A tailing pond settlement prediction method and system based on physical mechanism and improved SVR algorithm

Abstract

The present application belongs to the technical field of geotechnical engineering safety monitoring and disaster early warning, and particularly relates to a tailing pond settlement prediction method and system based on fusion of physical mechanism and improved SVR algorithm, which comprises the following steps: collecting and preprocessing engineering geological data of a tailing pond to be measured; constructing a "gray box" prediction model based on physical mechanism and improved SVR algorithm; and obtaining tailing pond settlement prediction results based on the preprocessed engineering geological data and the "gray box" prediction model. The present application effectively solves the problems of difficulty in obtaining parameters of traditional physical models and poor generalization ability of single machine learning models under small samples, and significantly improves the precision and robustness of tailing pond settlement prediction under complex geological conditions.

Description

Technical Field

[0001] This invention belongs to the field of geotechnical engineering safety monitoring and disaster early warning technology, specifically involving a tailings dam settlement prediction method and system based on the fusion of physical mechanisms and improved SVR algorithm for the settlement deformation of deep and weak mud layers in waterlogged areas during the closure and operation of large mine tailings dams. Background Technology

[0002] The settling process of tailings ponds is an extremely complex fluid-solid-chemical coupling process. Unlike typical natural soft soil, tailings mud is deposited under hydraulic separation, exhibiting significant sorting and stratification. Currently, the widely used settlement calculation method in engineering design is the layered summation method based on Terzaghi's one-dimensional consolidation theory. This method suffers from drawbacks such as difficulty in parameter selection, neglect of rheology, underestimation of settlement, and high computational cost. While existing physical models can provide a general trend of settlement, they often have significant deviations in quantitative accuracy. Especially for deep and weak tailings, physical calculations are often used as a "lower limit" reference and cannot be directly used to guide refined construction. To overcome the difficulties in obtaining physical model parameters and the numerous simplifying assumptions, researchers have attempted to introduce artificial intelligence algorithms.

[0003] Backpropagation (BP) neural networks were among the earliest algorithms applied to settling prediction. By simulating the connections of neurons in the human brain, they possess powerful nonlinear mapping capabilities. However, in tailings dam settling prediction, BP networks have revealed significant drawbacks: 1. Poor overfitting and generalization: The number of monitoring points at tailings dam sites is limited, with the effective sample size typically ranging from tens to hundreds, making it a typical "small sample" problem. BP networks are prone to overfitting under small sample conditions, leading to significant prediction biases for new data. 2. Vanishing gradients and local minima: Gradient descent-based training algorithms are prone to getting trapped in local minima, and the training speed is greatly affected by initial weights, resulting in unstable results. Support vector regression (SVR), based on the principle of structural risk minimization in statistical learning theory, is theoretically more suitable for small sample regression problems than neural networks, exhibiting better robustness. However, SVR performance is highly dependent on the selection of two core parameters: the penalty factor (C) and the kernel function parameter (γ). Current technologies typically employ grid search, genetic algorithms (GA), or standard particle swarm optimization (PSO) to find the optimal parameters.

[0004] Current technologies either lean towards pure physical calculations or pure data mining, lacking a fusion model that effectively embeds prior knowledge of soil mechanics into machine learning algorithms. Particularly in the specific scenario of tailings ponds, how to use physically calculated values ​​to "guide" machine learning models, enabling them to focus on learning "nonlinear residuals" that cannot be explained by physical formulas, is a pressing technical challenge. Furthermore, while existing technologies have improved the PSO algorithm's premature convergence problem in SVR parameter optimization (e.g., by introducing mutation operations), few in-depth improvement schemes combining chaos theory and adaptive inertia weights have been applied to geotechnical settlement prediction.

[0005] Specifically, the advantage of a pure physical model (hierarchical summation method) lies in its clear mechanical mechanism and physical boundaries, providing a stable trend output that conforms to the laws of gravity and compressibility. However, its accuracy is limited by parameter simplification and the inability to quantify complex time-dependent deformations. The advantage of a pure data-driven model (SVR) lies in its ability to capture complex nonlinear characteristics from monitoring data, but it lacks physical constraints and is prone to distortion in small sample sizes or with high data noise. The two models are naturally complementary in their technical approaches: the physical model provides a physical benchmark for the main body's settlement, while SVR can compensate for nonlinear residuals that the physical model cannot explain. However, no tailings dam settlement prediction scheme that deeply integrates the two has yet been found in the current technology. How to construct a "gray box" prediction framework that maintains the interpretability of the physical mechanism while utilizing machine learning to compensate for the insufficient accuracy of the physical model is a pressing technical challenge in this field.

[0006] In conclusion, developing a settlement prediction method that integrates physical mechanisms and improved optimization algorithms has significant engineering value and scientific significance for improving the safety monitoring level of tailings dams. Summary of the Invention

[0007] To address the following issues in existing technologies: 1. Insufficient accuracy of physical models: Traditional layered summation methods cannot quantify the rheological and lateral extrusion effects of tailings sludge, leading to underestimation of predicted values; 2. Poor interpretability of black-box models: Pure machine learning models lack physical constraints, and prediction results are easily distorted and unreliable when data noise is high or operating conditions change abruptly; 3. Difficulty in optimizing algorithm parameters: The parameters of standard SVR models are difficult to determine, and conventional optimization algorithms (such as PSO) are prone to getting trapped in local optima, failing to maximize the model's performance; This invention provides a tailings dam settlement prediction method and system based on the fusion of physical mechanisms and improved SVR algorithms. By constructing a gray-box model architecture of "physical benchmark + intelligent residual" and using an improved particle swarm optimization algorithm (IPSO) with Logistic-Tent composite chaotic mapping and adaptive weights to optimize SVR parameters, high-precision, robust, and physically meaningful settlement prediction is achieved.

[0008] To achieve the above objectives, the present invention provides the following solution: A tailings dam settlement prediction method based on the fusion of physical mechanisms and an improved SVR algorithm, the method comprising: Collect and preprocess engineering geological data of the tailings dam to be tested; A "grey box" prediction model is constructed based on physical mechanisms and an improved SVR algorithm; Based on the preprocessed engineering geological data and the "gray box" prediction model, the tailings dam settlement prediction results were obtained.

[0009] Preferably, the "gray box" prediction model includes: a physical basis settlement calculation module, a residual correction module, and a fusion output module; The physical basis settlement calculation module is used to construct a benchmark settlement calculation model based on the physical mechanism of the hierarchical summation method and obtain the predicted physical settlement value. The residual correction module is used to construct a residual correction model based on the improved SVR algorithm and obtain the SVR settlement prediction value. The fusion output module is used to fuse the physical settlement prediction value and the SVR settlement prediction value to obtain the final settlement prediction result.

[0010] Preferably, the method for obtaining the predicted physical settlement value includes: ; in, This is the predicted value of physical settlement. Empirical coefficients for settlement calculation. The initial void ratio under self-weight stress. The void ratio after compression under total stress. For layer thickness, For the number of base levels, This is the number of the current computing layer.

[0011] Preferred methods for constructing residual correction models based on improved SVR algorithms include: An improved SVR algorithm is obtained by adopting a Logistic-Tent composite chaotic mapping initialization strategy, a nonlinear adaptive inertia weight and learning factor coordinated adjustment strategy, and an adaptive mutation and reverse learning perturbation strategy based on fitness variance. An improved SVR algorithm is used to optimize the parameters of the SVR model and obtain the optimal parameter combination. An SVR model is constructed and trained based on the optimal parameter combination to obtain a residual correction model.

[0012] Preferably, the nonlinear adaptive inertia weight and learning factor co-adjustment strategy includes: Formula for calculating nonlinear adaptive inertia weight: ; in, For the first Inertia weights during iterations This represents the maximum value of the inertia weight. This represents the minimum inertial weight. Let the current iteration algebra be... The maximum number of iterations; Learning factor adaptive adjustment strategies include: ; ; in, For the first The cognitive learning factor of the generation controls the weight of the particle learning to its own historical best position; For the first The social learning factor of the generation controls the weight of particles learning to the global optimal position of the group; , These are the initial and final values ​​of the cognitive learning factor, respectively. , These are the starting and ending values ​​for the social learning factor, respectively.

[0013] Preferred adaptive mutation and reverse learning perturbation strategies based on fitness variance include: The formula for calculating the variance of population fitness is: ; in, For population fitness variance; Population size; For the first The fitness value of each particle; This represents the average fitness value of the population. when hour, To trigger the mutation perturbation mechanism by setting a preset diversity threshold: ; in, For the first The particle in the first Dimensional position value, For variable asynchronous long coefficients, an adaptive decay strategy is adopted. This causes the variation amplitude to gradually decrease as the iteration progresses; These are random numbers distributed according to a standard normal distribution. The initial variable asynchronous length; The attenuation coefficient; When a particle's position exceeds the search boundary, a reverse learning strategy is used for "rebirth": ; In the formula, The position of the reborn particle; The original position of the particle that has exceeded the boundary; , For the first The lower and upper bounds of the dimensional search space.

[0014] The present invention also provides a tailings dam settling prediction system based on the fusion of physical mechanisms and improved SVR algorithm. The system is used to implement the aforementioned method and includes: an acquisition and processing module, a model building module, and a prediction module. The data acquisition and processing module is used to acquire and preprocess the engineering geological data of the tailings dam to be tested. The model building module is used to build a "grey box" prediction model based on physical mechanisms and an improved SVR algorithm; The prediction module is used to obtain tailings dam settlement prediction results based on preprocessed engineering geological data and the "gray box" prediction model.

[0015] Preferably, the model building module includes: a physical basis settlement calculation unit, a residual correction unit, and a fusion output unit; The physical basis settlement calculation unit is used to construct a benchmark settlement calculation model based on the physical mechanism of the hierarchical summation method and obtain the predicted value of physical settlement. The residual correction unit is used to construct a residual correction model based on the improved SVR algorithm and obtain the SVR settlement prediction value. The fusion output unit is used to fuse the physical settlement prediction value and the SVR settlement prediction value to obtain the final settlement prediction result.

[0016] Preferably, the process of constructing a residual correction model based on the improved SVR algorithm includes: An improved SVR algorithm is obtained by adopting a Logistic-Tent composite chaotic mapping initialization strategy, a nonlinear adaptive inertia weight and learning factor coordinated adjustment strategy, and an adaptive mutation and reverse learning perturbation strategy based on fitness variance. An improved SVR algorithm is used to optimize the parameters of the SVR model and obtain the optimal parameter combination. An SVR model is constructed and trained based on the optimal parameter combination to obtain a residual correction model.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Prediction accuracy has been significantly improved. Experimental data show that the mean squared error (MSE) of the "grey box" prediction model of this invention on the test set is much lower than that of single physical computation or traditional SVR models. The IPSO algorithm initialized with a Logistic-Tent composite chaotic map can find the globally optimal parameters of SVR with a higher probability, resulting in a higher R-value for residual prediction. 2 The accuracy has been improved to over 0.96. The nonlinear adaptive weighting strategy results in higher convergence accuracy and effectively captures minute nonlinear changes during the settling process.

[0018] 2. Extremely strong generalization ability and robustness At tailings dam sites, borehole data and monitoring samples are often scarce (small sample conditions). Pure neural network models are prone to overfitting under such conditions. This invention utilizes a physical mechanism-based benchmark settlement calculation model to provide a stable "benchmark constraint," while the SVR algorithm itself, based on structural risk minimization, has a natural advantage for small samples. The combination of these two ensures the generalization ability of the "gray box" prediction model under different geological units and loading rates, avoiding prediction distortion caused by data noise.

[0019] 3. Clear physical meaning and high engineering safety. The output of this invention always includes physical calculation components, which ensures that the prediction results conform to basic soil mechanics laws (such as monotonic settlement and convergence trend). Even if the machine learning module fails due to data anomalies, the physics module can still provide a minimum settlement estimate, avoiding the "catastrophic forgetting" or predictions that violate physical common sense (such as predicting settlement rebound) that may occur with pure black box models, thus providing a higher safety redundancy for engineering decisions.

[0020] 4. The algorithm has a fast convergence speed and strong global optimization ability. This invention employs a Logistic-Tent composite chaotic mapping initialization strategy, effectively avoiding the periodic degradation problem of a single chaotic mapping in the later stages of iteration. This significantly increases the diversity of the particle swarm and avoids blind random search in the early stages of iteration. Combined with a nonlinear adaptive weight and learning factor co-adjustment strategy, and an adaptive mutation and reverse learning perturbation strategy based on fitness variance, the algorithm can effectively escape local extremum traps, greatly enhancing its global optimization capability. Experimental results show that the improved IPSO algorithm reduces the number of iterations to reach the optimal solution by approximately 40%-50% compared to the traditional PSO algorithm, while achieving higher convergence accuracy and significantly shortening model training time, thus meeting the needs of real-time dynamic inversion in engineering projects.

[0021] 5. Wide range of applications The framework proposed in this invention is universal. Although the embodiments are aimed at tailings dam settlement, by adjusting the calculation formula of the physical layer (such as adapting to the consolidation theory of different soil types), this method can be widely applied to various geotechnical engineering deformation prediction fields such as coastal reclamation, deep foundation pit excavation, and roadbed settlement, and has broad prospects for promotion and application. Attached Figure Description

[0022] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram illustrating the complete workflow of the tailings dam settling prediction method according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the specific operation flow of data preprocessing and feature engineering in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the construction process of the improved particle swarm optimization algorithm for support vector regression (IPSO-SVR) residual correction model in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the fusion process of prediction results from the physical mechanism model and the data-driven model in an embodiment of the present invention. Detailed Implementation

[0024] 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.

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] Example 1 This invention provides a tailings dam settlement prediction method based on the fusion of physical mechanisms and an improved SVR algorithm, comprising: Collect and preprocess engineering geological data of the tailings dam to be tested; A "grey box" prediction model is constructed based on physical mechanisms and an improved SVR algorithm; Based on the preprocessed engineering geological data and the "gray box" prediction model, the tailings dam settlement prediction results were obtained.

[0027] like Figures 1-4 As shown, the specific implementation process of the present invention is as follows: This invention addresses the problems in existing tailings dam closure pre-closure water accumulation area treatment projects, such as the complex settlement and deformation mechanism of deep and weak tailings mud layers during graded ore discharge and silt squeezing, the scarcity of on-site borehole exploration samples, and the insufficient accuracy of traditional calculation methods. It proposes a "gray box" prediction model that integrates soil mechanics physical mechanisms and improved machine learning algorithms.

[0028] The application scenario of this embodiment is a large-scale tailings dam closure and remediation project. This project involves graded ore discharge and silt removal in waterlogged areas and the enhancement of the bearing capacity of the tailings beach. The geological environment at the site is complex, the tailings particles on the sedimentary beach are extremely fine, and they are in a suspended, fluid state, with a large area of ​​water accumulation.

[0029] The "gray box" prediction model proposed in this invention mainly includes three core modules: a physical basis settlement calculation module based on the hierarchical summation method (physical mechanism layer), a residual correction module based on improved particle swarm optimization support vector regression (IPSO-SVR) (data-driven layer), and a fusion output module. Specifically: The physical basis settlement calculation module is used to construct a benchmark settlement calculation model (physical mechanism model) based on the physical mechanism of the hierarchical summation method and obtain the predicted physical settlement value. The residual correction module is used to build a residual correction model (data-driven model) based on the improved SVR algorithm and obtain the SVR settlement prediction value. The fusion output module is used to fuse the physical settlement prediction value and the SVR settlement prediction value to obtain the final settlement prediction result.

[0030] The core concept of this invention lies in breaking the traditional paradigm of independent physical models and data-driven models, and constructing a gray-box model architecture of "physical benchmark + intelligent residual". Specifically, firstly, a benchmark settlement calculation model (physical mechanism model) is established using the classic soil mechanics theory of hierarchical summation method to extract the main trend and physical laws of tailings dam settlement. Although the accuracy of this output is limited, it has clear physical meaning and stability, and can provide a "benchmark constraint" that conforms to the basic laws of soil mechanics for the entire prediction. On this basis, the deviation between the physical model calculation value and the field measured value is defined as the "residual". This residual essentially reflects the settlement information that the physical model cannot explain due to parameter simplification, neglect of rheological effects, and lateral extrusion. Subsequently, the improved particle swarm optimization (IPSO) algorithm is used to optimize the support vector regression (SVR) model, and nonlinear time-series fitting and compensation are performed on the residual sequence.

[0031] In this embodiment, the synergy and complementarity between the physical mechanism model and the improved SVR algorithm are reflected in the following three aspects: First, the division of labor is clear, and each component performs its specific function. The physical mechanism model is responsible for calculating the "linear basic component" of the settlement—this part of the settlement can be explained by classical consolidation theory and has a clear mechanical mechanism; the SVR residual correction model is responsible for learning the "nonlinear residual component"—this part of the settlement originates from complex factors such as the rheological properties of tailings sludge, the destruction of flocculated structures, and lateral extrusion. The physical mechanism model is difficult to quantify precisely, but SVR, with its powerful modeling ability for small samples and nonlinear problems, is able to capture these implicit patterns from the monitoring data.

[0032] Second, physical constraints ensure the reliability of predictions. Purely data-driven models are prone to producing predictions that violate physical principles when there is high data noise or insufficient samples (such as predicting settlement rebound or infinite growth). This invention uses the output of the physical mechanism model as a reference component to ensure that the final prediction always contains a physical basis that conforms to the laws of gravity and compressibility. Even if the machine learning module deviates under extreme conditions, the physical mechanism layer can still provide a baseline settlement estimate, avoiding the "catastrophic forgetting" problem that may occur with "black box" models.

[0033] Third, residual modeling reduces the learning difficulty. If the SVR is directly trained on the original settlement data, the settlement curve itself has complex nonlinear characteristics and is affected by multiple coupled factors, requiring the model to have extremely high complexity to achieve good accuracy. However, this invention strips away the parts that the physical mechanism model can explain, allowing the SVR to learn only the "residuals that the physical mechanism model cannot explain." This effectively reduces the learning objective and modeling difficulty of the SVR, enabling high-precision residual fitting results to be obtained even under small sample conditions.

[0034] The implementation process relies on high-performance computing workstations and on-site monitoring systems. The computing workstations are configured to run machine learning models for large settlement deformation, while the on-site monitoring system includes displacement sensors and pore water pressure gauges to acquire measured data. The specific implementation plan is as follows: Step 1: Acquisition and Preprocessing of Engineering Geological Data Before conducting predictive modeling, it is necessary to first establish a geological generalization model of tailings mud in the waterlogged area. This specifically includes the following steps: (1) Data Acquisition: Underwater surveying equipment was used to measure the current underwater topography of the tailings dam to determine the elevation of the top surface of the silt layer and the extent of the water accumulation area. Physical and mechanical parameters of each soil layer were obtained as engineering geological data through on-site drilling and sampling and indoor geotechnical tests (moisture content test, density test, standard consolidation test, etc.), including the thickness of each layer, initial void ratio, compression coefficient, compression modulus, and specific gravity of soil particles. At the same time, the cumulative settlement of the surface was collected in a time series using settlement plates and displacement sensors buried on-site.

[0035] (2) Geological stratification and parameter assignment: Macroscopic stratification is performed based on the genetic type and differences in physical and mechanical properties of the strata. On this basis, the foundation is divided into several horizontal calculation layers according to the accuracy requirement of (where (the width of the bottom of the siltation body)). The interfaces between different soil properties serve as the stratification boundaries. The stratification structure is expressed as follows: In the formula, For the first Layer thickness, in meters (m); , The first Elevation of the upper and lower interfaces of the layer, in meters; For the first Stratigraphic type label; This represents the total number of layers within the foundation compression layer.

[0036] Each computational layer is assigned a vector of physical and mechanical parameters: In the formula, For the first The initial void ratio of the soil layer was measured by indoor consolidation tests and is dimensionless. For the first The compressibility coefficient of the soil layer is calculated from the ep curve, and the unit is MPa-1; For the first The compression modulus of the soil layer is calculated from the consolidation test and is expressed in MPa. For the first The specific gravity of soil particles in the soil layer is measured by a specific gravity bottle test and is dimensionless. For the first The natural unit weight of the soil layer was determined by the ring sampler test, and the unit is kN / m³. 3 ; For the first The natural moisture content of the soil layer was determined by the oven-drying test, and the unit is % . For the first The cohesion of the soil layer is measured by direct shear or triaxial test, and the unit is kPa; For the first The internal friction angle of the soil layer is measured by direct shear or triaxial test, and the unit is °; For the first The permeability coefficient of the soil layer is measured by permeability tests or estimated empirically, and the unit is cm / s. The parameters of each layer are organized into a layer parameter matrix according to the layer sequence. For layers lacking test data, the weighted average method of adjacent layers is used for interpolation.

[0037] (3) Determination of stress state and boundary conditions: Calculate the initial self-weight stress field based on the unit weight and thickness of each layer, and determine the planar range of the water accumulation area and the calculated depth of the compression layer (based on the additional stress). With self-weight stress The ratio is the control standard: generally taken from soil soft soil The model incorporates geometric and mechanical boundary conditions, including the location of the groundwater level. Ultimately, a generalized geological model is formed, encompassing sequence stratigraphy, thickness, parameter matrices, and stress distribution, serving as the unified input for subsequent physical settlement calculations and SVR residual corrections. This model is dynamically corrected and updated as the project progresses and new borehole data is acquired.

[0038] (4) Data preprocessing: To eliminate the influence of different physical units (such as time in days, load in kPa, and pore pressure in kPa) on the convergence speed of the algorithm, the feature data of the input model are normalized and mapped to the [0,1] interval: In the formula, The original data, and These are the minimum and maximum values ​​of the feature data, respectively. This is the normalized data.

[0039] The second step: Constructing a benchmark settlement calculation model based on physical mechanisms (Layer 1: Physical Mechanism Layer) This step aims to calculate the principal consolidation settlement under siltation load using the classic soil mechanics theory—the layered summation method. This calculated value constitutes the "physical lower limit" of the prediction results, ensuring that the output of the baseline settlement calculation model conforms to the laws of gravity and compressibility. The specific implementation process is as follows: 1. Foundation stratification: The weak tailings mud foundation at the bottom of the siltation layer is divided into... Each layer is a horizontal computing layer. The layering principle satisfies: layer thickness... ( The bottom width of the silt mass is used as the boundary between different soil types, and the interface between different soil types is used as the stratification boundary.

[0040] 2. Stress calculation: (1) Calculate the first Self-weight stress at the midpoint of the layer : In the formula, For the first The effective unit weight of the soil layer, in kN / m³ 3 ; For the first The thickness of the soil layer, in meters (m). For the first The effective unit weight of the soil layer, in kN / m³ 3 ; For the first The thickness of the soil layer, in meters (m). This is the number of the current computing layer. .

[0041] (2) Calculate the first Additional stress at the midpoint of the layer caused by the upper siltation fill (load). Based on the geometry of the sludge mass (usually simplified to trapezoidal or triangular loads), the Boussinesq solution or corner point method coefficients (stress coefficients) of elasticity are used. Perform the calculation: In the formula, For the first Additional stress (kPa) at the midpoint of the layer caused by the upper load; The base pressure (kPa) applied to the extruded material; The stress coefficient is determined by the Boussinesq solution (dimensionless) and depends on the load distribution and the location of the calculation point.

[0042] 3. Settlement calculation: Determine the self-weight stress based on the ep curve (void ratio-pressure curve) obtained from the indoor consolidation test. Initial porosity under action and total stress Porosity after compression under action . No. Compression deformation of soil layers for: The total final settlement of the foundation The sum of the settlement amounts of each layer: in, This is the predicted physical settlement value, i.e., the total final settlement of the foundation. An empirical coefficient for settlement calculation is used, with values ​​typically ranging from 1.1 to 1.4, based on regional experience. The initial void ratio under self-weight stress. The void ratio after compression under total stress. For layer thickness, For the number of local base levels.

[0043] The third step: Constructing a residual correction model based on the improved PSO-SVR (Layer 2 & 3: Data-driven and algorithm optimization layers) Because tailings sludge exhibits a flocculated structure during underwater deposition, and the siltation process is accompanied by strong disturbances and nonlinear rheology, simple physical calculations cannot quantify this deviation. This step establishes a residual correction model to predict the residual between the "physically calculated value" and the "measured value".

[0044] 1. Residual Definition and Sample Construction: Definition The settlement residual at time t is: in, To represent the true physical settlement, key factors influencing nonlinear settlement are selected to form the input vector. ,in: Consolidation time; This represents the current height of the silt-displacing and backfilling. The dissipation rate of excess pore water pressure; Settlement rate. Construct a training set. ,in The total number of training samples, This is a sample index variable.

[0045] 2. Support Vector Regression (SVR) Model Construction: Introduction - An insensitive loss function transforms the regression problem into a convex quadratic programming problem of finding the optimal hyperplane. Objective function: In the formula, the hyperparameters of the SVR objective function are explained as follows: The normal vector (weight vector) of the hyperplane; The bias term (threshold) for the hyperplane; ≥0; The penalty factor (>0) controls the intensity of the penalty for samples outside the error band. The larger the value, the more complex the model and the higher the fit. This represents the total number of training samples.

[0046] Constraints: In the formula, Indicates the first The true residual values ​​of each training sample, in mm; The total number of training samples; This is the penalty factor, which controls the degree of penalty for samples that exceed the error band; , These are slack variables; This is a mapping function that maps the input space to a high-dimensional feature space. A radial basis function (RBF) is used as the kernel function. ,in Let be the value of the basis kernel function, representing the th . training samples With the training samples The similarity or inner product in the high-dimensional feature space has a range of (0,1]; These are kernel parameters that control the distribution breadth of the RBF function; For input vectors and The squared Euclidean distance between the two samples measures their similarity in the original input space. The final regression function is: In the formula, , The Lagrange multipliers are obtained by solving the above convex quadratic programming problem, and both are non-negative real numbers. This represents the input feature vector of the i-th sample in the training set, which is known data that has already been used for model training; The input feature vector represents the unknown data for which settlement prediction needs to be performed.

[0047] 3. Improved Particle Swarm Optimization (IPSO) Parameter Optimization: The prediction accuracy of the SVR model is highly dependent on the parameter combination. The standard PSO algorithm has the following three main defects: (1) It uses a random method to initialize the population, resulting in uneven particle distribution, insufficient search space coverage, and easy trapping in local optima; (2) The inertial weight usually adopts a fixed value or linear decreasing strategy, which cannot adapt to the complex nonlinear optimization process, and it is difficult to coordinate the global exploration and local development capabilities; (3) The population diversity is lost in the later stage of iteration, and there is a lack of effective escape mechanism. Once trapped in a local extreme value, it is difficult to get out. In view of the above defects, this invention makes synergistic improvements from three levels: initialization strategy, weight adjustment strategy and diversity maintenance strategy, to obtain an improved SVR algorithm. The specific improvement process includes: Improvement 1: Initialization strategy for Logistic-Tent composite chaotic mapping (1) Deficiencies of existing technology: The standard PSO algorithm initializes the population in a random manner. Due to the inherent randomness of random numbers, it is difficult to ensure uniform coverage of the initial particle distribution in the search space. Particle aggregation or sparseness is likely to occur, causing the algorithm to deviate from the global optimal region in the early stage of iteration. Furthermore, the single Logistic chaotic mapping has a periodic degradation problem in the later stage of iteration, which affects the traversal quality of the chaotic sequence.

[0048] (2) Improved solution of the present invention: The present invention introduces a Logistic-Tent composite chaotic mapping to replace random initialization, and uses the ergodicity and randomness of chaotic motion to generate an initial population, so that the particles are more evenly distributed in the search space, effectively avoiding the periodic degradation problem that occurs in the later stages of a single Logistic mapping. The formula for generating composite chaotic variables is as follows: when hour: ; when hour: ; In the formula, For the first The chaotic variable of the step takes a value range of (0, 1); For control parameters, this embodiment takes... This value ensures that the system is in a completely chaotic state without a periodic window. The Logistic-Tent composite chaotic map combines the ergodicity of the Logistic map and the uniform distribution of the Tent map, resulting in chaotic sequences with better uniformity and randomness in the search space.

[0049] (3) Specific implementation steps: Step 1: Randomly generate a dimensional vector ( ); Step 2: Iteratively generate chaotic sequences using the above formula. , , ..., ,in Population size; Step 3: Convert chaotic variables Carrier to Parameter Within the search range: In the formula, For the first The particle in the first Dimensional position value (corresponding) or ); , The first The lower and upper bounds of the dimensional search space; For the first The particle in the first The value of the chaotic variable corresponding to the dimension.

[0050] The resulting initial population combines ergodicity and randomness, fundamentally improving the quality of the initial solution and laying a solid foundation for subsequent iterative searches.

[0051] Improvement 2: Coordinated adjustment strategy of nonlinear adaptive inertia weights and learning factors (1) Deficiencies of existing technology: inertial weight The learning factor is a key parameter for balancing the global exploration capability and local exploitation capability of the PSO algorithm. However, standard PSO typically uses fixed weights or linearly decreasing weights. The former cannot achieve a dynamic transition from global exploration to local exploitation, while the latter, although an improvement, has a constant decreasing rate, making it difficult to adapt to the differentiated requirements of weight change rates at different stages in complex nonlinear optimization processes. At the same time, the learning factor is usually set to a fixed value, which cannot coordinate the particle's "self-awareness" and "social learning" capabilities with the iteration process.

[0052] (2) Improved solution of the present invention: The present invention designs an adaptive weight strategy that decreases nonlinearly quadratically with the number of iterations, so that the weight change rate is dynamically adjusted with the iteration process. At the same time, an adaptive adjustment strategy for the learning factor is introduced, so that the algorithm focuses on individual exploration in the early stage and on group convergence in the later stage. The formula for calculating the nonlinear adaptive inertia weight is as follows: In the formula, For the first Inertia weight during iteration (dimensionless). This represents the maximum value of the inertia weight (0.9 in this embodiment), corresponding to the global exploration phase in the early stage of the algorithm; This is the minimum value of the inertia weight (0.4 in this embodiment), corresponding to the local fine-tuning stage in the later stage of the algorithm; Let this be the current iteration algebra; This represents the maximum number of iterations.

[0053] This quadratic decreasing strategy keeps the weights at a high level in the early stages and decreases slowly to fully explore the search space; in the later stages, the weights drop rapidly to a lower level to accelerate local convergence. Compared to the linear decreasing strategy, the quadratic decreasing strategy better meets the actual needs of the PSO algorithm: "a long period of thorough exploration in the early stages and rapid and accurate convergence in the later stages."

[0054] The formula for the adaptive adjustment strategy of the learning factor is as follows: In the formula, For the first The cognitive learning factor of the generation controls the weight of the particle learning to its own historical best position; For the first The social learning factor of the generation controls the weight of particles learning to the global optimal position of the group; , These are the starting and ending values ​​of the cognitive learning factor (2.5 and 0.5 in this example). , These are the starting and ending values ​​of the social learning factor (0.5 and 2.5 in this example).

[0055] This adjustment strategy makes It gradually decreases as the number of iterations increases. The value gradually increases with the number of iterations, achieving a smooth transition from focusing on individual experience exploration in the early stages of the algorithm to focusing on group information sharing in the later stages.

[0056] Improvement 3: Adaptive mutation and reverse learning perturbation strategy based on fitness variance (1) Deficiencies of existing technologies: In the later stages of the standard PSO algorithm, individuals in the population gradually gather in local optimum regions, leading to a rapid loss of population diversity. At this time, the search ability of the algorithm drops significantly, and it is difficult to escape once trapped in a local optimum. Traditional improvement methods mostly use random mutation with fixed probability, and the timing and intensity of mutation lack specificity. In addition, the handling of outbound particles usually adopts a simple boundary truncation method, which discards the useful information that outbound particles may carry.

[0057] (2) Improved solution of the present invention: The present invention introduces a real-time monitoring mechanism for population diversity based on fitness variance to accurately determine whether the population is trapped in a local optimum. Mutation perturbation is triggered only when necessary. At the same time, an adaptive decay variable asynchronous length is adopted to gradually reduce the mutation amplitude as the iteration progresses, so as to take into account the stability in the later stage. For out-of-bounds particles, a reverse learning strategy is used for "rebirth" instead of simple truncation.

[0058] The formula for calculating the variance of population fitness is: In the formula, Population fitness variance reflects the degree of dispersion among individuals in the population; Population size; For the first The fitness value of each particle; This represents the average fitness value of the population.

[0059] when hour( To preset the diversity threshold, this embodiment uses 10. -4 This indicates that the population is becoming homogeneous, and the algorithm may have gotten stuck in a local optimum, at which point the mutation perturbation mechanism is triggered. In the formula, For variable asynchronous long coefficients, an adaptive decay strategy is adopted. This allows the variation amplitude to gradually decrease as the iteration progresses, in order to take into account the stability in the later stages; These are random numbers distributed according to a standard normal distribution. The initial variable asynchronous length is (in this embodiment, it is 0.1 times the search range). This is the attenuation coefficient (2.0 in this embodiment); Let this be the current iteration algebra; This represents the maximum number of iterations.

[0060] When a particle's position exceeds the search boundary, a reverse learning strategy is used for "rebirth": In the formula, The position of the reborn particle; The original position of the particle that has exceeded the boundary; , For the first The lower and upper bounds of the dimensional search space.

[0061] This strategy maps out-of-bounds particles to their opposite positions in the search space, ensuring the effectiveness of the population and enhancing its exploration capabilities by generating "opposite solutions." This synergistic effect with the aforementioned mutation perturbation mechanism helps prevent premature convergence of the algorithm.

[0062] Improvement point 4: Fitness function and optimal parameter output Fitness function design: The mean squared error (MSE) of the SVR model on the training set is used as the fitness function for the particles. In the formula, This represents the particle's fitness value; a smaller value indicates that the particle is more fitness-oriented. The better the parameter combination; The total number of training samples; For the first The true residual values ​​(mm) of each training sample; For the first SVR model prediction residuals (mm) for training samples.

[0063] The IPSO algorithm searches for a way to make particles appear larger by continuously updating their velocity and position. Minimize the optimal combination of parameters Combine the optimal parameters Substitute the values ​​into the SVR model to obtain the residual correction model and then obtain the SVR settlement prediction values.

[0064] Step 4: Fusing Predictions and Model Output The first step is residual prediction, which involves combining the optimal parameters. Substitute the feature vectors of the time to be predicted into the SVR model. The predicted nonlinear residual values ​​are obtained. This refers to the SVR settlement prediction value. Then, the results are overlaid, combining the physical settlement prediction values. Compared with SVR settlement prediction value The results are superimposed to obtain the final settlement prediction results. : Finally, there is the model update. As the project progresses, the residual correction model is incrementally trained regularly using newly collected field monitoring data to continuously correct the model parameters and achieve dynamic inversion prediction.

[0065] In summary, this invention provides a tailings dam settlement prediction method based on the fusion of physical mechanisms and an improved SVR algorithm. First, based on engineering geological survey data, a physical mechanism model is constructed using the hierarchical summation method to calculate the linear basic components of settlement. Second, a nonlinear residual sequence is constructed using the deviation between field monitoring data and the physical model's calculated values. Subsequently, consolidation time, fill height, excess pore water pressure dissipation, and settlement rate are selected to construct a multidimensional feature vector. An improved particle swarm optimization (IPSO) algorithm is used to optimize the penalty factor and kernel parameters of Support Vector Regression (SVR) to establish a nonlinear residual correction model. Finally, the calculated values ​​of the physical model are superimposed with the predicted residual values ​​of the correction model to output the final settlement prediction result. This invention effectively solves the problems of difficulty in obtaining parameters for traditional physical models and the poor generalization ability of single machine learning models under small sample conditions, significantly improving the accuracy and robustness of tailings dam settlement prediction under complex geological conditions.

[0066] Example 2 Based on the same inventive concept, the present invention also provides a tailings dam settling prediction system based on the fusion of physical mechanisms and improved SVR algorithm, for implementing the method described in the foregoing embodiments. The system includes: an acquisition and processing module, a model building module, and a prediction module. The acquisition and processing module is used to acquire engineering geological data of the tailings dam to be tested and perform preprocessing, including: obtaining physical and mechanical parameters such as the thickness of each layer, initial void ratio, compression coefficient, compression modulus and soil particle specific gravity by using borehole sampling and indoor geotechnical tests; collecting the actual cumulative settlement of the ground surface by using settlement plates and displacement sensors buried on site in a time series; and normalizing all input feature data and mapping them to the [0,1] interval.

[0067] The model building module is used to construct a "grey box" prediction model based on physical mechanisms and an improved SVR algorithm. This module further includes a physical basis settlement calculation unit, a residual correction unit, and a fusion output unit. The physical basis settlement calculation unit is used to calculate the predicted physical settlement values ​​based on the hierarchical summation method. The physical settlement prediction value is obtained; the residual correction unit is used to construct the residual correction model based on the improved SVR algorithm. First, the settlement residual is defined. The consolidation time, fill height, excess pore water pressure dissipation, and settlement rate are selected to construct the input feature vector. Then, the penalty factor for the SVR model is applied using an improved particle swarm optimization algorithm. and kernel parameters Optimization is performed, and finally, an SVR model is constructed and trained based on the optimal parameter combination to obtain SVR settlement prediction values; the fusion output unit is used to superimpose the physical settlement prediction values ​​and the SVR settlement prediction values. To obtain the final settlement prediction results.

[0068] The prediction module is used to obtain the tailings dam settlement prediction results based on the preprocessed engineering geological data and the "gray box" prediction model, and to dynamically update the model using newly collected field monitoring data as the project progresses.

[0069] The specific implementation principles and technical effects of the above modules have been explained in detail in Example 1, and will not be repeated here.

[0070] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A tailings dam settlement prediction method based on the fusion of physical mechanisms and an improved SVR algorithm, characterized in that, The method includes: Collect and preprocess engineering geological data of the tailings dam to be tested; A "grey box" prediction model is constructed based on physical mechanisms and an improved SVR algorithm; Based on the preprocessed engineering geological data and the "gray box" prediction model, the tailings dam settlement prediction results were obtained.

2. The method according to claim 1, characterized in that, The "gray box" prediction model includes: a physical basis settlement calculation module, a residual correction module, and a fusion output module; The physical basis settlement calculation module is used to construct a benchmark settlement calculation model based on the physical mechanism of the hierarchical summation method and obtain the predicted physical settlement value. The residual correction module is used to construct a residual correction model based on the improved SVR algorithm and obtain the SVR settlement prediction value. The fusion output module is used to fuse the physical settlement prediction value and the SVR settlement prediction value to obtain the final settlement prediction result.

3. The method according to claim 2, characterized in that, Methods for obtaining predicted physical settlement values ​​include: ； in, This is the predicted value of physical settlement. Empirical coefficients for settlement calculation. The initial void ratio under self-weight stress. The void ratio after compression under total stress. For layer thickness, For the number of base levels, This is the number of the current computing layer.

4. The method according to claim 2, characterized in that, Methods for constructing residual correction models based on the improved SVR algorithm include: An improved SVR algorithm is obtained by adopting a Logistic-Tent composite chaotic mapping initialization strategy, a nonlinear adaptive inertia weight and learning factor coordinated adjustment strategy, and an adaptive mutation and reverse learning perturbation strategy based on fitness variance. An improved SVR algorithm is used to optimize the parameters of the SVR model and obtain the optimal parameter combination. An SVR model is constructed and trained based on the optimal parameter combination to obtain a residual correction model.

5. The method according to claim 4, characterized in that, Nonlinear adaptive inertia weight and learning factor co-adjustment strategies include: Formula for calculating nonlinear adaptive inertia weight: ； in, For the first Inertia weights during iterations This represents the maximum value of the inertia weight. This represents the minimum inertial weight. Let the current iteration algebra be... The maximum number of iterations; Learning factor adaptive adjustment strategies include: ； ； in, For the first The cognitive learning factor of the generation controls the weight of the particle learning to its own historical best position; For the first The social learning factor of the generation controls the weight of particles learning to the global optimal position of the group; , These are the initial and final values ​​of the cognitive learning factor, respectively. , These are the starting and ending values ​​for the social learning factor, respectively.

6. The method according to claim 5, characterized in that, Adaptive mutation and reverse learning perturbation strategies based on fitness variance include: The formula for calculating the variance of population fitness is: ； in, For population fitness variance; Population size; For the first The fitness value of each particle; This represents the average fitness value of the population. when hour, To trigger the mutation perturbation mechanism by setting a preset diversity threshold: ； in, For the first The particle in the first Dimensional position value, For variable asynchronous long coefficients, an adaptive decay strategy is adopted. This causes the variation amplitude to gradually decrease as the iteration progresses; These are random numbers distributed according to a standard normal distribution. The initial variable asynchronous length; The attenuation coefficient; When a particle's position exceeds the search boundary, a reverse learning strategy is used for "rebirth": ； In the formula, The position of the reborn particle; The original position of the particle that has exceeded the boundary; , For the first The lower and upper bounds of the dimensional search space.

7. A tailings dam settlement prediction system based on the fusion of physical mechanisms and an improved SVR algorithm, the system being used to implement the method described in any one of claims 1-6, characterized in that, The system includes: an acquisition and processing module, a model building module, and a prediction module; The data acquisition and processing module is used to acquire and preprocess the engineering geological data of the tailings dam to be tested. The model building module is used to build a "grey box" prediction model based on physical mechanisms and an improved SVR algorithm; The prediction module is used to obtain tailings dam settlement prediction results based on preprocessed engineering geological data and a "gray box" prediction model.

8. The system according to claim 7, characterized in that, The model building module includes: a physical basis settlement calculation unit, a residual correction unit, and a fusion output unit; The physical basis settlement calculation unit is used to construct a benchmark settlement calculation model based on the physical mechanism of the hierarchical summation method and obtain the predicted value of physical settlement. The residual correction unit is used to construct a residual correction model based on the improved SVR algorithm and obtain the SVR settlement prediction value. The fusion output unit is used to fuse the physical settlement prediction value and the SVR settlement prediction value to obtain the final settlement prediction result.

9. The system according to claim 7, characterized in that, The process of constructing a residual correction model based on the improved SVR algorithm includes: An improved SVR algorithm is obtained by adopting a Logistic-Tent composite chaotic mapping initialization strategy, a nonlinear adaptive inertia weight and learning factor coordinated adjustment strategy, and an adaptive mutation and reverse learning perturbation strategy based on fitness variance. An improved SVR algorithm is used to optimize the parameters of the SVR model and obtain the optimal parameter combination. An SVR model is constructed and trained based on the optimal parameter combination to obtain a residual correction model.

Images