An evaluation method and system for the running state of a diaphragm wall based on an artificial neural network

Abstract

This invention relates to the field of anti-seepage wall technology and discloses a method and system for evaluating the operational status of anti-seepage walls based on artificial neural networks. It aims to address the problems of poor efficiency, accuracy, interpretability, and physical consistency in existing methods. The solution mainly includes: collecting multi-source data to construct a sample set containing standardized input and output vectors; designing a multi-layer feedforward neural network as a surrogate model; and training the model using a composite loss function that integrates engineering physical constraints and multi-task consistency. This enables the model to learn accurate mappings from five indicators: geometry, water level, soil parameters, stress, displacement, leakage, deformation rate, and outlet gradient coefficient. In application, the current operating condition parameters are input into the trained model, and a comprehensive evaluation result can be quickly output through a single forward inference. The health level is automatically classified according to standard thresholds. This invention achieves efficient, objective, reliable, and physically credible intelligent evaluation of the operational status of anti-seepage walls.

Description

Technical Field

[0001] This invention relates to the field of anti-seepage wall technology, specifically to a method and system for evaluating the operational status of anti-seepage walls based on artificial neural networks. Background Technology

[0002] As a crucial underground seepage prevention structure in water conservancy projects, the operational status of cutoff walls directly affects the safety and stability of the entire project. Traditional methods for evaluating the operational status of cutoff walls mainly rely on two types of technical means: one is numerical simulation analysis based on the finite element method, which involves establishing a detailed soil-structure interaction model, inputting soil physical and mechanical parameters, cutoff wall geometric parameters, and boundary conditions, and calculating indicators such as stress, deformation, and seepage; the other is empirical formulas based on field monitoring data and expert experience analysis, making qualitative or semi-quantitative judgments based on historical data patterns and engineering analogies.

[0003] However, the aforementioned existing technologies have significant limitations in practical engineering applications. While the finite element numerical simulation method is theoretically rigorous, its modeling process is complex, computationally resource-intensive, and time-consuming, making it difficult to meet the needs of real-time and rapid evaluation during operation. Furthermore, this method is extremely sensitive to the accuracy of input parameters; even small fluctuations in soil parameters can lead to significant deviations in the calculation results, and the interpretation of the results is highly dependent on the experience of professionals. On the other hand, while experience-based analysis methods based on monitoring data are relatively quick, they are highly subjective, lack consistency, struggle to quantify the coupled effects of multiple factors, and lack the ability to predict future changes in behavior.

[0004] In recent years, with the development of artificial intelligence technology, especially the rise of machine learning methods, some studies have attempted to apply it to structural condition prediction. However, traditional machine learning networks still have significant shortcomings when applied to predicting the condition of such complex engineering structures: First, traditional networks usually optimize and predict for a single indicator, making it difficult to synchronously and coordinately output multiple interrelated key performance indicators such as stress, deformation, and leakage, thus failing to form a systematic assessment of the overall health of the structure; Second, as purely data-driven models, their training process often aims only at reducing prediction errors, lacking the embedding and adherence to fundamental physical laws and engineering safety principles, which may lead to prediction results that are physically unreasonable or contradict common sense in engineering, i.e., poor interpretability and physical consistency. Summary of the Invention

[0005] This invention aims to address the problems of poor efficiency, accuracy, interpretability, and physical consistency in existing methods for evaluating the operational status of cut-off walls. It proposes a method and system for evaluating the operational status of cut-off walls based on artificial neural networks.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0007] In a first aspect, the present invention provides a method for evaluating the operational status of a cutoff wall based on an artificial neural network, the method comprising:

[0008] Historical engineering data, operational monitoring data, and numerical simulation results are collected to construct a multi-source sample set. Each sample in the multi-source sample set is processed into structured data containing a standardized input vector and a standardized target output vector. The standardized input vector includes the thickness of the cutoff wall, the depth of penetration into the rock, the thickness of the overburden layer, the average water level upstream and downstream, the rate of water level change, the net pumping volume, and a set of soil mechanical parameters. The standardized target output vector is a quintuple containing stress index, displacement index, leakage index, deformation rate index, and outlet gradient coefficient index.

[0009] A multi-layer feedforward neural network is constructed as a surrogate model. The surrogate model takes a standardized input vector as input and maps the output to a five-tuple prediction result with the same dimension and meaning as the standardized target output vector.

[0010] The surrogate model is iteratively trained using the backpropagation algorithm based on the multi-source sample set and its corresponding standardized input vector and standardized target output vector. During the training process, the model parameters are optimized by minimizing a composite loss function, which includes a prediction error term and a regularization term used to constrain the model output to conform to engineering physical laws and consistency among multiple indicators.

[0011] The current operating parameters of the cutoff wall to be evaluated are obtained, processed into corresponding standardized input vectors, and then input into the trained surrogate model. Through one forward inference, the corresponding five-tuple prediction results are obtained. Based on each index in the five-tuple prediction results and the preset safety threshold, the health status level of the cutoff wall is obtained.

[0012] Furthermore, the set of soil mechanical parameters includes tensile modulus, Poisson's ratio, density, average permeability, and porosity;

[0013] The outlet gradient coefficient in the standardized target output vector is the ratio of the outlet gradient at the downstream monitoring section of the anti-seepage wall to the critical gradient of the soil material.

[0014] Furthermore, the construction process of the proxy model includes:

[0015] A multi-layer feedforward neural network is constructed, comprising an input layer, at least one hidden layer, and an output layer; wherein the number of neurons in the input layer is the same as the dimension of the normalized input vector; the output layer is configured with five output neurons, corresponding to the stress index, displacement index, leakage index, deformation rate index, and outlet gradient coefficient index in the normalized target output vector, respectively.

[0016] Furthermore, the output value of the hidden layer of the proxy model is calculated by the following formula:

[0017] ;

[0018] in, Indicates the hidden layer number 1 The output value of each neuron Indicates the input layer's first... The input value of each neuron. Indicates the number of input layers. The first neuron is connected to the hidden layer. The connection weights of each neuron Indicates the hidden layer number 1 Bias terms for each neuron, Indicates the activation function;

[0019] The output value of the output layer of the proxy model is calculated by the following formula:

[0020] ;

[0021] Indicates the output layer number The output value of each neuron Indicates from the hidden layer The nth neuron to the output layer The connection weights of each neuron Indicates the output layer number Bias terms for each neuron.

[0022] Furthermore, during the iterative training process, a composite loss function is constructed based on the prediction error between the model's predicted value and the true value, and the composite loss function is used as the overall objective function for model optimization.

[0023] The formula for calculating the prediction error is as follows:

[0024] ;

[0025] in, Indicates the prediction error. The output layer of the proxy model is represented by the first... Output value of each neuron That is, the proxy model for the first The predicted values ​​of each indicator, For the corresponding sample in the multi-source sample set The true target value of each indicator;

[0026] The composite loss function is as follows:

[0027] ;

[0028] in, Indicates compound loss. Represents engineering physical constraints. This represents a regular expression for multi-task consistency. Indicates hyperparameters, This represents the learnable parameter vector of the proxy model. This represents the L2 norm.

[0029] Furthermore, the calculation formula for the engineering physical constraint term is as follows:

[0030] ;

[0031] in, This indicates the number of samples in a multi-source sample set. The proxy model represents the first The predicted export decline for each sample. Indicates the first The critical permeability gradient of the soil material corresponding to each sample. Represents the linear rectified function. Mathematical variables representing linear rectifier functions;

[0032] The calculation formula for the multi-task consistency regularization term is as follows:

[0033] ;

[0034] in, This indicates the number of samples in a multi-source sample set. The proxy model represents the first A sample at time... The predicted value of deformation rate, and These represent the proxy model for the first... A sample at time... Compared with the previous observation time The predicted displacement value, Indicates the observation time step.

[0035] Furthermore, during training, when the composite loss exceeds a set threshold, backpropagation of the error is initiated, and gradient descent is used to adjust the connection weights of the surrogate model. The formula for calculating the adjustment increment for any connection weight in the surrogate model is as follows:

[0036] ;

[0037] in, This represents the adjustment increment of the connection weights. Indicates the learning rate, Indicates composite loss For connection weights The partial derivatives of .

[0038] Furthermore, it also includes introducing a momentum term based on the gradient descent method to adjust the adjustment increment of the connection weights:

[0039] For the input layer The first neuron is connected to the hidden layer. Connection weights of each neuron In its first Adjustment amount during the next training iteration The calculation formula is as follows:

[0040] ;

[0041] For from the hidden layer The nth neuron to the output layer Connection weights of each neuron In its first Adjustment amount during the next training iteration The calculation formula is as follows:

[0042] ;

[0043] in, This indicates the hidden layer during the backpropagation of the error. The error gradient of each neuron. This indicates the output layer during the backpropagation of the error. The error gradient of each neuron. This represents the momentum coefficient.

[0044] Furthermore, during the backpropagation of the error, the error gradient between the hidden layer and the output layer neurons is calculated using the following formula:

[0045] ;

[0046] ;

[0047] in, Indicates the hidden layer number 1 Net input to each neuron, Indicates the output layer number Net input to each neuron, Indicates activation function The derivative of .

[0048] Secondly, the present invention provides a system for evaluating the operational status of a cutoff wall based on an artificial neural network, used to implement the method for evaluating the operational status of a cutoff wall based on an artificial neural network as described in the first aspect, the system comprising:

[0049] The data acquisition module is used to collect historical engineering data, operational monitoring data, and numerical simulation results to construct a multi-source sample set. Each sample in the multi-source sample set is processed into structured data containing a standardized input vector and a standardized target output vector. The standardized input vector includes the thickness of the cutoff wall, the depth of penetration into the rock, the thickness of the overburden layer, the average water level upstream and downstream, the rate of water level change, the net pumping volume, and a set of soil mechanical parameters. The standardized target output vector is a quintuple containing stress index, displacement index, leakage index, deformation rate index, and outlet gradient coefficient index.

[0050] The model building module is used to build a multi-layer feedforward neural network as a proxy model. The proxy model takes a standardized input vector as input and maps the output to a five-tuple prediction result with the same dimension and meaning as the standardized target output vector.

[0051] The model training module is used to iteratively train the surrogate model using the backpropagation algorithm based on the multi-source sample set and its corresponding standardized input vector and standardized target output vector; after training, the performance of the surrogate model is verified using the reserved validation set and test set.

[0052] The status evaluation module is used to obtain the current working condition parameters of the cutoff wall to be evaluated, process them into corresponding standardized input vectors, and input them into the trained and validated surrogate model to obtain the corresponding five-tuple prediction results; based on each index in the five-tuple prediction results and the preset safety threshold, the health status level of the cutoff wall is obtained.

[0053] The beneficial effects of this invention are as follows: The method and system for evaluating the operational status of cut-off walls based on artificial neural networks provided by this invention overcome the shortcomings of traditional finite element methods, such as computational complexity, time consumption, and reliance on expert experience, as well as the poor physical consistency and weak interpretability of traditional deep learning models. By constructing a neural network proxy model that integrates physical constraints, this invention achieves efficient, automatic, and comprehensive evaluation of the operational status of cut-off walls. Based on standardized multi-source data, this invention uses a composite loss function that includes physical and consistency constraints for training, ensuring that the prediction results conform to engineering realities and safety criteria. Finally, multiple key indicators can be output simultaneously through a single forward inference, and a rapid, objective, and interpretable intelligent judgment of health status is achieved based on standardized thresholds, significantly improving the timeliness, reliability, and engineering practicality of the evaluation. Attached Figure Description

[0054] Figure 1A flowchart illustrating the method for evaluating the operational status of an anti-seepage wall based on an artificial neural network, provided for an embodiment.

[0055] Figure 2 A schematic diagram of the strain change curve of a uniaxial strain gauge inside the seepage barrier wall provided in the embodiment;

[0056] Figure 3 A schematic diagram of the structure of a multilayer feedforward neural network provided in the embodiment;

[0057] Figure 4 This is a schematic diagram of the structure of the anti-seepage wall operation status evaluation system based on artificial neural networks provided in the embodiment. Detailed Implementation

[0058] Traditional methods for evaluating the operational status of anti-seepage walls suffer from low computational efficiency and strong subjective dependence, while traditional deep learning models lack the embedding of physical laws and the synergistic constraints of multiple indicators, resulting in poor physical consistency and weak interpretability of their prediction results. Both are difficult to achieve efficient and reliable comprehensive evaluation.

[0059] Based on this, the technical solution of this invention is proposed. In this invention, a multi-source sample set containing standardized input vectors and standardized target output vectors is constructed by collecting multi-source data, and a multi-layer feedforward neural network is constructed as a surrogate model. The model is trained using an error backpropagation algorithm with the goal of minimizing a composite loss function containing prediction error terms and regularization terms used to constrain physical laws and consistency of multiple indicators. This allows the surrogate model to learn the mapping relationship between five key indicators: the geometry of the cutoff wall, water level, soil parameters, stress, displacement, leakage, deformation rate, and outlet gradient coefficient. In practical applications, the current working condition parameters of the cutoff wall to be evaluated are processed into standardized input vectors of the same specification and input into the trained surrogate model. The five-tuple prediction result is obtained through one forward inference. Finally, by comparing each predicted indicator with a preset safety threshold, a rapid, automated, and intelligent evaluation of the health status of the cutoff wall is achieved.

[0060] The technical solutions in this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0061] Figure 1 A flowchart illustrating a method for evaluating the operational status of cutoff walls based on artificial neural networks is shown. Please refer to [link / reference]. Figure 1 The method includes the following steps:

[0062] Step 1: Data Acquisition and Preprocessing

[0063] Historical engineering data, operational monitoring data, and numerical simulation results are collected to construct a multi-source sample set. Each sample in the multi-source sample set is processed into structured data containing a standardized input vector and a standardized target output vector. The standardized input vector includes the thickness of the cutoff wall, the depth of penetration into the rock, the thickness of the overburden layer, the average water level upstream and downstream, the rate of water level change, the net pumping volume, and a set of soil mechanical parameters. The standardized target output vector is a quintuple containing stress index, displacement index, leakage index, deformation rate index, and outlet gradient coefficient index.

[0064] In this embodiment, the set of soil mechanical parameters includes tensile modulus, Poisson's ratio, density, average permeability coefficient, and porosity; the outlet gradient coefficient index in the standardized target output vector is the ratio of the outlet gradient at the downstream monitoring section of the cutoff wall to the critical gradient of the soil material.

[0065] In practical applications, multi-source sample data is first collected from three dimensions: first, historical engineering data, such as design and construction information; second, operational monitoring data, which may include strain time history data from uniaxial strain gauges (e.g., ... Figure 2 The data consist of three parts: (1) wall deformation cloud maps or key point displacements during the completion period; (2) vertical stress distribution and leakage monitoring sequences on the upstream and downstream sides of the wall; and (3) numerical simulation results obtained through traditional numerical simulation methods such as the finite element method. These data together constitute a multi-source sample set covering various possible working conditions and responses.

[0066] Subsequently, the multi-source sample set underwent crucial structuring processing. Each sample was processed into a structured data pair. Its input part was defined as a standardized input vector, which integrated key factors influencing the performance of the cutoff wall, specifically including: the cutoff wall's own geometric parameters (thickness, depth into the rock, thickness of the overburden layer), environmental load parameters (average upstream and downstream water levels, water level change rate over the past 7 days, net pumping volume), and a set of soil mechanical parameters characterizing the foundation properties. In this embodiment, this parameter set specifically includes: elastic modulus, Poisson's ratio, density, average permeability coefficient, and porosity. The output part was defined as a standardized target output vector, which is a quintuple with clear physical meaning, designed to comprehensively describe the operational performance of the cutoff wall, including: stress indices. Displacement (deformation) index Leakage index Deformation rate index The text also mentions an outlet gradient coefficient, used to characterize permeability stability. The outlet gradient coefficient is defined as the actual outlet gradient at the downstream monitoring section of the cutoff wall. Critical gradient of soil material ratio Its value directly reflects the degree of risk of piping or seepage downstream. That is, the quintuple is: .

[0067] Through the above steps, the original engineering data from multiple sources and heterogeneity are transformed into structured samples with a unified format, consistent dimensions, and complete input-output mapping relationships.

[0068] Step 2, Proxy Model Construction:

[0069] A multi-layer feedforward neural network is constructed as a surrogate model. The surrogate model takes a standardized input vector as input and maps the output to a quintuple prediction result with the same dimension and meaning as the standardized target output vector.

[0070] In this embodiment, the construction process of the proxy model includes:

[0071] A multi-layer feedforward neural network is constructed, comprising an input layer, at least one hidden layer, and an output layer; wherein the number of neurons in the input layer is the same as the dimension of the normalized input vector; the output layer is configured with five output neurons, corresponding to the stress index, displacement index, leakage index, deformation rate index, and outlet gradient coefficient index in the normalized target output vector, respectively.

[0072] Specifically, this step is used to construct a multi-layer feedforward neural network based on the error backpropagation algorithm, and to use it as a surrogate model, such as... Figure 3 As shown, the number of neurons in the input layer of this multi-layer feedforward neural network corresponds exactly to the dimension of the defined standardized input vector, ensuring that all engineering parameters can be fully received. The output layer is specifically designed with five neurons, each designed to correspond one-to-one with a quintuple, aiming to simultaneously predict five key performance indicators: stress, displacement, leakage, deformation rate, and outlet gradient coefficient. By introducing at least one hidden layer between the input and output layers and selecting an appropriate activation function, the model possesses the ability to learn the nonlinear mapping relationship between input parameters and complex multi-indicator responses, thus providing a foundation for subsequent rapid and accurate automated evaluation.

[0073] In this embodiment, the output value of the hidden layer of the proxy model is calculated by the following formula:

[0074] ;

[0075] in, Indicates the hidden layer number 1 The output value of each neuron Indicates the input layer's first... The input value of each neuron. Indicates the number of input layers. The first neuron is connected to the hidden layer. The connection weights of each neuron Indicates the hidden layer number 1 Bias terms for each neuron, Indicates the activation function;

[0076] The output value of the output layer of the proxy model is calculated by the following formula:

[0077] ;

[0078] Indicates the output layer number The output value of each neuron Indicates from the hidden layer The nth neuron to the output layer The connection weights of each neuron Indicates the output layer number Bias terms for each neuron.

[0079] Specifically, the forward computation process of the surrogate model is achieved through weighted summation and nonlinear activation between neurons. Hidden layer number... Output value of each neuron From all its input values (Each element of the corresponding standardized input vector) and the corresponding connection weights The weighted sum, plus the bias term Then, it is activated by the function. Obtained by nonlinear transformation (such as ReLU or Tanh), i.e. This process transforms the original input features into higher-dimensional, more expressive hidden layer features. Subsequently, the output layer... Output value of each neuron (corresponding to one indicator in the quintuple) Calculated in a similar manner, it is expressed as the hidden layer output value. As input, through weights Weighted summation plus bias Then, after processing by the activation function, that is... This layer-by-layer forward propagation mechanism enables the surrogate model to learn and implement a nonlinear mapping from complex multidimensional inputs to multidimensional performance indices.

[0080] Step 3: Train the surrogate model:

[0081] The surrogate model is iteratively trained using the backpropagation algorithm based on the multi-source sample set and its corresponding standardized input vector and standardized target output vector. During training, the model parameters are optimized by minimizing a composite loss function, which includes a prediction error term and a regularization term used to constrain the model output to conform to engineering physics laws and consistency among multiple indicators.

[0082] During the iterative training process, a composite loss function is constructed based on the prediction error between the model's predicted value and the true value, and the composite loss function is used as the overall objective function to optimize the model.

[0083] The formula for calculating the prediction error is as follows:

[0084] ;

[0085] in, Indicates the prediction error. The output layer of the proxy model is represented by the first... Output value of each neuron That is, the proxy model for the first The predicted values ​​of each indicator, For the corresponding sample in the multi-source sample set The true target value of each indicator;

[0086] The composite loss function is as follows:

[0087] ;

[0088] in, Indicates compound loss. Represents engineering physical constraints. This represents a regular expression for multi-task consistency. Indicates hyperparameters, This represents the learnable parameter vector of the proxy model. This represents the L2 norm.

[0089] Specifically, this embodiment incorporates a domain-knowledge-based training mechanism, enabling the surrogate model to not only learn data patterns but also adhere to engineering physical constraints. Training employs an error backpropagation algorithm, aiming to minimize the composite loss function. This composite loss function includes not only a prediction error term measuring the difference between predicted and true values ​​but also an engineering physical constraint term. and multi-task consistency regularization Furthermore, L2 regularization is used to control model complexity. This is achieved by balancing hyperparameters. This training process, while pursuing data fitting accuracy, incorporates key physical laws such as the seepage stability of the anti-seepage wall and the inherent coordination between multiple indicators as hard constraints into the model. This ensures that the prediction results of the finally trained surrogate model have both data accuracy and physical credibility, overcoming the problem of violating engineering common sense that may occur in traditional pure data-driven models.

[0090] In this embodiment, the calculation formula for the engineering physical constraint term is as follows:

[0091] ;

[0092] in, This indicates the number of samples in a multi-source sample set. The proxy model represents the first The predicted export decline for each sample. Indicates the first The critical permeability gradient of the soil material corresponding to each sample. Represents the linear rectified function. Mathematical variables representing linear rectifier functions.

[0093] Specifically, engineering physical constraints Using the ReLU function to reduce the export ratio predicted by the model Exceeding the critical gradient of the soil In cases where a one-sided penalty is imposed, the agent model is forced to automatically avoid physically unsafe situations during learning (i.e., The prediction ensures that the output conforms to common sense of engineering safety.

[0094] The calculation formula for the multi-task consistency regularization term is as follows:

[0095] ;

[0096] in, This indicates the number of samples in a multi-source sample set. The proxy model represents the first A sample at time... The predicted value of deformation rate, and These represent the proxy model for the first... A sample at time... Compared with the previous observation time The predicted displacement value, Indicates the observation time step.

[0097] Specifically, multi-task consistency regularization To ensure the inherent consistency among the multiple outputs of the surrogate model, its calculation formula requires the deformation rate predicted by the surrogate model. Its corresponding displacement prediction value By satisfying the numerical derivative relationship on the time series, logical contradictions between different output indicators (displacement and velocity) are avoided, enhancing the overall physical rationality and self-consistency of the surrogate model's output. These two constraints, together with the prediction error term, constitute a composite loss function, guiding the surrogate model to learn a comprehensive mapping relationship that both fits the data and obeys physical laws.

[0098] During training, when the composite loss exceeds a set threshold, backpropagation of the error is initiated, and gradient descent is used to adjust the connection weights of the surrogate model. The formula for calculating the adjustment increment for any connection weight in the surrogate model is as follows:

[0099] ;

[0100] in, This represents the adjustment increment of the connection weights. Indicates the learning rate, Indicates composite loss For connection weights The partial derivatives of .

[0101] Specifically, in model training, when the composite loss... When the preset threshold is exceeded, the error backpropagation process will be initiated, iteratively optimizing the connection weights in the surrogate model based on the gradient descent principle. This process calculates the composite loss. For each connection weight partial derivatives Determine the direction in which the loss decreases the fastest, and then multiply it by the learning rate. By controlling the update step size, the final adjustment increment of the weights is obtained. This mechanism enables the surrogate model to automatically and sequentially adjust its internal parameters during training, continuously approaching the optimal mapping relationship that simultaneously satisfies data fitting accuracy, engineering physical constraints, and multi-task consistency, thereby gradually improving its overall prediction performance.

[0102] In this embodiment, a momentum term is introduced based on the gradient descent method to adjust the adjustment increment of the connection weights:

[0103] For the input layer The first neuron is connected to the hidden layer. Connection weights of each neuron In its first Adjustment amount during the next training iteration The calculation formula is as follows:

[0104] ;

[0105] For from the hidden layer The nth neuron to the output layer Connection weights of each neuron In its first Adjustment amount during the next training iteration The calculation formula is as follows:

[0106] ;

[0107] in, This indicates the hidden layer during the backpropagation of the error. The error gradient of each neuron. This indicates the output layer during the backpropagation of the error. The error gradient of each neuron. This represents the momentum coefficient.

[0108] To improve training stability and accelerate convergence during gradient descent optimization, this embodiment introduces a momentum term into the standard weight update rule. This term applies to the weights from the input layer to the hidden layer. Its adjustment amount Not only includes based on the current error gradient and input values Instantaneous adjustment item It also retains a certain proportion (based on the momentum coefficient). The previous iteration adjustment amount of the control) As the direction of inertia; weights from the hidden layer to the output layer gradient amount Similarly, calculate, that is This mechanism of introducing historical update directions helps to smooth the optimization path, enabling the model to more effectively traverse flat regions or escape local minima during training, thereby achieving more stable and efficient convergence when optimizing complex composite loss functions.

[0109] In this embodiment, during the backpropagation of errors, the error gradient between the hidden layer and the output layer neurons is calculated using the following formula:

[0110] ;

[0111] ;

[0112] in, Indicates the hidden layer number 1 Net input to each neuron, Indicates the output layer number Net input to each neuron, Indicates activation function The derivative of .

[0113] Specifically, the error gradient of the output layer neurons Due to composite loss The output of this neuron partial derivatives Multiply by the derivative of the activation function at that point This error is obtained; it is then backpropagated to the hidden layer, the th hidden layer. Error gradient of each neuron Errors of all output layer neurons connected to it Weight After weighted summation, it is also multiplied by the derivative of the activation function of the hidden layer neuron. This chain-law-based error backpropagation mechanism ensures that the composite loss, which includes data errors, physical constraints, and consistency requirements, is accounted for. The gradient information can be precisely assigned to each weight in each layer of the network, thereby guiding the model to be effectively optimized in a direction that makes the prediction results more accurate and more in line with the laws of engineering physics.

[0114] Step 4: Evaluation of the operational status of the anti-seepage wall:

[0115] The current operating parameters of the cutoff wall to be evaluated are obtained, processed into corresponding standardized input vectors, and then input into the trained surrogate model. Through one forward inference, the corresponding five-tuple prediction results are obtained. Based on each index in the five-tuple prediction results and the preset safety threshold, the health status level of the cutoff wall is obtained.

[0116] Specifically, when a condition assessment of a cutoff wall needs to be performed, the following steps are taken: First, various parameters under its current operating conditions are obtained, including geometric dimensions, real-time water level, monitoring data, and foundation soil parameters. These parameters are processed into standardized input vectors according to the rules defined in step 1, ensuring that their format is completely consistent with the training data. Then, this vector is input into the already trained surrogate model. Through a single efficient forward inference calculation, the model can output the corresponding five-tuple prediction results within milliseconds, namely the predicted values ​​of five indicators: stress, displacement, leakage, deformation rate, and outlet gradient coefficient.

[0117] Then, the prediction results are interpreted and comprehensively evaluated in an engineering manner. Each predicted index value is compared with a safety threshold pre-set based on design specifications, historical safety baselines, and expert experience. By calculating the utilization coefficient or safety margin of each index and selecting the most unfavorable value as the basis for comprehensive scoring, the state of the cutoff wall is ultimately mapped to a health status level with clear engineering significance. This results in the output of intuitive level conclusions and key indicators, greatly reducing reliance on manual interpretation.

[0118] In summary, the artificial neural network-based method for evaluating the operational status of cut-off walls provided in this embodiment achieves millisecond-level, automated comprehensive evaluation of the cut-off wall's operational performance by constructing a multi-layer feedforward neural network surrogate model that integrates engineering physical constraints and multi-task consistency requirements. It transforms the complex numerical simulation process into a single, efficient forward inference, simultaneously outputting five key indicators: stress, displacement, leakage, deformation rate, and outlet gradient coefficient. These indicators are directly mapped to a defined health level by comparison with standard thresholds. This significantly improves the timeliness, objectivity, and engineering practicality of the evaluation while ensuring the physical reliability of the prediction results and the coordination of multiple indicators. It provides reliable technical support for realizing routine intelligent monitoring and early warning of cut-off walls.

[0119] Based on the above technical solutions, this embodiment also proposes an artificial neural network-based anti-seepage wall operation status evaluation system to implement the artificial neural network-based anti-seepage wall operation status evaluation method as described in the embodiment. Please refer to [link to relevant documentation]. Figure 4 The system includes:

[0120] The data acquisition module is used to collect historical engineering data, operational monitoring data, and numerical simulation results to construct a multi-source sample set. Each sample in the multi-source sample set is processed into structured data containing a standardized input vector and a standardized target output vector. The standardized input vector includes the thickness of the cutoff wall, the depth of penetration into the rock, the thickness of the overburden layer, the average water level upstream and downstream, the rate of water level change, the net pumping volume, and a set of soil mechanical parameters. The standardized target output vector is a quintuple containing stress index, displacement index, leakage index, deformation rate index, and outlet gradient coefficient index.

[0121] The model building module is used to build a multi-layer feedforward neural network as a proxy model. The proxy model takes a standardized input vector as input and maps the output to a five-tuple prediction result with the same dimension and meaning as the standardized target output vector.

[0122] The model training module is used to iteratively train the surrogate model using the backpropagation algorithm based on the multi-source sample set and its corresponding standardized input vector and standardized target output vector; after training, the performance of the surrogate model is verified using the reserved validation set and test set.

[0123] The status evaluation module is used to obtain the current working condition parameters of the cutoff wall to be evaluated, process them into corresponding standardized input vectors, and input them into the trained and validated surrogate model to obtain the corresponding five-tuple prediction results; based on each index in the five-tuple prediction results and the preset safety threshold, the health status level of the cutoff wall is obtained.

[0124] It is understood that since the artificial neural network-based anti-seepage wall operation status evaluation system described in this embodiment is a system for implementing the artificial neural network-based anti-seepage wall operation status evaluation method described in the embodiment, the system disclosed in the embodiment is relatively simple to describe because it corresponds to the method disclosed in the embodiment. For relevant parts, please refer to the description of the method, and it will not be repeated here.

Claims

1. A method for evaluating the operational status of a cutoff wall based on an artificial neural network, characterized in that, The method includes: Historical engineering data, operational monitoring data, and numerical simulation results are collected to construct a multi-source sample set. Each sample in the multi-source sample set is processed into structured data containing a standardized input vector and a standardized target output vector. The standardized input vector includes the thickness of the cutoff wall, the depth of penetration into the rock, the thickness of the overburden layer, the average water level upstream and downstream, the rate of water level change, the net pumping volume, and a set of soil mechanical parameters. The standardized target output vector is a quintuple containing stress index, displacement index, leakage index, deformation rate index, and outlet gradient coefficient index. A multi-layer feedforward neural network is constructed as a surrogate model. The surrogate model takes a standardized input vector as input and maps the output to a five-tuple prediction result with the same dimension and meaning as the standardized target output vector. The surrogate model is iteratively trained using the backpropagation algorithm based on the multi-source sample set and its corresponding standardized input vector and standardized target output vector. During the training process, the model parameters are optimized by minimizing a composite loss function, which includes a prediction error term and a regularization term used to constrain the model output to conform to engineering physical laws and consistency among multiple indicators. The current operating parameters of the cutoff wall to be evaluated are obtained, processed into corresponding standardized input vectors, and then input into the trained surrogate model. Through one forward inference, the corresponding five-tuple prediction results are obtained. Based on each index in the five-tuple prediction results and the preset safety threshold, the health status level of the cutoff wall is obtained.

2. The method for evaluating the operational status of an anti-seepage wall based on an artificial neural network according to claim 1, characterized in that, The set of soil mechanical parameters includes physical modulus, Poisson's ratio, density, average permeability, and porosity; The outlet gradient coefficient in the standardized target output vector is the ratio of the outlet gradient at the downstream monitoring section of the anti-seepage wall to the critical gradient of the soil material.

3. The method for evaluating the operational status of an anti-seepage wall based on an artificial neural network according to claim 1, characterized in that, The construction process of the proxy model includes: A multi-layer feedforward neural network is constructed, comprising an input layer, at least one hidden layer, and an output layer; wherein the number of neurons in the input layer is the same as the dimension of the normalized input vector; the output layer is configured with five output neurons, corresponding to the stress index, displacement index, leakage index, deformation rate index, and outlet gradient coefficient index in the normalized target output vector, respectively.

4. The method for evaluating the operational status of an anti-seepage wall based on an artificial neural network according to claim 3, characterized in that, The output value of the hidden layer of the proxy model is calculated by the following formula: ； in, Indicates the hidden layer number 1 The output value of each neuron Indicates the input layer's first... The input value of each neuron, Indicates the number of input layers. The first neuron is connected to the hidden layer. The connection weights of each neuron Indicates the hidden layer number 1 Bias terms for each neuron, Indicates the activation function; The output value of the output layer of the proxy model is calculated by the following formula: ； Indicates the output layer number The output value of each neuron Indicates from the hidden layer The nth neuron to the output layer The connection weights of each neuron Indicates the output layer number Bias terms for each neuron.

5. The method for evaluating the operational status of an anti-seepage wall based on an artificial neural network according to claim 4, characterized in that, During the iterative training process, a composite loss function is constructed based on the prediction error between the model's predicted value and the true value, and the composite loss function is used as the overall objective function to optimize the model. The formula for calculating the prediction error is as follows: ； in, Indicates the prediction error. The output layer of the proxy model is represented by the first... Output value of each neuron That is, the proxy model for the first The predicted values ​​of each indicator, For the corresponding sample in the multi-source sample set The true target value of each indicator; The composite loss function is as follows: ； in, Indicates compound loss. Represents engineering physical constraints. This represents a regular expression for multi-task consistency. Indicates hyperparameters, This represents the learnable parameter vector of the proxy model. This represents the L2 norm.

6. The method for evaluating the operational status of an anti-seepage wall based on an artificial neural network according to claim 5, characterized in that, The calculation formula for the engineering physical constraints is as follows: ； in, This indicates the number of samples in a multi-source sample set. The proxy model represents the first The predicted export decline for each sample. Indicates the first The critical permeability gradient of the soil material corresponding to each sample. Represents the linear rectified function. Mathematical variables representing linear rectifier functions; The calculation formula for the multi-task consistency regularization term is as follows: ； in, This indicates the number of samples in a multi-source sample set. The proxy model represents the first Each sample at time... The predicted value of deformation rate, and These represent the proxy model for the first... Each sample at time... Compared with the previous observation time The predicted displacement value, Indicates the observation time step.

7. The method for evaluating the operational status of an anti-seepage wall based on an artificial neural network according to claim 5, characterized in that, During training, when the composite loss exceeds a set threshold, backpropagation of the error is initiated, and gradient descent is used to adjust the connection weights of the surrogate model. The formula for calculating the adjustment increment for any connection weight in the surrogate model is as follows: ； in, This represents the adjustment increment of the connection weights. Indicates the learning rate, Indicates composite loss For connection weights The partial derivatives of .

8. The method for evaluating the operational status of an anti-seepage wall based on an artificial neural network according to claim 7, characterized in that, This also includes introducing a momentum term based on gradient descent to adjust the increment of the connection weights: For the input layer The first neuron is connected to the hidden layer. Connection weights of each neuron In its first Adjustment amount during the next training iteration The calculation formula is as follows: ； For from the hidden layer The nth neuron to the output layer Connection weights of each neuron In its first Adjustment amount during the next training iteration The calculation formula is as follows: ； in, This indicates the hidden layer during the backpropagation of the error. The error gradient of each neuron. This indicates the output layer during the backpropagation of the error. The error gradient of each neuron. This represents the momentum coefficient.

9. The method for evaluating the operational status of an anti-seepage wall based on an artificial neural network according to claim 8, characterized in that, During the backpropagation of the error, the error gradient between the hidden layer and the output layer neurons is calculated using the following formula: ； ； in, Indicates the hidden layer number 1 Net input to each neuron, Indicates the output layer number Net input to each neuron, Indicates activation function The derivative of .

10. A system for evaluating the operational status of a cutoff wall based on an artificial neural network, characterized in that, The system is used to implement the method for evaluating the operational status of a cutoff wall based on an artificial neural network as described in any one of claims 1 to 9, the system comprising: The data acquisition module is used to collect historical engineering data, operational monitoring data, and numerical simulation results to construct a multi-source sample set. Each sample in the multi-source sample set is processed into structured data containing a standardized input vector and a standardized target output vector. The standardized input vector includes the thickness of the cutoff wall, the depth of penetration into the rock, the thickness of the overburden layer, the average water level upstream and downstream, the rate of water level change, the net pumping volume, and a set of soil mechanical parameters. The standardized target output vector is a quintuple containing stress index, displacement index, leakage index, deformation rate index, and outlet gradient coefficient index. The model building module is used to build a multi-layer feedforward neural network as a proxy model. The proxy model takes a standardized input vector as input and maps the output to a five-tuple prediction result with the same dimension and meaning as the standardized target output vector. The model training module is used to iteratively train the surrogate model using the backpropagation algorithm based on the multi-source sample set and its corresponding standardized input vector and standardized target output vector; after training, the performance of the surrogate model is verified using the reserved validation set and test set. The status evaluation module is used to obtain the current working condition parameters of the cutoff wall to be evaluated, process them into corresponding standardized input vectors, and input them into the trained and validated surrogate model to obtain the corresponding five-tuple prediction results; based on each index in the five-tuple prediction results and the preset safety threshold, the health status level of the cutoff wall is obtained.

Images