Gravel pile construction parameter optimization method and system based on data mining

Abstract

The application provides a gravel pile construction parameter optimization method and system based on data mining, relates to the technical field of engineering optimization, and comprises the following steps: obtaining stratum information and mechanical parameters under a gravel pile construction position; performing historical data matching according to the stratum information to generate initial recommended parameters; obtaining a construction parameter sequence, and performing evaluation according to the stratum information and the construction parameter sequence; performing construction parameter optimization according to the evaluation result, the stratum information and the initial recommended parameters to obtain optimized construction parameters; and verifying the construction parameters and controlling according to the verified construction parameters, so that the reliability of gravel pile construction can be improved.

Description

Technical Field

[0001] This invention relates to the field of engineering optimization technology, and more specifically, to a method and system for optimizing construction parameters of crushed stone piles based on data mining. Background Technology

[0002] Crushed stone piles are a widely used reinforcement technology in soft soil foundation treatment. They form dense piles within the soil layer through vibration or impact, creating a composite foundation with the soil between the piles, significantly improving bearing capacity and shear strength. However, the quality of crushed stone pile construction is highly dependent on the rational selection of construction parameters. Traditional construction methods rely mainly on engineering experience or fixed process parameters, making them difficult to adapt to complex and variable geological conditions. Due to the significant randomness and spatial variability of underground soil layer distribution, fixed parameters often lead to reduced pile diameter and discontinuity in weak layers, while insufficient compaction or excessive energy consumption occurs in hard strata. Furthermore, the construction process is highly concealed, and quality inspection is delayed (usually requiring dynamic penetration testing or load tests several weeks after pile completion). Once quality problems arise, rework costs are extremely high. The sparse geological exploration drilling cannot provide continuous stratigraphic information, further exacerbating the blindness of parameter adjustments. In recent years, the Internet of Things (IoT) and data mining technologies have been gradually applied to the field of geotechnical engineering. However, existing solutions are mostly limited to the analysis of single data sources or post-hoc statistical regression, lacking the ability to dynamically perceive, make real-time decisions, and control the entire construction process, making it difficult to achieve the transition from "experience-driven" to "data-driven." Therefore, there is an urgent need for a method that can identify geological strata online and adaptively optimize construction parameters to improve the reliability of crushed stone pile construction. Summary of the Invention

[0003] In view of this, the present invention proposes a method and system for optimizing construction parameters of crushed stone piles based on data mining, so as to solve the problems existing in the prior art.

[0004] To achieve the above objectives, this invention proposes a data mining-based method for optimizing construction parameters of crushed stone piles, including: Obtain the geological information and mechanical parameters of the crushed stone pile construction location; perform historical data matching based on the geological information to generate initial recommended parameters; obtain the construction parameter sequence and evaluate it based on the geological information and construction parameter sequence; based on the evaluation results, optimize the construction parameters based on the geological information and initial recommended parameters to obtain optimized construction parameters; verify the construction parameters and control them based on the verified construction parameters.

[0005] Optionally, the process of acquiring the stratigraphic information includes: Acquire raw sensor data at the construction location of the crushed stone piles and extract feature vectors from the raw sensor data; acquire cluster data, match the feature vectors in the cluster data to obtain the corresponding cluster categories, where the cluster category is the soil layer type; integrate the soil layer type and depth to obtain the stratum information.

[0006] Optionally, the process of obtaining the clustering data includes: Historical feature data is obtained, and the historical feature data is clustered using an incremental fuzzy C-means clustering algorithm to obtain clustered data.

[0007] Optionally, the historical data matching process includes: Obtain the stratigraphic sequence of the exploration site and construct a construction knowledge graph; search the construction knowledge graph based on the stratigraphic sequence and stratigraphic information to obtain the corresponding construction parameter set, and correct it based on the construction parameter set to obtain the initial recommended parameters.

[0008] Optionally, the evaluation process based on stratigraphic information and construction parameter sequences includes: The geological information and construction parameter sequence are evaluated using a deep learning model to obtain a quality score and early warning signal.

[0009] Optionally, the process of optimizing construction parameters includes: The initial recommended parameters and construction parameter sequences are combined to form an initial population. The initial population is then optimized using optimization methods to obtain the optimized construction parameters.

[0010] Optionally, the construction parameters can be verified using a digital twin model, wherein the physical model in the digital twin model is determined based on the geological information and mechanical parameters.

[0011] On the other hand, the present invention provides a data mining-based system for optimizing the construction parameters of crushed stone piles, for performing the above-described method.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves real-time geological identification during drilling through multi-source sensing and incremental clustering algorithms, transforming traditional passive exploration into active perception. It can dynamically update the 3D geological model with centimeter-level resolution, providing accurate prior geological information for parameter adjustments and effectively avoiding the blind construction caused by incomplete or changing geological data. By constructing a construction knowledge graph and integrating a deep convolutional attention network for real-time quality prediction, pile quality shifts from "post-construction detection" to "process controllability," enabling early warning of potential problems such as diameter reduction and pile breakage during construction, significantly reducing rework rates and material waste. A multi-objective genetic algorithm is used to optimize four interdependent indicators—quality, efficiency, cost, and environment—outputting a Pareto-optimal parameter combination, which is then physically verified using a digital twin model. This ensures construction quality, improves efficiency and economic benefits, and reduces environmental impacts such as vibration and noise. The closed-loop control system incorporates Shewhart control charts for real-time monitoring of key parameters, automatically triggering re-optimization in case of anomalies. This forms a complete intelligent closed loop of "identification—recommendation—verification—execution—monitoring—re-optimization," significantly enhancing the system's adaptability and robustness to complex geological conditions. Finally, the continuous self-evolution of the knowledge graph and deep model allows the system to accumulate engineering experience during use. As the number of projects increases, the accuracy of initial parameter recommendations and the precision of quality predictions gradually improve, enabling the transferability and reusability of construction techniques. Overall, this system elevates crushed stone pile construction from manual operation relying on personal experience to a data-driven intelligent decision-making process, resulting in substantial progress in pile qualification rate, construction efficiency, material conservation, and environmental protection. Attached Figure Description

[0013] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a schematic diagram of the method flow in an embodiment of the present invention. Detailed Implementation

[0014] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0015] The core of this invention lies in constructing a closed-loop optimization scheme that integrates construction as exploration, dynamic adaptive knowledge graph reuse of parameters, and more. This solves the problem that traditional crushed stone pile construction relies on fixed parameters or manual experience, making it difficult to adapt to complex and changing geological conditions, resulting in unstable pile quality, material waste, and low efficiency.

[0016] To address these challenges, this solution proposes the following design principles: First, it transforms passive exploration into proactive perception, utilizing multi-source data such as current, torque, and vibration during drilling. An incremental fuzzy C-means clustering algorithm is used to identify geological interfaces in real time, enabling real-time identification of the drilling process and providing high-resolution geological information for parameter adjustments. Second, it shifts from experience-based decision-making to data-driven and knowledge-guided approaches. A historical construction knowledge graph is constructed, combining deep convolutional attention network quality prediction with a non-dominated sorting genetic algorithm with an elite strategy to automatically find the Pareto optimal parameter combination among quality, efficiency, cost, and environment. Third, it transforms open-loop control into a digital twin closed-loop system. Before parameter issuance, a simplified model based on the circular hole expansion theory is used to quickly simulate pile formation. During execution, Shewhart control charts are used to monitor key indicators in real time, automatically triggering re-optimization when anomalies occur, forming a closed loop of recommendation, verification, execution, monitoring, and re-optimization. The overall system architecture is divided into a perception layer, analysis layer, decision-making layer, and control layer, with edge and cloud collaboration to achieve intelligent construction with "one policy per pile, real-time adaptation."

[0017] This embodiment proposes a method and system for optimizing construction parameters of crushed stone piles based on data mining, such as... Figure 1 As shown, it includes: 1) Multi-source sensing and online formation identification: Continuously runs, outputting mechanical parameters and formation interface events at the current depth.

[0018] 2) Historical case matching and initial parameter recommendation: This is performed before construction begins or when new strata are exposed. Based on the stratigraphic sequence output in the first step, the knowledge graph is retrieved, and the initial construction parameter vector is output.

[0019] 3) Real-time prediction of construction process quality: During the construction process, it is triggered at fixed depth intervals, receives the stratum information and real-time construction parameter sequence of the first step, and outputs the quality score of the currently completed section; when the score is lower than the threshold, the fourth step of optimization is triggered.

[0020] 4) Multi-objective dynamic parameter optimization: Receives formation information from step 1, initial parameters from step 2, quality warnings from step 3, anomaly feedback from step 6, and verification rejection feedback from step 5, and outputs the optimized construction parameter vector. The optimization process employs a mutual exclusion mechanism, processing only one optimization request at a time. If proactive optimization and warning optimization occur simultaneously, the warning optimization takes priority.

[0021] 5) Digital twin verification: Receive the recommended parameters from step 4 and the current soil mechanical parameters from step 1, and output "pass" or "reject" and the deviation value; if rejected, return to step 4 to add physical constraints and re-optimize (set the maximum number of re-optimizations, and if exceeded, switch to manual intervention).

[0022] 6) Parameter issuance and closed-loop control: Execute the parameters passed in step 5 and issue instructions to the programmable logic controller; use Shewhart control charts to monitor the encrypted current during construction, and if an abnormality is detected, pause and return to step 4 for re-optimization.

[0023] Post-pile completion data closure: On-site test results are entered into the database, the knowledge graph is updated, and used for incremental model training in step 3. During initial construction, if there is no historical data to calculate control limits, a preset proportion of the equipment's rated current is used as a temporary limit. Once sufficient data is accumulated, it is switched to adaptive statistical control limits.

[0024] Specifically, for step 1) multi-source sensing and online stratum identification, the input source is raw sensor data. A Hall current sensor, depth encoder, hydraulic pressure sensor, three-axis MEMS accelerometer, and BeiDou positioning module are deployed on the vibratory shock absorber.

[0025] All sensor data are aggregated to the edge computing industrial control computer via fieldbus, and timestamps are aligned according to depth to extract and form a multidimensional feature vector for each depth point. This multidimensional feature vector includes the mean current, penetration velocity: the derivative of depth with respect to time, peak torque, the main frequency energy distribution extracted after the vibration signal is subjected to fast Fourier transform, and the hydraulic pressure fluctuation coefficient, etc.

[0026] Incremental fuzzy C-means clustering algorithm is used to achieve real-time stratigraphic identification. First, standard fuzzy C-means clustering is performed using the feature vectors of an initial set of depth points. The optimal initial number of clusters is then determined using the XB exponent.

[0027] Take the smallest VB value Used as the initial cluster number, the initial cluster centers are saved. And the membership matrix. Where c represents the number of clusters, and n represents the number of different soil layers identified. u ik `x` represents the fuzzy membership degree, indicating the degree to which the i-th depth point belongs to the k-th class. `m` represents the fuzziness index, controlling the degree of fuzziness in clustering, and is typically greater than 1. i The feature vector of the i-th depth point (including current, penetration velocity, vibrational frequency band energy, etc.). k Let represent the cluster center vector of the k-th class, which represents the typical characteristics of this type of soil layer. This represents the Euclidean distance, which measures the similarity between feature points and cluster centers. This represents the minimum distance between different cluster centers and is used to evaluate the separability of clusters.

[0028] Then for each new depth point Calculate its distance to existing cluster centers Weighted Euclidean distance Weights of each feature dimension It is learned offline from historical data by the ReliefF algorithm:

[0029] Where F represents the total dimension of the feature vector, and f represents the feature label.

[0030] Dynamic threshold The system adaptively adjusts based on the mean of the average distance within each category (e.g., taking a multiple of the average distance within each category). minimum distance If the point is located in the nearest category, then the soil layer category of the corresponding cluster is determined.

[0031] When adding new data to a clustering method, the new data is treated as new depth points, and a learning rate that decays as the total number of samples assigned to that class increases is used. Update cluster center :

[0032] This represents the old cluster centers. Among them... Exponential decay or reciprocal decay strategies can be used. If Then create a new category and initialize the cluster center as . Number of categories Increasing the number of categories is used when adding a large amount of historical data. If real-time data cannot meet the requirements for clustering into historical data, then proceed with the regular construction until the above requirements are met.

[0033] Simultaneously, a fixed-length sliding window is set to continuously track the classification of several recent depth points. Continuous data over a period of time is used to determine whether the current stratum is stable. If it is stable, subsequent optimization content is generated; otherwise, no stratum interface event is generated, and no further optimization is performed, until the fixed-length window is determined to be a stable stratum, at which point further operations are initiated. When the proportion of points belonging to the same newly created category within the window exceeds a set percentage, and the Mahalanobis distance between the center of the newly created category and the center of the previous stable category is... If the value exceeds the predicted separation threshold, it is determined to be a formation interface.

[0034] in This is the combined covariance matrix of the two classes of samples.

[0035] After determination, a stratum interface event is generated, including the depth value and soil layer type (soil layer name). The mechanical parameters (elastic modulus) of the soil layer type at the current depth, mapped from the built-in geotechnical database, are output. Undrained shear strength (If real-time data from static cone penetration testing or pressure gauges is available on-site, the measured values ​​will be used first), and ground interface events (depth, soil type). These outputs are pushed to steps 2, 3, and 4 in real time.

[0036] Specifically, regarding step 2) historical case matching and initial parameter recommendation, the input sources are the exploration stratigraphic sequence obtained for the new terrain and the stratigraphic interface events identified in real time in step 1. The exploration stratigraphic sequence includes: stratigraphic stratification order, basic physical properties of the soil, mechanical parameters, groundwater information, and other key engineering properties such as permeability coefficient, sensitivity, particle size distribution curve, and organic matter content. These parameters affect the compaction effect of the vibratory compactor and the quality of pile formation.

[0037] Construct and maintain a construction knowledge graph based on a graph database (such as Neo4j). The nodes in the graph include: geological units (storing stratigraphic sequence strings, SPT blow count ranges, water content intervals, and particle size distribution descriptions), equipment models (vibratory compactor power, excitation force parameters), construction parameter sets (refinement current threshold, single filler volume, vibration dwell time, number of back-interpolation attempts, and back-interpolation depth), and test results (dynamic penetration correction values, load test settlement, and pile diameter uniformity index).

[0038] Relationships such as "contains," "uses," and "generates" are defined between nodes, and similar edges between different geological sequences are pre-calculated. When a new project requires recommendations for initial construction parameters, the user inputs a layered geological sequence (or it is automatically identified in the first step). The graph query engine first performs index filtering based on the type of the first layer of soil, then performs a dynamic time warping algorithm on the candidate cases to eliminate differences in stratum thickness between different cases, aligning the two sequences to nodes of the same depth.

[0039] in The function is determined by combining the semantic similarity of soil layer names (obtained from the geotechnical terminology system) and the normalized value of the difference in standard penetration test blow counts. : The name or attribute of the i-th layer after the current data is aligned. : The name or attribute of the j-th layer after historical case alignment. dist( ): The dissimilarity function between two strata, combined with semantics and the difference in SPT blow counts. min( ): Take the minimum value of adjacent subproblems, corresponding to three different alignment paths: Align the i-th layer of the new sequence with the j-th layer of the case sequence, and allow the new sequence to have thickness redundancy. Align the i-th layer of the new sequence with the j-th layer of the case sequence, and allow the case sequence to have thickness redundancy. Align the current layers of the two sequences directly, and advance each one layer forward. This recursion starts at i=1, j=1, and gradually fills a matrix, ultimately... The value represents the total cost of the best match between the two stratigraphic sequences (the smaller the value, the more similar the sequences). Through this "bent" alignment method, DTW can effectively handle thickness variations or local gaps in the same stratigraphic sequence across different boreholes, making subsequent weighted cosine similarity calculations more reasonable.

[0040] After alignment, calculate the weighted cosine similarity and depth weight. Using exponential decay (maximum weight at the surface, decreasing downwards):

[0041] in This represents the total number of strata after alignment. For the first The matching coefficient of the layer (1 for a perfect match, and reduced for a partial match based on the attribute distance). This represents the depth weight corresponding to the i-th layer.

[0042] The system selects several historical cases with high similarity rankings and extracts their corresponding complete construction parameter sets. This corresponds to the encryption current, filler quantity, vibration dwell time, number of reverse insertions, and reverse insertion depth.

[0043] Then, a pre-trained Gaussian process regression model is invoked. The input to this model is the difference vector of formation mechanical parameters between the current project and the case. (Such as differences in SPT blow count, moisture content, and compression modulus), the output is the correction amount for each construction parameter. The kernel function uses a squared exponent kernel. :

[0044] Hyperparameters It is learned from historical data. represents the Kronecker function, which is 1 when i=j and 0 otherwise; Measure the similarity between two difference vectors. It represents the signal variance and controls the amplitude of the kernel function output. This represents the length scale parameter, which controls the rate at which the correlation of the difference vector decays with distance. Noise variance.

[0045] When making predictions, given a new difference vector Calculate the posterior mean of the parameter corrections:

[0046] in The kernel vectors of the new sample and the training sample. The covariance matrix between training samples, Let I be the observed parameter offset values ​​of the training samples (i.e., the difference between the actual optimal parameters and the case parameters); I is the identity matrix. y represents the observed parameter offset values ​​(the difference between the actual optimal parameters and the case parameters) in the training samples.

[0047] The recommended final initial parameter values ​​are If a parameter exceeds a preset physical constraint (e.g., exceeds the equipment's rated value), it is truncated to the boundary. The output is the initial construction parameter vector. The population is fed into the fourth step as the initial population center for the genetic algorithm, or it can be used directly for trial planting, waiting for the third step to trigger optimization.

[0048] Specifically, for step 3), real-time prediction of construction process quality. The input sources are the soil layer type and mechanical parameter range at the current depth output in step 1, and the parameter sequence P (current, filler volume, vibration retention time, number of back-interpolations and depth, etc.) collected in real time during construction.

[0049] A deep convolutional attention network was used as the quality prediction model. The network was pre-trained in the cloud using a large amount of historical construction sequences and their corresponding subsequent field inspection data (dynamic penetration test blow count, pile diameter variation coefficient).

[0050] The network structure, from input to output, is as follows: a first layer of one-dimensional convolution (followed by batch normalization, ReLU activation function, and max pooling), a second layer of one-dimensional convolution (also followed by batch normalization, ReLU, and max pooling), followed by a spatial attention module. A fully connected layer calculates scalar weights for the convolutional features at each depth location, and the weights at all locations are normalized using Softmax. The weighted features are then summed to obtain a global feature vector. Finally, this vector is mapped to a single output value through several fully connected layers (reducing the number of neurons layer by layer), and converted to a quality score (between 0 and 1) using the Sigmoid function. The training dataset consists of paired construction sequences and subsequent detection results. Labels are normalized to the 0-1 range based on the ratio of measured metrics to design requirements. The loss function used is binary cross-entropy. Real-time early warning mechanism: During construction, after each fixed length (e.g., every 0.5 meters) of drilling or densification is completed, a fixed length window (e.g., the most recent few meters range) is extracted from the real-time parameter sequence along the depth direction. The window contains multi-dimensional features of each sampling point within that depth range (current change curve, cumulative filler volume, vibration time distribution, reverse insertion position marker, etc.).

[0051] The window data is input into the network, which automatically performs forward calculations and outputs a quality score for the currently completed segment. The system continuously monitors the trend of this score: if the score falls below the preset acceptable threshold multiple times consecutively, an alert is triggered, and the alert information, along with the current stratigraphic identification results, construction parameter snapshots, and the depth location with the highest attention weight (problem segment), is sent to step 4 of the optimization process. The final output is the quality score. And warning signals (including the location of the problematic section) trigger the fourth step of optimization.

[0052] Specifically, for step 4), multi-objective dynamic parameter optimization, the input sources include the latest formation identification results (current depth and formation interface events) provided in step 1, and the initial parameters recommended in step 2. Step 3: Real-time acquisition of construction parameter sequences The transmitted quality warning information (including the location of the problem section), the Shewhart control chart anomaly records from step 6 (such as current offset trends), and the physical constraint violations (virtual pile diameter deviation) from step 5 are also included. ).

[0053] The decision variables are a vector of construction parameters. , representing the encryption current threshold, single filler quantity, vibration dwell time, number of inversions, and inversion depth, respectively. Real number encoding is used, with integer variables rounded after cross-mutation. Each variable has a defined physical range (e.g., current not exceeding the equipment's rated value, filler quantity limited by hopper capacity, etc.).

[0054] Simultaneously optimize four objective functions: quality objective When calling the deep convolutional attention network in step three for prediction, the parameter vector needs to be... Convert it into a virtual feature sequence. The conversion rules are as follows: based on the encryption current... Set the current value in the characteristic sequence (keeping it constant); based on the oscillation time... Set the high current duration; based on the number of reverse insertions. and inversion depth Add an inversion marker at the corresponding depth position of the sequence; filler volume The cumulative change rate of the filler material in the sequence is affected.

[0055] efficiency goals The estimated total time required to complete this section of construction includes the time for lowering the penetration point, filling the gap, densification, and reverse insertion. The duration of each action is given by an empirical regression model. Economic Objectives This includes the cost of crushed stone materials (positively correlated with the total amount of filler) and the cost of electricity consumption (the integral of current over time multiplied by the electricity price). Environmental goals. The peak ground vibration velocity (PDV) prediction model is composed of a weighted sum of peak ground vibration velocity and equivalent noise level. It can be constructed using regression trees or empirical formulas. The inputs are densified current, the corresponding stratum stiffness and other physical parameters in the stratum information, and distance.

[0056] The optimization algorithm employs a non-dominated sorting genetic algorithm with an elitist strategy (NSGA-II). The dominance relationship is defined as follows: if an individual is superior to or equal to another individual in all target metrics, and at least better than the other individual, then the solution is considered complete. Dominate If and only if for all targets have And there is at least one target. Make .

[0057] The specific process includes: Population initialization: Randomly generate a set of individuals (each individual represents a set of construction parameters). Non-dominated sorting: Divide the population into multiple Pareto fronts (layers) based on the objective function values ​​of all individuals. The first layer consists of individuals not dominated by any other individuals; the second layer consists of individuals dominated only by individuals in the first layer, and so on. Crowding distance calculation: For each individual within the same front, calculate the density (crowding distance) of its surrounding individuals. The larger the distance, the sparser the solutions near that individual, in order to maintain diversity. Selection, crossover, and mutation: Select the parent generation through tournament selection (prioritizing individuals with lower front layers and larger crowding distances), and generate the offspring population using simulated binary crossover. Elite retention: Merge the parent and offspring generations, perform non-dominated sorting and crowding distance calculation again, and take the top N individuals as the new generation population. Iterative convergence: Repeat the above steps until the preset number of generations or target convergence is reached.

[0058] The initial population was constructed using the construction parameter vector collected in real time in step 3. A random perturbation is performed around the center to generate a set number of individuals; simultaneously, the construction parameters obtained in step 2 are... As a supplement, construction parameters from historical cases that are highly similar to the current geology can be extracted from the knowledge graph and directly inserted into the initial population as elite individuals to accelerate convergence.

[0059] Genetic operator details: Selection uses tournament selection, randomly selecting several individuals each time, and choosing the best one to enter the mating pool. Crossover uses simulated binary crossover; for two parent individuals, crossover is determined by the crossover probability, resulting in two offspring. Mutation uses multinomial mutation. For each variable, mutation is determined by the mutation probability, and the mutation magnitude is controlled by a multinomial distribution.

[0060] Elite preservation strategy: In each generation, the parent and offspring are merged, and a fast non-dominated sort is performed. New populations are then filled in hierarchically, and within the same hierarchical level, crowding distance is used to preserve diversity. After iteration to convergence (such as reaching a preset number of generations or when the fitness change is small), a set of Pareto non-dominated solutions is obtained, including the Pareto front, which is the subsequent multiple dominant subsets.

[0061] Multi-attribute decision-making (FAHP+TOPSIS): To select the final recommended parameters from the Pareto front, the system integrates a multi-attribute decision-making method. First, fuzzy hierarchical analysis is used to compare each objective pairwise, constructing a fuzzy judgment matrix and calculating the weight vector of each objective. And perform a consistency check.

[0062] Then, the Top-Optimal Solution Ranking (TOPSIS) method is used: the distance from each solution in the Pareto front to the positive ideal solution (the virtual solution where all objectives are optimal) is calculated. Distance to the negative ideal solution (the worst-case virtual solution for all objectives) The relative closeness is :

[0063] Final choice The largest solution is used as the recommendation parameter. .

[0064] Re-optimization limit: If step 5 fails the verification, the system returns to re-optimization, recording the number of re-optimizations. If the system still fails to pass the verification after reaching the preset maximum number of re-optimizations (e.g., 3 times), automatic optimization will terminate and the system will switch to manual intervention mode, where engineers will manually select parameters based on system suggestions. The output will show recommended parameters. Then, it is sent to step 5 for digital twin verification.

[0065] Specifically, regarding step 5) digital twin verification, the input source is the recommended parameters output from step 4. And the soil mechanical parameters (elastic modulus) at the current depth provided in step 1. Undrained shear strength wait).

[0066] Before the recommended parameters are officially issued to the implementing agency, a mandatory "virtual construction" verification is conducted using a lightweight physical model simulator. This simulator is based on the theory of circular hole expansion and has been simplified for the vibration-based compaction process of crushed stone piles. Key empirical coefficients in the model (such as equipment conversion coefficients) are used in the simulation. Hole expansion index Calibration is required for specific vibratory impactors.

[0067] Calibration process: In a uniform stratum, test piles were constructed according to multiple sets of different parameters, and the pile diameter was measured. The least squares method was used to fit the model parameters to minimize the sum of squared errors between the predicted diameter and the measured diameter.

[0068] The simulator will use the encrypted current recommended in step four. Converted to equivalent radial expansion pressure:

[0069] Then, substitute the simplified borehole enlargement formula to estimate the pile diameter:

[0070] in The initial diameter of the casing. For reference current, The empirical index obtained from calibration is used to calculate the predicted diameter and the design pile diameter. Relative deviation:

[0071] Meanwhile, a small perturbation is made near the nominal value of the parameter (e.g., each parameter changes independently by a certain positive or negative proportion), and the upper and lower bounds of the predicted diameter are calculated to obtain the confidence interval width.

[0072] like If the percentage exceeds a preset threshold (e.g., 15%) or the confidence interval width exceeds the design allowable range, the simulator determines that the reliability of the set of parameters is insufficient and refuses to distribute them.

[0073] Output and Feedback: If the validation passes, Marked as "verified", proceed to step 6; if verification fails, generate a rejection signal and attach the absolute value of the relative deviation. Return to step 4. Step 4 adds a penalty term to the objective function. , This indicates the preset maximum deviation. This represents the penalty coefficient, which causes solutions that deviate significantly from the design pile diameter to be gradually eliminated in the non-dominated sorting process, and a new search is conducted for parameter combinations that satisfy physical feasibility.

[0074] Specifically, regarding step 6) parameter distribution and closed-loop control, the input source is the parameters that passed verification in step 5. And the current depth and actual value of encrypted current updated in real time in step 1.

[0075] The optimized construction parameters are sent to the field programmable logic controller (PLC) or vehicle-mounted human-machine interface (HMI) via an Industrial Internet of Things (IIoT) protocol. The PLC automatically adjusts the current setpoint of the vibratory compactor, the start-stop time of the vibratory compactor, the valve opening of the automatic packing system, and the sinking depth and number of insertions of the reverse insertion mechanism based on the received parameters. The control program includes a PID controller: the deviation between the actual current and the set encrypted current is calculated using proportional-integral (PI) and output to the frequency converter to control the speed of the vibratory compactor motor. Simultaneously, statistical process control is initiated at the edge to monitor the critical encrypted current in real time. This module dynamically calculates the control limits of the Shewhart control chart. Centerline Take the moving average value of the current in the current pile's completed densification section. The length of the moving window can be adaptively adjusted according to the smoothness of construction (lengthen the window if the process is stable, and shorten it otherwise).

[0076] Upper limit of control chart and lower limit Based on the 3σ principle:

[0077] coefficient Take a typical value (e.g., 3). The online estimate of the standard deviation is obtained using the moving range method:

[0078] in It is an exponentially weighted moving average of the moving range (the absolute value of the difference between two adjacent current segments). It is a constant related to the subgroup size, i.e., the number of observations in each sample group.

[0079] Initial control limit handling: In the early stages of construction, when the number of completed densification sections is insufficient to calculate the statistical control limit, the system uses a preset proportion of the equipment's rated current (e.g., ± a certain percentage) as a temporary control limit. Once sufficient data is accumulated (e.g., after completing 3 sections), it automatically switches to the aforementioned adaptive statistical control limit.

[0080] After each actual current sampling value is marked on the control chart, abnormal patterns are detected, including: a single point exceeding the control limit, several consecutive points all located on the same side of the center line (trend deviation), several consecutive points monotonically rising or falling (trend drift), and the slope of the fitted straight line being too large (trend strength exceeding the standard).

[0081] Once any abnormal pattern is detected, the system immediately issues an audible and visual alarm on the operation interface and sends a pause command to the programmable logic controller to automatically interrupt the current encryption action; at the same time, it returns the abnormal information (current current value, trend type) to step 4 to trigger re-optimization.

[0082] On-site operators can view the system's recommended new parameters via a touchscreen, or choose to manually confirm and continue construction, or switch entirely to manual operation mode. The control panel includes an emergency stop button and a "manual priority" switch. When the switch is set to manual, the programmable logic controller ignores all automatic parameter updates, but the system continues to record data and display recommendations. Once the switch is restored to automatic, it prompts whether to use the cached recommended parameters, ensuring that personnel have the highest decision-making authority in any abnormal situation. The output is the executed action, and it records actual construction data (including current, depth, time, etc.) for subsequent model updates.

[0083] Finally, there is the data closure loop and system self-evolution. After the piles are completed, the field test results (dynamic penetration test, load test) are input into the system to update the knowledge graph: new case nodes (including geological sequence, equipment, construction parameters, and test results) are added, and similar edges are recalculated.

[0084] Simultaneously, the deep convolutional attention network is optimized, and incremental training with new data (e.g., several rounds of gradient descent with small batch sizes) is performed to update the weights of the edge models. As the number of construction projects increases, the initial recommendation accuracy and quality prediction capabilities of the system continuously improve, achieving a self-evolving capability that improves with use.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for optimizing construction parameters of crushed stone piles based on data mining, characterized in that, include: Obtain the geological information and mechanical parameters of the crushed stone pile construction location; perform historical data matching based on the geological information to generate initial recommended parameters; obtain the construction parameter sequence and evaluate it based on the geological information and construction parameter sequence; based on the evaluation results, optimize the construction parameters based on the geological information and initial recommended parameters to obtain optimized construction parameters; verify the construction parameters and control them based on the verified construction parameters.

2. The method according to claim 1, characterized in that, The process of acquiring the stratigraphic information includes: Acquire raw sensor data at the construction location of the crushed stone piles and extract feature vectors from the raw sensor data; acquire cluster data, match the feature vectors in the cluster data to obtain the corresponding cluster categories, where the cluster category is the soil layer type; integrate the soil layer type and depth to obtain the stratum information.

3. The method according to claim 2, characterized in that, The process of obtaining the clustering data includes: Historical feature data is obtained, and the historical feature data is clustered using an incremental fuzzy C-means clustering algorithm to obtain clustered data.

4. The method according to claim 1, characterized in that, The process of historical data matching includes: Obtain the stratigraphic sequence of the exploration site and construct a construction knowledge graph; search the construction knowledge graph based on the stratigraphic sequence and stratigraphic information to obtain the corresponding construction parameter set, and correct it based on the construction parameter set to obtain the initial recommended parameters.

5. The method according to claim 1, characterized in that, The evaluation process based on stratigraphic information and construction parameter sequences includes: The geological information and construction parameter sequence are evaluated using a deep learning model to obtain a quality score and early warning signal.

6. The method according to claim 1, characterized in that, The process of optimizing construction parameters includes: The initial recommended parameters and construction parameter sequences are combined to form an initial population. The initial population is then optimized using optimization methods to obtain the optimized construction parameters.

7. The method according to claim 1, characterized in that, The construction parameters are verified using a digital twin model, in which the physical model is determined based on geological information and mechanical parameters.

8. A data mining-based optimization system for crushed stone pile construction parameters, characterized in that, Used to perform the method described in any one of claims 1-7.

Images