Shield tunneling parameter multi-step continuous recursive prediction method, electronic equipment and storage medium

Abstract

The invention provides a multi-step continuous recursive prediction method for shield tunneling parameters, electronic equipment and a storage medium, and belongs to the technical field of intelligent construction of tunnel engineering. Comprising the following steps: preprocessing tunneling parameter historical data; machine-ground comprehensive index conversion, and data set and normalization criterion construction; constructing and training a tunneling parameter long short-term memory neural network multi-step prediction model; constructing a starting input data set, importing the starting input data set into a prediction model, outputting to obtain a # imgabs0 # group and # imgabs1 # tunneling parameter prediction data set, and converting the data set into a # imgabs2 # group historical-prediction data mixed recursive prediction input data set; and circularly inputting the results into the prediction model of each tunneling parameter to finish continuous prediction of the tunneling parameters in the limited area in front of the shield. According to the method, the operation process of manually setting the tunneling parameters can be reduced, the tunneling parameter multi-step prediction model is constructed based on the long-short-term memory network, and the long-distance continuous prediction of the tunneling parameters is realized by adopting time sequence multi-step recursion processing.

Description

Technical Field

[0001] The invention belongs to the technical field of intelligent construction of tunnel engineering, and relates to a multi-step continuous recursive prediction method of shield tunneling parameters, electronic equipment and storage medium. Background Art

[0002] The shield method has become one of the most widely used construction methods in my country's tunnel construction due to its many advantages such as fast construction speed, high degree of mechanization and strong adaptability to the strata. It is widely used in tunnel construction of infrastructure such as rail transit, road transportation and water diversion projects.

[0003] During the operation of the shield machine, the equipment operator often needs to rely on experience to set the shield tunneling control parameters section by section, and then manually adjust and optimize according to the equipment operating parameter status. However, the geological conditions of underground projects are highly complex and the heterogeneity changes rapidly. The mismatch between the tunneling parameters and the real-time formation conditions will increase the wear of the cutterhead and affect the tunneling efficiency. The unreasonable setting of tunneling parameters in sensitive areas can easily cause mud stagnation accidents and even machine jams and shutdowns. Modern shield machines equipped with integrated sensor systems have complete operating process monitoring capabilities. Advance prediction and preset recommendations of shield machine control parameters and operating parameters based on shield equipment conditions and massive monitoring data can further provide a basis for reducing construction costs, shortening construction periods, and improving intelligent shield tunneling construction technology.

[0004] At present, the intelligent prediction of tunneling parameters is mainly carried out by regression analysis, machine learning and deep learning. However, simple regression fitting analysis requires more assumptions and has low generalization and robustness in complex strata or facing large amounts of data. The nonlinearity, high parallelism and high fault tolerance of machine learning methods make them very attractive in intelligent construction. Artificial neural networks (ANN), support vector machines (SVM), regression trees (RT), random forests (RF) and other methods have been tried as core algorithms for predicting tunneling parameters. However, existing non-sequential machine learning methods such as traditional BP neural networks or single structures are difficult to fully extract the historical change characteristics of tunneling data for training and learning, and the accuracy of future time series prediction of tunneling parameters is difficult to guarantee. More complex deep / recurrent / convolutional neural network models are gradually being applied to tunneling parameter prediction models, such as convolutional neural networks (CNN), recurrent neural networks (RNN), generative adversarial networks (GAN) and other methods to improve the prediction accuracy and efficiency in more complex strata.

[0005] At present, the prediction input of tunneling parameters is usually based on current formation parameter information, historical tunneling parameters or other operation data of the shield machine body. A high-dimensional pure data-driven neural network prediction model is constructed to perform single-step / multi-step real-time prediction on the main tunneling parameters. However, the time required for training and optimization is generally long, and it is also difficult to achieve continuous prediction. Moreover, due to the high data dimension and the large difficulty in obtaining specialized real-time input data, especially when facing complex geological conditions, the generalization degree and transferability of the prediction model are relatively low, and the error rate or failure rate is relatively large, resulting in limited engineering application scenarios.

[0006] Generally speaking, at present, the prediction input dimension of tunneling parameters is large, the training and optimization time is long, it is difficult to achieve continuous prediction, and in the case of facing special complex geological conditions, the prediction generalization degree and transferability are low, and the error rate or failure rate is relatively large. Summary of the Invention

[0007] The present invention provides a multi-step continuous recursive prediction method for shield tunneling parameters, including the following steps: Step 1, preprocess the historical data of tunneling parameters; Step 2, convert the preprocessed tunneling parameters into shield-stratum comprehensive state indicators; Taking the shield machine-stratum comprehensive state indicator of each step as the input and setting the multi-step prediction tunneling parameters as the output to construct a multi-step prediction training input set tuple of the long short-term memory neural network model and a multi-step prediction training output set tuple of different categories of tunneling parameters ; Normalize the multi-step prediction training input set tuple and the multi-step prediction training output set tuple of different categories of tunneling parameters respectively to obtain a normalized tuple of the multi-step prediction training input set and a normalized tuple of the multi-step prediction training output set of different categories of tunneling parameters ; Store and record the normalization criterion of the normalized tuple of the multi-step prediction training input set as and store and record the anti-normalization criterion of the normalized tuple of the multi-step prediction training output set of different categories of tunneling parameters as ; Step 3, construct a long short-term memory neural network model based on the preprocessed tunneling parameter training set, and train the long short-term memory neural network model to obtain a long short-term memory neural network multi-step prediction model for each tunneling parameter; Step 4, select the normalized tuple of the multi-step prediction training input set The last set of data in [[]] forms the starting input data set for multi-step prediction. , and the constructed starting input data set is imported into the long short-term memory neural network multi-step prediction model that has been trained to obtain the normalized predicted values of the multi-step tunneling parameters for the th group; Step Five: Based on the normalization criterion and the inverse normalization criterion , the normalized predicted values of the multi-step tunneling parameters for the th group are gradually subjected to inverse normalization processing of the output data set, shield machine-stratum comprehensive state index conversion, and normalization processing of the input data set to obtain the th group of normalized input data sets ; Step Six: Using the starting input data set for multi-step prediction as the basic data set, the first steps of data in the basic data set are deleted and then the th group of normalized input data sets are supplemented to obtain the th group of historical-prediction data mixed recursive prediction input data sets ; Using the th group of historical-prediction data mixed recursive prediction input data sets as the reference set, the first steps of data in the reference set are deleted and then the th group of normalized input data sets are added to the reference set to obtain the th group of historical-prediction data mixed recursive prediction input data sets ; Step Seven: Based on the th group of historical-prediction data mixed recursive prediction input data sets , it is cyclically input into the long short-term memory neural network multi-step prediction models of each tunneling parameter to complete the continuous prediction of the tunneling parameters within a limited area in front of the shield.

[0008] Further, the process of preprocessing the historical data of the tunneling parameters in Step One includes using the box plot method to identify and remove outliers from the true time series data of the tunneling parameters; The specific process is as follows: S1.11: Using the upper normal value and the lower normal value of the historical data of the tunneling parameters determined by the box plot; S1.12. Consider the historical data of tunneling parameters with values higher than the upper limit of the normal value or lower than the lower limit of the normal value in the historical data of tunneling parameters as abnormal values; S1.13. Calculate the replacement values for replacing the abnormal values , and then based on the replacement values , use the moving window average method to replace the abnormal values in the historical data of tunneling parameters.

[0009] Furthermore, the process of preprocessing the historical data of tunneling parameters in step one also includes using discrete wavelet transform to denoise the tunneling parameters after abnormal value removal and replacement; The specific method is as follows: S1.21. Use wavelet transform to decompose the time series data in the tunneling parameters into approximation coefficients and detail coefficients at different scales; S1.22. Perform threshold processing on the detail coefficients to suppress noise and retain the main features in the tunneling parameters simultaneously, obtaining the preprocessed time series data of tunneling parameters; specifically: ①. Use soft threshold to process the detail coefficients to obtain the adjusted equivalent detail coefficients ; ②. Replace the detail coefficients with the adjusted equivalent detail coefficients , and then perform inverse wavelet transform to reconstruct the denoised time series data , that is, obtain the preprocessed tunneling parameters; the preprocessed tunneling parameters include the preprocessed tunneling speed, the preprocessed cutter head rotation speed, the preprocessed cutter head torque, the preprocessed total thrust force, and the preprocessed cutter head torque.

[0010] Furthermore, the specific process of obtaining the long short - term memory neural network multi - step prediction model for each tunneling parameter is as follows: S3.1. Based on the preprocessed tunneling parameters, construct the corresponding long short - term memory neural network model; that is: based on the preprocessed tunneling speed, construct a long short - term memory neural network model including the tunneling speed; based on the preprocessed cutter head rotation speed, construct a long short - term memory neural network model including the cutter head rotation speed; based on the preprocessed cutter head torque, construct a long short - term memory neural network model including the cutter head torque; based on the preprocessed total thrust force, construct a long short - term memory neural network model including the total thrust force; based on the preprocessed cutter head torque, construct a long short - term memory neural network model including the cutter head torque; S2.2. For each layer and each gating mechanism of each long short-term memory neural network model, the error backpropagation algorithm is used to calculate the transfer error over its entire time series and based on the transfer error iterative training is performed until the number of iterations reaches the maximum number of iterations, and a long short-term memory neural network multi-step prediction model for each tunneling parameter is obtained.

[0011] Furthermore, the architecture of a single long short-term memory neural network model is specifically as follows: ①. Create an input feature layer and set the format of this input feature layer according to the data sequence features of the constructed multi-step prediction training normalized input set. ②. Create a data flattening layer to further flatten the multi-dimensional input data passing through the input feature layer into one-dimensional data to better conform to the training of the LSTM model. ③. Create an LSTM layer responsible for feature learning, determine the number of hidden units in this layer, and use the He initialization method to initialize the initial weights and input weights. ④. Create an LSTM layer used to output only the state of the last time step in the sequence, determine the number of hidden units in this layer, and also specify the initialization method of the initial and input weights as He initialization. ⑤. Create a dropout layer to randomly discard a certain proportion of hidden units during the training process, and determine the random discard ratio through grid search trial calculation. ⑥. Create a fully connected layer and determine the number of output hidden units according to the number of output variables (m steps) of the constructed multi-step prediction training normalized output set. ⑦. Create an output feature layer to realize the output of multi-step prediction values.

[0012] Furthermore, the specific process of training a single long short-term memory neural network model is as follows: ①. Determine the hyperparameters of the long short-term memory neural network model, and the hyperparameters include the number of hidden layers, the maximum number of training times, the initial learning rate, the learning rate adjustment factor, and the dropout rate. ②. Based on the multi-step prediction training input data set and the multi-step prediction training output data set obtained in step one, use the adam optimization algorithm to train and learn the multi-step prediction model respectively, and obtain the connection weights and bias terms between the layers of the multi-step prediction model including the tunneling speed, the connection weights and bias terms between the layers of the multi-step prediction model including the cutter head rotation speed, the connection weights and bias terms between the layers of the multi-step prediction model including the cutter head torque, and the connection weights and bias terms between the layers of the multi-step prediction model including the total thrust force. ③. For each layer and each gating mechanism of the tunneling parameter multi-step prediction model, the error backpropagation algorithm is used to calculate the transfer error over the entire time series Calculation; ④. Based on the transmission error Iteratively train each long short - term memory neural network model until the number of iterations reaches the maximum number of iterations, and obtain the long short - term memory neural network multi - step prediction model for each tunneling parameter.

[0013] Further, the specific process of the group - normalized input dataset is as follows: S5.1. Denormalize the output dataset using the normalized predicted values of the multi - step tunneling parameters in the group to obtain the multi - step predicted tunneling parameters in the group; ; S5.2. Convert the multi - step predicted tunneling parameters in the group into the shield - formation comprehensive state index of the group; The shield - formation comprehensive state index includes displacement index , strength index, and energy index; S5.3. Convert the shield - formation comprehensive state index of the group into the group - normalized input dataset of the long short - term memory neural network model .

[0014] Further, the specific process of completing the continuous prediction of tunneling parameters within a limited area in front of the shield is as follows: S7.1. Import the recursive prediction input dataset of the historical - prediction data mixture in the group into the trained multi - step long short - term memory neural network model to obtain the prediction results in the group; S7.2. Based on the prediction results in the group, return to step five and repeat step five and step six to form a new recursive prediction input dataset again; S7.3. Assume the number of prediction steps in front of the shield is

[0015] The present invention also provides an electronic device comprising a memory, one or more processors and one or more programs stored in the memory, wherein the one or more programs include instructions for executing the multi-step continuous recursive prediction method for shield tunneling parameters as described above.

[0016] The present invention also provides a storage medium, comprising one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs include instructions for executing the multi-step continuous recursive prediction method for shield tunneling parameters as described above.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The method for continuous prediction of excavation parameters provided by the present invention can reduce the operation process of manually setting excavation parameters, construct a multi-step prediction model of excavation parameters based on a long short-term memory network, and adopt a time-series multi-step recursive process to achieve long-distance continuous prediction of excavation parameters; It can also improve prediction efficiency while ensuring prediction accuracy, while reducing the difficulty of pre-training learning and the error transmission rate of the deep neural network model, providing theoretical and technical support for intelligent construction and parameter pre-control of shield tunneling.

[0018] Compared with the best existing technologies, the main differences and advantages of the present invention are that the predicted input of tunneling parameters no longer relies on the parameters provided by the shield tunneling machine, but adopts coupled parameter indicators that can reflect the stratum information and the tunneling status, which has stronger stability and portability for the prediction process; the existing technology can only predict one or more values ​​at a time, and the recursive prediction method considered in the present invention can realize the forward continuous multi-step prediction of tunneling parameters, and can more effectively predict the tunneling parameters at farther locations.

[0019] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 It is a schematic diagram of the overall process of a multi-step continuous recursive prediction method for shield tunneling parameters in an embodiment of the present invention; Figure 2 Schematic diagram of the gating mechanism architecture in a single long short-term memory neural network model in an embodiment of the present invention; Figure 3 It is a schematic diagram of the result of identifying, eliminating and replacing abnormal values ​​in the excavation parameters in an embodiment of the present invention; Figure 4(a) is a schematic diagram of the original data of the total propulsion force in the embodiment of the present invention; Figure 4(b) is a schematic diagram of the result after wavelet denoising processing of the original data shown in Figure 4(a); Figure 5 It is a schematic diagram of the multi-step continuous recursive prediction result of the cutter head torque in the embodiment of the present invention; Figure 6 It is a schematic diagram of the multi-step continuous recursive prediction result of the total propulsion force in the embodiment of the present invention; Figure 7 It is a schematic diagram of the multi-step continuous recursive prediction result of the propulsion speed in the embodiment of the present invention; Figure 8 It is a schematic diagram of the multi-step continuous recursive prediction result of the cutter head rotation speed in the embodiment of the present invention. Specific Embodiments

[0021] To make the above objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is provided in conjunction with the accompanying drawings. It should be noted that the accompanying drawings of the present invention are all in simplified forms and use non-precise scales, only for the convenience of clearly assisting in the description of the embodiments of the present invention. Embodiment:

[0022] Refer to Figure 1 As shown, a multi-step continuous recursive prediction method for shield tunneling parameters provided by the present invention mainly relates to the multi-step continuous prediction of tunneling parameters in a limited area in front during the stable construction of a shield machine; the prediction method includes the following steps: Step 1. Preprocess the historical data of tunneling parameters; specifically: use the box plot method to remove and replace outliers from the true time series data of the historical data of tunneling parameters generated during the shield tunneling construction of a certain existing project (the true time series data generally has a large number of step data, related to time), and use discrete wavelet transform to remove the noise in the historical data of shield tunneling to obtain the preprocessed time series data .

[0023] Preferably, the tunneling parameters include tunneling speed, cutter head rotation speed, cutter head torque, total propulsion force, etc.

[0024] Furthermore, the specific process of using the box plot method to identify and remove outliers from the true time series data of tunneling parameters and replace them with the moving window average method (the results of identifying, removing, and replacing outliers in the tunneling parameters are shown in Figure 3 ) is as follows: S1.11. Determine the upper limit of the normal value and the lower limit of the normal value of the historical data of tunneling parameters using the box plot; Upper limit of normal value of tunneling historical data and lower limit of normal value are expressed as follows: ; ; Among them, is the first quartile of the normal value range of tunneling parameters, is the third quartile of the normal value range of tunneling parameters, is the interquartile range of the normal value range of tunneling parameters; S1.12. Consider the tunneling parameter historical data with values higher than the upper limit of normal value or lower than the lower limit of normal value as outliers; S1.13. Calculate the replacement value for the outlier, and then replace the outlier in the tunneling parameter historical data with the replacement value using the moving window average method; The replacement value is expressed as follows: ; Among them, is the position of the outlier in the tunneling parameter historical data, is the radius of the moving average window, is the th value of the moving window, and is the specific data of the tunneling parameter historical data (i.e., is any one of tunneling speed, cutterhead rotation speed, cutterhead torque, total thrust, etc.).

[0025] Furthermore, the specific method for noise reduction processing of the tunneling parameters after outlier removal and replacement using discrete wavelet transform (the total thrust in the tunneling parameters in the original data without outlier removal and replacement is shown in Figure 4(a), and the result of noise reduction processing of the tunneling parameters after outlier removal and replacement using discrete wavelet transform is shown in Figure 4(b)) is as follows: S1.21. Use wavelet transform to decompose the time series data (time - series data) in the tunneling parameters into approximation coefficients (the approximation coefficient is the low - frequency signal) and detail coefficients (the detail coefficient is the high - frequency signal); The expression for decomposing the time series data is as follows: ; Among them, is the scaling function, is a wavelet function. Both the scaling function and the wavelet function are determined by the selected wavelet basis. In this embodiment, the scaling function and the wavelet function are specifically generated using the Daubechies wavelet (db8); is the approximation coefficient, is the detail coefficient; the subscript " " is the scale index, and the subscript " " is the position index (meaning the specific position of the tunneling parameter in the time domain), and the subscript " is the coarsest scale of the decomposition (meaning the lowest frequency); S1.22. Perform threshold processing on the detail coefficient to suppress noise and retain the main features in the tunneling parameters, obtaining the preprocessed time-series data of the tunneling parameters. Specifically: ①. Use the soft threshold to process the detail coefficient to obtain the adjusted equivalent detail coefficient ; The adjusted equivalent detail coefficient has the following expression: ; where, is the positive or negative sign of the detail coefficient ; is the soft threshold. In this embodiment, the soft threshold uses the universal threshold for calculation, is the standard deviation of the noise, is the length of the time-series data.

[0026] ②. Replace the detail coefficient with the adjusted equivalent detail coefficient , and then perform the inverse wavelet transform to reconstruct the denoised time-series data , that is, obtain the preprocessed tunneling parameters (including the preprocessed tunneling speed, the preprocessed cutterhead rotation speed, the preprocessed cutterhead torque, the preprocessed total thrust, and the preprocessed total extrusion force).

[0027] Step 2. Convert the preprocessed tunneling parameters into displacement indexes , strength indexes and energy indexes of the shield-ground comprehensive state, and form the multi-step prediction training input set and output set of the long short-term memory neural network model after constructing the stride, and perform data normalization and labeling processing on the input set and output set respectively, obtaining the normalization criteria and inverse normalization criteria of the input set and output set respectively.

[0028] The specific process is as follows: S2.1. Convert the pre - processed tunneling parameters into shield - formation comprehensive state indicators with steps. The shield - formation comprehensive state indicators include displacement indicators , strength indicators and energy indicators . The expression of the displacement indicator is as follows: . The expression of the strength indicator is as follows: . The expression of the energy indicator is as follows: . Among them, the unit of the displacement indicator is , and its physical meaning is the tunneling speed of the shield per unit cutterhead rotation speed, that is, the depth of the cutterhead cutting into the formation per revolution is used as the displacement indicator; the unit of the strength indicator is respectively , and its physical meaning is to define the total thrust of the shield machine under unit penetration; the unit of the energy indicator is , and its physical meaning is the ratio of the actual energy consumption of the shield tunneling per unit formation to the work done by the shield machine; is the pre - processed tunneling speed, with the unit of ; is the pre - processed cutterhead rotation speed, with the unit of ; is the pre - processed total thrust, with the unit of ; is the pre - processed cutterhead torque, with the unit of ; is the cutterhead diameter, with the unit of .

[0029] S2.2. Determine a number of steps as the construction step size (that is, using steps of shield - formation comprehensive state indicators as input, and the construction step size is generally much smaller than the total number of steps ), and set multi - step predicted tunneling parameters as output to construct the multi - step prediction training input set tuple and the multi - step prediction training output set tuple of different categories of tunneling parameters; The multi - step prediction training input set tuple The expression is as follows: ; Tuple of the multi-step prediction training output set for different categories of tunneling parameters The expression is as follows: ; Where: is the shield machine - stratum comprehensive state index matrix, including displacement index, strength index and energy index; is the tunneling parameter, is the category of different tunneling parameters; is the multi-step prediction number of steps for tunneling parameters, is the total number of training groups for tunneling parameter prediction.

[0030] S2.3. Respectively, perform normalization processing on the multi-step prediction training input set tuple and the multi-step prediction training output set tuple for different categories of tunneling parameters to obtain the normalized tuple of the multi-step prediction training input set and the normalized tuple of the multi-step prediction training output set for different categories of tunneling parameters ; The expression of the normalized tuple of the multi-step prediction training input set is as follows: ; The expression of the normalized tuple of the multi-step prediction training output set for different categories of tunneling parameters is as follows: ; Among them, are different machine - stratum comprehensive state indicators (displacement indicator, strength indicator and energy indicator); is the minimum value in this matrix, is the maximum value in this matrix.

[0031] S2.4. Store and record the normalization criterion of the normalized tuple of the multi-step prediction training input set as , store and record the normalization criterion of the normalized tuple of the multi-step prediction training output set for different categories of tunneling parameters as , store and record the anti-normalization criterion of the normalized tuple of the multi-step prediction training input set as and store and record the anti-normalization criterion of the normalized tuple of the multi-step prediction training output set for different categories of tunneling parameters as .

[0032] Preferably, in this embodiment, the construction stride is set to 100, the multi-step prediction step number is set to 2, and the total number of training groups is set to 4750.

[0033] Step 3: Construct a long short-term memory neural network model based on the preprocessed tunneling parameters, and train the long short-term memory neural network model to obtain a long short-term memory neural network multi-step prediction model for each tunneling parameter.

[0034] The specific process is as follows: S3.1: Based on the preprocessed tunneling parameters, construct a corresponding long short-term memory neural network model; that is, based on the preprocessed tunneling speed, construct a long short-term memory neural network model including the tunneling speed; based on the preprocessed cutter head rotation speed, construct a long short-term memory neural network model including the cutter head rotation speed; based on the preprocessed cutter head torque, construct a long short-term memory neural network model including the cutter head torque; based on the preprocessed total thrust force, construct a long short-term memory neural network model including the total thrust force; based on the preprocessed cutter head torque, construct a long short-term memory neural network model including the cutter head torque; S2.2: For each long short-term memory neural network model, the error backpropagation algorithm is used between each layer and each gating mechanism to calculate the transfer error over its entire time series, and iterative training is performed based on the transfer error until the number of iterations reaches the maximum number of iterations, obtaining a long short-term memory neural network multi-step prediction model for each tunneling parameter.

[0035] Furthermore, the architecture of a single long short-term memory neural network model is specifically as follows: ①. Create an input feature layer, and set the format of this input feature layer according to the data sequence features of the constructed multi-step prediction training normalized input set; ②. Create a data flattening layer to further flatten the multi-dimensional input data passing through the input feature layer into one-dimensional data to better conform to the training of the LSTM model (for the specific architecture of the LSTM model, see Figure 2 as shown); ③. Create an LSTM layer responsible for feature learning, determine the number of hidden units in this layer, and use the He initialization method to initialize the initial weights and input weights; ④. Create an LSTM layer used to output only the state of the last time step in the sequence, determine the number of hidden units in this layer, and also specify the initialization method of the initial and input weights as He initialization; ⑤. Create a dropout layer to randomly discard a certain proportion of hidden units during the training process, and determine the random discard proportion through grid search trial calculations; ⑥. Create a fully connected layer and determine the number of output hidden units based on the number of output variables (m steps) of the constructed multi-step prediction training normalized output set. ⑦. Create an output feature layer to output the multi-step prediction values.

[0036] Furthermore, the specific process of training a single long short-term memory neural network model is as follows: ①. Determine the hyperparameters of the long short-term memory neural network model, where the hyperparameters include the number of hidden layers, the maximum number of training times, the initial learning rate, the learning rate adjustment factor, and the dropout rate. ②. Based on the multi-step prediction training input dataset and the multi-step prediction training output dataset obtained in step ①, use the adam optimization algorithm to train and learn the multi-step prediction model respectively, and obtain the connection weights and bias terms between the layers of the multi-step prediction model including the tunneling speed, the connection weights and bias terms between the layers of the multi-step prediction model including the cutterhead rotation speed, the connection weights and bias terms between the layers of the multi-step prediction model including the cutterhead torque, and the connection weights and bias terms between the layers of the multi-step prediction model including the total thrust force. ③. The error backpropagation algorithm is used for both the multi-step prediction model of tunneling parameters between layers and each gating mechanism to calculate the transmission error over the entire time series. Calculate. The transmission error is expressed as follows: ; Where: is the true value of the th tunneling parameter (i.e., the th value in the normalized tuple of the multi-step prediction training output set of different tunneling parameter categories), is the predicted value of the th tunneling parameter, is the total number of training samples, is the th layer structure of the multi-step prediction model, is the total number of layers of the multi-step prediction model, is the L2 regularization coefficient; ④. Based on the transmission error iteratively train each long short-term memory neural network model until the number of iterations reaches the maximum number of iterations, and obtain the long short-term memory neural network multi-step prediction models for each tunneling parameter; The long short-term memory neural network multi-step prediction models for each tunneling parameter are recorded as ; Where, For different categories of tunneling parameters (tunneling parameters include tunneling speed, cutterhead rotation speed, cutterhead torque, total thrust force, and total extrusion force).

[0037] Step Four: Select the last set of data in the multi-step prediction training input set normalization tuple to form the starting input data set for multi-step ( step) prediction , and import the constructed starting input data set into the multi-step prediction model of the long short-term memory neural network that has been trained to obtain the normalized predicted values of the multi-step tunneling parameters ; The expression of the starting input data set for multi-step prediction is as follows: ; The normalized predicted values of the multi-step tunneling parameters for the group are expressed as follows: .

[0038] Step Five: Based on the normalization criterion and the inverse normalization criterion , perform inverse normalization processing of the output data set, shield machine-stratum comprehensive state index conversion, and normalization processing of the input data set on the normalized predicted values of the multi-step tunneling parameters for the group step by step to obtain the group of normalized input data sets ; The specific process is as follows: S5.1. Perform inverse normalization processing of the output data set on the normalized predicted values of the multi-step tunneling parameters for the group to obtain the group of multi-step predicted tunneling parameters ; S5.2. Convert the group of multi-step predicted tunneling parameters into the group of shield machine-stratum comprehensive state indexes ; The shield machine-stratum comprehensive state index includes displacement index, strength index, and energy index; S5.3. Convert the group of shield machine-stratum comprehensive state indexes into the group of normalized input data sets of the long short-term memory neural network model ; The expression of the first - group multi - step prediction tunneling parameters ; The expression of the first - group shield - formation comprehensive state index ; The expression of the first - group normalized input data set .

[0039] Step 6: Using the starting input data set for multi - step prediction as the basic data set, delete the first steps of data in the basic data set, and then supplement the first - group normalized input data set to obtain the first - group historical - prediction data - mixed recursive prediction input data set ; Using the first - group historical - prediction data - mixed recursive prediction input data set as the reference set, delete the first steps of data in the reference set, and then supplement the first - group normalized input data set to the reference set to obtain the second - group historical - prediction data - mixed recursive prediction input data set ; The expression of the second - group historical - prediction data - mixed recursive prediction input data set .

[0040] Step 7: Based on the second - group historical - prediction data - mixed recursive prediction input data set , complete the continuous prediction of tunneling parameters within a limited area in front of the shield.

[0041] The specific process is as follows: S7.1: Import the second - group historical - prediction data - mixed recursive prediction input data set into the multi - step long short - term memory neural network model that has been trained, and obtain the second - group prediction results; S7.2: Based on the Return to Step 5 with the prediction results of the group, and repeat Steps 5 and 6 to form a new recursive prediction input data set again. ; S7.3. Set the number of prediction steps in front of the shield tunneling machine to (in this embodiment, it is set to 250). Based on the new recursive prediction input data set repeat S7.1 and S7.2 until the predetermined number of prediction steps is completed, that is, complete the continuous prediction of the tunneling parameters within a limited area in front of the shield tunneling machine (see the multi-step recursive and continuous prediction results of the tunneling speed, cutterhead rotation speed, cutterhead torque, and total thrust force in the tunneling parameters in Figures 5 to 8 as shown).

[0042] Further, the number of prediction steps is determined after convergence analysis according to the mean absolute percentage error of calculating the multi-step recursive prediction set ; The expression of the mean absolute percentage error is as follows: ; Where: is the predicted value of the th prediction of the multi-step recursive prediction set for the th tunneling parameter; is the original true value corresponding to the predicted value of the th prediction of the multi-step recursive prediction set for the th tunneling parameter.

[0043] Further, in the process of continuously looping S7.1 and S7.2, the anti-normalized true numerical results of the tunneling parameter predictions are stored and recorded each time a loop is completed until the predetermined number of prediction steps is completed.

[0044] As a further embodiment of the present invention, the present invention also provides an electronic device, including: One or more processors; A storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the foregoing method.

[0045] In specific use, a user can interact with a server, which is also an electronic device, through an electronic device serving as a terminal device and based on a network to implement functions such as receiving or sending messages. The terminal device is generally various electronic devices equipped with a display device and used based on a human-machine interface, including but not limited to smartphones, tablets, laptop computers, desktop computers, etc. Various specific application software can be installed on the terminal device as needed, including but not limited to web browser software, instant messaging software, social platform software, shopping software, etc.

[0046] Furthermore, the server is a network server for providing various services, such as a background server that provides corresponding computing services for the received tunneling parameters, long short-term neural network multi-step prediction models, etc. transmitted from the terminal device, so as to implement the processing of continuous prediction of tunneling parameters within a limited area in front of the shield, calculate the specific values of the tunneling parameters in a certain area in front of the shield tunneling, and finally return them to the terminal device.

[0047] As a further embodiment of the present invention, the present invention also provides a storage medium, including one or more programs for execution by one or more processors of an electronic device, and the one or more programs include instructions for executing the multi-step continuous recursive prediction method of shield tunneling parameters as described above.

[0048] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A multi-step continuous recursive prediction method for shield tunneling parameters, characterized in that: The following steps are involved: Step 1: preprocess the historical data of tunneling parameters; Step 2: converting the pre-processed excavation parameters into shield-stratum comprehensive state indicators; by The multi-step prediction training input set tuple of the long short-term memory neural network model is constructed by taking the comprehensive state index of the shield machine-stratum as input and setting the multi-step prediction excavation parameters as output. The output set of multi-step prediction training for different categories of tunneling parameters is tuple ; For multi-step prediction training input set tuples The output set of multi-step prediction training for different categories of tunneling parameters is tuple Perform normalization to obtain the normalized tuple of the multi-step prediction training input set Normalized tuple of multi-step prediction training output set for different categories of tunneling parameters ; Normalize the multi-step forecast training input set into a tuple The normalization criteria for storing records is And normalize the multi-step prediction training output set of different categories of tunneling parameters into tuples The anti-normalization criterion is stored as ; Step 3: construct a long short-term memory neural network model based on the preprocessed excavation parameter training set, and train the long short-term memory neural network model to obtain a long short-term memory neural network multi-step prediction model for each excavation parameter; Step 4: Select the normalized tuple of the multi-step prediction training input set The last set of data in forms the starting input data set for multi-step prediction , and construct the starting input data set Import it into the already trained long short-term memory neural network multi-step prediction model to obtain the first Normalized prediction values ​​of multi-step tunneling parameters ; Step 5: Based on the normalization criterion and the anti-normalization criterion For Normalized prediction values ​​of multi-step tunneling parameters The output data set is denormalized, the shield machine-stratum comprehensive state index is converted, and the input data set is normalized. Group normalized input dataset ; Step 6: Input dataset for multi-step forecasting As the basic data set, the previous After deleting the first step data, Group normalized input dataset Add in and get Recursive forecasting input dataset with mixed historical and forecast data ; First Recursive forecasting input dataset with mixed historical and forecast data As the benchmark set, the previous After deleting the first data, Group normalized input dataset Add to the benchmark set and get Recursive forecasting input dataset with mixed historical and forecast data ; Step 7: Based on Recursive forecasting input dataset with mixed historical and forecast data , and cyclically input into the long short-term memory neural network multi-step prediction model of each excavation parameter to complete the continuous prediction of the excavation parameters in the limited area in front of the shield.

2. The shield tunneling parameter multi-step continuous recursive prediction method according to claim 1 is characterized in that: The process of preprocessing the historical data of the excavation parameters in step 1 includes using a box plot method to identify and remove abnormal values ​​from the real time series data of the excavation parameters; The whole process is as follows: S1.

11. Upper limit of normal value of historical data of tunneling parameters determined by box plot and lower limit of normal ; S1.

12. The values ​​in the historical data of the excavation parameters are higher than the upper limit of the normal value. Or below the lower limit of normal The historical data of tunneling parameters are regarded as abnormal values; S1.

13. Calculate replacement values ​​for outliers , and then based on the replacement value The sliding window average method is used to replace abnormal values ​​in the historical data of tunneling parameters.

3. The shield tunneling parameter multi-step continuous recursive prediction method according to claim 2 is characterized in that: The process of preprocessing the historical data of the excavation parameters in step 1 also includes using discrete wavelet transform to perform noise reduction processing on the excavation parameters after the outliers are removed and replaced; The specific method is: S1.

21. Use wavelet transform to decompose the time series data in the tunneling parameters into approximate coefficients of different scales and detail factor ; S1.22, detail coefficient Threshold processing is performed to suppress noise while retaining the main features of the excavation parameters to obtain the preprocessed excavation parameter time series data; specifically: ① Use soft threshold to adjust detail coefficient Processing is performed to obtain the adjusted equivalent detail coefficient ; ②、The adjusted equivalent detail coefficient Replace detail factor , and then perform inverse wavelet transform to reconstruct the denoised time series data , that is, the excavation parameters after preprocessing are obtained; The pre-processed excavation parameters include the pre-processed excavation speed, the pre-processed cutter head rotation speed, the pre-processed cutter head torque, the pre-processed total propulsion force and the pre-processed cutter head torque.

4. The shield tunneling parameter multi-step continuous recursive prediction method according to claim 3 is characterized in that: The specific process of obtaining the long short-term memory neural network multi-step prediction model of each tunneling parameter is as follows: S3.

1. Based on the preprocessed excavation parameters, a corresponding long short-term memory neural network model is constructed; that is, based on the preprocessed excavation speed, a long short-term memory neural network model including the excavation speed is constructed; based on the preprocessed cutter head speed, a long short-term memory neural network model including the cutter head speed is constructed; based on the preprocessed cutter head torque, a long short-term memory neural network model including the cutter head torque is constructed; Based on the preprocessed total propulsion force, a long short-term memory neural network model including the total propulsion force is constructed; based on the preprocessed cutter head torque, a long short-term memory neural network model including the cutter head torque is constructed; S3.

2. Each layer and each gating mechanism of each long short-term memory neural network model uses the error back propagation algorithm to transfer the error over the entire time series. Calculate and based on the transmission error Iterative training is performed until the number of iterations reaches the maximum number of iterations, and a long short-term memory neural network multi-step prediction model for each tunneling parameter is obtained.

5. The shield tunneling parameter multi-step continuous recursive prediction method according to claim 4 is characterized in that: The architecture of a single long short-term memory neural network model is as follows: ①. Create an input feature layer and set the format of the input feature layer with the data sequence features of the constructed multi-step prediction training normalized input set; ② Create a data flattening layer to further flatten the multi-dimensional input data that has passed through the input feature layer into one-dimensional data to better suit the training of the LSTM model; ③. Create an LSTM layer responsible for feature learning, determine the number of hidden units in this layer, and use the He initialization method to initialize the initial weights and input weights; ④. Create an LSTM layer that only outputs the state of the last time step in the sequence, and determine the number of hidden units in this layer. Also specify the initialization method of the initial and input weights as He initialization; ⑤. Create a random dropout layer, randomly discard a certain proportion of hidden units during training, and determine the random dropout ratio through grid search trial calculation; ⑥ Create a fully connected layer to determine the number of output hidden units based on the number of output variables of the constructed multi-step prediction training normalized output set; ⑦. Create an output feature layer to output multi-step prediction values.

6. The shield tunneling parameter multi-step continuous recursive prediction method according to claim 5, characterized in that: The specific process of training a single long short-term memory neural network model is as follows: ①, determine the hyperparameters of the long short-term memory neural network model, the hyperparameters include the number of hidden layers, the maximum number of training times, the initial learning rate, the learning rate adjustment factor, and the random inactivation rate; ②, according to the multi-step prediction training input data set and the multi-step prediction training output data set obtained in step 1, the multi-step prediction model is trained and learned respectively by using the adam optimization algorithm, and the connection weights and bias items between each layer of the multi-step prediction model including the tunneling speed, the connection weights and bias items between each layer of the multi-step prediction model including the cutterhead speed, the connection weights and bias items between each layer of the multi-step prediction model including the cutterhead torque, and the connection weights and bias items between each layer of the multi-step prediction model including the total propulsion force are obtained respectively; ③、The error back propagation algorithm is used between each layer and each gating mechanism of the tunneling parameter multi-step prediction model to transfer the transmission error of the entire time series. calculate; ④ Based on transmission error Each long short-term memory neural network model is iteratively trained until the number of iterations reaches the maximum number of iterations, and a long short-term memory neural network multi-step prediction model of each tunneling parameter is obtained.

7. The shield tunneling parameter multi-step continuous recursive prediction method according to claim 6, characterized in that: No. Group normalized input dataset The specific process is as follows: S5.

1. Normalized prediction values ​​of multi-step tunneling parameters The normalized predicted values ​​of the multi-step tunneling parameters are denormalized to obtain the output data set. Multi-step prediction of tunneling parameters ; S5.

2. Multi-step prediction of tunneling parameters Convert to TBM-stratum comprehensive status index of the group ; Shield machine-comprehensive stratum status indicators include displacement indicators , intensity index and energy index; S5.

3. TBM-stratum comprehensive status index of the group The first step of transforming into a long short-term memory neural network model Group normalized input dataset .

8. The shield tunneling parameter multi-step continuous recursive prediction method according to claim 7, characterized in that: The specific process of completing the continuous prediction of tunneling parameters in the limited area in front of the shield is as follows: S7.

1. Recursive forecasting input dataset with mixed historical and forecast data Import it into the already trained long short-term memory neural network multi-step prediction model to obtain the first The predicted results of the group; S7.2, based on The prediction results of the group are returned to step 5, and steps 5 and 6 are repeated to form a new recursive prediction input data set ; S7.3, let the predicted number of steps ahead of the shield be , based on the new recursive prediction input dataset Repeat S7.1 and S7.2 until the predetermined number of prediction steps is completed, that is, the continuous prediction of the excavation parameters in the limited area in front of the shield is completed.

9. An electronic device, characterized in that: It comprises a memory, one or more processings and one or more programs stored in the memory, wherein the one or more programs comprise instructions for executing the multi-step continuous recursive prediction method for shield tunneling parameters as described in any one of claims 1-8.

10. A storage medium, characterized in that: It comprises one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs include instructions for executing the multi-step continuous recursive prediction method for shield tunneling parameters as described in any one of claims 1-8.

Images