayesian optimization iterative process includes the following steps: first, evaluating the objective function value based on the current parameter combination; second, updating the Gaussian process surrogate model using the existing evaluation results; third, calculating the expected improvement value for each candidate parameter combination based on the surrogate model; and fourth, selecting the parameter combination with the largest expected improvement value for the next round of evaluation. The iteration terminates when the objective function value does not decrease significantly for five consecutive rounds and the decrease is less than 0.5%, or when the preset maximum number of iterations is reached, such as fifty iterations.

[0059] After parameter adjustments, the model was retrained using an expanded training set containing data from the new working conditions. The expanded training set consisted of a 9:1 mixture of the original training data and the measured data from the new working conditions, ensuring the model maintained its original generalization ability while adapting to the characteristics of the new conditions. The retraining process used the same optimizer and learning rate decay strategy as the initial training, but reduced the maximum number of training epochs to 50% of the original number to improve computational efficiency. After retraining, a comprehensive physical equation residual verification was performed: the mean residuals of the mass conservation equation, energy conservation equation, and Biot consolidation equation were calculated on all training and validation samples to confirm that all residuals were below the engineering allowable error standard. Simultaneously, theoretical comparison verification was conducted, applying the model to ideal working conditions with analytical solutions, such as single-layer homogeneous soil, to verify the consistency between the prediction results and the analytical solutions of Terzaghi consolidation theory. Only when both the physical equation residual verification and theoretical comparison verification met the requirements were the adjusted model parameters saved and applied to subsequent predictions. Through the above-mentioned automated deviation diagnosis, parameter optimization and model retraining process, the prediction deviation problem caused by new working condition data is systematically solved, enabling the model to dynamically adjust according to the actual working conditions and continuously maintain physical rationality, which significantly improves the model's adaptability, prediction reliability and engineering application value under complex and variable geological conditions.

[0060] Example 8: Please refer to Figure 1 The specific method for step S7 is as follows: S7.1. Based on the S5 inference chain report, extract technical parameters to generate calculation documents. The calculation documents include a stratigraphic profile, a settlement calculation report using the layered summation method, a consolidation degree time history curve, and a safety factor verification report. The stratigraphic profile extracts the depth of the S1-1 stratum and the soil type of S2-1 to draw a vertical profile and label six types of parameters, including thickness h, unit weight γ, compression modulus Es, permeability coefficient k, internal friction angle φ, and cohesion c. The layered summation method calculation report is prepared according to GB50007 format and includes four parts: self-weight stress, additional stress, compression amount, and total settlement. The consolidation degree curve extracts the data from the nine time nodes predicted by S3 to draw a Ut relationship curve and overlay the Terzaghi theoretical curve, labeling the t90 time point. The safety factor verification report is prepared according to GB50007 and GB50157 to calculate the anti-sliding safety factor Fssliding and the anti-uplift safety factor Fsuplift, and to determine whether Fssliding is greater than or equal to 1.2 and whether Fsuplift is greater than or equal to 1.5. S7.2. Conduct a compliance review of the technical documents generated in S7-1. The review includes three aspects: 1) Compliance review of the calculation method, verifying whether the layered summation method complies with the standard; 2) Checking the influence of groundwater level on the calculated buoyant unit weight γ' (taking 10 kN / m³); 3) Checking the stress diffusion angle range of 20 to 30 degrees; 4) Checking the layered calculation when the difference in compression modulus Es is greater than 2 times; 5) Reviewing the rationality of parameter values, verifying the empirical range of Es for cohesive soil (3 to 15 MPa, φ for 10 to 30 degrees) and Es for sandy soil (10 to 30 MPa, φ for 25 to 40 degrees), verifying that the parameter consistency deviation is less than 5%; 6) Checking that Fs slip is not less than 1.2 and Fs uplift is not less than 1.5; 7) Compliance review of the conclusion statement, verifying the final settlement amount, settlement rate, and settlement stabilization time t90. After the review is passed, output the technical documents in PDF and Word formats, along with the S5-2 inference chain report, indicating the generation time, project name, coordinates, signature, and settlement calculation safety verification conclusion according to the project overview and geological conditions.

[0061] In this embodiment: Based on the key parameters and calculation results extracted from the aforementioned inference chain report, a complete technical document system conforming to engineering technical specifications is automatically generated. This document system includes four core documents: stratigraphic profile, layered summation method calculation report, consolidation degree time history curve, and safety factor verification report. The stratigraphic profile is drawn using professional drawing software or geographic information system tools. The vertical axis represents stratigraphic depth, and the horizontal axis indicates soil layer type. Six key parameters are marked on the map: thickness of each soil layer, natural unit weight, compression modulus, initial void ratio, permeability coefficient, and groundwater level location. Different colors or fill patterns distinguish different soil layer types such as cohesive soil, sandy soil, and confined aquifers, providing accurate geological data support for subsequent settlement calculations. The layered summation method calculation report is strictly prepared in accordance with the technical requirements of the "Code for Design of Building Foundations" GB50007, including the basis for soil layer division, the stress calculation process for each soil layer, the compression calculation formula, and a summary table of final settlement. The calculations used buoyant unit weight for the soil layer below the groundwater level, which is equal to the natural unit weight minus the unit weight of water. The diffusion angle of the additional stress was verified to ensure it was between 20 and 30 degrees, consistent with the reasonable range of elasticity theory. The consolidation degree time history curve plotted time on the horizontal axis and consolidation degree on the vertical axis showed the predicted consolidation degree over time. Simultaneously, the analytical solution curve of Terzaghi's one-dimensional consolidation theory was overlaid on the graph. The comparison of the two curves verified the consistency between the model's predictions and classical theory.

[0062] To ensure that the generated technical documents comply with engineering specifications and physical rationality requirements, the system employs a three-tiered automatic review mechanism to comprehensively check the document content. The first tier is a review of the calculation method, checking whether the criteria for dividing each soil layer in the stratified summation method calculation are clear, whether the range of compressible layer thickness is reasonable, and verifying whether more detailed stratification has been performed when the difference in compressibility modulus between adjacent soil layers exceeds three times, ensuring that the calculation method adapts to complex geological characteristics. It also checks whether key calculation assumptions such as stress diffusion angle and basement pressure distribution meet specification requirements, avoiding inaccurate results due to improper calculation model selection. The second tier is a review of parameter values, verifying the rationality of the soil physical and mechanical parameters in the geological profile and calculation report. This includes whether the compressibility modulus is within the range of 0.5 MPa to 50 MPa, the permeability coefficient is within the range of 0.001 μm / s to 100 μm / s, and the initial void ratio is within the range of 0.3 to 1.5. When parameter values ​​exceed reasonable ranges, the system automatically marks them as abnormal and requires manual review, ensuring the physical authenticity of the input data. Simultaneously, check whether geometric parameters such as groundwater level depth and soil layer thickness are consistent with actual survey data to prevent data entry errors. The third step is to review the conclusions, checking whether the safety factor calculation report clearly gives a qualified or unqualified evaluation conclusion, whether the consolidation degree time history curve indicates the time required to reach 90% consolidation degree, and whether the layered summation method calculation report indicates the final settlement value and its engineering significance, ensuring that the document contains complete decision support information and avoiding the impact of missing key conclusions on engineering judgment and design decisions.

[0063] Technical documents that pass the above review are output in a unified standard format. The document naming convention includes identifying information such as the project name, station number or coordinates, and generation date, facilitating document management and retrieval. Each technical document is accompanied by a complete inference chain report as an attachment. This report details the entire calculation path from input parameters to the final prediction result, including the contribution analysis of key input features, the extraction process of intermediate physical parameters, the verification results of physical equation constraints, and the historical record of parameter adjustments, establishing a complete traceability system from raw data to engineering conclusions. Technical documents are output in portable document format or editable text processing format, ensuring that they can be directly used for engineering design review and construction guidance, while also facilitating necessary editing and supplementation by technical personnel as needed. The system also generates a document generation log, recording the document generation time, the model version used, the source of input parameters, and the review results, providing a basis for quality traceability and accountability. The above technical solutions systematically solve problems such as the difficulty in converting artificial intelligence prediction results into technical documents that conform to engineering specifications, the lack of a complete traceability mechanism in the prediction process, and the low efficiency of relying on manual review for document quality. They realize the automated conversion from intelligent prediction to standardized engineering documents, significantly improve the professionalism, reliability and engineering application value of technical achievements, and effectively connect artificial intelligence technology with the traditional engineering specification system.

[0064] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0065] 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 described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting surface subsidence using a drilling device based on artificial intelligence assistance, characterized in that: The specific steps are as follows: S1. During the drilling process, the drilling device simultaneously collects parameters such as formation depth, soil density, water content, and penetration resistance. At the formation interface, it obtains the total stress and pore water pressure data required for effective stress calculation through pore water pressure sensors and permeability coefficient testing modules, and records the pressure dissipation curve according to the time intervals required by Biot consolidation theory. S2. Physical constraint labeling of the collected data, calculation of seepage velocity and labeling of seepage direction according to Darcy's law, labeling of stress-strain relationship according to Duncan-Chang constitutive model, labeling of soil stability level using Mohr-Coulomb criterion, and establishing mapping relationship between data labels and geomechanical equations; S3. Construct a physical information neural network architecture, embed the mass conservation equation, energy conservation equation and Biot consolidation equation constraint terms into the loss function, set physical constraint weight coefficients, so that network training can minimize data fitting error and physical equation residuals at the same time. S4. Input the ideal formation model with known analytical solution to verify the conformity between the predicted settlement curve and the Terzaghi consolidation theory, and check the consistency of the superposition of layered settlement and the physical correctness of the direction of the hydraulic head gradient of the confined aquifer. S5. The decision path is traced in reverse by using the layer-by-layer correlation propagation algorithm to extract key input features and intermediate physical parameters, and generate a complete reasoning chain including pore water pressure identification, consolidation degree calculation and settlement inference. S6. When new operating condition data is detected, analyze the physical causes of the prediction deviation, use Bayesian optimization to adjust the physical constraint weights or constitutive model parameters, retrain and verify the convergence of the equation residuals. S7. Based on the inference chain, automatically generate technical documents that meet the requirements of the specifications, including stratigraphic profiles, calculation process of the layered summation method, consolidation degree time history curves, and safety factor verification sheets.

2. The method for detecting surface subsidence using an artificial intelligence-assisted drilling device according to claim 1, characterized in that: The specific method of step S1 is as follows: S1.1 During the drilling process, data is collected by a sensor array integrated on the outer wall of the drill pipe. The sensor array includes a depth encoder, a gamma ray density probe, a capacitive moisture content sensor, and a dynamic penetration resistance meter. The depth encoder records the drilling depth h at a resolution of 1 mm. The density probe measures the soil density ρ and converts it to unit weight γ. The moisture content sensor measures the mass moisture content w. The penetration resistance meter records the standard penetration test blow count N. A set of data is collected every 10 cm ± 2 cm of drilling, with a collection time interval of no more than 5 seconds. Each set of data includes depth coordinates, soil physical and mechanical parameters, and a timestamp, forming a basic parameter dataset for the stratigraphic profile. S1.

2. When a formation interface is detected, the test is initiated. The interface determination criteria are that the difference in N value between adjacent measuring points is greater than 5 blows or the density change rate exceeds 15%. The pore water pressure sensor penetrates the soil to a depth of not less than 20cm and is left to stand for 3 minutes before measuring the initial pore water pressure u0 with an accuracy of not less than 0.5kPa. Based on the data in S1-1, the total stress σ is calculated using the layered summation method to obtain a pair of dual-parameter data. The permeability coefficient testing module measures the permeability coefficient k. The pore water pressure u is recorded at time points of 0.5 minutes, 1 minute, 2 minutes, 4 minutes, 9 minutes, 16 minutes, 25 minutes, and 36 minutes. t The process continues until the dissipation rate reaches 90% or more, or the pressure change rate at three consecutive measuring points is less than 2%, or the monitoring lasts for 60 minutes, to form a time-series pressure dissipation curve dataset.

3. The method for detecting surface subsidence using an artificial intelligence-assisted drilling device according to claim 2, characterized in that: The specific method of step S2 is as follows: S2.

1. Based on the data in S1, physical parameters are calculated. According to Darcy's law, the seepage velocity v is calculated using the permeability coefficient k and the hydraulic gradient i. The hydraulic gradient i is calculated by the ratio of the difference in pore water pressure between adjacent measuring points to the vertical distance. The unit weight of water γw is taken as 10 kN / m³. When i is greater than 0, it is marked as upward seepage; when i is less than 0, it is marked as downward seepage. According to the Duncan-Chang model, the tangent modulus Et is calculated using the modulus parameters K and n, and the stress-strain relationship type is marked. The estimated range of K is 100 to 400, and the range of n is 0.4 to 0.

8. The safety factor Fs is calculated using the Mohr-Coulomb criterion using the cohesion c, the internal friction angle φ, and the effective stress σ'. When Fs is greater than or equal to 1.5, it is marked as high stability; when Fs is greater than or equal to 1.2 and less than 1.5, it is marked as medium stability; when Fs is less than 1.2, it is marked as low stability. S2.2 Establish a database mapping data labels to geomechanical equations. Map seepage parameters to the boundary conditions of the mass conservation equation, constraining the consistency of the seepage direction with the hydraulic gradient and the conservation of mass flux. Map stress-strain parameters to the Duncan-Chang incremental equation, constraining the stress path not to exceed the failure envelope. Map stability parameters to the Mohr-Coulomb yield criterion and force balance equation, constraining the yield function value to be less than or equal to zero and the resultant force to be zero. Generate a composite label vector for each set of data. The composite label vector contains the equation type code, parameter list, constraint expression, and penalty weight. The equation type code is 1 for mass conservation equation, 2 for constitutive relation equation, and 3 for force balance equation. The penalty weight ranges from 0.1 to 10, forming a structured training dataset containing original measurements, physical parameters, equation types, and constraint conditions.

4. The method for detecting surface subsidence using a drilling device based on artificial intelligence assistance according to claim 3, characterized in that: The specific method of step S3 is as follows: S3.1 Construct a multi-layer feedforward neural network. The input layer receives features from the S2 dataset, including seven parameters: formation depth h, soil weight γ, water content w, permeability coefficient k, pore water pressure u, effective stress σ', and time t. The hidden layer is a fully connected structure of 3 to 5 layers, with 50 to 200 neurons per layer, using the tanh activation function. The output layer includes total settlement s, settlement distribution s(z), and degree of consolidation U(t). Parallel physical constraint calculation branch modules are set up. The partial derivatives are calculated using automatic differentiation technology and substituted into the continuity equation to calculate the mass conservation residual R_conservation, where the water density ρw is taken as 1000 kg / m³. The energy residual R_energy is substituted into the energy conservation equation, and the consolidation residual R_consolidation is substituted into the Biot consolidation equation. Three types of residual values ​​are output. S3.2 Construct a composite loss function comprising a data fitting loss term Ldata and physical constraint loss terms Lconservation, Lenergy, and Lconsolidation. The data loss is calculated using the average of the squared differences between predicted and measured values, while the physical loss is calculated using the average of the squared residuals. The number of sampling points is 2 to 5 times that of the training samples. Initial weights are λconservation = 0.5, λenergy = 0.3, and λconsolidation = 1.

0. When the mean residual Ri is greater than a threshold, the coefficients are adjusted according to the weight update formula, with an adjustment rate α ranging from 0.1 to 0.

5. The thresholds are 0.01, 0.05, and 0.01, respectively. The Adam optimizer is used with a learning rate of 0.001 to 0.01, batch size of 64 to 256, and iterations of 5000 to 20000. Termination conditions include a loss change rate less than 10 for 50 consecutive epochs. -4 The residual R is conserved < 0.01, R energy < 0.05, and R consolidation < 0.

01. The validation set RMSE is less than 10% of the measured mean.

5. The method for detecting surface subsidence using an artificial intelligence-assisted drilling device according to claim 4, characterized in that: The specific method of step S4 is as follows: S4.1 Construct a single-layer homogeneous saturated clay model with a soil layer thickness H of 10m, a compression modulus Es of 5MPa, and a permeability coefficient k of 1×10⁻⁶. -8 With a speed of m / s, Poisson's ratio μ of 0.35, and an applied load q of 100 kPa, the boundary condition is that the top surface is permeable and the bottom surface is impermeable. The parameters are input into the PINN model trained on S3 to obtain the predicted degree of consolidation time history curve. The theoretical analytical solution is calculated based on Terzaghi consolidation theory. The time factor is calculated using the consolidation coefficient, time, and drainage distance. The consolidation coefficient is calculated using the permeability coefficient, water unit weight, and volumetric compressibility coefficient. The water unit weight is taken as 10 kN / m³. 3 The relative error was less than 5% within the consolidation degree range of 0% to 90%. A three-layer foundation model was constructed, consisting of a 5m thick silty clay layer with a compression modulus of 8MPa, an 8m thick silty clay layer with a compression modulus of 3MPa, and a 6m thick silty sand layer with a compression modulus of 15MPa. The settlement of each layer and the total settlement were predicted, and the superposition relative deviation was verified to be less than 3%. S4.2 Construct a model containing a confined aquifer, with an upper impermeable clay layer 5m thick and a permeability coefficient of 1×10⁻⁶. -9 m / s, with a pressure sand layer thickness of 10m and a permeability coefficient of 1×10⁻⁶. -4 m / s, bottom layer is impermeable bedrock, pore water pressure of top plate is 150 kPa and bottom plate is 200 kPa. Input PINN to predict the internal pressure distribution and seepage velocity distribution of the confined layer. Calculate the hydraulic gradient based on the predicted pressure distribution. The hydraulic gradient is calculated by dividing the partial derivative of pore pressure with depth by the unit weight of water, where the unit weight of water is taken as 10 kN / m. 3 The experiment verified that a hydraulic gradient greater than 0 indicates upward seepage. The theoretical seepage velocity was calculated using Darcy's law by multiplying the permeability coefficient by the hydraulic gradient. The relative error between the predicted velocity and the theoretical value was less than 10%. The dot product of the velocity vector and the head gradient vector was verified to be greater than 0. A control volume was selected to calculate the inflow and outflow mass fluxes. The mass fluxes were calculated by multiplying the water density, flow velocity, and cross-sectional area. The water density was taken as 1000 kg / m³. 3 The relative deviation in flux is less than 2%.

6. The method for detecting surface subsidence using a drilling device based on artificial intelligence assistance as described in claim 5, characterized in that: The specific method of step S5 is as follows: S5.

1. The PINN model trained by S3 is back-analyzed using a layer-by-layer correlation propagation algorithm. The correlation score is propagated back from the output layer to the input layer. The initial correlation score in the output layer is set to be equal to the predicted settlement value. The correlation score in each hidden layer is allocated to the node of the previous layer according to the connection weight ratio. The contribution value of each input feature is calculated. The top five key input features with the highest contribution are identified, including formation depth h, pore water pressure u, permeability coefficient k, effective stress σ', and water content w. The physical parameters activated in the intermediate hidden layers are extracted, including the degree of consolidation U, soil compression Δs, and seepage velocity v. The physical parameters are obtained by substituting the activated values ​​of the hidden layers into the constraint terms of the S3 physical equation and solving in reverse. The numerical range, activation intensity, and influence weight of each physical parameter are recorded. S5.

2. Based on the key features extracted in S5.1 and the physical parameters of the intermediate layer, a reasoning chain is constructed according to the geomechanical causal relationship. The first step is pore water pressure identification. When the ratio of pore pressure u to hydrostatic pressure is greater than 1.2, it is identified as a confined aquifer. The water unit weight γw is taken as 10 kN / m³. The second step is consolidation degree calculation. The consolidation degree U is calculated based on the effective stress change Δσ' and the pore pressure dissipation rate combined with the residual value of the Biot consolidation equation. The consolidation degree is obtained by dividing the difference between the initial pore pressure and the current pore pressure by the initial pore pressure. When U is greater than 60%, it is identified as the main stage of consolidation. The third step is settlement estimation. The final settlement s is estimated based on the consolidation degree U, soil layer thickness h, and volume compressibility coefficient mv combined with the tangent modulus Et. When s exceeds the design allowable value, it is identified as a high-risk area. The thickness is 30 mm for subway tunnels and 50 mm for building foundations. A reasoning chain report containing physical parameter identification results, calculation process, and risk conclusions is generated. The source of parameters and theoretical basis are marked in the report.

7. The method for detecting surface subsidence using an artificial intelligence-assisted drilling device according to claim 6, characterized in that: The specific method of step S6 is as follows: S6.1 When new working condition data is detected, it is input into the PINN model trained in S3 to obtain the predicted settlement s. The actual settlement s is collected and measured. The prediction deviation Δs is calculated. When the absolute value of the deviation is greater than 15% of the measured value, the physical cause analysis is initiated. The analysis includes three aspects. First, the soil parameters of the new working condition are compared with the statistical range of the S2 training set to identify parameters that exceed the range and the degree of deviation is recorded. Second, the residuals R conservation, R energy, and R consolidation of the three types of physical equations of the new working condition are calculated. The convergence criteria of 0.01, 0.05, and 0.01 are compared to identify equations that exceed the standard and the excess multiple is recorded. Third, the stress-strain relationship data is extracted and the deviation from the theoretical curve of the Duncan-Chang model is calculated. When the deviation is greater than 20% or the stress ratio is greater than 0.85, it is determined that the constitutive parameters need to be corrected and the strategy for adjusting the physical constraint weights or constitutive model parameters is determined. S6.

2. Based on the analysis results of S6-1, Bayesian optimization method is used to adjust the parameters. The objective function is to minimize the weighted sum of prediction bias and physical residual. The optimization variables include the weight coefficient λ conservation, λ energy, λ consolidation, and constitutive parameters K, n, and Rf. The search range is weight 0.1 to 10, K 50 to 500, n 0.2 to 1.0, and Rf 0.7 to 0.

95. A probabilistic surrogate model is constructed using a Gaussian process. The next evaluation point is selected by the expected improvement EI criterion. Iterative execution is performed to evaluate the objective function value, update the surrogate model, calculate the EI value, and select the parameter combination with the maximum EI. The termination condition is that the improvement amount is less than 0.01 for 5 consecutive times, or after 50 iterations, or the objective function is less than 0.

05. After obtaining the optimal parameters, the model is retrained using the S3 architecture. The training set is the original data from S2 combined with the new working condition data. The verification is that the mean of the three types of residuals is less than 0.01, 0.05, and 0.01, and the test relative error is less than 10%. After the conditions are met, the parameters of the S3 model are updated.

8. The method for detecting surface subsidence using an artificial intelligence-assisted drilling device according to claim 7, characterized in that: The specific method of step S7 is as follows: S7.

1. Based on the S5 inference chain report, extract technical parameters to generate calculation documents. The calculation documents include a stratigraphic profile, a settlement calculation report using the layered summation method, a consolidation degree time history curve, and a safety factor verification report. The stratigraphic profile extracts the depth of the S1-1 stratum and the soil type of S2-1 to draw a vertical profile and label six types of parameters, including thickness h, unit weight γ, compression modulus Es, permeability coefficient k, internal friction angle φ, and cohesion c. The layered summation method calculation report is prepared according to GB50007 format and includes four parts: self-weight stress, additional stress, compression amount, and total settlement. The consolidation degree curve extracts the data from the nine time nodes predicted by S3 to draw a Ut relationship curve and overlay the Terzaghi theoretical curve, labeling the t90 time point. The safety factor verification report is prepared according to GB50007 and GB50157 to calculate the anti-sliding safety factor Fssliding and the anti-uplift safety factor Fsuplift, and to determine whether Fssliding is greater than or equal to 1.2 and whether Fsuplift is greater than or equal to 1.

5. S7.

2. Conduct a compliance review of the technical documents generated in S7-1. The review includes three aspects: 1) Compliance review of the calculation method, verifying whether the layered summation method complies with the standard; 2) Checking the influence of groundwater level on the calculated buoyant unit weight γ' (taking 10 kN / m³); 3) Checking the stress diffusion angle range of 20 to 30 degrees; 4) Checking the layered calculation when the difference in compression modulus Es is greater than 2 times; 5) Reviewing the rationality of parameter values, verifying the empirical range of Es for cohesive soil (3 to 15 MPa, φ for 10 to 30 degrees) and Es for sandy soil (10 to 30 MPa, φ for 25 to 40 degrees), verifying that the parameter consistency deviation is less than 5%; 6) Checking that Fs slip is not less than 1.2 and Fs uplift is not less than 1.5; 7) Compliance review of the conclusion statement, verifying the final settlement amount, settlement rate, and settlement stabilization time t90. After the review is passed, output the technical documents in PDF and Word formats, along with the S5-2 inference chain report, indicating the generation time, project name, coordinates, signature, and settlement calculation safety verification conclusion according to the project overview and geological conditions.

Images