An intelligent prediction method for shield tunneling speed based on a hybrid neural network

Abstract

An intelligent prediction method for shield tunneling speed based on hybrid neural network. Firstly, the important parameters of the shield advancement process (including cutterhead speed RS, cutterhead torque TOR, total thrust TH, bearing capacity characteristic value F) are obtained through preliminary geological survey and continuous monitoring. a , compression modulus E S , tunneling speed AR), remove the empty values and outliers from the obtained parameters, obtain reliable input big data, and perform normalization. Secondly, build an Attention‑ResNet‑LSTM hybrid neural network model, and establish the optimal model through the data in the training set and the validation set. Use the trained optimal model to predict the shield tunneling speed on the test set, and repeat the above steps until the prediction accuracy meets the actual application requirements. This intelligent prediction method can predict the shield tunneling speed in real time, assist relevant operators in matching the optimal shield construction parameters, and is conducive to the safety and efficiency of shield construction.

Description

Technical Field

[0001] The present invention relates to an intelligent prediction method for shield tunneling speed based on an Attention-ResNet-LSTM hybrid neural network, aiming to use shield construction parameters and formation parameters to predict the speed during shield tunneling in real time, assist in matching the optimal shield construction parameters, and make targeted adjustments to the tunneling strategy. Background Art

[0002] In recent years, with the large-scale development and utilization of the economic circles along the rivers and coasts of China, the development of tunnel engineering has been growing rapidly. Among various construction methods, the shield method has been widely used due to its many advantages such as high mechanization degree, high speed, and small environmental disturbance. However, due to the complex interaction between the shield machine and the formation, the behavioral performance of the shield machine will continuously fluctuate during the tunneling process. To ensure the safe and efficient shield construction, it is particularly important to predict the working performance of the shield machine. Among them, the tunneling speed reflects the interaction state between the shield and the soil, and is closely related to the construction period and the total cost. It is one of the most concerned indicators in tunnel construction.

[0003] Most traditional methods predict the performance of earth pressure or slurry shields through some empirical and theoretical models. These methods are often only applicable to specific formations. While costing relatively high, it is difficult to simulate the complex interaction between the shield and the formation, and there is a large difference from the actual working conditions. Therefore, it is difficult to accurately predict key parameters such as the shield tunneling speed and to feedback the rationality of the matching of construction parameters. Summary of the Invention

[0004] Based on the background art, the present invention proposes a new intelligent prediction method for shield tunneling speed based on an Attention-ResNet-LSTM hybrid neural network. The feature vectors input into the neural network are selected. At the same time, by introducing the ResNet structure on the basis of the LSTM network structure, the feature extraction ability of the data network during the shield tunneling process is enhanced. At the same time, the Attention mechanism is introduced to enable the neural network to adaptively update the data weight matrix. The existing positive sample strategies accumulated in health and safety engineering are used to ensure safety, and more extensive engineering samples are accumulated to make the neural network have strong robustness and generalization ability. It is a more intelligent and accurate prediction method.

[0005] The object of the present invention is to use the spatio-temporal construction parameters and formation parameters collected during the tunneling process of a shield machine as inputs, input these data into a trained Attention-ResNet-LSTM hybrid neural network model, obtain the real-time predicted value of the shield tunneling speed at a future moment, and provide reference and guarantee for the safety in actual shield construction. The shield construction data source used in the present invention comes from shield projects with safe, reliable and normal construction, and it is required that no obvious safety accidents have occurred during the shield construction process, ensuring that the model trained with the big data source can reflect the "machine-soil" interaction relationship under the safe tunneling of the shield machine. With the improvement of the big data of the shield safety tunneling database, the adaptability of the model to various formations will be better and better, and the robustness will also be continuously enhanced, which will provide higher guarantee for the safe application of the prediction model.

[0006] In order to achieve the object of the invention, the following technical solutions are adopted:

[0007] An intelligent prediction method for shield tunneling speed based on an Attention-ResNet-LSTM hybrid neural network, characterized in that it specifically includes the following steps:

[0008] S1. Data collection:

[0009] During the shield tunneling process, the shield operation parameters are recorded in real time at a fixed sampling frequency f, and at the same time, the soil quality parameters during the shield tunneling process are recorded;

[0010] S2. Data preprocessing:

[0011] 1) Empty data detection

[0012] 2) Outlier detection

[0013] The Mahalanobis Distance is used as the criterion for outlier discrimination; for a multivariate input sequence x = (x1, x2, x3, x4, x5, x6) with a mean of μ = (μ1, μ2, μ3, μ4, μ5, μ6) T , and a covariance matrix of S T , the calculation of its Mahalanobis distance is shown in Formula 2:

[0014] Formula 2:

[0015] In the above formula, x refers to the input variables cutter head rotation speed RS, cutter head torque TOR, total thrust TH, characteristic value of bearing capacity F a , compression modulus E S and tunneling speed AR;

[0016] 3) Data normalization

[0017] Normalize the data and map it to the interval [0, 1];

[0018] S4. Dataset splitting:

[0019] Divide the preprocessed data into a training set, a validation set, and a test set;

[0020] S5. Construct an Attention-ResNet-LSTM hybrid neural network model:

[0021] The normalized input data is fed into the ResNet structure for feature extraction, and different weights are assigned to the feature maps of each channel after extraction by the ResNet structure through the channel attention mechanism. The weighted feature maps are input into the LSTM network, and a temporal attention mechanism is added to the last layer of the LSTM structure to assign weight values to the hidden layer outputs at different times. The temporal vector after weighted summation is concatenated with the output of the last time step as the final output after the attention mechanism processing;

[0022] S6. Model training and performance evaluation:

[0023] Based on the above S5 neural network model, the sequence data of the past p historical moments of each selected feature is used as the model input, and the output is the tunneling speed at time t. The relationship between the input and output parameters is as follows:

[0024] Formula Six: AR| t = f((RS, TH, TOR, Es, Fa, AR)| t-1,t-2,…,t-p )

[0025] Among them, the function f is the mapping relationship to be fitted by the deep learning model, p is the length of the historical time series data, and the optimal value can be determined by numerical experiments;

[0026] Evaluate the model prediction performance. Select the mean absolute percentage error MAPE and the root mean square error RMSE as evaluation indicators, and the combination of the two evaluates the performance of the model. The calculation method is as follows:

[0027] Formula Eight:

[0028] Formula Nine:

[0029] The smaller the values of the two indicators MAPE and RMSE, the smaller the deviation between the predicted value and the true value, and the higher the model fitting accuracy;

[0030] Use the trained optimal model to predict the tunneling speed on the test set. If the prediction accuracy meets the requirements, the model is applied to the actual project to predict the shield tunneling speed in real time; if the prediction accuracy of the model does not meet the requirements, repeat the above steps until the prediction accuracy meets the actual application needs.

[0031] This application is an intelligent prediction method for shield tunneling speed based on the Attention-ResNet-LSTM hybrid neural network, which is mainly applicable to the shield construction scenario and is targeted at shield machine operators and construction management personnel. By collecting, extracting, and processing the construction parameters during the tunneling process, training the Attention-ResNet-LSTM hybrid neural network model, and predicting the shield tunneling speed based on this, it helps relevant operators to match and select shield construction parameters, which is beneficial to the safe and efficient shield construction.

[0032] Beneficial effects:

[0033] 1. It can solve the problem that it is difficult to accurately and efficiently predict the tunneling speed due to the complex interaction relationship between the shield and the formation during the shield tunneling process.

[0034] 2. It fully considers the important factors that may affect the tunneling speed during the shield tunneling process and gives a concise and clear model training and prediction method, which helps relevant operators and management personnel to match and select shield construction parameters, ensuring that the shield tunnel can be excavated safely and quickly. Description of the drawings

[0035] Figure 1 is a common convolutional neural network architecture

[0036] Figure 2 is a schematic diagram of the basic unit of the residual neural network (ResNet)

[0037] Figure 3 is the basic structure of the LSTM unit

[0038] Figure 4 is the fitting result of the EPR algorithm during the exploration process

[0039] Figure 5 is the structure diagram of the Attention-ResNet-LSTM hybrid neural network model

[0040] Figure 6 Schematic diagram of the intelligent control system for shield tunneling speed

[0041] Figure 7 is the prediction effect of the Attention-ResNet-LSTM hybrid neural network model in Example 1

[0042] Figure 8 Relationship curves between AR and RS, AR and TH, and AR and TOR in Example 2 Detailed implementation mode

[0043] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and gives detailed implementation manners and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments.

[0044] General introduction and its application in the network structure design of the present invention:

[0045] 1) Convolutional neural network and residual neural network

[0046] The basic composition structure of a convolutional neural network (CNN) is: convolutional layer, pooling layer, fully connected layer, and output layer. The convolutional layer is the core layer for constructing a CNN. Each neuron in it makes local connections (local perception) with the input data of the previous layer through a convolutional kernel, and extracts different features of the input through continuous convolutional operations. Moreover, the same depth slice shares the same set of weights and biases, greatly reducing the number of parameters in the network. The pooling layer is generally located after the convolutional layer, and its function is to reduce the spatial size of the data body and play a role in secondary feature extraction. The commonly used methods are max pooling and mean pooling. The fully connected layer in a CNN is similar to a traditional neural network. Each neuron makes full connections with all neurons in the previous layer, so as to integrate the local information in the convolutional layer or pooling layer, and the final output value is passed to an output layer. A common convolutional neural network architecture is as Figure 1 shown.

[0047] As the network depth increases, the feature extraction ability of a CNN continuously enhances, but the training of a deep convolutional neural network often faces the problem of performance degradation: the accuracy of the model on the training set decreases instead as the number of network layers increases. When building the network in the present invention, the idea of residual learning is introduced into the residual neural network (ResNet), and the problem of performance degradation of deep networks is solved by learning the more easily fitted residual mapping. Its basic unit is as Figure 2 shown. The present invention uses the ResNet module to extract features from the input data.

[0048] 2) Long short-term memory neural network

[0049] The long short-term memory neural network (LSTM) is a recurrent neural network with long-term memory ability, and is often used in the field of natural language processing. The gating mechanism in its structure deletes and writes the information stored in the unit, solving the problems of gradient disappearance and gradient explosion that are prone to occur in RNNs in long time series problems. The basic structure of an LSTM unit is asFigure 3 as shown (in the figure, x s in the present invention refers to the input feature at time s). The present invention uses the LSTM module to further extract the time series information contained in the features.

[0050] 3) Attention mechanism

[0051] After traditional neural networks are trained, they will obtain a fixed weight matrix, which remains unchanged even when the network receives completely different inputs. After introducing the attention mechanism, the network will assign corresponding weights to different input features, enabling the neural network to have the opportunity to focus on more important information in the task.

[0052] Parameter combinations and applications adapted to the prediction network structure of the present invention:

[0053] Select input parameters:

[0054] The preprocessed dataset contains a large number of parameters. If all of them are used as features to input into the prediction model, it will make the complexity of the model too high and greatly increase the calculation time. However, if too few input parameters are selected, it is not enough for the model to learn the mutual relationship between features and cannot guarantee the prediction accuracy. Therefore, selecting the right and appropriate input parameters is of great significance for the application of deep learning models.

[0055] The exploration process of the present invention:

[0056] The present invention uses the Evolutionary Polynomial Regression algorithm (EPR) to select the best combination of input parameters for the prediction model. EPR is a hybrid regression method that combines traditional numerical regression and gene coding technology and can be used to describe the correlation between multiple input variables and output variables. The EPR algorithm can be summarized into two steps:

[0057] In the first step, use an optimization algorithm (such as genetic algorithm, particle swarm algorithm, etc.) to search for the symbolic expression of the polynomial. The transformed variables can be expressed as:

[0058] Formula Four: where x i represents the i-th input variable (through the exploration process, finally in the present invention, the input variables refer to the cutter head rotation speed RS, cutter head torque TOR, total thrust TH, bearing capacity characteristic value F a and compression modulus E S ), k is the number of input variables (through the exploration process, finally in the present invention, the value range is k = 1 - 5). ES m×k represents the exponent matrix searched by the genetic algorithm. z j represents the j-th transformed variable, and its number can be determined in advance.

[0059] Step 2: Estimate the regression coefficients of each term in the polynomial through least-squares linear regression. The final EPR expression is as follows:

[0060] Formula 5: where y is the predicted result (through the exploration process, the predicted result in the present invention finally refers to the shield tunneling speed AR), a j is the regression coefficient of the j-th term, and a0 is an optional bias term.

[0061] In the research stage, select the cutterhead rotation speed RS, cutterhead torque TOR, and total thrust TH from the shield machine operation parameters (these parameters can be adjusted by the operator), and select the characteristic value of bearing capacity F a and compression modulus E S from the soil parameters. Take these variables as alternative combinations for predicting the shield tunneling speed AR, and use the EPR algorithm to obtain the correlation between different variable combinations and the prediction target.

[0062] Based on the fitting results of the above EPR algorithm (as Figure 4 shown), the optimal combination parameter set can be obtained, that is, the input combination of cutterhead rotation speed RS, cutterhead torque TOR, total thrust TH, characteristic value of bearing capacity F a and compression modulus E S achieves the best prediction effect. It should be noted that RS, TOR, and TH are all operation parameters representing the shield's own state, and F a and E S are geological parameters representing the soil state. This combination can more comprehensively reflect the "machine-soil" interaction. Therefore, the present invention selects these five parameters and the tunneling speed AR itself as the input parameter combination of the present invention.

[0063] It should be noted that the exploration process itself is not an integral part of the technical solution of the present invention, and the selection of the five parameters established through the exploration process is an essential part and technical contribution of the technical solution of the present invention.

[0064] The technical solution of the present invention is disclosed as follows:

[0065] An intelligent prediction method for shield tunneling speed based on an Attention-ResNet-LSTM hybrid neural network, characterized in that it specifically includes the following steps:

[0066] S1. Data collection:

[0067] During the shield tunneling process, record the shield operation parameters in real time at a fixed sampling frequency f, and record the soil parameters during the shield tunneling process. To ensure construction safety, that is, to ensure the safety of data samples, S2 step must be passed.

[0068] S2. Data preprocessing:

[0069] 4) Empty data detection

[0070] A large number of invalid empty push data are included in the original data. These empty data need to be deleted, and the remaining valuable data are extracted to form a data set. The empty data detection algorithm is as follows:

[0071] Formula 1: f i = TH i ·TOR i ·AR i ·RS i

[0072] Where TH i , TOR i , AR i , RS i respectively represent the thrust force, cutter head torque, tunneling speed and cutter head rotation speed of the shield machine at time i. When f i = 0, it means that the shield machine is in a stopped state, and the data record at this moment is deleted.

[0073] 5) Outlier detection

[0074] The data collected during the shield tunneling process are not all reliable. Affected by external interference or the sensors themselves, some outliers often occur. Here, the Mahalanobis Distance is used as the criterion for outlier discrimination. The Mahalanobis Distance is not affected by the dimension, and compared with the more familiar Euclidean Distance, the Mahalanobis Distance excludes the interference of the correlation between variables, making it suitable for outlier processing of multivariate time series. In this application, for a multivariate input sequence x = (x1, x2, x3, x4, x5, x6) T with a mean of μ = (μ1, μ2, μ3, μ4, μ5, μ6) T and a covariance matrix of S, the calculation of its Mahalanobis Distance is shown in Formula 2:

[0075] Formula 2:

[0076] In the above formula, x refers to the input variables cutter head rotation speed RS, cutter head torque TOR, total thrust force TH, characteristic value of bearing capacity F a , compression modulus E S and tunneling speed AR.

[0077] It can be seen from Formula 2 that when the variables are uncorrelated, the covariance matrix S is an identity matrix, and the Mahalanobis Distance is equal to the Euclidean Distance. Let the 0.9 quantile of D M be the threshold D t , and when D M > Dt As a condition for outlier discrimination, the data points that meet this condition are regarded as outliers and removed from the dataset.

[0078] 6) Data normalization

[0079] To enhance the convergence speed of the model and eliminate the order-of-magnitude differences between input parameters, all the data is normalized and mapped to the interval [0, 1]. For the input parameter x (the input parameters of the present invention are cutter head rotation speed RS, cutter head torque TOR, total thrust TH, characteristic value of bearing capacity F a , compression modulus E S and tunneling speed AR), the normalization method is shown in Formula 3:

[0080] Formula 3: where x min and x max are the minimum and maximum values of the variable x, respectively.

[0081] Note: For the neural network trained with normalized data, the predicted shield tunneling speed value finally output from the neural network should be restored from the normalized [0, 1] interval to the value in the initial sample space.

[0082] S4. Dataset splitting:

[0083] The preprocessed data is divided into a training set, a validation set, and a test set in the ratio of 8:1:1. The training set is used for model training, the validation set is used for optimizing the model hyperparameters, and the test set is used for evaluating the performance of the final model.

[0084] S5. Constructing an Attention-ResNet-LSTM hybrid neural network model:

[0085] The normalized input data is fed into the ResNet structure for feature extraction, and different weights are assigned to the feature maps of each channel extracted by the ResNet structure through the channel attention mechanism. The weighted feature maps are input into the LSTM network, and a temporal attention mechanism is added to the last layer of the LSTM structure to assign weight values to the hidden layer outputs at different times. The temporally weighted sum vector is concatenated with the output of the last time step as the final output after attention mechanism processing. Figure 5 In this, both m and n are model hyperparameters, and the optimal values can be determined through numerical experiments.

[0086] As Figure 5 shown in the Attention-ResNet-LSTM hybrid neural network model, the construction method and implementation strategy of the three major modules of the predicted model of the present invention are given:

[0087] First, input the sample data of p historical moments in the form of a multivariate time series and send it into the neural network. The ResNet module in the network model consists of m basic units, and each basic unit includes two convolutional layers and corresponding residual connections to extract and output the feature map;

[0088] Then, construct an LSTM module including n LSTM cells; and an attention mechanism including a channel attention module and a temporal attention module;

[0089] Among them, the Channel Attention of the channel attention module is composed of a global average pooling layer and two fully connected layers. The feature map extracted by the ResNet module first passes through a global average pooling layer to obtain the global information of each channel, and then passes through two fully connected layers to obtain the weight values of each channel, and inputs the weighted feature map into the LSTM module;

[0090] Among them, in the Temporal Attention of the temporal attention module, the similarity scores between the hidden state at the last moment and the hidden states at each moment are first calculated by a fully connected layer (s ranges from 1 to p), where v a and W a are both trainable weight matrices. Then, the attention weight values are obtained through the softmax function (s ranges from 1 to p). The weighted vector is concatenated with the output of the unit at the last moment of the LSTM module and finally passed to a fully connected layer Dense layer to obtain the final predicted output of the AR result at time t.

[0091] Figure 5 The temporal attention module in Figure 5 can be flexibly designed according to the actual situation, such as Bahdanau Attention, Luong Attention, etc.

[0092] S6. Model training and effect evaluation:

[0093] Based on the above S5 neural network model, the sequence data of each selected feature (including the tunneling speed to be predicted and output) in the past p historical moments is used as the model input, and the output is the tunneling speed at time t. The relationship between the input and output parameters is:

[0094] Formula Six: AR| t = f((RS, TH, TOR, Es, Fa, AR)| t-1,t-2,…,t-p )

[0095] Among them, the function f is the mapping relationship that the deep learning model needs to fit, and p is the length of the historical time series data, and the optimal value can be determined by numerical experiments.

[0096] Initialize the parameters in the Attention-ResNet-LSTM hybrid neural network model, select the mean squared error (MSE) as the loss function, and use the Adam optimization algorithm to train the model.

[0097] Formula Seven: Among them, MSE is the mean squared error, b is the total number of data, y i is the measured value of the data, is the predicted value of the model.

[0098] At the same time, use the random search algorithm to optimize the hyperparameters, and select the combination with the smallest error on the validation set as the optimal hyperparameter (m, n, learning rate, number of convolutional kernels, number of LSTM hidden layer neurons, etc.) combination of the model.

[0099] To prevent the model from overfitting, an early stopping strategy is adopted during training: when the loss of the model on the validation set no longer decreases after a certain number of iterations, stop the training of the model.

[0100] To evaluate the prediction effect of the model, select the mean absolute percentage error MAPE and the root mean square error RMSE as evaluation indicators. The combination of the two can effectively evaluate the performance of the model, and their calculation methods are as follows:

[0101] Formula Eight:

[0102] Formula Nine:

[0103] The smaller the values of the two indicators MAPE and RMSE, the smaller the deviation between the predicted value and the true value, and the higher the fitting accuracy of the model. Use the trained optimal model to predict the tunneling speed on the test set. If the prediction accuracy meets the requirements, the model can be applied to the actual project to predict the shield tunneling speed in real time; if the prediction accuracy of the model does not meet the requirements, repeat the above steps until the prediction accuracy meets the actual application requirements.

[0104] S7, Tunneling Application and Intelligent Evolution

[0105] Based on the optimal network model obtained in S6, carry out parameter sensitivity analysis, and obtain three relationship curves between AR and cutter head rotation speed RS, AR and total thrust TH, and AR and cutter head torque TOR respectively, which are used to provide to the PLC control system. Therefore, according to the predicted value of AR, the shield tunneling system can automatically match the optimal tunneling parameters.

[0106] With a highly secure guarantee by feeding safe samples into the "Attention-ResNet-LSTM hybrid neural network model", the construction application of efficient tunneling is realized. As the safety construction projects are gradually accumulated and the samples become diversified, the neural network is evolved, and the "Attention-ResNet-LSTM hybrid neural network model" has extremely strong network robustness and generalization. Therefore, the present invention is a more intelligent and accurate prediction method.

[0107] Furthermore, embodiments are given:

[0108] Embodiment 1

[0109] As Figure 6 shown, during the tunneling process of a certain shield tunnel, the key construction parameters and stratum parameters are monitored in real time through sensors. The parameters include: cutter head rotation speed RS, cutter head torque TOR, total thrust TH, characteristic value of bearing capacity F a , compression modulus E S , tunneling speed AR.

[0110] The collected data is detected for null data and outliers according to Formula 1 and Formula 2:

[0111] Formula 1: f i = TH i ·TOR i ·AR i ·RS i

[0112] Formula 2:

[0113] The detected null data and outliers are removed from the original data, and the data is normalized through Formula 3.

[0114] Formula 3:

[0115] The optimal input parameter combination is selected through Formula 4 and Formula 5 by means of the EPR algorithm:

[0116] Formula 4:

[0117] Formula 5:

[0118] The new data set is divided into a training set, a validation set and a test set according to the ratio of 8:1:1. According to Figure 4Build the Attention-ResNet-LSTM hybrid neural network model with the shown structural diagram, initialize the weight parameters, and train the model with the data in the training set until the error tends to be stable. Use the random search algorithm to optimize the hyperparameters (m, n, learning rate, number of convolutional kernels, number of LSTM hidden layer neurons, etc.), select the combination with the minimum error on the validation set as the optimal hyperparameter combination of the model, and denote the optimal model at this time as M optimal 。

[0119] Use M optimal to predict the shield tunneling speed on the test set. If the prediction accuracy meets the requirements, then M optimal can be applied to the actual project to predict the shield tunneling speed in real time; if the prediction accuracy of the model does not meet the requirements, then repeat the above steps to find a new optimal model M′ optimal until the prediction accuracy meets the actual application requirements. Based on M′ optimal carry out parameter sensitivity analysis to obtain the relationship curves between the tunneling speed AR and the cutterhead rotation speed RS, AR and the total thrust TH, and AR and the cutterhead torque TOR.

[0120] It should be noted that the prediction model of the present invention, that is, M′ optimal can be directly applied to a new data set for AR prediction, that is, it can be directly applied to the engineering prediction of the shield tunneling speed; if the generalization ability of the model is weak, the updated data of the shield project can also be used to train and optimize the model to continuously improve the robustness of the prediction model so that it can maintain a satisfactory accuracy when applied in the shield project.

[0121] Figure 7 This is the prediction effect of the Attention-ResNet-LSTM model in this embodiment. It can be observed from the figure that the model has a good prediction effect on the tunneling speed, the predicted values are generally basically consistent with the measured values, the root mean square error RMSE of the model is 1.308 mm / min, and the mean absolute percentage error is 1.529%, showing a high prediction accuracy. In the actual project, shield operators and managers can use this intelligent prediction method to carry out real-time and accurate prediction of the shield tunneling speed, and know in advance the change trend of the shield tunneling work performance, so as to match and select the best shield construction parameters to ensure the safe and rapid tunneling of the shield tunnel.

[0122] Embodiment 2:

[0123] As Figure 6 shown, further disclose an intelligent control system for the tunneling speed of a shield tunneling method tunnel, including

[0124] The sensing system includes a rotational speed sensor and a force sensor, which are respectively arranged inside the large shield machine gear of the main drive shaft and on the master valve block of the grouped propulsion cylinders, and are used to output the cutter head rotational speed RS and the total propulsion force TH. The cutter head torque TOR is provided by the frequency conversion cabinet of the shield machine main drive motor, and the characteristic value of bearing capacity F a , compression modulus E S are obtained from the geological exploration report; a sample library is constructed and provided to the prediction system;

[0125] The computer prediction system includes a data preprocessing module and a network module; input parameter combinations are input from the sample library, and the predicted tunneling speed AR is output through the network module. The relationship curves between AR and the cutter head rotational speed RS, AR and the total propulsion force TH, and AR and the cutter head torque TOR are obtained through parameter sensitivity analysis (as shown in Figure 8 ), and provided to the PLC control system;

[0126] The PLC control system, based on Figure 8 the predicted tunneling speed AR and the relationship curves between AR and the above shield operation parameters (which can be obtained by the network module through parameter correlation analysis), automatically adjusts the relevant shield operation parameters, namely the cutter head rotational speed RS, the cutter head torque TOR, and the total propulsion force TH. The PLC control system is used under these three curves to control and match the tunneling equipment in real time at the optimal tunneling speed AR on the premise of safety, so as to reach the expected value, that is, safe and efficient.

[0127] It should be noted that Figure 8 the curves only originate from the Yangtze River shield crossing project of the China-Russia Eastern Gas Pipeline (Yongqing - Shanghai), and the formation type is mainly sandy clay. With the further expansion of the future data set, the correlation curves can be extended to various different formations, and their robustness and accuracy will be further improved.

[0128] Figure 6 This is the logic block diagram of the shield tunneling speed intelligent control system introduced in Embodiment 2. Relying on this system, the automatic control of the shield tunneling speed can be realized, and the optimal tunneling parameters can be intelligently matched, so as to ensure the safety and efficiency of shield construction.

Claims

1. An intelligent prediction method for the shield tunneling speed based on a hybrid neural network, characterized in that, Specifically, it includes the following steps: S1. Data collection: During the shield tunneling process, the shield operation parameters are recorded in real time at a fixed sampling frequency f, and the soil parameters during the shield tunneling process are also recorded. S2. Data preprocessing: 1) Detection of empty data 2) Detection of outliers Use the Mahalanobis Distance as the criterion for outlier discrimination; for a multivariate input sequence \(x=(x_1,x_2,x_3,x_4,x_5,x_6)\) with a mean of \(\mu = (\mu_1,\mu_2,\mu_3,\mu_4,\mu_5,\mu_6)\) T , and a covariance matrix \(S\) T , the calculation of its Mahalanobis Distance is shown in Equation 2: Formula 2: In the above formula, x refers to the input variables of cutter head rotation speed RS, cutter head torque TOR, total thrust force TH, characteristic value of bearing capacity F a , compression modulus E S and tunneling speed AR; 3) Data normalization Normalize the data and map it to the interval [0, 1]. S4. Dataset segmentation: Divide the preprocessed data into a training set, a validation set, and a test set. S5. Construct an Attention-ResNet-LSTM hybrid neural network model: The normalized input data is fed into the ResNet structure for feature extraction, and different weights are assigned to the feature maps of each channel extracted by the ResNet structure through the channel attention mechanism. The weighted feature maps are input into the LSTM network, and a temporal attention mechanism is added to the last layer of the LSTM structure to assign weight values to the hidden layer outputs at different times. The temporal vector after weighted summation is concatenated with the output of the last time step as the final output processed by the attention mechanism. S6. Model training and performance evaluation: Based on the above neural network model in S5, the sequence data of the past p historical moments of each selected feature is used as the model input, and the output is the tunneling speed at time t. The relationship between the input and output parameters is as follows: Formula VI: AR| t = f((RS, TH, TOR, Es, Fa, AR)| t-1,t-2,…,t-p ) Among them, the function f is the mapping relationship to be fitted by the deep learning model, and p is the length of the historical time series data. The optimal value can be determined by numerical experiments. Evaluate the prediction performance of the model. Select the mean absolute percentage error MAPE and the root mean square error RMSE as evaluation indicators. The combination of the two evaluates the performance of the model, and their calculation methods are as follows: Formula VIII: Formula Nine: The smaller the values of the two indicators MAPE and RMSE, the smaller the deviation between the predicted value and the true value, and the higher the fitting accuracy of the model. Use the trained optimal model to predict the tunneling speed on the test set. If the prediction accuracy meets the requirements, the model is applied to the actual project to predict the shield tunneling speed in real time. If the prediction accuracy of the model does not meet the requirements, repeat the above steps until the prediction accuracy meets the actual application requirements.

2. The method according to claim 1, wherein in S2, the detection algorithm for the empty data is as follows: Formula 1: f i = TH i · TOR i · AR i · RS i Among them, TH i , TOR i , AR i , RS i respectively represent the propulsion force, cutter head torque, tunneling speed and cutter head rotation speed of the shield machine at the i-th moment; when f i = 0, it means that the shield machine is in a shutdown state, and the data record at this moment is deleted.

3. The method according to claim 1, wherein For Formula 2, when each variable is uncorrelated, the covariance matrix S is an identity matrix, and the Mahalanobis distance is equal to the Euclidean distance; set the 0.9 quantile of D M as the threshold D t , and use D M >D t as the condition for outlier discrimination. Data points that meet this condition are regarded as outliers and removed from the dataset.

4. The method according to claim 1, wherein in S4: The training set is used for model training, the validation set is used for optimizing the model hyperparameters, and the test set is used for evaluating the final model performance.

5. The method according to claim 1, wherein in S5: For the Attention-ResNet-LSTM hybrid neural network model, its construction method and implementation strategy: First, input the sample data of p historical moments in the form of a multivariate time series into the neural network. The ResNet module in the network model consists of m basic units, and each basic unit includes two convolutional layers and corresponding residual connections to extract and output a feature map; Then, construct an LSTM module including n layers of LSTM units; and the attention mechanism includes a channel attention module and a temporal attention module. Among them, the Channel Attention module consists of a global average pooling layer and two fully connected layers. The feature map extracted by the ResNet module first passes through a global average pooling layer to obtain the global information of each channel, and then passes through two fully connected layers to obtain the weight values of each channel. The weighted feature map is input into the LSTM module; Among them, in the Temporal Attention module, the similarity score between the hidden state at the last moment and the hidden states at each moment is first calculated by a fully connected layer , where s ranges from 1 to p, and where v a and W a are both trainable weight matrices; then the attention weight values are obtained through the softmax function , where s ranges from 1 to p. The weighted vectors are concatenated with the unit output at the last moment of the LSTM module, and finally passed to a Dense layer to obtain the final predicted output of the AR result at time t 6. The method according to claim 5, characterized in that, Initialize the parameters in the Attention-ResNet-LSTM hybrid neural network model. The mean squared error (MSE) is selected as the loss function, and the Adam optimization algorithm is used to train the model; Formula VII: Among them, MSE is the mean square error, b is the total number of data, and y i is the measured value of the data, and is the predicted value of the model; At the same time, the random search algorithm is used for hyperparameter optimization, and the combination with the smallest error on the validation set is selected as the optimal hyperparameter combination of the model.

7. The method according to claim 6, wherein The hyperparameters include m, n, learning rate, the number of convolutional kernels, and the number of hidden neurons in the LSTM.

Images