A real-time lithology prediction method in tunnel boring machine tunneling process

Abstract

The present application relates to the technical field of intelligent control of tunnel boring machine, and particularly relates to a lithology real-time prediction method in the tunneling process of tunnel boring machine, comprising: detecting effective data in the tunneling process of tunnel boring machine; inputting the effective data into a lithology prediction model to obtain an output lithology prediction result; during the process, dividing the effective data into at least one feature block according to the feature block size; obtaining a time domain attention matrix, a frequency domain attention matrix, a frequency domain feature and a time domain feature corresponding to the feature block according to the set LSTM network and the mixed attention mechanism; obtaining the lithology prediction result based on the time domain attention matrix, the frequency domain attention matrix, the frequency domain feature and the time domain feature; and then judging the working state of the tunnel boring machine in the next tunneling cycle according to the lithology prediction result. The beneficial effect is that the real-time lithology prediction is realized in the TBM construction process, and the local features in the address data are analyzed more carefully, so as to improve the sensitivity to lithology change.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for tunnel boring machines, and in particular to a method for real-time prediction of rock properties during the tunneling process of a tunnel boring machine. Background Technology

[0002] Tunnel Boring Machines (TBMs), as highly efficient underground engineering machinery, play a vital role in modern underground engineering construction and are widely used in water conservancy tunnels, mine roadways, subway projects, and other large-scale infrastructure projects. The advantages of TBMs lie in their ability to approach complex geological conditions with relatively high efficiency, while ensuring construction safety through automated control systems. Compared to traditional drill-and-blast methods, TBM automated operation significantly reduces reliance on human labor and also lowers environmental impact, such as less noise and dust pollution. However, TBMs still have certain limitations in terms of geological adaptability; their performance is easily affected by different geological conditions, thus limiting their widespread application.

[0003] Under complex geological conditions, TBM performance varies significantly. For example, in hard rock environments, the wear rate of the TBM cutterhead increases significantly, while in soft rock, it faces increased tunneling resistance. Therefore, lithology becomes a key factor affecting TBM construction efficiency, cutterhead wear, and construction safety. Traditional lithology prediction methods rely on geological surveys, advanced drilling, and geophysical exploration, which cannot achieve real-time lithology prediction, and these methods require substantial manpower and resources. To improve the adaptability of TBMs under different geological conditions, real-time lithology prediction has become an important technical means. 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 real-time lithology prediction method during the tunnel boring machine excavation process. It solves the technical problems of traditional lithology prediction methods that rely on geological surveys, advanced drilling and geophysical exploration, which cannot achieve real-time lithology prediction, and require a lot of manpower and material resources.

[0006] (II) Technical Solution

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

[0008] In a first aspect, embodiments of the present invention provide a method for real-time prediction of lithology during the tunnel boring machine excavation process, comprising:

[0009] S11. Real-time detection of valid data corresponding to each tunneling cycle during the tunnel boring machine's tunneling process;

[0010] S12. Input the obtained valid data into the pre-trained lithology prediction model to obtain the lithology prediction results output by the lithology prediction model.

[0011] During this period, the characteristic parameters corresponding to the effective data are obtained based on the effective data corresponding to each tunneling cycle;

[0012] Based on the feature parameters corresponding to the obtained valid data and the pre-set feature block size, the feature parameters corresponding to the valid data are divided into at least one feature block, and the time domain signal and frequency domain signal corresponding to each feature block are obtained.

[0013] The frequency domain signal and time domain signal corresponding to each feature block are respectively input into a pre-set LSTM network to obtain the frequency domain feature and time domain feature corresponding to that feature block;

[0014] Based on the frequency domain features and time domain features corresponding to each feature block, and the pre-set hybrid attention mechanism, the time domain attention matrix and frequency domain attention matrix corresponding to each feature block are obtained.

[0015] Based on the temporal attention matrix, frequency domain attention matrix, frequency domain features, and temporal features corresponding to each feature block, the lithology prediction results corresponding to the effective data are obtained.

[0016] S13. Based on the lithology prediction results corresponding to the valid data, the working status of the tunnel boring machine in the next tunneling cycle is determined.

[0017] Optionally, each tunneling cycle includes an ascending phase and a stabilizing phase;

[0018] The valid data corresponding to each tunneling cycle includes: the state parameters of each unit time in the first preset time period of the rising segment of the tunneling cycle and the state parameters of each unit time in the second preset time period of the stable segment;

[0019] Then S11 includes:

[0020] The system monitors the status parameters corresponding to each unit of time within each tunneling cycle during the tunnel boring machine's tunneling process in real time, and filters all the status parameters within each tunneling cycle to obtain the effective data corresponding to each tunneling cycle.

[0021] Each status parameter includes: cutterhead rotation speed, tunneling speed, total thrust, cutterhead torque, shield pressure, support shoe pressure, support shoe pump pressure, main conveyor pump pressure, and main conveyor rotation speed.

[0022] Optionally, the feature parameters include batch size, sequence length, and feature dimension;

[0023] Then, in step S12, based on the feature parameters corresponding to the obtained valid data and the pre-set feature block size, the valid data is divided into at least one feature block, and the time-domain signal and frequency-domain signal corresponding to each feature block are obtained, including:

[0024] Based on the sequence length and feature dimension corresponding to the obtained valid data, as well as the pre-set feature block size and pre-set Formula 1, the valid data is divided into at least one feature block, each feature block having its corresponding sequence length and feature dimension; Formula 1 is:

[0025] ;

[0026] Where P is the number of feature blocks corresponding to valid data, T is the sequence length corresponding to valid data, and L is the pre-set feature block size;

[0027] Based on the sequence length and feature dimension corresponding to each feature block, and according to the pre-set Formula 2, each feature block is reshaped to obtain the time-domain signal corresponding to each feature block; Formula 2 is:

[0028] X time =X[:,:P×L,:].view(B,P,L,D);

[0029] Among them, X time B is the time-domain signal corresponding to the feature block, D is the batch size corresponding to the effective data, view is the reshaping function, and X[:,:P×L,:] is the slicing operation on the effective data.

[0030] Based on the time-domain signal corresponding to each feature block and the pre-set Formula 3, the frequency-domain signal corresponding to each feature block is obtained; Formula 3 is:

[0031] X freq =Re(FFT(X time ,dim=2));

[0032] Among them, X freq The frequency domain signal corresponding to the feature block, Re represents taking the real part of the complex number, FFT is the Fast Fourier Transform, dim=2 indicates that the operation is performed in the second dimension, X time This is the time-domain signal corresponding to the feature block.

[0033] Optionally, the LSTM network includes a time-domain LSTM network and a frequency-domain LSTM network;

[0034] Then, in step S12, the frequency domain signal and time domain signal corresponding to each feature block are respectively input into a pre-set LSTM network to obtain the frequency domain features and time domain features corresponding to that feature block, including:

[0035] The frequency domain signal corresponding to the feature block is input into a pre-set frequency domain LSTM network to capture the frequency domain features corresponding to the feature block; the processing procedure of the frequency domain LSTM network is as follows:

[0036] L freq ,(h freq ,c freq )=LSTM freq (X freq );

[0037] Among them, L freq For the frequency domain features corresponding to the feature blocks, LSTM freq For frequency domain LSTM networks, X freq h is the frequency domain signal corresponding to the feature block. freq c represents the short-term memory parameters in a frequency-domain LSTM network. freq These are the long-term memory parameters in the frequency domain LSTM network;

[0038] The time-domain signal corresponding to the feature block is input into a pre-set time-domain LSTM network to capture the time-domain features corresponding to the feature block; the processing procedure of the time-domain LSTM network is as follows:

[0039] L time ,(h time ,c time )=LSTM time (X time );

[0040] Among them, L time For the temporal features corresponding to the feature blocks, LSTM time For a time-domain LSTM network, X time h is the time-domain signal corresponding to the feature block. time c represents the short-term memory parameters in a time-domain LSTM network. time These are the long-term memory parameters in the time-domain LSTM network.

[0041] Optionally, the hybrid attention mechanism includes a multi-head attention mechanism and a multi-head convolutional projection attention mechanism;

[0042] Then, in step S12, based on the frequency domain features and time domain features corresponding to each feature block, and a pre-set hybrid attention mechanism, the time domain attention matrix and frequency domain attention matrix corresponding to each feature block are obtained, including:

[0043] Based on the time-domain and frequency-domain features corresponding to all feature blocks, the time-domain and frequency-domain features corresponding to each feature block are normalized respectively.

[0044] Based on the temporal features corresponding to each feature block, the pre-set hybrid attention mechanism, and the pre-set Formula 4, the temporal attention matrix corresponding to that feature block is obtained.

[0045] Formula four is as follows:

[0046] O time =α1·Attention time (Q r ,K r V r )+β1·Attention time (Q c ,K c V c );

[0047] Among them, O time The temporal attention matrix corresponding to the temporal features of the feature block. time (Q r ,K r V r ) represents the output matrix of the multi-head convolutional projection attention mechanism corresponding to the temporal features of the feature block; Attention time (Q c ,K c V c ) represents the output matrix of the multi-head attention mechanism corresponding to the temporal features of the feature block, α1 represents the pre-set multi-head convolutional projection attention weights of the temporal features, and β1 represents the pre-set multi-head attention weights of the temporal features;

[0048] Simultaneously, based on the frequency domain features corresponding to each feature block, the pre-set hybrid attention mechanism, and the pre-set Formula 5, the frequency domain attention matrix corresponding to that feature block is obtained;

[0049] Formula five is:

[0050] O freq =α2·Attention freq (Q r ,K r V r )+β2·Attention freq (Q c ,K c, V c );

[0051] Among them, O freq The frequency domain attention matrix corresponding to the frequency domain features of the feature block. freq (Q r ,K r V r) represents the output matrix of the multi-head convolutional projection attention mechanism corresponding to the frequency domain features of the feature block; Attention freq (Q c ,K c, V c ) is the output matrix of the multi-head attention mechanism corresponding to the frequency domain features of the feature block, α2 is the pre-set multi-head convolutional projection attention weight of the frequency domain features, and β2 is the pre-set multi-head attention weight of the frequency domain features.

[0052] Optionally, step S12, based on the temporal attention matrix, frequency domain attention matrix, frequency domain features, and temporal features corresponding to each feature block, obtains the lithology prediction result corresponding to the effective data, including:

[0053] Based on the temporal attention matrix, frequency domain attention matrix, temporal features, and frequency domain features corresponding to each feature block, the fusion features corresponding to each feature block are obtained;

[0054] Based on the fusion features corresponding to each feature block and the pre-set weighted attention mechanism, the global information of all feature blocks is obtained, that is, the global context vector corresponding to the effective data.

[0055] Based on the global context vector corresponding to the valid data, the lithology prediction result corresponding to the valid data is obtained.

[0056] Optionally, in step S12, based on the temporal attention matrix, frequency domain attention matrix, temporal features, and frequency domain features corresponding to each feature block, the fused features corresponding to each feature block are obtained, including:

[0057] The temporal attention matrix and frequency attention matrix corresponding to each feature block are respectively input into the pre-set first fully connected layer, and feature transformation is performed based on the pre-set formula six.

[0058] ;

[0059] Among them, A time W is the transformation result of the temporal attention matrix corresponding to the feature block. time For the pre-set time-domain weight matrix, O time A is the temporal attention matrix corresponding to the temporal features of the feature block; freq W is the transformation result of the frequency domain attention matrix corresponding to the feature block. freq For the pre-set time-domain weight matrix, O freq This is the frequency domain attention matrix corresponding to the frequency domain features of the feature block;

[0060] Subsequently, based on the pre-set Formula 7, the transformation result of the temporal attention matrix corresponding to each feature block is residually concatenated with the temporal feature corresponding to that feature block, and the transformation result of the frequency domain attention matrix corresponding to each feature block is residually concatenated with the frequency feature corresponding to that feature block; Formula 7 is:

[0061] ;

[0062] Among them, R time L represents the residual connection result of the feature blocks in the time domain. time R represents the temporal features corresponding to the feature blocks. freq L represents the residual connection result of the feature block in the frequency domain. freq These are the temporal features corresponding to the feature blocks;

[0063] Finally, based on the pre-set Formula 8, the residual connection results in the frequency domain and the residual connection results in the time domain of each feature block are concatenated to obtain the fused features corresponding to each feature block; Formula 8 is:

[0064] C = concat(R) time ,R freq ,dim=2);

[0065] Where C represents the fused feature corresponding to the feature block, concat is the concatenation operation, and dim=2 is the operation in the second feature dimension.

[0066] Optionally, in step S12, global information of all feature blocks is obtained based on the fused features corresponding to each feature block and a pre-set weighted attention mechanism, including:

[0067] Based on the fusion features corresponding to each feature block, the pre-set attention score weight matrix, and the pre-set Formula 9, the attention score matrix corresponding to each feature block is obtained; Formula 9 is:

[0068] S=CW score ;

[0069] Where S is the attention score matrix corresponding to the feature block, C is the fusion feature corresponding to the feature block, and W is the sum of the attention scores. score A pre-set attention score weight matrix;

[0070] Subsequently, the attention score matrix corresponding to each feature block is normalized by a pre-set softmax function to generate the attention weights corresponding to each feature block.

[0071] Finally, based on the attention weights and fusion features corresponding to each feature block, and the pre-set Formula 10, the global context vector corresponding to the effective data is obtained; Formula 10 is:

[0072] ;

[0073] Where A is the attention weight corresponding to the feature block, c is the global context vector corresponding to the effective data, C is the fusion feature corresponding to the feature block, i is the index of the feature block, and P is the number of feature blocks corresponding to the effective data.

[0074] Optionally, step S12, based on the global context vector corresponding to the valid data, obtains the lithology prediction result corresponding to the valid data, including:

[0075] The fused feature corresponding to the last feature block in all feature blocks of the valid data is concatenated with the global context vector corresponding to the valid data to obtain the connection vector corresponding to the valid data; the specific formula is as follows:

[0076] p=concat(c,H t );

[0077] Where p is the join vector corresponding to the valid data, concat is the concatenation operation, c is the global context vector corresponding to the valid data, and H is the global context vector corresponding to the valid data. t This refers to the fused feature corresponding to the last feature block.

[0078] Subsequently, the fusion features corresponding to the valid data are input into a pre-set second fully connected layer, and the lithology prediction results corresponding to the valid data are obtained based on the pre-set formula eleven; the formula eleven is:

[0079] ;

[0080] Among them, W f and W p All are pre-set weights of the second fully connected layer, v is the intermediate feature generated through nonlinear transformation, p is the connection vector corresponding to the effective data, and O is the lithology prediction result corresponding to the effective data.

[0081] Optionally, the process may further include the following steps before step S11:

[0082] S10. Based on the pre-set model training set, train the lithology prediction model to obtain the trained lithology prediction model.

[0083] During this period, the hyperparameters of the lithology prediction model were optimized in real time based on the improved EWOA-GWO-MRFO and the pre-set model training strategy;

[0084] The model training strategy includes:

[0085] During the exploration phase, the optimal hyperparameters are sought based on a pre-set formula (Formula 12); formula 12 is:

[0086] ;

[0087] Among them, X t+1 and X t Let X represent the positions of the individual gray wolves at the (t+1)th iteration and the tth iteration, respectively. α For the individual with the best fitness, A is the step size dynamic adjustment factor, a is the dynamic control factor, r1 is a random number ranging from [0,1], sin(2πl) is the nonlinear perturbation term, and X... -☆ This is the globally optimal solution for the current population;

[0088] When the predation phase is underway, the optimal hyperparameters are sought based on a pre-set formula (Formula 13); Formula 13 is:

[0089] X t+1 =X β -A·|2·r2·X β -X t |;

[0090] Among them, X β For the second best-fit individual, r2 is a random number ranging from [0,1].

[0091] During the spiral predation phase, the optimal hyperparameters are sought based on a pre-set formula fourteen; formula fourteen is:

[0092] X t+1= |X ☆ -X t |·e b·l cos(2πl)+η·(X rand -X t )+γ·(X a -X t );

[0093] Where l is a random factor with a value range of [0,1], η and γ are the dynamic control factors for jumping and cooperation, respectively, b is a constant, and X rand It refers to any individual in the population.

[0094] (III) Beneficial Effects

[0095] The beneficial effects of the present invention are as follows: The present invention provides a real-time lithology prediction method during the tunnel boring machine (TBM) excavation process. By using a lithology prediction model to predict the current lithology in real time during TBM operation, the data is decomposed into small segments for analysis. Compared with the prior art, it can achieve real-time lithology prediction during TBM construction and analyze the local features in geological data in more detail, thereby improving the sensitivity to lithology changes. Attached Figure Description

[0096] Figure 1 This is a flowchart of a method for real-time prediction of lithology during the tunnel boring machine excavation process. Detailed Implementation

[0097] 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.

[0098] This invention proposes a real-time lithology prediction method during tunnel boring machine (TBM) excavation. It employs a novel lithology prediction model, including a patch mechanism, a time-domain and frequency-domain decomposed LSTM network, a hybrid attention mechanism, residual connections and normalization, and a global attention mechanism. By decomposing the data into small segments, the patch mechanism allows the model to analyze local features in the geological data more meticulously, thereby improving its sensitivity to lithological changes. Fourier transform is used to convert the time-domain data into frequency-domain data to capture periodic features and frequency information in the geological data. Subsequently, a time-domain and frequency-domain decomposed LSTM network is employed. The time-domain LSTM network can effectively memorize and utilize long-term time-dependency information, while the frequency-domain LSTM network can identify periodic and oscillatory features in the data. The hybrid attention mechanism combines dynamic convolutional projection and multi-head attention to achieve comprehensive capture of data features. Through the local connectivity of convolutional layers, the dynamic convolutional projection module effectively extracts spatial and frequency features from the data. Multi-head attention allows the model to focus on multiple feature dimensions simultaneously, enhancing its ability to analyze complex data. The hybrid attention mechanism can dynamically adjust the focus of attention to cope with different geological features and variations. To further improve model stability and training efficiency, residual connections and normalization techniques are incorporated. Residual connections alleviate the problems of vanishing and exploding gradients, while batch normalization ensures the stability of the data distribution and reduces internal covariance shift. The global attention mechanism allows the model to consider global contextual information when processing data, rather than just local features.

[0099] 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.

[0100] Example 1

[0101] This embodiment proposes a method for real-time prediction of lithology during the tunnel boring machine excavation process, such as... Figure 1 As shown, it includes:

[0102] S11. Real-time detection of valid data corresponding to each tunneling cycle during the tunnel boring machine's tunneling process;

[0103] S12. Input the obtained valid data into the pre-trained lithology prediction model to obtain the lithology prediction results output by the lithology prediction model.

[0104] During this period, the characteristic parameters corresponding to the effective data are obtained based on the effective data corresponding to each tunneling cycle;

[0105] Based on the feature parameters corresponding to the obtained valid data and the pre-set feature block size, the feature parameters corresponding to the valid data are divided into at least one feature block, and the time domain signal and frequency domain signal corresponding to each feature block are obtained.

[0106] The frequency domain signal and time domain signal corresponding to each feature block are respectively input into a pre-set LSTM network to obtain the frequency domain feature and time domain feature corresponding to that feature block;

[0107] Based on the frequency domain features and time domain features corresponding to each feature block, and the pre-set hybrid attention mechanism, the time domain attention matrix and frequency domain attention matrix corresponding to each feature block are obtained.

[0108] Based on the temporal attention matrix, frequency attention matrix, frequency features, and temporal features corresponding to each feature block, the lithology prediction results corresponding to the effective data are obtained; the lithology prediction results include limestone, granite, tuffaceous sandstone, diorite, etc.

[0109] S13. Based on the lithology prediction results corresponding to the valid data, the working status of the tunnel boring machine in the next tunneling cycle is determined.

[0110] This embodiment provides a method for real-time lithology prediction during the tunnel boring machine (TBM) excavation process. By using a lithology prediction model to predict the current lithology in real time during TBM operation, the data is decomposed into small segments for analysis. Compared with existing technologies, this method can achieve real-time lithology prediction during TBM construction while analyzing local features in geological data in greater detail, thereby improving the sensitivity to lithology changes.

[0111] Example 2

[0112] This embodiment provides a method for real-time prediction of lithology during the tunnel boring machine excavation process, including:

[0113] S11. Real-time detection of valid data corresponding to each tunneling cycle during the tunnel boring machine's tunneling process;

[0114] S12. Input the obtained valid data into the pre-trained lithology prediction model to obtain the lithology prediction results output by the lithology prediction model;

[0115] S13. Based on the lithology prediction results corresponding to the valid data, the working status of the tunnel boring machine in the next tunneling cycle is determined.

[0116] The tunneling cycle of a tunnel boring machine typically includes an ascending phase and a steady phase. The ascending phase refers to the stage that the tunnel boring machine experiences at the beginning of its excavation, also known as the "start-up phase" or "advancement phase." The steady phase refers to the stage where the tunnel boring machine approaches the target position during the excavation process. In the ascending phase, the tunnel boring machine gradually accelerates from a stationary state and begins to excavate the tunnel in the underground rock and soil. In the steady phase, the tunnel boring machine approaches the target position and needs to stabilize its excavation, adapt to the underground rock and soil conditions, and make necessary adjustments.

[0117] The valid data corresponding to each tunneling cycle includes: the state parameters of each unit time in the first preset time period of the ascending phase and the state parameters of each unit time in the second preset time period of the stabilizing phase of the tunneling cycle; usually, the state parameters of the first 10 seconds of the ascending phase and the first 15 seconds of the stabilizing phase are selected, that is, including these 25 seconds, the state parameters corresponding to each second; the state parameters per second include: cutterhead rotation speed, tunneling speed, total propulsion force, cutterhead torque, shield pressure, support shoe pressure, support shoe pump pressure, main conveyor pump pressure, and main conveyor rotation speed, etc.

[0118] Specifically, S11 includes: real-time detection of the state parameters corresponding to each unit time (usually 1 second) within each tunneling cycle during the tunnel boring machine's tunneling process, and filtering of all state parameters within each tunneling cycle to obtain valid data corresponding to each tunneling cycle.

[0119] In S12 above, the processing procedure of the lithology prediction model specifically includes:

[0120] The feature parameters corresponding to each valid data point are obtained based on the valid data for each tunneling cycle; the feature parameters include batch size, sequence length, and feature dimension.

[0121] Based on the sequence length and feature dimension corresponding to the obtained valid data, as well as the pre-set feature block size and pre-set Formula 1, the valid data is divided into at least one feature block, each feature block having its corresponding sequence length and feature dimension; Formula 1 is:

[0122] ;

[0123] Where P is the number of feature blocks corresponding to valid data, T is the sequence length corresponding to valid data, and L is the pre-set feature block size;

[0124] Based on the sequence length and feature dimension corresponding to each feature block, and according to the pre-set Formula 2, each feature block is reshaped to obtain the time-domain signal corresponding to each feature block; Formula 2 is:

[0125] X time =X[:,:P×L,:].view(B,P,L,D);

[0126] Among them, X time B is the time-domain signal corresponding to the feature block, D is the batch size corresponding to the effective data, view is the reshaping function, and X[:,:P×L,:] is the slicing operation on the effective data.

[0127] Based on the time-domain signal corresponding to each feature block, a Fast Fourier Transform is used to convert the time-domain signal into a frequency-domain signal, enabling us to capture the periodicity and trend information in the data. Specifically, this is achieved using Formula 3 to obtain the frequency-domain signal corresponding to each feature block. Formula 3 is as follows:

[0128] X freq =Re(FFT(X time ,dim=2));

[0129] Among them, X freq The frequency domain signal corresponding to the feature block, Re represents taking the real part of the complex number, FFT is the Fast Fourier Transform, dim=2 indicates that the operation is performed in the second dimension, X time This is the time-domain signal corresponding to the feature block.

[0130] The frequency domain signal and time domain signal corresponding to each feature block are respectively input into a pre-set frequency domain LSTM network and a frequency domain LSTM network;

[0131] That is, the frequency domain signal corresponding to the feature block is input into a pre-set frequency domain LSTM network to capture the frequency domain features corresponding to the feature block; the processing procedure of the frequency domain LSTM network is as follows:

[0132] L freq ,(h freq ,c freq )=LSTM freq (X freq );

[0133] Among them, L freq For the frequency domain features corresponding to the feature blocks, LSTM freq For frequency domain LSTM networks, X freq h is the frequency domain signal corresponding to the feature block. freq c represents the short-term memory parameters in a frequency-domain LSTM network. freqThese are the long-term memory parameters in the frequency domain LSTM network;

[0134] The time-domain signal corresponding to the feature block is input into a pre-set time-domain LSTM network to capture the time-domain features corresponding to the feature block; the processing procedure of the time-domain LSTM network is as follows:

[0135] L time ,(h time ,c time )=LSTM time (X time );

[0136] Among them, L time For the temporal features corresponding to the feature blocks, LSTM time For a time-domain LSTM network, X time h is the time-domain signal corresponding to the feature block. time c represents the short-term memory parameters in a time-domain LSTM network. time These are the long-term memory parameters in the time-domain LSTM network.

[0137] Based on the time-domain and frequency-domain features corresponding to all feature blocks, the time-domain and frequency-domain features corresponding to each feature block are normalized respectively.

[0138] Based on a pre-set hybrid attention mechanism, the time-domain features and frequency-domain features corresponding to each feature block are processed separately to obtain the time-domain attention matrix and frequency-domain attention matrix corresponding to the feature block.

[0139] For time-domain features, they are processed using Formula 4, which is:

[0140] O time =α1·Attention time (Q r ,K r V r )+β1·Attention time (Q c ,K c V c );

[0141] Among them, O time The temporal attention matrix corresponding to the temporal features of the feature block. time (Q r ,K r V r ) represents the output matrix of the multi-head convolutional projection attention mechanism corresponding to the temporal features of the feature block. Attention time (Q c ,K c Vc ) represents the output matrix of the multi-head attention mechanism corresponding to the temporal features of the feature block, α1 represents the pre-set multi-head convolutional projection attention weights of the temporal features, and β1 represents the pre-set multi-head attention weights of the temporal features;

[0142] For frequency domain features, they are processed using Formula 5, which is:

[0143] O freq =α2·Attention freq (Q r ,K r V r )+β2·Attention freq (Q c ,K c, V c );

[0144] Among them, O freq The frequency domain attention matrix corresponding to the frequency domain features of the feature block. freq (Q r ,K r V r ) represents the output matrix of the multi-head convolutional projection attention mechanism corresponding to the frequency domain features of the feature block; Attention freq (Q c ,K c, V c ) is the output matrix of the multi-head attention mechanism corresponding to the frequency domain features of the feature block, α2 is the pre-set multi-head convolutional projection attention weight of the frequency domain features, and β2 is the pre-set multi-head attention weight of the frequency domain features.

[0145] The temporal attention matrix and frequency attention matrix corresponding to each feature block are respectively input into the pre-set first fully connected layer, and feature transformation is performed based on the pre-set formula six.

[0146] ;

[0147] Among them, A time W is the transformation result of the temporal attention matrix corresponding to the feature block. time For the pre-set time-domain weight matrix, O time A is the temporal attention matrix corresponding to the temporal features of the feature block; freq W is the transformation result of the frequency domain attention matrix corresponding to the feature block. freq For the pre-set time-domain weight matrix, O freq This is the frequency domain attention matrix corresponding to the frequency domain features of the feature block;

[0148] Subsequently, based on the pre-set Formula 7, the transformation result of the temporal attention matrix corresponding to each feature block is residually concatenated with the temporal feature corresponding to that feature block, and the transformation result of the frequency domain attention matrix corresponding to each feature block is residually concatenated with the frequency feature corresponding to that feature block; Formula 7 is:

[0149] ;

[0150] Among them, R time L represents the residual connection result of the feature blocks in the time domain. time R represents the temporal features corresponding to the feature blocks. freq L represents the residual connection result of the feature block in the frequency domain. freq These are the temporal features corresponding to the feature blocks;

[0151] Finally, based on the pre-set Formula 8, the residual connection results in the frequency domain and the residual connection results in the time domain of each feature block are concatenated to obtain the fused features corresponding to each feature block; Formula 8 is:

[0152] C = concat(R) time ,R freq ,dim=2);

[0153] Where C represents the fused feature corresponding to the feature block, concat is the concatenation operation, and dim=2 is the operation in the second feature dimension.

[0154] Based on the fusion features corresponding to each feature block, the pre-set attention score weight matrix, and the pre-set Formula 9, the attention score matrix corresponding to each feature block is obtained; Formula 9 is:

[0155] S=CW score ;

[0156] Where S is the attention score matrix corresponding to the feature block, C is the fusion feature corresponding to the feature block, and W is the sum of the attention scores. score A pre-set attention score weight matrix;

[0157] Subsequently, the attention score matrix corresponding to each feature block is normalized by a pre-set softmax function to generate the attention weights corresponding to each feature block.

[0158] Finally, based on the attention weights and fusion features corresponding to each feature block, and the pre-set Formula 10, the global context vector corresponding to the effective data is obtained; Formula 10 is:

[0159] ;

[0160] Where A is the attention weight corresponding to the feature block, c is the global context vector corresponding to the effective data, C is the fusion feature corresponding to the feature block, i is the index of the feature block, and P is the number of feature blocks corresponding to the effective data.

[0161] The fused feature corresponding to the last feature block in all feature blocks of the valid data is concatenated with the global context vector corresponding to the valid data to obtain the connection vector corresponding to the valid data; the specific formula is as follows:

[0162] p=concat(c,H t );

[0163] Where p is the join vector corresponding to the valid data, concat is the concatenation operation, c is the global context vector corresponding to the valid data, and H is the global context vector corresponding to the valid data. t This refers to the fused feature corresponding to the last feature block.

[0164] Subsequently, the fusion features corresponding to the valid data are input into a pre-set second fully connected layer, and the lithology prediction results corresponding to the valid data are obtained based on the pre-set formula eleven; the formula eleven is:

[0165] ;

[0166] Among them, W f and W p All are pre-set weights of the second fully connected layer, v is the intermediate feature generated by nonlinear transformation, p is the connection vector corresponding to the effective data, O is the lithology prediction result corresponding to the effective data, tanh is the hyperbolic tangent function, and dropout is regularization.

[0167] Traditional CNNs are limited to focusing primarily on local information, making it difficult to capture long-range dependencies in data, especially when dealing with complex time-series data. While 1D-CNNs can be applied to time-series data, they still cannot fully capture global temporal dependencies. LSTM networks assign the same weights to inputs at different time steps, making it difficult to flexibly adjust the importance of each time step. Transformers are inadequate in capturing local details because they tend to distribute attention evenly across the entire dataset.

[0168] The lithology prediction model provided in this embodiment includes a patch mechanism, a time-domain and frequency-domain decomposed LSTM network, a hybrid attention mechanism, residual connections and normalization, and a global attention mechanism. By decomposing the data into small segments, the patch mechanism allows the model to analyze local features in geological data in greater detail, thereby improving its sensitivity to lithological changes. Fourier transform is used to convert time-domain data into frequency-domain data to capture periodic features and frequency information in the geological data. Subsequently, a time-domain and frequency-domain decomposed LSTM network is employed. The time-domain LSTM network can effectively memorize and utilize long-term time-dependent information, while the frequency-domain LSTM network can identify periodic and oscillatory features in the data. The hybrid attention mechanism combines dynamic convolutional projection and multi-head attention to achieve comprehensive capture of data features. Specifically, through the local connectivity characteristics of convolutional layers, the dynamic convolutional projection module can effectively extract spatial and frequency features from the data; multi-head attention allows the model to focus on multiple feature dimensions simultaneously, enhancing its ability to analyze complex data; and the hybrid attention mechanism can dynamically adjust the focus of attention to cope with different geological features and changes. To further improve model stability and training efficiency, residual connections and normalization techniques were incorporated. Residual connections alleviate the vanishing and exploding gradient problems, while batch normalization ensures the stability of the data distribution and reduces internal covariance shift. A global attention mechanism allows the model to consider global contextual information when processing data, rather than just local features.

[0169] Example 3

[0170] This embodiment provides a method for training a lithology prediction model, including:

[0171] Obtain a model training set for training the lithology prediction model, wherein the model training includes the state parameters of the tunnel boring machine under various lithological conditions;

[0172] Based on the model training set, the lithology prediction model is trained to obtain a well-trained lithology prediction model;

[0173] During model training, the hyperparameters of the lithology prediction model are optimized based on the improved EWOA-GWO-MRFO and the pre-set model training strategy.

[0174] The model training strategy includes: during the exploration phase, finding the optimal hyperparameters based on a pre-set formula 12; formula 12 is:

[0175] ;

[0176] Among them, X t+1 and X t Let X represent the positions of the individual gray wolves at the (t+1)th iteration and the tth iteration, respectively.a For the individual with the best fitness, A is the step size dynamic adjustment factor, a is the dynamic control factor, r1 is a random number ranging from [0,1], sin(2πl) is the nonlinear perturbation term, and X... ☆ This is the globally optimal solution for the current population;

[0177] When the predation phase is underway, the optimal hyperparameters are sought based on a pre-set formula (Formula 13); Formula 13 is:

[0178] X t+1 =X β -A·|2·r2·X β -X t |;

[0179] Among them, X β For the second best-fit individual, r2 is a random number ranging from [0,1].

[0180] During the spiral predation phase, the optimal hyperparameters are sought based on a pre-set formula fourteen; formula fourteen is:

[0181] X t+1= |X ☆ -X t |·e b·l cos(2πl)+η·(X rand -X t )+γ·(X a -X t );

[0182] Where l is a random factor with a value range of [0,1], η and γ are the dynamic control factors for jumping and cooperation, respectively, b is a constant, and X rand It refers to any individual in the population.

[0183] The model proposed in this embodiment has numerous hyperparameters, such as learning rate, number of neurons, and number of network layers. To achieve suitable performance, an improved EWOA-GWO-MRFO approach is used to optimize these hyperparameters. To improve the performance of the Whale Optimization Algorithm (EWOA), the α-wolf and β-wolf guidance mechanisms of the Gray Wolf Optimization Algorithm (GWO) are introduced into the exploration and circling predation phases. Simultaneously, the vortex predation and chain predation behaviors of the Manta Ray Optimization Algorithm (MRFO) are introduced into the spiral predation phase, while retaining the whale spiral search formula. This improved scheme not only enhances global search capability and local search accuracy but also adds jump and mutation mechanisms to avoid getting trapped in local optima.

[0184] During the exploration phase, the whale optimization algorithm performs a global search by simulating random perturbations. To improve search efficiency, it incorporates the guiding behavior of the leader wolf (alpha wolf) in the Global Wilderness Optimization (GWO). Guided by the alpha wolf, individual whales can efficiently approach the global optimum.

[0185] During the predation phase, whales gradually approach their prey, leading to the introduction of the secondary leader wolf (β wolf) behavior from the gray wolf algorithm. This allows individual whales to more accurately approach the optimal solution through cooperation during local searches. The guidance behavior of the secondary leader wolf (β wolf) enhances the whale's search ability in local areas, preventing it from getting trapped in local optima.

[0186] In the spiral predation phase, the core formula of the whale optimization algorithm is retained, while the vortex predation and chain predation of MFO are combined to enhance the flexibility and jumping ability of the local search. The classic path formula of the whale algorithm is retained to simulate the spiral movement of a whale around its prey; vortex predation: through a random jumping mechanism, the search is made more flexible and avoids getting trapped in local optima; chain predation: through cooperative behavior among individuals, the accuracy of the local search is enhanced.

[0187] Example 4

[0188] This embodiment provides a method for real-time prediction of lithology during the tunnel boring machine excavation process, including:

[0189] In TBM construction, real-time monitoring and analysis of state parameters are crucial. These data are typically time-series data, including cutterhead rotation speed, tunneling speed, total thrust, cutterhead torque, shield pressure, support shoe pressure, support shoe pump pressure, main conveyor pump pressure, and main conveyor rotation speed during tunnel excavation. To better capture the local patterns and periodic features in this time-series data, we segmented the input data (i.e., the effective data) into multiple feature blocks and extracted frequency domain features. This approach helps the model capture the changing patterns of the data at different time scales, enhancing the understanding and analysis of TBM construction time-series data. The selected input data consisted of the first 15 seconds of continuous data from the rising phase and the first 10 seconds of continuous data from the stabilizing phase, totaling 25 seconds of data.

[0190] First, the input valid data X has dimensions (B, T, D), where B represents the batch size / batch, T represents the sequence length, and D represents the input feature dimension. To process this data, the input valid data is divided into multiple feature blocks based on a pre-set feature block size, as shown in the following formula:

[0191] ;

[0192] Where P is the number of feature blocks corresponding to valid data, T is the sequence length corresponding to valid data, and L is the pre-set feature block size;

[0193] Subsequently, the segmented feature blocks are reshaped accordingly using the view function, namely:

[0194] X time =X[:,:P×L,:].view(B,P,L,D);

[0195] Among them, X time denoted as , where is the time-domain signal corresponding to the feature block, B is the batch size corresponding to the valid data, D is the feature dimension corresponding to the feature block, view is the reshaping function, and X[:,:P×L,:] is the slicing operation performed on the valid data.

[0196] To more accurately capture the periodic characteristics in TBM construction time-series data, a Fourier transform is performed on each feature block to extract frequency domain features. The Fourier transform converts the time-domain signal into a frequency-domain signal, enabling the capture of periodicity and trend information in the data. The formula is as follows:

[0197] ;

[0198] Where x(n) represents the time-domain signal, X(f) represents the complex representation of the frequency-domain signal, N is the sequence length of the feature block, f is the frequency, and j is the imaginary unit of the complex imaginary part.

[0199] Here, the complex part of the Fourier transform result retains only the real part, that is:

[0200] X freq =Re(FFT(X time ,dim=2));

[0201] Among them, X freq The frequency domain signal corresponding to the feature block, Re represents taking the real part of the complex number, FFT is the Fast Fourier Transform, dim=2 indicates that the operation is performed in the second dimension, X time This is the time-domain signal corresponding to the feature block.

[0202] By using the above data preprocessing methods, we can better capture the local and periodic features of the TBM construction process and enhance the model's ability to identify and understand these features.

[0203] Furthermore, the frequency and time domain features of these feature blocks are processed separately using an LSTM network. Specifically, when processing TBM construction time-series data, the LSTM network captures long-term and short-term dependencies through a gating mechanism. To process this time-series data, an LSTM network is used to process both time-domain and frequency-domain signals, capturing information from different domains.

[0204] The specific process of LSTM is as follows: the input gate controls the number of input states to the cell:

[0205] i t =σ(W i x t +U i h t-1 +b i );

[0206] Where t usually represents the current time step in the sequence, i t For the output of the input gate, b i h is the bias vector of the input gate. t-1 W is the hidden state from the previous moment. i U is the weight matrix of the input gate. i σ is the weight matrix in the input gate that is related to the hidden state in the previous time step, and σ is the Sigmoid activation function.

[0207] The forgetting gate controls the amount of cell state retained from the previous time step:

[0208] f t =σ(W f x t +U f h t-1 +b f );

[0209] Among them, f t For the output of the forget gate, W f U is the weight matrix of the forget gate. f It is the weight matrix related to the hidden state in the previous time step in the forget gate, b f This is the bias vector for the forget gate.

[0210] o t =σ(W o x t +U o h t-1 +b o );

[0211] Among them, o t W is the output of the output gate. o U is the weight matrix of the output gate. o b is the weight matrix in the output gate that is related to the hidden state in the previous time step. o This is the bias vector for the output gate.

[0212] Calculation of new candidate cell states:

[0213] ;

[0214] in, W represents the current memory state. c U is the weight matrix for the state candidates.c It is the weight matrix b in the state candidates that is related to the hidden state of the previous time step. c The bias vector is the state candidate.

[0215] Cell status update:

[0216] ;

[0217] c t This represents the current state of the cell.

[0218] Hidden status update:

[0219] h t =o t ⊙tanh(c t );

[0220] Here, ⊙ represents element-wise multiplication.

[0221] Specifically,

[0222] Time-domain LSTM network processing:

[0223] L time ,(h time ,c time )=LSTM time (X time );

[0224] Among them, L time For the temporal features corresponding to the feature blocks, LSTM time For a time-domain LSTM network, X time h is the time-domain signal corresponding to the feature block. time c represents the short-term memory parameters in a time-domain LSTM network. time These are the long-term memory parameters in the time-domain LSTM network.

[0225] Frequency domain LSTM network processing:

[0226] L freq ,(h freq ,c freq )=LSTM freq (X freq );

[0227] Among them, L freq For the frequency domain features corresponding to the feature blocks, LSTM freq For frequency domain LSTM networks, X freq h is the frequency domain signal corresponding to the feature block. freq c represents the short-term memory parameters in a frequency-domain LSTM network. freq These are the long-term memory parameters in the frequency domain LSTM network;

[0228] The time-domain signal corresponding to the feature block is input into a pre-set time-domain LSTM network to capture the time-domain features corresponding to the feature block; the processing procedure of the time-domain LSTM network is as follows:

[0229] When processing TBM construction time-series data, the differences in data distribution between different batches can affect the training effect of the model. In order to improve the stability of training and the convergence speed of the model, we performed batch normalization on the features within each batch to avoid excessive differences in data distribution between different batches, thereby improving the generalization ability of the model.

[0230] Batch standardization standardizes the features within each batch, as shown in the following formula:

[0231] ;

[0232] Where, μ batch and Let L be the mean and variance of the batch, respectively; ε is a small constant set for numerical stability; and L is the mean and variance of the batch. x For L time or L freq .

[0233] To preserve original feature information and improve training stability when processing TBM construction time series data, we add attention outputs in the time and frequency domains to the features processed by LSTM, forming residual connections. Residual connections effectively alleviate the vanishing gradient problem, improve training performance, and preserve original feature information.

[0234] First, the attention outputs of time-domain features and frequency-domain features are calculated, that is, the attention matrix of frequency-domain features and time-domain features is calculated through a hybrid attention mechanism.

[0235] In TBM construction time-series data, the extraction of local features is crucial because these features may contain key information about changes in construction parameters. To improve local feature extraction in the multi-head attention mechanism and enable the model to better capture local patterns, we use convolution operations to generate queries (Q... c ), key (K) c ) and value matrix (V c Convolution operations extract local features by learning the kernel weights, as shown in the following equation:

[0236] Q c =Conv1D(X),K c =Conv1D(X),V c =Conv1D(X);

[0237] The key is to utilize local perception to extract features from adjacent time steps, Q c K c and V c The calculation process is similar, using Q c For example, let's demonstrate the calculation process:

[0238] ;

[0239] Where K is the size of the convolution kernel, W convk It is the weight of the k-th convolutional kernel, x i+k It is the (i+k)th time step of the input sequence.

[0240] Next, the attention score is calculated using convolutional projection:

[0241] ;

[0242] T is the transpose, d k The dimension of the key.

[0243] To compute dependencies between different locations globally, a multi-head attention mechanism is used to generate query, key, and value matrices and compute attention weights, thereby capturing long-distance dependencies in the input sequence. The multi-head attention mechanism improves the model's expressive power by processing information from different subspaces in parallel through multiple attention heads.

[0244] First, a linear transformation generates a query, key, and value matrix, as shown in the following equation:

[0245] ;

[0246] Among them, W Q W K and W V These are the weight parameters for the query, key, and value matrices, respectively.

[0247] Next, the dot product of the query and the key is calculated and scaled by the square root of the key's dimension to obtain the scaled attention matrix, as shown below:

[0248] ;

[0249] T is the transpose, d k Q is the dimension of the key. r K r T It is the dot product of the query and key matrices.

[0250] To combine local and global features, we employ a dynamic weighted fusion of the results from convolutional projection and traditional attention, enabling the model to adaptively balance the contributions of local and global information. This hybrid projection attention mechanism can simultaneously utilize detailed information within local windows and long-range dependencies in the global sequence, thereby enhancing a comprehensive understanding of the construction process.

[0251] First, we compute the attention results of convolutional projection and traditional attention mechanisms, and then obtain the final attention output through weighted fusion.

[0252] For the time domain, it is calculated using the following formula:

[0253] O time =α1·Attention time (Q r ,K r V r )+β1·Attention time (Q c ,K c V c );

[0254] Among them, O time The temporal attention matrix corresponding to the temporal features of the feature block. time (Q r ,K r V r ) represents the output matrix of the multi-head convolutional projection attention mechanism corresponding to the temporal features of the feature block. Attention time (Q c ,K c V c ) represents the output matrix of the multi-head attention mechanism corresponding to the temporal features of the feature block, α1 represents the pre-set multi-head convolutional projection attention weights of the temporal features, and β1 represents the pre-set multi-head attention weights of the temporal features;

[0255] For the frequency domain:

[0256] O freq =α2·Attention freq (Q r ,K r V r )+β2·Attention freq (Q c ,K c, V c );

[0257] Among them, O freq The frequency domain attention matrix corresponding to the frequency domain features of the feature block.freq (Q r ,K r V r ) represents the output matrix of the multi-head convolutional projection attention mechanism corresponding to the frequency domain features of the feature block. Attention freq (Q c ,K c, V c ) is the output matrix of the multi-head attention mechanism corresponding to the frequency domain features of the feature block, α2 is the pre-set multi-head convolutional projection attention weight of the frequency domain features, and β2 is the pre-set multi-head attention weight of the frequency domain features.

[0258] Specifically, for example:

[0259] First, compute the convolutional projection query, key, and value matrix. Then, split the query, key, and value matrix into multiple headers:

[0260] ;

[0261] Q c K c and V c It was split into an h header;

[0262] Next, we compute the scaled dot product attention for each head:

[0263] ;

[0264] The scaled dot product attention of each head is given by Q. ch K ch V ch Let h represent the number of attention heads. Then, concatenate the outputs of all heads:

[0265] ;

[0266] Next, a similar approach will be taken for the traditional multi-head attention mechanism;

[0267] ;

[0268] Here, Q r K r V r Decompose it into the h header.

[0269] Next, we compute the scaled dot product attention for each head:

[0270] ;

[0271] The result is given by Q. rh K rh Vrh Let h represent the number of attention heads. Then, concatenate the outputs of all heads:

[0272] O r =concat(Attention(Q r1 ,K r1 V r1 ),...,Attention(Q rh ,K rh V rh ));

[0273] Finally, the results of convolutional projection and traditional attention are weighted and fused:

[0274] O f =αO r +βO c .

[0275] Through the first fully connected layer, the temporal attention matrix and frequency domain attention matrix corresponding to the obtained feature blocks are transformed, that is:

[0276] ;

[0277] Among them, A time W is the transformation result of the temporal attention matrix corresponding to the feature block. time For the pre-set time-domain weight matrix, O time A is the temporal attention matrix corresponding to the temporal features of the feature block; freq W is the transformation result of the frequency domain attention matrix corresponding to the feature block. freq For the pre-set time-domain weight matrix, O freq This is the frequency domain attention matrix corresponding to the frequency domain features of the feature block;

[0278] Next, residual connections are performed by adding the transformed attention output to the features processed by the LSTM; that is:

[0279] ;

[0280] Among them, R time L represents the residual connection result of the feature blocks in the time domain. time R represents the temporal features corresponding to the feature blocks. freq L represents the residual connection result of the feature block in the frequency domain. freq These are the temporal features corresponding to the feature blocks;

[0281] Finally, when generating the context vector, the residual connection results in the time and frequency domains are combined to form a comprehensive feature, which improves the model's representational ability and enables it to better capture the diversity of data, i.e.: C = concat(R time ,R freq ,dim=2);

[0282] Where C represents the fused feature corresponding to the feature block, concat is the concatenation operation, and dim=2 is the operation in the second feature dimension.

[0283] Then, the attention score is calculated: S=CW score ;

[0284] Where S is the attention score matrix corresponding to the feature block, C is the fusion feature corresponding to the feature block, and W is the sum of the attention scores. score A pre-set attention score weight matrix;

[0285] Generate context vectors: ;

[0286] Where A is the attention weight corresponding to the feature block, c is the global context vector corresponding to the effective data, C is the fusion feature corresponding to the feature block, i is the index of the feature block, and P is the number of feature blocks corresponding to the effective data.

[0287] Finally, to generate the final output, the context vector is concatenated with the hidden state of the previous time step and passed to the fully connected layer and the dropout layer: that is:

[0288] The fused feature corresponding to the last feature block in all feature blocks of the valid data is concatenated with the global context vector corresponding to the valid data to obtain the connection vector corresponding to the valid data; the specific formula is as follows:

[0289] p=concat(c,H t );

[0290] Where p is the join vector corresponding to the valid data, concat is the concatenation operation, c is the global context vector corresponding to the valid data, and H is the global context vector corresponding to the valid data. t This refers to the fused feature corresponding to the last feature block.

[0291] Subsequently, the fusion features corresponding to the valid data are input into a pre-set second fully connected layer, and the lithology prediction results corresponding to the valid data are obtained based on the pre-set formula eleven; the formula eleven is:

[0292] ;

[0293] Among them, W f and W pAll are pre-set weights of the second fully connected layer, v is the intermediate feature generated by nonlinear transformation, p is the connection vector corresponding to the effective data, O is the lithology prediction result corresponding to the effective data, tanh is the hyperbolic tangent function, and dropout is regularization.

[0294] This embodiment proposes a real-time lithology prediction method during the tunnel boring machine (TBM) excavation process. Based on the state parameters during TBM construction, the current lithology is predicted in real time, thereby enabling dynamic adjustment of parameters and construction plans during the excavation process based on the prediction results, effectively reducing the risk of equipment damage and sudden accidents.

[0295] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0296] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0297] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "over," or "on top" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," or "beneath" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0298] 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. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, 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.

[0299] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for real-time prediction of lithology during tunnel boring machine excavation, characterized in that, include: S11. Real-time detection of valid data corresponding to each tunneling cycle during the tunnel boring machine's tunneling process; S12. Input the obtained valid data into the pre-trained lithology prediction model to obtain the lithology prediction results output by the lithology prediction model. During this process, feature parameters corresponding to each effective data point are obtained based on the effective data for each tunneling cycle; the feature parameters include batch size, sequence length, and feature dimension. Based on the sequence length and feature dimension corresponding to the obtained valid data, as well as the pre-set feature block size and pre-set Formula 1, the valid data is divided into at least one feature block, each feature block having its corresponding sequence length and feature dimension; Formula 1 is: ； Where P is the number of feature blocks corresponding to valid data, T is the sequence length corresponding to valid data, and L is the pre-set feature block size; Based on the sequence length and feature dimension corresponding to each feature block, and according to the pre-set Formula 2, each feature block is reshaped to obtain the time-domain signal corresponding to each feature block; Formula 2 is: X time =X[:,:P×L,:].view(B,P,L,D); Among them, X time B is the time-domain signal corresponding to the feature block, D is the batch size corresponding to the effective data, view is the reshaping function, and X[:,:P×L,:] is the slicing operation on the effective data. Based on the time-domain signal corresponding to each feature block and the pre-set Formula 3, the frequency-domain signal corresponding to each feature block is obtained; Formula 3 is: X freq =Re(FFT(X time ,dim=2)); Among them, X freq The frequency domain signal corresponding to the feature block, Re represents taking the real part of the complex number, FFT is the Fast Fourier Transform, dim=2 indicates that the operation is performed in the second dimension, X time The time-domain signal corresponding to the feature block; The frequency domain signal and time domain signal corresponding to each feature block are respectively input into a pre-set LSTM network to obtain the frequency domain feature and time domain feature corresponding to that feature block; Based on the time-domain and frequency-domain features corresponding to all feature blocks, the time-domain and frequency-domain features corresponding to each feature block are normalized respectively. Based on the temporal features corresponding to each feature block, the pre-set hybrid attention mechanism, and the pre-set Formula 4, the temporal attention matrix corresponding to that feature block is obtained. Formula four is as follows: O time =α1·Attention time (Q r ,K r ,V r )+β1·Attention time (Q c ,K c ,V c ); Among them, O time The temporal attention matrix corresponding to the temporal features of the feature block. time (Q r ,K r V r ) represents the output matrix of the multi-head convolutional projection attention mechanism corresponding to the temporal features of the feature block. Attention time (Q c ,K c V c ) represents the output matrix of the multi-head attention mechanism corresponding to the temporal features of the feature block, α1 represents the pre-set multi-head convolutional projection attention weights of the temporal features, and β1 represents the pre-set multi-head attention weights of the temporal features; Simultaneously, based on the frequency domain features corresponding to each feature block, the pre-set hybrid attention mechanism, and the pre-set Formula 5, the frequency domain attention matrix corresponding to that feature block is obtained; Formula five is: O freq =α2·Attention freq (Q r ,K r ,V r )+β2·Attention freq (Q c ,K c, V c ); Among them, O freq The frequency domain attention matrix corresponding to the frequency domain features of the feature block. freq (Q r ,K r V r ) represents the output matrix of the multi-head convolutional projection attention mechanism corresponding to the frequency domain features of the feature block; Attention freq (Q c ,K c, V c ) is the output matrix of the multi-head attention mechanism corresponding to the frequency domain features of the feature block, α2 is the pre-set multi-head convolutional projection attention weight of the frequency domain features, and β2 is the pre-set multi-head attention weight of the frequency domain features; the temporal attention matrix and frequency domain attention matrix corresponding to each feature block are respectively input into the pre-set first fully connected layer, and feature transformation is performed based on the pre-set formula six; ； Among them, A time W is the transformation result of the temporal attention matrix corresponding to the feature block. time For the pre-set time-domain weight matrix, O time A is the temporal attention matrix corresponding to the temporal features of the feature block; freq W is the transformation result of the frequency domain attention matrix corresponding to the feature block. freq For the pre-set frequency domain weight matrix, O freq This is the frequency domain attention matrix corresponding to the frequency domain features of the feature block; Subsequently, based on the pre-set Formula 7, the transformation result of the temporal attention matrix corresponding to each feature block is residually concatenated with the temporal feature corresponding to that feature block, and the transformation result of the frequency domain attention matrix corresponding to each feature block is residually concatenated with the frequency feature corresponding to that feature block; Formula 7 is: ； Among them, R time L represents the residual connection result of the feature blocks in the time domain. time For the time-domain features corresponding to the feature blocks, R freq L represents the residual connection result of the feature block in the frequency domain. freq The frequency domain features corresponding to the feature blocks; Finally, based on the pre-set Formula 8, the residual connection results in the frequency domain and the residual connection results in the time domain of each feature block are concatenated to obtain the fused features corresponding to each feature block; Formula 8 is: C=concat(R time ,R freq ,dim=2); Where C represents the fused feature corresponding to the feature block, concat is the concatenation operation, and dim=2 is the operation in the second feature dimension; Based on the fusion features corresponding to each feature block, the pre-set attention score weight matrix, and the pre-set Formula 9, the attention score matrix corresponding to each feature block is obtained; Formula 9 is: S=CW score ; Where S is the attention score matrix corresponding to the feature block, C is the fusion feature corresponding to the feature block, and W is the sum of the attention scores. score A pre-set attention score weight matrix; Subsequently, the attention score matrix corresponding to each feature block is normalized by a pre-set softmax function to generate the attention weights corresponding to each feature block. Finally, based on the attention weights and fusion features corresponding to each feature block, and the pre-set Formula 10, the global context vector corresponding to the effective data is obtained; Formula 10 is: ； Where A is the attention weight corresponding to the feature block, c is the global context vector corresponding to the effective data, C is the fusion feature corresponding to the feature block, i is the index of the feature block, and P is the number of feature blocks corresponding to the effective data. Based on the global context vector corresponding to the valid data, the lithology prediction result corresponding to the valid data is obtained; S13. Based on the lithology prediction results corresponding to the valid data, the working status of the tunnel boring machine in the next tunneling cycle is determined.

2. The method for real-time prediction of lithology during tunnel boring machine excavation as described in claim 1, characterized in that, Each tunneling cycle includes an ascending phase and a stabilizing phase; The valid data for each tunneling cycle includes: the state parameters for each unit of time in the first preset time period during the rising phase of the tunneling cycle and the state parameters for each unit of time in the second preset time period during the stable phase. Then S11 includes: The system monitors the status parameters corresponding to each unit of time within each tunneling cycle during the tunnel boring machine's tunneling process in real time, and filters all the status parameters within each tunneling cycle to obtain the effective data corresponding to each tunneling cycle. Each status parameter includes: cutterhead rotation speed, tunneling speed, total thrust, cutterhead torque, shield pressure, support shoe pressure, support shoe pump pressure, main conveyor pump pressure, and main conveyor rotation speed.

3. The method for real-time prediction of lithology during tunnel boring machine excavation as described in claim 1, characterized in that, The LSTM network includes a time-domain LSTM network and a frequency-domain LSTM network; Then, in step S12, the frequency domain signal and time domain signal corresponding to each feature block are respectively input into a pre-set LSTM network to obtain the frequency domain features and time domain features corresponding to that feature block, including: The frequency domain signal corresponding to the feature block is input into a pre-set frequency domain LSTM network to capture the frequency domain features corresponding to the feature block; the processing procedure of the frequency domain LSTM network is as follows: L freq ,(h freq ,c freq )=LSTM freq (X freq ); Among them, L freq For the frequency domain features corresponding to the feature blocks, LSTM freq For frequency domain LSTM networks, X freq h is the frequency domain signal corresponding to the feature block. freq c represents the short-term memory parameters in a frequency-domain LSTM network. freq These are the long-term memory parameters in the frequency-domain LSTM network; The time-domain signal corresponding to the feature block is input into a pre-set time-domain LSTM network to capture the time-domain features corresponding to the feature block; the processing procedure of the time-domain LSTM network is as follows: L time ,(h time ,c time )=LSTM time (X time ); Among them, L time For the temporal features corresponding to the feature blocks, LSTM time For a time-domain LSTM network, X time h is the time-domain signal corresponding to the feature block. time c represents the short-term memory parameters in a time-domain LSTM network. time These are the long-term memory parameters in the time-domain LSTM network.

4. The method for real-time prediction of lithology during tunnel boring machine excavation as described in claim 1, characterized in that, S12, based on the global context vector corresponding to the valid data, obtains the lithology prediction result corresponding to the valid data, including: The fused feature corresponding to the last feature block in all feature blocks of the valid data is concatenated with the global context vector corresponding to the valid data to obtain the connection vector corresponding to the valid data; the specific formula is as follows: p=concat(c,H t ); Where p is the join vector corresponding to the valid data, concat is the concatenation operation, c is the global context vector corresponding to the valid data, and H is the global context vector corresponding to the valid data. t This refers to the fused feature corresponding to the last feature block. Subsequently, the fusion features corresponding to the valid data are input into a pre-set second fully connected layer, and the lithology prediction results corresponding to the valid data are obtained based on the pre-set formula eleven; the formula eleven is: ； Among them, W f and W p All are pre-set weights of the second fully connected layer, v is the intermediate feature generated by nonlinear transformation, p is the connection vector corresponding to the effective data, O is the lithology prediction result corresponding to the effective data, tanh is the hyperbolic tangent function, and dropout is regularization.

5. The method for real-time prediction of lithology during tunnel boring machine excavation as described in claim 1, characterized in that, The preceding steps, S11, also include: S10. Based on the pre-set model training set, train the lithology prediction model to obtain the trained lithology prediction model. During this period, the hyperparameters of the lithology prediction model were optimized in real time based on the improved EWOA-GWO-MRFO and the pre-set model training strategy; The model training strategy includes: During the exploration phase, the optimal hyperparameters are sought based on a pre-set formula (Formula 12); formula 12 is: ； Among them, X t+1 and X t Let X represent the positions of the individual gray wolves at the (t+1)th iteration and the tth iteration, respectively. α For the individual with the best fitness, A is the step size dynamic adjustment factor, a is the dynamic control factor, r1 is a random number ranging from [0,1], sin(2πl) is the nonlinear perturbation term, and X... ☆ This is the globally optimal solution for the current population; When the predation phase is underway, the optimal hyperparameters are sought based on a pre-set formula (Formula 13); Formula 13 is: X t+1 =X β -A·|2·r2·X β -X t |; Among them, X β For the second best-fit individual, r2 is a random number ranging from [0,1]. During the spiral predation phase, the optimal hyperparameters are sought based on a pre-set formula fourteen; formula fourteen is: X t+1= |X ☆ -X t |·e b·l cos(2πl)+η·(X rand -X t )+γ·(X α -X t ); Where l is a random factor with a value range of [0,1], η and γ are the dynamic control factors for jumping and cooperation, respectively, b is a constant, and X rand It refers to any individual in the population.

Images