Open-type TBM tunneling machine tunneling parameter prediction method based on reservoir calculation
Through the method based on reserve pool calculation, the problem of TBM excavation parameter setting depends on driver experience, and rapid and accurate excavation parameter prediction is achieved, construction safety and efficiency are improved, and the development of intelligent excavation technology is promoted.
Patent Information
- Application Number
- CN202510375090.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-27
- Publication Date
- 2025-07-18
AI Technical Summary
Traditional TBM excavation parameter setting relies on driver experience, resulting in inconsistency and lag, making it difficult to respond to changes in geological conditions in real time, affecting construction safety and efficiency.
Using a method based on reserve pool calculation, a reserve pool calculation model is constructed by collecting and denoising the excavation data, and the excavation parameters are adjusted in real time, including initializing the reserve pool state, sparse connection characteristics and training only the output layer weights. Combining Kalman filtering and zero regression algorithms, the calculation complexity and noise impact are reduced.
It realizes fast and accurate prediction of excavation parameters, reduces model training and testing time, improves real-time and safety of construction, reduces machine risk, and improves construction efficiency and accuracy.
Smart Images
Figure CN120337172A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of open TBM intelligent tunneling technology, and in particular to a method for predicting tunneling parameters of an open TBM tunneling machine based on reservoir computing. Background Art
[0002] As the core equipment for tunnel construction, the full-face hard rock tunnel boring machine (TBM) has the advantages of high-efficiency tunneling, good economy, and small environmental impact, and is widely used in various tunnel projects. However, in the traditional construction process, the setting of TBM tunneling parameters (such as cutterhead torque, cutterhead speed, total thrust, and propulsion speed) often depends on the driver's experience judgment. This manual operation mode has significant drawbacks: firstly, the experience differences of different drivers are likely to lead to inconsistent parameter settings, increasing the construction uncertainty; secondly, the manual adjustment has hysteresis and is difficult to respond to the changes in geological conditions in real time, easily causing jamming, abnormal wear of cutters, a decrease in tunneling efficiency, and even safety accidents. In addition, due to the complexity and unpredictability of geological conditions, it is difficult to achieve dynamic optimization of parameters based on experience alone, which seriously restricts the intelligent development of TBM. Therefore, in the tunnel construction of the full-face tunnel boring machine, the adaptive adjustment of TBM tunneling parameters becomes increasingly important.
[0003] In recent years, with the progress of artificial intelligence technology, machine learning algorithms have been introduced into the field of TBM tunneling parameter prediction. In the prior art, time series models such as long short-term memory network (LSTM), gated recurrent unit (GRU), and traditional recurrent neural network (RNN) are used for parameter prediction. However, due to the inherent limitations of traditional machine learning algorithms: some of these algorithms have complex network structures and many network layers, resulting in many model parameters, long training time, increased complexity, and being prone to overfitting. Moreover, in the process of predicting engineering data, they are easily affected by noise data and have a large response time difference, thereby leading to untimely parameter prediction, being unable to guide the project, and causing safety problems. This directly affects the accuracy and real-time performance of TBM tunneling parameter prediction and restricts the practical application of intelligent tunneling technology. Summary of the Invention
[0004] In order to solve the above technical problems, the present invention specifically adopts the following technical solutions.
[0005] Design a method for predicting tunneling parameters of an open TBM tunneling machine based on reservoir computing, including the following steps:
[0006] S1: Collect the tunneling data of the stable section of the TBM, including cutterhead torque, cutterhead speed, total thrust, and propulsion speed;
[0007] S2: Perform noise reduction processing on the tunneling data to generate smooth time series data;
[0008] S3: Divide the smoothed time series data into a training set and a test set, where part of the data in the training set is used to initialize the reservoir state;
[0009] S4: Construct a reservoir computing model and train the model based on the training set, specifically including:
[0010] a) Initialize the network structure parameters of the reservoir, and the network structure includes an input weight matrix, a recurrent weight matrix, and an output weight matrix;
[0011] b) Input the time series data of the training set into the reservoir, and dynamically update the reservoir state through a non-linear activation function;
[0012] c) Use the tunneling parameters at the next moment as the teacher signal of the reservoir computing model, and train the output weight matrix using a linear regression algorithm;
[0013] S5: Input the test set into the trained reservoir computing model, output the predicted values of the tunneling parameters, and adjust the operation parameters of the tunnel boring machine in real time according to the predicted values.
[0014] Preferably, in step S2, the noise reduction process uses the Kalman filter algorithm, with a window length of 21 and a polynomial order of 3.
[0015] Preferably, in step S3, the division ratio of the training set to the test set is 1:1, and the first 1 / 3 of the data in the training set is used to initialize the reservoir state.
[0016] Preferably, in step S4, the network structure parameters of the reservoir include: the number of reservoir nodes is 100, the reservoir connection sparsity is 30%, and the input weight matrix and the recurrent weight matrix are uniformly distributed from 0 to 1, randomly initialized and fixed.
[0017] Preferably, in step S4, the linear regression algorithm is zero regression, and the regularization factor λ is set to 0.1.
[0018] Preferably, in step S4, the non-linear activation function f uses the hyperbolic tangent function tanh; set the parameters of the reservoir computing model: among them, the learning rate of the cutter head torque is 0.01, and the learning rates of the cutter head speed, total thrust, and propulsion speed are 0.1. The update formula of the reservoir state is:
[0019] x t =(1 - α)x t-1 +αf(W in u t +W rec x t-1 ) (1);
[0020] Where, xt is the current reserve pool state, x t-1 is the reserve pool state at the previous moment, u t is the input tunneling parameter at the current moment, W in is the input weight matrix, Wrec is the recurrent weight matrix, α is the learning rate, u t is the test data;
[0021] Then u t+1 serves as the corresponding tunneling parameter for the next moment, and u t+1 serves as the teacher signal of the reserve pool calculation model, and the output weight matrix W out has the following calculation formula:
[0022] W out = UX T (XX T + λI) -1 (2);
[0023] where U is the teacher signal matrix composed of n training teacher signals, X is the reserve pool state matrix, I is the identity matrix, and X T represents the bias operation on the reserve pool state matrix;
[0024] When testing the model, directly input the test tunneling parameter u t in the test set into the already trained model, and obtain the predicted value y t of the test tunneling parameter u t = W out x t (3).
[0025] Preferably, in step S5, the evaluation indexes of the predicted value include the mean square error and the coefficient of determination, and the model training time is less than 0.2 s, and the test time is less than 0.01 s.
[0026] Preferably, the sampling time interval of the tunneling data in step S1 is 2 seconds.
[0027] The beneficial effects of the present invention are as follows:
[0028] 1. The present invention is modeled based on reservoir computing. Through the sparse connection characteristics of reservoir computing and the mechanism of only training the output layer weights, the computational complexity is greatly reduced. The model training time < 0.2 seconds, and the single-step test time < 0.01 seconds, meeting the real-time requirement of ≤ 2 seconds in TBM construction, and can quickly guide the adjustment of tunneling parameters to avoid jamming or safety accidents caused by response lag.
[0029] 2. In view of the problem that the existing model is vulnerable to noise interference, the present invention uses the Kalman filter algorithm to denoise the original data, and combines the random fixed weight design of the reservoir to effectively suppress the influence of noise, significantly improving the stability of the model in a complex engineering environment.
[0030] 3. The present invention fixes the input weight matrix and the recurrent weight matrix, and only trains the output matrix. At the same time, the zero regression algorithm (regularization factor λ = 0.1) is introduced to constrain the output weights, further reducing the risk of overfitting.
[0031] 4. The present invention separately establishes independent prediction models for four parameters, namely cutter head torque, rotational speed, total thrust and propulsion speed, and uses the maximum-minimum normalization and differential learning rate to achieve parameter-specific optimization. According to the evaluation index data of the predicted values of the four parameters of cutter head torque, rotational speed, total thrust and propulsion speed, the mean squared error of the total thrust prediction is as low as 0.0109, and the MSE (Mean Squared Error) of the cutter head rotational speed is 6.4037e-04, and the accuracy is significantly better than the traditional empirical judgment.
[0032] In summary, the present invention can quickly realize guiding the shield main driver to correctly set the tunneling parameters at the next moment, avoid safety problems caused by the wrong or incorrect setting of the main driver, and accurate prediction can achieve high-quality and high-efficiency tunneling, improve construction efficiency, increase revenue, save costs, and at the same time have the advantages of strong generalization ability, being able to adapt to different types of input data and noise environments. Therefore, during the project construction process, it can adapt to different project sites, reduce the influence of noise data on the results, more accurately learn the change relationship between data, improve the prediction accuracy of tunneling parameters, not only can guide the main driver to tunnel at the project site, but also make a technical preparation for unmanned tunneling, promoting the development of the domestic intelligent tunneling field. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 is a schematic diagram of the overall process of the present invention;
[0034] Figure 2 is a schematic diagram of the training process of the reservoir computing model in the present invention;
[0035] Figure 3 is the filtering effect diagram of the cutter head torque;
[0036] Figure 4 is the filtering effect diagram of the total thrust;
[0037] Figure 5 is the trend diagram of the test results of the cutter head torque;
[0038] Figure 6 Trend diagram of the test results of the total thrust;
[0039] Figure 7 Trend chart of the advancing speed test results;
[0040] Figure 8 Trend chart of the cutter head rotation speed test results; Specific implementation manners
[0041] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.
[0042] Embodiment 1
[0043] A prediction method for the tunneling parameters of an open TBM tunneling machine based on reservoir computing, as Figures 1 to 2 shown, includes the following steps:
[0044] S1: Collect the tunneling data of the stable section of the TBM, including cutter head torque, cutter head rotation speed, total thrust and advancing speed; the sampling time interval is 2 seconds.
[0045] The data acquisition device includes the TBM built-in sensor and the data recording system to ensure the continuity and integrity of the data.
[0046] S2: Perform noise reduction processing on the tunneling data by using the Kalman filter algorithm to generate smooth time series data, the window length of which is 21 and the polynomial order is 3. The filtering effect is as Figure 3 and Figure 4 shown. It can be seen from the figure that the noise is significantly reduced and the data trend is clear.
[0047] S3: Divide the smooth time series data into a training set and a test set according to a ratio of 1:1, and 1 / 3 of the data in the training set is used to initialize the reservoir state to avoid the influence of the initial transient on the model training.
[0048] S4: Build a reservoir computing model and train the model based on the training set, specifically including: initializing the network structure parameters of the reservoir, the network structure parameters of the reservoir including: the number of reservoir nodes is 100, the reservoir connection sparsity is 30%, the input weight matrix and the recurrent weight matrix are uniformly distributed from 0 to 1, randomly initialized and fixed; the network structure includes an input weight matrix, a recurrent weight matrix, and an output weight matrix;
[0049] Normalize each tunneling parameter by using the maximum-minimum method and map it to the interval [0, 1].
[0050] b) Input the time series data of the training set into the reservoir and dynamically update the reservoir state through the non-linear activation function f; the non-linear activation function f uses the hyperbolic tangent function tanh;
[0051] c) Use the tunneling parameters at the next moment as the teacher signal of the reservoir computing model, and train the output weight matrix using the linear regression algorithm; the linear regression algorithm is zero regression, and the regularization factor λ is set to 0.1;
[0052] Set the parameters of the reservoir computing model: among them, the learning rate of the cutterhead torque is 0.01, and the learning rates of the cutterhead rotation speed, total thrust, and propulsion speed are 0.1. The update formula for the reservoir state is:
[0053] x t = (1 - α)x t-1 + αf(W in u t + W rec x t-1 ) (1);
[0054] where, x t is the current reservoir state, x t-1 is the reservoir state at the previous moment, u t is the tunneling parameter input at the current moment. Then, u t+1 is used as the corresponding tunneling parameter at the next moment, and u t+1 is used as the teacher signal of the reservoir computing model. W in is the input weight matrix, Wrec is the recurrent weight matrix, α is the learning rate, and u t is the test data;
[0055] The calculation formula for the output weight matrix W out is:
[0056] W out = UX T (XX T + λI) -1 (2);
[0057] where, U is the teacher signal matrix composed of n training teacher signals, X is the reservoir state matrix, I is the identity matrix, and X T represents the bias operation on the reservoir state matrix;
[0058] S5: Input the test set into the trained reservoir computing model, and predict the tunneling parameters at the next moment point by point in chronological order. The comparison between the prediction result and the true value is as Figures 5 to 8 shown. Among them, the black in the figure is the predicted value, and the green is the actual value; the evaluation indicators of the predicted value include the mean square error and the coefficient of determination. And the model training time is less than 0.2s, and the test time is less than 0.01s, meeting the real-time requirements of TBM applications, which can greatly reduce the prediction time and guide the operation of the main driver. The specific evaluation indicators are as follows:
[0059] Cutter head torque: Mean square error MSE = 0.0290, coefficient of determination R 2 = 0.0305;
[0060] Total thrust: Mean square error MSE = 0.0109, R 2 = 0.7471;
[0061] Advance speed: Mean square error MSE = 0.0225, R 2 = 0.0283;
[0062] Cutter head rotation speed: Mean square error MSE = 6.4037e-04, R 2 = 0.8312;
[0063] When testing the model, directly input the test tunneling parameters u t in the test set into the already trained model, and obtain the predicted value y of the test tunneling parameter u t according to the above formulas (1) and (2) t = W out x t (3);
[0064] By comparing the predicted value with the teacher signal (true value), it can be obtained that the change trend of the predicted values of cutter head torque, cutter head rotation speed, total thrust and advance speed is the same as that of the true value. However, in the actual construction of TBM, if the predicted trend is different from the actual trend, it will lead to disasters such as TBM jamming and even safety accidents such as ground settlement.
[0065] Specific process: First, assume that 200 time points u t (t = 1...200) are input to form the internal state of the reservoir. After 200 points, start collecting the reservoir state x t , and form the reservoir state matrix X and the output teacher signal matrix U, and train the output weight matrix W out . In the test stage, input the tunneling data and directly obtain the prediction result. Based on the time series prediction calculated by the reservoir, the fast prediction of the time series can be realized, greatly reducing the training operation time and meeting the real-time requirements. In this embodiment, 4 TBM tunneling parameters are used as the input of the reservoir calculation model respectively to perform independent prediction of the 4 tunneling parameters.
[0066] In specific engineering applications, the prediction results are transmitted to the TBM control system in real time to guide the main driver to dynamically adjust parameters such as cutter head torque and rotation speed. The prediction system interface displays the trend comparison between the actual value and the predicted value (such as Figures 4 - 7 ), and an audible and visual alarm can be triggered when the deviation exceeds the threshold.
[0067] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features; any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.
Claims
1. An open TBM tunneling machine tunneling parameter prediction method based on reservoir computing, characterized in that It includes the following steps: S1: Collect the tunneling data of the stable section of the TBM, including cutterhead torque, cutterhead rotation speed, total thrust, and propulsion speed; S2: Perform noise reduction processing on the tunneling data to generate smooth time series data; S3: Divide the smooth time series data into a training set and a test set, and a part of the data in the training set is used to initialize the reservoir state; S4: Construct a reservoir computing model and train the model based on the training set; S5: Input the test set into the trained reservoir computing model, output the predicted values of the tunneling parameters, and adjust the operation parameters of the tunnel boring machine in real time according to the predicted values.
2. The prediction method for the tunneling parameters of an open TBM tunneling machine based on reservoir computing according to claim 1, wherein: In step S2, the noise reduction processing uses the Kalman filter algorithm, with a window length of 21 and a polynomial order of 3.
3. The prediction method of the tunneling parameters of the open TBM tunneling machine based on reservoir computing according to claim 1, wherein: In step S3, the division ratio of the training set to the test set is 1:1, and the first 1 / 3 of the data in the training set is used to initialize the reservoir state.
4. For the method for predicting tunneling parameters of an open TBM based on reservoir computing as described in claim 1, when training the reservoir computing model in step S4, it specifically includes the following steps: S1: Initialize the network structure parameters of the reservoir. The network structure includes an input weight matrix, a recurrent weight matrix, and an output weight matrix; S2: Input the time series data of the training set into the reservoir, and dynamically update the reservoir state through a non-linear activation function; S3: Use the tunneling parameters at the next moment as the teacher signal of the reservoir computing model, and train the output weight matrix using the linear regression algorithm.
5. The prediction method of the tunneling parameters of the open TBM tunneling machine based on reservoir computing according to claim 4, characterized in that: The network structure parameters of the reservoir include: the number of reservoir nodes is 100, the reservoir connection sparsity is 30%, and the input weight matrix and the recurrent weight matrix are uniformly distributed from 0 to 1, randomly initialized and fixed.
6. The prediction method for the tunneling parameters of the open TBM tunneling machine based on reservoir computing according to claim 4, characterized in that: The linear regression algorithm is zero regression, and the regularization factor λ takes a value of 0.
1.
7. The prediction method for the tunneling parameters of the open TBM tunneling machine based on reservoir computing according to claim 6, wherein: The non-linear activation function f uses the hyperbolic tangent function tanh; set the parameters of the reservoir computing model: among them, the learning rate of the cutterhead torque is 0.01, and the learning rates of the cutterhead rotation speed, total thrust, and propulsion speed are 0.
1. The update formula of the reservoir state is: x t = (1 - α)x t-1 + αf(W in u t + W rec x t-1 ); where x t is the current reserve pool state, x t-1 is the reserve pool state at the previous moment, u t is the input tunneling parameter at the current moment, W in is the input weight matrix, Wrec is the recurrent weight matrix, α is the learning rate, u t is the test data; u t+1 As the tunneling parameters corresponding to the next moment, let u t+1 be the teacher signal of the reservoir computing model, and obtain the output weight matrix W out The calculation formula of which is: W out = UX T (XX T + λI) -1 ; where \(U\) is a teacher signal matrix composed of \(n\) training teacher signals, \(X\) is a reservoir state matrix, \(I\) is an identity matrix, and \(X\) T represents a bias operation on the reservoir state matrix; When testing the model, directly input the test tunneling parameters u in the test set t to the model that has been trained, and obtain the predicted value y of the test tunneling parameters u t t = W out x t . 8. The prediction method for the tunneling parameters of an open TBM tunneling machine based on reservoir computing according to claim 1, wherein: In step S5, the evaluation indexes of the predicted values include the mean square error and the coefficient of determination, and the model training time is less than 0.2 s, and the test time is less than 0.01 s.
9. The prediction method for the tunneling parameters of an open TBM tunneling machine based on reservoir computing according to claim 1, wherein In step S1, the sampling time interval of the tunneling data is 2 seconds.
10. A training method for the reservoir computing model in the prediction method of tunneling parameters of the open TBM tunneling machine as described in claim 1, characterized in that, It specifically includes the following steps: S1: Initialize the network structure parameters of the reservoir. The network structure includes an input weight matrix, a recurrent weight matrix, and an output weight matrix; S2: Input the time series data of the training set into the reservoir, and dynamically update the reservoir state through a non-linear activation function; S3: Use the tunneling parameters at the next moment as the teacher signal of the reservoir computing model, and train the output weight matrix using the linear regression algorithm; S4: The model training is completed.