Subsidence prediction method and device, system and storage medium for subway line
The BiLSTM-Attention model, optimized by CEEMDAN multi-scale decomposition and the White Whale optimization algorithm, solves the problems of non-stationary feature processing and hyperparameter dependence in subway line settlement prediction, achieving high-precision and stable settlement prediction, and is suitable for multi-point prediction under complex geological conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WENLING CONSTRUCTION ENGINEERING QUALITY INSPECTION CO LTD
- Filing Date
- 2026-03-23
- Publication Date
- 2026-06-19
AI Technical Summary
Existing methods for predicting subway line settlement have shortcomings in areas such as handling non-stationary features, relying on manual adjustment of hyperparameters, insufficient utilization of time-series information, and weak ability to identify key stages, resulting in inaccurate prediction results and difficulty in promoting their application.
The CEEMDAN multi-scale decomposition technique is used to process the sedimentation time series data. The hyperparameters of the BiLSTM-Attention model are optimized by combining the Beluga optimization algorithm. Sedimentation prediction is performed through multi-channel time series input. The bidirectional long short-term memory network and attention mechanism are used to capture the bidirectional dependency relationship of the sedimentation process.
It significantly improves the accuracy and stability of subway line settlement prediction, enhances the model's generalization ability, and enables high-precision prediction under complex geological conditions and its application across points and sections.
Smart Images
Figure CN122241099A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of urban rail transit technology, specifically relating to a method, device, system, and storage medium for predicting subway line settlement. Background Technology
[0002] With the rapid development of urban rail transit construction, the construction and operation safety of subway lines in existing complex environments has become increasingly prominent. In particular, under the influence of external disturbances such as large-scale underground excavation, water conservancy renovation, or canal diversion projects, the structure and track foundation along the subway line are prone to long-term uneven settlement, posing a potential threat to structural safety and operational stability.
[0003] Subway line settlement prediction is a key technical aspect of ensuring the safe operation of urban rail transit, and its results directly affect the scientific validity of construction control, risk warning, and maintenance decisions. Currently, common settlement prediction methods mainly include empirical formula methods, numerical simulation methods, and machine learning methods.
[0004] Existing subway line settlement prediction technologies have the following main problems: Insufficient handling of non-stationary features: Traditional methods are mostly based on linear regression or single-scale time series models, which are difficult to effectively decompose the multi-frequency features in settlement data, resulting in prediction results that are not sensitive to short-term disturbances and do not accurately depict long-term trends.
[0005] Hyperparameters rely on manual tuning: Existing deep learning models such as LSTM and GRU rely on manual experience to set parameters such as the number of structural layers, the number of neurons, and the learning rate. This results in low optimization efficiency, a tendency to get stuck in local optima, and an impact on the model's generalization ability.
[0006] Insufficient utilization of temporal information: The one-way recursive network only considers historical data and cannot capture the bidirectional dependency relationship of the settlement sequence, resulting in incomplete correlation modeling of the three stages of "construction disturbance - settlement response - recovery stage".
[0007] Weak ability to identify key stages: Traditional models lack explicit attention mechanisms, making it difficult to identify the feature contribution of key time windows in the settlement process, which affects the reliability of predictions.
[0008] Limited generalization ability: Existing methods are only trained on a single monitoring point and lack prediction verification across points and sections, making it difficult to extend to different route locations or complex geological conditions. Summary of the Invention
[0009] To address the problems existing in the prior art, this invention provides a method, device, system, and storage medium for predicting subway line settlement.
[0010] To achieve the above objectives, the present invention provides the following solution: A method for predicting settlement of subway lines includes: Step S1: Preprocess the settlement time series data of the acquired subway line monitoring points; Step S2: Perform CEEMDAN multiscale decomposition on the preprocessed sedimentation sequence to obtain multiple IMF components and construct them as a multi-channel time series input; Step S3: Perform hyperparameter optimization on the sedimentation prediction model combining bidirectional long short-term memory network and attention mechanism to obtain the optimal parameters; Step S4: Input the multi-channel time-series input into the BiLSTM-Attention prediction model based on optimal parameters, and output the settlement prediction value at the corresponding time.
[0011] Preferably, in step S1, the settlement data undergoes preprocessing including outlier removal, missing value imputation, and normalization.
[0012] As a preferred option, in step S3, the beluga optimization algorithm is used to optimize the hyperparameters of the sedimentation prediction model that combines a bidirectional long short-term memory network with an attention mechanism to obtain the optimal parameters.
[0013] As a preferred embodiment, the method also includes: step S5, comparing and analyzing the predicted settlement curve with the measured curve, calculating the prediction accuracy index, and displaying the settlement change trend and early warning information in a graphical manner.
[0014] The present invention also provides a subway line settlement prediction device, comprising: The first processing module is used to preprocess the settlement time series data of the acquired subway line monitoring points; The second processing module is used to perform CEEMDAN multiscale decomposition on the preprocessed sedimentation sequence to obtain multiple IMF components and construct them as a multi-channel time series input. The third processing module is used to perform hyperparameter optimization on the sedimentation prediction model that combines bidirectional long short-term memory network with attention mechanism to obtain the optimal parameters; The fourth processing module is used to feed the multi-channel time-series input into the BiLSTM-Attention prediction model based on optimal parameters and output the settlement prediction value corresponding to the time.
[0015] Preferably, the first processing module performs preprocessing on the settlement data, including outlier removal, missing value imputation, and normalization.
[0016] As a preferred option, the third processing module uses the beluga optimization algorithm to optimize the hyperparameters of the sedimentation prediction model that combines a bidirectional long short-term memory network with an attention mechanism, and obtains the optimal parameters.
[0017] Preferably, it also includes: a fifth processing module, used to compare and analyze the predicted settlement curve with the measured curve, calculate the prediction accuracy index, and display the settlement change trend and early warning information in a graphical manner.
[0018] The present invention also provides a subway line settlement prediction system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a subway line settlement prediction method when executed by the processor.
[0019] The present invention also provides a storage medium storing a computer program, which executes a subway line settlement prediction method when running.
[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Significantly improved ability to model non-stationary signals: Multi-scale feature extraction is achieved through CEEMDAN decomposition, which effectively improves the model's accuracy and stability in modeling complex settlement time series data.
[0021] 2. Intelligent model parameter optimization: The White Whale optimization algorithm is used to realize automatic search and dynamic adjustment of hyperparameters, avoiding the uncertainty and inefficiency caused by manual parameter tuning.
[0022] 3. Bidirectional temporal and attention fusion: The BiLSTM structure captures the bidirectional dependencies of the sedimentation process, and the attention mechanism enhances the ability to identify key stages, making the model prediction more reliable.
[0023] 4. Significantly improved prediction accuracy and generalization performance: The model exhibits high R-values across multiple monitoring points. 2 With low RMSE, it has good engineering applicability.
[0024] 5. Significantly improved prediction accuracy and generalization performance: This invention can be embedded in a subway settlement monitoring platform to achieve integrated operation of data acquisition, prediction analysis and visualization, providing intelligent technical support for safety monitoring during subway construction and operation. Attached Figure Description
[0025] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0026] Figure 1 This is a flowchart of the subway line settlement prediction method according to an embodiment of the present invention. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0029] Example 1 like Figure 1 As shown, the present invention provides a method for predicting the settlement of subway lines, comprising: Step S1: Obtain the settlement time series data of the monitoring points of the subway line, and perform preprocessing on the settlement data, including outlier removal, missing value imputation and normalization. Step S2: Perform CEEMDAN multiscale decomposition on the preprocessed sedimentation sequence to obtain multiple IMF components and construct them as a multi-channel time series input; Step S3: Use the beluga optimization algorithm to optimize the hyperparameters of the sedimentation prediction model that combines bidirectional long short-term memory network with attention mechanism to obtain the optimal parameters; Step S4: Input the multi-channel time series input into the BiLSTM-Attention prediction model based on optimal parameters, and output the settlement prediction value at the corresponding time. Step S5: Compare and analyze the predicted settlement curve with the measured curve, calculate the prediction accuracy index, and display the settlement change trend and early warning information in a graphical manner.
[0030] As one embodiment of the present invention, in step S2, the preprocessed sedimentation sequence Perform CEEMDAN multiscale decomposition to obtain multiple intrinsic mode functions (IMFs). The calculation process is as follows: First generate Group white noise And construct a noise-assisted signal: in, Represents the original signal. For the first k The noise sequence used in the next perturbation. The noise intensity coefficient, This indicates the disturbance signal after the addition of noise.
[0031] right Perform first-order EMD decomposition to obtain first-order IMF components: in, For the first-order intrinsic mode function, This indicates the total number of noisy signals. This refers to the first-order component obtained by performing an empirical mode decomposition on the input signal. For the first The original signal without noise.
[0032] Calculate the first-order residual: Introducing adaptive noise into the residuals to construct the next-order input: Performing EMD on it yields the second-order IMF: The general formula is: The residual is updated as follows: in, Indicates the first The remaining amount after the second decomposition. This is the remaining sequence of the previous order. For the first The intrinsic mode functions extracted at order 1.
[0033] The final multi-scale decomposition sequence is obtained: Each IMF component is aligned according to time step to construct a multi-channel time series input, which is used for subsequent processing of the BiLSTM-Attention settlement prediction model.
[0034] In one embodiment of the present invention, in step S3, the hyperparameter set of the sedimentation prediction model combining a bidirectional long short-term memory network and an attention mechanism is denoted as... The hyperparameters include, but are not limited to: the number of neurons in the BiLSTM hidden layer, the number of BiLSTM layers, the learning rate, the batch size, the Dropout ratio, and the time window length; each hyperparameter All are limited to a given range of values. Inside. Among them, and They represent the first The lower and upper bounds of each hyperparameter, hyperparameter θj The search range is limited to the interval Inside.
[0035] Furthermore, step S3 includes: S3.1: Population Initialization Assume the beluga whale population size is The maximum number of iterations is The position vector of each beluga whale is mapped one-to-one with a set of hyperparameters to be optimized. The location of the beluga whale is indicated as follows: During initialization, an initial population is generated using a uniformly random method within each hyperparameter range: For hyperparameters that need to be integers, continuous values are mapped to feasible discrete values using rounding or integer functions.
[0036] S3.2: Fitness Function and Determination of Optimal Individual In the In the next iteration, for each beluga whale The corresponding hyperparameters It is configured into a BiLSTM-Attention model to learn from the training set and outputs settlement predictions on the validation set. .
[0037] Assume the number of samples in the validation set is The actual settlement value is Then the root mean square error (RMSE) is used as the fitness function: Alternatively, a comprehensive index can be used as the objective function, for example: in, These are the weighting coefficients; For the first In the nth iteration Objective function values for each candidate solution. This represents the corresponding hyperparameter vector. The root mean square error, The coefficient of determination.
[0038] Let the individual with the smallest fitness in the current iteration be: S3.3: Beluga Whale Location Update (Hyperparameter Search) The Beluga optimization algorithm updates the position through two behaviors: "hunting down prey" and "bubble web attack," while introducing random search to enhance global optimization capabilities.
[0039] Parameter update: In the In this iteration, the coefficient vector is calculated as follows: in, In order to be in A random vector that is uniformly distributed within an interval; These are control parameters that vary with the number of iterations. Indicates the current iteration This represents the maximum number of iterations.
[0040] Encircle the prey (development stage) ): when At that time, the beluga whale approaches the current global optimal solution. The updated formula is: in," " indicates element-wise multiplication, " represents the absolute value of each element.
[0041] Randomly search for prey (exploration phase) ): when At that time, a beluga whale was randomly selected from the population. Its update method is as follows: Bubble web spiral attack (local fine search): Introducing random numbers ,when Simulate the spiral ascent behavior of a beluga whale in a bubble network, and update the beluga whale's position in a spiral manner: in, To control the constant of the spiral shape, In order to be in Random numbers within the interval.
[0042] Boundary constraints and discrete mappings: For each dimension component of the beluga whale after update Perform boundary truncation: For integer or discrete hyperparameters, the mapping is performed as follows: Alternatively, specific values can be obtained through index mapping based on a pre-defined set of hyperparameter candidates.
[0043] S3.4: Termination Condition and Optimal Hyperparameter Output Repeat steps S32–S33 until any of the following conditions are met: The number of iterations reached the maximum value. ; The change in the optimal fitness value is less than a preset threshold in several consecutive iterations. .
[0044] Finally, the globally optimal beluga whale position vector is determined. The corresponding set of hyperparameters The optimal parameter configuration for the BiLSTM-Attention settlement prediction model is used for settlement prediction in the subsequent step S4.
[0045] As one embodiment of the present invention, in step S4, after obtaining the multi-channel timing input constructed after CEEMDAN decomposition, and the optimal hyperparameters obtained in step S3 Then, a BiLSTM-Attention network configured with these optimal parameters is used to generate the predicted settlement value for the corresponding time moment. Specifically, this includes: S4.1 Construction and Symbol Definition of Multi-Channel Timing Input Let the input sequence be: in, The length of the time window; This refers to the number of IMF components or channels. For the first The multi-channel input vector at time step 1; this sequence serves as the input to the BiLSTM-Attention network.
[0046] S4.2 Temporal Feature Extraction of Bidirectional LSTM Bidirectional LSTM consists of two parts: forward LSTM and backward LSTM, used to simultaneously extract the positive and negative dependencies of sedimentation sequences.
[0047] (1) Forward LSTM computation (from 1 to T) Forward LSTM at time step The hidden state is represented as: in, Indicates the forward LSTM at time... Hidden state For the current input, This is the hidden state from the previous moment. This represents the forward recursion function for Long Short-Term Memory (LSTM) units.
[0048] (2) Backward LSTM computation (from T to 1) same, Indicates the inverse LSTM at time... t The hidden state, For the current input, This is the hidden state for the next moment. This represents the backward recursive function for Long Short-Term Memory (LSTM) units.
[0049] (3) Two-way fusion output By concatenating the forward and backward hidden states, we obtain: in Forming a length of Bidirectional temporal feature sequences: S4.3 Attention Mechanism Key Moment Weight Calculation Attention mechanisms are used to identify key time steps in the sedimentation sequence that contribute the most to the prediction.
[0050] (1) Attention score is obtained by projecting hidden features. in; This is the attention layer weight matrix; For bias terms; ,and This represents intermediate attention.
[0051] (2) Attention weights are obtained by Softmax normalization. in, Indicates the first The importance of temporal characteristics for settlement prediction.
[0052] (3) Attention weighting yields the global feature vector. in, , which is the final feature representation at the sequence level.
[0053] S4.4 fully connected prediction layer outputs settlement prediction values The attention-weighted feature vector is input into the linear prediction layer to calculate the predicted settlement value: in; These are the output layer weights; For output layer bias; This represents the predicted settlement value for the corresponding time period.
[0054] The multi-channel temporal inputs generated by CEEMDAN decomposition are fed into a bidirectional long short-term memory network in chronological order. A forward LSTM extracts the forward evolution features of the sedimentation sequence from early to late time steps; a backward LSTM extracts the inverse correlation features of the sedimentation sequence from late time steps. By concatenating the forward and backward hidden states, a sequence feature representation containing bidirectional dependency information is obtained.
[0055] Subsequently, the bidirectional sequence features are input into the attention mechanism module. An attention score is calculated for the feature vector at each time step, and then Softmax normalization is performed to obtain the attention weight reflecting the importance of that time step. The attention weights are then used to perform a weighted summation of the feature vectors at all time steps to obtain a global feature vector representing the entire sedimentation sequence.
[0056] Finally, the global feature vector is input into the fully connected prediction layer, and the corresponding time-based settlement prediction value is calculated through linear transformation, thereby completing the settlement prediction process of the BiLSTM-Attention network structure of this invention.
[0057] This invention innovatively introduces the CEEMDAN algorithm for adaptive multi-scale decomposition of subway settlement sequences, effectively separating high-frequency disturbances from low-frequency trend components. This method overcomes the aliasing and noise leakage problems of traditional EEMD algorithms, providing clearer and more physically meaningful multi-scale input features for deep models.
[0058] This invention proposes an automatic hyperparameter optimization mechanism based on the Beluga Whale Optimization (BWO) algorithm, which can achieve the global optimal selection of model structure and learning parameters without human intervention, thereby improving model training efficiency and generalization performance.
[0059] This invention employs a composite structure combining a bidirectional long short-term memory network (BiLSTM) with an attention mechanism, which can simultaneously learn sequential information and focus on key settlement stages, thereby achieving high-precision dynamic prediction of complex subway settlement processes.
[0060] After completing the CEEMDAN multi-scale decomposition, this invention does not establish independent prediction models for each IMF component. Instead, it aligns and combines the multiple IMF components obtained from the decomposition according to time steps to construct a multi-dimensional temporal input feature, which is then input into a BiLSTM-Attention Network optimized by BWO, directly outputting the original settlement amount at the corresponding time. This approach avoids the error accumulation problem caused by predicting each IMF one by one and then reconstructing it, simplifies the prediction process, and improves the overall accuracy and temporal consistency of settlement prediction. Combined with the system's front-end visualization module, it can realize dynamic display and early warning output of the settlement time history.
[0061] Example 2 The present invention also provides a subway line settlement prediction device, comprising: The first processing module is used to preprocess the settlement time series data of the acquired subway line monitoring points; The second processing module is used to perform CEEMDAN multiscale decomposition on the preprocessed sedimentation sequence to obtain multiple IMF components and construct them as a multi-channel time series input. The third processing module is used to perform hyperparameter optimization on the sedimentation prediction model that combines bidirectional long short-term memory network with attention mechanism to obtain the optimal parameters; The fourth processing module is used to feed the multi-channel time-series input into the BiLSTM-Attention prediction model based on optimal parameters and output the settlement prediction value corresponding to the time.
[0062] As one embodiment of the present invention, the first processing module performs preprocessing on the settlement data, including outlier removal, missing value imputation, and normalization.
[0063] As one embodiment of the present invention, the third processing module uses the beluga optimization algorithm to optimize the hyperparameters of the sedimentation prediction model combining bidirectional long short-term memory network and attention mechanism to obtain the optimal parameters.
[0064] As one embodiment of the present invention, it further includes: a fifth processing module, used to compare and analyze the predicted settlement curve of the settlement prediction value with the measured curve, calculate the prediction accuracy index, and display the settlement change trend and early warning information in a graphical manner.
[0065] Example 3 The present invention also provides a subway line settlement prediction system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a subway line settlement prediction method when executed by the processor.
[0066] Example 4 The present invention also provides a storage medium storing a computer program, which executes a subway line settlement prediction method when running.
[0067] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for predicting settlement of subway lines, characterized in that, include: Step S1: Preprocess the settlement time series data of the acquired subway line monitoring points; Step S2: Perform CEEMDAN multiscale decomposition on the preprocessed sedimentation sequence to obtain multiple IMF components and construct them as a multi-channel time series input; Step S3: Perform hyperparameter optimization on the sedimentation prediction model combining bidirectional long short-term memory network and attention mechanism to obtain the optimal parameters; Step S4: Input the multi-channel time-series input into the BiLSTM-Attention prediction model based on optimal parameters, and output the settlement prediction value at the corresponding time.
2. The subway line settlement prediction method as described in claim 1, characterized in that, In step S1, the settlement data undergoes preprocessing, including outlier removal, missing value imputation, and normalization.
3. The subway line settlement prediction method as described in claim 2, characterized in that, In step S3, the beluga optimization algorithm is used to optimize the hyperparameters of the sedimentation prediction model that combines a bidirectional long short-term memory network with an attention mechanism to obtain the optimal parameters.
4. The subway line settlement prediction method as described in claim 3, characterized in that, It also includes: step S5, comparing and analyzing the predicted settlement curve with the measured curve, calculating the prediction accuracy index, and displaying the settlement change trend and early warning information in a graphical manner.
5. A subway line settlement prediction device, characterized in that, include: The first processing module is used to preprocess the settlement time series data of the acquired subway line monitoring points; The second processing module is used to perform CEEMDAN multiscale decomposition on the preprocessed sedimentation sequence to obtain multiple IMF components and construct them as a multi-channel time series input. The third processing module is used to perform hyperparameter optimization on the sedimentation prediction model that combines bidirectional long short-term memory network with attention mechanism to obtain the optimal parameters; The fourth processing module is used to feed the multi-channel time-series input into the BiLSTM-Attention prediction model based on optimal parameters and output the settlement prediction value corresponding to the time.
6. The subway line settlement prediction device as described in claim 5, characterized in that, The first processing module performs preprocessing on the settlement data, including outlier removal, missing value imputation, and normalization.
7. The subway line settlement prediction device as described in claim 6, characterized in that, The third processing module uses the beluga optimization algorithm to optimize the hyperparameters of the sedimentation prediction model that combines a bidirectional long short-term memory network with an attention mechanism, and obtains the optimal parameters.
8. The subway line settlement prediction device as described in claim 7, characterized in that, Also includes: The fifth processing module is used to compare and analyze the predicted settlement curve with the measured curve, calculate the prediction accuracy index, and display the settlement change trend and early warning information in a graphical manner.
9. A subway line settlement prediction system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program executing the metro line settlement prediction method as described in any one of claims 1 to 4 when run by the processor.
10. A storage medium, characterized in that, The storage medium stores a computer program, which executes the subway line settlement prediction method as described in any one of claims 1 to 4 when it runs.