Shield load prediction method based on deep learning

Abstract

The invention relates to a shield load prediction method based on deep learning. The method comprises the following steps: firstly, carrying out wavelet noise reduction on shield construction historical load, geology and tunnel structure data; secondly, recombining data, inputting the data into a CNN model fused with a multi-attention mechanism, performing pooling compression in a compression layer, calculating attention weight, updating an input matrix in a reconstruction layer, extracting a spatial feature vector through convolution operation, inputting the vector into a TCN model fused with the attention mechanism, extracting and optimizing time features through causal convolution and expansion convolution, and obtaining a TCN model fused with the attention mechanism; optimized time dependence characteristics are obtained, and shield thrust and torque predicted values are calculated based on the characteristics. According to the method, the defects that a traditional method is poor in universality, low in efficiency, difficult to denoise and incapable of synchronously capturing the spatial-temporal characteristics are overcome, the load prediction precision under the complex geological condition is improved, and support is provided for construction optimization and safety.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tunnel engineering shield construction, in particular to a shield load prediction method based on deep learning. BACKGROUND

[0002] In modern urban underground space development, shield tunnel construction has become the mainstream technology for subway and mountain tunnel engineering construction due to its advantages of high efficiency, safety and small environmental impact. However, in the process of shield construction, accurate prediction of shield load is crucial to ensure construction safety, optimize construction parameters and improve engineering quality. Traditional shield load prediction methods are mostly based on empirical formulas, numerical simulation and simplified mechanical models. Empirical formulas rely on experience data accumulated from specific engineering cases, have poor universality and are difficult to adapt to complex and variable geological conditions; numerical simulation methods can simulate and analyze the shield construction process to a certain extent, but the modeling process is complex, the calculation cost is high, and the selection of model parameters is often subjective; simplified mechanical models are difficult to accurately reflect the complex mechanical behavior of the interaction between soil and shield machine in the process of shield construction due to a large number of simplifications of actual working conditions.

[0003] Shield construction faces complex and diverse geological conditions, such as soft soil stratum, sand and gravel stratum, hard rock stratum, etc. The physical and mechanical properties of different strata differ significantly, and the shield tunneling parameters (such as cutterhead speed, advance speed, grouting pressure, etc.) and stratum parameters are coupled in the construction process, resulting in highly nonlinear, time-varying and uncertain characteristics of shield load. The prediction accuracy and reliability of traditional prediction methods are difficult to meet the actual engineering requirements when dealing with these complex factors. SUMMARY

[0004] Therefore, the purpose of the present application is to overcome the defects of traditional prediction methods in the prior art, such as poor universality, low calculation efficiency, difficulty in handling noise data, and inability to simultaneously capture the spatial correlation, time dependence and complex nonlinear coupling relationship of load parameters.

[0005] To solve the above technical problems, the present application provides a shield load prediction method based on deep learning, which comprises the following steps: S1: performing noise reduction processing on the original data obtained by collecting historical load data, geological parameters and tunnel structure in shield construction, to obtain a noise-reduced data set; S2: reorganizing the data set into an input matrix and inputting it into a CNN model fused with multiple attention mechanisms for spatial feature extraction, and converting the spatial features to obtain a spatial feature vector; S3: inputting the spatial feature vector into a TCN model fused with attention mechanisms for time feature optimization, to obtain optimized time-dependent features; S4: Based on the optimized time-dependent characteristics, the predicted values ​​of shield thrust and torque are obtained.

[0006] In one embodiment of the present invention, step S2, the method of reorganizing the dataset into an input matrix and then inputting it into a CNN model with a multi-attention mechanism for spatial feature extraction, includes the following steps: S21: In the compression layer of the CNN model, for the input matrix U, pooling compression is performed along the three dimensions of channel, width and height to obtain the channel identity matrix, width identity matrix and height identity matrix respectively; S22: Perform the attention mechanism in parallel on the channel identity matrix, width identity matrix and height identity matrix to obtain the channel domain weight matrix, width domain weight matrix and height domain weight matrix respectively; S23: In the feature reconstruction layer of the CNN model, the input matrix U is updated by weighting the channel domain weight matrix, the width domain weight matrix and the height domain weight matrix to obtain the updated input matrix X; S24: In the standard convolution module of the CNN model, convolution and pooling operations are performed on the updated input matrix X to complete spatial feature extraction.

[0007] In one embodiment of the present invention, the formula for calculating the channel domain weight matrix is ​​as follows: ; The formulas for calculating the width domain weight matrix and the height domain weight matrix are as follows: ; ; Where U is the input matrix, For average pooling function, For max pooling function, For fully connected layer functions, For activation function, This is the channel domain weight matrix. This is the width-domain weight matrix. This is the height domain weight matrix.

[0008] In one embodiment of the present invention, the method for obtaining the updated input matrix X is as follows: the input matrix is ​​updated by weighting it using the channel domain weight matrix, the width domain weight matrix, and the height domain weight matrix, and the calculation formula is: .

[0009] In one embodiment of the present invention, the formula for the convolution operation is: , in, For the first The first layer of convolutional network Each feature output, For activation function, For the first Input matrices, For the first The first layer of convolutional network Each feature output, For position kernel parameters, For the first The first layer of convolutional network The bias of the output of each feature; The formula for the pooling operation is: , in, For the first In layered pooling networks Each feature output, For activation function, For the first In layered pooling networks The weight parameters of each feature output. For downsampling function, For the first The first layer pooling network Each feature output, For the first In layered pooling networks The bias of each feature output.

[0010] In one embodiment of the present invention, in step S3, the method for obtaining the optimized time-dependent features is as follows: the spatial feature vector is input into the TCN model as a time series, and the time-dependent basic features are extracted from the spatial feature vector based on causal convolution; the temporal receptive field of the time-dependent basic features is expanded by dilated convolution, and dynamic weighted optimization is performed by using an attention mechanism to obtain the optimized time-dependent features.

[0011] In one embodiment of the present invention, the TCN model consists of stacked multi-layer causal convolutions and dilated convolutions, and its output calculation formula is as follows: , in, For the first The layer's output at time step t, Here, k is the activation function, and k is the kernel size. For the first The weight parameters of the convolutional kernel at position i, where d is the dilation factor. For the first The output of the layer, For the first The bias of the layer output.

[0012] In one embodiment of the present invention, an attention mechanism is used for dynamic weighted optimization, and the calculation formula for the optimization time-dependent feature S is as follows: , in, This represents the output of the TCN model at time step j. Let p be the attention weight at time step j, and p be the total number of time steps.

[0013] In one embodiment of the present invention, in step S1, the method for performing noise reduction is as follows: the original data is accurately represented by a scaling function, supplemented by a wavelet function, the mathematical expression of which is: , in, For scale parameters, The translation amount, For a moment, The input is the raw data signal. This is the mother wave function.

[0014] In one embodiment of the present invention, in step S4, the formulas for calculating the predicted values ​​of the shield thrust F and torque T are as follows: , , Where S represents the optimized time-dependent feature, and This is the weight matrix of the fully connected layer. and This is the bias for the fully connected layer.

[0015] The technical solution of the present invention has the following advantages compared with the prior art: This invention efficiently extracts the spatial correlation features of load parameters using a CNN model that integrates multiple attention mechanisms. It strengthens key sensor channels and spatially sensitive areas through attention mechanisms, significantly improving the ability to capture the spatial distribution patterns of shield loads. Simultaneously, it combines a TCN model with an integrated attention mechanism to extract long-term and short-term time-dependent features through causal convolution and dilatational convolution, and uses dynamic weighting to optimize key tunneling states in the time series, effectively overcoming the shortcomings of traditional methods that cannot simultaneously capture spatial and temporal features. Combined with wavelet denoising preprocessing to eliminate noise interference and outliers in the monitoring data, the model can still accurately characterize the highly nonlinear coupling relationship between shield loads and multi-source parameters under complex and variable geological conditions, ultimately achieving a significant improvement in the accuracy of shield thrust and torque prediction. Attached Figure Description

[0016] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0017] Figure 1 This is a flowchart illustrating a shield tunnel load prediction method based on deep learning provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the process of combining the CNN model and the attention mechanism in an embodiment of the present invention; Figure 3 This is a schematic diagram of CNN model calculation in an embodiment of the present invention; Figure 4 This is a schematic diagram of the TCN model structure in an embodiment of the present invention; Figure 5 This is a schematic diagram of the computational process combining the TCN model and the attention mechanism in an embodiment of the present invention; Figure 6 This is a schematic diagram of the MACT model structure in an embodiment of the present invention; Figure 7 This is the engineering plan view of Experiment 1 in Embodiment 2 of the present invention; Figure 8 This is an engineering cross-sectional view from Experiment 1 in Embodiment 2 of the present invention; Figure 9 This is a schematic diagram of the shield load prediction process in Experiment 1 of Embodiment 2 of the present invention; Figure 10 This is a schematic diagram of the dataset partitioning in Experiment 1 of Embodiment 2 of the present invention; Figure 11 These are thrust prediction curves for different models in Experiment 1 of Embodiment 2 of the present invention; Figure 12 This is a graph showing the torque prediction curves of different models in Experiment 1 of Embodiment 2 of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0019] Reference Figure 1 As shown, this invention provides a shield tunneling load prediction method based on deep learning, which includes the following steps: S1: Denoise the raw data obtained from collecting historical load data, geological parameters and tunnel structure data during shield tunneling construction to obtain a denoised dataset. S2: After reorganizing the dataset into an input matrix, input it into a CNN model that integrates a multi-attention mechanism to extract spatial features, and then transform the spatial features to obtain a spatial feature vector; S3: Input the spatial feature vector into the TCN model with the attention mechanism to optimize the temporal features and obtain optimized temporal dependency features; S4: Based on the optimized time-dependent characteristics, the predicted values ​​of shield thrust and torque are obtained.

[0020] This invention accurately extracts the spatial distribution features of shield tunneling loads by integrating a CNN (Convolutional Neural Networks) model with a multi-attention mechanism, and captures the time-dependent laws of load dynamic evolution by combining an attention mechanism-optimized TCN (Temporal Convolutional Networks) model. While effectively filtering out noise interference from the original data, it significantly improves the prediction accuracy of shield thrust and torque under complex geological conditions, providing reliable data support for dynamic adjustment of construction parameters and early warning of safety risks.

[0021] In step S1, multi-source heterogeneous data from the entire shield tunneling construction process are systematically collected, specifically including: historical operating parameters, equipment status parameters, environmental geological parameters, and tunnel structural parameters. After all data is aligned by timestamp, it is uniformly organized into a structured Excel dataset, with fields including time series ID, parameter type, numerical unit, and collection location, forming a standardized model training and validation dataset.

[0022] To address the noise interference problem in the original data, this invention employs wavelet transform for professional noise reduction. Considering the adverse factors such as construction vibration noise, sensor drift, and electromagnetic interference in the original data, this invention uses wavelet transform for noise reduction.

[0023] Specifically, the noise reduction method is as follows: the original data is accurately represented by a scaling function, supplemented by a wavelet function, the mathematical expression of which is: , in, For scale parameters, The translation amount, For a moment, The input is the raw data signal. This is the mother wave function.

[0024] The scaling function accurately represents the signal, capturing the low-frequency trend components of the original signal; the wavelet function compensates for the representation bias: to prevent the accumulation of deviations between the approximate value of the scaling function and the original signal at a large scale, the wavelet function is introduced to specifically represent the difference between the two.

[0025] Further, in step S2, the method of reorganizing the dataset into an input matrix and then inputting it into a CNN model with a multi-attention mechanism for spatial feature extraction includes the following steps: S21: In the compression layer of the CNN model, for the input matrix U, pooling compression is performed along the three dimensions of channel, width and height to obtain the channel identity matrix, width identity matrix and height identity matrix respectively; S22: Perform the attention mechanism in parallel on the channel identity matrix, width identity matrix and height identity matrix to obtain the channel domain weight matrix, width domain weight matrix and height domain weight matrix respectively; S23: In the feature reconstruction layer of the CNN model, the input matrix U is updated by weighting the channel domain weight matrix, the width domain weight matrix and the height domain weight matrix to obtain the updated input matrix X; S24: In the standard convolution module of the CNN model, convolution and pooling operations are performed on the updated input matrix X to complete spatial feature extraction.

[0026] Step S21 is a three-dimensional pooling compression, used to reduce the data dimensionality and extract statistical features of each dimension, providing a basis for attention calculation; Step S22 executes the attention mechanism in parallel on the three identity matrices to learn the importance weights of the three-dimensional features, solving the problem of insufficient attention to key information in traditional CNN models; Step S23 is used to perform feature weighted updates, enhance key features, suppress noise interference, and improve feature representation ability; Step S24 outputs the spatial feature vector, completing spatial feature extraction.

[0027] Reference Figure 2As shown, shield tunneling load data exhibits significant spatial distribution characteristics, and CNN models can effectively extract its spatial features through convolution and pooling operations. Meanwhile, shield tunneling loads are influenced by various parameters, including shield tunneling parameters, geological parameters, and tunnel structural parameters, and each parameter has a different degree of influence on the load. Considering the differences in the degree of influence, an attention mechanism is incorporated into the CNN model. The attention mechanism can focus on a small amount of important information from the raw data, which helps improve the model's prediction accuracy, enabling it not only to extract various features from the input but also to emphasize the importance of these features.

[0028] The original dataset is reorganized into an input matrix U, which can be obtained using the following formula: , in, For convolution operation, T is the original dataset matrix.

[0029] Furthermore, three compressed unit-length matrices are obtained through max pooling and min pooling. , and For the channel domain attention mechanism, each matrix is ​​passed through two fully connected layer structures ( After that, we obtain the new matrix with updated weights. The calculation formula is: ; in, For average pooling function, For max pooling function, For fully connected layer functions, For activation function, This is the channel domain weight matrix.

[0030] The input matrix U is updated using a new matrix in the channel domain direction: .

[0031] For the spatial attention mechanism, each identity matrix also goes through two fully connected layers ( We obtain two new matrices with updated weights, calculated using the following formula: ; ; in, This is the width-domain weight matrix. This is the height domain weight matrix.

[0032] The two new matrices in the spatial domain are merged, and then the input matrix U is updated: .

[0033] The matrices obtained from updating the spatial and channel domains separately are merged. The input matrix is ​​then updated using the channel domain weight matrix, width domain weight matrix, and height domain weight matrix to obtain the updated input matrix X, thus constructing a new input tensor. The calculation formula is as follows: .

[0034] Reference Figure 3 As shown, the updated input matrix X serves as the input to the CNN model. Convolution and pooling primarily function to effectively extract and compress useful features from the data. The formula for convolution is: , in, For the first The first layer of convolutional network Each feature output, For activation function, For the first Input matrices, For the first The first layer of convolutional network Each feature output, For position kernel parameters, For the first The first layer of convolutional network The bias of the output of each feature; The formula for calculating the pooling layer is: , in, For the first In layered pooling networks Each feature output, For activation function, For the first In layered pooling networks The weight parameters of each feature output. For downsampling function, For the first The first layer pooling network Each feature output, For the first In layered pooling networks The bias of each feature output.

[0035] Finally, a fully connected layer structure is used to transform the graph-structured data into vector form, which is then used as input to the TCN model. Where W is the weight and b is the bias.

[0036] Furthermore, shield load prediction is closely related to shield tunneling time, making it a traditional time series problem. Existing models mainly employ RNN (Recurrent Neural Network)-based methods, which can accurately capture the temporal characteristics of the data. However, since shield tunnel construction is a lengthy process involving a wide range of parameters and generating a large amount of data, it places high demands on parallel computing, stable gradients, and low memory usage. Traditional RNN-based models struggle to meet these requirements.

[0037] Reference Figure 4 As shown, optionally, the present invention employs a TCN model to extract the temporal features of the load data. The TCN is constructed using two interconnected residual blocks to create a residual structure. Each residual block contains a one-dimensional dilated causal convolution, weight normalization, an activation function, and a dropout layer.

[0038] Causal convolution ensures that the current output of a convolution operation depends only on the current and past inputs, thus preserving temporal order and preventing future information leakage. Causal convolution employs a unidirectional structure, ensuring that the current output is based on the current and previous inputs. Unlike CNN models, TCN simultaneously introduces dilated convolutions to expand the receptive field, improving computational efficiency. The range of input intervals sampled during convolution is determined by the size of the dilation factor in the dilated convolution.

[0039] Specifically, in step S3, the method for obtaining the optimized time-dependent features is as follows: the spatial feature vector is input into the TCN model as a time series, and the time-dependent basic features are extracted from the spatial feature vector based on causal convolution; the temporal receptive field of the time-dependent basic features is expanded by dilated convolution, and dynamic weighted optimization is performed using an attention mechanism to obtain the optimized time-dependent features.

[0040] The dilated causal convolution of a TCN can be obtained using the following formula: , in, The weights are the inputs before time step t; This is a predicted value.

[0041] The integration of causal convolution and dilated convolution allows convolutional layers to expand their receptive field and capture more information while adhering to strict time constraints. TCN models typically consist of multiple stacked layers of causal and dilated convolutions, and their output is calculated using the following formula: , in, For the first The layer's output at time step t, Here, k is the activation function, and k is the kernel size. For the first The weight parameters of the convolutional kernel at position i, where d is the dilation factor. For the first The output of the layer, For the first The bias of the layer output.

[0042] Reference Figure 5 As shown, the TCN network ignores the importance of different time features during training, resulting in insufficient learning of information at some important time points. To solve this problem and improve the accuracy of the model, an attention mechanism can be optionally implemented after the TCN model to determine the salience of each time point.

[0043] The output hidden state of TCN is Attention score It can be represented as: , Attention weights The calculation formula is: , The formula for calculating the optimized time-dependent feature S is as follows: , in, This represents the output of the TCN model at time step j. Let p be the attention weight at time step j, and p be the total number of time steps.

[0044] The output of the TCN model, which is the final output of the MCT model, can be obtained using the following formula: , in, Let be the attention weights at the i-th time step. This is the similarity calculation function.

[0045] The TCN model output is: , in, This is a fully connected layer function.

[0046] The formulas for calculating the predicted values ​​of shield thrust F and torque T are as follows: , , Where S represents the optimized time-dependent feature, and This is the weight matrix of the fully connected layer. and This is the bias for the fully connected layer.

[0047] Furthermore, the calculation results of the prediction model proposed in this invention are compared and analyzed with various benchmark models, and RMSE (root mean square error) and MAE (mean absolute error) are selected. The model's predictive performance is validated using the coefficient of determination (COD) and accuracy (ACC), and the mathematical formulas are as follows: ; ; ; ; in, This is a predicted value; Actual monitored value

[0048] Example 2: This invention also provides a shield tunneling load prediction system based on deep learning, such as... Figure 6 As shown, the system includes a MACT (Multi-attention-convolutional neural network-temporal convolutional network) model, which includes a CNN model and a TCN model that fuse attention mechanisms, and is used to implement the deep learning-based shield load prediction method described in Embodiment 1.

[0049] The above system includes the following modules: The data preprocessing and noise reduction module is used to perform noise reduction processing on the raw data obtained from historical load data, geological parameters and tunnel structure during shield tunneling construction to obtain a noise-reduced dataset. The multi-dimensional spatial feature extraction module is used to reorganize the dataset into an input matrix and then input it into a CNN model that integrates a multi-attention mechanism to extract spatial features, and to obtain a spatial feature vector from the spatial features. The temporal feature optimization module is used to input the spatial feature vector into the TCN model with fused attention mechanism to optimize the temporal features and obtain optimized temporal dependency features. The load prediction output module is used to obtain the predicted values ​​of shield thrust and torque based on the optimized time dependence characteristics.

[0050] Experiment 1: To verify the performance of the above MACT model, a comparative experiment was conducted, referring to... Figures 7-8As shown in the figure, in this experiment, the newly constructed shield tunnel needs to pass under the existing station structure. During the underpass process, the shield tunnel needs to pass through diaphragm walls and jet grouting piles, with a minimum clearance of 1.52 m between the tunnel top and the existing station floor. The natural soil is generally in a state of medium hardness or medium density. The shield machine is expected to pass through silty clay and silty clay. Different geological environments will affect the shield tunneling process. For example, in hard or gravelly strata, the shield machine may encounter wear and damage to the cutting tools; in loose sandy soil, the shield machine may have difficulty forming stable working face pressure before excavation, thus increasing the risk of ground settlement and collapse. Therefore, accurate prediction of shield load is crucial for timely adjustment of tunneling parameters and ensuring the construction safety of the project.

[0051] The shield load prediction process for this experiment is as follows: Figure 9 As shown, it is divided into three stages: Phase one is data collection and preprocessing. First, geological parameters, shield machine operating parameters, structural parameters and tunnel geometric parameters are integrated to construct a dataset. Data quality is then optimized through feature selection and data preprocessing (including cleaning, normalization, etc.). Phase two involves building a predictive model, referring to... Figure 10 As shown, the processed dataset is divided into training, validation, and test sets. Based on the training and validation subsets, various machine learning models (including RNN, CNN, LSTM, CNN-LSTM, MD-DNN, KF-CNN-BiGRU, and the proposed MACT model) are trained and their hyperparameters optimized to select the model with the best prediction performance. Phase three is model performance evaluation, which comprehensively assesses the model's predictive performance on the test set. Key metrics include MAE (mean absolute error) and RMSE (root mean square error). (Determination coefficient), while verifying computational efficiency and generalization ability, finally outputting reliable shield tunneling load prediction results. The load (thrust and torque) prediction curves of different models are shown in the figure. Figures 11-12 As shown.

[0052] To ensure the reasonableness of the results, the optimal hyperparameters of other models were first determined through experiments, and then the prediction results of each model were compared with the MACT model.

[0053] The optimal hyperparameters for different models are shown in Table 1.

[0054] Table 1:

[0055] The prediction results of different models are shown in Table 2. The MAE, RMSE, and other parameters of the MCT model are also shown. The accuracy and RMSE were 17.21, 19.03, 0.98, 97.86% and 18.34, 19.75, 0.97, 97.12%, respectively. The MACT model had the lowest MAE and RMSE among these models, while... The MCT model exhibits the highest accuracy and predictive performance compared to other baseline models. Furthermore, the RNN algorithm has the highest MAE and RMSE among these models. The lowest accuracy indicates that the RNN algorithm has the worst prediction performance, because the RNN algorithm is prone to gradient vanishing and exploding.

[0056] As shown in Table 2, the trend of the predicted values ​​obtained by the MCT algorithm is more consistent with the measured values.

[0057] Table 2:

[0058] Experiment 2: To analyze the computational efficiency of the MCT model, experiments were conducted using three datasets with different amounts of data (1000, 3000, and 5000).

[0059] Table 3 shows the computation time for each model. When the amount of data remains constant, the CNN-LSTM model has the longest computation time. This is because it is a combined model with a cascaded structure, so its computation time must consider both the CNN and LSTM sub-models simultaneously. RNN models (such as LSTM models) process data sequentially, and the computation at each time step depends on the result of the previous time step, which increases the model's computation time. In contrast, the MACT model uses a TCN model, enabling parallel computation, and its computation time is significantly reduced compared to the CNN-LSTM, KF-CNN-BiGRU, and MD-DNN models.

[0060] Table 3:

[0061] Experiment 3: To further verify the generalization ability of the prediction model provided by this invention for predicting shield tunneling loads under different geological conditions, shield tunneling load data from a construction section of Shenyang Metro Line 3 were selected as the research object. Geological parameter data were obtained from the geotechnical engineering investigation report, and shield tunneling operation data were obtained from the shield machine real-time monitoring platform. The experimental results are shown in Table 4. For thrust data, the MAE, RMSE, and [other parameters] obtained by the MCT model were [data missing]. The accuracy values ​​were 17.94, 20.32, 0.98, and 95.07%, respectively. Compared with the other six comparative models, the MCT model achieved the lowest MAE and RMSE values, as well as the highest accuracy. The accuracy and performance of the test results demonstrate its superior performance in predicting tunnel boring machine thrust. Similar conclusions can be drawn from the torque test results.

[0062] Table 4:

[0063] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0064] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0065] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0066] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0067] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for predicting shield tunnel loads based on deep learning, characterized in that, Includes the following steps: S1: Denoise the raw data obtained from collecting historical load data, geological parameters and tunnel structure data during shield tunneling construction to obtain a denoised dataset. S2: After reorganizing the dataset into an input matrix, input it into a CNN model that integrates a multi-attention mechanism to extract spatial features, and then transform the spatial features to obtain a spatial feature vector; S3: Input the spatial feature vector into the TCN model with the attention mechanism to optimize the temporal features and obtain optimized temporal dependency features; S4: Based on the optimized time-dependent characteristics, the predicted values ​​of shield thrust and torque are obtained.

2. The shield tunneling load prediction method based on deep learning according to claim 1, characterized in that: In step S2, the method of reorganizing the dataset into an input matrix and then inputting it into a CNN model with a multi-attention mechanism for spatial feature extraction includes the following steps: S21: In the compression layer of the CNN model, for the input matrix U, pooling compression is performed along the three dimensions of channel, width and height to obtain the channel identity matrix, width identity matrix and height identity matrix respectively; S22: Perform the attention mechanism in parallel on the channel identity matrix, width identity matrix and height identity matrix to obtain the channel domain weight matrix, width domain weight matrix and height domain weight matrix respectively; S23: In the feature reconstruction layer of the CNN model, the input matrix U is updated by weighting the channel domain weight matrix, the width domain weight matrix and the height domain weight matrix to obtain the updated input matrix X; S24: In the standard convolution module of the CNN model, convolution and pooling operations are performed on the updated input matrix X to complete spatial feature extraction.

3. The shield tunneling load prediction method based on deep learning according to claim 2, characterized in that: The formula for calculating the channel domain weight matrix is ​​as follows: ； The formulas for calculating the width domain weight matrix and the height domain weight matrix are as follows: ； ； Where U is the input matrix, For average pooling function, For max pooling function, For fully connected layer functions, For activation function, This is the channel domain weight matrix. This is the width-domain weight matrix. This is the height domain weight matrix.

4. The shield tunneling load prediction method based on deep learning according to claim 3, characterized in that: The method for obtaining the updated input matrix X is as follows: the input matrix is ​​updated by weighting it using the channel domain weight matrix, the width domain weight matrix, and the height domain weight matrix. The calculation formula is: 。 5. The shield tunneling load prediction method based on deep learning according to claim 2, characterized in that: The formula for the convolution operation is: ， in, For the first The first layer of convolutional network Each feature output, For activation function, For the first Input matrices, For the first The first layer of convolutional network Each feature output, For position kernel parameters, For the first The first layer of convolutional network The bias of the output of each feature; The formula for the pooling operation is: ， in, For the first In layered pooling networks Each feature output, For activation function, For the first In layered pooling networks The weight parameters of each feature output. For downsampling function, For the first The first layer pooling network Each feature output, For the first In layered pooling networks The bias of each feature output.

6. The shield tunneling load prediction method based on deep learning according to claim 1, characterized in that: In step S3, the method for obtaining the optimized time-dependent features is as follows: the spatial feature vector is input into the TCN model as a time series, and the basic time-dependent features are extracted from the spatial feature vector based on causal convolution; the temporal receptive field of the basic time-dependent features is expanded by dilated convolution, and dynamic weighted optimization is performed using an attention mechanism to obtain the optimized time-dependent features.

7. The shield tunneling load prediction method based on deep learning according to claim 1 or 6, characterized in that: The TCN model consists of stacked multi-layer causal convolutions and dilated convolutions, and its output is calculated using the following formula: ， in, For the first The layer's output at time step t, Here, k is the activation function, and k is the kernel size. For the first The weight parameters of the convolutional kernel at position i, where d is the dilation factor. For the first The output of the layer, For the first The bias of the layer output.

8. The shield tunneling load prediction method based on deep learning according to claim 1, characterized in that: Using an attention mechanism for dynamic weighted optimization, the formula for calculating the optimization time-dependent feature S is as follows: ， in, This represents the output of the TCN model at time step j. Let p be the attention weight at time step j, and p be the total number of time steps.

9. The shield tunneling load prediction method based on deep learning according to claim 1, characterized in that: In step S1, the noise reduction method is as follows: the original data is accurately represented by a scaling function, supplemented by a wavelet function, the mathematical expression of which is: ， in, For scale parameters, The translation amount, For a moment, The input is the raw data signal. This is the mother wave function.

10. The shield tunneling load prediction method based on deep learning according to claim 1, characterized in that: In step S4, the formulas for calculating the predicted values ​​of the shield thrust F and torque T are as follows: ， ， Where S represents the optimized time-dependent feature, and This is the weight matrix of the fully connected layer. and This is the bias for the fully connected layer.

Images