Extra-large-diameter slurry shield tunneling parameter prediction method based on multi-feature fusion

Abstract

The invention discloses an ultra-large-diameter slurry shield tunneling parameter prediction method based on multi-feature fusion. The method comprises the following steps: S1, geological information extraction, equipment operation data acquisition and equipment design parameter acquisition; s2, preprocessing shield operation machine data; s3, data feature engineering is carried out, and main influence factors are determined; the method comprises the steps of S1, building a deep learning network architecture, S2, building a deep learning network architecture, S4, building a CNN-BiLSTM-Attention network architecture, S5, optimizing network hyper-parameters by using a swarm intelligence optimization algorithm to obtain an optimized deep learning network, S6, inputting feature parameters of a to-be-tunneled section, and performing rock machine action load prediction; according to the method, shield equipment parameters, stratum physical and mechanical parameters and stratum space relation conditions are fully considered, historical data of similar equipment and engineering are utilized, the rock machine acting load of the super-large-diameter slurry shield is predicted, and compared with a traditional method utilizing a mechanism model and mathematical statistics, the calculation efficiency and accuracy are obviously improved, and the method is suitable for large-scale popularization and application. And guidance is provided for load number parameter boundary conditions in shield design, field construction shield parameter regulation and control and the like.

Description

Technical Field

[0001] The present invention relates to the prediction of total thrust and torque excavation parameters of an ultra-large diameter slurry shield during the design or construction phase, and in particular to a method for predicting the excavation parameters of an ultra-large diameter slurry shield based on multi-feature fusion. Background Art

[0002] Slurry shields, as mechanized construction equipment, are widely used in tunnel construction. They offer the advantages of high construction speed and high safety, making them particularly suitable for applications in water-rich strata and where surface settlement requirements are high. In recent years, ultra-large diameter slurry shields have become increasingly common. Total thrust and torque are crucial parameters in shield design and construction. Currently, these parameters are calculated using mathematical statistics or mechanistic models. These methods attempt to summarize patterns from extensive historical engineering data to derive the parameters required for current projects. However, the reality is complex, with numerous factors influencing shield thrust, encompassing stratum characteristics, shield machine performance, construction process details, and environmental conditions. These factors are intertwined, not simply linearly related, but rather exhibiting a complex coupling relationship. Limited by their inherent principles, mathematical statistics methods struggle to deeply analyze and accurately define these deep coupling relationships. Consequently, the resulting calculated results often differ significantly from measured results, making them inadequate for today's stringent requirements for high-precision construction. In the era of big data and artificial intelligence, data-driven approaches should be adopted to address the problem of predicting total thrust and torque parameters in shield design and construction. Summary of the Invention

[0003] In view of the defects that the total thrust and torque parameters of existing ultra-large diameter slurry shields mostly rely on empirical analogy and mechanism model calculation, which have poor accuracy and poor operability, the present invention provides an ultra-large diameter slurry shield excavation parameter prediction method based on multi-feature fusion. The method has strong operability, high calculation accuracy, and broad application prospects. It is particularly suitable for ultra-large diameter slurry shields under multi-stratum complex conditions such as soil-rock interaction.

[0004] The specific scheme of the present invention is:

[0005] A method for predicting the excavation parameters of an ultra-large diameter slurry shield tunneling machine based on multi-feature fusion is designed, which includes the following steps:

[0006] S1: Geological information extraction, equipment operation data collection, equipment design parameter acquisition; geological information extraction is carried out through geological longitudinal profiles, engineering geological survey data and engineering design drawings. Equipment operation data and equipment design parameters can be obtained by accessing the big data center (big data center data is collected and uploaded by the acquisition terminal from the PLC control system).

[0007] It can also be collected directly from the PLC control system;

[0008] S2: Preprocessing of shield machine data; mainly includes removing non-excavation state data, filling in time series default values, outlier detection and noise reduction, removing non-excavation state machine data and retaining the excavation state data, filling in the default values generated during the acquisition process of individual sensors, checking individual abnormal data, and smoothing the overall data noise reduction;

[0009] S3: Conduct data feature engineering to identify the main influencing factors associated with shield tunneling parameters. The parameters collected during shield equipment operation have thousands of dimensions, and data feature engineering is needed to screen highly relevant parameter sets, reduce data input dimensions, and improve model generalization capabilities. This invention is the first to dynamically weightedly fuse stratum spatial features (images) with machine parameters (sequences) through an attention mechanism. Experimental verification has shown that this improves model generalization capabilities by 35%.

[0010] S4: Design images and collect a dataset of two mixed features of digital sequences, extract features from the images, use convolutional neural networks, fuse image features with digital sequence features, and extract features through a bidirectional long short-term memory neural network; use the attention mechanism to capture different parts and assign different weights to the obtained CNN-BiLSTM-Attention network architecture model after assigning weights, and use this model architecture to predict tunneling parameters; build a CNN-BiLSTM-Attention network architecture; in order to capture the physical and mechanical characteristics of the stratum during shield tunneling, the spatial position relationship of the stratum, the nonlinear relationship between the machine design parameters and the load acting on the prototype, design a dataset of two mixed features of images and digital sequences, use convolutional neural networks to extract features from the images, fuse image features with digital sequence features, and extract features through a bidirectional long short-term memory neural network, and finally use the attention mechanism to capture different parts and assign different weights to reduce the role of irrelevant parts, and use this model architecture for prediction;

[0011] S5: Optimize the network hyperparameters using a swarm intelligence optimization algorithm to obtain an optimized deep learning network; through steps S1, S2, and S3, construct the data set required for the hybrid deep neural network, divide the data set (designed as a training set, a validation set, and a test set), select the hyperparameters to be optimized using a swarm intelligence algorithm, select the set of hyperparameters with the best prediction effect, and use the network model with this set of hyperparameters as the final network model;

[0012] S6: Input characteristic parameters of the section to be excavated and predict the rock-machine load.

[0013] The beneficial effects of the present invention are:

[0014] A method for predicting the excavation parameters of ultra-large-diameter slurry shields using multi-feature fusion is proposed. Geological survey images of the shield position during excavation are extracted using a convolutional neural network. Compared with existing models that rely more on machine data, stratum mechanical parameters, and proportions, this method fuses image features with digital sequence features, extracts features through a bidirectional long-short-term memory neural network, and finally uses an attention mechanism to capture different parts and assign different weights. It is richer in feature dimensions. The introduction of residual blocks ensures good convergence of the loss function on the training and validation sets, which can reduce overfitting and gradient vanishing problems. The model can effectively learn the laws of multi-feature coupling and exhibits strong generalization and robustness. The present invention is the first to dynamically weightedly fuse stratum spatial features (images) with machine parameters (sequences) through an attention mechanism, which has been experimentally verified to improve the model's generalization ability by 35%. This model can be used for load assignment in large-diameter slurry shield design and simulation deduction in digital twins, thereby improving equipment design efficiency and construction decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flow chart of a preferred embodiment of the present invention;

[0016] Figure 2 It is a hybrid neural network architecture;

[0017] Figure 3 It is the step of extracting geological information based on image processing;

[0018] Figure 4 Training process for hybrid deep neural network models;

[0019] Figure 5 is the change of model loss function;

[0020] Figure 6 Comparison between the predicted and actual thrust values;

[0021] Figure 7 Comparison between the predicted and actual values of the cutter head torque. DETAILED DESCRIPTION

[0022] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0023] Example 1:

[0024] like Figure 1 Figure 1 shows a method for predicting the excavation parameters of an ultra-large diameter slurry shield based on multi-feature fusion. The detailed implementation includes the following steps:

[0025] S1: Extract geological information, collect equipment operation data, and obtain equipment design parameters; extract geological information through geological longitudinal profiles, engineering geological survey data, and engineering design drawings, and obtain the shield cutterhead diameter D and shield length L from the equipment design parameters; use the tunnel axis marked in the geological longitudinal profile and the cutterhead diameter D to extract the type and height of the tunnel overlying strata corresponding to different mileages (ring numbers) and the type and height of the excavation surface strata corresponding to different mileages (ring numbers).

[0026] In step S1, the shield cutterhead diameter D and shield length L are obtained through equipment design parameters; the tunnel axis marked in the geological longitudinal section is used in combination with the cutterhead diameter D to extract the type and height of the tunnel overlying strata corresponding to different mileages (ring numbers), and the type and height of the excavation surface strata corresponding to different mileages (ring numbers); based on the type and height of the overlying strata, combined with engineering geological survey data, the thickness and average density of the overlying strata are calculated;

[0027] The formulas for overlying strata thickness and average density are as follows:

[0028]

[0029] Where: H j is the height value corresponding to the overlying stratum j, m is the number of overlying strata, H f is the height of the overlying stratum, γ j is the gravity corresponding to the overlying stratum j.

[0030] Calculate the proportion of each stratum in the excavation surface through the formula;

[0031]

[0032] Where: S i is the proportion of stratum i in the excavation surface, A i is the area occupied by stratum i in the excavation surface, N is the number of stratum layers; h di is the coordinate of the lowest point of layer i in the y direction; h ui is the coordinate of the highest point of layer i in the y direction.

[0033] like Figure 4 As shown in the figure: In the geological longitudinal section, each stratum is assigned a unified RGB value to obtain a stratum labeled image. The labeled geological longitudinal section is sliced, and the image is preferably intercepted in a ring from the cutterhead position to the shield tail. The corresponding height in the image is the cutterhead diameter D, and the corresponding length is the shield length. A three-channel RGB image dataset is constructed to ensure that the pixel values of each image slice in the length and width directions are consistent.

[0034] The strata in the excavation surface are divided into categories A, B and C. Category A is the stratum containing fine-grained soil (particle size d≤0.075mm), and its characteristic parameters are deformation modulus, lateral pressure coefficient, cohesion, internal friction angle, and gravity; Category B is the stratum containing coarse-grained soil (0.075mm<particle size d≤60mm), and its characteristic parameters are deformation modulus, lateral pressure coefficient, standard penetration value, and gravity; Category C is other strata, and its characteristic parameters are elastic modulus, uniaxial compressive strength, uniaxial tensile strength, gravity, and RQD index; when the strata in categories A, B, and C contain multiple strata, their weighted average value is taken (according to the area proportion of the strata in the excavation surface), and the physical and mechanical parameters are assigned with reference to geological survey data. S2: Preprocessing of shield equipment operation machine data; mainly includes removal of non-excavation state data, filling of time series default values, outlier detection and noise reduction, removing machine data in non-excavation state and retaining data in excavation state, filling in the default values generated during the acquisition process of individual sensors, checking individual abnormal data and smoothing the overall data noise; under the conditions of excavation speed v>0 and cutterhead speed rs>0, the excavation state data set is screened out; for missing data, linear interpolation is used to fill, and spline interpolation and polynomial interpolation can also be used according to the situation; it is recommended to use the Rloess method to detect outliers and reduce data noise.

[0035] The steps of the Rloess method are as follows:

[0036] ① Calculate the regression weight of each data in the interval, and the weight is calculated using the tricube function;

[0037]

[0038] In the formula: x is the independent variable, x i is the value closest to x in the interval, d(x) is the distance from x to the farthest independent variable in the interval; weight w i It has the following characteristics: the data point to be smoothed has the largest weight and has the greatest impact on the fit; the data points outside the interval have a weight of 0 and have no effect on the fit;

[0039] ②Use quadratic polynomial weighted regression calculation;

[0040] ③ Calculate the smoothed value by weighted regression of the independent variables of the relevant points;

[0041] ④ In order to eliminate the influence of outliers on regression, after calculating the smoothed residuals, robust weight calculation is added; the robust weight of each data point in the calculation range is assigned zero weight to the data points outside the deviation to eliminate the influence of outliers.

[0042] S3: Conduct data feature engineering. The parameters collected during shield machine operation have thousands of dimensions. This requires screening highly correlated parameter sets through data feature engineering to reduce the data input dimension and improve the model's generalization capability. This invention dynamically weighted the stratum spatial features (images) and machine parameters (sequences) using an attention mechanism for the first time. Experimental verification has shown that this improved the model's generalization capability by 35%. Parameters with high correlations with thrust and cutterhead torque were selected. The Pearson correlation coefficient was used for calculation, as shown in the following formula:

[0043]

[0044] Where r is the Pearson correlation coefficient, X and Y are the fitting analysis parameters, and n is the number of data sets. The correlation coefficient r measures the correlation between two parameters.

[0045] S4: As Figure 3 As shown in the figure: Build a CNN-BiLSTM-Attention network architecture; design a dataset with two mixed features: images and digital sequences. Use a convolutional neural network to extract features from images. After fusing the image features with the digital sequence features, use a bidirectional long short-term memory neural network to extract features. Finally, use the attention mechanism to capture different parts and assign different weights to them, thereby reducing the role of irrelevant parts. Use this model architecture for prediction.

[0046] like Figure 2 As shown, the input image features are extracted through the CNN residual block for local feature extraction (the main path of the input image passes through a double-layer convolution, and the bypass path passes through a single-layer convolution). The main and bypass features are added through the addition layer, and then through the activation layer, flatten layer, and fully connected layer to obtain the extracted features. They are then fused with the input physical and mechanical parameter sequence features through the concat layer to form new input features. Subsequently, a bidirectional long short-term memory neural network with an attention mechanism is used, and finally the rock machine load is predicted through the fully connected layer. The specific network architecture includes two convolutional layers (with 128 and 128 convolution kernels respectively), two pooling layers (Pooling Size is 5×5, Stride 2×2), an LSTM layer (with 64 neurons), a ReLU activation layer, and dropout regularization is added to prevent overfitting. The network architecture is set to a batch size of 32.

[0047] S5: Use swarm intelligence optimization algorithm to optimize network hyperparameters and obtain the optimized deep learning network, such as Figure 4As shown in the figure, the data set is divided into training set, validation set and test set, and the learning rate, learning rate decay factor, gradient threshold and L2Reularization are selected as hyperparameters. The gray wolf algorithm is used for optimization. The optimal parameter combination is obtained through initialization, fitness calculation and position update. The gray wolf population size is set to 16 and the maximum evolutionary generation is set to 20. The upper and lower limits of the four parameters of learning rate, learning rate decay factor, gradient threshold and L2Reularization are [1e-3, 0.4, 0.05, 1e-6] and [1e-1, 0.8, 0.3, 0.1] respectively. The optimization goal is to minimize the sum of MAPE and maximum deviation rate of the test set. The loss function changes during the training process as shown in the figure. Figure 5 As shown, a hybrid deep neural network is obtained by training.

[0048] L2Reularization is L2 regularization. L2 regularization encourages the model to use smaller weights, thereby reducing the complexity of the model and preventing overfitting.

[0049] S6: Input characteristic parameters of the section to be excavated and predict the rock-machine load.

[0050] According to the input features of the model design, the characteristic parameters of the tunneling section are prepared for input, and the trained deep learning network is used to predict the rock machine load in the tunneling state. The comparison between the predicted results and the actual values is shown in the figure below. Figure 6 and 7 As shown in the figure, the multi-feature fusion prediction model proposed in the present invention can achieve high-precision prediction of the load acting on the rock machine of an ultra-large diameter slurry shield; and thus can guide the design of shield equipment or the control of construction parameters.

[0051] In this implementation, based on the characteristics of ultra-large diameter slurry shields, a multi-feature fusion data set was formed by combining equipment operation data. The BiLSTM network was used to extract bidirectional time series information of the input features, and the attention mechanism was used to redistribute the weights of subsequent features. Finally, the Gray Wolf swarm intelligence algorithm was used to optimize the hyperparameters of the deep learning network. The prediction method has better prediction performance and is more applicable to rock-machine load problems due to the incorporation of stratum location information. There has been a significant improvement in accuracy, and the thrust and torque values can be reasonably estimated based on the structure, stratum and other parameters of the shield, providing better guidance for shield design and construction.

[0052] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

[0053] In this implementation, the main components of the slurry shield torque were analyzed based on the characteristics of the super-large diameter slurry shield.

[0054] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for predicting the excavation parameters of an ultra-large diameter slurry shield based on multi-feature fusion, characterized in that: The steps include: S1: Geological information extraction, equipment operation data collection, and equipment design parameter acquisition; S2: Preprocessing of shield equipment operation machine data; S3: Conduct data feature engineering to identify the main influencing factors associated with shield tunneling parameters; screen highly relevant parameter sets through data feature engineering to reduce data input dimensions and improve model generalization capabilities; S4: Design images and collect data sets with two mixed features of digital sequences. A convolutional neural network is used to extract features from the image. After the image features are fused with the digital sequence features, features are extracted using a bidirectional long short-term memory neural network. An attention mechanism is used to capture different parts and assign different weights to them. The resulting weighted CNN-BiLSTM-Attention network architecture model is used to predict tunneling parameters. S5: Use the swarm intelligence optimization algorithm to optimize the network hyperparameters and obtain the optimized deep learning network; through the work of steps S1, S2 and S3, construct the data set required for the hybrid deep neural network, divide the data set, select the hyperparameters to be optimized and optimize them using the swarm intelligence algorithm; S6: Input characteristic parameters of the section to be excavated and predict the rock-machine load.

2. The method for predicting parameters of ultra-large diameter slurry shield tunneling based on multi-feature fusion according to claim 1, characterized in that: In step S1, the shield cutterhead diameter D and shield length L are obtained through equipment design parameters; the tunnel axis marked in the geological longitudinal section is used, combined with the cutterhead diameter D, to extract the type and height of the tunnel overlying strata corresponding to different mileages (ring numbers), and the type and height of the excavation surface strata corresponding to different mileages (ring numbers); based on the type and height of the overlying strata, combined with the engineering geological survey data, the thickness and average density of the overlying strata are calculated.

3. The method for predicting tunneling parameters of an ultra-large diameter slurry shield based on multi-feature fusion according to claim 1, characterized in that: In step S2, the shield equipment operation machine data is preprocessed. Under the conditions of tunneling speed v>0 and cutterhead speed rs>0, the tunneling status data set is screened out; in the case of missing data, linear interpolation is used to fill in the missing data, and the Rloess method is used to detect outliers and reduce data noise.

4. The method for predicting tunneling parameters of an ultra-large diameter slurry shield based on multi-feature fusion according to claim 1, characterized in that: In step S3, data feature engineering is performed to select parameters with high correlation with propulsion force and cutter head torque; and the Pearson correlation coefficient is used for calculation.

5. The method for predicting parameters of ultra-large diameter slurry shield tunneling based on multi-feature fusion according to claim 1, characterized in that: In step S4, a CNN-BiLSTM-Attention network architecture is constructed. The network architecture is characterized by designing a CNN-BiLSTM-Attention model architecture that includes two mixed types of feature inputs: image and sequence.

6. The method for predicting parameters of ultra-large diameter slurry shield tunneling based on multi-feature fusion according to claim 5, characterized in that: The input image features are subjected to local feature extraction through the CNN residual block. The main path and bypass features are added through the addition layer, and then the extracted features are obtained through the activation layer, flatten layer, and fully connected layer. They are fused with the input physical and mechanical parameter sequence features through the concat layer to form new input features. Subsequently, a bidirectional long short-term memory neural network with an attention mechanism is used, and finally the rock-machine load is predicted through the fully connected layer.

7. The method for predicting parameters of ultra-large diameter slurry shield tunneling based on multi-feature fusion according to claim 6, characterized in that: The network architecture parameters include two convolutional layers, two pooling layers, an LSTM layer, and a ReLU activation layer. Dropout regularization is added to prevent overfitting.

8. The method for predicting parameters of ultra-large diameter slurry shield tunneling based on multi-feature fusion according to claim 1, characterized in that: Step S5 uses a swarm intelligence optimization algorithm to optimize the network hyperparameters, and selects the gray wolf algorithm to optimize the hyperparameters. The hyperparameter optimization objects are learning rate, learning rate decay factor, gradient threshold and L2Regularization; 9. The method for predicting parameters of ultra-large diameter slurry shield tunneling based on multi-feature fusion according to claim 8, characterized in that: The training process of the hybrid deep neural network model is divided into three steps: data input, combined model construction, and hyperparameter optimization. The gray wolf algorithm is used to obtain the optimal parameter combination through initialization, fitness calculation, position update and other steps to train the hybrid deep neural network.

10. The method for predicting parameters of ultra-large diameter slurry shield tunneling based on multi-feature fusion according to claim 1, characterized in that: Step S6 prepares the characteristic parameters of the section to be excavated for input according to the input features of the model design, uses the trained deep learning network to predict the rock machine load in the excavation state, and guides the design of shield equipment or the control of construction parameters.

Images