TBM pushing speed multi-step real-time prediction method

Abstract

The present application relates to a kind of TBM propelling speed multi-step real-time prediction method, the method comprises: S1, the feature data of pre-set time step before current time point is obtained;The feature data includes the data of the TBM feature with the grey correlation degree value of propelling speed greater than 0.5;S2, the feature data of the pre-set time step is input into the multi-step real-time prediction model pre-trained, and the multi-step real-time prediction model pre-trained outputs the predicted value of propelling speed in the second time period after current time point;The multi-step real-time prediction model includes: TCN model, first SENet, second SENet and fully connected layer;Wherein, the multi-step real-time prediction model established by pre-training training feature data set is obtained by pre-training the multi-step real-time prediction model.

Description

Technical Field

[0001] This invention relates to the field of TBM technology, and in particular to a method for multi-step real-time prediction of TBM propulsion speed. Background Technology

[0002] In interactions with surrounding rock, construction efficiency and cost depend on the cutterhead's load-bearing capacity and breaking efficiency. However, accurate advance speed setting is a prerequisite and key to both cutterhead load-bearing capacity and rock-breaking efficiency. Especially in highly abrasive hard rock formations, cutterheads are prone to abnormal wear, thus reducing TBM construction progress. To improve the efficiency of operator-set advance speed and achieve dynamic adjustments, multi-step real-time dynamic prediction is necessary. Although most existing studies provide suggested values ​​for advance speed (for the stable phase of the tunneling cycle), the dynamic adjustment process is not described. The advance speed setting is based on continuous adaptive changes to the geological conditions in front of the tunnel face. Since the advance speed is dynamically set according to geological conditions, dynamic multi-step real-time prediction is considered.

[0003] Existing models for predicting propulsion speed have many hidden layers, which increases the number of optimization parameters, the optimization time, and the overall time cost. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a multi-step real-time prediction method for TBM propulsion speed, which solves the technical problems of existing propulsion speed prediction models having many hidden layers, increasing optimization parameters, increasing model optimization time, and increasing time costs.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0008] This invention provides a method for multi-step real-time prediction of TBM propulsion speed, the method comprising:

[0009] S1. Obtain feature data of a preset time step before the current time point;

[0010] The feature data includes TBM feature data with a gray correlation value greater than 0.5 with propulsion speed;

[0011] S2. Input the feature data of the preset time step into the pre-trained multi-step real-time prediction model, and the pre-trained multi-step real-time prediction model outputs the predicted value of the propulsion speed in the second time period after the current time point.

[0012] The multi-step real-time prediction model includes: a TCN model, a first SENet, a second SENet, and a fully connected layer;

[0013] Specifically, the multi-step real-time prediction model is obtained by training a multi-step real-time prediction model using a pre-trained feature dataset.

[0014] Preferably,

[0015] The feature data of the preset time step includes: the cutterhead torque, total propulsion force, cutterhead power, and propulsion speed of the TBM within the preset time step.

[0016] Preferably,

[0017] The preset time step is 30-33 seconds.

[0018] Preferably, the process further includes the following step before S1:

[0019] S0. Obtain the pre-trained multi-step real-time prediction model by training a multi-step real-time prediction model using the trained feature dataset.

[0020] The training feature dataset includes feature data at multiple preset time steps within a historical time period that have been acquired in advance.

[0021] Preferably, S0 specifically includes:

[0022] S01. The feature data of each preset time step in the trained feature dataset is input into the TCN model in the established multi-step real-time prediction model for processing to obtain the processing result. The TCN model inputs the processing result into the first SENet.

[0023] S02. The first SENet processes the processing result to obtain a first operation result, and inputs the first operation result into the second SENet;

[0024] S03. The second SENet processes the result of the first operation to obtain the second operation result;

[0025] S04. For the feature data of the preset time step in the trained feature dataset, the first operation result and the second operation result are calculated using a preset ReLU function to obtain the calculation result;

[0026] S05. The calculation result is input into the fully connected layer, and the fully connected layer outputs the predicted value of the propulsion speed corresponding to the feature data of the preset time step in the trained feature dataset.

[0027] S06. Repeat S01-S05 a preset number of times to obtain the trained multi-step real-time prediction model.

[0028] Preferably,

[0029] The processing result is multiple feature maps corresponding to the feature data of the preset time step.

[0030] Preferably, S02 specifically includes:

[0031] S021. The first SENet uses formula (1) to compress each feature map through global average pooling to obtain the scalar corresponding to the feature map;

[0032] Formula (1) is as follows:

[0033]

[0034] z c This refers to the feature map u C The corresponding scalar;

[0035] Among them, feature map u C It is a feature map with a length of T corresponding to the feature data of the preset time step;

[0036] T is the preset time step;

[0037] i represents the i-th second in the preset time step;

[0038] S022, based on the scalar corresponding to each feature map, the weight corresponding to each feature map is calculated using formula (2);

[0039] Formula (2) is:

[0040] s=σ(W2δ(W1z));

[0041] in,

[0042] s C For feature map u c The corresponding weights;

[0043]

[0044] Where W1 and W2 are randomly generated weight matrices;

[0045]

[0046] Where r is a pre-set empirical value;

[0047] σ is the Sigmoid function; where

[0048] δ is the ReLU activation function; δ(x) = max(0, x);

[0049] S023. Based on the weights corresponding to each feature map, the first operation result is obtained using formula (3);

[0050] Formula (3) is as follows:

[0051]

[0052] In the formula,

[0053] The result of the first operation is

[0054]

[0055] For feature map u c The weighted feature map.

[0056] Preferably,

[0057] r equals 16.

[0058] Preferably, S03 specifically includes:

[0059] S031. The second SENet uses formula (4) to compress the weighted feature map of each feature map through global average pooling to obtain the first value corresponding to the weighted feature map.

[0060] Formula (4) is as follows:

[0061]

[0062] A c This refers to the weighted feature map. The corresponding value;

[0063] S032, based on the first value corresponding to each weighted feature map, the coefficient corresponding to each weighted feature map is calculated using formula (5);

[0064] Formula (5) is as follows:

[0065] k = σ(W3δ(W4A));

[0066] Where k = [k1, k2, ... k C ];

[0067] k C Weighted feature map The corresponding coefficient;

[0068] A = [A1, A2, ..., A C ];

[0069] Where W3 and W4 are randomly generated weight matrices;

[0070] S033. Based on the coefficients corresponding to each weighted feature map, the second operation result is obtained using formula (6);

[0071] Formula (6) is as follows:

[0072]

[0073] The result of the second operation is B;

[0074] B = [B1, B2, ... B] C ].

[0075] Preferably,

[0076] The preset number of times is 200.

[0077] (III) Beneficial Effects

[0078] The beneficial effects of this invention are as follows: The TBM propulsion speed multi-step real-time prediction method of this invention, by inputting feature data of a preset time step into a pre-trained multi-step real-time prediction model, outputs the predicted propulsion speed for a second time period after the current time point. The multi-step real-time prediction model includes a TCN model, a first SENet, a second SENet, and a fully connected layer. Compared with existing technologies, it can reduce the hidden layers of the model through the first SENet and the second SENet, thereby reducing the optimization time and improving efficiency. Attached Figure Description

[0079] Figure 1 This is a flowchart of a TBM propulsion speed multi-step real-time prediction method according to the present invention;

[0080] Figure 2 This is a schematic diagram of the multi-step real-time prediction model structure in an embodiment of the present invention. Detailed Implementation

[0081] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0082] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0083] The dynamic multi-step real-time prediction of advance speed takes into account the temporal characteristics of TBM construction data. Since the data dispersion of advance speed is also relatively large, enhancing the correlation of TBM construction data is key to improving model performance. To capture longer TBM input data lengths for multi-step prediction, a temporal convolutional neural network is considered. To further enhance the correlation between TBM construction data, a compressed excitation network is also considered. The multi-step real-time prediction of advance speed predicts the subsequent trend of advance speed based on the data just excavated within the same tunneling cycle; therefore, the results of the model predicting advance speed under different prediction steps are further explored. In addition to considering advance speed as an input feature, features affecting advance speed prediction are further optimized, and the adaptability of the multi-step real-time prediction model of advance speed established based on excavated projects is analyzed for transfer to other projects.

[0084] See Figure 1 This embodiment provides a method for multi-step real-time prediction of TBM propulsion speed, the method comprising:

[0085] S1. Obtain feature data of the preset time step before the current time point.

[0086] The feature data includes TBM feature data with a gray correlation value greater than 0.5 with propulsion speed.

[0087] The feature data of the preset time step includes: the cutterhead torque, total propulsion force, cutterhead power, and propulsion speed of the TBM within the preset time step.

[0088] The preset time step is 30-33 seconds.

[0089] In another implementation of this embodiment, the time step is 30 seconds.

[0090] S2. Input the feature data of the preset time step into the pre-trained multi-step real-time prediction model, and the pre-trained multi-step real-time prediction model outputs the predicted value of the propulsion speed in the second time period after the current time point.

[0091] See Figure 2 The multi-step real-time prediction model includes: a TCN model, a first SENet, a second SENet, and a fully connected layer.

[0092] Specifically, the multi-step real-time prediction model is obtained by training a multi-step real-time prediction model using a pre-trained feature dataset.

[0093] In this embodiment, the method further includes the following step before S1:

[0094] S0. Obtain the pre-trained multi-step real-time prediction model by training a multi-step real-time prediction model using the trained feature dataset.

[0095] The training feature dataset includes feature data at multiple preset time steps within a historical time period that have been acquired in advance.

[0096] Specifically, S0 includes:

[0097] S01. The feature data of each preset time step in the trained feature dataset is input into the TCN model in the established multi-step real-time prediction model for processing to obtain the processing result. The TCN model inputs the processing result into the first SENet.

[0098] The processing result is multiple feature maps corresponding to the feature data of the preset time step.

[0099] S02. The first SENet processes the processing result to obtain a first operation result, and inputs the first operation result into the second SENet.

[0100] Specifically, S02 includes:

[0101] S021. The first SENet uses formula (1) to compress each feature map by global average pooling to obtain the scalar corresponding to the feature map.

[0102] Formula (1) is as follows:

[0103]

[0104] z c This refers to the feature map u c The corresponding scalar.

[0105] Among them, feature map u c It is a feature map with a length of T corresponding to the feature data of the preset time step.

[0106] T is the preset time step.

[0107] i represents the i-th second in the preset time step.

[0108] S022, based on the scalar corresponding to each feature map, the weight corresponding to each feature map is calculated using formula (2).

[0109] Formula (2) is:

[0110] s=σ(W2δ(W1z));

[0111] in,

[0112] s C For feature map u c The corresponding weights.

[0113]

[0114] W1 and W2 are randomly generated weight matrices.

[0115]

[0116] Where r is a pre-set empirical value. r equals 16.

[0117] σ is the Sigmoid function; where

[0118] δ is the ReLU activation function; δ(x) = max(0, x).

[0119] S023. Based on the weights corresponding to each feature map, the first operation result is obtained using formula (3).

[0120] Formula (3) is as follows:

[0121]

[0122] In the formula,

[0123] The result of the first operation is

[0124]

[0125] For feature map u c The weighted feature map.

[0126] S03. The second SENet processes the result of the first operation to obtain the second operation result.

[0127] S04. For the feature data of the preset time step in the trained feature dataset, the first operation result and the second operation result are calculated using a preset ReLU function to obtain the calculation result.

[0128] S05. The calculation result is input into the fully connected layer, and the fully connected layer outputs the predicted value of the propulsion speed corresponding to the feature data of the preset time step in the trained feature dataset.

[0129] S06. Repeat S01-S05 a preset number of times to obtain the trained multi-step real-time prediction model. The preset number of times is 200.

[0130] Specifically, S03 includes:

[0131] S031. The second SENet uses formula (4) to compress the weighted feature map of each feature map through global average pooling to obtain the first value corresponding to the weighted feature map.

[0132] Formula (4) is as follows:

[0133]

[0134] A c This refers to the weighted feature map. The corresponding numerical value.

[0135] S032, based on the first value corresponding to each weighted feature map, the coefficient corresponding to each weighted feature map is calculated using formula (5).

[0136] Formula (5) is as follows:

[0137] k = σ(W3δ(W4A)).

[0138] Where k = [k1, k2, ... k C ].

[0139] k C Weighted feature map The corresponding coefficient.

[0140] A = [A1, A2, ..., A C ].

[0141] W3 and W4 are randomly generated weight matrices.

[0142] S033. Based on the coefficients corresponding to each weighted feature map, the second operation result is obtained using formula (6).

[0143] Formula (6) is as follows:

[0144]

[0145] The result of the second operation is B.

[0146] B = [B1, B2, ... B] C ].

[0147] This invention proposes a multi-step real-time prediction model for propulsion velocity prediction. It improves SENet to be suitable for processing time-series data, analyzes the input features that affect propulsion velocity prediction, obtains the impact of different historical time steps on model performance, and verifies the model's performance based on a new project.

[0148] Multi-step real-time prediction of propulsion speed is a regression problem, and mean absolute error (MAE) and mean square error (MSE) are used as evaluation metrics for the multi-step real-time prediction model.

[0149] The smaller the values ​​of MAE and MSE, the better the performance of the multi-step real-time prediction model.

[0150] For computational convenience, MSE is used as the fitness function to optimize the hyperparameters of the proposed model. The formulas for calculating MAE and MSE are shown below:

[0151]

[0152] In the formula, n is the number of samples, y i For the true value, It is a predicted value. It is the average of the true values.

[0153] The ascending and stabilizing phases are valuable data for the tunneling cycle. Predicting the advance speed of the ascending and stabilizing phases can ensure stable tunneling in the current tunneling cycle.

[0154] Most tunneling cycles have an effective construction time of more than 800 seconds, and machine learning models typically require an input of equal length. Considering these factors, we selected the first 800 seconds of data from the total time of the ascending and stabilizing phases of each tunneling cycle to explore the feasibility of multi-step real-time prediction of advance speed.

[0155] Historical time steps prior to the propagation speed are considered to predict future time steps. Therefore, propagation speed is chosen as the input feature.

[0156] The relationship between cutterhead torque, total propulsion force, cutterhead power, and propulsion speed was evaluated using grey relational analysis.

[0157] The higher the grey relational value, the more favorable it is for predicting the acceleration speed. A grey relational value greater than 0.5 for the above features indicates a strong relationship between the feature and the prediction target.

[0158] The proposed model employs an asynchronous, parallel, continuous halving algorithm for hyperparameter optimization. This algorithm simultaneously evaluates all hyperparameter combinations that can proceed to the next round within the current round. This strategy fully utilizes computational resources and significantly improves computational efficiency.

[0159] Based on the optimal hyperparameters, the number of training iterations for all models was set to 200, the batch size to 100, and the loss function to be mean squared error (MSE). To prevent overfitting (which typically occurs when the training set performs well but the test set performs poorly), early stopping was used in the model training process. That is, if the loss function has not decreased after 20 training iterations, the training was terminated.

[0160] The selected TBM construction data comprises 200 tunneling cycles, which are divided into training, validation, and test sets. The proportions of the training, validation, and test sets are 60%, 20%, and 20%, respectively. The training set is used to debug the neural network, the validation set is used to analyze the results of the optimized model, and the test set is used to evaluate the model's performance in predicting propulsion speed. The test set contains a total of 40 cycles.

[0161] The performance of different deep learning models is compared, and the results are shown in Table 1. Overall, for models based on recurrent neural networks (RNNs), gated recurrent unit networks (GRNNs) perform best, followed by long short-term memory (LSTM) networks. RNNs exhibit the worst generalization ability, especially for 9-step prediction, with an MSE of 36.703 and a MAE of 4.412. The TCN-SENet++ model (the multi-step real-time prediction model in this embodiment) incorporates a residual structure, enabling cross-layer information transfer, improving the generalization ability of temporal convolutional neural networks, and allowing different weights to be assigned to different feature maps, thereby enhancing model performance.

[0162] Table 1

[0163]

[0164] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0165] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0166] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0167] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0168] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0169] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

Claims

1. A TBM pushing speed multi-step real-time prediction method, characterized in that, The method includes: S1. Obtain feature data of a preset time step before the current time point; The feature data includes TBM feature data with a gray correlation value greater than 0.5 with propulsion speed; S2. Input the feature data of the preset time step into the pre-trained multi-step real-time prediction model, and the pre-trained multi-step real-time prediction model outputs the predicted value of the propulsion speed in the second time period after the current time point. The multi-step real-time prediction model includes: a TCN model, a first SENet, a second SENet, and a fully connected layer; Among them, the multi-step real-time prediction model is obtained by training a multi-step real-time prediction model in advance using a trained feature dataset; Before S1, it also includes: S0, obtaining the pre-trained multi-step real-time prediction model by training a multi-step real-time prediction model using the trained feature dataset; The training feature dataset includes feature data at multiple preset time steps within a pre-acquired historical time period. Specifically, S0 includes: S01. Input the feature data of each preset time step in the trained feature dataset into the TCN model in the established multi-step real-time prediction model for processing and obtain the processing result. The TCN model inputs the processing result into the first SENet. S02. The first SENet processes the processing result to obtain a first operation result, and inputs the first operation result into the second SENet; S03. The second SENet processes the result of the first operation to obtain the second operation result; S04. For the feature data of the preset time step, the first operation result and the second operation result in the trained feature dataset, feature fusion is performed to obtain fused features, and the fused features are calculated using a preset ReLU function to obtain the calculation result; S05. The calculation result is input into the fully connected layer, and the fully connected layer outputs the predicted value of the propulsion speed corresponding to the feature data of the preset time step in the trained feature dataset. S06. Repeat S01-S05 a preset number of times to obtain the trained multi-step real-time prediction model.

2. The method according to claim 1, characterized in that, The feature data of the preset time step includes: the cutterhead torque, total propulsion force, cutterhead power, and propulsion speed of the TBM within the preset time step.

3. The method according to claim 2, characterized in that, The preset time step is 30-33 seconds.

4. The method according to claim 3, characterized in that, The processing result is multiple feature maps corresponding to the feature data of the preset time step.

5. The method of claim 4, wherein, Specifically, S02 includes: S021. The first SENet uses formula (1) to compress each feature map through global average pooling to obtain the scalar corresponding to the feature map; Wherein, formula (1) is: ； z c refers to a feature map corresponding scalar; wherein the feature map is a Cth feature map of length T corresponding to the feature data of the preset time step. T is the preset time step; i represents the i-th second in the preset time step; S022, based on the scalar corresponding to each feature map, the weight corresponding to each feature map is calculated using formula (2); Formula (2) is: ； wherein ; characteristic map corresponding weights ； in and These are randomly generated weight matrices; ； ； Where r is a pre-set empirical value; is a Sigmoid function; where ; is a ReLU activation function; ; S023. Based on the weights corresponding to each feature map, the first operation result is obtained using formula (3); Formula (3) is as follows: ； In the formulae, ; wherein the first operation result is ; ； is a weighted feature map of the feature map.

6. The method according to claim 5, characterized in that, r equals 16.

7. The method of claim 6, wherein, Specifically, S03 includes: S031. The second SENet uses formula (4) to compress the weighted feature map of each feature map through global average pooling to obtain the first value corresponding to the weighted feature map. Formula (4) is as follows: ； A c refers to a weighted feature map corresponding numerical value; S032, based on the first value corresponding to each weighted feature map, the coefficient corresponding to each weighted feature map is calculated using formula (5); Formula (5) is as follows: ； wherein k = [k1, k2,... k C ] ; k C Weighted feature map The corresponding coefficient; A=[A1,A2,...,A C ] ; Where W3 and W4 are randomly generated weight matrices; S033. Based on the coefficients corresponding to each weighted feature map, the second operation result is obtained using formula (6); Formula (6) is as follows: ； The result of the second operation is B; B = [B1, B2,... B C ].

8. The method according to claim 7, characterized in that, The preset number of times is 200.

Images