A physical-data double-driven safety margin dynamic evaluation method for retaining wall

Abstract

Aiming at the problem that traditional methods cannot adapt to the performance degradation of retaining wall during service period, a physical-data dual-driven dynamic safety margin evaluation method for retaining wall is disclosed: 1, input the monitoring parameters during service period, calculate the stability coefficients of retaining wall, and map them into safety margin benchmark function; 2, construct a physical constraint learning subnetwork, guide the optimization of network parameters through a physical regularization loss function, and output the optimized safety margin benchmark value meeting the physical consistency requirement; 3, construct a performance degradation evolution subnetwork, capture the time sequence dependence of parameter degradation by using LSTM, guide the optimization of network parameters through a comprehensive loss function, and output the safety margin loss; 4, fuse the output results of each subnetwork, and establish the total safety margin function curve of retaining wall during service period; 5, establish a dynamic safety margin index and dynamically evaluate the safety margin of retaining wall during service period. Quantitative basis is provided for the operation and reinforcement decision and structural safety risk assessment of retaining wall.

Description

Technical Field

[0001] This invention relates to a data processing method for prediction purposes, which is applied to the safety assessment of geotechnical engineering and civil engineering structures, and specifically to a dynamic assessment method for the safety margin of retaining walls driven by both physical and data principles. Background Technology

[0002] Retaining walls, as a core support structure in geotechnical engineering, are widely used in engineering scenarios such as roads, slopes, and building foundation pits. During their service life, they face multiple factors including material aging, environmental erosion, and load fluctuations, leading to gradual performance degradation that directly threatens structural safety margins and operational stability. Existing dynamic assessment methods for retaining wall safety margins mainly fall into two categories: one relies solely on physical mechanism-based analytical models, which depend on idealized assumptions and parameter determinism, making it difficult to capture the complex and ever-changing parameter degradation patterns during service life and unable to dynamically correct deviations between the physical model and the actual structural response; the other relies solely on data-driven statistical models, which, due to a lack of physical mechanism constraints, are prone to extrapolation distortion and weak interpretability under limited monitoring data, making it difficult to guarantee engineering reliability. Furthermore, neither type of method effectively integrates physical constraints with performance degradation evolution characteristics, making it difficult to achieve dynamic tracking and accurate quantification of safety margins at different service stages.

[0003] Therefore, there is an urgent need for an evaluation method that integrates the advantages of physical mechanisms and data-driven approaches. This method should be based on physical models to uncover physical constraint relationships and use neural networks to uncover performance degradation patterns, thereby enabling dynamic evaluation of the safety margin of retaining walls throughout their entire life cycle and providing reliable support for operation and maintenance reinforcement decisions and safety risk assessments. Summary of the Invention

[0004] Given that traditional methods for assessing the safety margin of retaining walls cannot adapt to the performance degradation characteristics during service life and that there are discrepancies between physical models and actual responses, the present invention aims to provide a dynamic assessment method for the safety margin of retaining walls driven by both physical and data principles. By constructing a safety margin characterization index system to clarify the physical evaluation benchmark, and relying on the coupling mechanism of the physical constraint learning subnetwork and the performance degradation evolution subnetwork, the method automatically learns the physical correlation of multiple parameters and the performance degradation evolution law, thereby realizing the dynamic tracking and assessment of the safety margin at different service stages throughout the entire life cycle. This provides accurate and reliable quantitative support for retaining wall operation and maintenance reinforcement decisions and structural safety risk assessment.

[0005] To achieve the above objectives, this invention provides a dynamic assessment method for the safety margin of retaining walls driven by both physical and data principles, comprising the following steps:

[0006] S1, input the real-time monitoring parameters during the service period, complete the time discretization and parameter vectorization, calculate the stability coefficients of the retaining wall based on physical constraints, and map them to the safety margin benchmark function.

[0007] S2 constructs a physical constraint learning sub-network architecture consisting of an input layer, hidden layer, physical constraint layer, and output layer, using a physical regularization loss function. L total Guide network parameter optimization and establish a "parameter vector" that meets physical consistency requirements. x k "To "optimized safety margin benchmark function" M phys ( t k The mapping between ")".

[0008] S3 constructs a performance degradation evolution sub-network architecture of "input layer - LSTM layer - fully connected layer - output layer", and uses a comprehensive loss function. L decay Guide the network to optimize parameters and establish a "parameter vector". x k Compared with the optimized safety margin benchmark value M phys ( t k From “)” to “safety margin loss Δ” M ( t k The mapping between ")" outputs an effective safety margin loss function Δ. M ( t k ).

[0009] S4, using the superposition method to fuse the optimized safety margin benchmark function. M phys ( t k ) and safety margin loss function Δ M ( t k This generates a continuous total safety margin function that combines physical consistency and realistic attenuation. M total ( t k The curve of change.

[0010] S5, based on the total safety margin function M total ( t Calculate the baseline margin area A base ( t a , t b ) and margin remaining area A remain ( ta , t b Establish corresponding dynamic safety margin indicators. I ( t a , t b By comparing dynamic safety margin indicators I ( t a , t b ) and dynamic safety margin threshold I thre ( t a , t b This enables a dynamic and precise quantitative assessment of the safety margin of retaining walls during their service life.

[0011] S1 can be summarized as follows: Input the real-time monitoring parameters of the retaining wall during its service life, perform time discretization and parameter vectorization to obtain the time nodes of the service process. t k corresponding parameter vector x k Used to calculate the anti-sliding stability coefficient of the retaining wall at each time step. Q 1( t k ), overturning stability coefficient Q 2( t k ), shear stability coefficient Q 3( t k and foundation bearing capacity stability coefficient Q 4( t k And map each stability coefficient to a safety margin benchmark function at each time step. M ( t k ).

[0012] S1, the algorithm process is as follows:

[0013] S1.1 A strategy of dividing the continuous monitoring period of the retaining wall during its service life into equal time intervals is adopted. k Time nodes at each time step t k satisfy:

[0014] (1)

[0015] In the formula t 0 represents the initial service life time point; Δ t The discrete time step;k From 1 to n Natural numbers between.

[0016] S1.2 The service life monitoring parameters from S1.1 above are vectorized to obtain the parameter vector of the service process. x k satisfy:

[0017] (2)

[0018] In the formula G Due to the weight of the retaining wall itself, s G ( t k ) is the first k The distance from the toe of the retaining wall to the self-weight of each time step node. E az ( t k ) is the first k Vertical component of earth pressure corresponding to each time step node s az ( t k ) is the first k The distance from the vertical component of the earth pressure at each time step node to the wall toe. E ax ( t k ) is the first k The horizontal component of earth pressure corresponding to each time step node. s ax ( t k ) is the first k The distance from the toe of the wall to the horizontal component of the earth pressure corresponding to each time step node. m ( t k ) is the first k The coefficient of friction of the retaining wall base at each time step node. V ( t k ) is the first k The retaining wall at each time step node experiences the maximum shear force. f a ( t k ) is the first k Bearing capacity of the retaining wall foundation at each time step node p ( t k ) is the first k The retaining wall bottom pressure corresponding to each time step node;

[0019] From parameter vector x k The stability coefficients of the retaining wall at each time step can be calculated, wherein the anti-sliding stability coefficient is... Q 1( t k ), overturning stability coefficient Q 2( t k ), shear stability coefficient Q 3( t k and foundation bearing capacity stability coefficient Q 4( t k The calculations are as follows:

[0020] (3)

[0021] (4)

[0022] (5)

[0023] In the formula V ( t 0) represents the maximum shear force experienced by the retaining wall at the initial time point of its service life.

[0024] (6)

[0025] The safety margin benchmark function at each time step M ( t k The allowable value function for the stability coefficients of the retaining wall, taking into account all factors, is calculated as follows:

[0026] (7)

[0027] In the formula M ( t k ) is the first k The safety margin benchmark function of the retaining wall corresponding to each time step node; Q i ( t k ) represents the stability coefficients of each item. i = 1, 2, 3, 4; w i For each stability coefficient Q i ( t k The corresponding weighting coefficients are determined by expert evaluation and engineering experience.

[0028] S2 can be summarized as follows:

[0029] The subnetwork is learned through physical constraints, and the input parameter vector x k After training the subnetwork, an optimized safety margin benchmark function can be output. M phys ( t k Then, the loss function is adjusted using physical regularization. L total Guide network parameter optimization, if L total Less than or equal to the preset loss threshold d 1. Determine the optimized safety margin benchmark function. M phys ( t k If the physical consistency requirements are met, the process can proceed to the next step; otherwise, if... L total Greater than the preset loss threshold d If so, it is necessary to return to S2 to adjust the physical constraint layer weight matrix. W phys The weight matrix of the output layer of the physical constraint learning subnetwork W out and regularization weights l Retrain until the loss meets the threshold requirement. This is achieved by physically regularizing the loss function. L total The physical constraint learning subnetwork, after guiding network parameter optimization, can achieve the transformation from "parameter vector" to "physical constraint learning subnetwork". x k "To "optimized safety margin benchmark function" M phys ( t k The reliable mapping of ")" ensures that the output results are both physically interpretable and engineering-applicable.

[0030] S2, the specific method and process are as follows:

[0031] The physical constraint learning subnetwork described in S2.1 consists of an input layer, a hidden layer, a physical constraint layer, and an output layer, with the input parameter vector... x k After training the subnetwork, an optimized safety margin benchmark function can be output. M phys ( t k ).

[0032] The number of neurons and parameter vectors in the input layer of the physical constraint learning subnetwork xk The dimensions are consistent, and it is used to receive and input the parameter vector preprocessed by S1. x k .

[0033] The hidden layer of the physical constraint learning subnetwork has a two-layer fully connected structure. The first layer contains 48 neurons, and the second layer contains 32 neurons. Both layers use the ReLU activation function, which effectively alleviates the gradient vanishing problem. Through weighted summation of the two layers of neurons and activation function transformation, the hidden layer can extract features from the input data layer by layer and output the corresponding feature coefficient vector. H .

[0034] The physical constraint learning subnetwork has a physical constraint layer with 4 neurons and no activation function, used to directly output the stability coefficient vector derived in reverse derivation. Q total ( t k The physical constraint layer explicitly embeds mechanical constraints into the network architecture, forcing the network output to always derive from a reasonable mapping of mechanical coefficients, thus avoiding fitting distortions that deviate from the physical essence. Among these, the stability coefficient vector... Q total ( t k )for:

[0035] (8)

[0036] In the formula 、 、 and The first and second layers are derived from the physical constraint layers of the physical constraint learning subnetwork, respectively. k Mapped values ​​of the retaining wall's anti-sliding stability coefficient, anti-overturning stability coefficient, anti-shear stability coefficient, and foundation bearing capacity stability coefficient at each time step.

[0037] The stability coefficient vector Q total ( t k The calculation method for ) is as follows:

[0038] (9)

[0039] In the formula Q total ( t k The vector of stability coefficients is derived from the physical constraint layer of the physical constraint learning subnetwork. W physThis is the weight matrix of the physical constraint layer. H The feature coefficient vector output by the hidden layer. b phys This is the bias vector for the physical constraint layer.

[0040] The output layer of the physical constraint learning subnetwork contains one neuron and uses the Sigmoid activation function to output the stability coefficient vector of the physical constraint layer. Q total ( t k Mapped to the optimized safety margin benchmark function M phys ( t k This serves as a physical compliance reference benchmark for subsequent performance degradation evolution subnetworks.

[0041] The optimized safety margin benchmark function M phys ( t k The expression is:

[0042] (10)

[0043] In the formula M phys ( t k ) is the first k The optimized safety margin baseline value at each time step; s () is the Sigmoid activation function, which maps the output value to the interval [0,1]. W out To learn the weight matrix of the output layer of the physical constraint subnetwork, b out This is the bias vector for the output layer.

[0044] The physical constraint learning subnetwork described in S2.2 uses a physical regularization loss function. L total Guide network parameter optimization. Specifically, if... L total Less than or equal to the preset loss threshold d 1. Determine the optimized safety margin benchmark function. M phys ( t k If the physical consistency requirements are met, the process can proceed to the next step; otherwise, if... L total Greater than the preset loss threshold d If so, it is necessary to return to S2 to adjust the physical constraint layer weight matrix. Wphys The weight matrix of the output layer of the physical constraint learning subnetwork W out and regularization weights l Retrain until the loss meets the threshold requirement.

[0045] The physical regularization loss function L total Loss function fitted from data L MSE With physical constraint loss function L phys The weighted summation is used to ensure that the mapping relationship of network learning always closely follows the essence of mechanics.

[0046] The physical regularization loss function L total The expression is:

[0047] (11)

[0048] In the formula L MSE To fit the loss function to the data, L phys The physical constraint loss function, l These are regularization weights used to balance the accuracy of data fitting with the strength of physical consistency constraints.

[0049] The data fitting loss function L MSE The expression is:

[0050] (12)

[0051] In the formula n The total number of time steps. M phys ( t k ) is the first k The optimized safety margin baseline value at each time step M ( t k ) is the first digit calculated using physical formulas in S1. k Safety margin benchmark function at each time step.

[0052] The physical constraint loss function L phys The expression is:

[0053] (13)

[0054] In the formula i The value range is 1 to 4. The first physical constraint layer is derived from the physical constraint learning subnetwork by back-reaming. k Mapped values ​​of various stability coefficients of the retaining wall at each time step Q i ( t k ) is the first digit calculated using physical formulas in S1. k The stability coefficients of the retaining wall at each time step.

[0055] The preset loss threshold d The range of values ​​for 1 needs to be such that it can filter out invalid training results without being too strict and causing the training to fail to converge. It can be determined based on the engineering monitoring accuracy.

[0056] Furthermore, when L total ≤ d When 1 is reached, it indicates the optimized safety margin benchmark function. M phys ( t k If the physical consistency requirements are met, the process can proceed to the next step; when L total > d At step 1, it is necessary to return to S2 to adjust the physical constraint layer weight matrix. W phys The weight matrix of the output layer of the physical constraint learning subnetwork W out and regularization weights l After adjustments, restart network training until the requirements are met. L total ≤ d Requirements 1.

[0057] In S3, the performance degradation evolution subnetwork is established by creating a "parameter vector". x k Compared with the optimized safety margin benchmark value M phys ( t k From “)” to “safety margin loss Δ” M ( t k The mapping between ")" is the core objective;

[0058] The core components of this network include an input layer, an LSTM layer, a fully connected layer, and an output layer. The input "parameter vector" x k and optimized safety margin benchmark value M phys ( t kAfter training with the performance degradation evolution subnetwork, the "safety margin loss Δ" can be output. M ( t k )”;

[0059] Next, through the comprehensive loss function L decay Guide the network to optimize parameters, if L decay Less than or equal to the preset loss threshold d 2 and Δ M ( t k ) ≤ M phys ( t k Then the output safety margin loss Δ is determined. M ( t k If the condition is met, proceed to the next step; otherwise, if not, return to S3 to adjust the weight matrix of the fully connected layer. W fc Weight matrix of the output layer of the performance degradation evolution subnetwork W decay Physical constraint weights m and nonnegative constraint weights e Retrain the subnetwork until the loss meets the threshold requirement;

[0060] By comprehensive loss function L decay The performance degradation evolution subnetwork after guiding network parameter optimization can achieve the transformation from "parameter vector" to "parameter vector". x k Compared with the optimized safety margin benchmark value M phys ( t k From “)” to “safety margin loss Δ” M ( t k A reliable mapping between “)” ensures that the output results are consistent with the physical attenuation law.

[0061] Specifically, the performance degradation evolution subnetwork consists of an input layer, an LSTM layer, a fully connected layer, and an output layer, with the input being a "parameter vector". x k and the optimized safety margin benchmark value M phys ( t k After training through a performance degradation evolution subnetwork, the "safety margin loss Δ" is output. M ( t k)".

[0062] The input layer neurons of the performance degradation evolution subnetwork have a "physical baseline dimension + parameter vector dimension" and are used to receive and input two types of data: the optimized safety margin baseline value output by S2. M phys ( t k and the parameter vector after S1 preprocessing x k .

[0063] The LSTM (Long Short-Term Memory) layer of the performance degradation evolution subnetwork has a single-layer structure with 64 hidden units. It uses a combination of tanh and sigmoid activation functions to capture the temporal dependence and cumulative effect of parameter degradation. Its core state update formula satisfies:

[0064] (14)

[0065] (15)

[0066] (16)

[0067] (17)

[0068] (18)

[0069] (19)

[0070] In the formula i k For input gate, f k For the Gate of Oblivion g k In the candidate state of cells, o k For output gate, c k In cellular state, h k The hidden state vector; W ii , W if , W ig , W io These are the weight matrices from the input layer to each gate. W hi , W hf , W hg ,W ho These are the weight matrices from the hidden layer to each gate; b ii , b if , b ig , b io These are the bias vectors for each gate; x k The concatenated feature vectors from the input layer. h k-1 , c k-1 These are the hidden state vector and the cell state vector from the previous time step, respectively. s ( ) represents the Sigmoid activation function, and ⊙ represents the Hadamard product operation. This layer selectively memorizes the long-term trend of parameter decay through a gating mechanism to avoid the loss of temporal information.

[0071] The fully connected layer of the performance degradation evolution subnetwork has a two-layer structure, with 32 neurons in each layer. The ReLU activation function is used to alleviate the gradient vanishing problem. The fully connected layer is used to map the temporal features extracted by the LSTM layer into nonlinear loss features, the expression of which is:

[0072] (20)

[0073] In the formula, F decay The loss characteristics are those output by the fully connected layer. W fc This is the weight matrix of the fully connected layer. h k For the hidden state vector, b fc This is the bias vector for the fully connected layer.

[0074] The output layer of the performance degradation evolution subnetwork contains one neuron and uses the Softplus activation function to map the loss features to a non-negative safety margin loss Δ. M ( t k ).

[0075] The safety margin loss Δ M ( t k The expression for ) is:

[0076] (twenty one)

[0077] In the formula, Δ M ( t k) is the first k The safety margin loss at each time step, with a value ranging from [0, ... M phys ( t k The Softplus activation function ensures that the output is always non-negative. W decay The weight matrix of the output layer of the performance degradation evolution subnetwork. b decay This is the bias vector for the output layer.

[0078] The performance degradation evolution subnetwork uses a comprehensive loss function. L decay Guide the network to optimize parameters, if L decay Less than or equal to the preset loss threshold d 2 and Δ M ( t k ) ≤ M phys ( t k Then the output safety margin loss Δ is determined. M ( t k If the condition is met, proceed to the next step; otherwise, if not, return to S3 to adjust the weight matrix of the fully connected layer. W fc Weight matrix of the output layer of the performance degradation evolution subnetwork W decay Physical constraint weights m and nonnegative constraint weights e Retrain the subnetwork until the loss meets the threshold requirement.

[0079] The comprehensive loss function L decay Loss function fitted by loss L MSE-Δ Physical consistency constraint loss function L consist Non-negative constraint loss L nonneg The function is weighted and summed to ensure that the loss is consistent with the physical decay law and is non-negative.

[0080] The comprehensive loss function L decay The expression is:

[0081] (twenty two)

[0082] In the formula LMSE-Δ The loss function is the loss fit. L consist The loss function is the physical consistency constraint. L nonneg The loss function is a non-negativity constraint. m For physical constraint weights; e These are non-negative constraint weights, used to balance fitting accuracy and constraint strength.

[0083] The loss function of the loss fitting L MSE-Δ The expression is:

[0084] (twenty three)

[0085] In the formula n Δ is the total number of time steps. M ( t k ) is the first k The amount of safety margin loss at each time step M phys ( t k ) is the first k The optimized safety margin baseline value at each time step M meas ( t k ) is the first k The safety margin is obtained from on-site measurements at each time step.

[0086] The physical consistency constraint loss function L consist The expression is:

[0087] (twenty four)

[0088] In the formula i The value range of Δ is 1~4. M ( t k ) is the first k The amount of safety margin loss at each time step M phys ( t k ) is the first k The optimized safety margin baseline value at each time step The first physical constraint layer is derived from the physical constraint learning subnetwork by back-reaming. k Mapped values ​​of various stability coefficients of the retaining wall at each time step Q i ( t k) is the first digit calculated using physical formulas in S1. k The stability coefficients of the retaining wall at each time step.

[0089] The nonnegativity constraint loss function L nonneg The expression is:

[0090] (25)

[0091] In the formula, max{0,-Δ M ( t k Ensure that a penalty loss is incurred when the network output is negative, forcing Δ M ( t k ) ≥ 0.

[0092] The preset loss threshold d The value range of 2 needs to ensure that it can filter out invalid training results without being too strict and causing the training to be difficult to converge. It can be determined according to the engineering monitoring accuracy.

[0093] Furthermore, when the comprehensive loss function L decay ≤ d 2 and Δ M ( t k ) ≤ M phys ( t k When ), it indicates the loss of safety margin Δ in the output. M ( t k If the condition is met, proceed to the next step; otherwise, if not, return to S3 to adjust the weight matrix of the fully connected layer. W fc Weight matrix of the output layer of the performance degradation evolution subnetwork W decay Physical constraint weights m and nonnegative constraint weights e After adjustment, restart the subnetwork for training until the loss meets the threshold requirement.

[0094] In S4, the total safety margin M total ( t k The optimized safety margin baseline value output by S2 M phys ( t k The safety margin loss Δ of S3 output M (t k The operation essentially involves superimposing the attenuation loss as a reverse correction term onto the physical reference margin, thereby achieving the fusion of the reference term and the correction term.

[0095] The total safety margin M total ( t k The expression for ) is:

[0096] (26)

[0097] The total safety margin value at discrete time steps is integrated and smoothed to obtain the total safety margin function over the service life of the retaining wall. M total ( t )for:

[0098] (27)

[0099] In the formula n This represents the total number of time steps during the service period. k For [1, n Natural numbers between ] M total (t) is the total safety margin function over the entire service life, used for subsequent safety margin index calculations.

[0100] In S5, based on the total safety margin function M total ( t Calculate the baseline margin area A base ( t a , t b ) and margin remaining area A remain ( t a , t b Establish corresponding dynamic safety margin indicators. I ( t a , t b By comparing dynamic safety margin indicators I ( t a , t b ) and dynamic safety margin threshold I thre ( t a ,t b This enables dynamic assessment of the safety margin of retaining walls during their service life. t a and t b These refer to the left and right boundary time nodes of the selected rolling time window for evaluation.

[0101] The benchmark margin area A base ( t a , t b This refers to the theoretical remaining safety margin of a retaining wall during a certain period of its service life, where the remaining area of ​​the margin is... A remain ( t a , t b This refers to the actual remaining safety margin within that period. Based on the total safety margin function described in S4. M total ( t The reference margin area A base ( t a , t b ) and margin remaining area A remain ( t a , t b The calculations are as follows:

[0102] (28)

[0103] In the formula t a and t b These are the left and right boundary time nodes of the selected rolling time window for evaluation; M total ( t 0) is t The total safety margin function value corresponding to time 0.

[0104] (29)

[0105] The dynamic safety margin index I ( t a , t b ) is used to evaluate arbitrary rolling time windows. ta , t b The corresponding safety margin is calculated as follows:

[0106] (30)

[0107] In the formula, the dynamic safety margin index I ( t a , t b The larger the value of ∈ (0,1), the more sufficient the safety margin of the retaining wall.

[0108] The dynamic safety margin index threshold I thre ( t a , t b Used to determine dynamic safety margin indicators I ( t a , t b Does it meet the safety requirements? I thre ( t a , t b The calculation formula is as follows:

[0109] (31)

[0110] In the formula [ Q ] represents the stability coefficients of each term in formula (7). Q i ( t k The minimum value among the corresponding theoretical limits is determined by the safety factor specified in the retaining wall design code. M total ( t 0) is t The total safety margin function value corresponding to time 0.

[0111] Based on the aforementioned "dynamic safety margin index" I ( t a , t b A safety assessment of the retaining wall during its service life should be conducted. I ( t a , t b )∈( I thre (t a , t b ),1), indicates that the stability of the retaining wall exceeds the safety requirements of the specifications, the safety margin of the retaining wall is very sufficient, and routine monitoring is sufficient; when I ( t a , t b )∈(0, I thre ( t a , t b This indicates that the stability of the retaining wall is less than the safety requirements specified in the code, the safety margin is insufficient, and reinforcement or repair is required.

[0112] Compared with existing technologies, this invention integrates the core advantages of physical mechanisms and data-driven approaches, breaking through the limitations of traditional evaluation paradigms that rely on single physical modeling or pure data fitting. It proposes a dynamic evaluation method for the safety margin of retaining walls driven by both physical and data principles, offering the following beneficial effects:

[0113] (1) The physical mechanism is deeply coupled with the neural network, taking into account both interpretability and fitting ability: By embedding the physical constraint layer in the physical constraint learning sub-network and constructing the prior constraint relationship in combination with the physical regularization loss function, the output results are guaranteed to closely follow the mechanical laws such as anti-slip and anti-overturning, solving the problem that the pure data-driven model lacks physical basis and has weak interpretability; at the same time, the LSTM layer of the performance decay evolution sub-network is used to capture the temporal dependence and cumulative effect of parameter decay, making up for the defects of the pure physical model that it is difficult to adapt to dynamic decay and cannot correct the theoretical and actual deviations, thus achieving the dual guarantee of "physical mechanism" and "data accuracy";

[0114] (2) Both sub-networks incorporate loss functions to guide parameter optimization, improving the reliability and stability of computational results: the physical constraint learning sub-network determines the parameters through a loss threshold. L total ≤ d 1) Achieve physical consistency verification of baseline margin; the performance degradation evolution subnetwork uses dual-judgment ( L decay ≤ d 2 and DM ( t k ) ≤ M phys ( t kEnsuring the effectiveness and rationality of the loss amount. Using the loss function to guide network parameter optimization can significantly improve the reliability of the overall safety margin assessment, solving the problems of blind parameter optimization and output results that easily deviate from engineering reality in traditional models;

[0115] (3) Quantitative characterization of the performance and safety margin evolution law of retaining walls during service life. The remaining margin area comprehensively reflects the coupling effect of "functional loss amplitude - duration". The dynamic safety margin index realizes the real-time normalized characterization and quantification of the safety margin of retaining walls, providing intuitive and clear evaluation results for the safety margin of retaining walls, and solving the problem of fuzzy quantification of traditional evaluation indicators;

[0116] (4) It has high engineering application value and can provide accurate quantitative support for operation and maintenance decision-making: the output total safety margin curve can directly provide complete data for the calculation and evaluation of safety margin indicators, can quantitatively characterize the safety status at different service stages, and can provide reliable quantitative basis for the selection of the timing of the maintenance and reinforcement of retaining walls, the optimization of reinforcement schemes and the assessment of structural safety risks. Attached Figure Description

[0117] Figure 1 This is a schematic diagram of the scenario and system for the embodiment;

[0118] Figure 2 This is a flowchart of the dynamic assessment method for the safety margin of retaining walls based on a physical-data dual-drive approach, as described in this invention.

[0119] Figure 3 This is a schematic diagram of the network architecture of the physical constraint learning subnetwork in the embodiment;

[0120] Figure 4 This is a schematic diagram of the network architecture of the performance degradation evolution subnetwork in the embodiment;

[0121] Figure 5 This is a schematic diagram of the LSTM neural network unit in the embodiment;

[0122] Figure 6 This is a schematic diagram of the baseline margin area, remaining margin area, and total safety margin function of the retaining wall in the embodiment. Detailed Implementation

[0123] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0124] Example

[0125] This embodiment provides a physical-data-driven dynamic assessment method for the safety margin of retaining walls. This method dynamically tracks and quantifies the safety margin of retaining walls throughout their entire service life, enabling accurate determination of maintenance and reinforcement timing and safety risk assessment. For gravity-type retaining walls, their service environment faces multiple degradation factors such as seasonal rainfall erosion, temperature fluctuations, and material aging, requiring continuous monitoring and dynamic assessment to ensure service safety.

[0126] Figure 1 This is a schematic diagram of the implementation scenario and system. Three sets of fiber optic strain gauges are deployed along the height of the retaining wall to monitor wall displacement changes. The distance from the wall toe to the wall's toe, determined by the retaining wall's design dimensions, is calculated using the strain data. Vibrating wire earth pressure sensors are evenly deployed at different depths on the back of the wall to collect the horizontal and vertical components of the earth pressure. Two sets of pre-embedded friction coefficient sensors are deployed at the front toe, rear heel, and middle of the wall base to monitor the base friction characteristics. Simultaneously, pre-embedded pressure sensors are deployed at the bottom of the wall to collect the pressure at the base of the retaining wall. The bearing capacity of the foundation is obtained by combining this with the physical and mechanical parameters of the foundation soil. All sensors are connected to an industrial-grade data acquisition box deployed at the edge of the site. The acquisition box collects the monitoring data from each sensor in real time at one-hour intervals and transmits the service-life monitoring data to a remote cloud server via a 5G module. This provides continuous and complete raw data input for subsequent time discretization, parameter vectorization, and safety margin assessment processes, and the subsequent implementation steps S1-S5 are run on the cloud server.

[0127] like Figure 2 As shown in the figure, this embodiment presents a physical-data dual-driven dynamic assessment method for the safety margin of retaining walls. The specific implementation steps are as follows:

[0128] S1, input the real-time monitoring parameters during the service period, complete the time discretization and parameter vectorization, calculate the stability coefficients of the retaining wall based on physical constraints, and map them to the safety margin benchmark function.

[0129] S1 can be summarized as follows: Input the real-time monitoring parameters of the retaining wall during its service life, perform time discretization and parameter vectorization to obtain the time nodes of the service process. t k corresponding parameter vector x k Used to calculate the anti-sliding stability coefficient of the retaining wall at each time step. Q 1( t k ), overturning stability coefficient Q 2( t k ), shear stability coefficient Q 3( t kand foundation bearing capacity stability coefficient Q 4( t k And map each stability coefficient to a safety margin benchmark function at each time step. M ( t k ).

[0130] S2 constructs a physical constraint learning sub-network architecture consisting of an input layer, hidden layer, physical constraint layer, and output layer, using a physical regularization loss function. L total Guide network parameter optimization and establish a "parameter vector" that meets physical consistency requirements. x k "To "optimized safety margin benchmark function" M phys ( t k The mapping between ")".

[0131] S2 can be summarized as follows: A sub-network is learned through physical constraints, with the input parameter vector... x k After training the subnetwork, an optimized safety margin benchmark function can be output. M phys ( t k Then, the loss function is adjusted using physical regularization. L total Guide network parameter optimization, if L total Less than or equal to the preset loss threshold d 1. Determine the optimized safety margin benchmark function. M phys ( t k If the physical consistency requirements are met, the process can proceed to the next step; otherwise, if... L total Greater than the preset loss threshold d If so, it is necessary to return to S2 to adjust the physical constraint layer weight matrix. W phys The weight matrix of the output layer of the physical constraint learning subnetwork W out and regularization weights l Retrain until the loss meets the threshold requirement. This is achieved by physically regularizing the loss function. L total The physical constraint learning subnetwork, after guiding network parameter optimization, can achieve the transformation from "parameter vector" to "physical constraint learning subnetwork". x k "To "optimized safety margin benchmark function" M phys( t k The reliable mapping of ")" ensures that the output results are both physically interpretable and engineering-applicable.

[0132] S3 constructs a performance degradation evolution sub-network architecture of "input layer - LSTM layer - fully connected layer - output layer", and uses a comprehensive loss function. L decay Guide the network to optimize parameters and establish a "parameter vector". x k Compared with the optimized safety margin benchmark value M phys ( t k From “)” to “safety margin loss Δ” M ( t k The mapping between ")" outputs an effective safety margin loss function Δ. M ( t k ).

[0133] S4, using the superposition method to fuse the optimized safety margin benchmark function. M phys ( t k ) and safety margin loss function Δ M ( t k This generates a continuous total safety margin function that combines physical consistency and realistic attenuation. M total ( t k The curve of change.

[0134] S5, based on the total safety margin function M total ( t Calculate the baseline margin area A base ( t a , t b ) and margin remaining area A remain ( t a , t b Establish corresponding dynamic safety margin indicators. I ( t a , t b By comparing dynamic safety margin indicators I ( ta , t b ) and dynamic safety margin threshold I thre ( t a , t b This enables a dynamic and precise quantitative assessment of the safety margin of retaining walls during their service life.

[0135] In this embodiment, the above implementation steps are described as follows:

[0136] S1, the algorithm process is as follows:

[0137] S1.1 A strategy of dividing the continuous monitoring period of the retaining wall during its service life into equal time intervals is adopted. k Time nodes at each time step t k satisfy:

[0138] (1)

[0139] In the formula t 0 represents the initial service life time point; Δ t The discrete time step; k From 1 to n Natural numbers between.

[0140] S1.2 The service life monitoring parameters from S1.1 above are vectorized to obtain the parameter vector of the service process. x k satisfy:

[0141] (2)

[0142] In the formula G Due to the weight of the retaining wall itself, s G ( t k ) is the first k The distance from the toe of the retaining wall to the self-weight of each time step node. E az ( t k ) is the first k Vertical component of earth pressure corresponding to each time step node s az ( t k ) is the first k The distance from the vertical component of the earth pressure at each time step node to the wall toe. E ax ( t k) is the first k The horizontal component of earth pressure corresponding to each time step node. s ax ( t k ) is the first k The distance from the toe of the wall to the horizontal component of the earth pressure corresponding to each time step node. m ( t k ) is the first k The coefficient of friction of the retaining wall base at each time step node. V ( t k ) is the first k The retaining wall at each time step node experiences the maximum shear force. f a ( t k ) is the first k Bearing capacity of the retaining wall foundation at each time step node p ( t k ) is the first k The retaining wall bottom pressure corresponding to each time step node;

[0143] From parameter vector x k The stability coefficients of the retaining wall at each time step can be calculated, wherein the anti-sliding stability coefficient is... Q 1( t k ), overturning stability coefficient Q 2( t k ), shear stability coefficient Q 3( t k and foundation bearing capacity stability coefficient Q 4( t k The calculations are as follows:

[0144] (3)

[0145] (4)

[0146] (5)

[0147] In the formula V ( t 0) represents the maximum shear force experienced by the retaining wall at the initial time point of its service life.

[0148] (6)

[0149] The safety margin benchmark function at each time step M ( t k The allowable value function for the stability coefficients of the retaining wall, taking into account all factors, is calculated as follows:

[0150] (7)

[0151] In the formula M ( t k ) is the first k The safety margin benchmark function of the retaining wall corresponding to each time step node; Q i ( t k ) represents the stability coefficients of each item. i = 1, 2, 3, 4; w i For each stability coefficient Q i ( t k The corresponding weighting coefficients are determined by expert evaluation and engineering experience.

[0152] Furthermore, in S1.1, the monitoring parameters for the retaining wall during its service life include the retaining wall's own weight. G and its distance from the toe of the wall s G Vertical component of earth pressure E az and its distance from the toe of the wall s az Horizontal component of earth pressure E ax and its distance from the toe of the wall s ax Coefficient of friction of retaining wall base m The retaining wall is subjected to the maximum shear force V Bearing capacity of the foundation at the base of the retaining wall f a and the pressure at the base of the retaining wall p ;

[0153] The service life of the retaining wall is divided into time steps, i.e., time discretization, and the service life is divided into equal parts. n The time step, of which the first k The time node corresponding to each time step t k satisfy:

[0154] (1)

[0155] In the formula tk For the first k The time node corresponding to each time step. k From 1 to n Natural numbers between; t 0 represents the initial time point of the service life; Δt The time interval after dividing the service period into equal parts;

[0156] Real-time monitoring of parameters that change over time yields the first... k The distance from the toe of the retaining wall to the self-weight of each time step node. s G ( t k Vertical component of earth pressure E az ( t k ) and its distance from the toe of the wall s az ( t k ), horizontal component of earth pressure E ax ( t k ) and its distance from the toe of the wall s ax ( t k ), coefficient of friction of retaining wall base m ( t k The retaining wall experiences maximum shear force. V ( t k ), Bearing capacity of the foundation under the retaining wall f a ( t k ) and the pressure at the bottom of the retaining wall p ( t k ).

[0157] S2, the specific method and process are as follows:

[0158] The physical constraint learning subnetwork described in S2.1 consists of an input layer, a hidden layer, a physical constraint layer, and an output layer, with the input parameter vector... x k After training the subnetwork, an optimized safety margin benchmark function can be output. M phys ( t k ).

[0159] The number of neurons and parameter vectors in the input layer of the physical constraint learning subnetwork xk The dimensions are consistent, and it is used to receive and input the parameter vector preprocessed by S1. x k .

[0160] The hidden layer of the physical constraint learning subnetwork has a two-layer fully connected structure. The first layer contains 48 neurons, and the second layer contains 32 neurons. Both layers use the ReLU activation function, which effectively alleviates the gradient vanishing problem. Through weighted summation of the two layers of neurons and activation function transformation, the hidden layer can extract features from the input data layer by layer and output the corresponding feature coefficient vector. H .

[0161] The physical constraint learning subnetwork has a physical constraint layer with 4 neurons and no activation function, used to directly output the stability coefficient vector derived in reverse derivation. Q total ( t k The physical constraint layer explicitly embeds mechanical constraints into the network architecture, forcing the network output to always derive from a reasonable mapping of mechanical coefficients, thus avoiding fitting distortions that deviate from the physical essence. Among these, the stability coefficient vector... Q total ( t k )for:

[0162] (8)

[0163] In the formula 、 、 and The first and second layers are derived from the physical constraint layers of the physical constraint learning subnetwork, respectively. k Mapped values ​​of the retaining wall's anti-sliding stability coefficient, anti-overturning stability coefficient, anti-shear stability coefficient, and foundation bearing capacity stability coefficient at each time step.

[0164] The stability coefficient vector Q total ( t k The calculation method for ) is as follows:

[0165] (9)

[0166] In the formula Q total ( t k The vector of stability coefficients is derived from the physical constraint layer of the physical constraint learning subnetwork. W physThis is the weight matrix of the physical constraint layer. H The feature coefficient vector output by the hidden layer. b phys This is the bias vector for the physical constraint layer.

[0167] The output layer of the physical constraint learning subnetwork contains one neuron and uses the Sigmoid activation function to output the stability coefficient vector of the physical constraint layer. Q total ( t k Mapped to the optimized safety margin benchmark function M phys ( t k This serves as a physical compliance reference benchmark for subsequent performance degradation evolution subnetworks.

[0168] The optimized safety margin benchmark function M phys ( t k The expression is:

[0169] (10)

[0170] In the formula M phys ( t k ) is the first k The optimized safety margin baseline value at each time step; s () is the Sigmoid activation function, which maps the output value to the interval [0,1]. W out To learn the weight matrix of the output layer of the physical constraint subnetwork, b out This is the bias vector for the output layer.

[0171] The physical constraint learning subnetwork described in S2.2 uses a physical regularization loss function. L total Guide network parameter optimization. Specifically, if... L total Less than or equal to the preset loss threshold d 1. Determine the optimized safety margin benchmark function. M phys ( t k If the physical consistency requirements are met, the process can proceed to the next step; otherwise, if... L total Greater than the preset loss threshold d If so, it is necessary to return to S2 to adjust the physical constraint layer weight matrix. Wphys The weight matrix of the output layer of the physical constraint learning subnetwork W out and regularization weights l Retrain until the loss meets the threshold requirement.

[0172] The physical regularization loss function L total Loss function fitted from data L MSE With physical constraint loss function L phys The weighted summation is used to ensure that the mapping relationship of network learning always closely follows the essence of mechanics.

[0173] The physical regularization loss function L total The expression is:

[0174] (11)

[0175] In the formula L MSE To fit the loss function to the data, L phys The physical constraint loss function, l These are regularization weights used to balance the accuracy of data fitting with the strength of physical consistency constraints.

[0176] The data fitting loss function L MSE The expression is:

[0177] (12)

[0178] In the formula n The total number of time steps. M phys ( t k ) is the first k The optimized safety margin baseline value at each time step M ( t k ) is the first digit calculated using physical formulas in S1. k Safety margin benchmark function at each time step.

[0179] The physical constraint loss function L phys The expression is:

[0180] (13)

[0181] In the formula i The value range is 1 to 4. The first physical constraint layer is derived from the physical constraint learning subnetwork by back-reaming. k Mapped values ​​of various stability coefficients of the retaining wall at each time step Q i ( t k ) is the first digit calculated using physical formulas in S1. k The stability coefficients of the retaining wall at each time step.

[0182] The preset loss threshold d The range of values ​​for 1 needs to be such that it can filter out invalid training results without being too strict and causing the training to fail to converge. It can be determined based on the engineering monitoring accuracy.

[0183] Furthermore, when L total ≤ d When 1 is reached, it indicates the optimized safety margin benchmark function. M phys ( t k If the physical consistency requirements are met, the process can proceed to the next step; when L total > d At step 1, it is necessary to return to S2 to adjust the physical constraint layer weight matrix. W phys The weight matrix of the output layer of the physical constraint learning subnetwork W out and regularization weights l After adjustments, restart network training until the requirements are met. L total ≤ d Requirements 1.

[0184] like Figure 3 As shown, the physical constraint learning subnetwork in this embodiment adopts a network architecture of "input layer - hidden layer - physical constraint layer - output layer", and the specific implementation details are as follows:

[0185] S2.1.1 Network architecture parameter configuration:

[0186] 1) Input layer: Receives the parameter vector preprocessed by S1 x k .

[0187] 2) Hidden Layers: A two-layer fully connected structure is used. The first layer contains 48 neurons, and the second layer contains 32 neurons. Both layers use the ReLU activation function, and the initial values ​​of the weight matrix follow a normal distribution. N (0, 0.01), the initial value of the bias vector is set to 0; the hidden layer outputs the feature coefficient vector through layer-by-layer feature extraction. H .

[0188] 3) Physical constraint layer: Contains 4 neurons, no activation function, this layer outputs a stability coefficient vector. Q total ( t k ):

[0189] (8)

[0190] In the formula Q total ( t k The vector of stability coefficients is derived from the physical constraint layer of the physical constraint learning subnetwork. W phys This is the weight matrix of the physical constraint layer. H The feature coefficient vector output by the hidden layer. b phys This is the bias vector for the physical constraint layer.

[0191] 4) Output layer: Contains 1 neuron, using the Sigmoid activation function. This layer will... Q total ( t k Mapped to the optimized safety margin benchmark function M phys ( t k The expression is:

[0192] (9)

[0193] In the formula M phys ( t k ) is the first k The optimized safety margin baseline value at each time step; s (.) is the Sigmoid activation function, which maps the output value to the interval [0,1]. W out To learn the weight matrix of the output layer of the physical constraint subnetwork, b out This is the bias vector for the output layer.

[0194] S2.1.2 Network parameter optimization and physical consistency verification:

[0195] 1) Physical regularization loss function L total Configuration: Initial regularization weights l Set to 0.5, data fitting loss function LMSE With physical constraint loss function L phys The weighted summation expression is:

[0196] (10)

[0197] The data fitting loss function L MSE The expression is:

[0198] (11)

[0199] In the formula n The total number of time steps. M phys ( t k ) is the first k The optimized safety margin baseline value at each time step M ( t k ) is the first digit calculated using physical formulas in S1. k Safety margin benchmark function at each time step.

[0200] The physical constraint loss function L phys The expression is:

[0201] (12)

[0202] In the formula i The value range is 1 to 4. The first physical constraint layer is derived from the physical constraint learning subnetwork by back-reaming. k Mapped values ​​of various stability coefficients of the retaining wall at each time step Q i ( t k ) is the first digit calculated using physical formulas in S1. k The stability coefficients of the retaining wall at each time step.

[0203] 2) Training parameter settings: preset loss threshold d 1. When training reaches L total ≤ d Optimization stops at step 1, at which point the training results meet the physical consistency requirement, and the optimized safety margin benchmark function is output. M phys ( t k ).

[0204] like Figure 4As shown, the performance degradation evolution subnetwork in this embodiment adopts an architecture of "input layer - LSTM layer - fully connected layer - output layer", and the specific implementation details are as follows:

[0205] S3.1 Network Architecture Parameter Configuration:

[0206] 1) Input layer: The input is the parameter vector preprocessed by S1. x k The optimized safety margin baseline value output by S2 M phys ( t k ).

[0207] 2) LSTM layer: Contains 64 hidden units, using a combination of tanh and sigmoid activation functions. This layer captures the temporal dependence and cumulative effect of parameter decay. Figure 5 As shown, its core state update formula satisfies:

[0208] (14)

[0209] (15)

[0210] (16)

[0211] (17)

[0212] (18)

[0213] (19)

[0214] In the formula i k For input gate, f k For the Gate of Oblivion g k In the candidate state of cells, o k For output gate, c k In cellular state, h k The hidden state vector; W ii , W if , W ig , W io These are the weight matrices from the input layer to each gate. W hi , W hf, W hg , W ho These are the weight matrices from the hidden layer to each gate; b ii , b if , b ig , b io These are the bias vectors for each gate; x k The concatenated feature vectors from the input layer. h k-1 , c k-1 These are the hidden state vector and the cell state vector from the previous time step, respectively. s ( ) represents the Sigmoid activation function, and ⊙ represents the Hadamard product operation. This layer selectively memorizes the long-term trend of parameter decay through a gating mechanism to avoid the loss of temporal information.

[0215] 3) Fully Connected Layer: A two-layer structure is used, with 32 neurons in each layer. The ReLU activation function is employed to map the temporal features extracted by the LSTM layer into non-linear loss features. Its expression is:

[0216] (20)

[0217] In the formula, F decay The loss characteristics are those output by the fully connected layer. W fc This is the weight matrix of the fully connected layer. h k For the hidden state vector, b fc This is the bias vector for the fully connected layer.

[0218] 4) Output layer: The Softplus activation function is used to map the loss features to a non-negative safety margin loss Δ. M ( t k ):

[0219] (twenty one)

[0220] In the formula, Δ M ( t k ) is the first k The safety margin loss at each time step, with a value ranging from [0, ... M phys ( tk The Softplus activation function ensures that the output is always non-negative. W decay The weight matrix of the output layer of the performance degradation evolution subnetwork. b decay This is the bias vector for the output layer.

[0221] S3.2 Network parameter optimization and output validity determination:

[0222] 1) Comprehensive Loss Function L decay Configuration:

[0223] (twenty two)

[0224] In the formula L MSE-Δ The loss function is the loss fit. L consist The loss function is the physical consistency constraint. L nonneg The loss function is a non-negativity constraint. m For physical constraint weights; e These are non-negative constraint weights, used to balance fitting accuracy and constraint strength.

[0225] Loss function for loss fitting L MSE-Δ for:

[0226] (twenty three)

[0227] In the formula n Δ is the total number of time steps. M ( t k ) is the first k The amount of safety margin loss at each time step M phys ( t k ) is the first k The optimized safety margin baseline value at each time step M meas ( t k ) is the first k The safety margin is obtained from on-site measurements at each time step.

[0228] Physical consistency constraint loss function L consist for:

[0229] (twenty four)

[0230] In the formulai The value range of Δ is 1~4. M ( t k ) is the first k The amount of safety margin loss at each time step M phys ( t k ) is the first k The optimized safety margin baseline value at each time step The first physical constraint layer is derived from the physical constraint learning subnetwork by back-reaming. k Mapped values ​​of various stability coefficients of the retaining wall at each time step Q i ( t k ) is the first digit calculated using physical formulas in S1. k The stability coefficients of the retaining wall at each time step.

[0231] 2) Training parameter settings: preset loss threshold d 2. When training reaches the comprehensive loss function L decay ≤ d 2 and Δ M ( t k ) ≤ M phys ( t k Optimization stops when the training output's safety margin loss Δ is reached. M ( t k It is valid and can be used in subsequent S4 calculations.

[0232] S4, using the superposition method to fuse the optimized safety margin benchmark function. M phys ( t k ) and safety margin loss function Δ M ( t k This generates a continuous total safety margin function that combines physical consistency and realistic attenuation. M total ( t k The curve of change.

[0233] S4 Algorithm Process:

[0234] The optimized safety margin baseline value output by S2 M phys ( t kThe safety margin loss Δ of S3 output M ( t k The total safety margin is obtained by superimposing these values. M total ( t k ):

[0235] (26)

[0236] The total safety margin value at discrete time steps is integrated and smoothed to obtain the total safety margin function over the service life of the retaining wall. M total ( t )for:

[0237] (27)

[0238] In the formula n This represents the total number of time steps during the service period. k For [1, n Natural numbers between ] M total (t) is the total safety margin function over the entire service life, used for subsequent safety margin index calculations.

[0239] In S5, based on the total safety margin function M total ( t Calculate the baseline margin area A base ( t a , t b ) and margin remaining area A remain ( t a , t b Establish corresponding dynamic safety margin indicators. I ( t a , t b By comparing dynamic safety margin indicators I ( t a , t b ) and dynamic safety margin threshold I thre ( t a , t b This enables dynamic assessment of the safety margin of retaining walls during their service life.t a and t b These refer to the left and right boundary time nodes of the selected rolling time window for evaluation.

[0240] The benchmark margin area A base ( t a , t b This refers to the theoretical remaining safety margin of a retaining wall during a certain period of its service life, where the remaining area of ​​the margin is... A remain ( t a , t b This refers to the actual remaining safety margin within that period. Based on the total safety margin function described in S4. M total ( t The reference margin area A base ( t a , t b ) and margin remaining area A remain ( t a , t b The calculations are as follows:

[0241] (28)

[0242] In the formula t a and t b These are the left and right boundary time nodes of the selected rolling time window for evaluation; M total ( t 0) is t The total safety margin function value corresponding to time 0.

[0243] (29)

[0244] The dynamic safety margin index I ( t a , t b ) is used to evaluate arbitrary rolling time windows. t a , t bThe corresponding safety margin is calculated as follows:

[0245] (30)

[0246] In the formula, the dynamic safety margin index I ( t a , t b The larger the value of ∈ (0,1), the more sufficient the safety margin of the retaining wall.

[0247] The dynamic safety margin index threshold I thre ( t a , t b Used to determine dynamic safety margin indicators I ( t a , t b Does it meet the safety requirements? I thre ( t a , t b The calculation formula is as follows:

[0248] (31)

[0249] In the formula [ Q ] represents the stability coefficients of each term in formula (7). Q i ( t k The minimum value among the corresponding theoretical limits is determined by the safety factor specified in the retaining wall design code. M total ( t 0) is t The total safety margin function value corresponding to time 0.

[0250] Based on the aforementioned "dynamic safety margin index" I ( t a , t b A safety assessment of the retaining wall during its service life should be conducted. I ( t a , t b )∈( I thre ( t a , tb ),1), indicates that the stability of the retaining wall exceeds the safety requirements of the specifications, the safety margin of the retaining wall is very sufficient, and routine monitoring is sufficient; when I ( t a , t b )∈(0, I thre ( t a , t b This indicates that the stability of the retaining wall is less than the safety requirements specified in the code, the safety margin is insufficient, and reinforcement or repair is required.

[0251] Benchmark margin area A base ( t a , t b )for:

[0252] (28)

[0253] Marginal Remaining Area A remain ( t a , t b )for:

[0254] (29)

[0255] Dynamic safety margin index I ( t a , t b )for:

[0256] (30)

[0257] Dynamic safety margin threshold I thre ( t a , t b Used to determine dynamic safety margin indicators I ( t a , t b Does it meet the safety requirements? I thre ( t a , t b)for:

[0258] (31)

[0259] In the formula [ Q ]for w i Q i ( t k The stability coefficient corresponding to the minimum value Q i ( t k The corresponding theoretical limit value is determined by the safety factor specified in the retaining wall design code.

[0260] Based on the aforementioned dynamic safety margin index I ( t a , t b A safety assessment should be conducted on the retaining wall during its service life. I ( t a , t b )∈( I thre ( t a , t b ),1), indicates that the stability of the retaining wall exceeds the safety requirements of the specifications, the safety margin of the retaining wall is very sufficient, and routine monitoring is sufficient; when I ( t a , t b )∈(0, I thre ( t a , t b This indicates that the stability of the retaining wall is less than the safety requirements specified in the code, the safety margin is insufficient, and reinforcement or repair is required.

[0261] S5 In the embodiment, such as Figure 6 :

[0262] t a and t b These refer to the left and right boundary time nodes of the selected rolling time window for evaluation; points A and points B Each refers to t = t a andt = t b and M total ( t ) = M total ( t The intersection of 0); point C and points D Each refers to t = t a and t = t b and M total ( t The intersection of ) points; E and points F Each refers to t = t a and t = t b and M total ( t The intersection of () = 0; rectangle ABEF area S ABEF This represents the baseline margin area for that period. A base ( t a , t b );quadrilateral CDEF area S CDEF This indicates the margin remaining area for that period. A remain ( t a , t b ).

[0263] After dividing the continuous monitoring period of the retaining wall into equal time intervals, the total safety margin function obtained in S4 is... M total ( t Integrate the results. Figure 6 medium rectangle ABEF area S ABEF That is, the baseline margin area. A base ( t a , t b ), Figure 6Middle quadrilateral CDEF area S CDEF That is, the margin remaining area. A remain ( t a , t b ).

[0264] Benchmark margin area A base ( t a , t b )for:

[0265] (28)

[0266] Marginal Remaining Area A remain ( t a , t b )for:

[0267] (29)

[0268] Dynamic safety margin index I ( t a , t b )for:

[0269] (30)

[0270] Dynamic safety margin threshold I thre ( t a , t b Used to determine dynamic safety margin indicators I ( t a , t b Does it meet the safety requirements? I thre ( t a , t b )for:

[0271] (31)

[0272] In the formula [ Q [1] represents the stability coefficients of each item. Qi ( t k The minimum value among the corresponding theoretical limit values ​​is determined by the safety factor specified in the retaining wall design code.

[0273] Based on the aforementioned dynamic safety margin index I ( t a , t b A safety assessment should be conducted on the retaining wall during its service life. I ( t a , t b )∈( I thre ( t a , t b ),1), indicates that the stability of the retaining wall exceeds the safety requirements of the specifications, the safety margin of the retaining wall is very sufficient, and routine monitoring is sufficient; when I ( t a , t b )∈(0, I thre ( t a , t b This indicates that the stability of the retaining wall is less than the safety requirements specified in the code, the safety margin is insufficient, and reinforcement or repair is required.

[0274] This invention addresses the gradual performance degradation of retaining walls during their service life due to material aging, environmental erosion, and load fluctuations, as well as the engineering need for dynamic tracking and quantitative assessment of life-cycle safety margins. It constructs a safety margin assessment method that integrates physical mechanisms and neural networks. This method takes the time-discretization and parameter vectorization of service-life monitoring parameters as input. It maps four stability coefficients with practical physical meaning—anti-sliding, anti-overturning, shear, and foundation bearing capacity—to a safety margin benchmark function. A physical constraint learning subnetwork embeds physical constraints to ensure the physical consistency of the assessment results. An LSTM layer in the performance degradation evolution subnetwork captures the temporal dependence and cumulative effect of parameter degradation. A total safety margin function is generated by fusing the benchmark value and loss amount using a superposition method. Finally, a dynamic safety margin index is established to achieve quantitative determination of the safety status at any time period, providing an interpretable, high-precision, and dynamically updated engineering path for the life-cycle safety assessment of retaining walls.

[0275] Compared with existing technologies, this invention breaks through the traditional evaluation paradigm of single physical modeling or pure data fitting, while possessing both physical interpretability and data-driven accuracy. On the one hand, it uses a stable physical mechanism to constrain the network learning process, ensuring that the evaluation results are consistent with engineering criteria. On the other hand, through the loss function optimization mechanism of a dual-sub-network and the temporal modeling capability of LSTM, real-time monitoring information is transformed into adaptive correction of physical model deviations and precise capture of performance degradation patterns, significantly improving the evaluation stability and generalization ability under conditions of multi-factor coupling and dynamic parameter changes. This method can quantitatively characterize the evolution of safety margin at different service stages, providing more scientific and intelligent technical support and calculation methods for determining the timing of retaining wall maintenance and reinforcement and assessing risk levels.

Claims

1. A dynamic assessment method for the safety margin of retaining walls driven by both physical and data principles, characterized in that, Includes the following steps: S1, input the real-time monitoring parameters during the service period, complete the time discretization and parameter vectorization, calculate the stability coefficients of the retaining wall based on physical constraints, and map them to the safety margin benchmark function; S2 constructs a physical constraint learning sub-network architecture consisting of an "input layer - hidden layer - physical constraint layer - output layer," and uses a physical regularization loss function. L total Guide network parameter optimization and establish a "parameter vector" that meets physical consistency requirements. x k "To the optimized safety margin benchmark function" M phys ( t k Mapping between ")"; S3 constructs a performance degradation evolution sub-network architecture of "input layer - LSTM layer - fully connected layer - output layer", and uses a comprehensive loss function. L decay Guide the network to optimize parameters and establish a "parameter vector". x k Compared with the optimized safety margin benchmark value M phys ( t k From ")" to "safety margin loss Δ" M ( t k The mapping between ")" outputs an effective safety margin loss function Δ. M ( t k ); S4, using the superposition method to fuse the optimized safety margin benchmark function. M phys ( t k ) and safety margin loss function Δ M ( t k This generates a continuous total safety margin function that combines physical consistency and realistic attenuation. M total ( t k The curve of change; S5, based on the total safety margin function M total ( t Calculate the baseline margin area A base ( t a , t b ) and margin remaining area A remain ( t a , t b Establish corresponding dynamic safety margin indicators. I ( t a , t b By comparing dynamic safety margin indicators I ( t a , t b ) and dynamic safety margin threshold I thre ( t a , t b This enables a dynamic and precise quantitative assessment of the safety margin of retaining walls during their service life.

2. The method for dynamic evaluation of the safety margin of retaining walls driven by both physical and data approaches as described in claim 1, characterized in that, S1 represents the input of real-time monitoring parameters of the retaining wall during its service life. These parameters are then discretized over time and vectorized to obtain the time nodes of the service process. t k corresponding parameter vector x k Used to calculate the anti-sliding stability coefficient of the retaining wall at each time step. Q 1( t k ), overturning stability coefficient Q 2( t k ), shear stability coefficient Q 3( t k and foundation bearing capacity stability coefficient Q 4( t k And map each stability coefficient to a safety margin benchmark function at each time step. M ( t k ).

3. The method for dynamic evaluation of the safety margin of retaining walls driven by both physical and data approaches as described in claim 2, characterized in that, S1, the algorithm process is as follows: S1.1 A strategy of dividing the continuous monitoring period of the retaining wall during its service life into equal time intervals is adopted. k Time nodes at each time step t k satisfy: （1） In the formula t 0 represents the initial service life time point; Δ t The discrete time step; k From 1 to n Natural numbers between; S1.2 The service life monitoring parameters from S1.1 above are vectorized to obtain the parameter vector of the service process. x k satisfy: （2） In the formula G Due to the weight of the retaining wall itself, s G ( t k ) is the first k The distance from the toe of the retaining wall to the self-weight of each time step node. E az ( t k ) is the first k Vertical component of earth pressure corresponding to each time step node s az ( t k ) is the first k The distance from the vertical component of the earth pressure at each time step node to the wall toe. E ax ( t k ) is the first k The horizontal component of earth pressure corresponding to each time step node. s ax ( t k ) is the first k The distance from the toe of the wall to the horizontal component of the earth pressure corresponding to each time step node. μ ( t k ) is the first k The coefficient of friction of the retaining wall base at each time step node. V ( t k ) is the first k The retaining wall at each time step node experiences the maximum shear force. f a ( t k ) is the first k Bearing capacity of the retaining wall foundation at each time step node p ( t k ) is the first k The retaining wall bottom pressure corresponding to each time step node; From parameter vector x k The stability coefficients of the retaining wall at each time step can be calculated, wherein the anti-sliding stability coefficient is... Q 1( t k ), overturning stability coefficient Q 2( t k ), shear stability coefficient Q 3( t k and foundation bearing capacity stability coefficient Q 4( t k The calculations are as follows: （3） （4） （5） In the formula V ( t 0) represents the maximum shear force experienced by the retaining wall at the initial time point of its service life; （6） The safety margin benchmark function at each time step M ( t k The allowable value function for the stability coefficients of the retaining wall, taking into account all factors, is calculated as follows: （7） In the formula M ( t k ) is the first k The safety margin benchmark function of the retaining wall corresponding to each time step node; Q i ( t k ) represents the stability coefficients of each item. i = 1, 2, 3, 4; w i For each stability coefficient Q i ( t k The corresponding weighting coefficients are determined by expert evaluation and engineering experience.

4. The method for dynamic evaluation of the safety margin of retaining walls driven by both physical and data approaches as described in claim 3, characterized in that, Furthermore, in S1.1, the monitoring parameters for the retaining wall during its service life include the retaining wall's own weight. G and its distance from the toe of the wall s G Vertical component of earth pressure E az and its distance from the toe of the wall s az Horizontal component of earth pressure E ax and its distance from the toe of the wall s ax Coefficient of friction of retaining wall base μ The retaining wall is subjected to the maximum shear force V Bearing capacity of the foundation at the base of the retaining wall f a and the pressure at the base of the retaining wall p ; The service life of the retaining wall is divided into time steps, i.e., time discretization, and the service life is divided into equal parts. n The time step, of which the first k The time node corresponding to each time step t k satisfy: （1） In the formula t k For the first k The time node corresponding to each time step. k From 1 to n Natural numbers between; t 0 represents the initial time point of the service life; Δt The time interval after dividing the service period into equal parts; Real-time monitoring of parameters that change over time yields the first... k The distance from the toe of the retaining wall to the self-weight of each time step node. s G ( t k Vertical component of earth pressure E az ( t k ) and its distance from the toe of the wall s az ( t k ), horizontal component of earth pressure E ax ( t k ) and its distance from the toe of the wall s ax ( t k ), coefficient of friction of retaining wall base μ ( t k The retaining wall experiences maximum shear force. V ( t k ), Bearing capacity of the foundation under the retaining wall f a ( t k ) and the pressure at the bottom of the retaining wall p ( t k ).

5. The method for dynamic evaluation of the safety margin of retaining walls driven by both physical and data approaches as described in claim 1, characterized in that, S2 is: The subnetwork is learned through physical constraints, and the input parameter vector x k After training the subnetwork, an optimized safety margin benchmark function can be output. M phys ( t k Then, the loss function is adjusted using physical regularization. L total Guide network parameter optimization, if L total Less than or equal to the preset loss threshold δ 1. Determine the optimized safety margin benchmark function. M phys ( t k If the physical consistency requirements are met, the process can proceed to the next step; otherwise, if... L total Greater than the preset loss threshold δ If so, it is necessary to return to S2 to adjust the physical constraint layer weight matrix. W phys The weight matrix of the output layer of the physical constraint learning subnetwork W out and regularization weights λ Retrain until the loss meets the threshold requirement; then apply physical regularization to the loss function. L total The physical constraint learning subnetwork, after guiding network parameter optimization, can achieve the transformation from "parameter vector" x k "To the optimized safety margin benchmark function" M phys ( t k The reliable mapping of ")" ensures that the output results are both physically interpretable and engineering-applicable.

6. The method for dynamic evaluation of the safety margin of retaining walls driven by both physical and data approaches as described in claim 5, characterized in that, S2, the specific method and process are as follows: The physical constraint learning subnetwork described in S2.1 consists of an input layer, a hidden layer, a physical constraint layer, and an output layer, with the input parameter vector... x k After training the subnetwork, an optimized safety margin benchmark function can be output. M phys ( t k ); The number of neurons and parameter vectors in the input layer of the physical constraint learning subnetwork x k The dimensions are consistent, and it is used to receive and input the parameter vector preprocessed by S1. x k ; The hidden layer of the physical constraint learning subnetwork has a two-layer fully connected structure. The first layer contains 48 neurons, and the second layer contains 32 neurons. Both layers use the ReLU activation function, which can effectively alleviate the gradient vanishing problem. The hidden layer extracts features from the input data layer by layer through weighted summation of the two layers of neurons and activation function transformation, and outputs the corresponding feature coefficient vector. H ; The physical constraint learning subnetwork has a physical constraint layer with 4 neurons and no activation function, used to directly output the stability coefficient vector derived in reverse derivation. Q total ( t k The physical constraint layer explicitly embeds mechanical constraints into the network architecture, forcing the network output to always derive from a reasonable mapping of mechanical coefficients, thus avoiding fitting distortions that deviate from the physical essence; among them, the stability coefficient vector Q total ( t k )for: （8） In the formula 、 、 and The first and second layers are derived from the physical constraint layers of the physical constraint learning subnetwork, respectively. k Mapped values ​​of the anti-sliding stability coefficient, anti-overturning stability coefficient, anti-shear stability coefficient, and foundation bearing capacity stability coefficient of the retaining wall at each time step; The stability coefficient vector Q total ( t k The calculation method for ) is as follows: （9） In the formula Q total ( t k The vector of stability coefficients is derived from the physical constraint layer of the physical constraint learning subnetwork. W phys This is the weight matrix of the physical constraint layer. H The feature coefficient vector output by the hidden layer. b phys This is the bias vector for the physical constraint layer; The output layer of the physical constraint learning subnetwork contains one neuron and uses the Sigmoid activation function to output the stability coefficient vector of the physical constraint layer. Q total ( t k Mapped to the optimized safety margin benchmark function M phys ( t k This serves as a physical compliance reference benchmark for subsequent performance degradation evolution subnetworks; The optimized safety margin benchmark function M phys ( t k The expression is: （10） In the formula M phys ( t k ) is the first k The optimized safety margin baseline value at each time step; σ () is the Sigmoid activation function, which maps the output value to the interval [0,1]. W out To learn the weight matrix of the output layer of the physical constraint subnetwork, b out This is the bias vector for the output layer; The physical constraint learning subnetwork described in S2.2 uses a physical regularization loss function. L total Guide network parameter optimization; specifically, if L total Less than or equal to the preset loss threshold δ 1. Determine the optimized safety margin benchmark function. M phys ( t k If the physical consistency requirements are met, the process can proceed to the next step; otherwise, if... L total Greater than the preset loss threshold δ If so, it is necessary to return to S2 to adjust the physical constraint layer weight matrix. W phys The weight matrix of the output layer of the physical constraint learning subnetwork W out and regularization weights λ Retrain until the loss meets the threshold requirement; The physical regularization loss function L total Loss function fitted from data L MSE With physical constraint loss function L phys The weighted summation is used to ensure that the mapping relationship in network learning always closely adheres to the essence of mechanics; The physical regularization loss function L total The expression is: （11） In the formula L MSE To fit the loss function to the data, L phys The physical constraint loss function, λ These are regularization weights used to balance the accuracy of data fitting with the strength of physical consistency constraints. The data fitting loss function L MSE The expression is: （12） In the formula n The total number of time steps. M phys ( t k ) is the first k The optimized safety margin baseline value at each time step M ( t k ) is the first digit calculated using physical formulas in S1. k Safety margin benchmark function at each time step; The physical constraint loss function L phys The expression is: （13） In the formula i The value range is 1 to 4. The first physical constraint layer is derived from the physical constraint learning subnetwork by back-reaming. k Mapped values ​​of various stability coefficients of the retaining wall at each time step Q i ( t k ) is the first digit calculated using physical formulas in S1. k The stability coefficients of the retaining wall at each time step; Furthermore, when L total ≤ δ When 1 is reached, it indicates the optimized safety margin benchmark function. M phys ( t k If the physical consistency requirements are met, the process can proceed to the next step; when L total > δ At step 1, it is necessary to return to S2 to adjust the physical constraint layer weight matrix. W phys The weight matrix of the output layer of the physical constraint learning subnetwork W out and regularization weights λ After adjustments, restart network training until the requirements are met. L total ≤ δ Requirements 1.

7. The method for dynamic evaluation of the safety margin of retaining walls driven by both physical and data approaches as described in claim 1, characterized in that, In S3: The performance degradation evolution subnetwork is established by creating a "parameter vector". x k Compared with the optimized safety margin benchmark value M phys ( t k From ")" to "safety margin loss Δ" M ( t k The mapping between ")" is the core objective; The core components of this network include an input layer, an LSTM layer, a fully connected layer, and an output layer. The input "parameter vector" x k and optimized safety margin benchmark value M phys ( t k After training with the performance degradation evolution subnetwork, the "safety margin loss Δ" can be output. M ( t k )”; Next, through the comprehensive loss function L decay Guide the network to optimize parameters, if L decay Less than or equal to the preset loss threshold δ 2 and Δ M ( t k ) ≤ M phys ( t k Then the output safety margin loss Δ is determined. M ( t k If the condition is met, proceed to the next step; otherwise, if not, return to S3 to adjust the weight matrix of the fully connected layer. W fc Weight matrix of the output layer of the performance degradation evolution subnetwork W decay Physical constraint weights μ and nonnegative constraint weights ε Retrain the subnetwork until the loss meets the threshold requirement; By comprehensive loss function L decay The performance degradation evolution subnetwork after guiding network parameter optimization can achieve the transformation from "parameter vector" x k Compared with the optimized safety margin benchmark value M phys ( t k From ")" to "safety margin loss Δ" M ( t k A reliable mapping between “)” ensures that the output results are consistent with the physical attenuation law.

8. The method for dynamic evaluation of the safety margin of retaining walls driven by both physical and data approaches as described in claim 7, characterized in that, S3 specifically refers to: The performance degradation evolution subnetwork consists of an input layer, an LSTM layer, a fully connected layer, and an output layer, with the input being a "parameter vector". x k and the optimized safety margin benchmark value M phys ( t k After training through a performance degradation evolution subnetwork, the "safety margin loss Δ" is output. M ( t k )”; The input layer neurons of the performance degradation evolution subnetwork have the number of "physical baseline dimension + parameter vector dimension", and are used to receive and input two types of data: the optimized safety margin baseline value output by S2. M phys ( t k and the parameter vector after S1 preprocessing x k ; The LSTM layer of the performance degradation evolution subnetwork has a single-layer structure with 64 hidden units. It uses a combination of tanh and sigmoid activation functions to capture the temporal dependence and cumulative effect of parameter degradation. Its core state update formula satisfies: （14） （15） （16） （17） （18） （19） In the formula i k For input gate, f k For the Gate of Oblivion g k In the candidate state of cells, o k For output gate, c k In cellular state, h k The hidden state vector; W ii , W if , W ig , W io These are the weight matrices from the input layer to each gate. W hi , W hf , W hg , W ho These are the weight matrices from the hidden layer to each gate; b ii , b if , b ig , b io These are the bias vectors for each gate; x k The concatenated feature vectors from the input layer. h k-1 , c k-1 These are the hidden state vector and the cell state vector from the previous time step, respectively. σ ( ) represents the Sigmoid activation function, and ⊙ represents the Hadamard product operation; this layer selectively memorizes the long-term trend of parameter decay through a gating mechanism to avoid the loss of temporal information; The fully connected layer of the performance degradation evolution subnetwork has a two-layer structure, with 32 neurons in each layer. The ReLU activation function is used to alleviate the gradient vanishing problem. The fully connected layer is used to map the temporal features extracted by the LSTM layer into nonlinear loss features, the expression of which is: （20） In the formula, F decay The loss characteristics are those output by the fully connected layer. W fc This is the weight matrix of the fully connected layer. h k For the hidden state vector, b fc This is the bias vector of the fully connected layer; The output layer of the performance degradation evolution subnetwork contains one neuron and uses the Softplus activation function to map the loss features to a non-negative safety margin loss Δ. M ( t k ); The safety margin loss Δ M ( t k The expression for ) is: （21） In the formula, Δ M ( t k ) is the first k The safety margin loss at each time step, with a value ranging from [0, ... M phys ( t k )]; The Softplus activation function ensures that the output is always non-negative; W decay The weight matrix of the output layer of the performance degradation evolution subnetwork. b decay This is the bias vector for the output layer; The performance degradation evolution subnetwork uses a comprehensive loss function. L decay Guide the network to optimize parameters, if L decay Less than or equal to the preset loss threshold δ 2 and Δ M ( t k ) ≤ M phys ( t k Then the output safety margin loss Δ is determined. M ( t k If the condition is met, proceed to the next step; otherwise, if not, return to S3 to adjust the weight matrix of the fully connected layer. W fc Weight matrix of the output layer of the performance degradation evolution subnetwork W decay Physical constraint weights μ and nonnegative constraint weights ε Retrain the subnetwork until the loss meets the threshold requirement; The comprehensive loss function L decay Loss function fitted by loss L MSE-Δ Physical consistency constraint loss function L consist Non-negative constraint loss L nonneg The function is weighted and summed to ensure that the loss is consistent with the physical decay law and is non-negative. The comprehensive loss function L decay The expression is: （22） In the formula L MSE-Δ The loss function is the loss fit. L consist The loss function is the physical consistency constraint. L nonneg The loss function is a non-negativity constraint. μ For physical constraint weights; ε These are non-negative constraint weights, used to balance fitting accuracy and constraint strength; The loss function of the loss fitting L MSE-Δ The expression is: （23） In the formula n Δ is the total number of time steps. M ( t k ) is the first k The amount of safety margin loss at each time step M phys ( t k ) is the first k The optimized safety margin baseline value at each time step M meas ( t k ) is the first k The safety margin obtained from on-site measurements at each time step; The physical consistency constraint loss function L consist The expression is: （24） In the formula i The value range of Δ is 1~4. M ( t k ) is the first k The amount of safety margin loss at each time step M phys ( t k ) is the first k The optimized safety margin baseline value at each time step The first physical constraint layer is derived from the physical constraint learning subnetwork by back-reaming. k Mapped values ​​of various stability coefficients of the retaining wall at each time step Q i ( t k ) is the first digit calculated using physical formulas in S1. k The stability coefficients of the retaining wall at each time step; The nonnegativity constraint loss function L nonneg The expression is: （24） In the formula, max{0,-Δ M ( t k Ensure that a penalty loss is incurred when the network output is negative, forcing Δ M ( t k ) ≥ 0; The preset loss threshold δ The range of values ​​for 2 needs to be such that it can filter out invalid training results without being too strict and causing the training to fail to converge. It can be determined based on the engineering monitoring accuracy. Furthermore, when the comprehensive loss function L decay ≤ δ 2 and Δ M ( t k ) ≤ M phys ( t k When ), it indicates the loss of safety margin Δ in the output. M ( t k If the condition is met, proceed to the next step; otherwise, if not, return to S3 to adjust the weight matrix of the fully connected layer. W fc Weight matrix of the output layer of the performance degradation evolution subnetwork W decay Physical constraint weights μ and nonnegative constraint weights ε After adjustment, restart the subnetwork for training until the loss meets the threshold requirement.

9. The method for dynamic evaluation of the safety margin of retaining walls driven by both physical and data approaches as described in claim 1, characterized in that, In S4, the total safety margin M total ( t k The optimized safety margin baseline value output by S2 M phys ( t k The safety margin loss Δ of S3 output M ( t k The operation essentially involves superimposing the attenuation loss as a reverse correction term onto the physical reference margin, thereby achieving the fusion of the reference term and the correction term. The total safety margin M total ( t k The expression for ) is: （26） The total safety margin value at discrete time steps is integrated and smoothed to obtain the total safety margin function over the service life of the retaining wall. M total ( t )for: （27） In the formula n This represents the total number of time steps during the service period. k For [1, n Natural numbers between ] M total (t) is the total safety margin function over the entire service life, used for subsequent safety margin index calculations.

10. The method for dynamic evaluation of the safety margin of retaining walls driven by both physical and data approaches as described in claim 1, characterized in that, In S5, based on the total safety margin function M total ( t Calculate the baseline margin area A base ( t a , t b ) and margin remaining area A remain ( t a , t b Establish corresponding dynamic safety margin indicators. I ( t a , t b By comparing dynamic safety margin indicators I ( t a , t b ) and dynamic safety margin threshold I thre ( t a , t b This enables dynamic assessment of the safety margin of retaining walls during their service life. t a and t b These refer to the left and right boundary time nodes of the selected rolling time window for evaluation, respectively. The benchmark margin area A base ( t a , t b This refers to the theoretical remaining safety margin of a retaining wall during a certain period of its service life, where the remaining area of ​​the margin is... A remain ( t a , t b This refers to the actual remaining safety margin during that period. Based on the total safety margin function described in S4 M total ( t The reference margin area A base ( t a , t b ) and margin remaining area A remain ( t a , t b The calculations are as follows: （28） In the formula t a and t b These are the left and right boundary time nodes of the selected rolling time window for evaluation; M total ( t 0) is t The total safety margin function value at time 0; （29） The dynamic safety margin index I ( t a , t b ) is used to evaluate arbitrary rolling time windows. t a , t b The corresponding safety margin is calculated as follows: （30） In the formula, the dynamic safety margin index I ( t a , t b )∈(0,1), the larger the value, the more sufficient the safety margin of the retaining wall; The dynamic safety margin index threshold I thre ( t a , t b Used to determine dynamic safety margin indicators I ( t a , t b Does it meet the safety requirements? I thre ( t a , t b The calculation formula is as follows: （31） In the formula [ Q ] represents the stability coefficients of each term in formula (7). Q i ( t k The minimum value among the corresponding theoretical limits is determined by the safety factor specified in the retaining wall design code. M total ( t 0) is t The total safety margin function value at time 0; Based on the "Dynamic Safety Margin Index" I ( t a , t b A safety assessment of the retaining wall during its service life should be conducted. I ( t a , t b )∈( I thre ( t a , t b ),1), indicates that the stability of the retaining wall exceeds the safety requirements of the specifications, the safety margin of the retaining wall is very sufficient, and routine monitoring is sufficient; when I ( t a , t b )∈(0, I thre ( t a , t b This indicates that the stability of the retaining wall is less than the safety requirements specified in the code, the safety margin is insufficient, and reinforcement or repair is required.

Images