Underground pipe-jacking tunnel excavation pavement settlement prediction system and prediction method thereof

Abstract

The invention relates to the technical field of underground engineering construction safety monitoring, and discloses an underground pipe-jacking tunnel excavation pavement settlement prediction system and a prediction method thereof.The prediction system comprises an acquisition terminal and a computing server and is provided with a data preprocessing module, a model building module and other modules. The core lies in that a CA-CNN-LSTM hybrid neural network is constructed, geological spatial features are extracted by using parallel convolutional neural network branches, construction time sequence features are extracted by using bidirectional long short-term memory network branches, and deep fusion of spatial and temporal features is realized through a Cross-Attention module and a dynamic gating unit. The prediction method relates to multi-source data cleaning enhancement, model training and double-index grading early warning. According to the method, the problem of heterogeneous feature coupling is effectively solved, the settlement prediction precision is effectively improved, hierarchical control signals can be generated according to the prediction result, and the construction safety is effectively guaranteed.

Description

Technical Field

[0001] This invention relates to the field of underground engineering construction safety monitoring technology, specifically to a road surface settlement prediction system and prediction method for underground pipe jacking tunnel excavation. Background Technology

[0002] During the construction of underground pipe jacking tunnels, soil disturbance can cause road surface settlement, posing a threat to the safety of the surrounding environment. With the development of artificial intelligence technology, settlement prediction models based on deep learning are gradually replacing traditional empirical formulas and becoming an important means of engineering monitoring. However, traditional single neural network models have inherent limitations in dealing with such complex spatiotemporal coupling problems: convolutional neural networks are good at extracting spatial features, but they are difficult to capture long-term temporal dependencies; while long short-term memory networks perform well in processing time series data, they have limited ability to extract deep spatial features.

[0003] To overcome the limitations of single models, existing technical solutions attempt to combine different types of networks. However, most current hybrid prediction models use simple feature concatenation or chaining to process spatial and temporal data, failing to fully explore the inherent nonlinear relationships between spatiotemporal features. Due to the lack of effective dynamic feature selection and deep fusion mechanisms, these models are susceptible to interference from noise in field monitoring data. When faced with complex and variable geological conditions and unstable construction conditions, the models have insufficient generalization ability, poor stability, and difficulty in adapting to the differentiated needs of different engineering scenarios.

[0004] Furthermore, due to limitations in feature extraction and fusion capabilities, existing technologies have low sensitivity to abnormal settlement events, and prediction results often exhibit lag. This current technological state makes it difficult to meet the requirements of practical engineering projects for high-precision, high-robustness, and real-time early warning of settlement monitoring, posing safety risks to underground engineering construction. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a road surface settlement prediction system and method for underground pipe jacking tunnel excavation, which solves the technical problem that existing models cannot effectively integrate static geological spatial features and dynamic construction time sequence features, resulting in low road surface settlement prediction accuracy and insufficient generalization ability under complex working conditions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a road surface settlement prediction system and prediction method for underground pipe jacking tunnel excavation.

[0007] In a first aspect, the present invention provides a road surface settlement prediction system for underground pipe jacking tunnel excavation, including a data management module, a data preprocessing module, a model building module, a training and optimization module, and an online monitoring and early warning module;

[0008] The data management module is configured to receive multi-source heterogeneous data from the data acquisition terminal and generate a structured data table;

[0009] The data preprocessing module is configured to perform cleaning and feature enhancement operations on the structured data table;

[0010] The model building module is configured to establish a CA-CNN-LSTM hybrid neural network architecture, which includes a parallel spatial feature extraction branch and a temporal feature extraction branch, as well as a Cross-Attention feature fusion module connecting the outputs of the two branches.

[0011] The training optimization module is configured to update the weight parameters of the CA-CNN-LSTM hybrid neural network architecture using the backpropagation algorithm;

[0012] The online monitoring and early warning module is configured to generate graded warning signals based on the predicted values ​​output by the model and display them on the display terminal.

[0013] By adopting the above technical solutions, the system constructs parallel feature extraction channels, which can separately process the spatial features of geological conditions and the temporal features of construction parameters, avoiding feature expression conflicts caused by direct splicing. By using the Cross-Attention feature fusion module, a nonlinear mapping relationship between spatial features and temporal features is established in the deep feature space, which solves the problem of low coupling degree of multi-source heterogeneous features in traditional models and improves the accuracy of settlement prediction.

[0014] Preferably, the spatial feature extraction branch consists of a cascaded one-dimensional convolutional layer, a pooling layer, and a global average pooling layer, used to output a high-dimensional feature vector with implicit spatial correlation.

[0015] The temporal feature extraction branch is composed of a bidirectional long short-term memory network, which is used to output a temporal feature vector containing dynamic temporal information;

[0016] The Cross-Attention feature fusion module is internally divided into three logical units: a feature projection unit, a cross-attention calculation unit, and a dynamic gating fusion unit.

[0017] The feature projection unit is used to map the input features into a query vector, a key vector, and a value vector.

[0018] The cross-attention calculation unit is used to calculate the correlation between the query vector and the key vector and to weight the value vector;

[0019] The dynamic gating fusion unit is used to generate dynamic gating signals and perform feature mixing operations.

[0020] By adopting the above technical solutions, local spatial correlations are extracted using convolution operations, and forward and backward dependencies of time series are extracted using bidirectional loop structures. By projecting features of different dimensions onto a unified feature dimension, a data foundation is provided for subsequent feature interactions.

[0021] Preferably, the feature projection unit maps the temporal feature vector output by the temporal feature extraction branch to a query vector, and maps the high-dimensional feature vector output by the spatial feature extraction branch to a key vector and a value vector;

[0022] The cross-attention calculation unit calculates the correlation between the query vector and the key vector and weights the value vector;

[0023] The dynamic gating fusion unit generates a dynamic gating signal based on the temporal feature vector, and uses this signal to perform a weighted summation of the temporal feature vector and the attention features output by the cross-attention calculation unit.

[0024] By adopting the above technical solution, the correlation between the temporal variation characteristics of construction parameters and geological spatial characteristics is calculated and aggregated based on these characteristics. The dynamic gating mechanism, through adaptive weight allocation, retains the original construction temporal trend while introducing correlation-weighted geological spatial characteristic information, effectively balancing the influence of historical temporal information and current spatial characteristics on settlement prediction.

[0025] Preferably, the data preprocessing module integrates a quartile distance statistical algorithm, a cubic spline interpolation algorithm, and a feature space scaling transformation algorithm;

[0026] The interquartile range statistical algorithm is used to eliminate discrete outlier values;

[0027] The cubic spline interpolation algorithm is used to resample multi-source data to a uniform temporal resolution;

[0028] The feature space scaling transformation algorithm is used to apply a linear transformation to the feature vector under physical boundary constraints.

[0029] By adopting the above technical solutions, the alignment of the input model data in the time dimension and the stability of its numerical values ​​are guaranteed; the feature space scaling transformation enhances the model's generalization ability under small sample conditions and reduces the risk of overfitting by generating expanded samples within a reasonable physical range.

[0030] Preferably, the online monitoring and early warning module stores first-level, second-level, and third-level classification thresholds and is configured with comparison logic to compare the predicted cumulative settlement and the predicted settlement rate with the classification thresholds respectively, thereby generating corresponding first-level emergency stop, second-level adjustment, or third-level attention control signals.

[0031] By adopting the above technical solution, the numerical values ​​output by the model are converted into corresponding hierarchical control commands, realizing a closed loop from data prediction to engineering decision-making and ensuring construction safety.

[0032] Secondly, this invention provides a method for predicting road surface settlement during underground pipe jacking tunnel excavation, the method comprising:

[0033] Obtain geological condition data, pipe jacking construction parameters, environmental data, and historical settlement monitoring data;

[0034] The raw data undergoes structured cleaning, normalization, spatiotemporal alignment, and data augmentation.

[0035] A hybrid model containing convolutional neural network branches and bidirectional long short-term memory network branches is constructed, and spatial and temporal features are fused using a cross-attention mechanism;

[0036] Update the parameters of the hybrid model using the training dataset;

[0037] The trained model is used to output predicted road surface settlement values ​​and generate control signals.

[0038] By adopting the above technical solutions, the entire process implementation path from data acquisition to final prediction and control was clarified, ensuring the real-time nature and accuracy of settlement prediction.

[0039] Preferably, the normalization process and spatiotemporal alignment process include:

[0040] The Z-score standardization method was used to process geological condition data and pipe jacking construction parameters;

[0041] Box-Cox transform was used to process historical settlement monitoring data; all data sequences were standardized to Coordinated Universal Time (UTC).

[0042] The coordinates are uniformly projected to the universal transverse Mercator projection coordinate system, and the geological condition data are mapped to the settlement monitoring point locations using a spatial interpolation algorithm.

[0043] By adopting the above technical solution, the scale difference between data of different dimensions is eliminated, making the data distribution more concentrated, which is conducive to the convergence of neural networks; at the same time, it solves the inconsistency problem of multi-source data in terms of time and space benchmarks, and ensures the correspondence of input features in a physical sense.

[0044] Preferably, the execution process of the cross-attention mechanism includes:

[0045] Use temporal features as query vectors and spatial features as key and value vectors;

[0046] Calculate the dot product of the query vector and the transpose of the key vector, and perform scaling and Softmax normalization on the dot product result to obtain the attention weight matrix;

[0047] Attention features are obtained by weighting and aggregating the value vectors using an attention weight matrix.

[0048] Calculate dynamic gated scalars based on original temporal characteristics;

[0049] We use dynamic gating scalars and their complementary values ​​to weight the temporal features and attention features respectively, and then add the two results together.

[0050] By adopting the above technical solutions, deep feature fusion is achieved at the mathematical level. Specifically, the dot product operation of the query vector and the transpose of the key vector reflects the correlation between the current construction status and geological spatial conditions; the Softmax normalization operation transforms the correlation into weights in the form of a probability distribution; the weighted aggregation of value vectors realizes spatial information extraction based on correlation; and the calculation of dynamic gating scalars and weighted summation operations construct a feature adaptive adjustment mechanism, ensuring that the model can automatically adjust the contribution ratio of spatial features in the final prediction result according to the current input state, thereby adapting to complex and ever-changing engineering environments.

[0051] Preferably, based on the comparison results of the predicted cumulative settlement and predicted settlement rate with the stored first-level, second-level, and third-level criteria, control signals including emergency braking, parameter adjustment, key marking, or maintaining routine operations are generated.

[0052] By adopting the above technical solutions, a dual-indicator graded early warning system was established, which outputs specific engineering operation instructions for different risk levels, thereby improving the response speed and pertinence of on-site handling.

[0053] This invention provides a system and method for predicting road surface settlement during underground pipe jacking tunnel excavation. It offers the following advantages:

[0054] 1. This invention employs a CA-CNN-LSTM hybrid neural network architecture, utilizing parallel branches to process static geological spatial features and dynamic construction time-series features separately, avoiding feature interference caused by simple splicing. By utilizing a Cross-Attention mechanism and dynamic gating units, it achieves adaptive weighted fusion of geological spatial features based on the construction time-series status. This effectively solves the problem of traditional models struggling to capture spatiotemporal coupling features, enabling the model to maintain low prediction errors even when facing sudden changes in geological conditions or fluctuations in construction parameters.

[0055] 2. This invention integrates cubic spline interpolation, Box-Cox transformation, and feature space scaling transformation algorithms. By unifying the spatiotemporal benchmark and expanding the training samples within physical constraints, the quality and distribution of the input data are ensured, thereby reducing the model's dependence on a specific training set and enabling it to maintain stable predictive performance even under small sample conditions.

[0056] 3. This invention not only outputs predicted values ​​for road surface settlement but also establishes a dual-indicator graded early warning system that includes both cumulative settlement and settlement rate. The system can directly generate specific control signals for Level 1 emergency stop, Level 2 adjustment, or Level 3 monitoring based on the prediction results, transforming abstract data analysis results into executable on-site operation instructions. This shortens risk response time and effectively ensures road safety during pipe jacking construction. Attached Figure Description

[0057] Figure 1 This is a flowchart of the CA-CNN-LSTM hybrid neural network topology of the present invention;

[0058] Figure 2 This is a schematic diagram of the Cross-Attention feature fusion module of the present invention;

[0059] Figure 3 This is a flowchart of the road settlement prediction system of the present invention;

[0060] Figure 4 This is a flowchart illustrating the hierarchical risk warning logic of the present invention. Detailed Implementation

[0061] The technical solutions in 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.

[0062] To better understand the present invention, the above content will be described in detail below with reference to specific embodiments.

[0063] Please see the appendix Figure 1 - Appendix Figure 4 This invention provides a system and method for predicting road surface settlement during underground pipe jacking tunnel excavation.

[0064] This invention provides a road surface settlement prediction system for underground pipe jacking tunnel excavation. This prediction system achieves quantitative prediction of road surface settlement by deeply integrating geological spatial characteristics and construction time sequence characteristics. It is applicable to safety monitoring of underground engineering projects under various geological conditions such as silty clay.

[0065] The prediction system comprises data acquisition terminals, a transmission network, a computing server, and a display terminal at the hardware level. Data acquisition terminals are distributed across the construction site and monitoring points, equipped with various sensor interfaces to acquire real-time geological environmental parameters, pipe jacking machine operating parameters, and road settlement monitoring values. The transmission network establishes a communication link between the data acquisition terminals and the computing server, responsible for converting on-site analog signals into digital signals and transmitting them to the backend.

[0066] The computing server, as the core processing unit of the prediction system, is equipped with a central processing unit and a graphics processing unit, and deploys a deep learning computing framework to perform data storage, computation, and model inference tasks. The display terminal is connected to the computing server to display the graphical user interface, providing parameter input windows and result visualization windows for interactive operation by engineers.

[0067] The operational logic deployed within the computing server is functionally divided into five interconnected execution modules: a data management module, a data preprocessing module, a model building module, a training and optimization module, and an online monitoring and early warning module. These modules are connected via a data bus to enable the targeted transmission and processing of data streams.

[0068] The data management module connects to the underlying database and is responsible for the classification, storage, and indexing of multi-source heterogeneous data. It receives geological condition data, construction parameters, environmental data, and historical settlement data from data acquisition terminals and transforms them into structured data tables. The data management module supports rapid retrieval and reading of massive amounts of monitoring data, providing a data foundation for subsequent model training.

[0069] The data preprocessing module connects to the data management module to improve data quality and consistency. The data preprocessing module performs data cleaning procedures, using statistical criteria to identify and remove outliers and noise from the raw data. It also performs normalization and spatiotemporal alignment operations to eliminate scale differences between different physical units and map time-series data from different sampling frequencies onto a unified time axis, generating standardized model input tensors.

[0070] The model building module is used to construct a CA-CNN-LSTM hybrid neural network topology. Based on predefined algorithm logic, the module instantiates convolutional neural network layers, long short-term memory network layers, and cross-attention mechanism layers. The module defines the connection relationships, convolutional kernel size, number of neurons, and activation function types for each network layer, constructing a deep learning model capable of extracting and dynamically fusing spatial and temporal features in parallel.

[0071] The training and optimization module is used to solve for model parameters and perform performance tuning. It receives the partitioned training set data, calculates predicted values ​​through forward propagation, and calculates prediction bias using the mean squared error loss function. The module executes the backpropagation algorithm and gradient descent strategy, iteratively updating the weight matrices and bias parameters of each network layer in the model building module until the model converges or reaches the preset number of training epochs.

[0072] The online monitoring and early warning module is used to perform real-time settlement prediction and risk assessment. It calls upon a trained model, receives real-time construction and geological data, and outputs predicted pavement settlement values ​​for future times. The module has built-in tiered early warning thresholds set according to engineering specifications, comparing the calculated cumulative settlement and settlement rate with these thresholds. When the predicted indicators reach the risk limits, the module drives the display terminal to generate corresponding red, orange, yellow, and green warning signals.

[0073] The computer program instructions contained in the prediction system are stored in the storage medium of the computing server and executed by the processor to achieve specific functions. Logically, the prediction system is divided into a data management module, a data preprocessing module, a model building module, a training and optimization module, and an online monitoring and early warning module.

[0074] The data management module is configured to provide a human-computer interaction interface and underlying data indexing functionality. The module includes a parameter input area for receiving basic parameters entered by engineers or automatically collected by the system. These basic parameters include geological condition data, construction parameter data, environmental data, and monitoring data. Geological condition data includes soil layer type selection instructions, cohesion values, internal friction angle values, groundwater level values, and pipe jacking depth values; construction parameter data includes jacking speed values, grouting pressure values, grouting volume values, and real-time jacking distance values; environmental data includes daily average temperature values ​​and daily precipitation values. The data management module establishes a mapping relationship between each parameter and a timestamp, forming the original data sequence.

[0075] The data preprocessing module is connected to the data management module and is used to execute standardized data cleaning procedures. The data preprocessing module has a built-in interquartile range (IQR) algorithm configured to calculate the upper and lower quartiles of the original data sequence, identifying and removing outliers exceeding a preset discrete threshold. The data preprocessing module also integrates cubic spline interpolation and moving average algorithms to complete missing data and unify data sequences from different sources to a preset time resolution. The data preprocessing module is further configured with a data augmentation subroutine that introduces physically bounded random perturbations into the original data sequence through feature space scaling transformation and adaptive Gaussian noise injection algorithms, generating an expanded training sample set.

[0076] The model building module is configured to instantiate a CA-CNN-LSTM hybrid neural network architecture. It establishes parallel convolutional neural network (CNN) and long short-term memory (LSTM) network branches. The CNN branch consists of multiple concatenated one-dimensional convolutional layers and pooling layers, used to extract spatial feature vectors of geological parameters; the LSTM branch consists of bidirectional recurrent units, used to extract temporal feature vectors of construction parameters. The module further constructs a cross-attention feature fusion layer, defines the mapping rules between query vectors, key vectors, and value vectors, and establishes adaptive gating units to calculate the fusion weights of spatial and temporal features, generating a high-dimensional feature representation.

[0077] The training optimization module is configured to manage the parameter iteration process of the model. It stores the calculation formulas for the mean squared error loss function and the L2 regularization term, used to evaluate the deviation between the model output value and the actual settling value. The module integrates the Adam optimizer algorithm, updating the weight parameters of each network layer in the model building module based on the calculated gradient information. The module also includes early stopping monitoring logic; when the validation set loss value fails to decrease for a preset number of consecutive epochs, it generates a training termination command and saves the current optimal model parameters.

[0078] The online monitoring and early warning module is configured to perform real-time inference and risk assessment tasks. It inputs real-time collected data into a trained model and outputs predicted road settlement values. The module incorporates engineering-compliant judgment logic to calculate cumulative settlement and settlement rate indices, comparing these indices with preset safety thresholds. Based on the comparison results, the module generates tiered early warning signals. If the predicted indices exceed the first-level threshold, the corresponding monitoring point is marked on the display terminal with a red high-risk warning; if the predicted indices are within the second or third-level threshold range, orange or yellow risk warnings are displayed respectively; and if the predicted indices are below the safety threshold, a green normal status warning is displayed.

[0079] This invention provides a method for predicting road surface settlement during the excavation of underground pipe jacking tunnels. The prediction method executes a series of computer-readable instructions to achieve a complete closed loop from multi-source data input to settlement risk early warning, mainly including data acquisition step S1, data preprocessing step S2, model building step S3, model training and optimization step S4, and settlement prediction and early warning step S5.

[0080] S1. Perform systematic collection of multi-source heterogeneous data. Step S1 specifically includes steps S11 to S13.

[0081] S11. Obtain geological condition data, which characterizes the static physical properties of the construction area, specifically including soil layer type coding and soil cohesion. internal friction angle Groundwater level and the burial depth of the pipe jacking tunnel .

[0082] S12. Obtain the pipe jacking construction parameters, which characterize the dynamic disturbance features during tunnel excavation, specifically including the jacking speed. Grouting pressure Unit grouting volume and real-time jacking distance .

[0083] S13. Acquire environmental data and settlement monitoring data, wherein the environmental data includes average daily temperature. With daily precipitation The settlement monitoring data includes road surface settlement at historical time steps. Regarding the specific methods for collecting the aforementioned data, those skilled in the art can directly read the data using standard sensing equipment such as earth pressure gauges, flow meters, total stations, and meteorological monitoring stations deployed at the construction site, or obtain it through remote transmission interfaces. These are well-known technologies in the field and will not be elaborated upon here.

[0084] S2. The collected raw data undergoes structured cleaning and enhancement to construct a high-quality dataset suitable for deep learning models. Step S2 first executes data cleaning step S21, using the Pandas library to convert multi-source data into structured data frames and applying the interquartile range (IQR) criterion to identify outliers. For identified outliers, nearest neighbor interpolation or mean substitution is used for correction. Subsequently, data normalization step S22 is executed, using Z-score normalization to map feature data of different dimensions to a standard normal distribution space, and performing Box-Cox transformation on skewed settlement features to improve the normalization of the data distribution.

[0085] Step S2 also includes the crucial spatiotemporal alignment step S23. To address the issue of inconsistent sampling frequencies from different sensors, all time-series data are unified to Coordinated Universal Time (UTC). The time zone was determined, and a cubic spline interpolation algorithm was used to resample the data to a uniform time resolution. (For example, 60 minutes). To address the spatial mismatch issue, the coordinates of geological exploration boreholes and road settlement monitoring points are uniformly projected onto the UTM coordinate system, and geological parameters are mapped to the locations of each settlement monitoring point using a spatial interpolation algorithm.

[0086] Furthermore, step S2 includes a physically constrained data augmentation step S24 to expand the training sample size and improve the model's generalization ability. Step S24 employs a feature space scaling transformation strategy to scale the input feature vector. Applying a linear transformation to generate enhanced samples The calculation formula is: Among them, the scaling factor With offset The values ​​are strictly limited by the physical boundaries of soil mechanics and the limiting parameters of construction equipment to avoid generating invalid samples that violate physical laws. Simultaneously, an adaptive Gaussian noise injection strategy is employed to superimpose zero-mean Gaussian noise into the feature data. Generate noisy features The noise term follows a distribution. , Differentiated noise intensity coefficients are set for different feature types. is the standard deviation of the corresponding feature.

[0087] S3. Construct a CA-CNN-LSTM hybrid neural network model based on parallel architecture and cross-attention mechanism. Step S3 includes establishing two parallel data processing branches: a spatial feature extraction branch and a temporal feature extraction branch. The spatial feature extraction branch is configured as a convolutional neural network (CNN) structure, containing multiple one-dimensional convolutional layers, used to process geological condition data and tunnel structure parameters to extract implicit spatial correlation features. The temporal feature extraction branch is configured as a bidirectional long short-term memory (Bi-LSTM) network structure, using gated recurrent units to process construction parameters and historical settlement sequences to capture long-term and short-term time dependencies.

[0088] Step S3 focuses on constructing the Cross-Attention feature fusion module, which connects the outputs of the two parallel branches mentioned above. During the fusion process, the temporal features output by the Bi-LSTM branch are mapped to the query vector. Mapping the spatial features output by the CNN branch to key vectors AND value vector Calculate using the scaled dot product attention mechanism. and The correlation matrix between them, and based on this, the correlation matrix between them. Weighted aggregation is performed to dynamically filter key geological spatial features using construction time sequence. Subsequently, an adaptive gating mechanism is used to generate a gating signal, and the original time sequence features and the aggregated attention features are weighted and summed to generate the final fused feature vector.

[0089] S4. Perform iterative training and parameter optimization of the model. Input the preprocessed training set data into the constructed model and perform forward propagation to calculate the predicted settlement value. Calculate the mean square error (MSE) between the predicted and actual settlement values, and add an L2 regularization term to construct the total loss function. The Adam optimizer is used to calculate the gradient based on the loss function, and the convolutional kernel weights, LSTM unit weights, and attention projection matrix parameters in the network are updated through backpropagation. An early stopping strategy is implemented, terminating training when the validation set loss decreases for several consecutive epochs, thus saving the optimal model parameters.

[0090] S5. Utilize the trained model for online settlement prediction and graded early warning. Input the real-time geological and construction data, processed in step S2, into the model, and output the predicted pavement settlement values ​​for future time steps. Calculate the cumulative settlement based on the predicted values. With settling rate The calculation results are compared with the control thresholds specified in the "Technical Standard for Monitoring of Building Foundation Pit Engineering". If the index is lower than the warning threshold, it is determined to be in a safe state; if the index exceeds the set threshold, the corresponding level of warning signal is triggered according to the degree of exceedance, and the risk area and values ​​are displayed intuitively through the system interface to assist engineers in making decisions on grouting reinforcement or adjusting jacking parameters.

[0091] In step S11, geological condition data is acquired to characterize the static physical and mechanical properties of the construction area. The geological condition data specifically includes soil layer type and soil cohesion. internal friction angle Groundwater level and the burial depth of the pipe jacking tunnel Soil layer types are obtained by consulting geological survey reports and converted into numerical vectors according to preset coding rules. For example, silty clay, muddy soil, and sandy soil are assigned different discrete integer codes or subjected to unique thermal coding. Soil cohesion. With internal friction angle Groundwater level, a key parameter characterizing the shear strength of soil, was determined through borehole sampling and laboratory geotechnical testing. Burial depth with tunnel axis The effective stress state and hydrostatic pressure boundary of the soil were defined. The geological data mentioned above are spatially distributed data, which will be mapped to the locations of each settlement monitoring point based on spatial coordinates in subsequent processing, serving as the input reference for the spatial feature extraction branch in the model.

[0092] In step S12, real-time acquisition of pipe jacking construction parameters is used to characterize the dynamic disturbance characteristics of the soil during the construction process. These construction parameters include the jacking speed. Grouting pressure Unit grouting volume and real-time jacking distance Among them, the jacking speed This reflects the rate at which the cutterhead cuts the soil and the squeezing effect on the soil in front; grouting pressure. This reflects the supporting effect of the thixotropic mud sleeve on the surrounding soil layers. Excessive or insufficient grouting pressure can lead to plastic deformation of the soil; unit grouting volume... It characterizes the degree of compensation for formation loss; real-time jacking distance. This is used to determine the relative positional relationship between the pipe jacking machine head and the monitoring point. The acquisition of the above construction parameters is achieved through a data interface connected to the pipe jacking machine control room or by adding pressure sensors, flow meters, and displacement sensors. The hardware connection and signal reading methods are well-known technologies in this field.

[0093] In step S13, environmental data is collected to quantify the impact of external climatic factors on shallow surface subsidence. The environmental data includes average daily temperature. With daily precipitation Daily precipitation It directly affects the water content and saturation of shallow soil, thereby altering the physical and mechanical properties of the soil; average daily temperature Changes in these conditions can cause thermal expansion and contraction of the road surface structure layers. This type of data is obtained through small weather stations set up at the construction site or by accessing the public data interface of the local meteorological bureau.

[0094] In step S14, road surface settlement monitoring data is acquired as supervision labels and autoregressive input features for model training. The monitoring data includes the cumulative road surface settlement over historical time steps. and settling rate Vertical displacement data at each monitoring point is recorded using a total station, electronic level, or automated hydrostatic leveling system at a preset sampling frequency. During the model training phase, historical settlement sequences are used to construct a time sliding window, enabling the model to learn the inertial trend of settlement development; during the prediction phase, real-time monitoring data is used to correct the initial state of the model.

[0095] The data collected in the above four steps together constitute a multi-source heterogeneous raw dataset, covering all dimensions from static geological properties to dynamic construction disturbances and external environmental influences, which can fully support the technical feature description of multi-source data fusion input in the claims. All collected data are marked with unified timestamps and spatial coordinates to facilitate subsequent data alignment and fusion processing.

[0096] In data cleaning step S21, the system performs outlier detection and missing value imputation. For outliers caused by sensor malfunctions or transmission errors, the interquartile range criterion is used for statistical identification. The upper quartiles of the data sequence are calculated. and lower quartile The interquartile range was obtained. Set the confidence interval for valid data as follows: Any value falling outside this range is considered an outlier. For detected outliers and missing values ​​in the original data, time-series interpolation is used for correction. Specifically, for short-term missing values, linear interpolation is used to fill in the gaps using observations from previous and subsequent time points; for long-term missing values, mean-based filling or correlation-based completion strategies based on similar sensors are used to ensure the continuity and integrity of the data sequence.

[0097] In the data normalization step S22, feature scaling and distribution transformation operations are performed to eliminate the interference of different physical dimensions on the model weight update. For geological parameters and construction parameters that follow or approximately follow a normal distribution, the Z-score normalization method is used. The Z-score normalization method calculates the mean of the original feature sequence. with standard deviation And according to the formula Original data Convert to standard distribution data with mean 0 and standard deviation 1. For pavement settlement monitoring data exhibiting a skewed distribution, a Box-Cox transformation is used for data correction to improve the prediction accuracy of the regression model. This transformation introduces parameters... For raw data The specific formula for performing power function conversion is as follows:

[0098] ;

[0099] The system searches for optimal parameters using the maximum likelihood estimation method. This makes the transformed data The distribution characteristics approximate the normal distribution to the greatest extent possible, thereby reducing the excessive influence of tail extrema on the model loss function.

[0100] In the spatiotemporal alignment step S23, the asynchronous nature of multi-source heterogeneous data in both time and space is addressed. Regarding the time dimension, since geological exploration data are static values, while construction parameters and monitoring data have different sampling frequencies, the system first converts the timestamps of all time-series data to Coordinated Universal Time (UTC). Subsequently, a unified time resolution is set. The cubic spline interpolation algorithm is used to downsample high-frequency data or upsample low-frequency data to a unified time step. By constructing a piecewise cubic polynomial, cubic spline interpolation fits the data curve while ensuring the continuity of the first and second derivatives, effectively preserving the trend characteristics of construction parameters during transient changes.

[0101] For the spatial dimension, step S23 performs a unified projection and spatial mapping of the coordinate system. The geographic coordinates of the geological exploration boreholes and the road settlement monitoring points are uniformly transformed to the universal transverse Mercator projection coordinate system, eliminating distance distortion caused by the Earth's curvature. Based on the projected planar coordinates, the Euclidean distance between the real-time position of the pipe jacking machine head and each settlement monitoring point is calculated. According to the regional geological profile model generated from the geological exploration report, a spatial interpolation algorithm is used to map the discrete borehole geological parameters to the specific coordinates of each settlement monitoring point, thereby constructing a corresponding static geological attribute vector for each monitoring point and completing the feature alignment in the spatial dimension.

[0102] This embodiment further elaborates on the physical constraint-driven data augmentation step S24 in the data preprocessing step S2. Given the high cost and scarcity of underground engineering samples, directly using the raw collected data to train a deep neural network can easily lead to overfitting. Step S24 is configured to expand the scale and diversity of the training samples through mathematical transformations while maintaining compliance with the physical meaning of the data. Specifically, this includes two strategies: feature space scaling transformation and adaptive Gaussian noise injection.

[0103] In the feature space scaling transformation strategy, the system sets physical boundary constraints based on soil mechanics principles and construction equipment performance indicators, and modifies the original feature vectors before normalization. Apply a linear transformation to generate enhanced samples The transformation process follows the calculation formula. .in, This is the scaling factor. This is the offset.

[0104] To ensure that the generated augmented samples conform to actual engineering logic, the scaling factor... With offset The range of values ​​is subject to strict physical constraints. Specifically, for geological parameters, the transformed values ​​must fall within an empirical range, such as the enhanced internal friction angle. The cohesion must not exceed the maximum theoretical value of the corresponding soil type under the limit state. The value must not be negative. For construction parameters, the transformed values ​​must not exceed the rated working limits of the construction equipment, such as the enhanced grouting pressure. The jacking speed must not exceed the maximum output pressure of the grouting pump. The maximum design speed of the pipe jacking machine must not be exceeded. This was generated through random sampling within the aforementioned physical constraints. and The model can simulate parameter fluctuations under different operating conditions, thereby learning more robust feature representations.

[0105] In the adaptive Gaussian noise injection strategy, the system simulates the unavoidable random errors and environmental interference during the field sensor acquisition process. This strategy superimposes zero-mean Gaussian noise onto the original feature data to generate noisy feature data. The calculation formula is: Among them, the noise term Follows Gaussian distribution In this formula, This represents the standard deviation of the current feature sequence. This is the noise intensity coefficient.

[0106] The key technical feature of this strategy lies in the noise intensity coefficient. Differentiated configurations. Considering the differences in measurement accuracy among different types of sensors, the system assigns different configurations to features in different dimensions. For example, for pavement settlement data measured using a high-precision total station, a smaller noise intensity coefficient is set to preserve subtle deformation trends; for construction parameters such as grouting flow rate or jacking speed, which fluctuate significantly, a larger noise intensity coefficient is set to cover the uncertainties of construction operations. This adaptive noise injection mechanism enhances the model's ability to resist input disturbances, enabling it to maintain stable predictive performance even when faced with noisy real-time monitoring data in actual engineering projects.

[0107] This embodiment details the specific implementation of the parallel feature extraction layer in model construction step S3. The hybrid neural network model constructed in this invention adopts a dual-channel parallel input topology, aiming to process static geological spatial data and dynamic construction time-series data with different data distribution characteristics respectively. Specifically, the parallel feature extraction layer consists of a convolutional neural network spatial feature extraction branch and a bidirectional long short-term memory network temporal feature extraction branch.

[0108] The spatial feature extraction branch of the convolutional neural network is configured to process geological condition data and tunnel structural parameters. Considering the continuous and layered spatial distribution of geological soil layers along the tunnel axis, this branch adopts a one-dimensional convolutional network architecture. Specifically, this branch consists of three cascaded one-dimensional convolutional layers with kernel numbers of 32, 64, and 64 respectively. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function layer to accelerate model convergence and introduce nonlinear transformation capabilities. Max pooling layers are placed between the convolutional layers to reduce feature dimensionality. After the last convolutional operation, a global average pooling layer is specifically set in this branch. The global average pooling layer calculates the average value of each feature channel over the entire time or spatial dimension, compressing the variable-length feature sequence into a fixed-length feature vector, thereby effectively reducing the number of model parameters and preventing overfitting.

[0109] To address the issue of spatial location information loss due to the translation invariance of convolutional neural networks, this branch introduces sinusoidal positional encoding at the input. Sinusoidal positional encoding maps position indices to trigonometric function values ​​of different frequencies, which are then superimposed on the input features. For geological sequences where the position is... Feature dimension index is or The element, its position encoding The calculation formula is:

[0110] ;

[0111] ;

[0112] in, Let be the dimension of the feature vector. Using the above formula, the model can utilize the periodicity of sine and cosine functions to perceive the relative distances and sequence order of different geological parameters along the tunnel axis. After three layers of convolution and pooling operations, this branch outputs a high-dimensional feature vector that implicitly contains geological spatial correlations, denoted as . This high-dimensional feature vector It not only encodes soil properties at a single location, but also covers contextual information about the geological conditions in the vicinity through the receptive field of the convolution kernel.

[0113] The temporal feature extraction branch of the bidirectional Long Short-Term Memory (LSTM) network is configured to process the pipe jacking construction parameter sequence and the historical pavement settlement sequence in parallel. Due to the time lag and cumulative effect of soil disturbance caused by construction, this branch employs a two-layer stacked bidirectional LSTM structure, with 64 hidden units in each layer. The bidirectional structure includes a forward computation layer and a backward computation layer; the forward layer starts from the time step... Read up to the current time The first layer captures the cumulative impact of historical construction operations on the current state; the second layer assists in capturing potential periodic patterns in the data sequence from the reverse time dimension.

[0114] Within a Bi-LSTM unit, the information flow is regulated through the synergistic action of the input gate, forget gate, and output gate. The forget gate determines how much information from the previous time step's unit state is retained, the input gate determines how much of the current construction input is written into the unit state, and the output gate determines the hidden state ultimately passed to the next layer. This gating mechanism effectively solves the gradient vanishing problem in long-sequence training, ensuring that the model can capture the long-range dependence of early grouting pressure fluctuations or sudden changes in jacking speed on later settlement. After processing by two layers of Bi-LSTM, this branch outputs a temporal feature vector containing rich temporal dynamic information, denoted as […]. .

[0115] The two parallel branches described above run independently and do not interfere with each other. The output of the spatial feature extraction branch... Dimensions It mainly characterizes the spatial background of where subsidence occurs; the output of the temporal feature extraction branch is... The dimensions were also adjusted to These two feature vectors primarily characterize the temporal trend of when settlement occurs. and As the basic input for the subsequent Cross-Attention feature fusion module, it provides a structured feature representation for realizing deep interaction between geological space and construction time sequence.

[0116] This embodiment details the construction and operation mechanism of the core Cross-Attention feature fusion module in the CA-CNN-LSTM model. Located after the parallel feature extraction layer and before the prediction output layer, the feature fusion module is configured to perform a deep nonlinear fusion of geological spatial features and construction time-series features. The feature fusion module mainly consists of a feature projection unit, a cross-attention calculation unit, and a dynamic gating fusion unit.

[0117] The feature projection unit is configured to unify the dimension of the output features of the parallel branches and establish a query system for cross-mapping. Specifically, the feature projection unit receives the temporal feature vector output by the Bi-LSTM branch and defines it as the original temporal feature. Simultaneously, the feature projection unit receives spatial feature vectors output from the CNN branches. To establish correlations between features of different properties, the feature projection unit maps the input features to a unified dimension through a linear projection layer. In this mapping process, we will include temporal features that contain rich time-dependent information. Mapped to query vector This maps spatial features containing geological spatial structure information into key vectors. Sum value vector This feature allocation method constructs a spatiotemporal correlation mechanism, that is, it utilizes query vectors containing dynamic temporal information. For key vectors containing spatial distribution information Correlation calculations are performed to determine the weight distribution of geological features at different spatial locations in the current time step.

[0118] The cross-attention computation unit is configured to perform scaled dot product attention operations to quantify the importance weights of different geological conditions at the current construction moment. The cross-attention computation unit first computes the query vector. With key vector transpose matrix The dot product is calculated, and the result characterizes the correlation strength between the construction status and geological conditions. The dot product result is then divided by a scaling factor. To prevent the vanishing gradient problem caused by excessive dimensionality, the scaled values ​​are input into the Softmax normalization function to generate an attention weight matrix in the form of a probability distribution. Finally, this weight matrix is ​​used to adjust the value vector. Perform weighted summation to generate attention features. The calculation process strictly follows the following formula:

[0119] ;

[0120] in, This represents the query vector originating from the LSTM branch. This represents the key vector originating from a CNN branch. This represents a value vector originating from a CNN branch. The scaling constant represents the feature dimension. The function ensures that the sum of all weights is 1.

[0121] The dynamic gating fusion unit is configured to address the information redundancy and conflict issues in multi-source feature fusion. This is done to balance the original temporal features. Historical inertia and cross-attention characteristics To mitigate the impact of spatial perturbations, the dynamic gating fusion unit introduces an adaptive gating mechanism. First, a fully connected layer combined with a sigmoid activation function generates a dynamic gating signal with a value between 0 and 1 based on the current feature state. Dynamic gating signal Acting as an information filter determines the proportion of original information retained.

[0122] Subsequently, the dynamic gating fusion unit performs the final feature mixing operation. The feature mixing operation utilizes gating signals. Features of the original LSTM output Weighting is performed using complementary signals. Cross-attention calculation results Weighting is performed, and the two results are added together to obtain the final higher-order feature vector. The fusion process follows the formula below:

[0123] ;

[0124] in, This is the final high-order feature vector input to the fully connected layer. For dynamically generated scalar gated signals. This represents element-wise multiplication. Using this formula, the model can automatically determine, without human intervention, whether to rely more on historical settlement trends or focus more on the specific impacts of current geological conditions based on the current construction context, thus achieving physical mechanism-driven adaptive feature fusion.

[0125] This embodiment further elaborates on the specific structure and working principle of the prediction output module in model construction step S3. The prediction output module is located at the end of the entire deep learning network and is connected to the output of the Cross-Attention feature fusion module. The prediction output module is configured to map the high-dimensional feature vector after multi-layer feature extraction and adaptive fusion into a one-dimensional continuous value, that is, the road settlement prediction value for the next time step.

[0126] The prediction output module first contains a fully connected layer, which acts as a nonlinear regressor. The fully connected layer receives the output vector from the dynamically gated fusion unit. As input. To enhance the model's non-linear expressiveness and prevent overfitting, fully connected layers are typically designed as multilayer perceptron structures, containing at least one hidden layer and one output layer. The hidden layer is configured with a ReLU activation function to filter negative responses and introduce sparsity. Between the hidden and output layers, a Dropout regularization layer is placed, which randomly disconnects neuron connections with a preset probability during training, thereby forcing the network to learn more robust distributed feature representations.

[0127] The final layer of the prediction output module is a linear regression layer, which has a single neuron and does not use a non-linear activation function. The linear regression layer projects the hidden features to the real number space through a weighted summation operation, generating a predicted value at a normalized scale. The mapping process follows the matrix operation formula:

[0128] ;

[0129] in, The weight matrix of the output layer. The output feature vector of the hidden layer. This is the bias term for the output layer. The representative model's predicted value under a standardized data distribution.

[0130] To obtain settlement values ​​with practical engineering physical significance, the prediction output module also integrates an inverse normalization processing unit. This unit stores the statistical parameters of the settlement data calculated in data preprocessing step S2, including the mean value. Standard deviation and Box-Cox transform parameters The inverse normalization unit performs the opposite mathematical transformation to the preprocessing step, transforming... Restored to the actual road surface settlement prediction value in millimeters For data that has undergone Z-score standardization, the inverse transform formula is: If a Box-Cox transformation has been performed, then a power-law restoration operation is further executed. The final output is... This is the system's quantitative prediction of the road surface condition at future moments, which is directly used for subsequent risk warning and judgment.

[0131] This embodiment details the loss function design and parameter iterative update mechanism in step S4 of the model training optimization. To ensure that the constructed CA-CNN-LSTM hybrid neural network model accurately approximates the nonlinear ground settlement patterns under pipe jacking construction conditions, step S4 is configured to construct a composite objective function. This objective function not only focuses on the accuracy of the predicted values ​​but also introduces constraints on model complexity to prevent overfitting on limited engineering sample data.

[0132] The training optimization module uses mean squared error as the main loss term to evaluate the deviation between the model's predicted values ​​and the actual monitored values. For each training batch, the system calculates the average of the sum of squared prediction errors for all sample points. Main Loss Term The calculation follows the formula:

[0133] ;

[0134] in, This represents the size of the training batch. Represents the sample index in the current batch. Representing the The actual road surface settlement monitoring value corresponding to each sample This represents the predicted settlement value output by the model for this sample. This function is a convex function that is differentiable everywhere, providing a stable convergence direction for the gradient descent algorithm.

[0135] Given that the model constructed in this invention includes convolutional layers, LSTM layers, and an attention mechanism, involving a large number of weight parameters, and that construction site data often contains noise interference, an L2 regularization term is introduced in the training optimization module to improve the model's generalization ability, based on the main loss term. The L2 regularization term penalizes the sum of squares of all trainable weight parameters in the network, forcing the model to tend to choose smaller weight values ​​during training, thereby generating smoother decision boundaries. Regularization term The calculation formula is:

[0136] ;

[0137] in, Represents the total number of all weight parameters in the network. Representing the network The values ​​of each weight parameter.

[0138] Based on the above two indicators, the system constructs the final composite objective function. Composite objective function By introducing a regularization coefficient This is to balance the relationship between fitting accuracy and model complexity. The final loss calculation formula is defined as:

[0139] ;

[0140] In this formula, This is a hyperparameter used to control the strength of the regularization penalty. If If the value is too large, it will lead to underfitting of the model; if the value is too small, it will not be able to effectively suppress overfitting. In this embodiment... The specific values ​​will be explained in the subsequent parameter settings section. Coefficient This is to cancel out the coefficient 2 generated by the squared term when calculating the gradient by differentiation, thus simplifying the calculation process.

[0141] In each training iteration, the system calculates... The value is then used to execute the backpropagation algorithm using the chain rule. The backpropagation algorithm calculates the objective function relative to each weight parameter in the network. gradient The calculated gradient information is passed to the optimizer to update the kernel weights in the convolutional neural network branch, the gate unit weights in the bidirectional LSTM branch, and the projection matrix parameters in the Cross-Attention module, so that the model gradually converges to the global optimum or local optimum in the multidimensional feature space.

[0142] This embodiment further discloses the specific hyperparameter configuration scheme in the model training optimization step S4. To ensure that the model can converge to the optimal solution quickly and stably on the multidimensional non-convex loss function surface, this invention does not use a fixed gradient descent step size, but instead configures an adaptive moment estimation optimizer.

[0143] Regarding the learning rate configuration, the initial learning rate is set to 0.001. To quickly approach the optimal solution in the early stages of training and perform refined searching later, the system employs a learning rate decay strategy. When the loss function value on the validation set fails to decrease within five consecutive training epochs, the decay mechanism is triggered, reducing the current learning rate to 50% of its original value, until the learning rate drops to a set lower limit. .

[0144] Regarding the configuration of data input dimensions, based on the hysteresis effect analysis of soil disturbance during pipe jacking construction, the time window length is set to 96. This means that when the model makes a single prediction, it will backtrack on geological changes and construction operation records over the past 96 time steps. This time span covers the typical cycle of soil disturbance from its inception to the dissipation of pore water pressure and eventual stabilization. The training batch size is set to 32, which ensures that stochastic gradient descent has sufficient noise to escape local minima while maximizing the parallel computing efficiency of the graphics processor.

[0145] To prevent the model from overfitting to the training data, the system implements an early stopping strategy. Based on dividing the dataset into training and validation sets, the system monitors the changes in the validation set loss value in real time during training. A tolerance parameter of 15 is set; if the validation set loss does not show an improvement trend within 15 consecutive training epochs, the system will automatically terminate the training process. At this point, the system does not retain the parameters from the last iteration, but instead rolls back and saves the model weight parameters corresponding to the minimum validation set loss as the final deployment model. This mechanism ensures that the output model has optimal generalization performance, rather than simply being a memorization of the training data.

[0146] This embodiment details the operation mechanism of the online real-time prediction submodule in the settlement prediction and early warning step S5. The online real-time prediction submodule is configured to deploy the trained CA-CNN-LSTM hybrid neural network model on a computing terminal at the engineering site, performing a rapid mapping from real-time data streams to predicted settlement values. This online prediction process specifically includes a feature window rolling update step S51, a model forward inference step S52, and a prediction result visualization feedback step S53.

[0147] In the feature window rolling update step S51, the system maintains a time sliding window queue of fixed length, whose length is consistent with the sequence length during model training. Whenever the monitoring system collects a new current time... When processing geological parameters, construction parameters, and environmental data, the system first calls the normalization and alignment procedure in data preprocessing step S2 to standardize the new data. Then, the system performs a first-in, first-out queue update operation, removing historical data from the front of the time window and updating the newly processed time window. The data is appended to the end of the window. A queue update operation is used to construct a window with dimension [missing information]. The real-time input feature tensor, where 1 represents the batch size and 96 represents the time step. This represents the total number of characteristic channels. This step ensures that the model is always based on the latest construction conditions and geological background.

[0148] In the model forward inference step S52, the system loads the optimal model weight parameters saved in Example 5. The constructed real-time input feature tensor is input into the loaded CA-CNN-LSTM network. Matrix multiplication and nonlinear transformation operations are performed using the computer's graphics processing unit or central processing unit. Since the model does not need to calculate gradients or backpropagation during the inference phase, the computation time of the entire forward propagation process is controlled at the millisecond level, meeting the stringent real-time requirements of engineering practice. The model output layer outputs the normalized predicted value, which, after inverse normalization, yields the next time step value. Predicted road settlement value .

[0149] In the prediction result visualization feedback step S53, the system displays the calculated road settlement prediction value. The data is written to the database and pushed to the monitoring terminal interface. The system plots a real-time prediction curve in a time-settlement coordinate system. The real-time prediction curve stitches together historical measured settlement data with future predicted data, visually presenting the development trend of road settlement. The system also calculates the difference between the current predicted value and the measured value at the previous moment, quantifying the settlement rate. This feedback mechanism allows construction personnel to know about road deformation in advance, thus gaining a time window to adjust construction parameters before settlement actually occurs.

[0150] This embodiment details the implementation of the graded risk early warning logic submodule in the settlement prediction and early warning step S5. To transform a single numerical prediction into a decision signal guiding on-site construction, the graded risk early warning logic submodule incorporates a multi-dimensional threshold determination algorithm based on current national standards. This multi-dimensional threshold determination algorithm specifically comprises three steps: early warning index calculation step S54, risk level determination step S55, and differentiated response step S56.

[0151] In the early warning indicator calculation step S54, the system calculates the road surface settlement prediction value for the next time step based on the model output. Real-time calculation of two key risk control indicators: predicted cumulative settlement. With predicted settlement rate Among them, the predicted cumulative settlement That is The absolute value of the value represents the total amount of pavement deformation. It is used to predict the settlement rate. By calculating the predicted value and the current measured value The difference divided by the time interval To obtain, that is This characterizes the degree of abruptness in pavement deformation. The system has a preset cumulative settlement control threshold. (Set at 20mm) and the settlement rate control threshold (Set at 2mm / d), these two thresholds are strictly set in accordance with the protection requirements of the "Technical Standard for Monitoring of Building Foundation Pit Engineering" and the project design documents.

[0152] In the risk level determination step S55, the system performs a four-level early warning logic operation with two indicators in parallel.

[0153] A Level 1 warning (red warning) corresponds to an extremely dangerous state for the project. When the conditions are met... The event is triggered when either the predicted cumulative settlement or the settlement rate exceeds the control threshold.

[0154] A Level 2 warning (orange warning) corresponds to a high-risk state for the project. When the conditions are met... Triggered when the indicator value has reached 80% to 100% of the control threshold.

[0155] A Level 3 warning (yellow warning) corresponds to a situation where the project requires increased attention. When the conditions are met... It is triggered when the indicator value is in the range of 60% to 80% of the control threshold.

[0156] A Level 4 warning (green normal) corresponds to a safe and controllable state for the project. It is triggered when none of the above conditions are met, i.e., when all indicator values ​​are less than 60% of the control threshold.

[0157] In the differentiated response step S56, the system outputs corresponding control commands based on the determined warning level. If a Level 1 warning is triggered, the system generates a Level 1 emergency stop control signal, which is configured to trigger the emergency braking logic of the pipe jacking machine control system and outputs command data suggesting the initiation of an emergency grouting reinforcement program. If a Level 2 warning is triggered, the system generates a Level 2 adjustment control signal, which includes parameter adjustment commands suggesting a reduction in jacking speed and an increase in synchronous grouting volume, and simultaneously outputs a sampling command to increase the monitoring frequency to twice the original frequency. If a Level 3 warning is triggered, the system generates a Level 3 attention signal, marking the current monitoring point as a key focus object and outputting a prompt message to verify the stability of geological parameters. If a Level 4 warning is triggered, the system outputs a normal operation signal to maintain the current construction parameters.

[0158] Through this graded early warning mechanism, the present invention not only realizes the prediction of road settlement values, but also realizes the automated transformation from data to engineering decisions, enabling construction parties to take preventive measures based on clear grade signals before settlement disasters occur.

Claims

1. A road surface settlement prediction system for underground pipe jacking tunnel excavation, characterized in that, The system comprises a data management module, a data preprocessing module, a model construction module, a training optimization module, and an online monitoring and early warning module. The data management module is configured to receive multi-source heterogeneous data from the data acquisition terminal and generate a structured data table. The data preprocessing module is configured to perform cleaning and feature enhancement operations on the structured data table. The model construction module is configured to establish a CA-CNN-LSTM hybrid neural network architecture, which includes a spatial feature extraction branch and a time series feature extraction branch arranged in parallel, and a Cross-Attention feature fusion module connected to the output ends of the two branches. The training optimization module is configured to update the weight parameters of the CA-CNN-LSTM hybrid neural network architecture using a backpropagation algorithm. The online monitoring and early warning module is configured to generate a graded warning signal based on the predicted value output by the model and display it on a display terminal.

2. The system for predicting ground settlement due to underground pipe jacking tunnel excavation according to claim 1, wherein, The spatial feature extraction branch is composed of a one-dimensional convolution layer, a pooling layer, and a global average pooling layer, which are connected in cascade, and is used to output a high-dimensional feature vector with implicit spatial correlation. The time series feature extraction branch is composed of a bidirectional long short-term memory network, which is used to output a time series feature vector containing time dynamic information. The Cross-Attention feature fusion module is divided into three logical units: a feature projection unit, a cross-attention calculation unit, and a dynamic gating fusion unit. The feature projection unit is used to map the input features into query vectors, key vectors, and value vectors. The cross-attention calculation unit is used to calculate the correlation between the query vectors and the key vectors and weight the value vectors. The dynamic gating fusion unit is used to generate a dynamic gating signal and perform feature mixing operations.

3. The system for predicting ground settlement due to underground pipe jacking tunneling according to claim 2, wherein, The feature projection unit is connected to the output ends of the time series feature extraction branch and the spatial feature extraction branch, and maps the time series feature vector into the query vector, the high-dimensional feature vector into the key vector, and the value vector. The dynamic gating fusion unit includes a fully connected layer and a Sigmoid activation function, which is used to generate the dynamic gating signal with a value range of 0 to 1 based on the time series feature vector. The dynamic gating fusion unit is configured with a weighted summation logic, which is used to weight the time series feature vector with the dynamic gating signal, weight the attention features output by the cross-attention calculation unit with a complementary signal, and add them together as the fusion output features.

4. The system for predicting ground settlement due to underground pipe jacking tunneling according to claim 1, wherein, The data preprocessing module integrates quartile range calculation algorithm, cubic spline interpolation algorithm, and feature space scaling transformation algorithm. The quartile range calculation algorithm is used to identify and remove abnormal values. The cubic spline interpolation algorithm is used to resample data sequences with different sampling frequencies to a unified time resolution. The feature space scaling transformation algorithm is used to apply linear transformation to the feature vector within the physical boundary constraint range to expand the training sample set.

5. The system for predicting ground settlement due to underground pipe jacking tunneling according to claim 1, wherein, The online monitoring and early warning module stores hierarchical thresholds set according to engineering specifications, and the hierarchical thresholds include a first-level threshold, a second-level threshold, and a third-level threshold; The online monitoring and early warning module is configured with a comparison logic for comparing the calculated predicted cumulative settlement and predicted settlement rate with the hierarchical thresholds respectively; The online monitoring and early warning module is configured to generate a first-level emergency stop control signal when the index exceeds the first-level threshold, a second-level adjustment control signal when the index is between the first-level threshold and the second-level threshold, and a third-level attention signal when the index is between the second-level threshold and the third-level threshold.

6. The system for predicting ground surface settlement due to underground pipe jacking tunneling according to claim 1, wherein, The sensor interface of the data acquisition terminal is configured to collect the following data: Geological condition data, including soil layer type code, soil cohesion, internal friction angle, groundwater level, and pipe jacking tunnel depth; Pipe jacking construction parameters, including jacking speed, grouting pressure, unit grouting volume, and real-time jacking distance; Environmental data, including daily average temperature and daily precipitation.

7. A method of predicting ground surface settlement due to underground pipe jacking tunnel excavation, characterized by, An underground pipe jacking tunnel excavation pavement settlement prediction system according to any one of claims 1-6, comprising the following steps: S1, obtaining geological condition data, pipe jacking construction parameters, environmental data, and historical settlement monitoring data of the construction site; S2, performing structured cleaning, normalization processing, spatio-temporal alignment processing, and data enhancement processing on the obtained raw data to generate a standardized training data set; S3, constructing a hybrid model including a convolutional neural network branch and a bidirectional long short-term memory network branch, and using a cross-attention mechanism to fuse the spatial features extracted by the convolutional neural network branch and the temporal features extracted by the bidirectional long short-term memory network branch; S4, inputting the training data set into the hybrid model, calculating the deviation between the predicted value and the true value, and updating the weight matrix and bias parameters of the network layer using an optimizer; S5, using the trained hybrid model to receive real-time data streams, output the pavement settlement prediction value at the next time, and generate the corresponding control signal according to the calculated settlement index.

8. The method of claim 7, wherein, The normalization processing and spatio-temporal alignment processing in step S2 specifically include: Using Z-score standardization method to process the geological condition data and pipe jacking construction parameters; Using Box-Cox transformation to process the historical settlement monitoring data; Unifying the time stamps of all data sequences to the coordinated universal time standard; Projecting the geological exploration hole coordinates and pavement settlement monitoring point coordinates to the universal transverse Mercator projection coordinate system; Using spatial interpolation algorithm to map the geological condition data to the settlement monitoring point position.

9. The method of claim 7, wherein, The execution process of the cross-attention mechanism in step S3 includes: Taking the temporal features as query vectors and the spatial features as key vectors and value vectors; Calculating the dot product of the transpose matrix of the query vectors and the key vectors, and performing scaling operation and Softmax normalization operation on the dot product result to obtain an attention weight matrix; Using the attention weight matrix to perform weighted aggregation on the value vectors to generate attention features; Calculating a dynamic gating scalar based on the original temporal features; The dynamic gating scalar is used to weight the timing feature, a complementary value of the dynamic gating scalar is used to weight the attention feature, and the two weighted results are added to obtain a fusion feature.

10. The method of claim 7, wherein, The step S5 specifically comprises: calculating a predicted accumulated settlement and a predicted settlement rate based on the road surface settlement prediction value; determining whether the predicted accumulated settlement and the predicted settlement rate meet a stored first-level criterion, and if so, generating a first-level emergency stop control signal containing an emergency braking instruction; determining whether the predicted accumulated settlement and the predicted settlement rate meet a stored second-level criterion, and if so, generating a second-level adjustment control signal containing a parameter adjustment instruction; determining whether the predicted accumulated settlement and the predicted settlement rate meet a stored third-level criterion, and if so, generating a third-level attention signal containing a key marking instruction; if none of the above criteria are met, a regular operation signal maintaining the current parameters is generated.

Images