Urban subsidence prediction method based on transfer learning

Abstract

The present application relates to the technical field of building informatization, more specifically, to a city underground space settlement prediction method and system based on transfer learning, comprising: obtaining geological feature data, construction technology data and historical settlement monitoring data of a target city; obtaining geological feature data, construction technology data and historical settlement monitoring data of a source domain city with rich settlement data; based on the source domain city data, constructing and training a source domain neural network model; based on the target city data and the source domain neural network model, constructing a target domain neural network model; according to the target domain neural network model and the source domain neural network model, performing a transfer learning process to obtain a fusion model; based on the fusion model, generating a target city underground space settlement prediction result; by introducing the transfer learning technology, the present application can effectively utilize the settlement knowledge of a data-rich city to improve the prediction accuracy of a data-scarce city.

Description

Technical Field

[0001] This invention relates to the field of building information technology, and more specifically, to a method and system for predicting urban underground space settlement based on transfer learning. Background Technology

[0002] With the continuous advancement of urbanization, underground space development has become an important means to alleviate urban land shortages and improve land use efficiency. However, underground space development inevitably causes land subsidence, posing a potential threat to the safety of urban buildings and infrastructure. Therefore, accurately predicting urban underground space subsidence is crucial for ensuring urban safety and sustainable development.

[0003] Traditional settlement prediction methods mainly include empirical formulas, numerical simulations, and artificial intelligence (AI) methods. Empirical formulas are simple and intuitive, but they struggle to account for complex geological conditions and construction factors, resulting in limited prediction accuracy. While numerical simulations can effectively simulate settlement mechanisms, they are computationally complex, time-consuming, and require high parameter accuracy. In recent years, with the development of machine learning technology, AI-based settlement prediction methods have been widely applied. These methods can automatically learn complex nonlinear relationships and possess strong predictive capabilities.

[0004] However, existing AI-based prediction methods still have some problems. First, these methods typically require a large amount of historical data to train the model, but in newly developed cities or regions, there is often a lack of sufficient settlement monitoring data, making model training difficult and resulting in low prediction accuracy. Second, the geological conditions and engineering characteristics of different cities vary greatly, and when a model trained in one city is directly applied to another, it often results in significant prediction bias. Furthermore, existing methods mostly focus on single-point settlement prediction, lacking comprehensive analysis and risk assessment of the settlement trend across the entire region. Summary of the Invention

[0005] This invention aims to solve the aforementioned technical problems by proposing a method for predicting urban underground space subsidence based on transfer learning. This method effectively addresses issues such as insufficient data and poor model generalization ability by innovatively integrating multi-source data and transfer learning techniques, while simultaneously enabling regional subsidence trend analysis and risk warning.

[0006] This invention provides a method for predicting urban underground space subsidence based on transfer learning, comprising:

[0007] The acquisition steps include:

[0008] Obtain geological feature data, construction technology data, and historical settlement monitoring data for the target city;

[0009] Acquire geological characteristic data, construction technology data, and historical settlement monitoring data of source cities with abundant settlement data;

[0010] The processing steps include:

[0011] Based on the source domain city data, a source domain neural network model is constructed and trained;

[0012] Based on the target city data and the source domain neural network model, a target domain neural network model is constructed;

[0013] Based on the target domain neural network model and the source domain neural network model, a transfer learning process is performed to obtain a fusion model;

[0014] Output steps, including:

[0015] Based on the fusion model, the predicted results of underground space subsidence for the target city are generated.

[0016] Preferably, the acquisition step specifically includes:

[0017] The geological feature data, including soil stratification, thickness, and subsidence trend, are obtained through geological surveys.

[0018] The construction process data, including support structure type, excavation method, and support type, is obtained through engineering records;

[0019] The historical settlement monitoring data is obtained through on-site monitoring equipment, including the relative settlement between the measuring point and the horizontal benchmark, the cumulative settlement of the monitoring point, and the settlement rate.

[0020] Preferably, the process of constructing the source domain neural network model specifically includes:

[0021] The source city data is preprocessed, including data cleaning, standardization, and outlier handling.

[0022] A deep neural network structure with multiple hidden layers is constructed, each layer equipped with a non-linear activation function; the source domain neural network model is trained using the backpropagation algorithm until the loss function converges. Preferably, the construction of the target domain neural network model in the processing step specifically includes: replicating the structure of the source domain neural network model as an initial target domain neural network model; fine-tuning the initial target domain neural network model using the target city data; and updating the parameters of the target domain neural network model based on the gradient of the loss function.

[0023] Preferably, the transfer learning process in the processing steps specifically includes:

[0024] Calculate the difference in data distribution between the source and target domains;

[0025] Based on the aforementioned differences, the degree of source domain knowledge transfer is dynamically adjusted;

[0026] The parameters of the fusion model are optimized by combining the loss functions of the source and target domains.

[0027] Preferably, the degree of dynamic adjustment of source domain knowledge transfer specifically includes:

[0028] Introduce a weighting coefficient λ, where 0 ≤ λ ≤ 1;

[0029] When λ = 0, the target domain data is used entirely; when λ = 1, the source domain data is used entirely; the optimal value of λ is determined through iterative optimization.

[0030] Preferably, a model evaluation step is also included:

[0031] The prediction accuracy of the fusion model was evaluated using validation set data;

[0032] Calculate evaluation indicators such as mean absolute percentage error;

[0033] Based on the evaluation results, determine whether further optimization of the fusion model is needed.

[0034] As a preferred option, the inversion parameter step is also included:

[0035] Based on the historical settlement monitoring data, the key parameters affecting settlement are obtained by inversion.

[0036] The key parameters are input as additional features into the fusion model;

[0037] The fusion model was updated to improve its ability to capture sedimentation mechanisms.

[0038] Preferably, the output step further includes:

[0039] Generate a regional subsidence trend map for the target city;

[0040] Identify areas with a high risk of subsidence;

[0041] Output settlement early warning information.

[0042] Preferably, a model update step is also included:

[0043] Regularly obtain the latest subsidence monitoring data for the target city;

[0044] Based on the latest data, update the parameters of the fusion model;

[0045] Repeat the output step to generate updated settlement prediction results.

[0046] The beneficial effects of this invention are mainly reflected in the following aspects:

[0047] First, by introducing transfer learning technology, this invention can effectively utilize subsidence knowledge from data-rich cities to improve prediction accuracy in data-scarce cities. This knowledge transfer mechanism not only solves the problem of insufficient data but also greatly improves the model's generalization ability, enabling it to adapt to the geological conditions and engineering characteristics of different cities.

[0048] Secondly, this invention employs a multimodal data fusion strategy, comprehensively considering factors such as geological characteristics, construction techniques, and historical settlement data. This comprehensive data utilization approach enables the model to more accurately capture various factors affecting settlement, thereby improving the accuracy and reliability of predictions.

[0049] Furthermore, this invention achieves precise adaptation of the model to the characteristics of the target city by dynamically adjusting the degree of source domain knowledge transfer. This adaptive mechanism ensures that the model can fully learn the specific features of the target city while retaining useful prior knowledge, thereby achieving optimal knowledge transfer between different cities.

[0050] Furthermore, this invention not only provides single-point settlement prediction but also generates regional settlement trend maps and risk warning information. This comprehensive output provides full decision support for urban planning, engineering design, and risk management, greatly improving the practicality and value of the prediction results.

[0051] Finally, this invention introduces a continuous model update mechanism to ensure that the prediction model maintains its accuracy and applicability over time. This dynamic update method can better adapt to changes in geological conditions, engineering activities, and other factors, providing long-term reliable settlement prediction support.

[0052] In summary, the urban underground space settlement prediction method based on transfer learning proposed in this invention effectively solves many problems existing in current technologies through the organic combination of multiple innovative technologies, significantly improving the accuracy, reliability, and practicality of settlement prediction. The application of this method will provide strong technical support for the safe development and management of urban underground space, and is of great significance for promoting sustainable urban development. Attached Figure Description

[0053] Figure 1 This is a flowchart of the method of the present invention.

[0054] Figure 2 This is a logic block diagram of the data acquisition steps of the present invention.

[0055] Figure 3 This is a logical block diagram of the data processing and model building of the present invention.

[0056] Figure 4 This is a logical block diagram of the transfer learning process of the present invention.

[0057] Figure 5 This is a logic block diagram for model evaluation and updating in this invention. Detailed Implementation

[0058] Please refer to Figure 1-5 This invention provides a method for predicting urban underground space subsidence based on transfer learning. This method, by innovatively integrating multi-source data and transfer learning techniques, effectively solves the problems of insufficient data and low prediction accuracy in urban underground space subsidence prediction. The specific implementation of this invention will be described in detail below.

[0059] The method of the present invention includes an acquisition step, a processing step, and an output step.

[0060] In the data acquisition step, this method first acquires geological feature data, construction technology data, and historical settlement monitoring data for the target city. These data form the foundational inputs for the prediction model. Simultaneously, this method also acquires corresponding data from source city areas with abundant settlement data. Source city areas are typically those that have undergone extensive underground space development and possess well-established monitoring systems; their data can provide valuable prior knowledge for the prediction of the target city.

[0061] In the processing steps, this method first constructs and trains a source domain neural network model based on source domain city data. This model captures the general patterns of underground space subsidence in the source domain cities. Next, based on target domain data and the source domain neural network model, a target domain neural network model is constructed. This step achieves initial knowledge transfer. Then, this method performs a transfer learning process to obtain a fusion model. This process achieves deep transfer and adaptation of source domain knowledge to the target domain.

[0062] In the output step, this method generates underground space subsidence prediction results for the target city based on a fusion model. This result comprehensively considers source domain knowledge and target domain characteristics, providing reliable decision support for urban underground space development.

[0063] Furthermore, the acquisition steps of the present invention include more specific data acquisition methods.

[0064] For geological feature data, this method obtains the data through geological surveys. This data includes soil stratification, thickness, and settlement trends. For example, soil stratification may include multiple layers such as topsoil, silty clay, and sandy clay, with each layer varying in thickness from several meters to tens of meters. Settlement trends may exhibit linear, non-linear, or periodic variations.

[0065] For construction process data, this method obtains the data through engineering records. This data includes the type of support structure, excavation method, and support type. For example, the support structure may be a soil nailing wall, pile-anchor support, or diaphragm wall; the excavation method may be open-cut, cut-and-cover, or tunnel; and the support type may be internal bracing, cable anchors, or no bracing, etc.

[0066] For historical settlement monitoring data, this method acquires the data using on-site monitoring equipment. This data includes the relative settlement between the measuring point and the horizontal baseline, the cumulative settlement at the monitoring point, and the settlement rate. For example, the relative settlement may be on the order of millimeters to centimeters, the cumulative settlement may reach tens of centimeters, and the settlement rate may vary from a few millimeters to a few centimeters per day.

[0067] This detailed data acquisition method ensures the comprehensiveness and accuracy of the input data, laying a solid foundation for subsequent model training and prediction.

[0068] This invention employs a series of innovative processing methods when constructing the source domain neural network model.

[0069] First, this method preprocesses the source city data. This includes data cleaning, standardization, and outlier removal. Data cleaning removes erroneous or irrelevant data points, such as outlier readings caused by equipment malfunctions. Standardization maps data from different scales to the same scale, typically using z-score standardization.

[0070]

[0071] Where z is the standardized value, x is the original value, μ is the mean, and σ is the standard deviation.

[0072] Outlier handling can employ the 3σ rule, which considers data exceeding the mean by three standard deviations as outliers and removes or corrects them.

[0073] Next, this method constructs a deep neural network structure containing multiple hidden layers, each equipped with a non-linear activation function. A typical network structure may contain 4-6 hidden layers, with the number of neurons in each layer ranging from tens to hundreds. Commonly used non-linear activation functions include ReLU, Sigmoid, and Tanh. For example, the expression for the ReLU function is:

[0074] f(x) = max(0,x)

[0075] This deep structure can effectively capture the nonlinear relationships in the complex process of underground space subsidence.

[0076] Finally, this method uses the backpropagation algorithm to train the source domain neural network model until the loss function converges. The loss function is typically the mean squared error (MSE).

[0077]

[0078] Where n is the number of samples, y i For the true value, These are predicted values.

[0079] Preferably, when the rate of change of the loss function is less than 1e -6 When the threshold is reached, the model can be considered converged. This threshold is chosen based on experience and can be fine-tuned according to specific circumstances in practical applications.

[0080] Through the above steps, this method can construct a source domain neural network model with powerful predictive capabilities, laying the foundation for subsequent transfer learning. Compared to traditional single-city prediction models, this method can better utilize existing knowledge, improving prediction accuracy and generalization ability.

[0081] This invention employs an efficient transfer strategy when constructing the target domain neural network model.

[0082] First, this method replicates the structure of the source domain neural network model as the initial target domain neural network model. This approach is based on the assumption that although the sedimentation patterns of the source and target domains differ, they share certain similarities. By replicating the source domain model structure, an initial model with some predictive ability can be quickly obtained, significantly accelerating the model's convergence speed.

[0083] Next, this method fine-tunes the initial target domain neural network model using target city data. This step is central to transfer learning, allowing the model to adaptively adjust based on the specific circumstances of the target city. During fine-tuning, typically only the parameters of the later layers of the network are updated, while the parameters of the earlier layers remain unchanged. This is because the earlier layers of the neural network usually capture more general features, while the later layers are more task-specific.

[0084] Preferably, in one embodiment of the present invention, only the parameters of the last two layers of the network can be updated. This approach can fully adapt to the characteristics of the target domain while preserving knowledge of the source domain. Of course, the specific number of layers of parameters to be updated can be adjusted according to the similarity between the source and target domains.

[0085] Finally, this method updates the parameters of the target domain neural network model based on the gradient of the loss function. The gradient descent method is used here, and its update rule can be expressed as:

[0086]

[0087] Where, θ t Let L(θ) represent the parameters at the t-th iteration, η represent the learning rate, and L(θ) represent the learning rate.t ) represents the loss function. This represents the gradient operator.

[0088] Preferably, the learning rate η can be set to a small value, such as 0.001, to ensure that the model can converge stably. In practical applications, adaptive learning rate methods, such as the Adam optimizer, can also be used to further improve the training effect of the model.

[0089] Through the above steps, this method can effectively transfer knowledge from the source domain to the target domain while maintaining sensitivity to the characteristics of the target domain, thereby constructing a neural network model that can both utilize prior knowledge and adapt to the characteristics of the target city.

[0090] This invention employs a dynamic adjustment strategy during the transfer learning process to achieve optimal transfer of knowledge from the source domain.

[0091] First, this method calculates the difference in data distribution between the source and target domains. This step is crucial because it determines the extent to which knowledge from the source domain can be directly applied to the target domain. There are various methods for calculating the distribution difference; this invention preferably uses the Maximum Mean Discrepancy (MMD) method. The formula for calculating MMD is as follows:

[0092]

[0093] Where X and Y represent the data in the source and target domains, respectively, and φ is a feature mapping function. It is the regenerating nucleus Hilbert space.

[0094] Next, this method dynamically adjusts the degree of source domain knowledge transfer based on the calculated distribution differences. Specifically, when the distribution differences are small, more reliance can be placed on source domain knowledge; when the distribution differences are large, more reliance needs to be placed on target domain data. This dynamic adjustment strategy can be implemented through a weight coefficient λ, where the value of λ ranges from [0,1].

[0095] Finally, this method combines the loss functions of the source and target domains to optimize the parameters of the fusion model. The fused loss function can be expressed as:

[0096] L total =λL source +(1-λ)L target +γMMD 2 (X,Y)

[0097] Among them, L source and L target These are the loss functions for the source and target domains, respectively, and γ is a balancing parameter used to control the degree of influence of the distribution difference term.

[0098] Preferably, γ can be set to a value between 0.1 and 1, and the specific value can be determined through cross-validation. This design of the fusion loss function allows the model to fully utilize source domain knowledge while learning features of the target domain, thereby achieving effective knowledge transfer.

[0099] Through the above steps, this method can adaptively adjust the degree of knowledge transfer based on the specific circumstances of the source and target domains, thereby constructing a fusion model that can fully utilize prior knowledge and adapt well to the characteristics of the target city. Compared to simple transfer or complete retraining, this method can better balance knowledge transfer and target adaptation, thus improving the accuracy and reliability of settlement prediction.

[0100] This invention introduces a crucial weighting coefficient λ when dynamically adjusting the degree of knowledge transfer from the source domain. The appropriate setting of this coefficient is vital to the effectiveness of transfer learning.

[0101] First, this method defines the range of λ as [0,1]. When λ = 0, the model uses the target domain data entirely, which is equivalent to training a new model from scratch on the target domain; when λ = 1, the model uses the source domain data entirely, which is equivalent to directly applying the source domain model to the target domain. In practical applications, the optimal λ value usually lies between these two extremes.

[0102] Next, this method determines the optimal λ value through iterative optimization. Preferably, a grid search method can be used, which involves uniformly selecting a series of λ values ​​in the interval [0,1], such as {0,0.1,0.2,...,0.9,1}, then training a model for each λ value, and finally selecting the λ value that performs best on the validation set.

[0103] Another more efficient method is Bayesian optimization. This method intelligently selects the next value of λ to try by constructing a probabilistic model between λ and model performance, thus finding the optimal solution faster. The objective function of Bayesian optimization can be defined as:

[0104] f(λ) = -MSE val (λ)

[0105] Among them, MSE val (λ) represents the mean squared error of the model trained on the validation set using the value of λ.

[0106] Preferably, in one embodiment of the present invention, the number of iterations for Bayesian optimization can be set to 50, which is usually sufficient to find a good λ value. Of course, if computational resources permit, the number of iterations can be increased to obtain more accurate results.

[0107] Through the above steps, this method can find an optimal λ value, achieving precise control over the degree of source domain knowledge transfer. Compared to a fixed knowledge transfer strategy, this method can better adapt to the differences in geological conditions and engineering characteristics among different cities, thereby improving the generalization ability and prediction accuracy of the settlement prediction model.

[0108] The present invention also includes an important model evaluation step for verifying the predictive performance of the fusion model.

[0109] First, this method uses validation set data to evaluate the prediction accuracy of the fusion model. The validation set is a dataset independent of the training set, used to simulate the model's performance in real-world applications. Preferably, the total dataset can be divided into training, validation, and test sets in a 7:2:1 ratio. This division ensures sufficient training data while providing enough data for model validation and final testing.

[0110] Next, this method calculates evaluation metrics such as Mean Absolute Percentage Error (MAPE). The formula for calculating MAPE is:

[0111]

[0112] Among them, A i This is the actual value. is the predicted value, and n is the number of samples.

[0113] Preferably, in one embodiment of the present invention, a MAPE of less than 10% can be used as a standard for good model performance. This standard is set based on the actual needs of underground space engineering, because in most cases, an error within 10% is sufficient to meet the requirements of engineering design and risk assessment.

[0114] In addition to MAPE, this method can also calculate indicators such as root mean square error (RMSE) and coefficient of determination (R^2) to comprehensively evaluate the predictive performance of the model.

[0115] Finally, based on the evaluation results, this method determines whether further optimization of the fusion model is needed. If the model performance does not meet the preset standards, the following optimization strategies can be adopted:

[0116] 1. Adjust the network structure, such as increasing or decreasing the number of hidden layers, or changing the number of neurons in each layer.

[0117] 2. Try different activation functions, such as switching from ReLU to Leaky ReLU or ELU.

[0118] 3. Adjust the learning rate or adopt a learning rate decay strategy.

[0119] 4. Add regularization terms, such as L1 / L2 regularization or Dropout, to prevent overfitting.

[0120] 5. Increase training data or use data augmentation techniques.

[0121] Through this comprehensive model evaluation and optimization process, this method ensures that the final fusion model has good predictive performance and generalization ability, providing reliable technical support for urban underground space subsidence prediction.

[0122] The invention also includes an innovative inversion parameter step designed to further improve the model’s ability to capture settlement mechanisms.

[0123] First, this method uses historical settlement monitoring data to invert and obtain key parameters affecting settlement. These key parameters typically include soil compression index, secondary compression index, and prestressing. The inversion process can employ optimization algorithms, such as genetic algorithms or particle swarm optimization algorithms. Taking a genetic algorithm as an example, its objective function can be defined as:

[0124]

[0125] Where θ represents the parameter set to be inverted. It is the measured settlement value at the i-th time point. It is the theoretical settlement value calculated using parameter θ, and n is the total number of time points.

[0126] Preferably, in one embodiment of the present invention, the population size of the genetic algorithm can be set to 100 and the number of iterations to 500. These parameter settings can usually achieve a good balance between computational efficiency and optimization effect.

[0127] Next, this method uses the key parameters obtained from the inversion as additional features and inputs them into the fusion model. This step combines the physical model and the data-driven model, significantly improving the model's interpretability and prediction accuracy. Specifically, these parameters can be concatenated with the original feature vector to form a new input vector:

[0128] X new =[X original ,θ inverted ],

[0129] Among them, X original It is the original eigenvector, θ inverted It is the parameter vector obtained by inversion.

[0130] Finally, this method updates the fusion model to improve its ability to capture the sedimentation mechanism. This step can be achieved by retraining the model, but it should be noted that in order to maintain the model's sensitivity to the original features, a small learning rate, such as 0.0001, can be used.

[0131] Through the above steps, this method achieves an organic combination of physical model and data-driven model. It utilizes the flexibility and adaptability of data-driven methods while incorporating the prior knowledge of physical models, thereby constructing a more reliable and interpretable settlement prediction model.

[0132] The output steps of this invention include not only generating single-point settlement prediction results, but also generating more comprehensive settlement information.

[0133] First, this method generates a regional subsidence trend map of the target city. This trend map can visually display the subsidence distribution of the entire city or a specific area. Preferably, Kriging interpolation can be used to generate a continuous subsidence distribution map. The basic formula for Kriging interpolation is as follows:

[0134]

[0135] Among them, Z * (x0) is the estimated value Z(x) of the point to be predicted. i ) represents the observed value of a known point, and λ represents the observed value of that point. i These are weighting coefficients, satisfying...

[0136] Preferably, in one embodiment of the invention, ordinary kriging can be selected, and a spherical semivariogram model can be used. This setup performs well in most geostatistical applications.

[0137] Next, this method identifies areas with high settlement risk. This can be achieved by setting a settlement threshold. For example, areas with an annual settlement rate exceeding 20 mm / year can be marked as high-risk areas. This threshold is based on engineering experience, but can be adjusted according to specific circumstances in practical applications.

[0138] The algorithm for identifying high-risk areas can be represented as:

[0139]

[0140] Where R(x,y) represents the risk status of location (x,y), with 1 indicating high risk and 0 indicating low risk. S(x,y) represents the annual subsidence rate of location (x,y), S threshold It is the set threshold.

[0141] Finally, this method outputs settlement early warning information. The warning information may include the location, area, and maximum settlement rate of high-risk areas. Preferably, the warning information can be divided into three levels: yellow warning (annual settlement rate 20-30 mm), orange warning (30-50 mm), and red warning (>50 mm).

[0142] Through the above steps, this method not only provides detailed settlement prediction results, but also generates intuitive settlement trend maps and targeted risk warning information. These outputs can provide comprehensive decision support for urban planning, engineering design, and risk management.

[0143] The invention also includes an important model update step to ensure that the predictive model can maintain its accuracy and applicability over time.

[0144] First, this method acquires the latest settlement monitoring data for the target city periodically. The specific meaning of "periodically" may vary from project to project, but it can usually be set to monthly or quarterly.

[0145] Preferably, in one embodiment of the present invention, an automated data acquisition system, such as a distributed optical fiber sensor network, can be used to achieve real-time or near-real-time data updates.

[0146] Next, this method updates the parameters of the fusion model based on the latest data. This step employs the idea of ​​online learning, allowing the model to continuously adjust its parameters according to new data. Specifically, the parameter update can be performed using the mini-batch gradient descent method. The update formula is as follows:

[0147]

[0148] Where, θ t These are the current model parameters, η is the learning rate, L is the loss function, and X is... t:t+n and Y t:t+n These are the features and labels of the newly acquired n samples.

[0149] Preferably, the learning rate η can be set to a small value, such as 0.001, to ensure that the model can be updated stably without drastic fluctuations. Simultaneously, a learning rate decay strategy can be employed, gradually reducing the learning rate over time, such as:

[0150] The learning rate decay strategy can be expressed as:

[0151]

[0152] Where η0 is the initial learning rate, decay_rate is the decay rate, and t is the number of updates.

[0153] The performance degradation calculation formula in the performance monitoring mechanism is:

[0154]

[0155] If Performance_Drop exceeds a certain threshold, such as 10%, the model retraining process will be triggered.

[0156] Through this continuous updating and monitoring mechanism, this method ensures that the settlement prediction model maintains its accuracy and applicability, providing reliable technical support for the long-term safety management of urban underground spaces. Compared to static models, this dynamic updating method can better adapt to changes in geological conditions, engineering activities, and other factors, thus providing more accurate and timely settlement prediction results.

[0157] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for urban underground space settlement prediction based on transfer learning, characterized in that, include: The acquisition steps include: Obtain geological feature data, construction technology data, and historical settlement monitoring data for the target city; Acquire geological characteristic data, construction technology data, and historical settlement monitoring data of source cities with abundant settlement data; The processing steps include: Based on the source domain city data, a source domain neural network model is constructed and trained; Based on the target city data and the source domain neural network model, a target domain neural network model is constructed; Based on the target domain neural network model and the source domain neural network model, a transfer learning process is performed to obtain a fusion model; Output steps, including: Based on the fusion model, the predicted results of underground space subsidence for the target city are generated; The specific steps of constructing the target domain neural network model in the processing steps include: The structure of the source domain neural network model is copied as the initial target domain neural network model; The initial target domain neural network model was fine-tuned using the target city data. The parameters of the target domain neural network model are updated based on the gradient of the loss function. The transfer learning process in the processing steps specifically includes: Calculate the difference in data distribution between the source and target domains; Based on the aforementioned differences, the degree of source domain knowledge transfer is dynamically adjusted; The parameters of the fusion model are optimized by combining the loss functions of the source domain and the target domain. The degree of dynamic adjustment of source domain knowledge transfer specifically includes: Introduce a weighting coefficient λ, where 0 ≤ λ ≤ 1; When λ=0, the target domain data is used entirely; when λ=1, the source domain data is used entirely. The optimal value of λ is determined through iterative optimization.

2. The method of claim 1, wherein, The acquisition steps specifically include: The geological feature data, including soil stratification, thickness, and subsidence trend, are obtained through geological surveys. The construction process data, including support structure type, excavation method, and support type, is obtained through engineering records; The historical settlement monitoring data is obtained through on-site monitoring equipment, including the relative settlement between the measuring point and the horizontal benchmark, the cumulative settlement of the monitoring point, and the settlement rate.

3. The method of claim 1, wherein, The specific steps of constructing the source domain neural network model in the processing steps include: The source city data is preprocessed, including data cleaning, standardization, and outlier handling. Construct a deep neural network structure containing multiple hidden layers, each equipped with a non-linear activation function; The source domain neural network model is trained using the backpropagation algorithm until the loss function converges.

4. The method of claim 1, wherein, It also includes a model evaluation step: The prediction accuracy of the fusion model was evaluated using validation set data; Calculate evaluation indicators such as mean absolute percentage error; Based on the evaluation results, determine whether further optimization of the fusion model is needed.

5. The method of claim 1, wherein, It also includes the inversion parameter step: Based on the historical settlement monitoring data, the key parameters affecting settlement are obtained by inversion. The key parameters are input as additional features into the fusion model; The fusion model was updated to improve its ability to capture sedimentation mechanisms.

6. The method according to claim 1, characterized in that, The output step further includes: Generate a regional subsidence trend map for the target city; Identify areas with a high risk of subsidence; Output settlement early warning information.

7. The method of claim 1, wherein, It also includes the model update step: Regularly obtain the latest subsidence monitoring data for the target city; updating parameters of the fusion model based on the latest settlement monitoring data; re-executing the outputting step to generate an updated settlement prediction result.

Images