Shield tunneling attitude prediction method and system based on CNN-BiLSTM-STDAM combined model, and storage medium

Abstract

The invention provides a shield tunneling attitude prediction method and system based on a CNN-BiLSTM-STDAM combined model, and a storage medium. Comprising the following steps: constructing a multi-modal spatiotemporal data set, preprocessing multi-source heterogeneous data, inputting the preprocessed data into a CNN-BiLSTM-STDAM combined model for attitude prediction, outputting four deviation predicted values of a shield attitude, and triggering a multi-level early warning mechanism according to a prediction result. Through the integrated design of deep space-time modeling, dynamic attention focusing and intelligent early warning, the technical bottlenecks of a traditional neural network method in the aspects of prediction precision, generalization ability and the like are overcome, and a high-reliability, self-adaptive and low-cost attitude prediction solution is provided for projects such as urban subways and cross-river tunnels. And the digitization and intellectualization development of underground engineering construction is promoted.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction technology for tunnel engineering, and in particular to a shield tunneling attitude prediction method, system and storage medium based on a CNN-BiLSTM-STDAM combined model (STDAM (Spatio-Temporal Decoupled Attention Module) includes a spatial attention sub-module (multi-head self-attention) and a temporal attention sub-module (causal convolution), and BiLSTM is a bidirectional long short-term memory network). It is applicable to intelligent construction control scenarios under complex geological conditions such as urban subways, integrated utility tunnels, and river-crossing tunnels. Background Technology

[0002] With the rapid development of underground spaces such as urban subways, river-crossing tunnels, and integrated utility tunnels, the shield tunneling method has become the mainstream construction method due to its advantages of safety, efficiency, and environmental friendliness. Real-time and accurate prediction of the shield tunneling posture (horizontal and vertical deviations of the shield head / tail) is a core element in ensuring the accuracy of the tunnel alignment, avoiding excessive ground settlement, and preventing structural damage.

[0003] In recent years, tunnel boring machines (TBMs) have been widely equipped with multi-source sensors such as ground-penetrating radar, earth pressure sensors, hydraulic sensors, and rotary encoders, which can generate high-dimensional, heterogeneous, and high-noise construction data in real time. However, how to effectively integrate these multi-source heterogeneous data and achieve high-precision, real-time, and generalizable attitude prediction under complex geological conditions remains a technical challenge that the industry urgently needs to solve.

[0004] Existing technologies mainly fall into the following categories, but all of them have significant shortcomings: (a) Single Neural Network Method CNN models can only extract local spatial features and cannot depict the long-term dependence of shield tunneling parameters on the evolution of time, resulting in low accuracy in predicting dynamic working conditions.

[0005] LSTM model: unidirectional time series modeling ignores inverse causality and lacks spatial feature extraction capabilities, and is insufficient for mining multi-sensor coupling relationships.

[0006] (II) Traditional Statistical Regression Methods

[0007] Least squares regression assumes a linear relationship between variables, making it difficult to characterize the strong nonlinear coupling between geology and machinery; it is sensitive to noise and has poor robustness; and it cannot adaptively adjust to geological changes.

[0008] Combined prediction method: CNN-LSTM splicing model: large number of parameters and high computational complexity; simple splicing of spatiotemporal features can easily cause feature confusion, which in turn reduces prediction performance; lacks attention mechanism for shield tunneling conditions.

[0009] (iii) Manual experience method

[0010] Relying on engineers' subjective judgment results in a delayed response (>15 min), which cannot meet the tunnel boring machine's advance speed requirement of 0.5–1 m per minute; and there is a high risk of misjudgment when sensors fail or geological changes occur.

[0011] Existing technologies share the following common drawbacks: Data silos: The sampling frequencies of ground-penetrating radar (spatial high resolution), hydraulic system (high frequency time series), and settlement monitoring (low frequency discrete) are inconsistent with the coordinate reference, making it difficult to unify modeling.

[0012] Spatiotemporal modeling is fragmented: existing technologies either focus on spatial features or on time series, lacking a unified spatiotemporal coupling framework.

[0013] Weak generalization ability: The error fluctuation of the same model across geological scenes is greater than 50%, requiring repeated data collection and offline retraining, which cannot meet the needs of rapid engineering transition.

[0014] Delayed response to sudden working conditions: Traditional models are updated in "hours-days", which cannot respond to extreme working conditions such as sudden changes in soft soil and crossing of isolated boulders in real time.

[0015] In view of this, it is particularly necessary to develop a system and method that overcomes the bottlenecks in accuracy, real-time performance, and generalization ability, and meets the requirements for high-precision dynamic prediction and intelligent control of shield tunnel attitude under complex geological conditions such as urban subways and river-crossing tunnels. Summary of the Invention

[0016] This invention proposes a method, system, and storage medium for predicting the attitude of tunnel boring machines based on a CNN-BiLSTM-STDAM combined model, solving the problems mentioned in the background art. The technical solution of this invention is implemented as follows: A shield tunneling attitude prediction method based on a CNN-BiLSTM-STDAM combined model includes: Step S1: Collect geological parameters, mechanical state parameters, and environmental parameters during shield tunneling construction using a multi-source sensor array to construct a multimodal spatiotemporal dataset; Step S2: Preprocess the multi-source heterogeneous data, including: performing spatiotemporal alignment, outlier removal, missing value filling, normalization, and redundant feature dimensionality reduction on the multimodal spatiotemporal dataset to generate a standardized feature dataset. Step S3: Input the preprocessed data into the CNN-BiLSTM-STDAM combined model for attitude prediction. The CNN module extracts the spatial features of the data; the BiLSTM module models the temporal series features of the data in both forward and reverse directions; the STDAM module decouples the spatial and temporal features and dynamically weights and fuses them; and outputs four predicted deviations of the shield attitude: shield head horizontal deviation, shield head vertical deviation, shield tail horizontal deviation, and shield tail vertical deviation. Step S4: Trigger a multi-level early warning mechanism based on the prediction results.

[0017] Furthermore, the spatiotemporal preprocessing in step S2 specifically includes: Step S21: Use an adaptive sliding window for time series reconstruction and compensate for missing values ​​using cubic spline interpolation; Multi-scale standardization of parameters with different dimensions is performed based on the Min-Max normalization function, and the specific formula is as follows:

[0018] in, This represents the normalized eigenvalues, ranging from [0, 1]. This represents the maximum value of the feature across all samples. This represents the minimum value of the feature across all samples. This represents a specific feature value in the raw sensor data, such as earth pressure or torque. Step S22: Remove outlier data using a dynamic threshold filtering algorithm and fill the window with the mean of a sliding window, as shown in the following formula: ; in, This represents the average value within the i-th sliding window, used to fill in missing values; i represents the starting position of the current calculation, and its value range is 1 ≤ i ≤ n. k+1; k represents the width of the sliding window; Step S23: Based on Pearson correlation coefficient

[0019] in, The Pearson correlation coefficient represents the relationship between feature X and target variable Y, with a value range of [-1, 1].

[0020] xi,yi represent the observed values ​​of feature X and target variable Y in the i-th sample; , The sample mean of the representative feature X and the target variable Y; n represents the total number of samples; Calculate the feature and total attitude deviation

[0021] The correlation is used to generate a feature importance ranking matrix to eliminate redundant features; in, This represents the total deviation between the shield centerline and the tunnel axis. , These are the horizontal deviation of the shield head and the vertical deviation of the shield head. These are the horizontal deviation of the shield tail and the vertical deviation of the shield tail.

[0022] Furthermore, in step S3, The CNN module uses 64 sets of 3×1 convolutional kernels to extract spatial features of earth pressure distribution and hydraulic gradient; The BiLSTM module captures the temporal patterns of the tunneling direction and the lag effect of geological parameters through bidirectional LSTM layers. The STDAM module processes spatiotemporal features separately through a spatiotemporal decoupling attention mechanism: Spatial attention employs a multi-head self-attention mechanism to establish dynamic correlations between sensor parameters; Temporal attention employs causal convolution constraints to fuse only historical temporal information; Furthermore, the spatiotemporal decoupling attention mechanism of the STDAM module satisfies the following: the spatial attention submodule dynamically weights the characteristics of the cutterhead torque, soil chamber pressure, and thrust along the sensor dimension; the temporal attention submodule ensures that the attention weights conform to the physical causality law through causal convolution constraints.

[0023] Furthermore, it also includes the following online model update process based on the event-triggered mechanism (ETM): When the total prediction error

[0024] Exceeding the threshold At that time, the initial model was incrementally trained and its parameters were fine-tuned using the latest construction data. This is an adjustment factor used to adjust the tolerance for errors. Based on engineering experience, it is set to 0.1-0.3, or dynamically adjusted through Bayesian optimization. It is the total error in attitude prediction. This represents the true value of the tunnel boring machine's attitude. These are the predicted values ​​from the model.

[0025] Furthermore, the performance of the combined model is measured by the mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²).2 To conduct an assessment:

[0026]

[0027] ; in, For the number of test samples, For the first The true value of each sample For the first The model prediction value for each sample. The mean value is the average value of the sample, and it satisfies the following conditions: shield attitude prediction error ≤ 6mm, R2 ≥ 0.94.

[0028] A shield tunneling attitude prediction system, used in the method, includes: Multi-source data acquisition module: a sensor array deployed in key parts of the tunnel boring machine; Standardized processing unit: performs spatiotemporal alignment, outlier correction, and feature filtering; CNN-BiLSTM-STDAM prediction engine: Loads pre-trained models and outputs pose biases in real time; Adaptive update module: dynamically fine-tunes model parameters based on event-triggered mechanisms; Multi-level early warning execution unit: Links the electromechanical control system to implement a multi-level response strategy.

[0029] Furthermore, the prediction engine further includes: Batch normalization layers and the LeakyReLU activation function are used to solve the gradient vanishing problem. Dropout layer suppresses noise overfitting and generalizes abrupt geological conditions.

[0030] Furthermore, the adaptive update module includes: The base model pre-trained in the historical engineering database; Incremental learning interface, real-time integration of construction data to update weights.

[0031] Furthermore, the standardized processing unit specifically includes: The geological parameter module integrates data from ground-penetrating radar, earth pressure sensor, and pore water pressure sensor. The mechanical status module integrates data from strain gauges, rotary encoders, hydraulic sensors, and six-dimensional force sensors; The environmental monitoring module is linked to data from surface subsidence monitoring stations, distributed fiber optic sensing systems, and temperature and vibration composite sensors. The multi-level early warning execution unit executes a three-level response strategy: Level 1 warning: Absolute attitude deviation ≤ 10mm, triggering audible and visual alarm; Level 2 warning: Offset 10mm–25mm, generate multi-target optimized correction scheme; Level 3 warning: If the offset is greater than 25mm, activate the hard-wire emergency stop protection.

[0032] A computer-readable storage medium storing computer instructions, the steps of the method when the instructions are executed by the shield tunneling attitude prediction system.

[0033] Compared with existing technologies, this solution has the following advantages: (1) The prediction accuracy is significantly improved: the CNN-BiLSTM-STDAM model can stably control the shield attitude error within 6mm, which is 26%~52% higher than the traditional method, providing a reliable basis for the accurate assembly of tunnel segments; (2) Millisecond-level real-time response: edge inference combined with event-triggered online learning, model updates and correction instructions are completed in seconds under sudden working conditions, greatly reducing response latency; (3) Efficient fusion of multi-source data: A unified spatiotemporal alignment and normalization framework connects heterogeneous data links such as geological radar, hydraulic, and subsidence data, eliminating dimensional differences and data silos; (4) Rapid cross-scenario adaptation: The transfer + fine-tuning mechanism requires only a small number of new geological samples to reuse the model, which significantly reduces the cost of repeated training and deployment cycle; (5) Closed-loop safety control: Three-level early warning linkage with actuators such as cutterhead thrust and hydraulic cylinder zone pressure to achieve full-process automation of "perception-prediction-control" and effectively avoid the risk of over-limit tunneling. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a technical structure diagram of the method of the present invention; Figure 2 This is a flowchart illustrating the technical process of the method of the present invention. Figure 3 Comparison of attitude prediction curves for different prediction models; Figure 4 A comparison of the errors in attitude prediction results from different prediction models; Figure 5 Comparison of the mean absolute error of attitude prediction results from different prediction models; Figure 6 A data feature distribution map after data feature visualization; Figure 7 A distribution plot of the target variable after visualizing the data features; Figure 8 A correlation heatmap after visualizing data features; Figure 9 The scatter distribution and index values ​​of the prediction results for the test set; Figure 10 Comparison curves for prediction results on the test set; Figure 11 A comparison chart of prediction errors for the test set; Figure 12 This is a schematic diagram of the three-level early warning response for shield tunneling attitude. Detailed Implementation

[0036] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0037] Reference Figures 1-5 The present invention provides step S1: collecting geological parameters, mechanical state parameters and environmental parameters during shield tunneling construction through a multi-source sensor array to construct a multimodal spatiotemporal dataset; Step S2: Preprocess the multi-source heterogeneous data, including: performing spatiotemporal alignment, outlier removal, missing value imputation, normalization, and redundant feature dimensionality reduction on the multimodal spatiotemporal dataset to generate a standardized feature dataset; the spatiotemporal preprocessing specifically includes: Step S21: Use an adaptive sliding window for time series reconstruction and compensate for missing values ​​using cubic spline interpolation; Multi-scale standardization of parameters with different dimensions is performed based on the Min-Max normalization function, and the specific formula is as follows:

[0038] in, This represents the normalized eigenvalues, ranging from [0, 1]. This represents the maximum value of the feature across all samples. This represents the minimum value of the feature across all samples. This represents a specific feature value in the raw sensor data, such as earth pressure or torque. Step S22: Remove outlier data using a dynamic threshold filtering algorithm and fill the window with the mean of a sliding window, as shown in the following formula: ; in, This represents the average value within the i-th sliding window, used to fill in missing values; i represents the starting position of the current calculation, and its value range is 1 ≤ i ≤ n. k+1; k represents the width of the sliding window; Step S23: Based on Pearson correlation coefficient

[0039] in, The Pearson correlation coefficient represents the relationship between feature X and target variable Y, with a value range of [-1, 1].

[0040] xi,yi represent the observed values ​​of feature X and target variable Y in the i-th sample; , The sample mean of the representative feature X and the target variable Y; n represents the total number of samples; Calculate the feature and total attitude deviation

[0041] The correlation is used to generate a feature importance ranking matrix to eliminate redundant features; in, This represents the total deviation between the shield centerline and the tunnel axis. , These are the horizontal deviation of the shield head and the vertical deviation of the shield head. These are the horizontal deviation of the shield tail and the vertical deviation of the shield tail.

[0042] Step S3: Input the preprocessed data into the CNN-BiLSTM-STDAM combined model for attitude prediction. The CNN module extracts the spatial features of the data; the BiLSTM module models the temporal series features of the data bidirectionally; the STDAM module decouples the spatial and temporal features and dynamically weights and fuses them; the output is four predicted deviations of the shield attitude: shield head horizontal deviation, shield head vertical deviation, shield tail horizontal deviation, and shield tail vertical deviation, with 55...

[0043] The CNN module uses 64 sets of 3×1 convolutional kernels to extract spatial features of earth pressure distribution and hydraulic gradient; The BiLSTM module captures the temporal patterns of the tunneling direction and the lag effect of geological parameters through bidirectional LSTM layers. The STDAM module processes spatiotemporal features separately through a spatiotemporal decoupling attention mechanism: Spatial attention employs a multi-head self-attention mechanism to establish dynamic correlations between sensor parameters; Temporal attention employs causal convolution constraints to fuse only historical temporal information; Furthermore, the spatiotemporal decoupling attention mechanism of the STDAM module satisfies the following: the spatial attention submodule dynamically weights the characteristics of the cutterhead torque, soil chamber pressure, and thrust along the sensor dimension; the temporal attention submodule ensures that the attention weights conform to the physical causality law through causal convolution constraints.

[0044] Furthermore, it also includes the following online model update process based on the event-triggered mechanism (ETM): When the total prediction error

[0045] Exceeding the threshold

[0046] At that time, the initial model was incrementally trained and its parameters were fine-tuned using the latest construction data. Here, is an adjustment factor used to adjust the tolerance for error, is the total error of the predicted attitude, is the true value of the shield tunneling attitude, and is the predicted value of the model.

[0047] Step S4: Trigger a multi-level early warning mechanism based on the prediction results.

[0048] The following section will elaborate on several systems and mechanisms.

[0049] Multi-source heterogeneous data standardization processing system

[0050] This solution constructs a multimodal data acquisition system covering all elements of tunnel boring machine (TBM) construction, forming a three-dimensional perception system encompassing geology, machinery, and environment, providing a high-fidelity data foundation for attitude prediction. Specifically, the system deploys multi-type sensor arrays at key locations of the TBM. The geological parameter module uses forward-looking ground-penetrating radar to detect the distribution of rock strata interfaces and the location of weak interlayers within a range three times the tunnel diameter in front of the cutterhead in real time. It also monitors the pressure distribution at the excavation face using earth pressure sensors and integrates a pore water pressure sensor array to capture dynamic changes in groundwater levels. The machinery status module uses high-precision strain gauges and rotary encoders to collect the torque fluctuation curves of the cutterhead drive system in real time. It employs a distributed hydraulic sensor network to capture the pressure gradient distribution of the propulsion cylinders and installs a six-dimensional force sensor array at the hinge points to monitor the eccentric stress during attitude adjustment, while simultaneously acquiring parameters related to the screw conveyor speed and excavated soil flow. The environmental monitoring module relies on automatic ground settlement monitoring stations to construct a surface deformation early warning network. It combines a distributed fiber optic sensing system to perceive changes in groundwater levels around the tunnel axis in real time and deploys temperature and vibration composite sensors inside the TBM's internal compartments to capture environmental parameters related to equipment operation.

[0051] To address the spatiotemporal asynchronous problem of multi-source heterogeneous data, the system designs a spatiotemporal alignment preprocessing unit, employs an adaptive sliding window method to reconstruct the asynchronous data in time series, and uses a cubic spline interpolation algorithm to compensate for missing values ​​caused by sensor failures or communication interruptions. At the data standardization level, the system constructs a multi-scale normalization model, designing Min-Max normalization functions for parameters of different dimensions such as earth pressure, torque, and displacement, with the calculation formula as shown in equation (1). A dynamic threshold filtering algorithm is introduced to automatically identify and remove abnormal signals caused by interference or mechanical vibration. After removing abnormal data, the sliding window average is used to fill the gap to prevent data loss, calculated as shown in equation (2), where i represents the starting position of the current calculation, ranging from 1 to n-k+1. To further optimize the feature space, the system develops feature correlation analysis based on the Pearson correlation coefficient. By calculating the nonlinear dependency between attitude parameters and the original sensor data, a feature importance ranking matrix is ​​constructed, automatically removing redundant feature dimensions with low mutual information entropy values, ultimately forming an optimized feature dataset that retains key working condition information while reducing computational complexity.

[0052]

[0053]

[0054] The Pearson correlation coefficient, also known as the Pearson product-moment correlation coefficient, is used to measure the linear correlation between two variables X and Y. Its value ranges between -1 and 1. The closer the Pearson correlation coefficient is to 1, the stronger the positive correlation between the two characteristics; conversely, the closer it is to -1, the stronger the negative correlation. When its value is close to 0, it indicates a weak or no correlation between the two variables. Defined as the quotient of the covariance and standard deviation of two variables X and Y, the calculation formula is as follows (3).

[0055]

[0056] To select the feature variables that are highly correlated with the four attitudes of the tunnel boring machine (TBM), and considering the limitations of actual data types and ease of calculation, the total deviation of the TBM attitude is defined as follows.

[0057]

[0058] In the above formula, This represents the total deviation between the shield centerline and the tunnel axis. , These are the horizontal deviation of the shield head and the vertical deviation of the shield head. These are the shield tail horizontal deviation and the shield tail vertical deviation. Therefore, by simply calculating the Pearson correlation coefficient between each feature variable and the total deviation, the optimal input features can be quantitatively selected.

[0059] The specific steps include: S1: Geological, mechanical, and environmental multi-source data acquisition and storage S2: Data Spatiotemporal Alignment Processing S3: Data cleaning and noise reduction (outlier correction, missing value imputation, wavelet transform denoising, normalization) S4: Data Feature Visualization (Feature Distribution, Target Variable Distribution, Correlation Heatmap) (see details below) Figure 6 , Figure 7 and Figure 8 ).

[0060] Spatiotemporal feature extraction and decoupling attention mechanism

[0061] A CNN-BiLSTM-STDAM network architecture deeply integrates spatiotemporal features, forming a multi-level intelligent prediction engine for feature extraction and dynamic focusing. In the convolutional feature extraction layer design, the system adopts a 1D-CNN channel architecture, deploying 64 sets of 3×1 convolutional kernels for sliding window scanning based on the characteristics of the shield tunneling construction time-series data stream. Spatial feature patterns such as earth pressure distribution and hydraulic gradient are extracted layer by layer through local receptive fields. Each convolutional layer is cascaded with batch normalization layers to perform zero-mean processing on the feature map, and a nonlinear mapping is constructed using the LeakyReLU activation function, effectively solving the gradient vanishing problem in deep network training. Based on this, the system innovatively designs a bidirectional temporal modeling unit. The forward LSTM layer extracts the temporal evolution of parameters such as hydraulic system pressure and cutterhead speed along the tunneling direction, while the backward LSTM layer captures the lag effect of geological parameters through backpropagation. A Dropout layer is introduced at the output of the bidirectional LSTM to suppress overfitting to noisy features by randomly masking neuron connections, enhancing the model's generalization ability under abrupt geological conditions.

[0062] To dynamically focus on key working condition features, the system employs a spatiotemporal decoupled attention mechanism to extract spatiotemporal features from the data. Spatial attention utilizes a multi-head self-attention mechanism, establishing dynamic correlations between spatial features such as cutterhead torque, soil chamber pressure, and thrust along the sensor parameter dimensions. Temporal attention introduces causal convolution constraints, allowing only historical time-series information to participate in the current prediction, ensuring that the attention weights conform to the physical causality of the tunnel boring process. Finally, the spatiotemporal features are adaptively fused, avoiding gradient interference caused by the confusion in modeling spatiotemporal features and reducing computational complexity through parameter sharing. Ultimately, a tunnel boring machine attitude prediction model based on a CNN-BiLSTM-STDAM network structure is generated.

[0063] To visually demonstrate the effectiveness of the neural network model, this study selected Mean Squared Error (MSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²) as evaluation metrics for model training performance. Mean Squared Error is a metric that measures the difference between predicted and true values, directly reflecting the overall deviation of the model's predictions. Its calculation formula is shown in equation (5). A smaller MSE indicates a more accurate model. Mean Absolute Error is another important metric. Compared to Mean Squared Error, it is more robust to outliers in the data. Its calculation formula is shown in equation (6). The Coefficient of Determination is an important indicator for evaluating model fit, quantifying the model's ability to explain changes in the target variable. Its calculation formula is shown in equation (7). The closer its value is to 1, the higher the model's fit to the data.

[0064]

[0065]

[0066]

[0067] In the above formula, For the number of test samples, For the first The true value of each sample For the first The model prediction value for each sample. This represents the average value of the samples. The training results of the CNN-BiLSTM-STDAM network model are as follows: Figures 9-11 As shown.

[0068] In addition to the training results mentioned above, two commonly used neural network models were selected for comparison: the CNN-BiLSTM-AM model and the CNN-BiLSTM model. The comparison results are shown in the figures at the end of the article. According to the results, compared to the CNN-BiLSTM model without an attention mechanism, the STDAM model and the single AM ​​model both show significant improvements in prediction performance. Compared to the traditional AM model, the STDAM model is only slightly inferior in predicting HDST (head vertical deviation), while all other results show slight improvements. The mean absolute errors for predicting head horizontal deviation, tail horizontal deviation, and tail vertical deviation are reduced by 9.8%, 6.6%, and 51.8%, respectively. Compared to the CNN-BiLSTM model, the model proposed in this study performs better in predicting all four orientations, with the mean absolute errors for predicting head horizontal deviation, head vertical deviation, tail horizontal deviation, and tail vertical deviation reduced by 26.0%, 37.9%, 45.1%, and 45.7%, respectively.

[0069] Table 1 Comparison of metrics for different network models

[0070] Note: The plus or minus sign in the first column of the table indicates the comparison of the prediction results for the same pose. For example, "4.9150 (-+)" means that the result is worse than the prediction result of CNN-BiLSTM-AM, but better than the prediction result of CNN-BiLSTM.

[0071] Online learning-based model adaptive update system

[0072] This invention establishes a continuous optimization capability for the shield tunneling attitude prediction model by constructing an adaptive model fine-tuning training mechanism. The system first trains a basic model (initial model) on an engineering history database, which can predict the current shield tunneling attitude. Simultaneously, an event-triggered mechanism (ETM)-based model self-update mechanism is introduced, expressed as follows (8): It is an adjustment factor used to adjust the tolerance for error. It is the total error of the predicted attitude, defined by equation (9), where This represents the true value of the tunnel boring machine's attitude. This represents the model's predicted value. The meaning of this formula is that when the total prediction error of the initial model exceeds a set threshold, meaning the initial model cannot meet the accuracy requirements of subsequent construction, a second training phase will be initiated, using the latest construction data to fine-tune the model.

[0073]

[0074]

[0075] Attitude deviation early warning mechanism

[0076] By constructing an early warning system that integrates multimodal perception and intelligent decision-making, the system achieves real-time monitoring, dynamic evaluation, and adaptive response capabilities for shield tunneling attitude deviation. When the real-time attitude trajectory exceeds a safety threshold, the system activates a multi-level response mechanism, such as: Level 1 warning stage (absolute attitude deviation less than or equal to 10mm), Level 2 warning stage (absolute attitude deviation within the range of 10mm-25mm), and Level 3 warning stage (absolute attitude deviation exceeding 25mm), as follows. Figure 12 As shown.

[0077] Specifically, the steps include the following: S1: Multi-source sensor data acquisition and transmission S2: Data Preprocessing S3: Load the neural network model S4: Output predicted pose S5: Prediction Result Determination and Alarm Response Further, the multi-source sensor data acquisition in step S1 includes: soil type data, shield attitude data, earth pressure data, cutterhead rotation speed, cutterhead torque, jack pressure, total thrust, hinge stroke, etc. In step S2, the acquired multi-source sensor data is preprocessed, including spatiotemporal alignment, outlier / missing value handling, normalization, redundant data removal, and data feature visualization to ensure data integrity and accuracy. In step S3, the trained combined neural network model is loaded, and the data processed in step S2 is input into the neural network model to output the final prediction result. In step S5, the prediction result is judged, and an alarm response is executed.

[0078] application server

[0079] The shield tunneling attitude prediction technology server based on a combined neural network model specifically includes: a data acquisition and integration module, a storage and logging module, a data preprocessing module, a data caching and visualization module, a neural network pre-training module, a model transfer and fine-tuning module, a result visualization module, a model update / save module, a model loading and output module, an early warning module, and a data analysis and reporting module.

[0080] The data acquisition and integration module integrates raw data from various multi-sensor acquisitions; the storage and logging module saves and records raw data and construction operations; the data preprocessing module is a prerequisite for neural network training, reducing noise and correcting raw data through preprocessing to improve the network's prediction accuracy; the model transfer and fine-tuning module is crucial for improving the prediction accuracy of the prediction model under complex geological conditions, improving the model's generalization performance through fine-tuning training; the results visualization module displays the network training results, including but not limited to error curves and root mean square error; the model update / save module saves the network training results and incremental training results for easy user access later; the model loading and output module predicts an output result based on the collected real-time data using the trained model; the early warning module determines whether the early warning conditions are met based on the output result according to a preset threshold, triggering an alarm when the threshold is reached; finally, the data analysis and reporting module records attitude prediction deviations and actual shield deviations, providing data support for subsequent compensation of prediction errors and real-time correction of prediction results.

[0081] This invention aims to provide a shield tunneling attitude prediction method and early warning system based on a CNN-BiLSTM-STDAM neural network model, addressing the problems of low accuracy, insufficient real-time performance, difficulty in fusing multi-source data features, and weak model generalization ability in existing technologies. Traditional shield tunneling attitude prediction methods mainly rely on empirical formulas or single sensor data analysis, resulting in defects such as weak model generalization ability, insufficient capture of nonlinear features, and lack of spatiotemporal correlation modeling, making it difficult to cope with the dynamic changes during shield tunneling under complex geological conditions. Furthermore, existing methods lack an effective fusion mechanism for multimodal data (such as earth pressure parameters, hydraulic system status, and shield machine parameters), leading to significant data silos and prediction results lagging behind actual working conditions. Under sudden ground disturbances or abnormal mechanical conditions, traditional linear models struggle to respond quickly, easily causing safety hazards such as excessive shield attitude deviation and segment misalignment, seriously affecting the safety and construction efficiency of underground engineering.

[0082] This application proposes a solution that overcomes the limitations of traditional methods by constructing a deep learning model that integrates a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and a spatiotemporally decoupled attention mechanism (STDAM). Addressing the need for multi-dimensional spatiotemporal feature extraction of tunnel boring machine (TBM) attitude, the CNN module effectively mines the spatial correlations of multi-source data such as geological parameters and mechanical states; the BiLSTM module captures the long- and short-term dynamic evolution patterns during TBM advancement through bidirectional temporal modeling; and the spatiotemporally decoupled attention mechanism separately processes the temporal and spatial features of the data, preventing model performance degradation caused by feature confusion. By establishing an end-to-end TBM attitude prediction framework, high-precision real-time prediction of TBM construction attitude is achieved, providing advanced decision-making basis for dynamically adjusting tunneling parameters.

[0083] This invention proposes a unified, standardized multimodal data processing workflow to address the fusion challenges caused by differences in sampling frequencies and inconsistencies in units of measurement among heterogeneous sensor data, and constructs a multidimensional feature database encompassing geology, machinery, and the environment. Combined with model transfer and fine-tuning techniques, the model can quickly adapt to different geological conditions, significantly improving the generalization ability of the prediction system. By introducing an online learning mechanism, the system can update model parameters in real time by integrating the latest construction data, enhancing its adaptability to unknown working conditions.

[0084] This proposed solution can be widely applied to shield tunneling scenarios such as urban subways, river-crossing tunnels, and integrated utility tunnels, providing technical support for achieving intelligent tunneling control, reducing construction risks, and saving engineering costs. By linking the prediction results with the early warning system, a closed-loop solution of "perception-prediction-early warning" can be formed, promoting the transformation and upgrading of underground engineering construction towards digitalization and self-adaptation.

[0085] The innovativeness of this application lies in: (1) Spatiotemporal data alignment preprocessing mechanism: unify data from different sensors with different spatiotemporal resolutions, sampling frequencies or reference bases into a consistent spatiotemporal framework so as to carry out subsequent joint analysis, modeling or visualization; (2) Spatiotemporal feature “decoupling-fusion” mechanism: an attention mechanism for spatiotemporal feature decoupling is introduced to process the temporal and spatial features of shield tunneling historical data separately to prevent feature confusion, and finally feature fusion is performed; (3) Enhanced engineering adaptability: Design a multimodal data coupling architecture of geology-machine-environment to improve the prediction stability under complex geological conditions;

[0086] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the attitude of tunnel boring machines based on a CNN-BiLSTM-STDAM combined model, characterized in that, include: Step S1: Collect geological parameters, mechanical state parameters, and environmental parameters during shield tunneling construction using a multi-source sensor array to construct a multimodal spatiotemporal dataset; Step S2: Preprocess the multi-source heterogeneous data, including: performing spatiotemporal alignment, outlier removal, missing value filling, normalization, and redundant feature dimensionality reduction on the multimodal spatiotemporal dataset to generate a standardized feature dataset. Step S3: Input the preprocessed data into the CNN-BiLSTM-STDAM combined model for attitude prediction. The CNN module extracts the spatial features of the data; the BiLSTM module models the temporal series features of the data in both forward and reverse directions; the STDAM module decouples the spatial and temporal features and dynamically weights and fuses them; and outputs four predicted deviations of the shield attitude: shield head horizontal deviation, shield head vertical deviation, shield tail horizontal deviation, and shield tail vertical deviation. Step S4: Trigger a multi-level early warning mechanism based on the prediction results.

2. The shield tunneling attitude prediction method based on the CNN-BiLSTM-STDAM combined model according to claim 1, characterized in that, The spatiotemporal preprocessing in step S2 specifically includes: Step S21: Use an adaptive sliding window for time series reconstruction and compensate for missing values ​​using cubic spline interpolation; Multi-scale standardization of parameters with different dimensions is performed based on the Min-Max normalization function, and the specific formula is as follows: ; in, This represents the normalized eigenvalues, ranging from [0, 1]. This represents the maximum value of the feature across all samples. This represents the minimum value of the feature across all samples. This represents a specific feature value in the raw sensor data, such as earth pressure or torque. Step S22: Remove outlier data using a dynamic threshold filtering algorithm and fill the window with the mean of a sliding window, as shown in the following formula: ； in, This represents the average value within the i-th sliding window, used to fill in missing values; i represents the starting position of the current calculation, and its value range is 1 ≤ i ≤ n. k+1; k represents the width of the sliding window; Step S23: Based on Pearson correlation coefficient in, The Pearson correlation coefficient represents the relationship between feature X and target variable Y, with a value range of [-1, 1]. xi,yi represent the observed values ​​of feature X and target variable Y in the i-th sample; , The sample mean of the representative feature X and the target variable Y; n represents the total number of samples; Calculate the feature and total attitude deviation The correlation is used to generate a feature importance ranking matrix to eliminate redundant features; in, This represents the total deviation between the shield centerline and the tunnel axis. , These are the horizontal deviation of the shield head and the vertical deviation of the shield head. These are the horizontal deviation of the shield tail and the vertical deviation of the shield tail.

3. The shield tunneling attitude prediction method based on the CNN-BiLSTM-STDAM combined model according to claim 1, characterized in that, In step S3 The CNN module uses 64 groups of 3×1 convolutional kernels to extract spatial features of earth pressure distribution and hydraulic gradient. The BiLSTM module captures the temporal patterns of the tunneling direction and the lag effect of geological parameters through bidirectional LSTM layers. The STDAM module processes spatiotemporal features separately through a spatiotemporal decoupling attention mechanism: Spatial attention employs a multi-head self-attention mechanism to establish dynamic correlations between sensor parameters; Temporal attention employs causal convolution constraints to fuse only historical temporal information; Furthermore, the spatiotemporal decoupling attention mechanism of the STDAM module satisfies the following: the spatial attention submodule dynamically weights the characteristics of the cutterhead torque, soil chamber pressure, and thrust along the sensor dimension; The temporal attention submodule ensures that the attention weights conform to physical causality through causal convolution constraints.

4. The shield tunneling attitude prediction method based on the CNN-BiLSTM-STDAM combined model according to claim 1, characterized in that, Also includes: The online model update process based on the event-triggered mechanism (ETM) is as follows: When the total prediction error Exceeding the threshold At that time, the initial model was incrementally trained and its parameters were fine-tuned using the latest construction data. It is an adjustment factor used to adjust the tolerance for error. It is the total error in attitude prediction. This represents the true value of the tunnel boring machine's attitude. These are the predicted values ​​from the model.

5. The shield tunneling attitude prediction method based on the CNN-BiLSTM-STDAM combined model according to claim 1, characterized in that, The performance of the combined model is measured by mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). 2 To conduct an assessment: ； in, For the number of test samples, For the first The true value of each sample For the first The model prediction value for each sample. The mean value is the average value of the sample, and it satisfies the following conditions: shield attitude prediction error ≤ 6mm, R2 ≥ 0.

94.

6. A shield tunneling attitude prediction system, used to implement the method described in any one of claims 1-5, characterized in that, include: Multi-source data acquisition module: a sensor array deployed in key parts of the tunnel boring machine; Standardized processing unit: performs spatiotemporal alignment, outlier correction, and feature filtering; CNN-BiLSTM-STDAM prediction engine: Loads pre-trained models and outputs pose biases in real time; Adaptive update module: dynamically fine-tunes model parameters based on event-triggered mechanisms; Multi-level early warning execution unit: Links the electromechanical control system to implement a multi-level response strategy.

7. The shield tunneling attitude prediction system according to claim 6, characterized in that, The prediction engine further includes: Batch normalization layers and the LeakyReLU activation function are used to solve the gradient vanishing problem. Dropout layer suppresses noise overfitting and generalizes abrupt geological conditions.

8. A shield tunneling attitude prediction system according to claim 6, characterized in that, The adaptive update module includes: The base model pre-trained in the historical engineering database; Incremental learning interface, real-time integration of construction data to update weights.

9. The shield tunneling attitude prediction system according to claim 6, characterized in that, The standardization processing unit specifically includes: The geological parameter module integrates data from ground-penetrating radar, earth pressure sensor, and pore water pressure sensor. The mechanical status module integrates data from strain gauges, rotary encoders, hydraulic sensors, and six-dimensional force sensors; The environmental monitoring module is linked to data from surface subsidence monitoring stations, distributed fiber optic sensing systems, and temperature and vibration composite sensors. The multi-level early warning execution unit executes a three-level response strategy: Level 1 warning: Absolute attitude deviation ≤ 10mm, triggering audible and visual alarm; Level 2 warning: Offset 10mm–25mm, generate multi-target optimized correction scheme; Level 3 warning: If the offset is greater than 25mm, activate the hard-wire emergency stop protection.

10. A computer-readable storage medium storing computer instructions, characterized in that, When the instruction is executed by the shield tunneling attitude prediction system as described in claims 6 to 9, it implements the steps of the method described in any one of claims 1 to 5.

Images