Full-face tunnel boring machine cutterhead torque long-term prediction method and system

Abstract

The present invention provides a method and system for long-term prediction of cutterhead torque of a full-face tunnel boring machine, including: collecting cutterhead torque signals during the tunneling process of the full-face tunnel boring machine and preprocessing them to obtain a cutterhead torque sequence; decomposing the cutterhead torque sequence into high-frequency and low-frequency parts by using a wavelet packet decomposition matrix; decomposing the low-frequency part into several subsequences and a residual sequence; decomposing the high-frequency part into several subsequences; constructing a multi-step long-term prediction neural network model for cutterhead torque and training it; normalizing several subsequences respectively by using the min-max method and transmitting them to the trained multi-step long-term prediction neural network model for cutterhead torque to obtain several prediction results; adding the several prediction results to obtain the cutterhead torque value at the predicted time t; calculating the mean absolute percentage error, root mean square error and mean absolute error respectively according to the multiple cutterhead torque values obtained by prediction, and evaluating the prediction performance of the cutterhead torque.

Description

Technical Field

[0001] The present invention relates to the technical field of parameter prediction and optimization. Specifically, it relates to a long-term prediction method and system for the cutterhead torque of a full-face tunnel boring machine. More specifically, it relates to a long-term prediction method and system for the cutterhead torque of a full-face tunnel boring machine based on adaptive multi-level decomposition. Background Art

[0002] A full-face tunnel boring machine is a large-scale device integrating mechanical manufacturing, new materials, new technologies, electronic information, and automation technologies. Shield tunneling construction has the advantages of high excavation quality, safety, little influence on ground settlement and the environment, and its excavation efficiency is 3 to 10 times that of the traditional drill-and-blast method. Therefore, full-face tunnel boring machines are widely used in tunnel projects such as subways, railways, and highways. To ensure the safe and efficient construction of a full-face tunnel boring machine, it is necessary to adjust the operating parameters of the equipment according to the geological environment. However, it is currently difficult to accurately predict the geological conditions before excavation. Compared with geological condition prediction, the prediction of the operating parameters of a full-face tunnel boring machine is more feasible in actual tunnel construction. The cutterhead torque is an important operating parameter of a full-face tunnel boring machine. Accurately predicting the cutterhead torque helps operators adjust the operating parameters in advance, which is beneficial to avoiding cutterhead jamming and ensuring the safe and efficient tunnel construction.

[0003] Patent document CN112347580A (application number: CN202011220548.9) discloses a real-time prediction method and system for the cutterhead torque of a shield machine, including: selecting the shield machine operating parameters that meet the preset requirements for the influence on the cutterhead torque during the actual operation of the shield machine, and performing preprocessing; establishing a residual CNN-LSTM neural network cutterhead torque prediction model based on the preprocessed operating parameters and training it; evaluating the prediction accuracy of the trained residual CNN-LSTM neural network cutterhead torque prediction model for the cutterhead torque on different data sets.

[0004] Patent document CN113221458A (application number: 202110534801.6) discloses a multi-step prediction method and system for shield cutterhead torque, including: collecting cutterhead torque signals and preprocessing them into a cutterhead torque sequence; using the VMD decomposition method to decompose the cutterhead torque sequence into multiple subsequences and a residual sequence, and further decomposing the residual sequence through the EWT decomposition method; normalizing the torque subsequences and feeding them into an LSTM neural network; constructing a multi-step prediction neural network model for shield cutterhead torque and training it; predicting the cutterhead torque value at a preset future moment; calculating the root mean square error, mean absolute error, and mean absolute percentage error respectively according to the cutterhead torque value at the preset future moment to test the prediction accuracy of the cutterhead torque. This patent directly performs VMD decomposition on the original torque signal and further decomposes the residual sequence through the EWT decomposition method, making it difficult to achieve a relatively fine decomposition of complex-frequency torque signals. However, the present invention performs adaptive hierarchical decomposition on the original torque signal, achieving a relatively fine decomposition of complex-frequency torque signals, greatly reducing the signal complexity and extracting the essential features of the signal, thereby helping to improve the prediction accuracy. Summary of the Invention

[0005] Aiming at the defects in the prior art, the purpose of the present invention is to provide a long-term prediction method and system for the cutterhead torque of a full-face tunnel boring machine.

[0006] According to a long-term prediction method for the cutterhead torque of a full-face tunnel boring machine provided by the present invention, it includes:

[0007] Step S1: Collect the cutterhead torque signal during the tunneling process of the full-face tunnel boring machine and perform preprocessing to obtain a cutterhead torque sequence;

[0008] Step S2: Use a wavelet packet decomposition matrix to decompose the cutterhead torque sequence into a high-frequency cutterhead torque sequence and a low-frequency cutterhead torque sequence;

[0009] Step S3: The low-frequency cutterhead torque sequence is decomposed into several low-frequency subsequences and a residual sequence by an improved adaptive variational mode decomposition method; the high-frequency cutterhead torque sequence is decomposed into several high-frequency subsequences by an empirical wavelet transform method;

[0010] Step S4: Based on a gated recurrent neural network, use the Keras package under the TensorFlow framework to construct a multi-step long-term prediction neural network model for cutterhead torque and train it to obtain a trained multi-step long-term prediction neural network model for cutterhead torque;

[0011] Step S5: Normalize a number of high-frequency subsequences and a number of low-frequency subsequences respectively by using the min-max method, and transmit the normalized high-frequency subsequences and low-frequency subsequences to the trained multi-step long-term prediction neural network model of cutterhead torque to obtain a number of prediction results; add up the number of prediction results to obtain the cutterhead torque value at the predicted time t;

[0012] Step S6: Calculate the mean absolute percentage error, root mean square error and mean absolute error respectively according to the obtained multiple cutterhead torque values, and evaluate the prediction performance of the cutterhead torque;

[0013] The improved adaptive variational mode decomposition method adaptively determines the number of modes and the corresponding center frequency and penalty factor of each mode according to the empty-scale transformation of the signal to be decomposed, and on this basis, performs a more refined variational mode decomposition on the signal to be decomposed.

[0014] Preferably, the step S2 adopts:

[0015] W pc =W p ·T( small

[0016]

[0017]

[0018] where T(t) represents the original torque signal; W p represents the wavelet packet decomposition matrix; W pc represents the coefficient matrix obtained by decomposing the original torque signal T(t) by the wavelet packet decomposition matrix; W p1 represents the first column of W p ; W p2 represents the second column of W p ; W pN represents the Nth column of W p ; W pk represents the kth column of W p ; Γ1 represents the first column of the identity matrix of signal length N; Γ2 represents the second column of the identity matrix of signal length N; Γ k represents the kth column of the identity matrix of signal length N; Γ N represents the Nth column of the identity matrix of signal length N; λ represents the number of layers of wavelet packet decomposition; db6 represents the wavelet basis function Daubechies 6; the functions wpdee() and wpcoef() are the decomposition function and wavelet packet coefficient acquisition function in the matlab wavelet packet toolbox.

[0019] Preferably, the improved adaptive variational mode decomposition method adopts:

[0020] Step S3.1: Obtain the spatial scale transformation representation L(f) of the low-frequency cutterhead torque sequence f(t);

[0021] Step S3.2: Obtain the number of modes K of the low-frequency cutterhead torque sequence according to the spatial scale transformation representation L(f), the central frequency ω corresponding to each mode i, i and the penalty factor α i ;

[0022] Step S3.3: Using the number of modes K, the central frequency ω corresponding to each mode i, i and the penalty factor α i as the initial values, perform empirical mode decomposition on the low-frequency cutterhead torque sequence, so as to decompose the low-frequency cutterhead torque sequence into several low-frequency subsequences and a residual sequence.

[0023] Preferably, the cutterhead torque multi-step long-term prediction neural network model further includes a fully connected neural network, and the prediction result is output through the fully connected neural network.

[0024] Preferably, the step S6 adopts:

[0025]

[0026]

[0027]

[0028] where MAPE represents the mean absolute percentage error; RMSE represents the root mean square error; MAE represents the mean absolute error; X = {x1, x2, …, x n} is the true value; is the predicted value; n is the number of predicted values; j is the number of sequences; the smaller the MAPE value, the higher the prediction accuracy; the smaller the RMSE value, the higher the prediction accuracy; the smaller the MAE value, the higher the prediction accuracy.

[0029] According to a long-term cutterhead torque prediction system for a full-face tunnel boring machine provided by the present invention, it includes:

[0030] Module M1: Collect the cutterhead torque signal during the tunneling process of the full-face tunnel boring machine and perform preprocessing to obtain the cutterhead torque sequence;

[0031] Module M2: Decompose the cutterhead torque sequence into a high-frequency cutterhead torque sequence and a low-frequency cutterhead torque sequence by using a wavelet packet decomposition matrix;

[0032] Module M3: The low-frequency cutterhead torque sequence is decomposed into several low-frequency subsequences and a residual sequence by using an improved adaptive variational mode decomposition method; the high-frequency cutterhead torque sequence is decomposed into several high-frequency subsequences by using an empirical wavelet transform method;

[0033] Module M4: Construct a multi-step long-term prediction neural network model for cutterhead torque using the Keras package under the TensorFlow framework based on a gated recurrent neural network and train it to obtain a trained multi-step long-term prediction neural network model for cutterhead torque;

[0034] Module M5: Normalize a number of high-frequency subsequences and a number of low-frequency subsequences respectively using the min-max method, and transmit the normalized high-frequency subsequences and low-frequency subsequences to the trained multi-step long-term prediction neural network model for cutterhead torque to obtain a number of prediction results; Add the number of prediction results to obtain the cutterhead torque value at the predicted time t;

[0035] Module M6: Calculate the mean absolute percentage error, root mean square error and mean absolute error respectively based on the predicted cutterhead torque values to evaluate the prediction performance of the cutterhead torque;

[0036] The improved adaptive variational mode decomposition method adaptively determines the number of modes, the corresponding central frequency and penalty factor of each mode according to the empty-scale transformation of the signal to be decomposed, and on this basis, performs a more refined variational mode decomposition on the signal to be decomposed.

[0037] Preferably, in the said module M2:

[0038] W pc = W p ·T( small

[0039]

[0040]

[0041] Among them, T(t) represents the original torque signal; W p represents the wavelet packet decomposition matrix; W pc represents the coefficient matrix obtained by decomposing the original torque signal T(t) by the wavelet packet decomposition matrix; W p1 represents the first column of W p ; W p2 represents the second column of W p ; W pN represents the Nth column of W p ; W pk represents the kth column of W p ; Γ1 represents the first column of the identity matrix of signal length N; Γ2 represents the second column of the identity matrix of signal length N; Γ k represents the kth column of the identity matrix of signal length N; Γ NThe Nth column of the identity matrix representing the signal length N; λ represents the number of layers of wavelet packet decomposition; db6 represents the Daubechies 6 wavelet basis function; the functions wpdee() and wpcoef() are the decomposition function and the wavelet packet coefficient acquisition function in the matlab wavelet packet toolbox.

[0042] Preferably, the improved adaptive variational mode decomposition method adopts:

[0043] Module M3.1: Obtain the spatial scale transformation representation L(f) of the low-frequency cutter head torque sequence f(t);

[0044] Module M3.2: Obtain the number of modes K of the low-frequency cutter head torque sequence according to the spatial scale transformation representation L(f), the center frequency ω corresponding to each mode i i and the penalty factor α i ;

[0045] Module M3.3: Using the number of modes K, the center frequency ω corresponding to each mode i i and the penalty factor α i as the initial values, perform empirical mode decomposition on the low-frequency cutter head torque sequence, so as to decompose the low-frequency cutter head torque sequence into several low-frequency subsequences and a residual sequence.

[0046] Preferably, the cutter head torque multi-step long-term prediction neural network model further includes a fully connected neural network, and the prediction result is output through the fully connected neural network.

[0047] Preferably, in the module M6:

[0048]

[0049]

[0050]

[0051] Among them, MAPE represents the mean absolute percentage error; RMSE represents the root mean square error; MAE represents the mean absolute error; X = {x1, x2,..., x n} is the true value; is the predicted value; n is the number of predicted values; j is the number of sequences; the smaller the MAPE value, the higher the prediction accuracy; the smaller the RMSE value, the higher the prediction accuracy; the smaller the MAE value, the higher the prediction accuracy.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. The present invention constructs a wavelet packet decomposition matrix for decomposing the original cutterhead torque sequence into low-frequency and high-frequency parts. The wavelet packet decomposition matrix can be generated and stored in advance, effectively shortening the decomposition time.

[0054] 2. The present invention proposes an improved adaptive variational mode decomposition method, which can perform more refined decomposition on the low-frequency part.

[0055] 3. The present invention performs adaptive hierarchical decomposition on the original torque signal, greatly reducing the signal complexity and extracting the essential features of the signal. It realizes high-precision multi-step long-time prediction of the cutterhead torque of a full-face tunnel boring machine under complex geological and working conditions, helps guide the driver to adjust the operation parameters of the shield machine in advance, realizes efficient and safe propulsion, and thus improves the automation and intelligence level of the full-face tunnel boring machine. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] By reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings, other features, objects, and advantages of the present invention will become more apparent:

[0057] Figure 1 FIG. is the structural diagram of a neural network model for multi-step long-time prediction of cutterhead torque based on adaptive multi-level decomposition proposed by the present invention.

[0058] Figure 2 FIG. is the actual cutterhead torque diagram of the neural network model for multi-step long-time prediction of cutterhead torque based on adaptive multi-level decomposition proposed by the present invention in the test set of dataset 1.

[0059] Figure 3 FIG. is the first-step cutterhead torque prediction diagram of the neural network model for multi-step long-time prediction of cutterhead torque based on adaptive multi-level decomposition proposed by the present invention in the test set of dataset 1. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those of ordinary skill in the art can make several changes and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0061] Aiming at the problems of low prediction accuracy and weak generalization ability existing in the current cutterhead torque prediction methods, the present invention provides a long-time prediction method and system for the cutterhead torque of a full-face tunnel boring machine based on adaptive multi-level decomposition.

[0062] Example 1

[0063] A long-term prediction method for the cutterhead torque of a full-face tunnel boring machine provided by the present invention includes:

[0064] Step S1: Collect the cutterhead torque signal during the tunneling process of the full-face tunnel boring machine and perform preprocessing to obtain the cutterhead torque sequence;

[0065] Step S2: Decompose the cutterhead torque sequence into a high-frequency cutterhead torque sequence and a low-frequency cutterhead torque sequence by using the multi-wavelet packet decomposition matrix (MWPD);

[0066] The specific implementation of step S2 is as follows:

[0067] W pc = W p ·T(t);

[0068]

[0069]

[0070] Among them, T(t) represents the original torque signal; W p represents the multi-wavelet packet decomposition matrix; W pc represents the coefficient matrix obtained by decomposing the original torque signal T(t) through the multi-wavelet packet decomposition matrix; W p1 represents the first column of W p ; W p2 represents the second column of W p ; W pN represents the Nth column of W p ; W pk represents the kth column of W p ; Γ1 represents the first column of the identity matrix with signal length N; Γ2 represents the second column of the identity matrix with signal length N; Γ k represents the kth column of the identity matrix with signal length N; Γ N represents the Nth column of the identity matrix with signal length N; λ represents the number of decomposition layers of the multi-wavelet packet decomposition; db6 represents the Daubechies 6 wavelet basis function; the functions wpdec() and wpcoef() are the decomposition function and the multi-wavelet packet coefficient acquisition function in the matlab multi-wavelet packet toolbox.

[0071] Step S3: Decompose the low-frequency cutterhead torque sequence into several low-frequency subsequences and a residual sequence by using the improved adaptive variational mode decomposition method (IDVMD); decompose the high-frequency cutterhead torque sequence into several high-frequency subsequences by using the empirical wavelet transform method (EWT);

[0072] The specific implementation of the improved adaptive variational mode decomposition method is as follows:

[0073] Step S3.1: Obtain the spatial scale transformation representation L(f) of the low-frequency cutter head torque sequence f(t);

[0074] Step S3.2: Obtain the number of modes K of the low-frequency cutter head torque sequence according to the spatial scale transformation representation L(f), the central frequency ω corresponding to each mode i i and the penalty factor α i ;

[0075] Step S3.3: Using the number of modes K, the central frequency ω corresponding to each mode i i and the penalty factor α i as the initial values, perform empirical mode decomposition (VMD) on the low-frequency cutter head torque sequence, so as to decompose the low-frequency cutter head torque sequence into several low-frequency subsequences and a residual sequence.

[0076] Step S4: Based on the gated recurrent neural network (GRU), use the Keras package under the TensorFlow framework to construct a multi-step long-term prediction neural network model for the cutter head torque and train it to obtain a trained multi-step long-term prediction neural network model for the cutter head torque;

[0077] The multi-step long-term prediction neural network model for the cutter head torque specifically includes: matrix wavelet packet decomposition, improved adaptive variational mode decomposition, EWT decomposition, and GRU neural network; the matrix wavelet packet decomposition decomposes the original cutter head torque sequence into low-frequency and high-frequency parts; the improved adaptive variational mode decomposition performs a more refined decomposition on the low-frequency part with complex change characteristics to obtain subsequences and a residual series with simple change characteristics; the EWT decomposition further decomposes the high-frequency part to reduce its complexity; the gated recurrent neural network includes a preset layer, and each layer of GRU neural network has a preset number of neurons; the GRU neural network includes 3 layers, the number of neurons in the first layer of GRU network is 50, the number of neurons in the second layer of GRU network is 30, and the number of neurons in the third layer of GRU network is 10; use the multi-step long-term prediction neural network model for the cutter head torque to extract the time-varying characteristics of the normalized subsequences; the multi-step long-term prediction neural network model for the cutter head torque also includes a fully connected neural network, and the prediction results are output through the fully connected neural network.

[0078] Step S5: Use the min-max method to normalize several high-frequency subsequences and low-frequency subsequences respectively, and transmit the normalized high-frequency subsequences and low-frequency subsequences to the trained multi-step long-term prediction neural network model for the cutter head torque to obtain several prediction results; add the several prediction results to obtain the cutter head torque value at the predicted time t;

[0079] Step S6: Calculate the mean absolute percentage error, root mean square error, and mean absolute error respectively based on the predicted multiple cutterhead torque values to evaluate the prediction performance of the cutterhead torque;

[0080] The specific implementation of step S6 is as follows:

[0081]

[0082]

[0083]

[0084] Among them, MAPE represents the mean absolute percentage error; RMSE represents the root mean square error; MAE represents the mean absolute error; X = {x1, x2, …, x n} is the true value; is the predicted value; n is the number of predicted values; j is the sequence number; the smaller the MAPE value, the higher the prediction accuracy; the smaller the RMSE value, the higher the prediction accuracy; the smaller the MAE value, the higher the prediction accuracy.

[0085] The improved adaptive variational mode decomposition method adaptively determines the number of modes, the corresponding central frequency and penalty factor of each mode according to the empty-scale transformation of the signal to be decomposed, and on this basis, performs a more refined variational mode decomposition on the signal to be decomposed.

[0086] The multi-step long-term prediction neural network model for cutterhead torque includes matrix wavelet packet decomposition, improved adaptive variational mode decomposition, EWT decomposition, and GRU neural network; the cutterhead torque signal at a preset moment is used to obtain the predicted cutterhead torque value through the multi-step long-term prediction neural network model for cutterhead torque.

[0087] First, the cutterhead torque signal during the on-site construction of the full-face tunnel boring machine is used as the input quantity of the prediction model. Then, a multi-step long-term prediction neural network model for cutterhead torque based on adaptive multi-level decomposition is established, using the cutterhead torque of the previous 10 historical moments as the input of the model and the cutterhead torque values of the next 5 moments as the output. And the data in the cutterhead torque parameter database of the full-face tunnel boring machine during operation is used for model training. The trained model can realize real-time multi-step prediction of the cutterhead torque, so as to guide the driver to adjust the operation parameters of the shield machine in advance and realize the efficient and safe tunneling construction of the full-face tunnel boring machine.

[0088] According to a long-term prediction system for cutterhead torque of a full-face tunnel boring machine provided by the present invention, it includes:

[0089] Module M1: Collect the cutterhead torque signal during the tunneling process of the full-face tunnel boring machine and perform preprocessing to obtain the cutterhead torque sequence;

[0090] Module M2: The cutter head torque sequence is decomposed into a high-frequency cutter head torque sequence and a low-frequency cutter head torque sequence by using the wavelet packet decomposition matrix (MWPD);

[0091] Specifically, Module M2 adopts:

[0092] W pc = W p ·T(r);

[0093]

[0094]

[0095] Among them, T(t) represents the original torque signal; W p represents the wavelet packet decomposition matrix; W pc represents the coefficient matrix obtained by decomposing the original torque signal T(t) through the wavelet packet decomposition matrix; W p1 represents the first column of W p ; W p2 represents the second column of W p ; W pN represents the Nth column of W p ; W pk represents the kth column of W p ; Γ1 represents the first column of the identity matrix with signal length N; Γ2 represents the second column of the identity matrix with signal length N; Γ k represents the kth column of the identity matrix with signal length N; Γ N represents the Nth column of the identity matrix with signal length N; λ represents the number of layers of wavelet packet decomposition; db6 represents the Daubechies 6 wavelet basis function; The functions wpdec() and wpcoef() are the decomposition function and the wavelet packet coefficient acquisition function in the matlab wavelet packet toolbox.

[0096] Module M3: The low-frequency cutter head torque sequence is decomposed into several low-frequency subsequences and a residual sequence by using the improved adaptive variational mode decomposition method (IDVMD); The high-frequency cutter head torque sequence is decomposed into several high-frequency subsequences by using the empirical wavelet transform method (EWT);

[0097] Specifically, the improved adaptive variational mode decomposition method adopts:

[0098] Module M3.1: Obtain the spatial scale transformation representation L(f) of the low-frequency cutter head torque sequence f(t);

[0099] Module M3.2: Obtain the number of modes K of the low-frequency cutter head torque sequence according to the spatial scale transformation representation L(f), and the center frequency ω i and the penalty factor α i ;

[0100] Module M3.3: According to the number of modes K, the central frequency ω corresponding to each mode i i and the penalty factor α i As the initial values, perform empirical mode decomposition (VMD) on the low-frequency cutter head torque sequence, so as to decompose the low-frequency cutter head torque sequence into several low-frequency subsequences and a residual sequence.

[0101] Module M4: Based on the gated recurrent neural network (GRU), use the Keras package under the TensorFlow framework to construct a multi-step long-term prediction neural network model for the cutter head torque and perform training to obtain a trained multi-step long-term prediction neural network model for the cutter head torque;

[0102] The multi-step long-term prediction neural network model for the cutter head torque specifically includes: matrix wavelet packet decomposition, improved adaptive variational mode decomposition, EWT decomposition, and GRU neural network; the matrix wavelet packet decomposition decomposes the original cutter head torque sequence into low-frequency and high-frequency parts; the improved adaptive variational mode decomposition performs a more refined decomposition on the low-frequency part with complex change characteristics to obtain subsequences and a residual series with simple change characteristics; the EWT decomposition further decomposes the high-frequency part to reduce its complexity; the gated recurrent neural network includes a preset layer, and each layer of GRU neural network has a preset number of neurons; the GRU neural network includes 3 layers, the number of neurons in the first layer of GRU network is 50, the number of neurons in the second layer of GRU network is 30, and the number of neurons in the third layer of GRU network is 10; use the multi-step long-term prediction neural network model for the cutter head torque to extract time-varying features of the normalized subsequences; the multi-step long-term prediction neural network model for the cutter head torque also includes a fully connected neural network, and outputs the prediction results through the fully connected neural network.

[0103] Module M5: Use the min-max method to normalize several high-frequency subsequences and low-frequency subsequences respectively, and transmit the normalized high-frequency subsequences and low-frequency subsequences to the trained multi-step long-term prediction neural network model for the cutter head torque to obtain several prediction results; add the several prediction results to obtain the cutter head torque value at the predicted time t;

[0104] Module M6: According to the multiple cutter head torque values obtained by prediction, calculate the mean absolute percentage error, root mean square error, and mean absolute error respectively to evaluate the prediction performance of the cutter head torque;

[0105] The specific adoption of Module M6 is:

[0106]

[0107]

[0108]

[0109] Among them, MAPE represents the mean absolute percentage error; RMSE represents the root mean square error; MAE represents the mean absolute error; X = {x1, x2, …, x n} is the true value; is the predicted value; n is the number of predicted values; j is the sequence number; the smaller the MAPE value, the higher the prediction accuracy; the smaller the RMSE value, the higher the prediction accuracy; the smaller the MAE value, the higher the prediction accuracy.

[0110] The improved adaptive variational mode decomposition method adaptively determines the number of modes, the corresponding central frequency and penalty factor of each mode according to the empty-scale transformation of the signal to be decomposed, and on this basis, performs a more refined variational mode decomposition on the signal to be decomposed.

[0111] The cutter head torque multi-step long-term prediction neural network model includes matrix wavelet packet decomposition, improved adaptive variational mode decomposition, EWT decomposition and GRU neural network; the cutter head torque signal at a preset moment is used to obtain the predicted cutter head torque value through the cutter head torque multi-step long-term prediction neural network model.

[0112] First, the cutter head torque signal from the on-site construction of the full-face tunnel boring machine is used as the input quantity of the prediction model. Then, a cutter head torque multi-step long-term prediction neural network model based on adaptive multi-level decomposition is established, using the cutter head torque of the previous 10 historical moments as the input of the model, and the cutter head torque values of the next 5 moments as the output. And the data in the cutter head torque parameter database of the full-face tunnel boring machine during operation is used to train the model. The trained model can realize real-time multi-step prediction of the cutter head torque, so as to guide the driver to adjust the operation parameters of the shield machine in advance and realize the efficient and safe tunneling construction of the full-face tunnel boring machine.

[0113] Example 2

[0114] Embodiment 2 is a variant of Embodiment 1.

[0115] Reference Figures 1 to 3 , the present invention provides a long-term prediction method for the cutter head torque of a full-face tunnel boring machine based on adaptive multi-level decomposition, including the following steps:

[0116] Step 1: Select the cutter head torque signal in the actual working process of the full-face tunnel boring machine and perform preprocessing, and use the operation parameter data of the shield machine in the previous 10 historical moments to predict the cutter head torque values of the next 5 moments;

[0117] Step 2: Use the wavelet packet decomposition matrix (MWPD) to decompose the cutter head torque sequence into high-frequency and low-frequency parts;

[0118] Step 3: The low-frequency part is further decomposed into several subsequences and a residual sequence by the improved adaptive variational mode decomposition method (IDVMD), while the high-frequency part is further decomposed into several subseries by the empirical wavelet transform (EWT);

[0119] Step 4: Based on the GRU neural network, a multi-step long-term prediction neural network model for the cutterhead torque is constructed using the Keras package under the TensorFlow framework. The GRU neural network consists of 3 layers. The number of neurons in the first layer of the GRU network is 50, the number of neurons in the second layer of the GRU network is 30, and the number of neurons in the third layer of the GRU network is 10, as Figure 1 shown; then it is trained. The learning rate is set to 0.002. The root mean square error (RMSE) and the Adam optimizer are used as the loss function and the optimizer respectively. The number of model training times and the batch size are both set to 100. The training set includes 7000 rows of cutterhead torque data of a full-face tunnel boring machine, and the test set includes 3000 rows of cutterhead torque data of a full-face tunnel boring machine, obtaining the trained multi-step long-term prediction neural network model for the cutterhead torque;

[0120] Step 5: The minimum-maximum method is used to normalize the torque subsequences obtained by the improved adaptive variational mode decomposition and the empirical wavelet transform and send them to the trained multi-step long-term prediction neural network model for the cutterhead torque;

[0121] Step 6: The trained multi-step long-term prediction neural network model for the cutterhead torque is used to predict the cutterhead torque values at the next 5 moments;

[0122] Step 7: The root mean square error, the mean absolute error, and the mean absolute percentage error are calculated respectively according to the cutterhead torque values at several future moments, and the prediction accuracy of the cutterhead torque is tested according to the calculated root mean square error, mean absolute error, and mean absolute percentage error.

[0123] From Figure 2 and Figure 3 it can be seen that the predicted cutterhead torque values by the proposed long-term prediction model for the cutterhead torque of a full-face tunnel boring machine based on adaptive multi-level decomposition are in good agreement with the actual cutterhead torque values. On this dataset, the prediction accuracies of the first to fifth step predictions are 98.898%, 98.347%, 98.049%, 97.799%, and 97.586% respectively, the RMSE (KN·m) are 24.449, 38.282, 46.132, 53.728, and 58.985 respectively, and the MAE (KN·m) are 17.795, 27.260, 32.455, 36.252, and 39.701 respectively, indicating that the proposed long-term prediction method for the cutterhead torque of a full-face tunnel boring machine based on adaptive multi-level decomposition has a high multi-step prediction accuracy.

[0124] Those skilled in the art know that in addition to implementing the systems, devices and their respective modules provided by the present invention in the form of pure computer-readable program code, it is entirely possible to logically program the method steps so that the systems, devices and their respective modules provided by the present invention are implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, etc. to achieve the same program. Therefore, the systems, devices and their respective modules provided by the present invention can be considered as a kind of hardware components, and the modules included therein for implementing various programs can also be regarded as the structures within the hardware components; the modules for implementing various functions can also be regarded as either software programs for implementing the methods or the structures within the hardware components.

[0125] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A long-term prediction method for the cutter head torque of a full-face tunnel boring machine, characterized in that, Including: Step S1: Collect the cutter head torque signal during the tunneling process of the full-face tunnel boring machine and perform preprocessing to obtain the cutter head torque sequence; Step S2: Use the wavelet packet decomposition matrix to decompose the cutter head torque sequence into a high-frequency cutter head torque sequence and a low-frequency cutter head torque sequence; Step S3: The low-frequency cutter head torque sequence is decomposed into several low-frequency subsequences and a residual sequence by the improved adaptive variational mode decomposition method; the high-frequency cutter head torque sequence is decomposed into several high-frequency subsequences by the empirical wavelet transform method; Step S4: Based on the gated recurrent neural network, use the Keras package under the TensorFlow framework to construct a multi-step long-term prediction neural network model for the cutter head torque and perform training to obtain the trained multi-step long-term prediction neural network model for the cutter head torque; Step S5: Normalize several high-frequency subsequences and several low-frequency subsequences respectively by the min-max method, and transmit the normalized several high-frequency subsequences and several low-frequency subsequences to the trained multi-step long-term prediction neural network model for the cutter head torque to obtain several prediction results; add the several prediction results to obtain the cutter head torque value at the predicted time t; Step S6: According to the multiple cutter head torque values obtained by prediction, calculate the mean absolute percentage error, root mean square error, and mean absolute error respectively to evaluate the prediction performance of the cutter head torque; The improved adaptive variational mode decomposition method adaptively determines the number of modes, the corresponding central frequency and penalty factor of each mode according to the empty-scale transformation of the signal to be decomposed, and on this basis, performs a more refined variational mode decomposition on the signal to be decomposed; The Step S2 adopts: Among them, T(t) represents the original torque signal; W p represents the wavelet packet decomposition matrix; W pc represents the coefficient matrix obtained by decomposing the original torque signal T(t) by the wavelet packet decomposition matrix; W p1 represents the first column of W p ; W p2 represents the second column of W p ; W pN represents the Nth column of W p ; W pk represents the kth column of W p ; Γ1 represents the first column of the identity matrix of signal length N; Γ2 represents the second column of the identity matrix of signal length N; Γ k represents the kth column of the identity matrix of signal length N; Γ N represents the Nth column of the identity matrix of signal length N; λ represents the number of layers of wavelet packet decomposition; db6 represents the Daubechies 6 wavelet basis function; the functions wpdec() and wpcoef() are the decomposition function and the wavelet packet coefficient acquisition function in the matlab wavelet packet toolbox; The improved adaptive variational mode decomposition method adopts: Step S3.1: Obtain the space-scale transformation representation L(f) of the low-frequency cutter head torque sequence f(t); Step S3.2: Obtain the number of modes K of the low-frequency cutterhead torque sequence according to the spatial scale transformation representation L(f), the center frequency ω corresponding to each mode i i and the penalty factor α i ; Step S3.3: Based on the number of modes K, the central frequency ω corresponding to each mode i i and the penalty factor α i are used as initial values to perform empirical mode decomposition on the low-frequency cutterhead torque sequence, thereby decomposing the low-frequency cutterhead torque sequence into several low-frequency subsequences and a residual sequence.

2. The full-face tunnel boring machine cutter head torque long-term prediction method according to claim 1, characterized in that The multi-step long-term prediction neural network model for the cutter head torque also includes a fully connected neural network, and the prediction results are output through the fully connected neural network.

3. The full-face tunnel boring machine cutter head torque long-term prediction method according to claim 1, characterized in that, The Step S6 adopts: Among them, MAPE represents the mean absolute percentage error; RMSE represents the root mean square error; MAE represents the mean absolute error; X = {x1, x2, …, x n} is the true value; is the predicted value; n is the number of predicted values; j is the number of sequences; the smaller the MAPE value, the higher the prediction accuracy; the smaller the RMSE value, the higher the prediction accuracy; the smaller the MAE value, the higher the prediction accuracy.

4. A long-term prediction system for the cutter head torque of a full-face tunnel boring machine, characterized in that, Including: Module M1: Collect the cutter head torque signal during the tunneling process of the full-face tunnel boring machine and perform preprocessing to obtain the cutter head torque sequence; Module M2: Use the wavelet packet decomposition matrix to decompose the cutter head torque sequence into a high-frequency cutter head torque sequence and a low-frequency cutter head torque sequence; Module M3: The low-frequency cutter head torque sequence is decomposed into several low-frequency subsequences and a residual sequence by the improved adaptive variational mode decomposition method; the high-frequency cutter head torque sequence is decomposed into several high-frequency subsequences by the empirical wavelet transform method; Module M4: Based on the gated recurrent neural network, use the Keras package under the TensorFlow framework to construct a multi-step long-term prediction neural network model for the cutter head torque and perform training to obtain the trained multi-step long-term prediction neural network model for the cutter head torque; Module M5: Normalize several high-frequency subsequences and several low-frequency subsequences respectively by the min-max method, and transmit the normalized several high-frequency subsequences and several low-frequency subsequences to the trained multi-step long-term prediction neural network model for the cutter head torque to obtain several prediction results; add the several prediction results to obtain the cutter head torque value at the predicted time t; Module M6: Calculate the mean absolute percentage error, root mean square error, and mean absolute error respectively based on multiple cutterhead torque values obtained from the prediction, and evaluate the prediction performance of the cutterhead torque; The improved adaptive variational mode decomposition method adaptively determines the number of modes, the corresponding center frequency and penalty factor of each mode according to the empty-scale transformation of the signal to be decomposed, and performs a more refined variational mode decomposition on the signal to be decomposed on this basis; In the module M2: Among them, T(t) represents the original torque signal; W p represents the wavelet packet decomposition matrix; W pc represents the coefficient matrix obtained by decomposing the original torque signal T(t) through the wavelet packet decomposition matrix; W p1 represents the first column of W p ; W p2 represents the second column of W p ; W pN represents the Nth column of W p ; W pk represents the kth column of W p ; Γ1 represents the first column of the identity matrix of signal length N; Γ2 represents the second column of the identity matrix of signal length N; Γ k represents the kth column of the identity matrix of signal length N; Γ N represents the Nth column of the identity matrix of signal length N; λ represents the number of layers of wavelet packet decomposition; db6 represents the Daubechies 6 wavelet basis function; the functions wpdec() and wpcoef() are the decomposition function and the wavelet packet coefficient acquisition function in the matlab wavelet packet toolbox; The improved adaptive variational mode decomposition method adopts: Module M3.1: Obtain the space-scale transformation representation L(f) of the low-frequency cutterhead torque sequence f(t); Module M3.2: Obtain the number of modes K of the low-frequency cutterhead torque sequence based on the spatial scale-transformed representation L(f), the center frequency ω corresponding to each mode i i and the penalty factor α i ; Module M3.3: According to the number of modes K, the central frequency ω corresponding to each mode i i and the penalty factor α i are used as initial values to perform empirical mode decomposition on the low-frequency cutterhead torque sequence, so as to decompose the low-frequency cutterhead torque sequence into several low-frequency subsequences and a residual sequence.

5. The full-face tunnel boring machine cutter head torque long-term prediction system according to claim 4, characterized in that The multi-step long-term prediction neural network model of the cutterhead torque also includes a fully connected neural network, and the prediction result is output through the fully connected neural network.

6. The full-face tunnel boring machine cutter head torque long-term prediction method according to claim 4, characterized in that In the module M6: Among them, MAPE represents the mean absolute percentage error; RMSE represents the root mean square error; MAE represents the mean absolute error; X = {x1, x2, …, x n} is the true value; is the predicted value; n is the number of predicted values; j is the number of sequences; the smaller the MAPE value, the higher the prediction accuracy; the smaller the RMSE value, the higher the prediction accuracy; the smaller the MAE value, the higher the prediction accuracy.

Images