An assembled subway station structure response prediction and risk assessment method

Abstract

The application provides an assembled subway station structure response prediction and risk assessment method, which acquires multi-source monitoring data such as stress, environmental variables and load variables of the assembled subway station structure, and adopts a sliding time window method to construct a supervised learning sample; the original non-stationary monitoring time series is subjected to VMD multi-scale decomposition to obtain a plurality of intrinsic mode function components and a residual trend item; the decomposition result is input into a CNN-LSTM-SE prediction model, local time sequence features are extracted through CNN, feature channel re-labeling is performed through an SE attention mechanism, time sequence dependence modeling is performed through LSTM, and a future 24-hour key structure response prediction value is output; then, the prediction value is compared with a normal working condition benchmark value and a critical working condition reference value, a risk proximity coefficient is calculated, and a risk grade is output. The application can improve the short-time prediction accuracy of the structure response of the assembled subway station structure, and realize the connection between the prediction result and the risk grading.

Description

Technical Field

[0001] This invention relates to the field of underground engineering structure health monitoring and safety assessment technology, and in particular to a method for predicting the response and assessing the risk of prefabricated subway station structures. Background Technology

[0002] With the development of urban rail transit construction, prefabricated subway stations have been widely adopted due to their high construction efficiency and controllable quality. During the construction and maintenance period from station construction to operation, temporary support systems such as anchor cables and diaphragm walls may experience performance degradation, or even localized failure, due to factors such as groundwater seepage, prestress relaxation, and material corrosion. Degradation of the support system alters the stress boundary conditions of the main structure, causing redistribution of internal forces and adjustments to deformation patterns, adversely affecting the long-term safety of the station. Therefore, accurately predicting the response of key structures in prefabricated subway stations based on long-term monitoring data and further assessing structural risk trends is of significant engineering importance.

[0003] Existing structural response prediction methods mainly include traditional time series models and deep learning models. Traditional time series models, such as autoregressive moving average models and grey prediction models, are usually used for extrapolation prediction of single monitoring indicators, but they are difficult to integrate the comprehensive influence of multiple factors such as seepage pressure, soil pressure, temperature, humidity, and rainfall on structural response. At the same time, their fitting ability to non-stationary and nonlinear monitoring data is limited, and their prediction accuracy is insufficient in local peak and stage fluctuation areas. Furthermore, it is difficult to establish a correspondence between the prediction results and the actual risk status of the structure.

[0004] Deep learning methods, such as Long Short-Term Memory networks and gated recurrent units, can improve the ability to learn temporal features, but they still have shortcomings in the monitoring of prefabricated subway station structures.

[0005] On the one hand, existing methods typically input the raw monitoring signal directly, failing to effectively separate high-frequency noise, mid-frequency fluctuations, and low-frequency trend components, thus affecting model learning efficiency and prediction stability;

[0006] On the other hand, existing models do not adequately consider the differences in importance of different monitoring indicators, making it difficult to adaptively enhance key sensitive features and suppress redundant noise features.

[0007] Existing methods mostly remain at the level of numerical prediction, lacking a mechanism to compare the prediction results with critical conditions such as anchor cable failure and stiffness degradation of diaphragm walls, making it difficult to form an engineering closed loop from prediction to risk classification.

[0008] Meanwhile, prefabricated subway stations are characterized by joint force transmission, arch effect, and sensitivity to stress in the connection area between the sidewall and the base slab. Parts such as the lower inner side of the right side wall, the inner and outer sides of the arch, and the tenon joints are particularly sensitive to support degradation and load changes. If the selection of prediction indicators and the risk assessment methods lack specificity, the prediction results may become disconnected from the actual weak points, affecting the engineering applicability of the risk assessment.

[0009] To address this, a method for predicting the structural response and assessing the risks of prefabricated subway stations is proposed. Summary of the Invention

[0010] In view of this, the present invention provides a method for predicting the structural response and assessing the risks of prefabricated subway stations, in order to solve or alleviate the technical problems existing in the prior art, and at least provide a beneficial option.

[0011] The technical solution of this invention is implemented as follows: A method for predicting the structural response and assessing the risk of prefabricated subway stations, comprising the following steps:

[0012] S1. Obtain multi-source monitoring data of prefabricated subway stations and construct supervised learning samples using the sliding time window method. The multi-source monitoring data includes the structural response history sequence, environmental variable sequence, and load variable sequence.

[0013] The structural response history sequence includes the inner stress of the lower part of the right wall, the inner stress of the arch crown, the outer stress of the arch crown, and the DET3 stress of the tenon joint; the environmental variable sequence includes air temperature, humidity, and rainfall; and the load variable sequence includes the seepage pressure and soil pressure at the arch crown. When constructing supervised learning samples using the sliding time window method, the window length... For a 24-hour period, the predicted step size is... The timeframe is 24 hours, meaning that multi-source monitoring data from the past 24 hours is used as input samples to predict key structural response indicators for the next 24 hours.

[0014] S2. Perform variational mode decomposition on the original non-stationary monitoring time series in the supervised learning samples to obtain multiple intrinsic mode function components and residual trend terms, and use the intrinsic mode function components and residual trend terms as multi-scale input features.

[0015] Specifically, the variational mode decomposition transforms the original non-stationary monitoring time series Decomposed into The method identifies eigenmode function components with different center frequencies and obtains the residual trend term based on the decomposition results; in this method... With a value of 6, in the decomposition results, IMF1 and IMF2 correspond to high-frequency disturbances and noise, IMF3 and IMF4 correspond to environmental coupling and short-term load fluctuations, and IMF5, IMF6 and residual trend terms correspond to mid-to-low frequency structural response adjustments and long-term trends.

[0016] S3. Construct a CNN-LSTM-SE prediction model, input the multi-scale input features into the convolutional neural network layer for local temporal feature extraction, input the feature map output by the convolutional neural network layer into the SE attention module for feature channel recalibration, and input the recalibrated features into the long short-term memory network layer for temporal dependency modeling.

[0017] Specifically, the convolutional neural network layer is a one-dimensional convolutional layer with a kernel size of 3, a kernel count of 64, a stride of 1, and an activation function of ReLU. A max-pooling layer with a pooling window of 2 is then applied after the one-dimensional convolutional layer. The SE attention module performs global average pooling on the feature map output by the convolutional neural network layer to obtain channel statistics. These statistics are then processed by two fully connected layers to generate channel weights, which are used to weight the feature map to obtain a recalibrated feature map. The long short-term memory network layer consists of two layers, each with 128 hidden units, and outputs predicted values ​​of key structural responses for the next 24 hours through a fully connected layer.

[0018] S4. Train the CNN-LSTM-SE prediction model using the historical monitoring dataset, and input the real-time collected multi-source monitoring data into the trained CNN-LSTM-SE prediction model to output the predicted values ​​of the key structure response for the next 24 hours.

[0019] Specifically, the historical monitoring dataset was divided into training, validation, and test sets in a ratio of 7:1.5:1.5. The Adam optimization algorithm was used for model training with an initial learning rate of 0.001, a batch size of 32, 200 training epochs, mean squared error as the loss function, a dropout rate of 0.2, and an early stopping mechanism was set up to stop training when the validation set loss did not decrease for 10 consecutive epochs.

[0020] S5. Compare the predicted response values ​​of the key structures with the baseline values ​​under normal operating conditions. Critical operating condition reference values Compare and calculate the risk proximity coefficient. The risk level of the prefabricated subway station is output based on the risk proximity coefficient λ.

[0021] Specifically, the critical condition reference value Rc is determined by a refined finite element numerical simulation under the asymmetric failure condition of the support system. The asymmetric failure condition of the support system includes the progressive failure critical condition A3 of the anchor cable and the asymmetric degradation critical condition B3 of the diaphragm wall. Among them, the progressive failure critical condition A3 of the anchor cable is the failure condition of the three rows of anchor cables on the right side, and the asymmetric degradation critical condition B3 of the diaphragm wall is the condition with a stiffness reduction of 80%.

[0022] The risk proximity coefficient λ is calculated according to the following formula:

[0023]

[0024] in, These are the predicted values ​​for the critical structural response. This is the baseline value under normal operating conditions. These are reference values ​​for critical operating conditions.

[0025] When λ < 0.5, output a safe state; when 0.5 ≤ λ < 0.8, output a state of concern; when λ ≥ 0.8, output a state of warning.

[0026] The embodiments of the present invention have the following advantages due to the adoption of the above technical solutions:

[0027] First, this invention uses VMD to decompose the original non-stationary monitoring time series into multiple scales, distinguishing high-frequency disturbances, mid-frequency fluctuations, and low-frequency trend components. This enables subsequent models to learn the response patterns at different physical scales, reducing the learning difficulty of directly processing the original aliased signals and improving the stability and accuracy of short-term prediction of structural responses.

[0028] Second, this invention combines CNN, LSTM and SE attention mechanism. CNN is used to extract local temporal features, LSTM is used to capture temporal dependencies in the monitoring sequence, and SE attention mechanism is used to adaptively weight different feature channels, so that the model can enhance the expressive ability of key sensitive features, while suppressing the interference of redundant noise features on the prediction results.

[0029] Third, this invention compares the predicted critical structural response values ​​with the normal operating condition reference value R0 and the critical operating condition reference value Rc, and classifies them into three levels: safety, concern, and early warning through the risk proximity coefficient λ, so that the prediction results can be connected with the critical operating condition of support degradation, thereby realizing the transformation from "numerical prediction" to "risk discrimination".

[0030] Fourth, this invention selects the inner stress of the lower part of the right wall, the inner stress of the arch, the outer stress of the arch, and the DET3 stress of the tenon joint as key structural response indicators, which can correspond to key force transmission areas such as the side wall, arch, and tenon joint of the prefabricated subway station, so that the prediction results and risk assessment results have a clear engineering orientation.

[0031] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 The image shows the optimized CNN-LSTM validation set results.

[0034] Figure 2 The image shows the LSTM test results.

[0035] Figure 3 The image shows the test results for CNN-GRU.

[0036] Figure 4 The image shows the test results for CNN-LSTM.

[0037] Figure 5 The image shows the test results of a traditional CNN-LSTM-SE.

[0038] Figure 6 A graph showing the changes in each mode of VMD in the monitoring sequence;

[0039] Figure 7 The image shows the results of the multi-scale time-domain decomposition of stress sequences based on VMD.

[0040] Figure 8 This is a schematic diagram of the Squeeze-and-Excitation attention mechanism.

[0041] Figure 9 The structure diagram of the CNN-LSTM-SE model with VMD. Detailed Implementation

[0042] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0043] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0044] like Figure 1-9As shown in the figure, this invention provides a method for predicting the structural response and assessing the risk of prefabricated subway stations. The method includes steps such as multi-source monitoring data acquisition, supervised learning sample construction, VMD multi-scale feature decomposition, CNN-LSTM-SE prediction model construction, model training and prediction, and risk trend assessment. This method performs multi-scale decomposition and deep learning prediction on long-term monitoring data of prefabricated subway stations, and compares the prediction results with critical operating condition reference values ​​to achieve structural response prediction and risk level determination.

[0045] Example 1

[0046] This embodiment uses Jiudingshan Station on Qingdao Metro Line 6 as an example. This station is a high-arch prefabricated metro station, with the Xinzhuang North River adjacent to the right side of the foundation pit, exhibiting complex hydrogeological conditions. This makes it suitable for verifying the application effect of the method of this invention in predicting the structural response and assessing the risks of prefabricated metro stations.

[0047] Multi-source monitoring data acquisition and sample construction:

[0048] This embodiment collects long-term monitoring data from Jiudingshan Station of Qingdao Metro Line 6 from January 2023 to December 2023. The sampling frequency is 1 hour / time, and a total of 8760 sets of valid monitoring data are obtained.

[0049] The multi-source monitoring data includes structural stress data, environmental variable data, and load variable data. The structural stress data includes the stress on the inner side of the lower part of the right wall, the stress on the inner side of the arch crown, the stress on the outer side of the arch crown, and the stress at the key measuring point DET3 on the tenon. Specifically, the stress on the inner side of the lower part of the right wall corresponds to measuring point C3, the stress on the inner side of the arch crown corresponds to measuring point E1, and the stress on the outer side of the arch crown corresponds to measuring point E1′. Environmental variable data includes air temperature, humidity, and rainfall. Load variable data includes seepage pressure and soil pressure at the arch crown.

[0050] In the sample construction process, a sliding time window method is used to construct supervised learning samples. The window length L is set to 24 hours, and the prediction step size H is set to 24 hours. That is, multi-source monitoring data from the past 24 hours are used as input samples to predict key structural response indicators for the next 24 hours.

[0051] The input sample matrix is ​​represented as follows: ;

[0052] in, Indicates the first One input sample, Indicates the length of the time window. This represents the feature dimension. In this embodiment, L=24, and m corresponds to the feature dimension formed by the structural stress data, environmental variable data, and load variable data.

[0053] The 8760 sets of valid monitoring data were divided into training, validation, and test sets in a ratio of 7:1.5:1.5. The training set consisted of 6132 sets, the validation set of 1314 sets, and the test set of 1314 sets.

[0054] Multi-scale feature decomposition based on VMD:

[0055] Long-term monitoring sequences of prefabricated subway stations typically exhibit non-stationary and multi-scale coupling characteristics. To separate high-frequency disturbances, mid-frequency fluctuations, and low-frequency trend components from the original monitoring signals, this embodiment employs variational mode decomposition to preprocess the original non-stationary monitoring time series.

[0056] The original non-stationary monitoring time series is denoted as... The input is then processed by the VMD module. VMD will... Decomposed into Eigenmode function components with different center frequencies And obtain the residual trend term. The components of each intrinsic mode function satisfy the following reconstruction relation:

[0057] In the VMD framework, each modal component is assumed to be a narrowband signal, revolving around its corresponding center frequency. Oscillation. To characterize the bandwidth of each modal component, the modal components are first... Perform a Hilbert transform to convert it into an analytic signal form:

[0058] in, Let be the impulse function. The imaginary unit, This represents the convolution operation.

[0059] Subsequently, we will analyze the signal and the exponential function. Multiplying these components shifts their spectrum to near the fundamental frequency, and the modal component bandwidth is measured by calculating the gradient norm. This leads to the following constrained variational problem:

[0060]

[0061] The constraints are:

[0062]

[0063] This constrained variational problem states that, under the condition that the linear superposition of each modal component can reconstruct the original monitoring time series, the bandwidth of each modal component should be minimized to achieve the separation of signals in different frequency bands.

[0064] To solve the above constrained optimization problem, Lagrange multipliers are introduced. and secondary penalty factor Construct the augmented Lagrangian function:

[0065]

[0066] in, For the inner product of Lagrange multipliers.

[0067] Using the alternating direction multiplier method , and Perform iterative updates until the convergence condition is met.

[0068] In this embodiment, the number of modes Set to 6, penalty factor The value is set to 2000. After VMD decomposition, six intrinsic mode function components (IMF1 to IMF6) and one residual trend term are obtained.

[0069] like Figure 6 and Figure 7 As shown, IMF1 and IMF2 mainly correspond to high-frequency disturbances and noise, IMF3 and IMF4 mainly correspond to environmental coupling and short-term load fluctuations, and IMF5, IMF6, and the residual trend term mainly correspond to mid-to-low frequency structural response adjustments and long-term trends. These intrinsic mode function components and residual trend terms serve as parallel inputs to subsequent prediction models.

[0070] CNN-LSTM-SE prediction model construction:

[0071] like Figure 8 and Figure 9 As shown, this embodiment constructs a CNN-LSTM-SE prediction model. The CNN-LSTM-SE prediction model includes a CNN layer, an SE attention layer, an LSTM layer, and an output layer.

[0072] First, the modal components decomposed by VMD are input into a one-dimensional convolutional layer, which extracts local temporal features. In this embodiment, the kernel size of the one-dimensional convolutional layer is 3, the number of kernels is 64, the stride is 1, and the activation function is ReLU.

[0073] The convolution operation is represented as:

[0074] in, For convolution output, For activation function, For convolution kernel weights, The data at the corresponding position in the input sequence. This is a bias term.

[0075] After completing the one-dimensional convolution process, a max pooling layer is applied with a pooling window of 2 to reduce the feature dimension and retain the main local response features.

[0076] Then, the feature map output from the CNN layer is input into the SE attention layer. Let the feature tensor after processing by the convolutional neural network be represented as:

[0077] in, The number of feature channels, This represents the length of the time series.

[0078] The SE attention layer first compresses each channel using global average pooling to extract the global statistical features of the channels, expressed as follows:

[0079]

[0080] in, For the first Statistics for each characteristic channel For the first Each feature channel in The characteristic value at time step.

[0081] Subsequently, a non-linear dependency relationship between channels is established through two layers of fully connected mapping:

[0082]

[0083] in, , is the channel statistical vector; , , Compression ratio; Represents the ReLU activation function; This represents the Sigmoid function.

[0084] In this embodiment, the compression ratio r = 16. The final weight vector is represented as follows:

[0085]

[0086] The weight vector represents the importance coefficient of each feature channel.

[0087] After completing the weight calculation, the original features are recalibrated via channels:

[0088]

[0089] in, For the recalibrated first Each channel features. Through the above channel recalibration, the model can enhance the expressive power of key feature channels and suppress the influence of redundant or noisy features.

[0090] Next, the recalibrated features are input into the LSTM layer. The LSTM layer controls the information flow through forget gates, input gates, and output gates to capture temporal dependencies in the monitored sequence. The core operations of the LSTM unit include:

[0091]

[0092]

[0093]

[0094]

[0095] in, For the Gate of Oblivion For input gate, In cellular state, In hidden state, Enter the current time. This is the hidden state from the previous moment. Candidate cell state, This is the output gate.

[0096] In this embodiment, the LSTM layer has two layers, with 128 hidden units in each layer. The temporal features output by the LSTM layer are processed by a fully connected layer to output the predicted values ​​of the critical structural responses for the next 24 hours.

[0097] In this embodiment, the model training parameters are shown in Table 1 below.

[0098] Table 1 Model training parameters

[0099]

[0100] During model training, the Adam optimization algorithm was used with an initial learning rate of 0.001, a batch size of 32, and 200 training epochs. The mean squared error (MSE) loss function was used. A Dropout rate of 0.2 was set during training, and an early stopping mechanism was introduced. Model training was stopped when the validation set loss did not decrease for 10 consecutive epochs.

[0101] After training, the real-time collected multi-source monitoring data is input into the trained CNN-LSTM-SE prediction model to obtain the predicted values ​​of the key structural responses for the next 24 hours.

[0102] V. Scale Setting and Performance Evaluation

[0103] To verify the prediction performance of the method in this embodiment, four comparative models were set up. Comparative model 1 is an LSTM model, comparative model 2 is a CNN-GRU model, comparative model 3 is a CNN-LSTM model, and comparative model 4 is a CNN-LSTM-SE model. The settings of each comparative model and the model in this embodiment are shown in Table 2 below.

[0104] Table 2 Comparative Model Settings

[0105]

[0106] The mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R²) were used as evaluation metrics. The predictive performance of each model on the test set is shown in Table 3.

[0107] Table 3 Comparison of Predictive Performance of Each Model

[0108]

[0109] As shown in Table 3, the optimized CNN-LSTM-SE model used in this embodiment outperforms the comparative model in terms of MAE, MSE, RMSE, and R². Specifically, the model in this embodiment has an MAE of 0.0143, an MSE of 0.0003, an RMSE of 0.0161, and an R² of... 2 It is 0.942.

[0110] Compared to the CNN-LSTM-SE model without VMD preprocessing, this embodiment decomposes the original non-stationary monitoring signal into modal components of different scales through VMD, enabling the model to learn the structural response variation patterns under different frequency bands. Compared to the CNN-LSTM model without SE attention mechanism, this embodiment improves the expressive power of key channel features by adaptively weighting different feature channels through the SE attention layer.

[0111] Figure 1 The optimized CNN-LSTM validation set results are shown. Figures 2 to 5 The test results for LSTM, CNN-GRU, CNN-LSTM, and the traditional CNN-LSTM-SE model are shown respectively. As can be seen from Table 3, the model in this embodiment can improve the accuracy and stability of predicting the response of key structures in prefabricated subway stations.

[0112] After obtaining the predicted value of the critical structural response for the next 24 hours, this embodiment compares the predicted value with the normal operating condition baseline value R0 and the critical operating condition reference value Rc, calculates the risk proximity coefficient λ, and outputs the risk level based on the value of λ.

[0113] In this embodiment, the stress on the inner side of the lower part of the right wall is used as an example for explanation. Normal operating condition reference value. -12.5 MPa is taken as the average value during the monitoring period; critical operating condition reference value. Take -18.0MPa, which corresponds to the simulated value of this measuring point under the condition that the stiffness of the diaphragm wall is reduced by 80%.

[0114] Risk proximity coefficient Calculate using the following formula:

[0115]

[0116] in, These are the predicted values ​​for the critical structural response. This is the baseline value under normal operating conditions. These are reference values ​​for critical operating conditions.

[0117] The risk level classification rules are as follows:

[0118] When λ < 0.5, it is considered a safe state;

[0119] When 0.5 ≤ λ < 0.8, it is determined to be in a state of concern;

[0120] When λ≥0.8, it is determined to be a warning state.

[0121] The stress on the inner side of the lower part of the right wall in the last 30 days of the test set was predicted in a rolling 24-hour forecast, and the risk was classified according to the risk proximity coefficient mentioned above. The results are shown in Table 4.

[0122] Table 4 Risk Trend Assessment Results

[0123]

[0124] As shown in Table 4, as the predicted stress value of the lower inner side of the right wall gradually approaches the critical working condition reference value... Risk proximity coefficient As the risk level gradually increases, it changes from a safe state to a state of concern, and then further to a state of warning. Therefore, this embodiment can identify the trend of the structural response approaching a critical state based on the structural response prediction results and output the corresponding risk level.

[0125] Critical operating condition reference value in this embodiment The conditions were determined by refined finite element numerical simulation under asymmetric failure conditions of the support system. These asymmetric failure conditions include the progressive failure critical condition A3 for anchor cables and the asymmetric degradation critical condition B3 for the diaphragm wall. Specifically, the progressive failure critical condition A3 for anchor cables is the failure of the three rows of anchor cables on the right side, and the asymmetric degradation critical condition B3 for the diaphragm wall is the condition with an 80% reduction in stiffness.

[0126] Implementation results:

[0127] This embodiment uses VMD to decompose the original non-stationary monitoring sequence at multiple scales, extracts local temporal features through a CNN layer, recalibrates different feature channels through an SE attention layer, models temporal dependencies through an LSTM layer, and outputs predicted values ​​of critical structural responses for the next 24 hours through a fully connected layer. Subsequently, the predicted values ​​are compared with the normal operating condition baseline value R0 and the critical operating condition reference value Rc to calculate the risk proximity coefficient λ, and output the safe status, the attention status, or the warning status.

[0128] As shown in Table 3, the MAE of the model in this embodiment is 0.0143, the MSE is 0.0003, the RMSE is 0.0161, and the R... 2 The accuracy is 0.942, indicating that the predictive performance is better than the comparative model. As shown in Table 4, this embodiment can identify the trend of risk level changes based on the changes in the predicted stress value of the lower inner side of the right wall, thus achieving a closed loop from structural response prediction to risk trend assessment.

[0129] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the structural response and assessing the risk of prefabricated subway stations, characterized in that: Includes the following steps: S1. Obtain multi-source monitoring data of prefabricated subway stations and construct supervised learning samples using the sliding time window method. The multi-source monitoring data includes the structural response history sequence, environmental variable sequence, and load variable sequence. S2. Perform variational mode decomposition on the original non-stationary monitoring time series in the supervised learning samples to obtain multiple intrinsic mode function components and residual trend terms, and use the intrinsic mode function components and residual trend terms as multi-scale input features; S3. Construct a CNN-LSTM-SE prediction model, input the multi-scale input features into the convolutional neural network layer for local temporal feature extraction, input the feature map output by the convolutional neural network layer into the SE attention module for feature channel recalibration, and input the recalibrated features into the long short-term memory network layer for temporal dependency modeling. S4. Train the CNN-LSTM-SE prediction model using the historical monitoring dataset, and input the real-time collected multi-source monitoring data into the trained CNN-LSTM-SE prediction model to output the predicted value of the key structure response for the next 24 hours. S5. Compare the predicted response values ​​of the key structures with the baseline values ​​under normal operating conditions. Critical operating condition reference values Compare and calculate the risk proximity coefficient. And based on the aforementioned risk proximity coefficient Output the risk level of prefabricated subway stations.

2. The method for predicting the structural response and assessing the risk of prefabricated subway stations according to claim 1, characterized in that, In step S1, the structural response history sequence includes the inner stress of the lower part of the right wall, the inner stress of the arch, the outer stress of the arch, and the DET3 stress of the tenon joint. The environmental variable series includes temperature, humidity, and rainfall; The load variable sequence includes crown seepage pressure and crown soil pressure.

3. The method for predicting the structural response and assessing the risk of prefabricated subway stations according to claim 1, characterized in that, In step S1, the window length of the sliding time window method For a 24-hour period, the predicted step size is... The input sample matrix of the supervised learning samples is represented as follows: (The time frame is 24 hours.) ; in, The time window length, For feature dimensions.

4. The method for predicting the structural response and assessing the risk of prefabricated subway stations according to claim 1, characterized in that, In step S2, the variational mode decomposition transforms the original non-stationary monitoring time series Decomposed into The intrinsic mode function components have different center frequencies, and the residual trend term is obtained based on the decomposition results. The intrinsic mode function components satisfy the linear superposition to reconstruct the original non-stationary monitoring time series.

5. The method for predicting the structural response and assessing the risk of prefabricated subway stations according to claim 4, characterized in that, The The value is 6, the The intrinsic mode function components include IMF1 to IMF6, where IMF1 and IMF2 correspond to high-frequency disturbances and noise, IMF3 and IMF4 correspond to environmental coupling and short-term load fluctuations, and IMF5, IMF6 and the residual trend term correspond to mid-to-low frequency structural response adjustments and long-term trends.

6. The method for predicting the structural response and assessing the risk of prefabricated subway stations according to claim 1, characterized in that, In step S3, the convolutional neural network layer is a one-dimensional convolutional layer with a kernel size of 3, a kernel count of 64, a stride of 1, and an activation function of ReLU. A max pooling layer is then connected after the one-dimensional convolutional layer, and the pooling window of the max pooling layer is 2.

7. The method for predicting the structural response and assessing the risk of prefabricated subway stations according to claim 1, characterized in that, In step S3, the SE attention module processes the feature map output by the convolutional neural network layer. Channel recalibration includes: Perform global average pooling on each feature channel to obtain the channel statistics z; The channel statistics z are input into a two-layer fully connected network, and channel weights s are generated using a compression ratio r=16; Based on the channel weights s, the feature map Weighting is performed to obtain the recalibrated feature map.

8. The method for predicting the structural response and assessing the risk of prefabricated subway stations according to claim 1, characterized in that, In step S3, the long short-term memory network layer has two layers, with 128 hidden units in each layer. The long short-term memory network layer performs time-dependent modeling on the recalibrated features through forget gate, input gate and output gate, and outputs the predicted value of the key structural response for the next 24 hours through a fully connected layer.

9. The method for predicting the structural response and assessing the risk of prefabricated subway stations according to claim 1, characterized in that, In step S4, the historical monitoring dataset is divided into a training set, a validation set, and a test set in a ratio of 7:1.5:1.

5. The CNN-LSTM-SE prediction model is trained using the Adam optimization algorithm with an initial learning rate of 0.001, a batch size of 32, 200 training epochs, a loss function of mean squared error, a dropout rate of 0.2, and an early stopping mechanism that stops training when the validation set loss does not decrease for 10 consecutive epochs.

10. The method for predicting the structural response and assessing the risk of prefabricated subway stations according to claim 1, characterized in that, In step S5, the critical operating condition reference value The asymmetric failure conditions of the support system were determined by refined finite element numerical simulation under the asymmetric failure conditions of the support system, including the critical condition of progressive failure of anchor cables A3 and the critical condition of asymmetric degradation of diaphragm walls B3. The progressive failure critical condition A3 of the anchor cable is the failure condition of the three rows of anchor cables on the right side, and the asymmetric degradation critical condition B3 of the diaphragm wall is the condition of stiffness reduction of 80%. The risk proximity coefficient Calculate using the following formula: in, These are the predicted values ​​for the critical structural response. This is the baseline value under normal operating conditions. This is a reference value for critical operating conditions; When λ < 0.5, output a safe state; When 0.5≤λ<0.8, output the attention status; when λ≥0.8, output the warning status.

Images