CN-121723571-B - Physical-data dual-drive retaining wall safety margin dynamic evaluation method

Abstract

Aiming at the problem that the traditional method cannot adapt to the service-period performance attenuation of the retaining wall, the physical-data dual-drive retaining wall safety margin dynamic evaluation method is disclosed, wherein 1, service-period monitoring parameters are input, each stability coefficient of the retaining wall is calculated and mapped into a safety margin reference function, 2, a physical constraint learning sub-network is constructed, network parameter optimization is guided through a physical regularization loss function, an optimized safety margin reference value meeting the physical consistency requirement is output, 3, a performance attenuation evolution sub-network is constructed, the time sequence dependence of LSTM capturing parameter attenuation is adopted, network parameter optimization is guided through a comprehensive loss function, the safety margin loss is output, 4, the output results of all sub-networks are fused, a total safety margin function curve of the retaining wall in the service period is built, and 5, dynamic safety margin indexes are built and the safety margin of the retaining wall in the service period is dynamically evaluated. And a quantitative basis is provided for retaining wall operation and maintenance reinforcement decision and structural safety risk assessment.

Inventors

• SHAN YAO
• Dong Yacheng
• LIU DONG
• ZHOU TONG
• WANG GUANKAI

Assignees

• 同济大学

Dates

Publication Date

20260512

Application Date

20260227

Claims (10)

1. 1. The physical-data double-driven retaining wall safety margin dynamic evaluation method is characterized by comprising the following steps of: s1, inputting real-time monitoring parameters in a service period, completing time dispersion and parameter vectorization, calculating each stability coefficient of a retaining wall based on physical constraint, and mapping the stability coefficients into a safety margin reference function; S2, constructing a physical constraint learning sub-network architecture of an input layer, a hidden layer, a physical constraint layer and an output layer, guiding network parameter optimization through a physical regularization loss function L total , and establishing a mapping between a parameter vector x k to an optimized safety margin reference function M phys (t k under the condition of meeting the physical consistency requirement; S3, constructing a performance attenuation evolution sub-network architecture of an input layer-LSTM layer-full connection layer-output layer, guiding a network to perform parameter optimization through a comprehensive loss function L decay , establishing a mapping between a parameter vector x k and an optimized safety margin reference value M phys (t k to a safety margin loss amount delta M (t k ), and outputting an effective safety margin loss function delta M (t k ); S4, fusing the optimized safety margin reference function M phys (t k ) and the safety margin loss function delta M (t k ) by adopting an superposition method to generate a continuous change curve of a total safety margin function M total (t k ) with physical consistency and attenuation authenticity; S5, calculating a reference margin area A base (t a ,t b and a margin residual area A remain (t a ,t b according to the total safety margin function M total (t), establishing a corresponding dynamic safety margin index I (t a ,t b ), and comparing the dynamic safety margin index I (t a ,t b ) with a dynamic safety margin index threshold I thre (t a ,t b to realize dynamic and accurate quantitative evaluation of the safety margin during service of the retaining wall.
2. 2. The method for dynamically evaluating the safety margin of a physical-data dual-drive retaining wall according to claim 1, wherein S1 is that real-time monitoring parameters of the retaining wall during service are input, time dispersion and parameter vectorization are carried out on the parameters to obtain parameter vectors x k corresponding to time nodes t k in the service process, the parameter vectors are used for calculating an anti-slip stability coefficient Q 1 (t k ), an anti-capsizing stability coefficient Q 2 (t k ), a shearing stability coefficient Q 3 (t k ) and a foundation bearing capacity stability coefficient Q 4 (t k of the retaining wall in each time step, and each stability coefficient is mapped into a safety margin reference function M (t k ) in each time step.
3. 3. The method for dynamically evaluating the safety margin of a physical-data double-driven retaining wall according to claim 2, wherein the algorithm process is as follows: S1.1, an equal time interval division strategy is adopted for a continuous monitoring period of the service period of the retaining wall, and a time node t k of a kth time step meets the following conditions: (1) Wherein t 0 is a service period initial time node, delta t is a discrete time step length, and k is a natural number between 1 and n; S1.2, carrying out parameter vectorization on the monitoring parameters of the service period of the S1.1 to obtain a parameter vector x k in the service process, wherein the parameter vector x k meets the following conditions: (2) Wherein G is self gravity of the retaining wall, s G (t k ) is the distance from the dead weight of the retaining wall corresponding to the kth time step node to the toe of the wall, E az (t k ) is the vertical component of the soil pressure corresponding to the kth time step node, s az (t k ) is the distance from the vertical component of the soil pressure corresponding to the kth time step node to the toe of the wall, E ax (t k ) is the horizontal component of the soil pressure corresponding to the kth time step node, s ax (t k ) is the distance from the horizontal component of the soil pressure corresponding to the kth time step node to the toe of the wall, mu (t k ) is the friction coefficient of the retaining wall base corresponding to the kth time step node, V (t k ) is the maximum shearing force applied to the retaining wall corresponding to the kth time step node, f a (t k ) is the bearing capacity of the retaining wall base corresponding to the kth time step node, and p (t k ) is the retaining wall base pressure corresponding to the kth time step node; Each stability coefficient of the retaining wall under each time step can be calculated by a parameter vector x k , wherein the anti-slip stability coefficient Q 1 (t k ), the anti-overturning stability coefficient Q 2 (t k ), the shear stability coefficient Q 3 (t k ) and the foundation bearing capacity stability coefficient Q 4 (t k ) are calculated as follows: (3) (4) (5) Wherein V (t 0 ) is the maximum shearing force applied to the retaining wall corresponding to the initial time node of the service period; (6) The safety margin reference function M (t k ) under each time step is an allowable value function comprehensively considering each stability coefficient of the retaining wall, and is calculated as follows: (7) Wherein M (t k ) is a retaining wall safety margin reference function corresponding to a kth time step node, Q i (t k ) is a weight coefficient corresponding to each stable coefficient, i=1, 2, 3, 4;w i is each stable coefficient Q i (t k ), and the weight coefficient is determined by expert evaluation and engineering experience.
4. 4. A physical-data dual-drive retaining wall safety margin dynamic assessment method according to claim 3, further characterized in that in S1.1, the retaining wall monitoring parameters during service include retaining wall self gravity G and distance from the toe of the retaining wall S G , soil pressure vertical component E az and distance from the toe of the retaining wall S az , soil pressure horizontal component E ax and distance from the toe of the retaining wall S ax , retaining wall base friction coefficient μ, retaining wall received maximum shear force V, retaining wall base bearing force f a and retaining wall base pressure p; performing time step division and time dispersion on the service period of the retaining wall, and equally dividing the service period into n time steps, wherein a time node t k corresponding to the kth time step satisfies the following conditions: (1) Wherein t k is a time node corresponding to the kth time step, k is a natural number between 1 and n, t 0 is a service period initial time node, and Δt is a time interval after service period equally dividing; The monitoring parameters changing along with time are monitored in real time, and the distance s G (t k between the dead weight of the retaining wall and the toe of the wall, the vertical component E az (t k ) of the soil pressure and the distance s az (t k between the dead weight and the toe of the wall, the horizontal component E ax (t k ) of the soil pressure and the distance s ax (t k between the dead weight and the toe of the wall, the friction coefficient mu (t k ) of the retaining wall base, the maximum shearing force V (t k ) of the retaining wall, the bearing force f a (t k of the foundation of the retaining wall and the pressure p (t k ) of the foundation of the retaining wall are obtained.
5. 5. The method for dynamically evaluating the safety margin of a physical-data double-driven retaining wall according to claim 1, wherein S2 is: The optimized safety margin reference function M phys (t k can be output through the physical constraint learning sub-network after the input parameter vector x k is trained through the sub-network, then the network parameter optimization is guided through the physical regularization loss function L total , if L total is smaller than or equal to a preset loss threshold delta 1 , the optimized safety margin reference function M phys (t k ) is determined to meet the physical consistency requirement, a subsequent flow can be entered, otherwise, if L total is larger than the preset loss threshold delta 1 , the physical constraint layer weight matrix W phys , the physical constraint learning sub-network output layer weight matrix W out and regularized weight lambda need to be returned S2 again, the physical constraint learning sub-network is retrained until the loss meets the threshold requirement, and the physical constraint learning sub-network after the network parameter optimization is guided through the physical regularization loss function L total , the reliable mapping from the parameter vector x k to the optimized safety margin reference function M phys (t k is ensured, and the output result has physical interpretability and engineering practicability.
6. 6. The method for dynamically evaluating the safety margin of a physical-data double-driven retaining wall according to claim 5, wherein the specific method comprises the following steps: s2.1, the physical constraint learning sub-network consists of an input layer, a hidden layer, a physical constraint layer and an output layer, and an input parameter vector x k can output an optimized safety margin reference function M phys (t k after being trained by the sub-network; the number of neurons of the input layer of the physical constraint learning sub-network is consistent with the dimension of the parameter vector x k , and the number of neurons is used for receiving and inputting the parameter vector x k after S1 preprocessing; The hidden layer of the physical constraint learning sub-network is provided with a 2-layer full-connection structure, the first layer comprises 48 neurons, the second layer comprises 32 neurons, and both the two layers adopt ReLU activation functions capable of effectively relieving the gradient vanishing problem; The physical constraint layer of the physical constraint learning sub-network comprises 4 neurons and no activation function, and is used for directly outputting a stability coefficient vector Q total (t k which is reversely deduced, the physical constraint layer forces a network output result to always originate from reasonable mapping of mechanical coefficients by explicitly embedding the mechanical constraint into a network architecture, so that fitting distortion which is separated from physical essence is avoided, and the stability coefficient vector Q total (t k ) is as follows: (8) In the middle of 、 、 And The anti-slip stability coefficient mapping value, the anti-capsizing stability coefficient mapping value, the shear stability coefficient mapping value and the foundation bearing capacity stability coefficient mapping value of the retaining wall at the kth time step are respectively obtained by reverse deduction of a physical constraint layer of a physical constraint learning sub-network; the calculation method of the stability coefficient vector Q total (t k ) comprises the following steps: (9) Wherein Q total (t k ) is a stability coefficient vector reversely deduced by a physical constraint layer of a physical constraint learning sub-network, W phys is a weight matrix of the physical constraint layer, H is a characteristic coefficient vector output by a hidden layer, and b phys is a bias vector of the physical constraint layer; The output layer of the physical constraint learning sub-network comprises 1 neuron, a Sigmoid activation function is adopted, and a stability coefficient vector Q total (t k ) output by the physical constraint layer is mapped into an optimized safety margin reference function M phys (t k ) for providing a physical compliance reference for the subsequent performance attenuation evolution sub-network; the optimized safety margin reference function M phys (t k ) has the expression: (10) M phys (t k ) is a safety margin reference value optimized in the kth time step, sigma () is a Sigmoid activation function, and can map an output value to a [0,1] interval, W out is a weight matrix of an output layer of a physical constraint learning sub-network, and b out is a bias vector of the output layer; S2.2, the physical constraint learning sub-network guides network parameter optimization through a physical regularization loss function L total , specifically, if L total is smaller than or equal to a preset loss threshold delta 1 , the optimized safety margin reference function M phys (t k ) is determined to meet the physical consistency requirement, a subsequent procedure can be entered, otherwise, if L total is larger than the preset loss threshold delta 1 , the physical constraint layer weight matrix W phys , the physical constraint learning sub-network output layer weight matrix W out and regularization weight lambda are required to be returned S2 again, and the training is carried out again until the loss meets the threshold requirement; The physical regularization loss function L total is formed by weighting and summing a data fitting loss function L MSE and a physical constraint loss function L phys , so as to ensure that the mapping relation of network learning always buckles the mechanical essence; The physical regularization loss function L total is expressed as: (11) Wherein L MSE is a data fitting loss function, L phys is a physical constraint loss function, lambda is a regularization weight, and the regularization weight is used for balancing data fitting precision and physical consistency constraint intensity; The data fitting loss function L MSE has the expression: (12) Wherein n is the total number of time steps, M phys (t k ) is the safety margin reference value optimized in the kth time step, and M (t k ) is the safety margin reference function in the kth time step calculated by a physical formula in S1; The physical constraint loss function L phys has the expression: (13) wherein the value range of i is 1 to 4, The mapping values of the stability coefficients of the retaining wall under the kth time step are reversely deduced by a physical constraint layer of a physical constraint learning sub-network, and Q i (t k ) is the stability coefficients of the retaining wall under the kth time step calculated by a physical formula in S1; Further, when L total ≤δ 1 , the optimized safety margin reference function M phys (t k ) meets the physical consistency requirement, the subsequent flow can be entered, when L total >δ 1 , the physical constraint layer weight matrix W phys , the physical constraint learning sub-network output layer weight matrix W out and the regularization weight lambda need to be returned to S2 again, and the network training is restarted after adjustment until the requirement of L total ≤δ 1 is met.
7. 7. The method for dynamically evaluating the safety margin of a physical-data double-driven retaining wall according to claim 1, wherein in S3: The performance attenuation evolution sub-network takes the establishment of the mapping between the parameter vector x k and the optimized safety margin reference value M phys (t k and the safety margin loss delta M (t k ) as a core target; The core component of the network comprises an input layer, an LSTM layer, a full connection layer and an output layer, wherein the input parameter vector x k and the optimized safety margin reference value M phys (t k can output the safety margin loss delta M (t k ) after being trained by a performance attenuation evolution sub-network; And then guiding the network to perform parameter optimization through a comprehensive loss function L decay , if L decay is smaller than or equal to a preset loss threshold delta 2 and delta M (t k ) ≤M phys (t k ), determining that the output safety margin loss delta M (t k ) is effective, and entering a subsequent flow, otherwise, returning to S3 to adjust a weight matrix W fc of the full-connection layer, a weight matrix W decay of the output layer of the performance attenuation evolution sub-network, a physical constraint weight mu and a non-negative constraint weight epsilon if the loss is not met, and retraining the sub-network until the loss meets the threshold requirement; The comprehensive loss function L decay is used for guiding the performance attenuation evolution sub-network after the network parameter optimization to realize the reliable mapping from the parameter vector x k and the optimized safety margin reference value M phys (t k to the safety margin loss delta M (t k ), and the output result is ensured to be consistent with the physical attenuation rule.
8. 8. The method for dynamically evaluating the safety margin of a physical-data double-driven retaining wall according to claim 7, wherein S3 is specifically: The performance attenuation evolution sub-network consists of an input layer, an LSTM layer, a full connection layer and an output layer, wherein the performance attenuation evolution sub-network is trained after a parameter vector x k and an optimized safety margin reference value M phys (t k are input, and a safety margin loss delta M (t k ) is output; the number of neurons of an input layer of the performance attenuation evolution sub-network is physical reference dimension plus parameter vector dimension, and the number of neurons of the input layer of the performance attenuation evolution sub-network is used for receiving and inputting two types of data, namely an optimized safety margin reference value M phys (t k output by S2 and a parameter vector x k preprocessed by S1; The LSTM layer of the performance attenuation evolution sub-network is provided with a 1-layer structure and comprises 64 hidden units, and a timing sequence dependence and accumulation effect of parameter attenuation are captured by adopting a tanh and Sigmoid dual-activation function combination, wherein a core state update formula of the LSTM layer comprises the following steps: (14) (15) (16) (17) (18) (19) Wherein i k is an input gate, f k is a forgetting gate, g k is a cell candidate state, o k is an output gate, c k is a cell state, h k is a hidden state vector, W ii 、W if 、W ig 、W io is a weight matrix from an input layer to each gate, W hi 、W hf 、W hg 、W ho is a weight matrix from the hidden layer to each gate, b ii 、b if 、b ig 、b io is a bias vector of each gate, x k is a characteristic vector after input layer splicing, h k-1 、c k-1 is a hidden state vector and a cell state vector of the last time step, sigma () is a Sigmoid activation function, as is a Hadamard product operation, the layer selectively memorizes a long-term trend of parameter attenuation through a gating mechanism, and loss of time sequence information is avoided; The full-connection layer of the performance attenuation evolution sub-network is provided with a 2-layer structure, each layer contains 32 neurons, the gradient disappearance problem is relieved by adopting a ReLU activation function, the full-connection layer is used for mapping the time sequence characteristics extracted by the LSTM layer into nonlinear loss characteristics, and the expression is as follows: (20) Wherein F decay is the loss characteristic of the output of the full-connection layer, W fc is the weight matrix of the full-connection layer, h k is the hidden state vector, and b fc is the bias vector of the full-connection layer; the output layer of the performance attenuation evolution sub-network comprises 1 neuron, and a Softplus activation function is adopted for mapping the loss characteristic into a non-negative safety margin loss quantity delta M (t k ); The expression of the safety margin loss amount Δm (t k ) is: (21) Wherein DeltaM (t k ) is the safety margin loss under the kth time step, the value range is [0, M phys (t k ) ], the Softplus activation function can ensure constant non-negative output, W decay is the weight matrix of the output layer of the performance attenuation evolution sub-network, and b decay is the bias vector of the output layer; The performance attenuation evolution sub-network guides the network to conduct parameter optimization through a comprehensive loss function L decay , if L decay is smaller than or equal to a preset loss threshold delta 2 and delta M (t k ) ≤M phys (t k ), the output safety margin loss delta M (t k ) is considered to be effective, a subsequent process can be entered, otherwise, if the loss is not satisfied, the sub-network is retrained until the loss meets the threshold requirement, the sub-network is retrained by returning to S3 to adjust a weight matrix W fc of the full-connection layer, a weight matrix W decay of the output layer of the performance attenuation evolution sub-network, a physical constraint weight mu and a non-negative constraint weight epsilon; The comprehensive loss function L decay is formed by weighted summation of a loss function L MSE-Δ , a physical consistency constraint loss function L consist and a non-negative constraint loss function L nonneg function of loss fitting, and is used for ensuring that the loss quantity is consistent with a physical attenuation rule and is non-negative; the expression of the comprehensive loss function L decay is as follows: (22) Wherein L MSE-Δ is a loss function of loss fitting, L consist is a physical consistency constraint loss function, L nonneg is a non-negative constraint loss function, mu is a physical constraint weight, epsilon is a non-negative constraint weight and is used for balancing fitting precision and constraint intensity; the expression of the loss function L MSE-Δ of the loss fitting is: (23) Wherein n is the total number of time steps, deltaM (t k ) is the safety margin loss amount in the kth time step, M phys (t k ) is the safety margin reference value optimized in the kth time step, and M meas (t k ) is the safety margin obtained by field actual measurement in the kth time step; The expression of the physical consistency constraint loss function L consist is: (24) Wherein the value range of i is 1-4, deltaM (t k ) is the safety margin loss under the kth time step, M phys (t k ) is the optimized safety margin reference value under the kth time step, The mapping values of the stability coefficients of the retaining wall under the kth time step are reversely deduced by a physical constraint layer of a physical constraint learning sub-network, and Q i (t k ) is the stability coefficients of the retaining wall under the kth time step calculated by a physical formula in S1; The expression of the non-negative constraint loss function L nonneg is: (24) Where max {0, - ΔM (t k ) } ensures penalty loss is generated when the network output is negative, forcing ΔM (t k ) > 0; The value range of the preset loss threshold delta 2 needs to ensure that not only invalid training results are filtered, but also training is difficult to converge due to excessive severity, and can be determined according to engineering monitoring precision; Further, when the loss function L decay ≤δ 2 and DeltaM (t k ) ≤M phys (t k ) are synthesized, the output safety margin loss DeltaM (t k ) is effective, and the subsequent flow can be entered, otherwise, if not, the method is required to return to S3 again to adjust the weight matrix W fc of the full-connection layer, the weight matrix W decay of the output layer of the performance attenuation evolution sub-network, the physical constraint weight mu and the non-negative constraint weight epsilon, and the sub-network is restarted for training after adjustment until the loss meets the threshold requirement.
9. 9. The method for dynamically evaluating the safety margin of a physical-data double-drive retaining wall according to claim 1, wherein in the step S4, the total safety margin M total (t k ) is formed by overlapping an optimized safety margin reference value M phys (t k ) output by the step S2 and a safety margin loss amount Δm (t k ) output by the step S3, and the operation essence is that attenuation loss is overlapped to the physical reference margin as an inverse correction term, so that fusion of the reference term and the correction term is realized; the expression of the total safety margin M total (t k ) is: (26) Performing time sequence integration and smoothing on the total safety margin value of the discrete time steps to obtain a total safety margin function M total (t) in the service period of the retaining wall, wherein the total safety margin function M total (t) is as follows: (27) wherein n is the total time step number of the service period, k is a natural number between [1, n ], and M total (t) is the total safety margin function in the whole service period, and is used for subsequent safety margin index calculation.
10. 10. The method for dynamically evaluating the safety margin of a retaining wall driven by two physical and data according to claim 1, wherein in S5, a reference margin area A base (t a ,t b ) and a margin remaining area A remain (t a ,t b ) are calculated according to the total safety margin function M total (t), a corresponding dynamic safety margin index I (t a ,t b ) is established, and the dynamic evaluation of the safety margin during the service period of the retaining wall is realized by comparing the dynamic safety margin index I (t a ,t b ) with a dynamic safety margin index threshold I thre (t a ,t b ), wherein t a and t b refer to the left boundary time node and the right boundary time node of a rolling time window selected for evaluation respectively; the reference margin area A base (t a ,t b ) refers to the theoretical residual safety margin in a certain period of the service period of the retaining wall, the margin residual area A remain (t a ,t b ) refers to the actual residual safety margin in the period, and according to the total safety margin function M total (t) of S4, the reference margin area A base (t a ,t b ) and the margin residual area A remain (t a ,t b ) are respectively calculated as follows: (28) Wherein t a and t b are respectively the left and right boundary time nodes of the rolling time window selected for evaluation, M total (t 0 ) is the total safety margin function value corresponding to the moment t 0 ; (29) The dynamic safety margin index I (t a ,t b ) is used for evaluating the safety margin corresponding to any rolling time window [ t a ,t b ], and is calculated as follows: (30) In the formula, a dynamic safety margin index I (t a ,t b ) epsilon (0, 1), wherein the larger the value is, the more sufficient the safety margin of the retaining wall is; The dynamic safety margin index threshold I thre (t a ,t b ) is used for judging whether the dynamic safety margin index I (t a ,t b ) meets the safety requirement, and the calculation formula of I thre (t a ,t b ) is as follows: (31) Wherein [ Q ] is the minimum value of theoretical limit values corresponding to each stability coefficient Q i (t k ) in the formula (7), the minimum value is determined by the safety coefficient specified by the design specification of the retaining wall, and M total (t 0 ) is the total safety margin function value corresponding to the moment t 0 ; And (3) carrying out safety evaluation on the retaining wall in the service period according to a dynamic safety margin index I (t a ,t b ), wherein when I (t a ,t b )∈(I thre (t a ,t b ) and 1), the stability of the retaining wall is larger than the standard safety requirement, the safety margin of the retaining wall is very sufficient, conventional monitoring is carried out, and when I (t a ,t b )∈(0,I thre (t a ,t b ), the stability of the retaining wall is smaller than the standard safety requirement, the safety margin is insufficient, and reinforcement or maintenance is needed.

Description

Physical-data dual-drive retaining wall safety margin dynamic evaluation method Technical Field The invention relates to a data processing method for prediction, which is applied to safety evaluation of geotechnical engineering and civil engineering structures, in particular to a physical-data dual-drive retaining wall safety margin dynamic evaluation method. Background The retaining wall is used as a core supporting structure in geotechnical engineering, is widely applied to engineering scenes such as roads, slopes, building foundation pits and the like, is subjected to multiple factors such as material aging, environmental erosion, load fluctuation and the like for a long time in the service period, is easy to realize progressive attenuation of performance, and directly threatens the safety margin of the structure and the operation stability of the engineering. The existing retaining wall safety margin dynamic evaluation method is mainly divided into two types, namely an analysis model based on a physical mechanism only depends on idealized hypothesis and parameter certainty, complex and changeable parameter attenuation rules in a service period are difficult to capture, deviation between the physical model and actual structure response cannot be dynamically corrected, and the other type is a statistical model based on data driving only, and due to the lack of physical mechanism constraint, problems of extrapolation distortion, poor interpretability and the like are easy to occur under the condition of limited monitoring data, and engineering reliability is difficult to guarantee. In addition, both methods can not effectively integrate physical constraint and performance attenuation evolution characteristics, and dynamic tracking and accurate quantification of safety margins in different service stages are difficult to realize. Therefore, an evaluation method integrating a physical mechanism and data driving advantages is needed, physical constraint relation is excavated based on a physical model, dynamic evaluation of safety margin of the whole life cycle of the retaining wall is realized by means of the neural network excavation performance attenuation law, and reliable support is provided for operation and maintenance reinforcement decision and safety risk evaluation. Disclosure of Invention In view of the problem that the traditional retaining wall safety margin assessment method cannot adapt to performance attenuation characteristics in a service period and deviation exists between a physical model and actual response, the invention aims to provide the physical-data dual-drive retaining wall safety margin dynamic assessment method, a physical assessment standard is defined by constructing a safety margin characterization index system, a multiparameter physical association and performance attenuation evolution law are automatically learned by depending on a physical constraint learning sub-network and a performance attenuation evolution sub-network coupling mechanism, dynamic tracking and assessment of the safety margins in different service periods in a whole life cycle are realized, and accurate and reliable quantitative support is provided for retaining wall operation and maintenance reinforcing decision and structural safety risk assessment. In order to achieve the above purpose, the present invention provides a physical-data dual-drive retaining wall safety margin dynamic assessment method, which comprises the following steps: s1, inputting real-time monitoring parameters in a service period, completing time dispersion and parameter vectorization, calculating each stability coefficient of the retaining wall based on physical constraint, and mapping the stability coefficients into a safety margin reference function. S2, constructing a physical constraint learning sub-network architecture of an input layer, a hidden layer, a physical constraint layer and an output layer, guiding network parameter optimization through a physical regularization loss function L total, and establishing a mapping between a parameter vector x k to an optimized safety margin reference function M phys(tk under the condition of meeting the physical consistency requirement. S3, constructing a performance attenuation evolution sub-network architecture of an input layer-LSTM layer-full connection layer-output layer, guiding a network to perform parameter optimization through a comprehensive loss function L decay, establishing a mapping between a parameter vector x k and an optimized safety margin reference value M phys(tk to a safety margin loss amount delta M (t k), and outputting an effective safety margin loss function delta M (t k). S4, fusing the optimized safety margin reference function M phys(tk) and the safety margin loss function delta M (t k) by adopting an superposition method, and generating a continuous change curve of the total safety margin function M total(tk) with physical consistency and attenuation authenti