Marine siliceous sand-geomembrane interfacial shear stress prediction model construction method

Abstract

The invention provides an ocean siliceous sand-geomembrane interface shear stress prediction model construction method, and relates to the technical field of ocean engineering. The method comprises the following steps: preparing an experimental sample; performing a wide-working-condition interface shearing experiment to obtain interface mechanical data; based on the interface mechanical data, dividing a training set and a test set, defining input and output parameters and performing normalization processing to form a standardized database; constructing four types of machine learning models; and evaluating the performance of the four types of machine learning models, and screening and determining an optimal prediction model as PSO-CNN. According to the method, a multi-working-condition experiment-high-precision model-engineering application multi-dimensional composite technology system is constructed, the whole-process support from experimental data acquisition to engineering parameter output can be realized, the efficiency and precision of sea sand-geomembrane interface shear stress research are effectively improved, the engineering practicability concept is deeply implemented, and the research on the sea sand-geomembrane interface shear stress is realized. The problems of incomplete data, low prediction precision and high cost of a traditional method are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of marine engineering technology, specifically to a method for constructing a predictive model for shear stress at the marine siliceous sand-geomembrane interface. Background Technology

[0002] In the field of marine engineering, the shear stress characteristics of the marine silica sand (hereinafter referred to as sea sand)-geomembrane interface are the core basis for the design and safety assessment of infrastructure such as port terminals and coastal roadbeds. The accuracy of its data directly determines the stability and service life of the geostructure. It should be emphasized that there are fundamental differences between sea sand and terrestrial silica sand in terms of material composition and engineering compatibility: marine silica sand is subjected to long-term sorting by marine dynamics, resulting in high particle roundness and concentrated particle size distribution. However, due to seawater immersion, it easily adsorbs chloride ions and contains shell fragments, forming nanoscale dissolution pores on the surface. Terrestrial silica sand is mostly derived from terrestrial weathering or river transport, with sharp-edged particles, poor sorting, low chloride ion content, and impurities mainly consisting of clay and feldspar fragments. When the two types of sand bodies interact with high-density polyethylene (HDPE) geomembranes, the interfacial friction angle of marine sand is 2° to 3° lower than that of terrestrial siliceous sand due to its particle morphology and surface characteristics. Furthermore, under the coupled effect of temperature and stress, the dispersion of shear stress at the marine sand-geomembrane interface (coefficient of variation 12% to 15%) is significantly greater than that at the terrestrial siliceous sand interface (coefficient of variation 6% to 8%). This is the core reason why marine engineering requires separate special research on the marine sand-geomembrane interface.

[0003] However, existing technologies for obtaining the shear stress at this interface have significant limitations in terms of "coverage of physical tests under wide operating conditions" and "adaptation to high-precision prediction models." In particular, under extreme temperatures in tropical coastal areas and coupled conditions of multi-grain size sea sand, it is difficult to output mechanical parameters that fit the actual engineering situation, thus limiting its supporting role in marine engineering design.

[0004] In terms of physical testing equipment, existing temperature-controlled interface shearing equipment mostly focuses on narrow temperature ranges (3℃-50℃), which cannot simulate the extreme environment of marine engineering from -5℃ to 80℃. Moreover, a single device can only adapt to the testing requirements of a single particle size and a single normal stress. To carry out coupled testing of multiple particle sizes, temperatures, and stresses, multiple discrete devices are required, which not only increases the purchase and maintenance costs but also reduces the comparability of data due to differences in the testing principles and accuracy of the equipment. At the same time, a single test needs to complete the entire process of temperature control, consolidation, and shearing, and the efficiency of acquiring data under multiple working conditions is extremely low, making it difficult to support the engineering requirements for rapid parameter measurement.

[0005] Regarding predictive models, traditional statistical methods cannot describe the strong nonlinear relationship between sea sand particle size, temperature, and normal and shear stress. Early machine learning models were prone to getting trapped in local optima, with a root mean square error of 9.63 and an average absolute percentage error of 11.3% on test sets. Moreover, most models did not incorporate the key parameter of sea sand particle size or only considered static temperature conditions, failing to adapt to the actual scenarios of diurnal temperature fluctuations of ±15℃ and dynamic changes in normal stress with load in engineering projects. This resulted in significant deviations between predicted and measured values. In engineering design, only conservative parameters could be used, which increased construction costs and could potentially create safety hazards due to parameter deviations.

[0006] With the increasing demand for accuracy and efficiency in acquiring interfacial shear stress data in marine engineering, the limitations of existing technologies have become a key factor restricting engineering safety design and economic optimization. Therefore, it is urgent to construct a method for predicting the interfacial shear stress of sea sand-geomembrane by integrating physical experiments under wide working conditions and high-precision algorithms, so as to solve the problems of low experimental efficiency and poor model adaptability, and provide reliable interfacial mechanical parameter support for marine engineering.

[0007] In summary, existing technologies generally suffer from limitations in temperature control, insufficient particle size adaptability, low functional integration, and a lack of temperature-particle-force coupling prediction capabilities. Experimental data and prediction models are disconnected, making it difficult to meet the research needs for accurate prediction of shear stress at the interface between geomembranes and sea sand of different particle sizes in marine engineering. Summary of the Invention

[0008] To address the limitations of existing technologies, such as insufficient temperature control, inadequate particle size adaptability, low functional integration, and lack of temperature-particle-force coupling prediction capabilities, resulting in a disconnect between experimental data and prediction models and difficulty in meeting the research needs for accurate prediction of shear stress at the interface between geomembranes and sea sand of different particle sizes in marine engineering, this invention proposes a method for constructing a sea sand-geomembrane interface shear stress prediction model that integrates wide-condition physical experiments and high-precision algorithms. This method aims to solve the above problems and provide reliable interface mechanical parameter support for marine engineering.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for constructing a prediction model for shear stress at the marine siliceous sand-geomembrane interface, the method comprising: Two or more marine silica sand samples with different particle sizes were screened, and geomembrane samples were prepared at the same time. The geomembrane samples were laid at the bottom of the upper shear box of a customized temperature-controlled interface shearing device. The marine silica sand samples were layered and filled into the upper surface of the geomembrane in the upper shear box to ensure stable contact between the marine sand samples and the geomembrane. Using the customized temperature-controlled interface shearing device, the device debugging, normal stress application and consolidation, shearing and data acquisition operations were completed in sequence to obtain interface mechanical data under different marine silica sand particle sizes, temperatures and normal stress conditions. Based on the interface mechanics data, a training set and a test set are divided, input and output parameters are defined and normalized to form a standardized database. Construct four types of machine learning models: BPANN, ELM, CNN, and PSO-CNN; The performance of the four types of machine learning models was evaluated, and the optimal prediction model was selected as PSO-CNN.

[0010] Compared with the prior art, the beneficial effects of the present invention are: Wide-temperature-range multi-particle-size coupling experiments overcome the limitations of scenario adaptation: Utilizing customized temperature-controlled interface shearing equipment, a wide temperature range of -5℃ to 80℃ is achieved (internal temperature difference of the sample ≤ ±1℃), accurately reproducing the tropical high-temperature (80℃) and cold-region low-temperature (-5℃) working conditions of marine engineering; simultaneously adaptable to two types of commonly used marine sand particle sizes in engineering, 1mm~2mm (S1) and 2mm~4mm (S2), combined with 15kPa, 25kPa, and 50kPa normal stress loading, covering the stress environment of shallow to deep marine foundations, filling the gap in traditional experimental equipment with narrow temperature range and only adaptable to single-size marine sand, which cannot carry out multi-factor coupled interface mechanics research of "temperature-particle-stress".

[0011] High-precision model optimization improves prediction accuracy by addressing strong nonlinear relationships: A particle swarm optimization convolutional neural network (PSO-CNN) model is constructed. The PSO algorithm is used to optimize the hyperparameters of the CNN (initial weights, number of convolutional channels, learning rate, etc.), effectively coupling the strong nonlinear relationship between temperature, sand particle size, normal stress, and shear displacement. Compared with the unoptimized CNN neural network, the RMSE of the PSO-CNN test set is reduced to 2.1, the MAPE is 8%, and the correlation coefficient R reaches 0.98, with an accuracy improvement of over 40%. This solves the problem that traditional statistical models or simple neural networks are difficult to accurately fit the mechanical laws of complex interfaces. An integrated experiment-model-engineering process enhances practicality and efficiency: It forms an integrated technical path of "wide-condition experimental data acquisition - standardized database construction - high-precision model training - rapid output of engineering parameters". A single customized device replaces multiple discrete instruments in the traditional method, reducing costs by more than 40%. The model can quickly output interface shear stress under different working conditions, shortening the time for acquiring parameters under multiple working conditions by 20%~25%. It avoids the defects of experiment and model disconnect and parameter acquisition delay in traditional methods, and directly provides efficient and accurate mechanical parameter support for the interface reinforcement design of coastal roadbeds, port terminals and other projects.

[0012] Multi-dimensional data support and strong generalization ability cover complex engineering scenarios: The model integrates 2,100 sets of valid experimental data to build a standardized database, and randomly divides the training set and test set in an 8:2 ratio to ensure uniform distribution of various working conditions such as "sea sand particle size-temperature-normal stress". Under extreme working conditions, the model's predicted value deviates from the measured value by ≤8%, which can accurately guide the design of sea sand particle size selection (such as prioritizing S2 sea sand for deep foundations to improve interface strength) and geomembrane layer laying spacing optimization in cold and tropical marine engineering. This solves the problem of weak generalization ability and inability to adapt to complex marine engineering scenarios of traditional models.

[0013] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method for constructing a prediction model for geomembranes and sea sand of different particle sizes according to the present invention; Figure 2 This is a characterization diagram of marine quartz sand samples according to the present invention; it includes three sub-figures: (a) showing the physical appearance of S1 (1mm~2mm) and S2 (2mm~4mm) sea sand, (b) showing the SEM microstructure, and (c) showing the particle size distribution curve. Figure 3 This is a characterization diagram of the HDPE geomembrane layer according to the present invention; Figure 4 This is a schematic diagram of a customized temperature control interface shearing device according to the present invention; Figure 5 This is the shear stress-displacement curve of the S1 sea sand-geomembrane interface according to the present invention; Figure 6 This is the shear stress-displacement curve of the S2 sea sand-geomembrane interface according to the present invention; Figure 7 This is a comparison chart of the prediction accuracy of the training sets of various models according to the present invention; Figure 8 This is a comparison chart of the prediction accuracy of each model on the validation set according to the present invention. Detailed Implementation

[0015] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings, so as to more clearly understand the purpose, features and advantages of this invention. It should be understood that the embodiments shown in the drawings are not intended to limit the scope of this invention, but are only for illustrating the essential spirit of the technical solutions of this invention. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0016] Unless the context requires otherwise, throughout the specification and claims, the word “comprising” and its variations, such as “including” and “having”, shall be understood to have an open, inclusive meaning, that is, to be interpreted as “including, but not limited to”.

[0017] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0018] The singular forms “a” and “the” used in this specification and the appended claims include plural references unless otherwise expressly stated herein. It should be noted that the term “or” is generally used to mean “and / or” unless otherwise expressly stated herein.

[0019] In the following description, in order to clearly demonstrate the structure and working method of the present invention, a number of directional terms will be used. However, terms such as "front", "back", "left", "right", "outside", "inside", "outward", "inward", "up", and "down" should be understood as convenient terms and not as limiting terms.

[0020] The implementation details of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following content is only for the convenience of understanding the implementation details and is not necessary for implementing this solution.

[0021] To address the shortcomings of existing studies on shear stress at the marine siliceous sand (hereinafter referred to as sea sand)-geomembrane interface, such as narrow temperature coverage of physical experimental equipment, suitability only for sea sand of a single particle size, low experimental efficiency, and the failure of existing prediction models to effectively couple the strong nonlinear relationship between "temperature-sea sand particle size-normal stress," this invention aims to provide a method for constructing a prediction model for shear stress at the marine siliceous sand-geomembrane interface that integrates physical experiments under wide working conditions with high-precision machine learning algorithms. This method enables accurate prediction of interfacial shear stress under multi-particle-size, wide-temperature-range, and multi-stress coupling scenarios, providing reliable mechanical parameter support for the reinforcement design of sea sand-geomembrane interfaces in marine engineering projects such as port terminals and coastal roadbeds.

[0022] Example 1 This invention constructs a high-precision sea sand-geomembrane interface shear stress prediction model through a technical path of "experimental sample preparation - wide-condition interface shear experiment - database construction - model training and optimization - model accuracy verification," overcoming the limitations of traditional interface shear stress methods. For example... Figure 1 As shown, this invention provides a method for constructing a predictive model of interfacial shear stress between a geomembrane layer and sea sand of different particle sizes, considering the influence of temperature, comprising the following steps: Step 1: Sample preparation: Marine quartz sand samples were prepared by sieving two commonly used engineering particle sizes. At the same time, geomembrane samples were prepared by using high-density polyethylene (HDPE). The geomembrane samples were laid at the bottom of the upper shear box of a customized temperature-controlled interface shearing device. The marine sand samples were layered and filled into the upper surface of the geomembrane in the upper shear box to ensure stable contact between the marine sand samples and the geomembrane.

[0023] In some specific embodiments, the sea sand sample preparation process is as follows: marine quartz sand is sieved using a standard vibrating sieve to obtain S1 sea sand with a particle size of 1mm~2mm and S2 sea sand with a particle size of 2mm~4mm respectively. The optimum moisture content of the two types of sea sand is determined and controlled to be 9.65%. A layered filling method is adopted, with each layer being 25mm thick, for a total of 3 layers. A shear box with a specification of 300mm×300mm×150mm is filled in. During the filling process, a light compactor is used to compact the material to ensure that the density deviation of each layer is ≤±2%.

[0024] As a better option, the sea sand screening uses a set of standard vibrating screens with apertures of 1mm, 2mm and 4mm, and the screening time is controlled at 10min~15min to ensure uniform particle size distribution; the sea sand moisture content control adopts a "weighing-spraying-stirring-retesting" cycle process, with a 5min interval between each cycle, until the moisture content error is ≤±0.5%.

[0025] As a better option, the mass of the sea sand sample is calculated based on the volume of the shear box and the density of the sea sand. The mass of the S1 sea sand is 14.09 kg and the mass of the S2 sea sand is 15.20 kg. After filling, the flatness of the top surface of the sample is measured using a dial gauge to ensure that the flatness deviation is ≤0.5 mm, so as to avoid the dispersion of experimental data due to the unevenness of the sample.

[0026] In some specific embodiments, the geomembrane sample preparation process is as follows: HDPE sheets with a thickness of 2mm and a density of 0.942g / cm³ are selected, and their surface three-dimensional roughness parameter Rs is measured and controlled to be 1.06 by a laser profilometer. The sheets are then cut into 280mm×460mm specifications using a CNC cutting machine and fixed with positioning pins when laid at the bottom of the upper shear box to ensure that the shearing area is stable at 300mm×300mm.

[0027] As a better alternative, the surface roughness of the geomembrane layer can be controlled by sandblasting. The sandblasting material is selected as quartz sand with a particle size of 0.1mm~0.3mm, and the sandblasting pressure is controlled at 0.2MPa~0.3MPa to further enhance the interfacial bonding effect.

[0028] In some specific embodiments, the customized temperature-controlled interface shearing device includes an external structure such as an external temperature control chamber, a normal stress loading device, and a shear stress loading device, as well as an internal testing device with an upper shear box, a lower shear box, and matching sensors. It can test the shear mechanical properties of the material interface by loading operations in the normal and shear directions in a controllable temperature environment, combined with data monitoring.

[0029] Step 2, Wide-condition interface shearing experiment: A customized temperature-controlled interface shearing device is used, such as... Figure 4 As shown, the equipment debugging, normal stress application and consolidation, shearing and data acquisition operations were completed in sequence to obtain interface mechanical data under different sea sand particle sizes, temperatures and normal stress conditions.

[0030] In some specific implementations, the equipment debugging process involves filling the upper shear box with either S1 or S2 sea sand samples, closing the environmental temperature control chamber, setting the experimental temperature, and maintaining the temperature for 2 hours using the temperature control system. The temperature is monitored by a platinum resistance temperature sensor embedded inside the sample to ensure that the internal temperature difference of the sample is ≤±1℃. Normal stress is applied using a servo motor loading system, with the application rate controlled at 5kPa / min~10kPa / min to avoid sudden stress increases that could cause sample disturbance. The consolidation time is maintained for 3 hours, and the consolidation deformation is monitored by a displacement sensor. Consolidation is considered complete when the deformation is ≤0.01mm within 1 hour. The shear rate is set to 1mm / min, which closely approximates the actual scenario of slow shear deformation of the foundation in marine engineering. Data acquisition uses a dynamic signal acquisition instrument with a sampling frequency of 10Hz to simultaneously record shear displacement, shear stress, and real-time temperature data. Each working condition is repeated 3 times, and outliers (data with deviations >10%) are removed before taking the average value to further reduce data dispersion.

[0031] As a better alternative, the temperature fluctuation range of the external temperature control chamber of the customized temperature control interface shearing device is controlled within ±0.5℃, the normal stress loading accuracy is ±0.1kPa, and the shear stress measurement accuracy is ±0.01kPa, ensuring the accuracy and reliability of experimental data and providing high-quality raw data for subsequent database construction.

[0032] Step 3: Database Construction: Integrate the experimental data obtained in Step 2, divide the data into training and test sets, define input and output parameters and perform normalization processing to form a standardized database.

[0033] In some specific implementations, the database construction process is as follows: integrate the experimental data of S1 sea sand and S2 sea sand under 5 temperatures and 3 normal stresses in 3 repeated experiments, remove invalid data, and form 2100 sets of valid datasets; divide the training set (1680 sets) and the test set (420 sets) in an 8:2 ratio, and use random sampling in the division process to ensure that the working condition distribution of the two sets of data is consistent.

[0034] The input parameters are defined as four characteristic indicators: sand type (sea sand particle size), normal stress, temperature, and shear displacement. The output parameter is the predicted shear stress value. All input and output parameters are standardized using the Min-Max normalization method to eliminate the influence of differences in dimensions and numerical ranges, ultimately forming a structural specification and numerically standardized shear stress prediction database.

[0035] The criteria for determining the validity of the data in the effective dataset are as follows: the shear stress-displacement curve must show a clear peak segment and residual segment, and the peak stress deviation of three repeated experiments must be ≤8%; the working condition distribution of the training and test sets must ensure that all combinations of variables "sea sand particle size-temperature-normal stress" are covered by experimental data, and the sample size of each combination must be proportionate to the overall dataset to meet the model training requirements, such as proportional allocation or allocation according to the actual working condition weights, to avoid insufficient model generalization ability due to missing data or imbalanced proportions of certain combinations. After normalization, data verification is performed using Excel or Matlab to ensure no calculation errors.

[0036] Step 4: Model Training and Optimization: Based on programming software, four types of machine learning models, namely BPANN, ELM, CNN, and PSO-CNN, are constructed. The algorithm parameters are set and the model is optimized using RMSE as the fitness function to improve the fitting ability to nonlinear relationships.

[0037] In some specific implementations, the BPANN model adopts a "4-10-1" network structure and uses the Sigmoid function as the activation function; the ELM adopts a "4-25-1" network structure, selects the Sigmoid function as the hidden layer activation function, sets the input layer to hidden layer weights and hidden layer thresholds to be randomly initialized, with a value range of [-1, 1], and uses the least squares method to directly solve for the output layer weights; the number of hidden layer neurons is fixed at 25, the input layer neurons correspond to 4 feature parameters (temperature, particle size, normal stress, shear displacement), and the output layer neurons correspond to the predicted shear stress value; the CNN model adopts a "4-32-64-16-1" network structure design, uses the ReLU function as the activation function to introduce nonlinear feature mapping; the initial learning rate is set to 0.005, the optimizer is AdamW, and a cosine annealing learning rate scheduling strategy is used to achieve dynamic adjustment of the learning rate, taking into account both the global search capability in the early stage of model training and the local convergence accuracy in the later stage.

[0038] PSO The CNN model is a hybrid intelligent model combining Particle Swarm Optimization (PSO) algorithm and Convolutional Neural Network (CNN). Its core is to use PSO to optimize the hyperparameters (such as the number of hidden layer nodes, kernel size, and learning rate) or weights of the CNN, balancing the feature extraction capabilities of CNN with the global optimization capabilities of PSO, achieving a balance between accuracy and efficiency in nonlinear prediction tasks. In the hyperparameter optimization stage of the PSO-CNN shear stress prediction model, precise constraints are imposed on the core control parameters of the PSO algorithm: the particle velocity search range is strictly limited to the [-2,2] interval to avoid excessive oscillations during the hyperparameter space search; simultaneously, the particle position range is constrained to the [-1,1] interval to ensure efficient optimization within the standardized search domain. Through multiple rounds of iterative testing and performance verification, the optimal number of hidden layer nodes for the PSO-CNN model was finally determined to be 10. Under this configuration, the PSO-CNN model achieves the optimal balance between fitting accuracy (such as R², RMSE) and computational efficiency (training time, memory usage) in the shear stress prediction task.

[0039] The core operating parameters of the PSO algorithm are specifically set as follows: the particle population size is 20, the maximum number of iterations is 100, the inertia weight adopts a dynamic adjustment strategy, the initial value is set to 0.98, and it decreases linearly to 0.4 as the iteration progresses, taking into account the global search capability in the early stage of the algorithm and the local optimization accuracy in the later stage; in addition, the cognitive learning factor (c1) and social learning factor (c2) of the algorithm are both set to 1.5, balancing the learning weight of particles on their own historical best position and global best position, ensuring the stability and convergence of hyperparameter optimization.

[0040] Step 5: Model accuracy verification and prediction formula construction: Using the correlation coefficient (R²) 2 The model performance was evaluated using the root mean square error (RMSE) and mean absolute percentage error (MAPE), and the optimal prediction model was selected and determined. Based on the optimal prediction model, a formula for predicting the shear stress at the sea sand-geomembrane interface was established.

[0041] In some specific implementations, the correlation coefficient (R²) between the training set and the test set of each model is calculated. 2 The root mean square error (RMSE) and mean absolute percentage error (MAPE) are used to select the optimal model, with R being the preferred choice. 2 The model that is close to 1 and minimizes RMSE and MAPE.

[0042] As a further optimization, model accuracy verification also requires an operational condition adaptability test. Typical high-temperature conditions (80℃, 15kPa) and low-temperature conditions (-5℃, 50kPa) in tropical coastal areas are selected, and the deviations between the model's predicted values ​​and the measured values ​​are compared. A deviation ≤8% is considered an optimal fit. Ultimately, the PSO-CNN model is determined as the optimal model, with training set RMSE=2.1, MAPE=8%, and R0=2.1. 2 =0.98, test set RMSE=2.2, MAPE=9%, R 2 =0.97, which can accurately capture the variation law of interfacial shear stress under the coupling of multiple factors.

[0043] Based on the established optimal PSO-CNN model, a formula for predicting the shear stress at the sea sand-geomembrane interface was constructed according to the coupling law of "temperature-particle size-normal stress-shear displacement-shear stress," and applied to engineering applications. Sensitivity analysis of the PSO-CNN model shows that the cumulative contribution rate of normal stress and sea sand particle size exceeds 60%, playing a dominant role in controlling the interfacial shear stress. Based on this sensitivity law, suggestions for adjusting engineering parameters are proposed, providing direct reference for engineers without machine learning backgrounds in design stages such as sea sand selection for roadbeds and optimization of geomembrane laying spacing. The formula for predicting the shear stress at the sea sand-geomembrane interface is as follows: Where Y is the shear stress, X1 is the temperature (reflecting the thermal state of the construction and operation environment), X2 is the particle size (representing a key indicator of sea sand gradation), X3 is the normal stress (reflecting the ballast effect of the subgrade or superstructure load on the interface), and X4 is the shear displacement (characterizing the relative displacement between the geomembrane and the sea sand interface). These X values ​​are the basic inputs for calculating the interface shear stress, and their values ​​directly affect the calculation result of the shear stress Y, and are also adjustable design parameters in engineering. C0 is a constant term in the formula, representing the basic value of the shear stress without depending on any characteristics; A i (i=1 to 4) are the coefficients of each characteristic linear term, corresponding to the linear influence weights of temperature X1, particle size X2, normal stress X3, and shear displacement X4 on shear stress, respectively; B1 is the coefficient of the quadratic shear displacement term, used to capture the nonlinear influence of shear displacement on shear stress; D k (k=1 to 2) are the coefficients of the interaction term, corresponding to the shear displacement X4 and other features. The influence weight when combining; and These are the characteristics involved in the interaction, specifically referring to the grain size and normal stress that interact with shear displacement.

[0044] This embodiment details the technical path for constructing a prediction model for the interfacial shear stress of marine siliceous sand and geomembrane. Specifically, two types of sea sand with different particle sizes are first obtained by sieving. The moisture content is controlled at 9.65%, and the sand is layered and filled into an upper shear box. Simultaneously, HDPE geomembrane samples of specific specifications and roughness are prepared and fixedly laid at the bottom of the upper shear box. Then, using a customized temperature-controlled interfacial shearing device, shear experiments are conducted under a wide temperature range of -5℃ to 80℃ and normal stresses of 15kPa / 25kPa / 50kPa. Equipment debugging, stress application and consolidation, shearing, and data acquisition are completed according to standard procedures. Subsequently, 2100 sets of valid experimental data are integrated, and the training and test sets are randomly divided at an 8:2 ratio. A standardized database was constructed through normalization. Then, four machine learning models—BPANN, ELM, CNN, and PSO-CNN—were built and their corresponding parameters (such as the number of particles and iterations for PSO-CNN) were set. Finally, the model performance was evaluated through R², RMSE, MAPE metrics and extreme condition adaptability tests. PSO-CNN was determined to be the optimal model, with training set RMSE=2.1, MAPE=8%, and R²=0.98, and test set RMSE=2.2, MAPE=9%, and R²=0.97. Furthermore, a shear stress prediction formula incorporating input parameters such as temperature, particle size, normal stress, and shear displacement was developed based on this model, providing support for engineering applications.

[0045] Example 2 This embodiment takes the prediction of the mechanical parameters of the sea sand-geomembrane interface in a tropical coastal roadbed project in Hainan (design depth 0.75m~2.5m) as an example to verify the effectiveness of the "method for constructing a prediction model of the interfacial shear stress between the geomembrane layer and sea sand of different particle sizes considering the influence of temperature". The specific steps are as follows: Step 1: Project Background and Parameter Determination Geological exploration revealed the characteristics of the marine sand in the coastal roadbed: the regional marine sand is mainly quartz sand, and sieving yielded two commonly used engineering particle sizes—S1 and S2 marine sand; the surface layer of the roadbed is affected by tropical high temperatures, while the deeper layers are affected by the constant temperature of groundwater, with extreme low temperatures reaching -5℃ in the cold season, requiring coverage for temperature conditions ranging from -5℃ to 80℃; combined with the roadbed depth, the normal stresses were determined to be 15kPa, 25kPa, and 50kPa. Experiments were conducted in typical areas of the roadbed to estimate material usage: S1 marine sand was used to fill shear boxes (300mm×300mm×150mm), and S2 marine sand was used to fill shear boxes; the HDPE geomembrane layer was cut to a size of 280mm×460mm (shear area 300mm×300mm).

[0046] Step 2: Preparation of experimental samples Sea sand sample preparation: S1 and S2 sea sand were controlled at the optimum moisture content of 9.65% and the humidity was controlled by a "weighing-spraying-stirring-retesting" cycle process with a 5-minute interval between each cycle. The moisture content error was ≤±0.5%. The shear box was filled with a layered filling method, with each layer being 25mm thick, for a total of 3 layers. After filling, the layers were compacted with a light compactor to ensure that the density deviation of each layer was ≤±2%. After filling, the flatness of the top surface of the sample was measured with a dial indicator, and the deviation was ≤0.5mm.

[0047] Geomembrane sample preparation: HDPE sheets were selected, and the three-dimensional surface roughness parameter Rs=1.06 (before the experiment) was measured and controlled by a laser profilometer. The sheets were cut into 280mm×460mm sizes using a CNC cutting machine and fixed with positioning pins when laid at the bottom of the upper shear box to avoid interface misalignment during the shearing process.

[0048] Step 3: Wide-condition interface shearing experiment like Figure 4 As shown, an experiment was conducted using a customized temperature-controlled shearing device. The specific operation is as follows: Equipment debugging: Fill the upper shear box with S1 or S2 sea sand sample, close the ambient temperature control chamber, set the experimental temperature (-5℃, 20℃, 40℃, 60℃, 80℃), and keep it at a constant temperature for 2 hours (monitor the internal temperature difference of the sample ≤±1℃ through a platinum resistance temperature sensor). Normal stress application and consolidation: 15 kPa, 25 kPa and 50 kPa normal stresses were applied by a servo motor loading system at an application rate of 8 kPa / min (to avoid disturbing the sample). Consolidation was carried out for 3 hours. When the deformation was ≤0.01 mm within 1 hour, consolidation was considered complete. Shearing and Data Acquisition: The shear box was moved horizontally at a shearing rate of 1 mm / min (maximum shearing distance 70 mm). A dynamic signal acquisition device was used to synchronously record shear displacement, interface shear stress, and real-time temperature data. Each working condition was repeated three times, and outliers with deviations >10% were removed; the average value was taken as the valid data. Figure 5 , Figure 6 As shown, the experimental results indicate that under conditions of 40℃ and 50kPa, the peak shear stress at the interface of S2 sea sand is 32.53kPa, which is 15.95% higher than that of S1 sea sand (27.34kPa). Under conditions of 80℃ and 15kPa, the peak shear stress of S1 sea sand is 13.77% lower than that at 20℃, which conforms to the rule of "interface strength decreases under high temperature and low stress, and strength increases under high temperature and high stress". By improving the interface data of sea sand with multiple particle sizes, the database is ensured to cover two types of commonly used sea sand in engineering, thus improving the model's adaptability to different particle sizes.

[0049] Step 4: Database Construction and Model Training Optimization Database construction: Effective data of S1 and S2 sea sand under 5 temperatures and 3 normal stresses were integrated to form a dataset of 2100 sets; the training set (1680 sets) and the test set (420 sets) were randomly divided at an 8:2 ratio to ensure that the "particle size-temperature-stress" working condition distribution of the two sets of data were consistent; the input parameters (sea sand particle size, temperature, normal stress, shear displacement) were normalized (value range [-1,1]), and the data were verified by Matlab to have no calculation errors.

[0050] Model Construction and Optimization: Constructing four types of models: BPANN, ELM, CNN, and PSO-CNN. BPANN employs a "4-10-1" network structure, adjusts the activation and training functions, and uses a pre-tuned learning rate.

[0051] ELM adopts a "4-25-1" network structure, selects the Sigmoid function as the hidden layer activation function, sets the input layer to hidden layer weights and hidden layer thresholds to be randomly initialized, with a value range of [-1,1], and uses the least squares method to directly solve the output layer weights; the number of hidden layer neurons is fixed at 25, the input layer neurons correspond to 4 feature parameters (temperature, particle size, normal stress, shear displacement), and the output layer neurons correspond to the predicted value of shear stress.

[0052] The CNN uses a "4-32-64-16-1" network structure, with adjustments to the activation function and training function (AdamW optimizer + cosine annealing scheduling).

[0053] PSO algorithm parameter settings: number of particles 20, maximum number of iterations 100, inertia weight 0.4~0.98, velocity range [-2,2]. The initial weights and threshold of the CNN were optimized using RMSE as the fitness function. Training results are shown in Table 1: The PSO-CNN model training set showed RMSE=2.1, MAPE=8%, and R=0.98, indicating the best fitting effect. Figure 7 As shown, it far surpasses BPANN (training set RMSE=9.63, MAPE=12.36%).

[0054] Step 5: Model Accuracy Verification and Engineering Application Accuracy verification: such as Figure 8As shown, the generalization ability of the model was verified using a test set (420 groups). The PSO-CNN model test set showed RMSE=2.2, MAPE=9%, and R=0.97, with the deviation between the predicted and measured values ​​≤8%. Extreme working conditions were selected for verification: at 80℃ and 15kPa, the predicted value of the interfacial shear stress of S2 sea sand was 14.1kPa, and the measured value was 14.3kPa, with a deviation of 1.4%; at -5℃ and 50kPa, the predicted value of S1 sea sand was 25.8kPa, and the measured value was 26.2kPa, with a deviation of 1.5%, which meets the engineering accuracy requirements.

[0055] The formula for predicting shear stress at the sea sand-geomembrane interface, built based on the PSO-CNN model, can provide parameter adjustment suggestions and can be directly used by engineers without machine learning background in design stages such as sea sand selection for roadbeds and optimization of geomembrane layer laying spacing.

[0056] Engineering Applications: Guiding Subgrade Design Based on Model Output Parameters For deep subgrade (50kPa), S2 sea sand is preferred because its peak interfacial shear stress at 80℃ is 15.95% higher than that of S1 sea sand, which can improve the subgrade's anti-slip ability. In tropical high-temperature areas (80℃), the recommended spacing for laying geomembrane layers is ≤1.5m (based on the stress attenuation law predicted by the model) to avoid interface shear slip. After 6 months of on-site monitoring, the subgrade settlement was ≤4cm, which meets the design requirements and verifies the guiding value of the model for engineering practice.

[0057] Example 2 uses a tropical coastal roadbed project in Hainan as an application scenario to verify the effectiveness of the proposed prediction model construction method. Specifically, the following steps are taken: First, geological exploration is used to determine the required particle sizes of two types of sea sand (S1 and S2), as well as working parameters such as temperatures from -5℃ to 80℃ and normal stresses of 15kPa / 25kPa / 50kPa, to estimate material usage. Then, sea sand samples (controlling moisture content, layered filling, and ensuring compaction and flatness) and HDPE geomembrane samples (controlling roughness and specifications) are prepared according to specifications. Finally, a customized temperature-controlled interface shearing device is used to complete equipment debugging, normal stress application and consolidation, shearing, and data acquisition operations. The experiment was repeated three times under each working condition, and outliers were removed. Then, the data was integrated to form a dataset of 2100 sets, which were divided into training and test sets in an 8:2 ratio and normalized. Four types of machine learning models were constructed and their parameters were optimized. Among them, the PSO-CNN model had the best fitting effect. Finally, the model was verified on the test set and under extreme working conditions. The deviation between the predicted value and the measured value was ≤8%, which met the engineering accuracy requirements. Based on this, design suggestions were provided for the engineering, such as the selection of sea sand (S2 sea sand is preferred for deep subgrade) and the spacing of geomembrane laying (≤1.5m in high-temperature areas). The subgrade settlement monitored on site for 6 months met the design requirements, which verified the engineering practical value of the model.

[0058] This invention proposes a method for constructing a prediction model for the interfacial shear stress of marine silica sand and geomembrane. The core of this method involves preparing geomembrane samples by sieving two or more types of marine silica sand with different particle sizes. A customized temperature-controlled interfacial shearing device is used to conduct wide-condition (-5℃ to 80℃ temperature range, multiple particle sizes, multiple normal stresses) interfacial shearing experiments. After collecting valid data, training and testing sets are divided and normalized to construct a standardized database. Four types of machine learning models—BPANN, ELM, CNN, and PSO-CNN—are then constructed. The optimal PSO-CNN model is evaluated and selected using indicators such as correlation coefficient R², root mean square error RMSE, and mean absolute percentage error MAPE. A corresponding shear stress prediction formula is also established. This method overcomes the limitations of traditional experimental equipment, such as narrow temperature range and limited adaptability, and realizes multi-factor coupled research of "temperature-particle size-stress". The prediction accuracy of the PSO-CNN model is significantly improved (R² reaches 0.97 and MAPE is 9% on the test set). Moreover, the cost of a single customized device is reduced by more than 40%, and the time for acquiring parameters under multiple working conditions is shortened by 20% to 25%. It has constructed an integrated technical system of "multi-working condition experiment-high-precision model-engineering application", which effectively overcomes the problems of incomplete data, low prediction accuracy and high cost of traditional methods. It can provide efficient and accurate interface mechanical parameter support for marine engineering such as port terminals and coastal roadbeds, and effectively guide the design of marine sand selection, geomembrane laying optimization and other aspects in engineering.

[0059] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the invention by those skilled in the art without departing from the spirit and essence of the invention. Such modifications or substitutions should all fall within the scope of the invention, or any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the invention should be covered within the protection scope of the invention. Therefore, the protection scope of the invention should be determined by the scope of the claims.

Claims

1. A method for constructing a prediction model for shear stress at the marine siliceous sand-geomembrane interface, characterized in that, The method includes: Two or more marine silica sand samples with different particle sizes were screened, and geomembrane samples were prepared at the same time. The geomembrane samples were laid at the bottom of the upper shear box of a customized temperature-controlled interface shearing device. The marine silica sand samples were layered and filled into the upper surface of the geomembrane in the upper shear box to ensure stable contact between the marine silica sand samples and the geomembrane. Using the customized temperature-controlled interface shearing device, the device debugging, normal stress application and consolidation, shearing and data acquisition operations were completed in sequence to obtain interface mechanical data under different marine silica sand particle sizes, temperatures and normal stress conditions. Based on the interface mechanics data, a training set and a test set are divided, input and output parameters are defined and normalized to form a standardized database. Construct four types of machine learning models: BPANN, ELM, CNN, and PSO-CNN; The performance of the four types of machine learning models was evaluated, the optimal prediction model was selected and determined, and a prediction formula for the shear stress at the marine silica sand-geomembrane interface was built based on the optimal prediction model.

2. The method according to claim 1, characterized in that, The screening of marine silica sand samples with two or more different particle sizes includes: screening marine silica sand samples with two or more different particle sizes using a standard vibrating sieve with apertures of 1 mm, 2 mm, and 4 mm, with the screening time controlled at 10 min to 15 min; the moisture content of the marine silica sand is controlled by a "weighing-spraying-stirring-retesting" cyclic process, with a 5 min interval between each cycle, until the moisture content error is ≤ ±0.5%, and the moisture content of the marine silica sand sample is measured and controlled to be 9.65%.

3. The method according to claim 2, characterized in that, The layered filling process includes: filling each layer with marine silica sand samples of the same thickness, filling them into shear boxes of a certain size, and compacting them using a lightweight compactor during the filling process to ensure that the density deviation of each layer is ≤±2%; In addition, the filling mass of the marine silica sand sample is calculated based on the shear box volume and the density of the marine silica sand. After filling, the flatness of the top surface of the sample is measured using a dial gauge to ensure that the flatness deviation is ≤0.5mm.

4. The method according to claim 3, characterized in that, The preparation process of the geomembrane sample is as follows: a high-density polyethylene (HDPE) sheet with a thickness of 2 mm and a density of 0.942 g / cm³ is selected. The three-dimensional roughness parameter Rs of the HDPE sheet surface is measured and controlled to be 1.06 by a laser profilometer. The sample is cut into 280 mm × 460 mm specifications by a CNC cutting machine. When laid at the bottom of the upper shear box, it is fixed by positioning pins to ensure that the shearing area is stable at 300 mm × 300 mm. In addition, the surface roughness of the geomembrane sample is controlled by sandblasting. The sandblasting material is quartz sand with a particle size of 0.1 mm to 0.3 mm, and the sandblasting pressure is controlled to be 0.2 MPa to 0.3 MPa.

5. The method according to claim 4, characterized in that, The customized temperature-controlled interface shearing device sequentially completes equipment debugging, normal stress application and consolidation, shearing and data acquisition operations, specifically including: The equipment is debugged by filling the upper shear box with either the S1 or S2 marine silica sand sample, closing the environmental temperature control chamber, setting the experimental temperature, turning on the temperature control system to maintain the temperature for 2 hours, and monitoring the temperature through a platinum resistance temperature sensor embedded inside the sample to ensure that the internal temperature difference of the sample is ≤±1℃. The normal stress is applied using a servo motor loading system, with the application rate controlled between 5 kPa / min and 10 kPa / min to avoid sudden stress increases that could cause sample disturbance. The consolidation time is maintained at 3 hours, and the consolidation deformation is monitored by a displacement sensor. When the deformation is ≤0.01 mm within 1 hour, the consolidation is considered complete. The shearing rate was set to 1 mm / min; the data acquisition was performed using a dynamic signal acquisition instrument with a sampling frequency of 10 Hz, synchronously recording shear displacement, shear stress, and real-time temperature data. The experiment was repeated multiple times for each working condition, and the average value was taken after removing outliers, where the outliers were data with a deviation > 10%.

6. The method according to claim 5, characterized in that, The process of dividing the interface mechanics data into training and testing sets, defining input and output parameters, and performing normalization to form a standardized database specifically includes: We integrated repeated experimental data from S1 marine siliceous sand and S2 marine siliceous sand under various temperatures and normal stresses, removed invalid data, and formed multiple sets of valid datasets. The training set and the test set were divided in an 8:2 ratio, and the division process used random sampling to ensure that the working condition distribution of the two sets of data was consistent. The input parameters are defined as four characteristic indicators: marine silica sand particle size, normal stress, temperature, and shear displacement. The output parameter is the predicted value of shear stress. All input and output parameters are standardized using the Min-Max normalization method to eliminate the influence of differences in dimensions and numerical ranges, thus forming a structurally standardized and numerically standardized shear stress prediction database. The criteria for determining the validity of data in the effective dataset are as follows: the shear stress-displacement curve must show a peak segment and a residual segment, and the peak stress deviation of the three repeated experiments must be ≤8%; the working condition distribution of the training set and the test set must ensure that all combinations of variables "marine silica sand particle size-temperature-normal stress" are covered by experimental data, and the proportion of the sample size under each combination to the overall dataset must be adapted to the model training requirements; and the data must be verified using Excel or Matlab after normalization.

7. The method according to claim 6, characterized in that, The construction of the four types of machine learning models—BPANN, ELM, CNN, and PSO-CNN—specifically includes: The BPANN model uses a "4-10-1" network structure and selects the Sigmoid function as the activation function. The ELM model uses a "4-25-1" network structure, selects the Sigmoid function as the hidden layer activation function, sets the input layer to hidden layer weights and hidden layer thresholds to be randomly initialized, with values ​​ranging from [-1, 1], and uses the least squares method to directly solve for the output layer weights. The number of hidden layer neurons is fixed at 25. The input layer neurons correspond to four feature parameters: temperature, particle size, normal stress, and shear displacement, and the output layer neurons correspond to the predicted value of shear stress. The CNN model uses a "4-32-64-16-1" network structure design, selects the ReLU function as the activation function to introduce nonlinear feature mapping, sets the initial learning rate to 0.005, uses the AdamW optimizer, and uses a cosine annealing learning rate scheduling strategy to achieve dynamic adjustment of the learning rate. In the hyperparameter optimization stage of shear stress prediction, the PSO-CNN model constrains the core control parameters of the PSO algorithm: limiting the particle velocity search range to the [-2,2] interval; and simultaneously constraining the particle position range to the [-1,1] interval. Through multiple rounds of iterative testing and performance verification, the optimal number of hidden layer nodes in the PSO-CNN model was finally determined to be 10. The core operating parameters of the PSO algorithm are specifically set as follows: the particle population size is 20, the maximum number of iterations is 100; the inertia weight adopts a dynamic adjustment strategy, with an initial value of 0.98, which decreases linearly to 0.4 as the iteration progresses; in addition, the cognitive learning factor c1 and the social learning factor c2 of the algorithm are both set to 1.5 to balance the learning weight of particles on their own historical best position and global best position.

8. The method according to claim 7, characterized in that, The evaluation of the performance of the four types of machine learning models, and the selection and determination of PSO-CNN as the optimal prediction model, specifically includes: Calculate the correlation coefficient R between the training set and the test set of the four types of machine learning models. 2 Root mean square error (RMSE) and mean absolute percentage error (MAPE) are used to select R based on the results. 2 The model that is close to 1 and has the smallest RMSE and MAPE is the optimal model.

9. The method according to claim 8, characterized in that, The accuracy verification of the four types of machine learning models also needs to include a working condition adaptability test. A high-temperature working condition of 80℃ and an ambient pressure of 15kPa and a low-temperature working condition of -5℃ and an ambient pressure of 50kPa are selected, and the deviation between the model prediction value and the measured value is compared. If the deviation is ≤8%, it is considered to be adaptable.

10. The method according to claim 9, characterized in that, The method for predicting the shear stress at the marine siliceous sand-geomembrane interface based on the PSO-CNN specifically includes: constructing a shear stress prediction formula for the marine siliceous sand-geomembrane interface based on the coupling law of "temperature-particle size-normal stress-shear displacement-shear stress" of the determined optimal prediction model PSO-CNN, and applying it to engineering applications. The shear stress prediction formula for the marine siliceous sand-geomembrane interface is as follows: Where Y is shear stress; X1 is temperature, reflecting the thermal state of the construction and operation environment; X2 is particle size, representing a key indicator of marine siliceous sand gradation; X3 is normal stress, reflecting the ballast effect of the subgrade or superstructure load on the interface; X4 is shear displacement, characterizing the relative displacement between the geomembrane and the marine siliceous sand interface; C0 is a constant term in the formula, representing the basic value of shear stress without depending on any characteristics; A i These are the coefficients of each characteristic linear term, corresponding to the linear influence weights of temperature X1, particle size X2, normal stress X3, and shear displacement X4 on shear stress, respectively; B1 is the coefficient of the quadratic shear displacement term, used to capture the nonlinear influence of shear displacement on shear stress; D k These are the coefficients of the interaction term, corresponding to the shear displacement X4 and other features. The influence weight when combining; and These are the characteristics involved in the interaction, referring to the grain size and normal stress that act together with the shear displacement.

Images