A deep learning-based subgrade solidification effect prediction method and system

Abstract

This invention relates to the field of intelligent prediction technology, and discloses a method and system for predicting the effect of roadbed solidification based on deep learning. The method includes evaluating the matching degree between the soil and the solidifying agent and generating a score. If qualified, a preliminary strength trend curve is constructed based on the material ratio and ambient temperature, and its baseline slope is determined. Subsequently, the slope is analyzed in conjunction with environmental geological parameters to identify nonlinear critical nodes and potential inflection intervals in the solidification process. For this interval, the soil moisture content and ambient temperature are extracted, and the preliminary trend curve is dynamically coupled and corrected to generate an optimized strength growth model. The model is used to calculate simulated predicted values, and the deviation from the actual values ​​is used to determine iterative correction parameters, thereby optimizing and generating a corrected strength distribution map. This map is compared with a safety threshold to complete the solidification effect evaluation. Finally, an optimized ratio scheme is generated based on the evaluation conclusions. This method can improve the accuracy of predicting the inflection point of solidification strength.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction technology, and in particular to a method and system for predicting the effect of roadbed solidification based on deep learning. Background Technology

[0002] Currently, in the field of civil engineering, the research and application of roadbed consolidation technology plays a crucial role in ensuring the quality and service life of road projects. However, the prediction and evaluation of its effects have long relied on traditional experiments and empirical formulas, which are difficult to accurately cope with complex geological environments and variable conditions. This leads to deviations between the predicted results and the actual engineering performance, increasing the uncertainty and safety hazards in engineering design.

[0003] In a current technology, the prediction of subgrade consolidation effects mainly relies on traditional physical testing methods and empirical formulas derived from experimental data. This method typically involves conducting indoor consolidation tests on limited soil samples (e.g., particle composition, moisture content) to obtain strength data under specific conditions, and then establishing simplified prediction formulas through mathematical regression and other methods. However, this method, based on traditional tests and empirical formulas, heavily depends on specific, ideal test conditions for model construction, making it difficult to characterize and adapt to the complex and ever-changing actual engineering environment and the heterogeneity of the soil itself. Essentially, it is a static fit to limited historical test data, lacking the ability to deeply model and express the physicochemical mechanisms behind the consolidation process, particularly the dynamic matching and interaction between the soil's intrinsic properties and the proportion of consolidation materials. Therefore, when faced with real-world scenarios involving different soil types, complex environmental conditions (such as temperature and humidity variations), and their interactions, this method struggles to accurately grasp and deduce the entire process of intensity growth over time. In particular, it cannot effectively identify and predict nonlinear inflection points that may occur during intensity development. For example, under specific hydrothermal environments and mix proportions, the intensity growth of cohesive soil tends to plateau or prematurely enters a critical state of attenuation, leading to prediction deviations or failures at key decision points in practical engineering. In this context, the introduction of deep learning technology has become a powerful solution.

[0004] In summary, the existing technology suffers from low accuracy in predicting the inflection point of curing strength. Summary of the Invention

[0005] This invention provides a method and system for predicting the solidification effect of roadbed based on deep learning, in order to solve the problem of low accuracy in predicting the inflection point of solidification intensity.

[0006] In a first aspect, to address the aforementioned technical problems, this invention provides a method for predicting the effect of roadbed solidification based on deep learning, comprising: Obtain soil parameters, analyze the matching relationship between the soil parameters and the curing agent, and generate an initial matching score; If the initial matching score is higher than the preset matching threshold, the solidification reaction rate is calculated using the preset neural network perception model and trend analysis is performed to determine the preliminary trend curve and the baseline slope of the initial intensity growth. Environmental geological parameters are obtained, and a time-varying interaction sequence is constructed in combination with the benchmark slope. The time-varying interaction sequence is subjected to temporal feature extraction and pattern analysis through a temporal convolutional network to obtain the evolution trajectory of solidification intensity. Critical nodes are identified based on the evolution trajectory, and potential turning points of nonlinear changes are determined based on the critical nodes. Soil moisture content parameters are extracted from the potential turning point intervals, and the preliminary trend curve is coupled and corrected based on the soil moisture content parameters to generate an optimized strength growth curve. The simulation calculation is performed based on the intensity growth optimization curve to obtain the simulated intensity prediction value. The deviation of the simulated intensity prediction value is calculated, and based on the deviation, the number of iterations and correction coefficient required for deviation correction are determined. Based on the number of iterations and the correction coefficient, the simulated intensity prediction value is corrected to generate corrected intensity data. A corrected intensity distribution map is constructed based on the corrected intensity data. The intensity values ​​in the corrected intensity distribution map are compared with a preset safety threshold. Based on the comparison results, a final evaluation report is generated. Based on the final evaluation report and the revised intensity distribution map, the final proportion adjustment parameters are generated.

[0007] Secondly, the present invention provides a deep learning-based roadbed solidification effect prediction system, comprising: The data acquisition module is used to acquire soil parameters, analyze the matching relationship between the soil parameters and the curing agent, and generate an initial matching score. The trend analysis module is used to calculate the solidification reaction rate and perform trend analysis using a preset neural network perception model if the initial matching score is higher than the preset matching threshold, thereby determining the preliminary trend curve and the baseline slope of the initial intensity growth. The evolution analysis module is used to acquire environmental geological parameters, construct a time-varying interaction sequence in combination with the benchmark slope, extract time-series features and perform pattern analysis on the time-varying interaction sequence through a time-series convolutional network to obtain the evolution trajectory of solidification intensity, identify critical nodes based on the evolution trajectory, and determine potential turning intervals of nonlinear changes based on the critical nodes. The strength optimization module is used to extract soil moisture content parameters from the potential turning point, and to couple and correct the preliminary trend curve based on the soil moisture content parameters to generate an optimized strength growth curve. The deviation correction module is used to perform simulation calculations based on the intensity growth optimization curve to obtain the simulated intensity prediction value, calculate the deviation of the simulated intensity prediction value, and determine the number of iterations and correction coefficients required for deviation correction based on the deviation. The effect evaluation module is used to correct the simulated intensity prediction value according to the number of iterations and the correction coefficient, generate corrected intensity data, construct a corrected intensity distribution map based on the corrected intensity data, compare the intensity value in the corrected intensity distribution map with a preset safety threshold, and generate a final evaluation report based on the comparison results. The scheme output module is used to generate the final ratio adjustment parameters based on the final evaluation report and the modified intensity distribution map.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention generates a matching score by obtaining soil characteristic parameters and analyzing the matching degree between soil organic matter content and solidifier type; this process is based on the quantitative evaluation of key soil characteristics, realizing the accurate matching between solidifier selection and soil properties, and avoiding prediction deviation caused by material mismatch.

[0009] (2) When the matching degree score meets the requirements, the present invention constructs a preliminary trend curve and determines the baseline slope based on the material ratio and ambient temperature. It further combines environmental geological parameters to analyze the coupling relationship, identify critical nodes and potential turning points in the evolution process. Through multi-factor collaborative analysis, it can capture the key characteristics of the interaction between the initial trend of intensity growth and the geological environment, provide early warning of nonlinear change risks, and enhance the ability to predict the dynamic evolution law of solidification effect.

[0010] (3) The present invention calculates the simulated strength prediction value based on the strength growth optimization curve, determines the number of iterations and correction coefficients required for deviation correction by calculating the deviation between the simulated value and the actual value, optimizes the prediction value and generates a corrected strength distribution map, and generates the curing effect requirements and final evaluation report after comparing with the safety threshold. This iterative correction process realizes the dynamic calibration of the prediction results, and ensures the reliability of the evaluation conclusion and the safety of the project by continuously reducing the prediction error. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of a method for predicting the effect of roadbed solidification based on deep learning, provided in the first embodiment of the present invention. Figure 2 This is a schematic diagram of a roadbed solidification effect prediction system based on deep learning provided in the second embodiment of the present invention. Detailed Implementation

[0012] 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.

[0013] Reference Figure 1 The first embodiment of the present invention provides a method for predicting the effect of roadbed solidification based on deep learning, including the following steps: S11, Obtain soil parameters, analyze the matching relationship between the soil parameters and the curing agent, and generate an initial matching score; S12, if the initial matching score is higher than the preset matching threshold, the solidification reaction rate is calculated using the preset neural network perception model and trend analysis is performed to determine the preliminary trend curve and the baseline slope of the initial intensity growth. S13, obtain environmental geological parameters, construct a time-varying interaction sequence in combination with the benchmark slope, extract time-series features and perform pattern analysis on the time-varying interaction sequence through a time-series convolutional network to obtain the evolution trajectory of solidification intensity, identify critical nodes based on the evolution trajectory, and determine potential turning intervals of nonlinear changes based on the critical nodes. S14, extract soil moisture content parameters from the potential turning point interval, and couple and correct the preliminary trend curve based on the soil moisture content parameters to generate an optimized strength growth curve; S15, Perform simulation calculations based on the intensity growth optimization curve to obtain the simulated intensity prediction value, calculate the deviation of the simulated intensity prediction value, and determine the number of iterations and correction coefficients required for deviation correction based on the deviation. S16, Based on the number of iterations and the correction coefficient, the simulated intensity prediction value is corrected to generate corrected intensity data. A corrected intensity distribution map is constructed based on the corrected intensity data. The intensity values ​​in the corrected intensity distribution map are compared with a preset safety threshold. Based on the comparison results, a final evaluation report is generated. S17. Based on the final evaluation report and the modified intensity distribution map, generate the final proportion adjustment parameters.

[0014] In step S11, it is necessary to obtain soil parameters, analyze the matching relationship between the soil parameters and the curing agent, and generate an initial matching score, including: Soil characteristic parameters are collected, and principal component analysis is performed on the soil characteristic parameters to extract particle feature vectors and water content matrices; the soil characteristic parameters include soil particle distribution characteristics and soil water content parameters; Based on the particle feature vector and the moisture content matrix, historical data is extracted from a preset database and interpolation completion technology is used to complete the data to obtain a curing strength sequence. Based on the solidification strength sequence, the mapping relationship between soil organic matter content and different types of solidifying agents is analyzed to generate an initial matching score.

[0015] First, based on the characteristics of soil particle distribution, in-situ borehole sampling was conducted in specific strata to obtain undisturbed, undisturbed soil samples. These samples were then sent to the laboratory for physical dispersion and pretreatment according to geotechnical testing specifications. Next, a laser particle size analyzer was used to scan the prepared suspension. By measuring the diffraction angle of the laser light on the particles, the volume percentage of hundreds of consecutive particle size intervals across the entire particle size range, from clay to sand, was recorded in a single operation, thus generating the original particle size distribution curve data. For the moisture content parameter, measuring points were set at fixed intervals (e.g., every 0.5 meters) along the vertical direction of the survey section. A portable time domain reflectometer (TDR) probe was inserted into each measuring point, and the soil volumetric moisture content was calculated by measuring the propagation time of electromagnetic waves in the soil. Finally, a set of discrete moisture content values ​​were obtained by collecting data according to the depth sequence.

[0016] After data acquisition, principal component analysis (PCA) was used to process the original high-dimensional data. First, the original particle size distribution data acquired by the laser particle size analyzer was preprocessed using Z-score standardization to eliminate the influence of dimensions. Next, the covariance matrix of the standardized data was calculated and eigenvalue decomposition was performed. The eigenvalues ​​were sorted from largest to smallest, and the top k principal components with a cumulative variance contribution rate of 85% were selected. The eigenvectors corresponding to these principal components constitute the weight matrix for the projection transformation. The original particle size distribution curve data was multiplied by this weight matrix and projected onto the directions of these k principal components, thereby compressing and generating a low-dimensional particle size distribution eigenvector containing only k components. Meanwhile, for a series of moisture content data collected from different depths, they are arranged into column vectors according to the depth sequence, and the spatial change rate is quantified by calculating the difference between adjacent depth data, finally forming a 3-row, 2-column moisture content state matrix. The first column records the measured moisture content values ​​at three depths, and the second column records the moisture content change gradient between adjacent depths. In this way, while preserving the key physical and mechanical properties of the soil, the original discrete moisture content data is integrated into a structured matrix that can characterize the vertical distribution law of moisture, which significantly reduces the data dimension and complexity of subsequent calculations.

[0017] It is worth noting that the cumulative variance contribution rate threshold is set based on engineering experience. Extensive engineering practice has verified that selecting 85% can retain the main key information of the original data while reducing the data dimensionality to a level with lower computational difficulty, effectively balancing information retention and computational efficiency. This threshold can be fine-tuned according to the actual data characteristics.

[0018] Next, based on the generated particle size distribution feature vector and moisture content state matrix, records with similar soil characteristics are retrieved from a pre-set database. Specifically, the similarity between the feature vector of the current soil and the feature vector of historical cases is calculated (generally using cosine similarity as a metric), and several historical cases with the highest similarity are selected as references. Then, complete solidification strength-time correspondence sequence data of each record is extracted from these matching cases. When the retrieved historical case data is discontinuous or missing, the overall trend of the existing data points of the case can be analyzed, and then linear interpolation can be used to calculate the strength estimate of the missing day in a linear fashion over time. For missing values ​​at both ends of the sequence or more complex cases, a trained time series prediction model (such as the ARIMA model) may be called to infer the missing values, thereby generating a continuous and complete solidification strength sequence that reveals the law of soil strength growth over time.

[0019] It should be noted that the database can be constructed in the following way: First, systematically collect and input complete geotechnical test data (such as particle analysis and moisture content), corresponding material ratios, detailed construction environment records, and unconfined compressive strength data measured at various ages from numerous completed roadbed consolidation projects; then, clean, denoise, and standardize all collected multi-source and heterogeneous data, and encode and store the soil parameters of each case according to a unified feature vector model, thereby forming a structured historical case knowledge base.

[0020] Finally, the mapping relationship between organic matter content and different types of curing agents is analyzed using the obtained complete curing strength sequence. Specifically, the organic matter content in historical cases is used as a continuous feature variable, the type of curing agent is used as a categorical variable after one-heat encoding, and the curing effect index (such as strength or strength growth rate at a specific age) is used as the target variable to construct a structured training dataset. Subsequently, multiple linear regression is used to train this dataset to fit a mathematical model that can describe the complex relationship between input features (organic matter content, type of curing agent) and output target (curing effect index). Then, for the specific organic matter content of the current soil, the system inputs the data into the above-mentioned correlation model to calculate the expected performance index when using different types of curing agents. By comparing the degree of agreement between the expected index and the actual or expected evolution law reflected by the current soil curing strength sequence, for example, calculating the correlation coefficient between the predicted curve and the actual sequence, the final initial matching score is obtained.

[0021] In step S12, if the initial matching score is higher than a preset matching threshold, a preset neural network perception model is used to calculate the curing reaction rate and perform trend analysis to determine the preliminary trend curve and the baseline slope of the initial intensity growth, including: Obtain the material ratio and ambient temperature record values. If the initial matching score is higher than the preset matching threshold, the material ratio and ambient temperature record values ​​are mapped to a curing rate vector through a preset neural network perception model. The curing rate vector is input into a numerical model based on the curing kinetic equation to calculate the initial strength increment matrix. The strength increment of each time step is extracted from the matrix to form a strength increment sequence. The sliding window algorithm is used to smooth and fit the strength increment sequence to obtain a preliminary trend curve. Extract the set of tangent derivatives of the preliminary trend curve and calculate the arithmetic mean of the set of tangent derivatives to determine the baseline slope of the initial intensity growth.

[0022] It should be noted that after determining that the initial matching score is higher than the preset matching threshold, the system automatically retrieves and reads the recorded precise material proportion values ​​from the integrated construction management database or IoT sensor network, and simultaneously acquires the environmental temperature record sequence continuously monitored and uploaded by temperature and humidity sensors deployed at the construction site within a set time window. The preset matching threshold is determined as follows: First, the initial matching score of each completed project and its corresponding discrete label (e.g., qualified or unqualified) determined by domain experts or key performance indicators are extracted from the historical case library. Next, score distribution curves for qualified and unqualified cases are plotted using kernel density estimation, and an optimal split point is found by calculating the intersection of the two curves or maximizing the Youden index. This split point is fixed as the system's preset matching threshold and is periodically recalibrated as the case library expands to maintain its discriminative ability.

[0023] Next, a pre-defined neural network perception model is used to fuse and non-linearly map the two sets of input values ​​to generate a solidification rate vector.

[0024] The acquisition and training process of the neural network perception model is as follows: First, complete material proportions, environmental temperature time-series data during construction, and physicochemical parameters characterizing reaction rates (such as hydration heat curves at different ages) from completed engineering cases are extracted from the historical database as training sample sets, and all input and output data are normalized. The model adopts a multi-layer feedforward neural network structure. The number of nodes in the input layer is set to 8, corresponding to the material proportion components (mass percentage of cement, fly ash, and mineral powder) and temperature statistical features (daily average temperature, temperature rise rate, and effective accumulated temperature). The hidden layer is designed with 3 layers, with the number of nodes being 64, 32, and 16 respectively, using the ReLU activation function. Each layer is followed by a batch normalization layer and a Dropout layer (with a deactivation rate of 0.2). The number of nodes in the output layer is 6, corresponding to the parameters of each reaction stage of the curing rate vector. During training, the mean squared error is used as the loss function, the Adam optimizer is used, the initial learning rate is 0.001, the batch size is 32, and the maximum number of training rounds is 200. L2 regularization (weight decay of 1e-4) is introduced, and the training and validation sets are divided in a 7:3 ratio. Early stopping is used (the validation set loss stops after 20 consecutive rounds). Hyperparameters are determined through grid search combined with 5-fold cross-validation. The tuning range includes learning rate (0.0001, 0.0005, 0.001), batch size (16, 32, 64), hidden layer node combinations, and Dropout rate (0.1, 0.2, 0.3). The mean absolute error of the validation set is used as the evaluation metric to select the optimal parameter combination. Finally, the optimal model parameters, network structure, and data preprocessing standards are permanently saved and integrated into the system for real-time access.

[0025] Subsequently, the generated curing rate vector is input into a numerical model based on the curing kinetic equation to calculate the initial intensity increment matrix. During the calculation, the model gradually accumulates the theoretical intensity contribution value generated by the chemical reaction in each time interval at a preset time step (e.g., 12 hours), thereby outputting a series of discrete intensity prediction values ​​arranged in order of age, forming the initial intensity increment matrix. Next, in order to eliminate high-frequency fluctuations in the discrete data caused by measurement noise or random disturbances, the matrix needs to be smoothed. Specifically, the moving average method is used, setting a sliding window containing multiple consecutive time points (e.g., covering 3 time steps), moving sequentially on the time series, and using the arithmetic mean of all data points within the window as the new, smoothed intensity value at the center time point of the window. After performing this operation on all data points, the originally discrete intensity increment values ​​are transformed into a continuous, smooth curve, i.e., the initial trend curve.

[0026] It should be noted that the sliding window size is determined based on the data sampling frequency and noise characteristics. In this embodiment, the sampling interval is 12 hours, and the three windows correspond to 36 hours, which can effectively smooth short-term fluctuations caused by diurnal temperature differences without masking long-term trends. In practical applications, the optimal window length can be determined by analyzing the autocorrelation function of historical data.

[0027] The numerical model is constructed as follows: First, a constitutive equation framework describing strength development is established based on materials science theory (e.g., using a maturity function and reaction degree model that considers the influence of temperature). Next, key reaction kinetic parameters (such as reaction rate constant and activation energy) are obtained through numerous indoor standard tests (e.g., testing the compressive strength of cured bodies under different isothermal conditions), and the quantitative relationship between these parameters and temperature is established using the Arrhenius equation. Then, the equations and parameters calibrated through experiments are integrated to construct a computable mathematical model, which is then discretized into an iterative calculation format with a fixed time step using numerical methods (e.g., the Euler method). Finally, the model is validated and calibrated using independent historical engineering data that was not involved in parameter calibration. By comparing the degree of agreement between the model's predicted strength curve and the actual monitoring data, the empirical coefficients in the model are repeatedly adjusted until its prediction accuracy meets the requirements of engineering applications, thus completing the final model setup.

[0028] Subsequently, several key time points representing the initial growth stage (e.g., hours 24, 48, and 72) are selected on the curve. These points are chosen based on characteristics such as the uniformity of time distribution or curve inflection points. Next, at each selected point, the system calculates the slope of the tangent line using a numerical differentiation method. Specifically, taking a very small time interval before and after the selected point, the system uses the function values ​​of these two neighboring points on the curve to calculate the instantaneous slope (i.e., the first derivative value) of the tangent line at that point using the central difference formula. The instantaneous slope values ​​calculated for all selected points are collected to form the tangent derivative set. Finally, the system performs an arithmetic mean operation on all slope values ​​in this set, sums all slope values, divides by the total number of slopes, and determines the average value as the initial intensity growth benchmark slope characterizing the overall early intensity development rate under this working condition.

[0029] In step S13, environmental geological parameters need to be obtained, and a time-varying interaction sequence is constructed based on the benchmark slope. A temporal convolutional network is used to extract temporal features and perform pattern analysis on the time-varying interaction sequence to obtain the evolution trajectory of solidification intensity. Critical nodes are identified based on the evolution trajectory, and potential turning points in nonlinear changes are determined based on the critical nodes, including: Environmental geological parameters are obtained, and a time-varying interactive sequence of the geological environment is constructed by combining the baseline slope of the initial intensity growth. The environmental geological parameters include the range of environmental humidity fluctuations, historical deviation distribution patterns, solidification intensity sequence data, and geological complexity level. The time-varying interaction sequence is subjected to deep feature extraction using a temporal convolutional network, and the time-varying interaction sequence is decomposed into a time series to obtain the nonlinear evolution trajectory of the curing intensity. By performing second-order difference operations on the nonlinear evolution trajectory, critical nodes are identified, and based on the critical nodes, potential turning points in the nonlinear changes are determined.

[0030] It should be noted that after acquiring environmental geological parameters, including the range of environmental humidity fluctuations, historical deviation distribution patterns, solidification intensity sequence data, and geological complexity level, the qualitatively described geological complexity level (e.g., "Level III") is quantified into a specific environmental attenuation coefficient (e.g., 0.85) according to a preset mapping rule. This coefficient characterizes the degree to which adverse geological factors (e.g., groundwater seepage) weaken the ideal intensity growth trend. Next, the system performs a scalar multiplication operation on this attenuation coefficient and the baseline slope of the initial intensity growth to obtain the base growth rate corrected for geological disturbances. Then, using this corrected daily growth rate as a baseline, along the complete time span from the start to the end of solidification (e.g., 120 days), with "days" as the solidification period... The system recursively calculates a baseline intensity cumulative value for each time point in the sequence by a fixed time step, thereby generating a discrete baseline growth sequence. Simultaneously, the monitored environmental humidity fluctuation range (e.g., 65% to 82%) is modeled as a sine or cosine perturbation function with a period of 24 hours. The amplitude of this function is determined by the fluctuation range, and the phase is aligned with the starting conditions of environmental monitoring. Finally, at each time point, the system algebraically adds the instantaneous perturbation value (which can be positive or negative) calculated by the above periodic humidity perturbation function to the corresponding baseline intensity cumulative value, thereby generating a discrete data sequence that integrates long-term geological decay trends and short-term environmental periodic perturbations and changes continuously over time, namely, the geological environment time-varying interactive characteristic sequence.

[0031] It is worth noting that the mapping rules are established in the following way: First, a large number of engineering cases in the historical database are statistically analyzed to extract the complexity level (such as level 1 to level 5) defined by the geological exploration report of each case and the average deviation ratio between the final intensity observed in the actual project and the theoretically predicted intensity; then, statistical induction is used to establish the correspondence between qualitative levels and quantitative attenuation coefficients. For example, the average deviation ratio of the intensity of all historical cases belonging to level 3 complexity is calculated to be about 15%, so the attenuation coefficient is mapped to 1 minus this deviation ratio, i.e., 0.85. Finally, the correspondence between these levels and coefficients is solidified into a preset query mapping table or a simple piecewise linear function.

[0032] Subsequently, the system uses a pre-defined temporal convolutional network model to perform deep feature extraction on the acquired long-term trend term sequence, extracting deep abstract features that characterize the key dynamics of the nonlinear evolution of intensity. The specific structure of the temporal convolutional network model is as follows: A modular stacked architecture is used to build the network, which has a total of 8 layers. Each layer consists of a causal dilated convolutional module. The dilation coefficient of the dilated convolutional kernel is set from the first layer according to powers of 2, i.e., the dilation coefficient of the i-th layer is... The network employs a 3-dimensional convolutional kernel, with its weights initialized using a He normal distribution. After each convolutional operation, a normalization layer is immediately applied to standardize the output features, followed by a linear rectified activation function to introduce non-linearity. Simultaneously, each convolutional module has an identity mapping residual connection path, directly adding the module's input to the output after convolution, normalization, and activation, as the module's final output. The last layer of the network does not use an activation function, directly outputting the feature vector. During training, the entire network uses historical data, with mean squared error as the loss function, and is trained end-to-end using the Adam optimizer until convergence. The training data includes 500 historical engineering cases, each containing 120 days of continuous monitoring data, divided into training and validation sets in an 8:2 ratio. Early stopping is used to control model training.

[0033] Next, the seasonal and trend decomposition method is used to process the aforementioned time-varying interactive sequence. First, the period length of the periodic fluctuations in the sequence (e.g., 24 hours) is determined. Then, a moving average window with a length matching the period (e.g., a moving average filter with a window length of 24 for a daily periodic sequence with hourly intervals) is slid across the original time-varying interactive sequence. At each sliding position, the system calculates the arithmetic mean of all data points within the window and uses this mean as the initial estimate of the long-term trend term corresponding to the center position of the window. This operation effectively smooths and filters out short-term high-frequency fluctuations caused by daily changes in environmental humidity. After completing the moving average calculation for the entire sequence, a new, smoother sequence with the same length as the original sequence is obtained, namely the long-term trend term sequence. At the same time, the system separates the remaining part containing periodic fluctuations and irregular noise by subtracting this long-term trend term sequence from the original sequence, and then extracts the seasonal periodic term from it. Finally, the long-term trend term sequence obtained by the system is defined as the nonlinear evolution trajectory of solidification intensity.

[0034] Subsequently, to accurately capture the critical point where the growth rate abruptly changes in the nonlinear evolution trajectory, the system performs a second-order difference operation on the trajectory curve. First, the nonlinear evolution trajectory is considered as a discrete sequence arranged in chronological order. For each data point from the second to the last in the sequence, its first-order difference value is calculated to obtain the instantaneous velocity sequence. Then, for this instantaneous velocity sequence, its second-order difference value is calculated starting from the second value, thus obtaining the instantaneous acceleration sequence. Subsequently, the system sets a very small positive number ε (e.g., 1e-6) close to zero as a tolerance threshold for judging whether the acceleration has crossed zero, and sequentially scans the acceleration sequence: for each internal point (not the first or last point) in the sequence, the system checks whether the absolute value of its current acceleration value is less than ε, or whether its sign (positive / negative) is opposite to the sign of its previous acceleration value. When either of the above conditions is met, it indicates that near time point t, the intensity growth acceleration has undergone a fundamental change from positive to negative (inflection point of slowing growth) or from negative to positive (inflection point of accelerating growth), and the system marks this time point as a critical node.

[0035] Finally, the potential turning point interval is divided with the identified critical node as the center. The specific method is as follows: First, in the historical case database, all cases with similar geological complexity and soil conditions to the current project are retrieved, and the duration of the time from the identified critical node to the point where the intensity growth deviates significantly from the expected trajectory (such as a sharp drop or surge in the growth rate) in these cases is statistically analyzed. The statistical characteristic value (such as the median) of these time lengths is calculated as the benchmark span. Then, the system multiplies this benchmark span value by an engineering safety factor greater than 1 to include a certain early warning margin, thereby obtaining the specific buffer length extending before and after the critical node. Finally, the system uses the exact time of the critical node as the midpoint, subtracts the buffer length forward to obtain the starting point of the interval, and adds the buffer length backward to obtain the ending point of the interval. The time period defined in this way is precisely defined as the potential turning point interval. It is worth noting that the engineering safety factor can be determined as follows: First, analyze the dispersion of the actual strength fluctuation time length relative to its baseline span in the above-mentioned similar historical cases (generally measured by the coefficient of variation); simultaneously, the system will combine the risk tolerance corresponding to the current engineering safety level (such as ordinary engineering or key engineering) and preset different basic coefficient values ​​for different levels (such as 1.1 for ordinary engineering and 1.3 for key engineering); finally, the engineering safety factor is calculated using the following formula: ,in For the final engineering safety factor, These are the base coefficient values ​​corresponding to different levels. Given the current degree of dispersion, This represents the maximum acceptable volatility dispersion in historical data (typically taken as 0.3).

[0036] In step S14, it is necessary to extract soil moisture content parameters from the potential turning point interval, and couple and correct the preliminary trend curve based on the soil moisture content parameters to generate an optimized strength growth curve, including: Moisture content parameters are extracted from the potential transition intervals and combined with the recorded ambient temperature values ​​to generate dynamic coupling parameters; Based on the dynamic coupling parameters, a dynamic coupling analysis is performed on the preliminary trend curve to establish a mapping relationship; Based on the mapping relationship, the initial trend curve is adjusted to generate an intensity growth optimization curve.

[0037] First, for the identified potential turning points, based on the time boundaries of these intervals, the system extracts timestamp-aligned water content and temperature reading sequences from the original high-frequency monitoring database. For the water content sequence, a sliding window of a preset time length (e.g., 6 hours) is used to move across the sequence, calculating the arithmetic mean of the data within each window to smooth high-frequency noise. Then, the statistical median of the set of all window averages is calculated, and this median is mapped to the range of 0 to 1 using a min-max linear normalization method, resulting in a standardized water content parameter that quantifies the typical water content within the interval. For the temperature sequence, a temperature threshold (e.g., 0°C) is applied to each data point, and data below the threshold are considered invalid. Then, the valid temperature data is numerically integrated over time, specifically using the trapezoidal rule to calculate the area under the temperature-time curve, obtaining the effective accumulated temperature value in degrees per day. Finally, the system combines the standardized water content parameter and the effective accumulated temperature value into a binary feature vector, which serves as a dynamic coupling parameter characterizing the hydrothermal synergy.

[0038] The temperature threshold can be set in the following way: based on the theoretical minimum temperature required for the curing agent (such as cement) to undergo a significant hydration reaction, and referring to a large amount of data from indoor standard curing tests, the critical temperature point where strength growth almost stops is determined by analyzing the early strength development of the cured body under different constant temperatures; this critical temperature point is then increased by a safety margin (twice the standard deviation of the field monitoring temperature data) to cope with field fluctuations, and finally solidified into the temperature threshold of the material system, which is used to filter out low-temperature data that do not substantially contribute to strength development when calculating the effective accumulated temperature.

[0039] Subsequently, these dynamic coupling parameters are correlated with the preliminary trend curve. The standardized water content parameters and effective accumulated temperature values ​​obtained from the current calculations are used as a set of feature inputs and fed into a pre-defined nonlinear multiple regression model. Internally, the model performs nonlinear combinations and calculations on these features based on its learned weights and bias parameters, ultimately outputting a specific correction coefficient. This correction coefficient is a quantified proportional value, such as 1.05 or 0.92, which precisely characterizes the specific gain (coefficient greater than 1) or attenuation (coefficient less than 1) effect of the current hydrothermal coupling state on the theoretical intensity growth rate at any time point on the preliminary trend curve. By performing this calculation once for each discrete time point on the curve, the system establishes a one-to-one correspondence between environmental parameters and the intensity growth rate change coefficient at that time point; this is the required mapping relationship.

[0040] It is worth noting that the nonlinear multiple regression model can be constructed as follows: First, a gradient boosting decision tree is selected as the model framework, with standardized water content parameters and effective accumulated temperature as input features, and the intensity growth rate correction coefficient as the output target. The core hyperparameters of the model are determined through grid search: the number of trees is set to 100, the maximum depth is 5, the learning rate is 0.05, and the subsampling ratio is 0.8. During the training phase, training samples are extracted from a historical engineering database. Each sample contains a pair of hydrothermal environment parameters (water content, effective accumulated temperature) and the corresponding measured correction coefficient (i.e., the ratio of the actual intensity growth rate to the theoretical baseline growth rate) within a specific time period. The dataset is divided into a training set and a validation set in a 7:3 ratio, and the mean squared error is used as the loss function. During training, the current residual is fitted in each iteration to gradually construct a new decision tree; at the same time, early stopping (stopping when the validation set loss does not decrease for 10 consecutive rounds) and five-fold cross-validation are used to evaluate the model's generalization ability and prevent overfitting.

[0041] Finally, the preliminary trend curve is reconstructed point by point based on the above mapping relationship. This involves multiplying the values ​​of each original data point on the preliminary trend curve by the correction coefficient determined by the corresponding mapping relationship, and then recombining the corrected data points in the original order to obtain the final intensity growth optimization curve.

[0042] In step S15, simulation calculations are performed based on the intensity growth optimization curve to obtain simulated intensity prediction values. The deviation of the simulated intensity prediction values ​​is calculated, and based on the deviation, the number of iterations and correction coefficients required for deviation correction are determined, including: Based on the intensity growth optimization curve, multiple simulations are performed by introducing random perturbation variables to generate simulated intensity prediction values. The simulated intensity predictions are matched with historical intensity data, and the geological complexity level is quantified based on the deviation distribution pattern of the matching results. If the geological complexity level exceeds a preset complexity threshold, the dispersion of the multiple sets of simulated intensity prediction values ​​relative to the central trend is calculated to obtain the deviation magnitude; Based on the deviation magnitude and in conjunction with the preset mapping relationship between the deviation magnitude and the correction parameters, the number of iterations required for deviation correction and the corresponding correction coefficient are determined.

[0043] First, the simulated intensity prediction value is calculated based on the intensity growth optimization curve. Spatial variation statistical characteristics of key soil parameters (such as moisture content and dry density) are extracted from historical databases or geological survey reports, including the mean, standard deviation, and possible probability distribution types (such as normal distribution) for each parameter. Next, the system generates a large number of random number sequences that conform to this distribution, with each set representing a possible spatial distribution scenario for the soil parameters. Then, the system uses each set of random parameter scenarios as input and independently calls the intensity growth optimization curve for parallel calculation, thereby obtaining an intensity prediction time series corresponding to that parameter scenario. The system generates 500 sets of intensity prediction sequences through Monte Carlo simulation, selecting 5 representative outputs (such as the 10th, 30th, 50th, 70th, and 90th percentiles) as input for subsequent analysis, i.e., the simulated intensity prediction value.

[0044] Subsequently, these multiple sets of simulated strength predictions are matched with actual strength monitoring data under the same or similar soil conditions in the historical engineering database. The cosine similarity between key parameters such as the particle size distribution feature vector and organic matter content of the current soil and the corresponding parameters in historical cases is calculated. A batch of reference cases with the closest soil conditions is then selected from the historical database. Next, for each generated set of simulated strength prediction sequences, the system calculates the residual values ​​at the same solidification age time point, comparing it with the measured strength sequences of each successfully matched reference case. This results in a set containing a large number of residual data points. Then, the standard deviation of the residual set is calculated and compared with a preset level threshold interval to obtain the final geological complexity level parameters.

[0045] It is worth noting that the level threshold range can be determined as follows: First, collect the standard deviation dataset of all completed cases in the historical database after the above matching and residual calculation, forming a population sample of prediction bias dispersion. Next, perform statistical analysis on this dataset, using the quantile partitioning method from unsupervised learning, combined with engineering expert experience, to divide the continuous dispersion range into several intervals (e.g., levels I to V) with clear engineering significance. For example, by calculating the 20th, 40th, 60th, and 80th percentiles of the dataset, dispersion below the 20th percentile is defined as level I (simple), dispersion between the 20th and 40th percentiles is level II (medium), and so on, with dispersion above the 80th percentile defined as level V (extremely complex). These partition points (e.g., 0.4 MPa) are then fixed as the preset level thresholds within the system.

[0046] Next, the obtained geological complexity level is compared with a preset complexity threshold (e.g., Level II). If the current level exceeds this threshold, the dispersion of multiple sets of simulated intensity prediction values ​​relative to the central trend is calculated: First, the median of the obtained multiple sets of simulated intensity prediction values ​​is determined as the central trend. Then, within the entire time span from the initial to the final stage of solidification, a series of representative key time points (e.g., days 7, 28, and 90) are selected. At each selected time point, the intensity value of each prediction sequence at that point is read to form a set of prediction values ​​for that moment. The range and standard deviation of this set are calculated and fused with equal weights to obtain the final deviation.

[0047] It is worth noting that the complexity threshold can be set in the following way: First, a retrospective analysis is performed on all cases in the historical engineering database to statistically analyze the incidence rate of cases where the final strength of the actual project does not meet the design requirements or where significant construction adjustments are required under each geological complexity level (Level I to Level V); then, the risk incidence rate corresponding to different levels is compared with the acceptable risk baseline of the project, and a critical level is determined through iterative analysis. When the level exceeds this critical value, the risk incidence rate will significantly exceed the baseline; finally, this critical level is solidified as the preset complexity threshold.

[0048] Finally, the number of iterations required for deviation correction and the corresponding correction coefficient are determined based on the calculated deviation magnitude. First, the method internally includes a pre-defined magnitude-iteration-coefficient mapping table trained using historical data. During application, the calculated current deviation magnitude is used as the lookup key, and the corresponding recommended iteration number N and a correction coefficient are directly obtained by searching this mapping table.

[0049] It is worth noting that the amplitude-iteration-coefficient mapping table can be constructed as follows: First, define the table as a discrete lookup table structure, where the input (key) is the discretized deviation amplitude range, and the output (value) is the corresponding recommended iteration number and correction coefficient, thus completing the table structure definition. Then, training samples are extracted from the historical database. Each sample contains the initial predicted deviation amplitude of a historical case, the validated minimum number of iterations required for prediction convergence, and the optimized single-step correction coefficient. These samples are then trained using cluster analysis such as k-means clustering. The mapping relationship between each cluster center is used as the final element to fill the table, and cross-validation is used to ensure the generalization ability of the rules.

[0050] In step S16, the simulated intensity prediction value needs to be corrected according to the number of iterations and the correction coefficient to generate corrected intensity data. A corrected intensity distribution map is constructed based on the corrected intensity data. The intensity values ​​in the corrected intensity distribution map are compared with a preset safety threshold. Based on the comparison results, a final evaluation report is generated, including: Based on the number of iterations and the correction coefficient, the simulated intensity prediction value is optimized to generate corrected intensity data, and a corrected intensity distribution map is constructed based on the corrected intensity data. Based on the corrected intensity distribution map, calculate the stability index of the simulated intensity prediction value; If the node strength values ​​of the modified strength distribution map are all within the preset safety threshold range, and the stability index is qualified, the mark meets the curing effect requirements, and the final evaluation report is output.

[0051] First, for each set of original simulated intensity prediction values, a loop is started, and the number of loops is equal to the preset number of iterations. In each loop, the system multiplies the current value of the data point by the correction coefficient and uses the calculation result as the input value for the next loop. This process is repeated until all the specified number of loops are completed, thereby obtaining a new intensity prediction sequence that has been adjusted step by step.

[0052] Next, the system introduces the Kriging spatial interpolation algorithm to correlate all optimized intensity data points with their specific coordinates in the three-dimensional space of the roadbed, forming a known set of sample points. Then, the algorithm analyzes the relationship between the spatial distance between these sample points and the corresponding intensity values, calculates and fits a mathematical model called the semi-variogram, to quantitatively describe the spatial correlation and variability of intensity. Based on this model, for any unknown location point to be predicted within the roadbed area, the algorithm comprehensively considers the spatial location and semi-variogram values ​​of multiple known sample points around it, and assigns an optimal weight to each known sample point by solving a set of linear equations. Finally, the intensity values ​​of these sample points are weighted and summed to calculate the intensity estimate of the unknown location point. By applying this process to all nodes to be predicted within the roadbed cross-section divided by a regular grid, a corrected intensity distribution map covering the entire target area with continuously changing intensity values ​​is finally generated.

[0053] Subsequently, a stability index for the simulated intensity predictions is calculated based on this corrected intensity distribution map. First, the intensity values ​​and spatial coordinates of all regular grid nodes in the corrected intensity distribution map are read. Next, the absolute intensity differences between each grid node and all its directly adjacent spatial nodes (such as front, back, left, and right adjacent nodes) are traversed, resulting in a set of differences containing the intensity variations of all adjacent nodes. Then, this set of differences is statistically analyzed to calculate its variance, which quantitatively characterizes the dispersion and fluctuation of intensity in spatial distribution. Finally, this calculated variance is compared with a preset acceptable threshold (e.g., 0.15). If the variance is less than or equal to the threshold, the stability index is considered acceptable, indicating a uniform intensity distribution; conversely, if the variance is greater than the threshold, it is considered unacceptable, indicating abrupt changes or unevenness in intensity distribution in space.

[0054] It is worth noting that the qualified threshold can be determined in the following way: Extract all successful solidification engineering cases from the historical engineering database, and perform the variance calculation of the strength difference between adjacent nodes on the final strength distribution map of each case to obtain a variance dataset of historical successful cases; then, perform statistical analysis on this dataset, usually taking a higher percentile (e.g., the 90th percentile) as a reference benchmark to ensure that the threshold can cover the fluctuation level of most successful conditions; at the same time, combine the minimum requirements of engineering specifications for roadbed uniformity to fine-tune and confirm the statistical benchmark value; finally, solidify this comprehensively determined value as the preset qualified threshold in the system. Finally, each grid node in the modified intensity distribution map is read sequentially to obtain its intensity value. This value is then compared with the preset lower and upper safety thresholds to check if it is simultaneously greater than or equal to the lower limit and less than or equal to the upper limit. If any node's intensity value fails to meet this condition, the traversal is immediately terminated and recorded as a failure. If all nodes pass the check and the stability index is confirmed to be qualified, a structured final evaluation report is automatically generated. The core content of this report includes the overall evaluation conclusion (qualified), key data summaries such as the lowest node intensity, the highest node intensity, stability index values, geological complexity level, original material mix parameters, and the timestamp of passing the verification.

[0055] The safety threshold can be set as follows: First, based on the design specifications and acceptance standards of the roadbed engineering, determine the minimum design strength value that the curing strength must meet, and use it as the initial benchmark for the lower limit of the threshold. Simultaneously, combine the theoretical maximum strength of the curing material used with the statistical values ​​of the actual upper limit of strength in historical projects where no damage has occurred, and determine the initial benchmark for the upper limit of the threshold. Next, extract a large number of engineering cases from the historical successful case database that have ultimately been verified as safe and stable in similar soil conditions and environments. Statistically analyze the actual distribution range of their full-line strength monitoring data. Calculate their statistical characteristics (such as taking the lower bound of the confidence interval corresponding to the design strength value, and the higher percentile of the actual strength data distribution, such as the upper percentile) to calibrate and fine-tune the initial upper and lower limit benchmarks. Finally, solidify the calibrated lower and upper limits as the preset safety threshold.

[0056] In step S17, based on the final evaluation report and the modified intensity distribution map, final mix proportion adjustment parameters need to be generated, including: Based on the final assessment report, combined with the corrected intensity distribution map and the material proportion values, a multidimensional correlation dataset for a specific geological complexity level is constructed. The original curing agent content data is extracted from the multidimensional associated dataset, and a ratio adjustment matrix is ​​generated through optimization calculation. The ratio adjustment matrix is ​​input into a preset simulation curing model to update the simulated intensity prediction value and calculate a new stability index; If the dispersion in the new stability index is lower than that of the original stability index, the final ratio adjustment parameters are determined according to the ratio adjustment matrix; otherwise, the ratio adjustment matrix and subsequent stability verification steps are recalculated.

[0057] First, the modified intensity distribution map is read to obtain the unique coordinate identifier of each spatial grid node in the map and its corresponding intensity prediction value. At the same time, the geological complexity level and original material ratio parameter information are read from the final assessment report. These data are integrated to obtain a data record. The record uses the node coordinates as an index and integrates the key information of the three dimensions of the node: intensity value, geological complexity level, and original material ratio parameter. The set of data records of all nodes constitutes a multidimensional associated dataset.

[0058] Next, the original curing agent content data is extracted from the multidimensional association dataset as the initial benchmark for optimization, and these original data are iteratively calculated by calling an intelligent optimization algorithm. For example, this invention uses a genetic algorithm as the intelligent optimization algorithm.

[0059] Specifically, the genetic algorithm used in this invention sets the population size to 200 based on the total number of roadbed grid nodes to ensure sufficient search diversity; the maximum number of iterations is set to 500 generations to ensure sufficient convergence; the crossover rate is set to 0.85, using a two-point crossover method; and the mutation rate is set to 0.02, using Gaussian random perturbation. During iterative optimization, the algorithm defines the original solidifier content of each spatial grid node as a "gene," and the content values ​​of all nodes together constitute an initial "population" of individuals.

[0060] Then, an iterative optimization process is performed. In each iteration, the algorithm first selects parent individuals using a roulette wheel selection method based on the individual fitness (i.e., the objective function value). Next, the selected parent individuals are paired according to the crossover rate, and gene fragments representing some grid nodes are randomly exchanged to generate offspring with new characteristics. Then, a small number of genes are randomly selected from all offspring individuals according to the mutation rate, and small random additions and subtractions are made within their value range. After each iteration, the algorithm retains the historically best individuals. In each iteration, better individuals are retained through selection, genes from different individuals are combined through crossover to explore new solutions, and the content of a small number of nodes is randomly fine-tuned through mutation to maintain population diversity. The convergence criterion is that the improvement in the optimal fitness of the population over 50 consecutive generations is less than 1e-5. This process is repeated until the algorithm converges to the Pareto optimal solution set that cannot be significantly improved under the given constraints. Finally, an optimal solution that balances performance and cost is selected from this solution set and decoded into a numerical matrix that corresponds one-to-one with the roadbed spatial grid and contains the adjusted curing agent content of each node, i.e., the ratio adjustment matrix.

[0061] It is worth noting that the fitness function in the above genetic algorithm is defined as follows. First, define the cost function:

[0062] in It is the variance of the absolute strength difference between all adjacent nodes of N spatial grid nodes on the roadbed cross section, which is the stability index of step S16. Let be the predicted intensity of the i-th node; These are the lower and upper limits of the safety threshold in S16, respectively; The adjusted curing agent content is generally taken as a reference, using the original curing agent ratio. The standard content of curing agent These are the penalty coefficients for the corresponding items. In the specific implementation, they are used to assign penalty intensity for strength exceeding safety limits and excessive deviation of the hardener adjustment scheme from the original ratio. Considering that safety and data magnitude differences should be prioritized when adjusting the ratio, they are set separately. The fitness function is then defined as follows:

[0063] This formula converts cost C into fitness F.

[0064] Subsequently, the system inputs this mix proportion adjustment matrix into a preset simulation solidification model for virtual construction drills. The model first reads the mix proportion adjustment matrix, which defines the adjusted material proportion parameters for each spatial grid node of the roadbed cross-section. Then, using these new, spatially differentiated mix proportion parameters as one of the core inputs, and combining them with soil characteristic parameters (such as particle size distribution and moisture content) and historical environmental data (such as temperature and humidity sequences) for each node, the model calls its internal strength prediction function, built based on reaction kinetics principles, to independently and in parallel calculate a new predicted strength-time curve for each node and extract the strength value at the target age (e.g., 28 days). After completing the calculations for all nodes, the system obtains an updated set of node strength values ​​covering the entire roadbed cross-section and generates a new simulated strength distribution map accordingly. Finally, based on this new strength distribution map, the system uses the same stability index calculation method as in S16 to calculate a new stability index.

[0065] It is worth noting that the acquisition and training process of the simulation solidification model is as follows: In the setup phase, the model framework is constructed based on the principles of material response dynamics and soil mechanics. Its inputs include soil parameters, environmental data, and material mix proportions, and the output is the strength prediction value. In the training phase, the system uses a large amount of complete "input-output" monitoring data from historical projects to train the model. The internal parameters of the model are continuously adjusted through optimization algorithms such as gradient descent to minimize the error between the model's predicted intensity and the actual monitored intensity. After training, the model achieves the required prediction accuracy on an independent validation set, and its final structure and parameters are solidified and saved, integrated into a computational module that can be called upon. Finally, the newly calculated stability index of the predicted value is compared with the stability index of the original scheme. First, the difference between the old and new stability indices is calculated, and it is checked whether the new index value is not only less than the original value, but also whether its reduction exceeds a preset effective optimization threshold (e.g., a reduction of no less than 10%). If this condition is met, the new mix design is officially deemed successfully optimized. Subsequently, the mix adjustment matrix that generated the successful scheme is locked as valid output. This matrix stores the specific material proportion parameters (such as curing agent content) of each spatial grid node of the roadbed cross-section after optimization calculation. Finally, through a data conversion and output module, all parameters in this matrix are serialized and encapsulated according to a preset engineering data format (such as a parameter list or visualization guide corresponding to construction zones), generating a final mix adjustment parameter file that can be directly used to guide on-site construction.

[0066] It should be noted that the effective threshold for optimization can be set as follows: Statistical analysis is conducted based on a historical optimization case library. Data on the relative improvement rate (i.e., reduction) of stability indicators of all historically successful ratio optimization schemes is collected. A benchmark value is determined by calculating a specific lower quantile (e.g., the 25th percentile) of this data distribution to ensure that the improvement of the new scheme reaches at least the medium level of historical successful cases. At the same time, combined with the experience definition of significant improvement by engineering experts (e.g., an improvement of at least 5% is considered engineering-worthy), this statistical benchmark value is calibrated and confirmed, ultimately forming a fixed percentage threshold that takes into account both statistical data patterns and engineering experience judgment.

[0067] Reference Figure 2 The second embodiment of the present invention provides a roadbed solidification effect prediction system based on deep learning, comprising: The data acquisition module is used to acquire soil parameters, analyze the matching relationship between the soil parameters and the curing agent, and generate an initial matching score. The trend analysis module is used to calculate the solidification reaction rate and perform trend analysis using a preset neural network perception model if the initial matching score is higher than the preset matching threshold, thereby determining the preliminary trend curve and the baseline slope of the initial intensity growth. The evolution analysis module is used to acquire environmental geological parameters, construct a time-varying interaction sequence in combination with the benchmark slope, extract time-series features and perform pattern analysis on the time-varying interaction sequence through a time-series convolutional network to obtain the evolution trajectory of solidification intensity, identify critical nodes based on the evolution trajectory, and determine potential turning intervals of nonlinear changes based on the critical nodes. The strength optimization module is used to extract soil moisture content parameters from the potential turning point, and to couple and correct the preliminary trend curve based on the soil moisture content parameters to generate an optimized strength growth curve. The deviation correction module is used to perform simulation calculations based on the intensity growth optimization curve to obtain the simulated intensity prediction value, calculate the deviation of the simulated intensity prediction value, and determine the number of iterations and correction coefficients required for deviation correction based on the deviation. The effect evaluation module is used to correct the simulated intensity prediction value according to the number of iterations and the correction coefficient, generate corrected intensity data, construct a corrected intensity distribution map based on the corrected intensity data, compare the intensity value in the corrected intensity distribution map with a preset safety threshold, and generate a final evaluation report based on the comparison results. The scheme output module is used to generate the final ratio adjustment parameters based on the final evaluation report and the modified intensity distribution map.

[0068] It should be noted that the deep learning-based roadbed solidification effect prediction system provided in this embodiment of the invention is used to execute all the process steps of the deep learning-based roadbed solidification effect prediction method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0069] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0070] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting the effect of roadbed consolidation based on deep learning, characterized in that, include: Obtain soil parameters, analyze the matching relationship between the soil parameters and the curing agent, and generate an initial matching score; If the initial matching score is higher than the preset matching threshold, the solidification reaction rate is calculated using the preset neural network perception model and trend analysis is performed to determine the preliminary trend curve and the baseline slope of the initial intensity growth. Environmental geological parameters are obtained, and a time-varying interaction sequence is constructed in combination with the benchmark slope. The time-varying interaction sequence is subjected to temporal feature extraction and pattern analysis through a temporal convolutional network to obtain the evolution trajectory of solidification intensity. Critical nodes are identified based on the evolution trajectory, and potential turning points of nonlinear changes are determined based on the critical nodes. Soil moisture content parameters are extracted from the potential turning point intervals, and the preliminary trend curve is coupled and corrected based on the soil moisture content parameters to generate an optimized strength growth curve. The simulation calculation is performed based on the intensity growth optimization curve to obtain the simulated intensity prediction value. The deviation of the simulated intensity prediction value is calculated, and based on the deviation, the number of iterations and correction coefficient required for deviation correction are determined. Based on the number of iterations and the correction coefficient, the simulated intensity prediction value is corrected to generate corrected intensity data. A corrected intensity distribution map is constructed based on the corrected intensity data. The intensity values ​​in the corrected intensity distribution map are compared with a preset safety threshold. Based on the comparison results, a final evaluation report is generated. Based on the final evaluation report and the revised intensity distribution map, the final proportion adjustment parameters are generated.

2. The method for predicting the effect of roadbed solidification based on deep learning according to claim 1, characterized in that, The process of acquiring soil parameters, analyzing the matching relationship between the soil parameters and the curing agent, and generating an initial matching score includes: Soil characteristic parameters are collected, and principal component analysis is performed on the soil characteristic parameters to extract particle feature vectors and water content matrices; the soil characteristic parameters include soil particle distribution characteristics and soil water content parameters; Based on the particle feature vector and the moisture content matrix, historical data is extracted from a preset database and interpolation completion technology is used to complete the data to obtain a curing strength sequence. Based on the solidification strength sequence, the mapping relationship between soil organic matter content and different types of solidifying agents is analyzed to generate an initial matching score.

3. The method for predicting the effect of roadbed solidification based on deep learning according to claim 1, characterized in that, If the initial matching score is higher than a preset matching threshold, a preset neural network perception model is used to calculate the solidification reaction rate and perform trend analysis to determine the preliminary trend curve and the baseline slope of the initial intensity growth, including: Obtain the material ratio and ambient temperature record values. If the initial matching score is higher than the preset matching threshold, map the material ratio and ambient temperature record values ​​into a curing rate vector through a preset neural network perception model. The curing rate vector is input into a numerical model based on the curing kinetic equation to calculate the initial strength increment matrix. The strength increment of each time step is extracted from the initial strength increment matrix to form a strength increment sequence. The sliding window algorithm is used to smooth and fit the strength increment sequence to obtain a preliminary trend curve. Extract the set of tangent derivatives of the preliminary trend curve and calculate the arithmetic mean of the set of tangent derivatives to determine the baseline slope of the initial intensity growth.

4. The method for predicting roadbed consolidation effect based on deep learning according to claim 1, characterized in that, The process involves acquiring environmental geological parameters, constructing a time-varying interactive sequence based on the baseline slope, extracting temporal features and performing pattern analysis on the time-varying interactive sequence using a temporal convolutional network to obtain the evolution trajectory of solidification intensity, identifying critical nodes based on the evolution trajectory, and determining potential turning points in nonlinear changes based on the critical nodes. This includes: Environmental geological parameters are obtained, and a time-varying interactive sequence of the geological environment is constructed by combining the baseline slope of the initial intensity growth. The environmental geological parameters include the range of environmental humidity fluctuations, historical deviation distribution patterns, solidification intensity sequence data, and geological complexity level. The time-varying interaction sequence is subjected to deep feature extraction using a temporal convolutional network, and the time-varying interaction sequence is decomposed into a time series to obtain the nonlinear evolution trajectory of the curing intensity. By performing second-order difference operations on the nonlinear evolution trajectory, critical nodes are identified, and based on the critical nodes, potential turning points in the nonlinear changes are determined.

5. The method for predicting the effect of roadbed solidification based on deep learning according to claim 3, characterized in that, The step of extracting soil moisture content parameters from the potential turning point interval and coupling and correcting the preliminary trend curve based on the soil moisture content parameters to generate an optimized strength growth curve includes: Moisture content parameters are extracted from the potential transition intervals and combined with the recorded ambient temperature values ​​to generate dynamic coupling parameters; Based on the dynamic coupling parameters, a dynamic coupling analysis is performed on the preliminary trend curve to establish a mapping relationship; Based on the mapping relationship, the initial trend curve is adjusted to generate an intensity growth optimization curve.

6. The method for predicting the effect of roadbed solidification based on deep learning according to claim 1, characterized in that, The step of performing simulation calculations based on the intensity growth optimization curve to obtain simulated intensity prediction values, calculating the deviation of the simulated intensity prediction values, and determining the number of iterations and correction coefficients required for deviation correction based on the deviation, includes: Based on the intensity growth optimization curve, multiple simulations are performed by introducing random perturbation variables to generate simulated intensity prediction values. The simulated intensity predictions are matched with historical intensity data, and the geological complexity level is quantified based on the deviation distribution pattern of the matching results. If the geological complexity level exceeds a preset complexity threshold, the dispersion of multiple sets of simulated intensity prediction values ​​relative to the central trend is calculated to obtain the deviation magnitude. Based on the deviation magnitude and in conjunction with the preset mapping relationship between the deviation magnitude and the correction parameters, the number of iterations required for deviation correction and the corresponding correction coefficient are determined.

7. The method for predicting the effect of roadbed solidification based on deep learning according to claim 1, characterized in that, The process involves correcting the simulated intensity prediction value based on the number of iterations and the correction coefficient to generate corrected intensity data. A corrected intensity distribution map is then constructed based on this data. The intensity values ​​in the corrected intensity distribution map are compared with a preset safety threshold. Based on the comparison results, a final evaluation report is generated, including: Based on the number of iterations and the correction coefficient, the simulated intensity prediction value is optimized to generate corrected intensity data, and a corrected intensity distribution map is constructed based on the corrected intensity data. Based on the corrected intensity distribution map, calculate the stability index of the simulated intensity prediction value; If the node strength values ​​of the modified strength distribution map are all within the preset safety threshold range, and the stability index is qualified, the mark meets the curing effect requirements, and the final evaluation report is output.

8. The method for predicting the effect of roadbed solidification based on deep learning according to claim 3, characterized in that, The step of generating final proportion adjustment parameters based on the final evaluation report and the revised intensity distribution map includes: Based on the final assessment report, combined with the corrected intensity distribution map and the material ratio values, a multidimensional correlation dataset for a specific level of geological complexity is constructed. The original curing agent content data is extracted from the multidimensional associated dataset, and a ratio adjustment matrix is ​​generated through optimization calculation. The ratio adjustment matrix is ​​input into a preset simulation curing model to update the simulated intensity prediction value and calculate a new stability index; If the dispersion in the new stability index is lower than that of the original stability index, the final ratio adjustment parameters are determined according to the ratio adjustment matrix; otherwise, the ratio adjustment matrix and subsequent stability verification steps are recalculated.

9. A deep learning-based system for predicting the effect of roadbed solidification, characterized in that, include: The data acquisition module is used to acquire soil parameters, analyze the matching relationship between the soil parameters and the curing agent, and generate an initial matching score. The trend analysis module is used to calculate the solidification reaction rate and perform trend analysis using a preset neural network perception model if the initial matching score is higher than the preset matching threshold, thereby determining the preliminary trend curve and the baseline slope of the initial intensity growth. The evolution analysis module is used to acquire environmental geological parameters, construct a time-varying interaction sequence in combination with the benchmark slope, extract time-series features and perform pattern analysis on the time-varying interaction sequence through a time-series convolutional network to obtain the evolution trajectory of solidification intensity, identify critical nodes based on the evolution trajectory, and determine potential turning intervals of nonlinear changes based on the critical nodes. The strength optimization module is used to extract soil moisture content parameters from the potential turning point, and to couple and correct the preliminary trend curve based on the soil moisture content parameters to generate an optimized strength growth curve. The deviation correction module is used to perform simulation calculations based on the intensity growth optimization curve to obtain the simulated intensity prediction value, calculate the deviation of the simulated intensity prediction value, and determine the number of iterations and correction coefficients required for deviation correction based on the deviation. The effect evaluation module is used to correct the simulated intensity prediction value according to the number of iterations and the correction coefficient, generate corrected intensity data, construct a corrected intensity distribution map based on the corrected intensity data, compare the intensity value in the corrected intensity distribution map with a preset safety threshold, and generate a final evaluation report based on the comparison results. The scheme output module is used to generate the final ratio adjustment parameters based on the final evaluation report and the modified intensity distribution map.

Images