Membrane bag sand bank slope settlement prediction method and system based on neural network

Abstract

The invention discloses a membrane bag sand levee slope settlement prediction method and system based on a neural network, and belongs to the technical field of prediction systems.The method comprises the steps that construction parameters, foundation monitoring data and geological exploration data of a membrane bag sand levee slope are obtained; performing feature fusion by analyzing correlation between the construction parameters and the compression amount of each foundation soil layer in the time dimension and between the foundation monitoring data and the compression amount of each foundation soil layer to obtain fusion time sequence features; and on the basis of the geological exploration data, using a settlement prediction model and fusion time sequence characteristics, predicting each compression amount by analyzing the pore drainage rate of each foundation soil layer, outputting a settlement prediction result of the membrane bag sand bank slope, identifying the settlement abnormal risk of the membrane bag sand bank slope, and outputting the settlement risk grade of a preset construction point. Therefore, by implementing the method and the device, the problem that the predicted total settlement amount of the film bag sand bank slope is not accurate enough due to the fact that the influence of the bank slope foundation structure on the settlement amount in the construction process is not completely considered in the prior art can be solved.

Description

Technical Field

[0001] This application belongs to the field of prediction system technology, specifically relating to a method and system for predicting settlement of membrane bag sand embankment slopes based on neural networks. Background Technology

[0002] Membrane bag sand embankments are flexible embankment structures formed by filling geomembrane bags with sand. They offer advantages such as high mechanization, fast construction speed, good overall stability, and strong adaptability, and are commonly used in embankment projects such as coastal protection and riverbank reinforcement. On soft soil foundations, membrane bag sand embankments can serve as a load-dispersing layer to reinforce the foundation and improve its bearing capacity. Therefore, the settlement of membrane bag sand embankments is a core indicator for assessing embankment stability and engineering safety. Large settlement over a long period may lead to risks such as embankment cracking and uneven settlement.

[0003] The commonly used layered summation method involves creating a geological profile of the foundation to obtain multiple foundation soil layers. The compression of each layer is calculated using construction parameters, and then these layers are summed to obtain the total settlement of the embankment slope. This method relies on the premise that the load across the entire foundation is uniformly distributed. However, in actual construction, structural differences in the foundation soil layers lead to varying drainage rates in different spaces under the same construction parameters, resulting in uneven load distribution. Therefore, the total settlement of the membrane bag sand embankment slope calculated using this method has significant errors and cannot provide effective data support for subsequent embankment reinforcement and adjustment of construction rates. Summary of the Invention

[0004] This application proposes a neural network-based method and system for predicting the settlement of membrane bag sand embankments. This method can solve the problem that existing technologies do not fully consider the impact of the embankment foundation structure on the settlement during construction, resulting in inaccurate predictions of the total settlement of membrane bag sand embankments.

[0005] The first aspect of this application provides a method for predicting settlement of membrane bag sand embankment slopes based on neural networks, the method comprising: Acquire construction parameters, foundation monitoring data, and geological exploration data of membrane bag sand embankment slopes at several preset time points; Using the compression amount of the foundation soil layer in each region as the target feature, feature fusion is performed by analyzing the correlation between the construction parameters and the foundation monitoring data and the target feature in the time dimension to obtain fused temporal features; Using the geological exploration data as constraints, the fused time-series features are input into the preset settlement prediction model. The compression of the foundation soil layer in each region is predicted by analyzing the pore drainage rate of the foundation soil layer in each region, and the settlement prediction results of the membrane bag sand embankment slope are output. Based on the settlement prediction results, the abnormal settlement risk of the membrane bag sand embankment is identified through a preset dynamic early warning mechanism, and the settlement risk level of the membrane bag sand embankment at each preset construction point is output.

[0006] The above scheme takes into account that under complex geological conditions, it is difficult to accurately predict the settlement of membrane bag sand embankments using only traditional empirical formulas and single models. Therefore, it uses historical construction parameters and foundation monitoring data related to embankment settlement. First, it combines data from different sources through feature fusion to obtain fused time-series features closely related to embankment settlement, effectively improving the accuracy of subsequent predictions. Then, a settlement prediction model is used to obtain the load intensity applied to the embankment through construction parameters and to calculate the pore drainage rate of the foundation soil layer through foundation monitoring data. This allows for analysis of the effective stress distribution directly related to settlement after pore water stress is eliminated, thus obtaining accurate compression of the foundation soil layer at each location after pore water drainage. Finally, a dynamic early warning mechanism is used to discriminate the settlement prediction results, gradually identifying settlement anomalies at various locations in the membrane bag sand embankment and assessing the anomaly risk level, providing decision support for the safety of subsequent construction.

[0007] In one possible implementation of the first aspect, construction parameters, foundation monitoring data, and geological exploration data of the membrane bag sand embankment slope are obtained at several preset time points, specifically as follows: Based on the structural form and geological profile results of the foundation of the membrane bag sand embankment slope, multiple preset measurement points were set on the foundation soil layers in each location; For the preset measurement points, pore water pressure data and horizontal displacement data are collected at each preset time point to obtain the foundation monitoring data, and the applied load intensity and construction rate corresponding to each preset time point are obtained to obtain the construction parameters. Based on the geological profile results, the thickness, initial void ratio, consolidation coefficient, and compression index of the foundation soil layers in various locations were measured to obtain the geological exploration data.

[0008] The above scheme, based on the foundation's structural form and the nature of the base layer in different locations, sets up multiple pre-defined measurement points in the subgrade soil layer for data collection and settlement prediction. Then, dynamic data such as foundation monitoring data and construction parameters are collected as influencing factors and feedback data for settlement changes, while static data such as geological exploration data, which does not change over time, are collected as inherent conditions for settlement changes to adjust for unreasonable situations in the prediction process. This makes subsequent settlement predictions more realistic and improves the reliability of the prediction results.

[0009] In one possible implementation of the first aspect, the compression amount of each foundation soil layer is taken as the target feature. Feature fusion is performed by analyzing the correlation between the construction parameters and the foundation monitoring data and the target feature in the time dimension, resulting in a fused temporal feature. Specifically: Based on a preset domain knowledge base and the target features, the construction parameters and the foundation monitoring data are filtered using a feature association method to obtain core associated features; Each of the core associated features is mapped onto a preset target time axis for feature alignment to obtain a core feature matrix; Based on the attention mechanism, the correlation score between each core associated feature is calculated through the core feature matrix, and the core associated features are fused to obtain the fused temporal features.

[0010] The above approach reduces data volume and improves the efficiency of subsequent prediction calculations by filtering core features highly correlated with the target features through a domain knowledge base. Then, feature alignment enhances data quality. Finally, attention mechanisms are used for temporal feature fusion, which automatically identifies the dynamic factors most relevant to settlement, reduces noise interference, and improves the specificity of feature representation and prediction accuracy.

[0011] In one possible implementation of the first aspect, based on a preset domain knowledge base and the target features, feature filtering is performed on the construction parameters and the foundation monitoring data using a feature association method to obtain core associated features, specifically: The settlement response curve related to the membrane bag sand embankment slope is extracted from the domain knowledge base. Based on the variables of the settlement response curve, the first associated feature that is directly related to the target feature is screened from the construction parameters and the foundation monitoring data. Extract causal relationships with the first associated feature as a variable from the domain knowledge base, and based on the causal relationships, filter out second associated features that are indirectly associated with the target feature from the construction parameters and the foundation monitoring data; Outlier removal and continuity correction are performed on the first and second related features to obtain the core related features.

[0012] The above scheme combines settlement response curves with causal relationships to screen features from both direct and indirect correlation levels, which can more comprehensively capture the explicit and implicit factors affecting settlement; through outlier removal and continuity correction, it improves data quality and enhances model stability.

[0013] In one possible implementation of the first aspect, based on an attention mechanism, the correlation score between each core associated feature is calculated through a core feature matrix, and feature fusion is performed on the core associated features to obtain the fused temporal features, specifically as follows: Based on each time step of the target time axis, the core feature matrix is ​​encoded by a capture encoder to obtain the initial temporal representation corresponding to each core associated feature; Calculate the correlation score between each of the initial temporal representations and the target feature, normalize the correlation score, and obtain the attention weight of each of the core associated features; If the attention weight is greater than the first threshold, the corresponding initial temporal representation is weighted and summed to perform feature fusion on the core associated features, and the fused temporal features are output.

[0014] The above scheme achieves feature fusion at different time steps through attention weights, highlights the influence of key features, makes time series features more reflective of the dynamic process of subsidence development, and improves the model's ability to model time series.

[0015] In one possible implementation of the first aspect, feature extraction and feature stitching are performed on the fused temporal features and geological exploration data. The feature stitching results are input into a preset settlement prediction model. The compression of the foundation soil layers in each location is predicted by analyzing the pore drainage rate of the foundation soil layers. The settlement prediction results of the membrane bag sand embankment slope are then output, specifically: The geological exploration data is encoded into semantic vectors of a preset dimension, and the semantic vectors are divided according to the type of foundation soil layer to obtain the static geological characteristics of each foundation soil layer. By analyzing the impact of the static geological features on the changing trend of the foundation monitoring data, the response features corresponding to the static geological features are extracted from the fused time series features. Based on a preset time step, the static geological features and the response features are spliced ​​together to obtain the feature splicing result. The feature splicing results are input into the settlement prediction model based on long short-term memory network. The compression of the foundation soil layer in each region is predicted by analyzing the pore drainage rate of the foundation soil layer in each region, and the settlement prediction results are output. The prediction process is adjusted by using the static geological features as constraints.

[0016] The aforementioned scheme encodes geological exploration data into static geological features and splices them with temporal features, enabling the model to simultaneously consider the spatial variability of geological conditions and the temporal changes in the construction process. In predicting future settlement changes, it considers not only the loads applied during construction but also the pore drainage rate due to soil compression. Dynamic assessment allows for the effective stress distribution of the foundation when the embankment slope settles, thus more accurately reflecting the impact of soil consolidation on settlement and achieving precise prediction of settlement under complex geological conditions.

[0017] In one possible implementation of the first aspect, the compression of the foundation soil layers in each location is predicted by analyzing the pore drainage rate of the foundation soil layers, and the settlement prediction result is output, specifically as follows: Based on the feature splicing results, according to the porosity of the foundation soil layer in various places and the applied load intensity on the membrane bag sand embankment slope, the groundwater level data in the future time period is predicted to obtain the predicted pore drainage rate. The construction rate is extracted from the feature splicing results, and the effective stress distribution of the foundation soil layer in the vertical direction is obtained by comparing the construction rate with the predicted pore drainage rate. Based on the effective stress distribution and the actual thickness of the foundation soil layer in each location, the compression distribution of the foundation soil layer in each location is predicted, and the settlement prediction results are output.

[0018] In one possible implementation of the first aspect, the prediction process is adjusted based on the static geological features as constraints, specifically as follows: During the prediction process, the static geological features are used to perform boundary detection on the predicted values ​​of the response features within a future time period; If the boundary detection fails, the predicted value is adjusted to a preset safety range, and the adjusted predicted value is used to predict the compression distribution of the foundation soil layer in various locations. The safety range is the set value range of the response feature.

[0019] The above scheme uses static geological features to constrain the predicted changes, avoiding model outputs that violate geological common sense or engineering safety limits, thus enhancing the reliability and safety of the prediction.

[0020] In one possible implementation of the first aspect, based on the settlement prediction results, a preset dynamic early warning mechanism is used to identify the abnormal settlement risk of the membrane bag sand embankment slope, and the settlement risk level of the membrane bag sand embankment slope at each preset construction point is output, specifically: Based on the settlement prediction results, draw a spatial continuous settlement distribution map of the predicted settlement at each preset construction point; The preset predicted settlement cloud map is superimposed on the settlement continuous distribution map. Based on the superposition result, the first preset construction point and its abnormal pattern that the settlement gradient exceeds the first threshold are identified, as well as the second preset construction point and its abnormal pattern that there is asymmetry in settlement on both sides of the dike axis; wherein, the settlement gradient is the difference in settlement within a fixed horizontal distance. Based on the dynamic early warning mechanism and the abnormal mode, the degree of structural hazard at the first preset construction point and the second preset construction point are assessed respectively to obtain the corresponding settlement risk level.

[0021] The above scheme can intuitively identify dangerous patterns such as abnormal settlement gradient and asymmetric settlement by overlaying settlement distribution maps and cloud maps, and assess the risk level in combination with the degree of structural hazard, so as to realize closed-loop management from prediction to early warning and improve the timeliness and accuracy of engineering risk prevention and control.

[0022] The second aspect of this application provides a neural network-based membrane bag sand embankment slope settlement prediction system, the system comprising: a data acquisition module, a feature fusion module, a settlement prediction module, and a risk prediction module; The data acquisition module is used to acquire construction parameters, foundation monitoring data and geological exploration data of the membrane bag sand embankment slope at several preset time points; The feature fusion module is used to take the compression amount of the foundation soil layer in each region as the target feature, and perform feature fusion by analyzing the correlation between the construction parameters and the foundation monitoring data and the target feature in the time dimension to obtain fused time-series features; The settlement prediction module is used to take the geological exploration data as a constraint, input the fused time series features into the preset settlement prediction model, predict the compression of the foundation soil layer in each place by analyzing the pore drainage rate of the foundation soil layer in each place, and output the settlement prediction results of the membrane bag sand embankment slope. The risk prediction module is used to identify the abnormal settlement risk of the membrane bag sand embankment slope through a preset dynamic early warning mechanism based on the settlement prediction results, and output the settlement risk level of the membrane bag sand embankment slope at each preset construction point. Attached Figure Description

[0023] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the specific process of a neural network-based method for predicting the settlement of a membrane bag sand embankment slope, provided in an embodiment of this application. Figure 2 This is a structural diagram of a neural network-based membrane bag sand embankment settlement prediction system provided in one embodiment of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0027] First Embodiment Predicting the settlement of membrane bag sand embankments is essentially predicting the compressive deformation of the soil under its own weight and external loads. During construction, soil settlement is fundamentally a process of pore water removal. Initially, the applied load is converted into pore water pressure. Subsequently, water is discharged from the soil under the pressure gradient, pore water pressure dissipates, and the applied load gradually transforms into effective stress, causing soil compression. However, in predicting the settlement of membrane bag sand embankments, the varying pore water pressure at different locations in the foundation due to different soil structures is often overlooked, affecting the drainage rate of the soil layers and resulting in uneven distribution of effective stress. Therefore, this application addresses the uneven load distribution phenomenon commonly seen in complex geological conditions by introducing foundation monitoring data and geological exploration data to analyze the pore drainage rate at different locations in the soil layers. This allows for accurate prediction of load distribution and accurate forecasting of the compression of foundation soil layers in various locations. Ultimately, this yields settlement prediction results and settlement risk levels for membrane bag sand embankments with smaller errors, providing effective data support for subsequent adjustments to construction parameters.

[0028] like Figure 1 As shown, to address the problem in existing technologies that fail to fully consider the impact of the foundation structure of the embankment slope on settlement during construction, resulting in inaccurate predictions of the total settlement of the membrane bag sand embankment, the first embodiment of this application provides a detailed flowchart of a neural network-based method for predicting the settlement of membrane bag sand embankments. This neural network-based method for predicting the settlement of membrane bag sand embankments includes steps S1 to S4, detailed below: Step S1: Obtain construction parameters, foundation monitoring data, and geological exploration data of the membrane bag sand embankment slope at several preset time points.

[0029] This application primarily uses a combination of three types of data—construction parameters, foundation monitoring data, and geological exploration data—to predict the settlement of membrane bag sand embankment slopes. To obtain this multi-source data, multiple pre-set measurement points are first established on each foundation soil layer based on the structural form and geological profile results of the underlying foundation. The number of these pre-set measurement points in each foundation soil layer is mainly determined by the properties of that soil layer. For some foundation soil layers with strong spatial variability, more pre-set measurement points are set to improve the accuracy of the collected data.

[0030] Then, for these preset measurement points, pore water pressure data and horizontal displacement data of the foundation soil layer are collected at multiple discrete preset time points during construction to obtain foundation monitoring data for each foundation soil layer. The pore water pressure data at different time points can show the increase and dissipation process of excess pore pressure at different depths and locations in the foundation, serving as dynamic data for predicting the pore drainage rate at various points in the foundation. The horizontal displacement data is used to determine the stability of the foundation soil layer, describing the horizontal displacement of the foundation soil layer at each preset measurement point during construction; because extremely high pore water pressure can drive the soil to form significant horizontal displacement, thus affecting the pore drainage rate, it is also one of the indicators used to predict the pore drainage rate.

[0031] Simultaneously, construction parameters corresponding to each preset time point were collected, including the load intensity applied to the foundation and the membrane bag sand embankment slope, as well as the construction rate. This provides direct data for analyzing the effective stress distribution on the foundation. In addition, based on the geological profile results, the thickness, initial void ratio, consolidation coefficient, compression index, and compression modulus of each foundation soil layer were measured to obtain geological exploration data.

[0032] Among them, the consolidation coefficient, compression index, and compression modulus are all physical and mechanical properties of the foundation. The consolidation coefficient is a physical and mechanical property that characterizes the rate at which pore water is expelled and the soil compacts under load in saturated soil. A larger consolidation coefficient indicates that pore water is expelled faster and the foundation is compressed faster. The compression index characterizes the relationship between the void ratio of the foundation soil and the pressure, and can be used to calculate the final compression. The compression modulus is the ratio of vertical stress to the corresponding vertical strain when the soil undergoes vertical compression, and can also be used to calculate the final compression.

[0033] Step S2: Taking the compression amount of the foundation soil layer in each region as the target feature, feature fusion is performed by analyzing the correlation between the construction parameters and the foundation monitoring data and the target feature in the time dimension to obtain the fused temporal feature.

[0034] To quickly obtain the influencing factors related to the compression of the foundation soil layer during construction, this application embodiment also establishes a domain knowledge base as a knowledge set specific to this domain. Structured knowledge within the domain knowledge base is used to analyze feature data related to the target characteristics.

[0035] This application extracts correlation features from both direct and indirect correlation perspectives, with the compression amount of the foundation soil layer as the target feature. Specifically, if experimental data showing an independent-dependent variable relationship with the compression amount of the foundation soil layer can be directly obtained from the domain knowledge base, then the corresponding variable can be extracted from this experimental data as the first correlation feature directly correlated with the target feature.

[0036] Specifically, the settlement response curve related to the membrane bag sand embankment slope is extracted from the domain knowledge base. The dependent variable of this curve is the target feature. If the corresponding independent variable exists in the construction parameters or the foundation monitoring data, the independent variable is extracted as the first associated feature.

[0037] For example, if a settlement response curve is used to show the relationship between the applied load intensity and the compression of the foundation soil layer, then the applied load intensity can be used as the first associated feature.

[0038] In addition, other data influence the compression of the foundation soil layer by affecting the first associated feature. In this embodiment, such features indirectly associated with the target feature are referred to as second associated features. Specifically, a response variable that has a causal relationship with the first associated feature is obtained from the domain knowledge base, and this response variable is used as the second associated feature.

[0039] For example, extracting structured knowledge from the domain knowledge base shows that there is a certain causal relationship between horizontal displacement data and applied load intensity. The greater the applied load intensity, the greater the horizontal displacement at a preset measurement point in the foundation soil layer.

[0040] The obtained first and second correlation features are sequentially subjected to outlier removal and continuity correction to obtain a set of core correlation features. In this embodiment, missing values ​​in the data are filled through continuity correction, and regression estimation is mainly performed using relationships with other strongly correlated features.

[0041] Next, each core associated feature is mapped onto a preset target time axis, and feature alignment is performed using time steps on the target time axis. After alignment, a core feature matrix is ​​constructed with time steps as the column coordinates and core associated features as the x-axis.

[0042] The core feature matrix is ​​input into an encoder and encoded for each time step to obtain the initial temporal representation corresponding to each core associated feature. The initial temporal representation for each time step contains compressed information of all features at that time step.

[0043] The correlation score between each initial time series representation and the compression of the foundation soil layer is calculated. This correlation score is used to characterize which core correlation features are most critical for predicting future settlement at which time steps. The correlation scores are then normalized to obtain the attention weights of each core correlation feature.

[0044] Optionally, this application embodiment uses an additive attention mechanism to map the feature information of the initial temporal representation at each time step and the foundation soil compression amount of the corresponding time period to the same hidden space through a learnable linear transformation matrix. The mapped feature information and the foundation soil compression amount are then added together, and the addition result is processed by a non-linear activation function to obtain the corresponding relevance score. The specific calculation formula is as follows: ; In the formula, e t,T The correlation score; v T Let be the compressibility of the subgrade soil layer, and tanh be the nonlinear activation function. W h , W y These are linear transformation matrices, h t The feature information of the initial time series representation at time step t, y T This represents the compression of the foundation soil layer at time step t. b This is a bias term.

[0045] By using the aforementioned attention weights, more attention can be paid to the time step at which the load begins to be applied and the time step at which the pore water pressure begins to dissipate and the effective stress begins to compress the soil layer when predicting the settlement (i.e., compression) of the foundation soil layer, resulting in more accurate settlement prediction results.

[0046] Furthermore, initial temporal representations with attention weights greater than a first threshold are extracted, and these initial temporal representations are weighted and summed to obtain context vectors for each core associated feature. The core associated features corresponding to the context vectors are then concatenated with the context vectors to achieve feature fusion of features strongly correlated with the target feature. The resulting fused temporal feature incorporates the most relevant information for settlement prediction extracted from the entire construction parameters and foundation monitoring data, providing accurate data support for subsequent settlement distribution prediction.

[0047] Step S3: Extract and stitch features from the fused time-series features and geological exploration data. Input the stitched feature results into the preset settlement prediction model. Predict the compression of the foundation soil layer in each region by analyzing the pore drainage rate of the foundation soil layer. Output the settlement prediction results of the membrane bag sand embankment slope.

[0048] To further improve prediction accuracy, this application's embodiments introduce geological exploration data to describe the structure of the foundation soil layers. First, the geological exploration data is encoded as a constant feature for each time step in the target time axis, obtaining a semantic vector representing the dimension of each time step. Then, based on the type of foundation soil layers in each region, data partitioning is used to obtain a semantic vector that accurately describes the structure of each foundation soil layer, thereby obtaining the static geological characteristics of the foundation soil layers in each region.

[0049] The influence of the static geological features on the changing trend of the foundation monitoring data is analyzed, and the response features corresponding to the static geological features are found from the fused time series features.

[0050] For example, the static geological characteristics show that the foundation soil layer A is a highly compressible silt layer, i.e., a low consolidation coefficient and a high compression index. Therefore, foundation soil layer A will experience major settlement and slow consolidation. Under these static geological characteristics, the pore water pressure in the soil layer dissipates slowly; therefore, the pore water pressure is the response characteristic.

[0051] Finally, based on the time step of the target time axis, the static geological features are concatenated with the response features in the fused temporal features. The concatenated feature result is then input into a settlement prediction model based on a long short-term memory network, so that the model can learn how static geological features affect the settlement process.

[0052] In the settlement prediction model, groundwater levels for future periods are predicted based on the porosity of the foundation soil layers and the applied load intensity on the membrane bag sand embankment slope during construction. This yields the pore drainage rate of the foundation soil layers. During foundation settlement, the construction rate can be compared with the pore drainage rate to determine how much of the applied load intensity is converted into effective stress that causes soil settlement. Furthermore, the stability of the foundation can be determined. Therefore, predicting the pore drainage rate provides effective data support for settlement prediction.

[0053] The construction rate from historical construction processes is then extracted from the feature splicing results. This construction rate is used as the construction rate for future time periods and compared with the predicted pore drainage rate to obtain the effective stress distribution of the foundation soil layers in the vertical direction. The construction process is divided into a loading period and an intermittent period. No construction operations are performed during the intermittent period. The comparison process is specific to the loading period, and the construction parameters during the loading period have a certain periodicity. Therefore, the construction rate is compared with the predicted pore drainage rate based on the time step of the loading period. If the loading rate exceeds the pore drainage rate, the newly applied load cannot be converted into effective stress in a timely manner, but is almost entirely converted into extremely high pore water pressure. At this time, the effective stress increases slowly, the instantaneous settlement of the soil increases, and the stability of the foundation is severely reduced, easily inducing landslides and instability. Correspondingly, the loading rate needs to be slowed down.

[0054] Based on the obtained effective stress distribution and the actual thickness of the foundation soil layers in each location, the compression distribution of the foundation soil layers in each location is predicted. Then, the compression distribution of all foundation soil layers is summed to obtain the total compression distribution of the foundation, and thus the settlement prediction result of the membrane bag sand embankment slope is obtained.

[0055] Meanwhile, during the prediction process, the static geological features serve as keywords for describing the foundation soil structure, guiding the model to pay more attention to the time points when pore water pressure disappears and load intensity begins to be applied, thus making the model more focused on the moment when settlement begins to occur.

[0056] As an improvement to the above scheme, this application embodiment also introduces static geological features as constraints in the prediction process to detect whether the value of the response feature in the prediction process exceeds the set theoretical value range, thereby improving the reliability of the prediction results.

[0057] Specifically, the predicted values ​​of the response features for future time periods are obtained, and boundary checks are performed on the predicted values ​​using constraints. If the predicted values ​​exceed the set theoretical range, the boundary check is deemed unqualified, and the predicted values ​​deviate from the theory. Therefore, the predicted values ​​are adjusted to a preset safety range, and the adjusted predicted values ​​are then used to predict the distribution of compressibility in the foundation soil layers in various locations. The safety range is the set value range of the response features.

[0058] Step S4: Based on the settlement prediction results, identify the settlement anomaly risk of the membrane bag sand embankment slope through a preset dynamic early warning mechanism, and output the settlement risk level of the membrane bag sand embankment slope at each preset construction point.

[0059] Based on the actual construction conditions, pre-defined construction points are extracted on the membrane bag sand embankment slope. Based on the obtained settlement prediction results, a corresponding continuous settlement distribution map is drawn within the spatial range of the pre-defined construction points, thus demonstrating the spatial distribution of the predicted settlement of the foundation under the pre-defined construction points. In the continuous settlement distribution map, the corresponding settlement amount is marked on each pre-defined construction point, and then the data from the discrete pre-defined construction points are made continuous in space.

[0060] The obtained continuous settlement distribution map is overlaid with a pre-set predicted settlement cloud map. Pre-set construction points where significant changes in settlement gradient occur are identified in the overlaid map. The predicted settlement cloud map uses a theoretical baseline to represent the pre-set, expected settlement amount of the construction project. In the overlay result, the difference between the predicted settlement and the expected settlement at the pre-set construction point within a fixed horizontal distance is used as the settlement gradient. If the settlement gradient exceeds a set threshold, it indicates concentrated settlement at that pre-set construction point, suggesting the presence of cavities or extremely thick soft soil layers in the membrane bag sand embankment slope at that point, thus identifying the abnormal pattern at that pre-set construction point.

[0061] In addition, by overlaying the results, the symmetry of settlement on both sides of the preset embankment axis was checked. If there is a significant asymmetry in settlement on both sides of the embankment axis at a certain preset construction point, it indicates that there is a risk of slope instability in the foundation soil layer at that preset construction point.

[0062] Finally, the pre-defined construction points with abnormal patterns are classified into corresponding types of structural risk points. The structural hazard level of each type of structural risk point is assessed through a pre-defined dynamic early warning mechanism to obtain the settlement risk level of each pre-defined construction point.

[0063] For example, areas with concentrated settlement are more prone to sudden collapse, thus posing a higher risk to the structure and requiring urgent investigation; areas with significant asymmetry in settlement are prone to cracking of the membrane bag sand embankment, posing a moderate risk to the structure and requiring the design of reinforcement structures.

[0064] Therefore, based on the settlement risk level of the membrane bag sand embankment at each preset construction point, abnormal areas of the membrane bag sand embankment can be monitored more intensively and further investigated to achieve proactive prevention and control of embankment settlement risk, ensure the safety of the entire project life cycle, and improve construction safety and reliability.

[0065] Implementing the embodiments of this application has the following beneficial effects: This application's embodiments consider that under complex geological conditions, it is difficult to accurately predict the settlement of membrane bag sand embankments using only traditional empirical formulas and single models. Therefore, it employs historical construction parameters and foundation monitoring data related to embankment settlement. First, it combines data from different sources through feature fusion to obtain fused time-series features closely related to embankment settlement, effectively improving the accuracy of subsequent predictions. Then, using a settlement prediction model, the load intensity applied to the embankment is obtained through construction parameters, and the pore drainage rate of the foundation soil is calculated using foundation monitoring data. This allows analysis of the effective stress distribution directly related to settlement after pore water stress is eliminated, thereby obtaining accurate compression of the foundation soil layers at each location after pore water drainage. Finally, a dynamic early warning mechanism is used to discriminate the settlement prediction results, progressively identifying settlement anomalies at various locations within the membrane bag sand embankment and assessing the anomaly risk level, providing decision support for the safety of subsequent construction.

[0066] Second Embodiment Furthermore, in order to implement the neural network-based membrane bag sand embankment slope settlement prediction system corresponding to the above method embodiments, and to achieve the corresponding functions and technical effects, Figure 2 A structural diagram of a neural network-based membrane bag sand embankment settlement prediction system is provided. For ease of explanation, only the parts relevant to this embodiment are shown. The neural network-based membrane bag sand embankment settlement prediction system provided in this application embodiment includes: The data acquisition module 201 is used to acquire construction parameters, foundation monitoring data and geological exploration data of the membrane bag sand embankment slope at several preset time points.

[0067] In this embodiment of the application, multiple preset measurement points are set on the foundation soil layers in each location according to the structural form and geological profile results of the foundation soil under the membrane bag sand embankment slope; For the preset measurement points, pore water pressure data and horizontal displacement data are collected at each preset time point to obtain the foundation monitoring data, and the applied load intensity and construction rate corresponding to each preset time point are obtained to obtain the construction parameters. Based on the geological profile results, the thickness, initial void ratio, consolidation coefficient, and compression index of the foundation soil layers in various locations were measured to obtain the geological exploration data.

[0068] The feature fusion module 202 is used to take the compression amount of the foundation soil layer in each region as the target feature, and perform feature fusion by analyzing the correlation between the construction parameters and the foundation monitoring data and the target feature in the time dimension to obtain fused time-series features.

[0069] In this embodiment of the application, based on a preset domain knowledge base and the target features, the construction parameters and the foundation monitoring data are filtered by feature association to obtain core associated features; Each of the core associated features is mapped onto a preset target time axis for feature alignment to obtain a core feature matrix; Based on the attention mechanism, the correlation score between each core associated feature is calculated through the core feature matrix, and the core associated features are fused to obtain the fused temporal features.

[0070] The settlement prediction module 203 extracts and stitches features from the fused time-series features and geological exploration data, inputs the feature stitching results into the preset settlement prediction model, predicts the settlement distribution by analyzing the pore drainage rate of the foundation soil layer in various places, and outputs the settlement prediction results of the membrane bag sand embankment slope.

[0071] In this embodiment of the application, the geological exploration data is encoded into a semantic vector of a preset dimension, and the semantic vector is divided according to the type of foundation soil layer to obtain the static geological characteristics of each foundation soil layer; By analyzing the impact of the static geological features on the changing trend of the foundation monitoring data, the response features corresponding to the static geological features are extracted from the fused time series features. Based on a preset time step, the static geological features and the response features are spliced ​​together to obtain the feature splicing result. The feature splicing results are input into the settlement prediction model based on long short-term memory network. The compression of the foundation soil layer in each region is predicted by analyzing the pore drainage rate of the foundation soil layer in each region, and the settlement prediction results are output. The prediction process is adjusted by using the static geological features as constraints.

[0072] The risk prediction module 204 is used to identify the abnormal settlement risk of the membrane bag sand embankment slope through a preset dynamic early warning mechanism based on the settlement prediction results, and output the settlement risk level of the membrane bag sand embankment slope at each preset construction point.

[0073] In this embodiment, pre-defined construction points are extracted on the membrane bag sand embankment slope based on the actual construction conditions. Based on the obtained settlement prediction results, a corresponding continuous settlement distribution map is drawn within the spatial range of the pre-defined construction points, thereby demonstrating the spatial distribution of the predicted settlement of the foundation under the pre-defined construction points. Specifically, in the continuous settlement distribution map, the corresponding settlement amount is marked on each pre-defined construction point, and then the data from the discrete pre-defined construction points are made continuous in space.

[0074] The obtained continuous settlement distribution map is overlaid with a pre-set predicted settlement cloud map. Pre-set construction points where significant changes in settlement gradient occur are identified in the overlaid map. The predicted settlement cloud map uses a theoretical baseline to represent the pre-set, expected settlement amount of the construction project. In the overlay result, the difference between the predicted settlement and the expected settlement at the pre-set construction point within a fixed horizontal distance is used as the settlement gradient. If the settlement gradient exceeds a set threshold, it indicates concentrated settlement at that pre-set construction point, suggesting the presence of cavities or extremely thick soft soil layers in the membrane bag sand embankment slope at that point, thus identifying the abnormal pattern at that pre-set construction point.

[0075] In addition, by overlaying the results, the symmetry of settlement on both sides of the preset embankment axis was checked. If there is a significant asymmetry in settlement on both sides of the embankment axis at a certain preset construction point, it indicates that there is a risk of slope instability in the foundation soil layer at that preset construction point.

[0076] Finally, the pre-defined construction points with abnormal patterns are classified into corresponding types of structural risk points. The structural hazard level of each type of structural risk point is assessed through a pre-defined dynamic early warning mechanism to obtain the settlement risk level of each pre-defined construction point.

[0077] For example, areas with concentrated settlement are more prone to sudden collapse, thus posing a higher risk to the structure and requiring urgent investigation; areas with significant asymmetry in settlement are prone to cracking of the membrane bag sand embankment, posing a moderate risk to the structure and requiring the design of reinforcement structures.

[0078] Therefore, based on the settlement risk level of the membrane bag sand embankment at each preset construction point, abnormal areas of the membrane bag sand embankment can be monitored more intensively and further investigated to achieve proactive prevention and control of embankment settlement risk, ensure the safety of the entire project life cycle, and improve construction safety and reliability.

[0079] In some embodiments, the data acquisition module 201 specifically comprises: This application primarily uses a combination of three types of data—construction parameters, foundation monitoring data, and geological exploration data—to predict the settlement of membrane bag sand embankment slopes. To obtain this multi-source data, multiple pre-set measurement points are first established on each foundation soil layer based on the structural form and geological profile results of the underlying foundation. The number of these pre-set measurement points in each foundation soil layer is mainly determined by the properties of that soil layer. For some foundation soil layers with strong spatial variability, more pre-set measurement points are set to improve the accuracy of the collected data.

[0080] Then, for these preset measurement points, pore water pressure data and horizontal displacement data of the foundation soil layer are collected at multiple discrete preset time points during construction to obtain foundation monitoring data for each foundation soil layer. The pore water pressure data at different time points can show the increase and dissipation process of excess pore pressure at different depths and locations in the foundation, serving as dynamic data for predicting the pore drainage rate at various points in the foundation. The horizontal displacement data is used to determine the stability of the foundation soil layer, describing the horizontal displacement of the foundation soil layer at each preset measurement point during construction; because extremely high pore water pressure can drive the soil to form significant horizontal displacement, thus affecting the pore drainage rate, it is also one of the indicators used to predict the pore drainage rate.

[0081] Simultaneously, construction parameters corresponding to each preset time point were collected, including the load intensity applied to the foundation and the membrane bag sand embankment slope, as well as the construction rate. This provides direct data for analyzing the effective stress distribution on the foundation. In addition, based on the geological profile results, the thickness, initial void ratio, consolidation coefficient, compression index, and compression modulus of each foundation soil layer were measured to obtain geological exploration data.

[0082] Among them, the consolidation coefficient, compression index, and compression modulus are all physical and mechanical properties of the foundation. The consolidation coefficient is a physical and mechanical property that characterizes the rate at which pore water is expelled and the soil compacts under load in saturated soil. A larger consolidation coefficient indicates that pore water is expelled faster and the foundation is compressed faster. The compression index characterizes the relationship between the void ratio of the foundation soil and the pressure, and can be used to calculate the final compression. The compression modulus is the ratio of vertical stress to the corresponding vertical strain when the soil undergoes vertical compression, and can also be used to calculate the final compression.

[0083] In some embodiments, the feature fusion module 202 specifically comprises: To quickly obtain the influencing factors related to the compression of the foundation soil layer during construction, this application embodiment also establishes a domain knowledge base as a knowledge set specific to this domain. Structured knowledge within the domain knowledge base is used to analyze feature data related to the target characteristics.

[0084] This application extracts correlation features from both direct and indirect correlation perspectives, with the compression amount of the foundation soil layer as the target feature. Specifically, if experimental data showing an independent-dependent variable relationship with the compression amount of the foundation soil layer can be directly obtained from the domain knowledge base, then the corresponding variable can be extracted from this experimental data as the first correlation feature directly correlated with the target feature.

[0085] Specifically, the settlement response curve related to the membrane bag sand embankment slope is extracted from the domain knowledge base. The dependent variable of this curve is the target feature. If the corresponding independent variable exists in the construction parameters or the foundation monitoring data, the independent variable is extracted as the first associated feature.

[0086] For example, if a settlement response curve is used to show the relationship between the applied load intensity and the compression of the foundation soil layer, then the applied load intensity can be used as the first associated feature.

[0087] In addition, other data influence the compression of the foundation soil layer by affecting the first associated feature. In this embodiment, such features indirectly associated with the target feature are referred to as second associated features. Specifically, a response variable that has a causal relationship with the first associated feature is obtained from the domain knowledge base, and this response variable is used as the second associated feature.

[0088] For example, extracting structured knowledge from the domain knowledge base shows that there is a certain causal relationship between horizontal displacement data and applied load intensity. The greater the applied load intensity, the greater the horizontal displacement at a preset measurement point in the foundation soil layer.

[0089] The obtained first and second correlation features are sequentially subjected to outlier removal and continuity correction to obtain a set of core correlation features. In this embodiment, missing values ​​in the data are filled through continuity correction, and regression estimation is mainly performed using relationships with other strongly correlated features.

[0090] Next, each core associated feature is mapped onto a preset target time axis, and feature alignment is performed using time steps on the target time axis. After alignment, a core feature matrix is ​​constructed with time steps as the column coordinates and core associated features as the x-axis.

[0091] The core feature matrix is ​​input into an encoder and encoded for each time step to obtain the initial temporal representation corresponding to each core associated feature. The initial temporal representation for each time step contains compressed information of all features at that time step.

[0092] The correlation score between each initial time series representation and the compression of the foundation soil layer is calculated. This correlation score is used to characterize which core correlation features are most critical for predicting future settlement at which time steps. The correlation scores are then normalized to obtain the attention weights of each core correlation feature.

[0093] Optionally, this application embodiment uses an additive attention mechanism to map the feature information of the initial temporal representation at each time step and the foundation soil compression amount of the corresponding time period to the same hidden space through a learnable linear transformation matrix. The mapped feature information and the foundation soil compression amount are then added together, and the addition result is processed by a non-linear activation function to obtain the corresponding relevance score. The specific calculation formula is as follows: ; In the formula, e t,T The correlation score; v T Let be the compressibility of the subgrade soil layer, and tanh be the nonlinear activation function. W h , W y These are linear transformation matrices, h t The feature information of the initial time series representation at time step t, y T This represents the compression of the foundation soil layer at time step t. b This is a bias term.

[0094] By using the aforementioned attention weights, more attention can be paid to the time step at which the load begins to be applied and the time step at which the pore water pressure begins to dissipate and the effective stress begins to compress the soil layer when predicting the settlement (i.e., compression) of the foundation soil layer, resulting in more accurate settlement prediction results.

[0095] Furthermore, initial temporal representations with attention weights greater than a first threshold are extracted, and these initial temporal representations are weighted and summed to obtain context vectors for each core associated feature. The core associated features corresponding to the context vectors are then concatenated with the context vectors to achieve feature fusion of features strongly correlated with the target feature. The resulting fused temporal feature incorporates the most relevant information for settlement prediction extracted from the entire construction parameters and foundation monitoring data, providing accurate data support for subsequent settlement distribution prediction.

[0096] In some embodiments, the settlement prediction module 203 specifically comprises: To further improve prediction accuracy, this application's embodiments introduce geological exploration data to describe the structure of the foundation soil layers. First, the geological exploration data is encoded as a constant feature for each time step in the target time axis, obtaining a semantic vector representing the dimension of each time step. Then, based on the type of foundation soil layers in each region, data partitioning is used to obtain a semantic vector that accurately describes the structure of each foundation soil layer, thereby obtaining the static geological characteristics of the foundation soil layers in each region.

[0097] The influence of the static geological features on the changing trend of the foundation monitoring data is analyzed, and the response features corresponding to the static geological features are found from the fused time series features.

[0098] For example, the static geological characteristics show that the foundation soil layer A is a highly compressible silt layer, i.e., a low consolidation coefficient and a high compression index. Therefore, foundation soil layer A will experience major settlement and slow consolidation. Under these static geological characteristics, the pore water pressure in the soil layer dissipates slowly; therefore, the pore water pressure is the response characteristic.

[0099] Finally, based on the time step of the target time axis, the static geological features are concatenated with the response features in the fused temporal features. The concatenated feature result is then input into a settlement prediction model based on a long short-term memory network, so that the model can learn how static geological features affect the settlement process.

[0100] In the settlement prediction model, groundwater levels for future periods are predicted based on the porosity of the foundation soil layers and the applied load intensity on the membrane bag sand embankment slope during construction. This yields the pore drainage rate of the foundation soil layers. During foundation settlement, the construction rate can be compared with the pore drainage rate to determine how much of the applied load intensity is converted into effective stress that causes soil settlement. Furthermore, the stability of the foundation can be determined. Therefore, predicting the pore drainage rate provides effective data support for settlement prediction.

[0101] The construction rate from historical construction processes is then extracted from the feature splicing results. This construction rate is used as the construction rate for future time periods and compared with the predicted pore drainage rate to obtain the effective stress distribution of the foundation soil layers in the vertical direction. The construction process is divided into a loading period and an intermittent period. No construction operations are performed during the intermittent period. The comparison process is specific to the loading period, and the construction parameters during the loading period have a certain periodicity. Therefore, the construction rate is compared with the predicted pore drainage rate based on the time step of the loading period. If the loading rate exceeds the pore drainage rate, the newly applied load cannot be converted into effective stress in a timely manner, but is almost entirely converted into extremely high pore water pressure. At this time, the effective stress increases slowly, the instantaneous settlement of the soil increases, and the stability of the foundation is severely reduced, easily inducing landslides and instability. Correspondingly, the loading rate needs to be slowed down.

[0102] Based on the obtained effective stress distribution and the actual thickness of the foundation soil layers in each location, the compression distribution of the foundation soil layers in each location is predicted. Then, the compression distribution of all foundation soil layers is summed to obtain the total compression distribution of the foundation, and thus the settlement prediction result of the membrane bag sand embankment slope is obtained.

[0103] Meanwhile, during the prediction process, the static geological features serve as keywords for describing the foundation soil structure, guiding the model to pay more attention to the time points when pore water pressure disappears and load intensity begins to be applied, thus making the model more focused on the moment when settlement begins to occur.

[0104] As an improvement to the above scheme, this application embodiment also introduces static geological features as constraints in the prediction process to detect whether the value of the response feature in the prediction process exceeds the set theoretical value range, thereby improving the reliability of the prediction results.

[0105] Specifically, the predicted values ​​of the response features for future time periods are obtained, and boundary checks are performed on the predicted values ​​using constraints. If the predicted values ​​exceed the set theoretical range, the boundary check is deemed unqualified, and the predicted values ​​deviate from the theory. Therefore, the predicted values ​​are adjusted to a preset safety range, and the adjusted predicted values ​​are then used to predict the distribution of compressibility in the foundation soil layers in various locations. The safety range is the set value range of the response features.

[0106] Implementing the embodiments of this application has the following beneficial effects: This application's embodiments consider that under complex geological conditions, it is difficult to accurately predict the settlement of membrane bag sand embankments using only traditional empirical formulas and single models. Therefore, it employs historical construction parameters and foundation monitoring data related to embankment settlement. First, it combines data from different sources through feature fusion to obtain fused time-series features closely related to embankment settlement, effectively improving the accuracy of subsequent predictions. Then, using a settlement prediction model, the load intensity applied to the embankment is obtained through construction parameters, and the pore drainage rate of the foundation soil is calculated using foundation monitoring data. This allows analysis of the effective stress distribution directly related to settlement after pore water stress is eliminated, thereby obtaining accurate compression of the foundation soil layers at each location after pore water drainage. Finally, a dynamic early warning mechanism is used to discriminate the settlement prediction results, progressively identifying settlement anomalies at various locations within the membrane bag sand embankment and assessing the anomaly risk level, providing decision support for the safety of subsequent construction.

[0107] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the scope of protection of this application. In particular, it should be noted that any modifications, equivalent substitutions, or improvements made by those skilled in the art within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for predicting settlement of membrane bag sand embankment slopes based on neural networks, characterized in that, include: Acquire construction parameters, foundation monitoring data, and geological exploration data of membrane bag sand embankment slopes at several preset time points; Using the compression amount of the foundation soil layer in each region as the target feature, feature fusion is performed by analyzing the correlation between the construction parameters and the foundation monitoring data and the target feature in the time dimension to obtain fused temporal features; Feature extraction and feature stitching are performed on the fused time series features and geological exploration data. The feature stitching results are input into the preset settlement prediction model. The compression of the foundation soil layer in each region is predicted by analyzing the pore drainage rate of the foundation soil layer in each region. The settlement prediction results of the membrane bag sand embankment slope are output. Based on the settlement prediction results, the abnormal settlement risk of the membrane bag sand embankment is identified through a preset dynamic early warning mechanism, and the settlement risk level of the membrane bag sand embankment at each preset construction point is output.

2. The method for predicting settlement of membrane bag sand embankment slopes based on neural networks according to claim 1, characterized in that, The acquisition of construction parameters, foundation monitoring data, and geological exploration data of the membrane bag sand embankment slope at several preset time points specifically includes: Based on the structural form and geological profile results of the foundation of the membrane bag sand embankment slope, multiple preset measurement points were set on the foundation soil layers in each location; For the preset measurement points, pore water pressure data and horizontal displacement data are collected at each preset time point to obtain the foundation monitoring data, and the applied load intensity and construction rate corresponding to each preset time point are obtained to obtain the construction parameters. Based on the geological profile results, the thickness, initial void ratio, consolidation coefficient, and compression index of the foundation soil layers in various locations were measured to obtain the geological exploration data.

3. The method for predicting settlement of membrane bag sand embankment slopes based on neural networks according to claim 1, characterized in that, The method uses the compressibility of the foundation soil layers in each region as the target feature. It then performs feature fusion by analyzing the correlation between the construction parameters and the foundation monitoring data and the target feature over time, resulting in a fused temporal feature. Specifically: Based on a preset domain knowledge base and the target features, the construction parameters and the foundation monitoring data are filtered using a feature association method to obtain core associated features; Each of the core associated features is mapped onto a preset target time axis for feature alignment to obtain a core feature matrix; Based on the attention mechanism, the correlation score between each core associated feature is calculated through the core feature matrix, and the core associated features are fused to obtain the fused temporal features.

4. The method for predicting settlement of membrane bag sand embankment slopes based on neural networks according to claim 3, characterized in that, Based on a preset domain knowledge base and the target features, the construction parameters and foundation monitoring data are filtered using a feature association method to obtain core associated features, specifically: The settlement response curve related to the membrane bag sand embankment slope is extracted from the domain knowledge base. Based on the variables of the settlement response curve, the first associated feature that is directly related to the target feature is screened from the construction parameters and the foundation monitoring data. Extract causal relationships with the first associated feature as a variable from the domain knowledge base, and based on the causal relationships, filter out second associated features that are indirectly associated with the target feature from the construction parameters and the foundation monitoring data; Outlier removal and continuity correction are performed on the first and second related features to obtain the core related features.

5. The method for predicting settlement of membrane bag sand embankment slopes based on neural networks according to claim 3, characterized in that, The attention-based mechanism calculates the correlation scores between core related features using a core feature matrix, and then performs feature fusion on these core related features to obtain the fused temporal features. Specifically: Based on each time step of the target time axis, the core feature matrix is ​​encoded by an encoder to obtain the initial temporal representation corresponding to each core associated feature; Calculate the correlation score between each of the initial temporal representations and the target feature, normalize the correlation score, and obtain the attention weight of each of the core associated features; If the attention weight is greater than the first threshold, the corresponding initial temporal representation is weighted and summed to perform feature fusion on the core associated features, and the fused temporal features are output.

6. The method for predicting settlement of membrane bag sand embankment slopes based on neural networks according to claim 1, characterized in that, The process involves extracting and stitching features from the fused temporal features and geological exploration data. The stitched feature results are then input into a pre-defined settlement prediction model. By analyzing the pore drainage rate of the foundation soil layers in different locations, the compression of the foundation soil layers in each location is predicted, and the settlement prediction results for the membrane bag sand embankment slope are output. Specifically: The geological exploration data is encoded into semantic vectors of a preset dimension, and the semantic vectors are divided according to the type of foundation soil layer to obtain the static geological characteristics of each foundation soil layer. By analyzing the impact of the static geological features on the changing trend of the foundation monitoring data, the response features corresponding to the static geological features are extracted from the fused time series features. Based on a preset time step, the static geological features and the response features are spliced ​​together to obtain the feature splicing result. The feature splicing results are input into the settlement prediction model based on long short-term memory network. The compression of the foundation soil layer in each region is predicted by analyzing the pore drainage rate of the foundation soil layer in each region, and the settlement prediction results are output. The prediction process is adjusted by using the static geological features as constraints.

7. The method for predicting settlement of membrane bag sand embankment slopes based on neural networks according to claim 6, characterized in that, The method of predicting the compression of foundation soil layers in various locations by analyzing the pore drainage rate of the foundation soil layers and outputting the settlement prediction results is as follows: Based on the feature splicing results, according to the porosity of the foundation soil layer in various places and the applied load intensity on the membrane bag sand embankment slope, the groundwater level data in the future time period is predicted to obtain the predicted pore drainage rate. The construction rate is extracted from the feature splicing results, and the effective stress distribution of the foundation soil layer in the vertical direction is obtained by comparing the construction rate with the predicted pore drainage rate. Based on the effective stress distribution and the actual thickness of the foundation soil layer in each location, the compression distribution of the foundation soil layer in each location is predicted, and the settlement prediction results are output.

8. The method for predicting settlement of membrane bag sand embankment slopes based on neural networks according to claim 6, characterized in that, The adjustment of the prediction process based on the aforementioned static geological features as constraints is specifically as follows: During the prediction process, the static geological features are used to perform boundary detection on the predicted values ​​of the response features within a future time period; If the boundary detection fails, the predicted value is adjusted to a preset safety range, and the adjusted predicted value is used to predict the compression distribution of the foundation soil layer in various locations. The safety range is the set value range of the response feature.

9. The method for predicting settlement of membrane bag sand embankment slopes based on neural networks according to claim 1, characterized in that, Based on the settlement prediction results, a preset dynamic early warning mechanism is used to identify the abnormal settlement risk of the membrane bag sand embankment slope, and the settlement risk level of the membrane bag sand embankment slope at each preset construction point is output, specifically: Based on the settlement prediction results, draw a spatial continuous settlement distribution map of the predicted settlement at each preset construction point; The preset predicted settlement cloud map is superimposed on the settlement continuous distribution map. Based on the superposition result, the first preset construction point and its abnormal pattern that the settlement gradient exceeds the first threshold are identified, as well as the second preset construction point and its abnormal pattern that there is asymmetry in settlement on both sides of the dike axis; wherein, the settlement gradient is the difference in settlement within a fixed horizontal distance. Based on the dynamic early warning mechanism and the abnormal mode, the degree of structural hazard at the first preset construction point and the second preset construction point are assessed respectively to obtain the corresponding settlement risk level.

10. A neural network-based system for predicting settlement of membrane bag sand embankment slopes, characterized in that, include: Data acquisition module, feature fusion module, settlement prediction module, and risk prediction module; The data acquisition module is used to acquire construction parameters, foundation monitoring data and geological exploration data of the membrane bag sand embankment slope at several preset time points; The feature fusion module is used to take the compression amount of the foundation soil layer in each region as the target feature, and perform feature fusion by analyzing the correlation between the construction parameters and the foundation monitoring data and the target feature in the time dimension to obtain fused time-series features; The settlement prediction module is used to extract and stitch features from fused time-series features and geological exploration data. The feature stitching results are input into a preset settlement prediction model. By analyzing the pore drainage rate of the foundation soil in various places, the compression of the foundation soil in various places is predicted, and the settlement prediction results of the membrane bag sand embankment slope are output. The risk prediction module is used to identify the abnormal settlement risk of the membrane bag sand embankment slope through a preset dynamic early warning mechanism based on the settlement prediction results, and output the settlement risk level of the membrane bag sand embankment slope at each preset construction point.

Images