Time series lstm-based dynamic safety warning method and system for bedding slope with soft interlayer

Abstract

The present application relates to the technical field of geotechnical slope safety, and particularly relates to a soft interlayer-containing bedding slope displacement dynamic safety early warning method and system based on time sequence LSTM. By constructing the step-by-step excavation process of the slope as a time sequence and introducing a long short-term memory neural network model capable of effectively learning long-term dependencies, the accurate simulation of the displacement dynamic cumulative response of the soft interlayer-containing bedding slope during the excavation process is realized. The method effectively overcomes the shortcomings of the existing technology, such as the inability of the limit equilibrium method to reflect time sequence effects, the low calculation efficiency and difficulty in real-time application of the numerical simulation method, and the inability of the static machine learning model to capture the step-by-step excavation path dependence characteristics. The advantage lies in short calculation time, fast iteration speed, and the ability to realize real-time and dynamic prediction of the displacement at each step during the excavation process, thereby providing timely and reliable technical support for slope safety management and risk early warning decision-making during engineering construction, greatly improving the timeliness and accuracy of early warning.

Description

Technical Field

[0001] This invention relates to the field of rock and soil slope safety technology, and in particular to a dynamic safety early warning method and system for displacement of bedding slopes containing weak interlayers based on time-series LSTM. Background Technology

[0002] Blended slopes with weak interlayers are a common type of slope in geotechnical engineering. Their structure typically consists of a relatively hard overlying rock layer and a weak underlying interlayer. During engineering construction (such as highways, railways, and mining excavations), the exposure or disturbance of the weak interlayer during excavation can easily induce slippage and deformation along this interlayer, creating an unfavorable mechanical pattern of a hard upper layer and a soft lower layer. It is important to note that such slopes do not necessarily develop into overall instability after excavation-induced deformation. When the cumulative displacement does not exceed the instability threshold, the slope may reach a new equilibrium state through stress adjustment. However, once the displacement exceeds the threshold, it can rapidly evolve into a catastrophic landslide, seriously threatening engineering safety and the safety of people and property.

[0003] Therefore, accurate and timely prediction of the displacement of bedding slopes containing weak interlayers during excavation is crucial for safety assessment and risk management during engineering construction. Currently, slope displacement prediction mainly relies on the following methods:

[0004] Limit equilibrium analysis method: This method calculates the slope safety factor based on static equilibrium conditions. Although it can make an overall assessment of stability, it cannot reflect the step-by-step and dynamic characteristics of the excavation process and is difficult to characterize the cumulative effect of displacement with the excavation sequence. Therefore, it has inherent limitations in predicting the dynamic displacement response induced by excavation.

[0005] Numerical simulation methods (such as the finite element method and the finite difference method) can simulate the excavation process and mechanical response by establishing a numerical model of the slope, and can reflect the displacement development process to a certain extent. However, this method is usually computationally time-consuming, has complex modeling, and is heavily dependent on the accuracy of the constitutive model and parameter selection. In actual engineering, due to limitations in computational efficiency and model simplification, it is difficult to achieve rapid and real-time prediction of displacement during excavation, and it is not suitable for construction scenarios that require frequent dynamic safety assessments.

[0006] Empirical formulas and case statistics: Empirical formulas based on historical case data have the advantages of being simple and fast, but they often fail to fully incorporate the complex nonlinear interactions and time-series dependencies between multiple factors such as excavation parameters and geotechnical parameters, resulting in limited extrapolation and prediction accuracy. In particular, they are not applicable to the special geological and excavation conditions of specific projects.

[0007] In recent years, machine learning technology has provided new insights for slope displacement prediction. Some studies have attempted to apply models such as support vector machines and neural networks to slope stability or displacement prediction. However, most existing machine learning-based prediction methods treat the excavation process as a static or holistic event, with input features typically being the final state parameters after excavation or a static snapshot at a certain time step. This fails to effectively consider the inherent temporal dynamics of the "step-by-step excavation" construction activity. For bedding slopes with weak interlayers, excavation is the core disaster-inducing factor. Each excavation step disturbs the slope's stress state, and displacement is the cumulative result of the mechanical effects of all historical excavation stages, exhibiting strong path dependence. Existing static prediction models cannot capture this crucial temporal feature, often leading to decreased accuracy, false alarms, or missed alarms when predicting the cumulative displacement in the later stages of excavation.

[0008] Long Short-Term Memory (LSTM) neural networks, as an improved model of recurrent neural networks, have demonstrated significant advantages in the field of time series forecasting due to their unique gating mechanism, which enables them to effectively learn long-term dependencies in time series data. They have also seen initial applications in geological hazard research, such as landslide prediction. This provides theoretical feasibility for constructing models capable of learning excavation time-series dynamics and accurately predicting the cumulative process of slope displacement.

[0009] Based on this, the present invention aims to address the aforementioned deficiencies in the existing technology by proposing a method and system that can fully consider the dynamic process of slope excavation and achieve high-precision dynamic displacement safety early warning. The technical problem to be solved by this invention is: how to provide an efficient and accurate prediction method to overcome the shortcomings of existing limit equilibrium methods in reflecting temporal effects, numerical simulation methods inefficient and difficult to apply in real time, and existing static machine learning models incapable of capturing the path-dependent characteristics of step-by-step excavation, thereby achieving dynamic prediction and real-time safety early warning of displacement response of bedding slopes containing weak interlayers during step-by-step excavation. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a dynamic safety early warning method and system for displacement of slopes with weak interlayers based on time-series LSTM.

[0011] In a first aspect, the present invention provides a dynamic safety early warning method for displacement of bedding slopes containing weak interlayers based on time-series LSTM, comprising the following steps: S1. Determine the model characteristic parameters and their value range: The model characteristic parameters include dynamic characteristic parameters that change during the excavation process and static soil and rock mass parameters that remain unchanged during the excavation process; S2. Within the range of the dynamic characteristic parameters, step-by-step excavation calculations are performed through finite element analysis simulation, and the corresponding cumulative displacement response is monitored to form N sets of time-series "characteristic-displacement" sample data, which constitute a time-series "characteristic-displacement" sample database. S3. Construct a long short-term memory neural network model, and obtain sample data from the sample database to train the long short-term memory neural network model to obtain a trained LSTM neural network model. S4. Input the actual excavation dynamic characteristic parameters of the target slope into the trained LSTM neural network model to predict the cumulative displacement of each excavation step; calculate the displacement warning threshold based on the static soil and rock parameters of the target slope, and compare the predicted displacement with the threshold to achieve dynamic safety warning.

[0012] Preferably, in step S1, the dynamic characteristic parameters include excavation geometric characteristic parameters, and the static soil and rock mass parameters include sliding surface mechanical strength parameters.

[0013] In this scheme, the dynamic characteristic parameters include: the excavation volume V corresponding to each step of excavation, the excavation length L along the slope direction, the excavation thickness D, and the excavation slope α; In this scheme, the static soil and rock parameters include: slope weight γ, sliding surface elastic modulus E, sliding surface Poisson's ratio μ, sliding surface cohesion c, and sliding surface internal friction angle φ.

[0014] Preferably, in step S2, the N sets of time-series "feature-displacement" sample data are generated using the Latin hypercube sampling method.

[0015] The specific process of generating sample data includes: S21. Normalize each feature parameter and divide it into N equally spaced intervals; S22. Randomly select a sample point from each interval, and use the inverse cumulative distribution function to map the sample point back to the actual parameter range. Then, randomly pair the dynamic parameters with the static parameters to form N sets of sample combinations.

[0016] Preferably, in step S2, the slope location monitored by the finite element analysis simulation process includes at least the top, middle, and toe of the slope.

[0017] Preferably, in step S3, the unit structure of the LSTM neural network model includes a forget gate, an input gate, and an output gate; the forget gate is used to learn and judge the degree of attenuation of the influence of historical excavation patterns on the current displacement; the input gate is used to identify the immediate and potential influence of the feature parameters of the current excavation step on the slope deformation; and the output gate is used to map the cumulative mechanical effects represented inside the model to the predicted cumulative displacement.

[0018] Preferably, in step S4, the displacement warning threshold is determined based on a reference threshold and an adjustment function that reflects the influence of the mechanical strength parameters of the sliding surface, wherein the reference threshold is obtained based on finite element simulation.

[0019] More preferably, the adjustment function is the product of the reference threshold and the adjustment coefficient, wherein the adjustment coefficient is determined based on the sliding surface mechanical strength parameter.

[0020] In this scheme, the displacement early warning threshold calculation model is as follows: ; in, The final displacement warning threshold, It is the benchmark threshold for bedding slopes with weak interlayers. It is 1 / 10 of the maximum displacement of the slope after excavation under natural conditions, as determined by finite element simulation. f(c, φ) is a regression function based on the cohesion c of the sliding surface and the internal friction angle φ of the sliding surface, which is used to adjust the benchmark threshold according to the actual state of the rock and soil.

[0021] Preferably, the safety warning is divided into multiple levels according to the ratio between the predicted displacement and the displacement warning threshold: safety level, attention level, warning level and danger level.

[0022] Preferably, the safety warning is specifically divided into four levels: In this scheme, the displacement is predicted. Y <0.8 At the current security level; 0.8 ≤ Y <0.9 Attention level; 0.9 ≤ Y <1.0 The current alert level is [level 1]. Y ≥1.0 The time is classified as dangerous.

[0023] Secondly, this invention discloses a dynamic safety early warning system for excavation-induced displacement of bedding slopes containing weak interlayers based on time-series LSTM, comprising: The parameter determination module is used to determine the dynamic characteristic parameters and static soil and rock parameters of the target slope and their value ranges. The sample database stores time-series feature-displacement sample data; The LSTM prediction model, or Long Short-Term Memory neural network model, is used to receive the input feature parameter sequence and output the predicted cumulative displacement sequence. The threshold calculation module is used to calculate the displacement warning threshold. The early warning judgment module is used to compare the predicted displacement output by the LSTM prediction model with the early warning threshold calculated by the threshold calculation module, and output safety early warning information according to the hierarchical rules.

[0024] Preferably, the parameter determination module, the sample database, the LSTM prediction model, the threshold calculation module, and the early warning judgment module are connected in sequence. The output of the parameter determination module is connected to the sample database and the LSTM prediction model, respectively. The output of the sample database is connected to the LSTM prediction model, and the outputs of the LSTM prediction model and the threshold calculation module are connected to the early warning judgment module, respectively.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing the step-by-step excavation process of a slope as a time series and introducing a Long Short-Term Memory (LSTM) neural network model capable of effectively learning long-term dependencies, accurate simulation of the dynamic cumulative response of displacement in bedding slopes with weak interlayers during excavation is achieved. This method effectively overcomes the shortcomings of existing technologies, such as the inability of the limit equilibrium method to reflect time-series effects, the low computational efficiency and difficulty in real-time application of numerical simulation methods, and the inability of static machine learning models to capture the path-dependent characteristics of step-by-step excavation. By effectively utilizing the structural advantages of the LSTM neural network, a model for the dynamic prediction of displacement in excavation-induced bedding slopes with weak interlayers is constructed, providing support for the dynamic realization of safety early warning.

[0026] 2. The advantage of the technical method of the present invention is that it does not require a large number of case samples as basic data, but rather it is a dynamic safety early warning system for a typical project. The technical method has the value of promotion and application, that is, any slope with weak interlayers can be dynamically warned according to this idea. Attached Figure Description

[0027] Figure 1 This is a flowchart of the displacement dynamic safety early warning method of the present invention; Figure 2 This is an information flow diagram of the long short-term memory neural network model of the present invention; Figure 3 This is a unit structure diagram of the long short-term memory neural network of the present invention; Figure 4 This is a finite element model of a bedding slope with weak interlayers and a layout diagram of displacement monitoring points according to the present invention; Figure 5 This is a schematic diagram illustrating the determination of the safety warning benchmark threshold for slope displacement containing weak interlayers according to the present invention.

[0028] Marked in the image: 1-Slope rock mass, 2-Weak interlayer, 3-Slope top monitoring point, 4-Slope middle monitoring point, 5-Slope toe monitoring point, 6-LSTM element. Detailed Implementation

[0029] The present invention will now be described in further detail with reference to specific embodiments. However, this should not be construed as limiting the scope of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0030] Example 1 This embodiment provides a dynamic safety early warning method for excavation-induced displacement of bedding slopes containing weak interlayers based on time-series LSTM, specifically combined with... Figure 1 As shown, it includes the following steps: S1. Determine the model characteristic parameters and their value range: The model characteristic parameters include dynamic characteristic parameters that change during the excavation process and static soil and rock mass parameters that remain unchanged during the excavation process; For a bedding slope with weak interlayers that is excavated in T-steps, each excavation stage includes 4 dynamic characteristic parameters that change with the excavation sequence and 5 static soil and rock parameters that remain unchanged throughout the excavation process, and sets the value range for each parameter. The dynamic characteristic parameters include excavation geometric characteristic parameters. Slope excavation is a three-dimensional disturbance of a natural slope. The slope gradient has a significant impact on the stability of the free face. The excavation disturbance can be expressed by four characteristic parameters, namely, excavation volume. V (Unit: m) 3 ), Excavation length along the slope direction L (Unit: m) Excavation thickness D (Unit: m) and excavation slope α (Unit: °).

[0031] In each excavation stage t ( t =1, 2, ..., T), define the above four parameters that change with the excavation process, and constitute t Dynamic vector of time step = ( V , L , D , α ), as a temporal dynamic feature.

[0032] Weak interlayers are typical sliding surfaces of bedding slopes. Excavation-induced displacement of bedding slopes is influenced by the physical and mechanical parameters of the soil and rock mass at the sliding surface. In engineering practice, the Mohr-Coulomb strength criterion is often used to characterize this displacement. These static soil and rock parameters include the sliding surface mechanical strength parameters: slope mass strength. γ (Unit: kN / m) 3 ), elastic modulus of the smooth surface E (Unit: MPa), Poisson's ratio of smooth surface μ (Dimensionless) Smooth surface cohesionc (Unit: kPa) and internal friction angle of the sliding surface φ (Unit: °). Five geotechnical parameters, as inherent properties of the soil and rock mass, remain unchanged throughout the excavation process, forming a static vector. =( γ , E , μ , c , φ ), where T represents the Tth step of excavation.

[0033] The value range of each parameter is determined as follows: The slope excavation process is divided into time-series steps, with different excavation parameters corresponding to different steps. Among the four excavation characteristic parameters, the excavation slope is the most important. α It is independent, and the excavation volume is... V Excavation length L Excavation thickness D It is a cumulative value.

[0034] Taking a slope excavated in two steps as an example, the second step of excavation corresponds to... V , L , D It is the sum of the excavation parameters from the first step and the increment caused by the second step of excavation, i.e.: V 2 =V 1 +ΔV, L 2 =L 1 +ΔL, D 2 =D 1 +ΔD.

[0035] By clearly defining four dynamic excavation characteristic parameters and five static geotechnical parameters, a comprehensive feature system characterizing slope excavation disturbance and its own geomechanical properties was constructed. This system can not only finely describe the dynamic parameters of each excavation step, but also stably reflect the static parameters of the slope slip surface, ensuring that the feature information of the input model has both temporal dynamics and geological representativeness, laying a solid foundation for the LSTM model to make high-precision and physically meaningful predictions.

[0036] Therefore, for a certain bedding slope containing weak interlayers, its excavation is divided into... T If the step is correct, then the excavation parameters for the slope are 4. T One, based on excavation volume V For example, including: V 1, V 2, ..., V T The physical and mechanical parameters of the soil and rock mass are five. Therefore, in this embodiment, the number of characteristic parameters of the bedding slope containing weak interlayers is four. T +5, as detailed below: (1) 4 T The baseline values ​​for the five parameters were determined through field investigations and indoor geotechnical tests. The distribution range of the characteristic parameters was set with a fluctuation range of 25% of the baseline value. x i-min , x i-max ), here x i-min For the first i The lower limit value of each parameter, x i-max For the first i The upper limit of each parameter.

[0037] For example, the excavation volume of the first excavation step. V If the baseline value is 4000 m³, then its distribution range is: V 1∈[3000,5000), that is V 1-min =3000 m³, V 1-max =5000 m³.

[0038] S2. Within the range of the dynamic characteristic parameters, step-by-step excavation calculations are performed through finite element analysis simulation, and the corresponding cumulative displacement response is monitored to form N sets of time-series "characteristic-displacement" sample data, which constitute a time-series "characteristic-displacement" sample database. Specifically, for each set of sample data, step-by-step excavation simulation is carried out through finite element analysis, and the cumulative displacement of the slope at at least three locations relative to the pre-excavation position is monitored to form N sets of time-series "feature-displacement" sample data. In this embodiment, the Latin hypercube sampling method is used to generate the N sets of time-series "feature-displacement" sample data. The specific process of generating the sample data includes: S21. Normalize each feature parameter and divide it into N equally spaced intervals; S22. Randomly select a sample point from each interval, and use the inverse cumulative distribution function to map the sample point back to the actual parameter range. Then, randomly pair the dynamic parameters with the static parameters to form N sets of sample combinations.

[0039] Specifically, in this embodiment, the Latin hypercube sampling method is used to generate 4T+5 different sets of feature parameters, each set containing... and These are the excavation parameters and geotechnical parameters, as detailed below: S21. Normalization of parameter distribution According to S1 T The dynamic vector of each excavation step For each parameter, the Z-Score method is used for normalization, mapping it to the [0, 1] interval, and then dividing it into... N A number of equally spaced intervals.

[0040] The normalization method is as follows: (2) In the formula, x i For the value of the i-th parameter, x i-max The upper limit value of the i-th parameter, x i-min Let be the lower limit value of the i-th parameter. x i-new It is the normalized value of the i-th parameter, where i = 1, 2, ..., 4T+5.

[0041] S22, Parameter Distribution Interval Division For each parameter x i Let u be a sample point randomly selected from each interval. j Let u be a random sample point in the j-th interval. j It follows a uniform distribution, that is: (3) In the formula, j = 1, 2, ... N N represents the number of intervals.

[0042] S23, Sample Parameter Reverse Mapping The uniformly distributed samples are mapped back to the actual parameter range using the inverse cumulative distribution function transformation method, as follows: (4) In the formula, u ij Let x be the sample point of the j-th interval of the i-th parameter. ij This is the actual parameter value corresponding to this value.

[0043] S24. Feature Parameter Sample Generation The dynamic parameters are randomly paired to form N sets of sample combinations as shown in equation (1).

[0044] S25. Generation of "Feature-Displacement" Sample Data The displacement distribution of a bedding slope containing weak interlayers induced by excavation was calculated using finite element analysis. Specifically, a finite element model was established, and the displacement distribution was calculated according to... and Excavation and geotechnical parameters were set, and a step-by-step excavation simulation calculation was performed on a bedding slope containing weak interlayers. The slope locations monitored during the finite element analysis simulation included the slope top, middle, and toe. The cumulative displacement of these three locations relative to the pre-excavation state was monitored during the process. =( Y a , Y b , Y c ), where Y a Y represents the displacement at the top of the slope. b Y represents the displacement at the mid-slope. c This indicates the displacement at the toe of the slope.

[0045] This is thus updated and formed. N There are N "feature-displacement" sample data (N≥200), and each sample data can be expressed as in equation (5): (5) S3. Construct a Long Short-Term Memory (LSTM) neural network model, and obtain sample data from the sample database to train the LSTM neural network model to obtain a trained LSTM neural network model. Specifically, the unit structure of the LSTM neural network model includes a forget gate, an input gate, and an output gate. The forget gate is used to learn and judge the degree of attenuation of the influence of historical excavation patterns on the current displacement. The input gate is used to identify the immediate and potential influence of the feature parameters of the current excavation step on the slope deformation. The output gate is used to map the cumulative mechanical effects represented by the model to the predicted cumulative displacement.

[0046] The model's input at any excavation time step t consists of the dynamic feature parameter vector of the current time step, the static soil and rock parameter vector, and the short-term memory h from the previous time step t-1. t-1 and long-term memory C from the previous time step t-1 t-1 The model outputs the predicted cumulative displacement vector for the current time step, and the short-term memory h passed to the next time step t+1. t and the long-term memory C of the next time step t+1 t ; The information flow of the LSTM neural network constructed in this invention is as follows: Figure 2 As shown, the unit structure is as follows Figure 3 As shown.

[0047] Figure 2 This demonstrates the information flow relationship of the LSTM model constructed in this embodiment during the time series prediction process. At any excavation time step t, the model input consists of two parts: Part 1: The dynamic vector of the current step =( V ,L , D , α ); Part Two: Static Vectors Describing the Inherent Properties of Slopes =( γ , E , μ , c , φ ).

[0048] At the same time, the model receives short-term memory h passed from the previous time step t-1. t-1 and long-term memory C t-1 .

[0049] The model's output also consists of two parts: 1) The predicted cumulative slope displacement vector Y corresponding to the current time step t t (including the top of the slope Y) at 、Slope Y bt , slope foot Y ct Displacement of Y; at Y represents the slope crest displacement at the current time step t. bt Y represents the mid-slope displacement at the current time step t. ct This represents the slope displacement at the current time step t.

[0050] 2) Update and pass the short-term memory h to the next time step t+1. t and the long-term memory C of the next time step t+1 t .

[0051] Figure 2 In the middle, X t-1 X represents the feature information of the previous time step t-1. t X represents the feature information of the current time step t. t+1 This represents the feature information for the next time step t+1. This figure intuitively illustrates the core mechanism of the LSTM model in handling temporal dependencies: the model not only makes predictions based on the current input, but also continuously learns and integrates all historical information (i.e., the excavation path) from the start of excavation to the current time step through the chained transmission of memory units C and hidden states H, thereby simulating the cumulative effect of displacement.

[0052] Figure 3 The internal structure and working principle of a single LSTM unit 6 are revealed in detail. Figure 2 The core of information processing. As shown in the figure, an LSTM unit mainly contains three gating structures and one cell state (long-term memory) transfer path: Figure 3In the process, short-term memory from the previous stage and current feature information are input into LSTM unit 6. After passing through a forgetting gate to filter out unimportant information, important information is further processed and divided into current long-term memory with long-term impact and current short-term memory that may have long-term impact, and then passed to the next stage. This figure illustrates how the LSTM model, through a sophisticated gating mechanism, selectively remembers, forgets, and outputs information, thereby possessing the ability to capture long-term dependencies during the data mining process.

[0053] At any time step t, the input to the LSTM model is a dynamic vector of the mining feature parameters at the current time step. static vector of soil and rock characteristic parameters And short-term memory h from the previous stage t-1 time step t-1 and long-term memory C t-1 The output of the LSTM model is the predicted cumulative displacement at the current time step. And the short-term memory h passed to the next stage t+1 time step t and long-term memory C t .

[0054] Specifically, in step 3, the core of LSTM unit 6 consists of three gating mechanisms, specifically the forget gate F. t Input gate I t and output gate O t Among them, the forgetting gate F t The model learns to determine the degree to which historical excavation patterns have diminished influencing the current displacement. For example, if the current excavation volume or thickness changes drastically, the model can use F... t Reducing reliance on memories from early excavation steps is beneficial for capturing the controlling changes in bedding slope displacements with weak interlayers caused by key excavation steps. t The functional relationship can be expressed as shown in equation (6): (6) In the formula, σ The activation function outputs values ​​between 0 and 1, where 1 represents "completely retained" and 0 represents "completely forgotten".

[0055] W f The weight matrix of the forget gate, b f The bias term h represents the forget gate. t-1 Represents the short-term memory at time step t-1. t represents the dynamic vector at the current time step t.

[0056] The input gate It is responsible for identifying the characteristic parameters of the current excavation step and their immediate and potential impact on slope deformation. For example, if the excavation length L and thickness D of a certain excavation step are relatively large, they will be strongly recorded in long-term memory C through It because they have a significant and lasting contribution to slope displacement. t Furthermore, the static vector of soil and rock characteristic parameters As an inherent property of bedding slopes containing weak interlayers, its influence is continuously and selectively incorporated into memory through the input gate. The functional relationship of It can be expressed as equation (7): (7) (8) In the formula, W i Let b represent the weight matrix of the input gate. i The input gate bias term; h t-1 Represents the short-term memory at time step t-1. t represents the dynamic vector at the current time step t.

[0057] Cell state updates can combine Ft and It information to update long-term memory. The code encodes all historical excavation actions on the slope from the start of excavation to the current time step t, under specific soil and rock conditions. The cumulative mechanical effect on slopes is not simply an accumulation of displacement, but rather a characterization of the internal state of slope damage or deformation trends.

[0058] The output gate Ot controls how the accumulated mechanical effects inside the slope are processed. Mapped to observable, specific, and unique predicted values. =( Y at , Y bt , Y ct Its value is implemented through a fully connected layer, i.e. The output gate ensures that the model outputs a reasonable set of cumulative displacements at each step, whose values ​​incorporate all information from the excavation history and the current excavation step.

[0059] In the formula, W y Let b represent the weight matrix of the output gate. y C represents the bias term of the output gate. t-1 This represents the long-term memory at time step t-1.

[0060] S4. Input the actual excavation dynamic characteristic parameters of the target slope into the trained LSTM neural network model to predict the cumulative displacement of each excavation step; calculate the displacement warning threshold based on the static soil and rock parameters of the target slope, and compare the predicted displacement with the threshold to achieve dynamic safety warning.

[0061] LSTM neural network model training and displacement dynamic threshold determination: The sample data is divided into training set, validation set and test set, and the LSTM model is trained and validated; a displacement early warning threshold calculation model combining the cohesion c of the weak interlayer sliding surface and the internal friction angle φ of the sliding surface is established; Figure 4 A typical three-dimensional finite element numerical model of a bedding slope with weak interlayers, built for generating training samples, is presented. The figure clearly shows the geometry and stratigraphic layering of the slope rock mass 1, and in particular, marks the location of the key weak interlayer 2. The step-by-step excavation process is illustrated by the mesh-divided variation regions. Displacement monitoring points ( Y a , Y b , Y c Three displacement monitoring points were set up at key locations on the slope surface: monitoring point 3 at the top of the slope, monitoring point 4 in the middle of the slope, and monitoring point 5 at the toe of the slope. The cumulative displacement of these three monitoring points (relative to the initial state before excavation) constitutes the displacement response vector Y = ( Y a , Y b , Y c Selecting these three points can effectively capture the deformation characteristics of the slope as a whole and its key parts.

[0062] This figure illustrates the numerical experimental method used to generate high-fidelity "feature-displacement" correspondence sample data, ensuring the reliability and representativeness of the subsequent LSTM model training data.

[0063] S41, Model Training The structure built in S2 N The time-series samples are divided into training, validation, and test sets in a ratio of 50%, 30%, and 20%, respectively. Mean squared error is used as the loss function, and the LSTM network parameters (W, b) are optimized through backpropagation and temporal backpropagation algorithms until the model's prediction accuracy on the validation set meets the requirements.

[0064] S42. Determination of dynamic displacement threshold Slopes containing weak interlayers often experience slip deformation along the weak interlayers. The control of slip along these weak interlayers primarily depends on the surface cohesion (c) and the internal friction angle (φ). During slope creep deformation, both surface cohesion (c) and internal friction angle (φ) will drop from their peak values ​​to their residual states due to shear damage. Therefore, the surface cohesion and internal friction angle of the weak interlayer are particularly important for the early warning threshold of bedding slopes. Furthermore, after rainfall infiltration, the weak interlayer undergoes a significant saturation softening effect, and the surface cohesion and internal friction angle decrease exponentially with increasing water content.

[0065] Therefore, this embodiment establishes a displacement warning threshold that takes into account the state of weak interlayers, as specifically shown in equation (9): (9) In the formula, This is the baseline threshold for a bedding slope containing weak interlayers. Its value corresponds to 1 / 10 of the maximum displacement of the bedding slope containing weak interlayers after excavation under natural conditions. It can be obtained through finite element simulation calculations, such as... Figure 5 . yes c and φ The function, where, c Indicates the cohesion of the slip surface. φ The internal friction angle of the slip surface can be determined through geotechnical tests on weak interlayers combined with regression analysis, and is used to adjust the baseline threshold. When c and φ At lower levels, If the threshold is less than 1, lower the warning threshold; otherwise, raise it.

[0066] Figure 5 The baseline threshold is illustrated by a typical cumulative displacement-excavation time step relationship curve. The method for determining the displacement is as follows: The horizontal axis represents the excavation time step, indicating the progress of the excavation. The vertical axis represents the cumulative displacement of key monitoring points on the slope (usually the points with the largest displacement). Figure 5 The baseline threshold is 6cm. The displacement development curve shows the cumulative displacement curve from the start to the end of excavation, obtained through finite element simulation calculation under a specific combination of parameters.

[0067] benchmark threshold Confirmed: such as Figure 5 As shown, after excavation, the displacement tends to stabilize, and its maximum cumulative displacement value is Y. max (The example in the figure is approximately 60cm). According to the method described in this invention, the reference threshold... Take 1 / 10 of the maximum displacement value, which is 6cm.

[0068] S43 Displacement warning for bedding slopes containing weak interlayers Displacement prediction value Yand displacement warning threshold In comparison, safety warnings for bedding slopes containing weak interlayers are classified into four levels, as follows: Y <0.8 Safety level, 0.8 ≤ Y <0.9 Note level, 0.9 ≤ Y <1.0 Warning level Y ≥1.0 Danger level.

[0069] Safety warnings are specifically divided into four levels, with clearly defined displacement ratio thresholds (0.8, 0.9, and 1.0 times the warning threshold). This tiered warning mechanism transforms continuous displacement predictions into discrete, easily understood, and operable risk levels, providing engineering managers with clear and intuitive decision-making support. It enables refined and tiered risk management, facilitating differentiated monitoring frequencies and response measures at different risk levels, thereby optimizing resource allocation and enhancing the scientific rigor and effectiveness of safety management.

[0070] The technical solution of this invention constructs the step-by-step excavation process of a slope as a time series and introduces a Long Short-Term Memory (LSTM) neural network model, which can effectively learn long-term dependencies, to accurately simulate the dynamic cumulative response of displacement in bedding slopes containing weak interlayers during excavation. This method effectively overcomes the shortcomings of existing technologies, such as the inability of the limit equilibrium method to reflect time-series effects, the low computational efficiency and difficulty in real-time application of numerical simulation methods, and the inability of static machine learning models to capture the path-dependent characteristics of step-by-step excavation. Its advantages lie in its short computation time and fast iteration speed, enabling real-time and dynamic prediction of displacement at each step of the excavation process. This provides timely and reliable technical support for slope safety management and risk early warning decisions during engineering construction, greatly improving the timeliness and accuracy of early warnings.

[0071] Example 2 This embodiment provides a dynamic safety early warning system for displacement of bedding slopes with weak interlayers based on time-series LSTM, used to implement the method described in Embodiment 1. Specifically, it includes: The parameter determination module is used to determine the dynamic characteristic parameters and static soil and rock parameters of the target slope and their value ranges. The sample database stores time-series feature-displacement sample data; The LSTM prediction model, or Long Short-Term Memory neural network model, is used to receive the input feature parameter sequence and output the predicted cumulative displacement sequence. The threshold calculation module is used to calculate the displacement warning threshold. The early warning judgment module is used to compare the predicted displacement output by the LSTM prediction model with the early warning threshold calculated by the threshold calculation module, and output safety early warning information according to the hierarchical rules.

[0072] The parameter determination module, sample database, LSTM prediction model, threshold calculation module, and early warning judgment module are connected in sequence. The output of the parameter determination module is connected to the sample database and the LSTM prediction model, respectively. The output of the sample database is connected to the LSTM prediction model, and the outputs of the LSTM prediction model and the threshold calculation module are connected to the early warning judgment module, respectively.

[0073] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic safety early warning method for displacement of bedding slopes with weak interlayers based on time-series LSTM, characterized in that, Includes the following steps: S1. Determine the model characteristic parameters and their value ranges: The model characteristic parameters include dynamic characteristic parameters that change during the excavation process and static soil and rock mass parameters that remain unchanged during the excavation process; The dynamic characteristic parameters include excavation volume V, excavation length L along the slope direction, excavation thickness D, and excavation slope α; The static soil and rock mass parameters include sliding surface cohesion c and sliding surface internal friction angle φ; S2. Within the range of the dynamic characteristic parameters, step-by-step excavation calculations are performed through finite element analysis simulation. The corresponding cumulative displacement response is monitored and obtained, forming N sets of time-series "characteristic-displacement" sample data, which constitute a time-series "characteristic-displacement" sample database. In the time-series "characteristic-displacement" sample data, each set of samples contains a sequence of characteristic parameters for T excavation time steps and a corresponding cumulative displacement sequence for T time steps. The cumulative displacement is the displacement of each monitoring point relative to the initial state before excavation. S3. Construct a Long Short-Term Memory (LSTM) neural network model, and train the LSTM neural network model using sample data obtained from the sample database to obtain a trained LSTM neural network model. The unit structure of the LSTM neural network model includes a forget gate, an input gate, and an output gate. The forget gate is used to learn and judge the degree of attenuation of the influence of historical excavation patterns on the current displacement. The input gate is used to identify the immediate and potential influence of the feature parameters of the current excavation step on the slope deformation. The output gate is used to map the cumulative mechanical effects represented inside the model to the predicted cumulative displacement. S4. Input the actual excavation dynamic characteristic parameters of the target slope into the trained LSTM neural network model to predict the cumulative displacement of each excavation step; calculate the displacement warning threshold based on the static soil and rock parameters of the target slope, and compare the predicted displacement with the threshold to achieve dynamic safety warning; the displacement warning threshold is determined based on the benchmark threshold and the adjustment function reflecting the influence of the sliding surface mechanical strength parameters, the adjustment function is the product of the benchmark threshold and the adjustment coefficient, and the adjustment coefficient is determined based on the sliding surface cohesion c and the sliding surface internal friction angle φ.

2. The method for dynamic safety early warning of displacement of bedding slopes with weak interlayers based on time-series LSTM according to claim 1, characterized in that, In step S1, the dynamic characteristic parameters include excavation geometric characteristic parameters, and the static soil and rock mass parameters include sliding surface mechanical strength parameters.

3. The method for dynamic safety early warning of displacement of bedding slopes with weak interlayers based on time-series LSTM according to claim 1, characterized in that, In step S2, the N sets of time-series "feature-displacement" sample data are generated using the Latin hypercube sampling method.

4. The method for dynamic safety early warning of displacement of bedding slopes with weak interlayers based on time-series LSTM according to claim 1, characterized in that, In step S2, the slope locations monitored during the finite element analysis simulation process include the top, middle, and toe of the slope.

5. The method for dynamic safety early warning of displacement of bedding slopes with weak interlayers based on time-series LSTM according to claim 1, characterized in that, In step S3, the unit structure of the LSTM neural network model includes a forget gate, an input gate, and an output gate; the forget gate is used to learn and judge the degree of attenuation of the influence of historical excavation patterns on the current displacement; the input gate is used to identify the immediate and potential influence of the feature parameters of the current excavation step on the slope deformation; and the output gate is used to map the cumulative mechanical effects represented inside the model to the predicted cumulative displacement.

6. The method for dynamic safety early warning of displacement of bedding slopes with weak interlayers based on time-series LSTM according to claim 1, characterized in that, In step S4, the displacement warning threshold is determined based on a reference threshold and an adjustment function that reflects the influence of the mechanical strength parameters of the sliding surface, wherein the reference threshold is obtained based on finite element simulation.

7. The method for dynamic safety early warning of displacement of bedding slopes with weak interlayers based on time-series LSTM according to claim 6, characterized in that, The adjustment function is the product of the reference threshold and the adjustment coefficient, which is determined based on the sliding surface mechanical strength parameter.

8. The method for dynamic safety early warning of displacement of bedding slopes with weak interlayers based on time-series LSTM according to claim 1, characterized in that, The safety warning is divided into several levels according to the ratio between the predicted displacement and the displacement warning threshold: safety level, attention level, warning level and danger level.

9. A dynamic safety early warning system for displacement of a bedding slope containing weak interlayers based on the early warning method described in any one of claims 1-8, characterized in that, include: The parameter determination module is used to determine the dynamic characteristic parameters and static soil and rock parameters of the target slope and their value ranges. The sample database stores time-series feature-displacement sample data; The LSTM prediction model, or Long Short-Term Memory neural network model, is used to receive the input feature parameter sequence and output the predicted cumulative displacement sequence. The threshold calculation module is used to calculate the displacement warning threshold. The early warning judgment module is used to compare the predicted displacement output by the LSTM prediction model with the early warning threshold calculated by the threshold calculation module, and output safety early warning information according to the hierarchical rules.

10. The system according to claim 9, characterized in that, The parameter determination module, the sample database, the LSTM prediction model, the threshold calculation module, and the early warning judgment module are connected in sequence. The output of the parameter determination module is connected to the sample database and the LSTM prediction model, respectively. The output of the sample database is connected to the LSTM prediction model, and the outputs of the LSTM prediction model and the threshold calculation module are connected to the early warning judgment module, respectively.

Images