Intelligent control method and system for cantilever tunnel boring

Abstract

This invention discloses an intelligent control method and system for cantilever tunneling. This invention relates to the technical field of full-face tunneling processes. The encoding layer receives an input sequence S and calculates candidate forward and backward hidden states using multiple forward and backward GRU (gated recurrent) units. This hidden state sequence H is input into the fuzzy processing layer for fuzzy operations, generating a modified hidden state sequence H' that is then sent to the decoding layer. The invention comprehensively considers various tunneling parameters, such as total thrust, cutterhead power, and torque, and uses heat encoding operations to construct a rich input sequence, providing comprehensive information to the model and improving the accuracy of tunneling speed prediction. The bidirectional Seq-Bi-GRU neural network can capture temporal dependencies and long-term memory effects in data. Compared to traditional BPNN neural networks, it has stronger generalization capabilities and prediction accuracy when processing complex sequence data.

Description

Technical Field

[0001] The present invention relates to the technical field of full-face tunneling processes, and more specifically to prediction and control technology for the PR value (penetration rate) of a full-face hard rock tunnel boring machine (TBM). This technology belongs to the field of model predictive control (MPC), and more particularly to an intelligent control method and system for cantilever tunneling. Background Art

[0002] With its significant advantages, including convenience, efficiency, cost-effectiveness, and environmental friendliness, TBMs have become an indispensable technical tool in a variety of fields, including urban transportation construction, coal mine tunnel excavation, and water conservancy and hydropower tunnel construction. In particular, in areas with significant terrain differences, such as water network branch line construction, the use of small-diameter TBMs has become an inevitable choice.

[0003] However, the construction performance of TBMs is extremely sensitive to construction geological conditions, especially when excavating in composite strata (composite strata refer to combinations of two or more strata with significant differences in geotechnical properties, engineering geological characteristics, and hydrogeological conditions within the tunnel excavation section or along the excavation extension direction).

[0004] When geological conditions don't match TBM excavation parameters, excavation efficiency plummets, leading to a series of problems such as escalating construction costs and extended project durations. For example, in small-diameter tunnel construction, the confined space makes setting TBM excavation parameters even more challenging. This not only increases construction uncertainty but also significantly widens the fluctuation range of the PR value. Therefore, as a key indicator of TBM construction efficiency, accurate prediction of excavation speed is crucial for TBM parameter control, construction risk management, and construction schedule estimation.

[0005] To this end, many existing technologies attempt to achieve this goal, such as:

[0006] (1) CN116446896A discloses a TBM cutterhead speed adaptive control method (publication date 2023-07-18). By establishing a relevant database and a random forest algorithm optimization model, the cutterhead speed is adjusted according to real-time parameters to achieve adaptive control of the cutterhead speed, which can improve the efficiency of construction and the accuracy of TBM operation. However, in the process of excavation in composite formations, the intelligent control of the cutterhead speed is not equivalent to improving the overall construction efficiency of the TBM, because the independent variables of the TBM excavation parameters include total thrust, cutterhead power, cutterhead torque, cutterhead penetration, FPI (field penetration), TPI (torque penetration index) and current excavation speed in addition to the cutterhead speed. (2) CN117404099B discloses an intelligent control method for TBM excavation speed based on the XGBoost algorithm (publication date 2024-08-13). Based on the XGBoost algorithm, an intelligent speed regulation model is established, and the recommended TBM excavation speed is pushed in real time. The recommended parameters are written into the PLC control system to achieve intelligent control of the TBM excavation speed. However, the model of this scheme is a regression task based on excavation environment parameters such as historical rock parameters. If the current construction environment does not have data that can perform the regression task or the deviation is large, the confidence level of the recommended excavation speed is not high.

[0007] (3) CN116702271A discloses a multi-objective optimization method for shield construction parameters based on machine learning and an improved genetic algorithm (publication date 2023-09-05). It establishes two SVM prediction models for shield construction parameters and shield horizontal and vertical postures, and an LGBM prediction model for the effect of shield construction parameters on final surface settlement. Based on the improved NSGA-III algorithm, it obtains a Pareto optimal solution set and determines the control range of shield construction parameters that meets the shield vertical and horizontal posture correction errors and surface settlement deformation limits. However, this solution is more inclined to the optimal solution for the TBM's excavation efficiency at different times; from a macro perspective, this solution that focuses on the optimal solution at every moment actually falls into the defect of the genetic algorithm's local optimal solution and is not a truly optimal solution for the overall construction and its excavation speed.

[0008] (4) CN118940579A discloses a shield machine excavation speed and surface settlement prediction method based on mechanism and data drive (publication date 2024-11-12), which is used to predict shield machine speed and surface settlement during shield machine construction. It includes starting from geotechnical mechanics after geological exploration, considering the mechanical properties of soil, establishing a force model of the cutter cutting soil, and further deriving the calculation formula of the cutter's horizontal cutting force; further introducing the cutting force as an input parameter into the dual-drive prediction model as an input parameter of the dual-drive prediction model, establishing an improved particle swarm algorithm to optimize the support vector machine excavation speed and surface settlement prediction model, realizing the prediction of excavation speed and surface settlement, and greatly improving the accuracy of subsequent excavation speed and surface settlement prediction. Although this solution is based on the force model of the cutter cutting soil, and derives the cutter's horizontal cutting force for optimization at every moment, this solution is only applicable to the single-input and single-output data-driven mode of "horizontal cutting force-surface settlement prediction". When considering multiple independent variables, the model does not support situations where the input and output sequences are of unequal length, and thus lacks robustness, making it difficult to support more complex data-driven processing and more accurate data predictions. Furthermore, during actual tunnel excavation, the dimensions and quantity of input data at different time points may vary due to changes in geological conditions and adjustments to the tunnel boring machine's operating strategy. The model fails to effectively handle the mapping relationship of such unequal-length sequences, limiting its application in complex and changing construction environments.

[0009] (5) The document "Yang Yaohong, Han Xingzhong, Zhang Zhixiao, et al. Prediction of the speed interval of a small-diameter tunnel boring machine in composite strata [J]. Science, Technology and Engineering, 2023, 23(34): 14638-14650." discloses a BPNN neural network that establishes a speed interval prediction model based on the Bootsurap resampling method to predict the PR value. However, this scheme also does not support the situation where the input and output sequences are of unequal length. Although this scheme uses the bootstrap sampling technology combined with random distribution to solve the defects of the regression model in the above-mentioned traditional technology, the Bootstrap resampling technology constructs multiple sub-sample sets by repeatedly extracting the original data set to increase the robustness of the model and reduce the risk of overfitting. However, this method is essentially a random resampling technology based on finite samples. When the sample size is limited or the sample distribution does not represent the overall distribution, the stability and representativeness of the results obtained by Bootstrap sampling will be questioned. Moreover, when combined with random distribution for statistical inference, although this method can generate estimates of the prediction interval, these estimates are often based on the empirical distribution of the sample data rather than the true overall distribution. As a result, the resulting statistical results (such as confidence intervals) may not accurately reflect the true distribution characteristics of actual PR values, significantly compromising the accuracy and reliability of statistical inference. Furthermore, the combination of bootstrap sampling and random distributions cannot directly provide an accurate description of the entire confidence interval distribution, making the model challenging to use when making statistical inferences. This limitation is particularly prominent when accurately assessing the uncertainty of PR values, conducting risk analysis, or making decisions.

[0010] To this end, the present invention proposes an intelligent control method and system for cantilever tunneling. Summary of the Invention

[0011] In light of this, the present invention provides an intelligent control method and system for cantilever tunneling to address or alleviate existing technical issues, namely, how to consider how to support multi-independent variable data drive (i.e., supporting input and output sequences of unequal lengths) during TBM tunneling tasks, further improve the confidence level of the estimated BP value (i.e., provide an accurate description of the entire confidence interval distribution) while avoiding falling into local optimal solutions, and further enhance the control effectiveness of the TBM. The technical solution of the present invention is achieved as follows:

[0012] First, intelligent control method for cantilever tunneling:

[0013] (I) Overview:

[0014] The present invention aims to predict and control the tunneling speed of a TBM (turbine machine brigade) through a data-driven approach. First, tunneling parameters are collected using the TBM sensor system as independent variables X to form an input sequence S, and the future tunneling speed PR is set as the dependent variable Y. Next, a bidirectional Seq-Bi-GRU neural network model is constructed, combined with the Seq2Seq architecture to handle the problem of unequal input and output lengths. A fuzzy processing layer is added to the model to reduce data noise and generate a modified hidden state sequence; the decoding layer uses the DS evidence theory algorithm to fuse multiple pieces of evidence, improve the prediction confidence, and ultimately output the most reliable tunneling speed prediction value Y'. Finally, based on the difference between the predicted value Y' and the actual tunneling speed PR', the control electrical signal is calculated by the PID controller to achieve precise control of the TBM tunneling speed. The integration of deep learning, fuzzy processing, and evidence theory effectively improves the accuracy of tunneling speed prediction and control precision, providing strong support for efficient and safe TBM tunneling.

[0015] (2) Technical solution:

[0016] To achieve the above objectives, the present invention chooses to perform the following steps.

[0017] 2.1 Step S1, Data Engineering:

[0018] At the current time step t, the TBM excavation parameter independent variable X(t) is collected through the TBM's own sensor system or control system, including the total thrust G composed by heat coding operation. t (unit kN), cutter head power C p (Unit: kW), cutter head torque C t (unit kN·m), cutterhead penetration C h (Unit mm·r -1 ), cutter head speed C s (Unit min·r -1 ), Field Penetration FPI (unit kN·cutter -1 ), torque penetration index TPI [unit (kN·m)·(mm·r -1 ) -1 ] and the current tunneling speed PR' (unit m·h -1 ); let the future excavation speed PR be the dependent variable Y; and construct the input sequence S with the independent variable X.

[0019] 2.1.1 Step S100, construct independent variable X(t):

[0020] At the current time step t, the independent variable X(t) collected by the TBM’s own sensor system or control system is: X(t)=[G t ,C p ,C t ,C h,C s ,FPI,TPI,PR′].

[0021] 2.1.2 Step S101, construct input sequence S:

[0022] The independent variable X(t) constitutes the input sequence S, which is represented as a set of independent variables at a series of time steps:

[0023] S=[X(1),X(2),…,X(t),…,X(T)];

[0024] Where T represents the total time step of the input sequence. The independent variable X(t) at each time step t contains the eight tunneling parameters mentioned above.

[0025] 2.2 Step S2, bidirectional Seq-Bi-GRU-DS fuzzy neural network operation:

[0026] To address the issue of unequal input and output lengths between the input sequence S and the dependent variable Y, a bidirectional Seq-Bi-GRU neural network is constructed by combining the Seq2Seq model with a bidirectional Bi-GRU neural network. This implements "encoding layer-decoding layer" processing to support unequal input and output sequence lengths. Furthermore, a fuzzy processing layer is added between the encoding and decoding layers to perform fuzzy operations on the hidden state sequence output by the encoding layer, forming a bidirectional Seq-Bi-GRU fuzzy neural network. This fuzzy processing reduces noise and uncertainty in the data and generates a corrected hidden state sequence. In the decoding layer, the DS evidence theory algorithm is used to calculate the joint confidence function, which can fuse multiple pieces of evidence (hidden states in the decoding layer) and calculate a comprehensive confidence level, thereby improving the confidence level of the prediction.

[0027] Therefore, the model of this step is a bidirectional Seq-Bi-GRU-DS fuzzy neural network as a whole. The process includes: the encoding layer receives the input sequence S, calculates the candidate hidden states of the forward and backward directions through multiple groups of forward and backward GRU (gated recurrent) units, and forms a hidden state sequence H, which is input to the fuzzy processing layer for fuzzy operation processing, and generates a modified hidden state sequence H' to the decoding layer; at the same time, a backtracking layer is added between the "encoding layer-decoding layer" to correct the hidden state sequence H' of the previous time step. t-1 , perform residual backtracking through the least squares method, use historical prediction errors for compensation, and generate an intermediate vector C; use the intermediate vector C as the input of the decoding layer;

[0028] The decoding layer is based on the intermediate vector C and the modified hidden state sequence H'. It also calculates the forward and backward candidate hidden states through multiple groups of forward and backward GRU units to form a prediction sequence S' of a set of dependent variables Y; the DS evidence theory algorithm is used to calculate the confidence of different dependent variables Y in the prediction sequence S', and the most credible final dependent variable Y' is further extracted.

[0029] 2.2.1 Step S200, coding layer processing:

[0030] For each time step t, the forward GRU unit receives the input sequence S and the hidden state of the previous time step , calculate the forward candidate hidden state for the current time step and the forward hidden state ;

[0031] Then similarly, for each time step t (from T to 1), the backward GRU unit receives the input sequence S and the hidden state of the next time step , calculate the backward candidate hidden state of the current time step and the backward hidden state ;

[0032] Finally, for each time step t, the forward hidden state and the backward hidden state Merge to form the final hidden state h(t).

[0033] 2.2.2.1 Forward calculation, backward calculation and their splicing:

[0034] (1) Forward hidden state : ;

[0035] (2) Backward hidden state : ;

[0036] (3) Hidden state sequence H: ;

[0037] in, , [·;·] represents vector concatenation operation.

[0038] 2.2.2.2 The internal data of the GRU unit is driven by:

[0039] (1) Update gate: ;

[0040] in, represents the sigmoid activation function, W z and U z is the weight matrix, bz is the bias vector; z t is the output of the update gate.

[0041] (2) Reset gate: ;

[0042] Among them, W r and U r is the weight matrix, b r is the bias vector; r t is the output of the reset gate.

[0043] (3) Candidate hidden states ;

[0044] Among them, tanh represents the hyperbolic tangent activation function, ⊙ represents element multiplication (Hadamard product), W h and U h is the weight matrix, b h is the bias vector.

[0045] (4) Actual hidden state .

[0046] 2.2.2 Step S201, fuzzy processing layer processing:

[0047] (1) Based on N fuzzy rules, each rule consists of a premise part (fuzzy set) and a conclusion part (linear transformation function); a fuzzy set is a set of "IF-THEN" rule statements:

[0048] ;

[0049] Among them, A i is a fuzzy set, F i is a linear transformation function.

[0050] (2) Fuzzy set A i is a fuzzy set defined on the hidden state space, including Gaussian fuzzy sets or triangular fuzzy sets. Gaussian fuzzy sets are:

[0051] ;

[0052] Among them, μ i and are the mean and standard deviation of the Gaussian fuzzy set respectively.

[0053] (3) Linear transformation: F i (h(t)) = W i h(t) + b i ;

[0054] Among them, W i is the weight matrix, bi is the bias vector.

[0055] (4) Fuzzy operation processing: For each time step t and each fuzzy rule R i , calculate h(t) belongs to fuzzy set A i The membership degree α i (t): ;

[0056] Then, using the membership degree α i (t) linear transformation F i (h(t)) is weighted averaged to obtain the corrected hidden state ;

[0057] Modify the hidden state sequence H' = [h'(1), h'(2), ..., h'(T)].

[0058] 2.2.3 Step S202, backtracking layer processing:

[0059] (1) Calculate the residual: Let E t−1 is the prediction error of the previous time step, that is, the actual value and predicted value The differences between: ;

[0060] (2) Least squares backtracking: Use the least squares method to fit the residuals and obtain the residual backtracking matrix W e , such that: ;

[0061] Where W is the backtracking matrix, ∥⋅∥ 2 is the square of the 2-norm of the vector.

[0062] Solving for W e Available through: ;

[0063] Among them, ┳ represents the transpose operation.

[0064] (3) Using the obtained residual backtracking matrix W e and the modified hidden state sequence H′ t−1 , generate the intermediate vector C: .

[0065] (4) Take the intermediate vector C as the input of the decoding layer: Combine the intermediate vector C with the encoding layer output H of the current time step t Combined as the input of the decoding layer. Here, the combination method can be concatenation method: H′′ t =f(H t ,C);

[0066] Where f represents the splicing operation function: H′′t=[H t ;C].

[0067] 2.2.4 Step S203, decoding layer processing:

[0068] (1) Similar to step S200, it also includes the calculation of the forward GRU unit, the calculation of the backward GRU, and the splicing operation. The difference is that this step is a decoding operation, while S200 is an encoding operation:

[0069] ;

[0070] ;

[0071] ;

[0072] in, represents the candidate hidden state computed by the forward GRU unit at time step t. → indicates that this is a forward (i.e., from left to right, or in chronological order) computation. h′ is the representation of the hidden state. f represents the output of the forward GRU. t is the index of the time step. GRU f Represents the data-driven operation of the forward GRU unit (see step S200 for details). represents the candidate hidden state calculated by the forward GRU unit at time step t−1. It is the hidden state of the previous time step of the current time step t and is used to pass information in the forward GRU. Represents the candidate hidden state calculated by the backward GRU unit at time step t. b Represents the data-driven operation of the backward GRU unit (see step S200 for details). Represents the candidate hidden state calculated by the backward GRU unit at time step t+1.

[0073] (2) Use hidden state representation h' t And the intermediate vector C generates the predicted sequence S' of the dependent variable Y, and a fully connected layer NN can be used: ;

[0074] S' is a sequence containing the predicted values ​​of multiple dependent variables Y: S' = [Y'1, Y'2, ..., Y' N ].

[0075] (3) Use the DS evidence theory algorithm to calculate the confidence of different dependent variables Y in the prediction sequence S': that is, for each dependent variable Y' i , calculate its confidence Bel(Y' i ):

[0076] (3.1) Identification framework Θ and its propositions: The identification framework Θ is defined as the set of all possible tunneling speed ranges. For example, the tunneling speed is divided into several intervals, each representing a proposition:

[0077] Let Θ = {Y1, Y2, ..., Y n}, where Y i The proposition that the excavation speed falls within the i-th interval.

[0078] (3.2) Execute the basic probability assignment mechanism: a method of assigning confidence to each proposition based on the available evidence. Here, we can assign a basic probability based on the information about the tunneling speed provided by the TBM sensor data, that is, calculate each proposition Y i The basic probability distribution m(Y i ). This basic probability is expressed as the likelihood or confidence that the tunneling speed falls within the interval:

[0079] Considering the existence of subset structure, and our essential idea is to calculate the confidence more accurately, we need to consider all the sets containing Y i The basic probability distribution of the subset of ′. Let P(Y i ) represents proposition Y i The set of all subsets of (including Y i itself), then the confidence Bel(Yi ′ ) is calculated as:

[0080] ;

[0081] Based on this idea, we can further expand on the framework of the DS evidence theory algorithm: not only is it limited to the information about the tunneling speed provided by the TBM sensor data, but we can also further obtain two sources of evidence based on external detection of the TBM tunneling speed (such as underground radar). Then, we need to use the Dempster combination rule to allow us to combine the basic probability distributions from different evidence sources to obtain a new, comprehensive basic probability distribution. This comprehensive basic probability distribution can then be used to calculate the confidence level. ;

[0082] Among them, m(Y i ) is the combined basic probability distribution. m1(A) and m2(B) are the basic probability distributions of the two evidence sources (the TBM sensor data and the external detection of the TBM's tunneling speed). A and B are propositions in the identification framework Θ, which come from the two evidence sources respectively. K is a normalization constant used to ensure that the sum of the combined basic probability distributions is 1: ;

[0083] Among them, Ø represents the empty set; represents the sum of the products of the basic probability assignments of all pairs of propositions that have no intersection.

[0084] Once the combined basic probability distribution m(Y i ), we can use it to calculate the confidence Bel(Y i ′). If the predicted value Y i ′ directly corresponds to a proposition Y i (depends on the number of propositions, that is, if conditions permit, the more propositions the better), then the confidence Bel(Y i ′) can be simply represented by m(Y i ) gives: Bel(Y' i ) = m(Y i );

[0085] If the predicted value Y i ′ does not correspond to a proposition Y i , then it must also be in Y i-1 and Y i+1 In the interval, we can determine whether it belongs to Y according to its trend. i-1 or Y i+1 .

[0086] Furthermore, by combining information from different sources, we can reduce the potential errors or uncertainties inherent in a single source of evidence, thereby improving the accuracy of advance rate predictions. In complex or uncertain environments, a single source of evidence may not provide sufficiently reliable information. By combining multiple sources of evidence, we can make the system more robust and better able to cope with various uncertainties and changes. Accurate advance rate predictions are crucial for TBM operation and maintenance. By combining information from multiple sources, we can provide decision makers with more comprehensive and accurate data support, helping them make more informed decisions.

[0087] (4) Based on the confidence level, extract the most credible final dependent variable Y', that is, the dependent variable with the highest confidence level: .

[0088] 2.3 Step S3, execution control:

[0089] Based on the difference and direction between the final dependent variable Y' and the current advance speed PR', a PID controller system is used to calculate and derive an electrical signal to control the TBM's advance speed and thereby control the TBM.

[0090] Error e(t) = Y' set -PR';

[0091] Integral term: I(t) = I(t-1) + e(t)·△t;

[0092] Differential term: ;

[0093] Proportional term: P(t) = K p ·e(t);

[0094] Control signal u(t) = P(t) +K i I(t) + K d D(t);

[0095] Among them, K p is the proportional gain; K i is the integral gain; K d is the differential gain; t is the time gain. The set value Y' is the target tunneling speed; I(t) is the integral term value at the current moment; I(t-1) is the integral term value at the previous sampling moment; e(t-1) is the error at the previous sampling moment; P(t) and D(t) are the differential term and proportional term values ​​at the current moment.

[0096] (3) Mechanism for resolving technical issues:

[0097] 3.1 Avoid inaccurate estimation when bootstrap sampling technology is combined with random distribution to generate prediction intervals:

[0098] The traditional bootstrap sampling technique combined with the random distribution method to generate prediction intervals often relies on a large number of repeated sampling and simulations to estimate the distribution of predicted values. However, when faced with high-dimensional, nonlinear or time-dependent data, this method may not be able to accurately capture the true distribution of the data, resulting in inaccurate prediction interval estimates. The present invention constructs a bidirectional Seq-Bi-GRU-DS fuzzy neural network to directly learn the complex mapping relationship between the input sequence and the output sequence without repeated sampling and simulation. At the same time, the DS evidence theory algorithm is used to fuse multiple evidences and calculate the joint trust function, which can more accurately describe the uncertainty of the predicted value, thereby generating a more reliable prediction interval.

[0099] 3.2 Overcoming the limitations of BPNN neural networks in processing complex sequence data:

[0100] When processing complex sequence data, BPNNs (back-propagation neural networks) often struggle to capture temporal dependencies and long-term memory effects. However, the bidirectional Seq-Bi-GRU-DS fuzzy neural network employed in this paper, through forward and backward GRU units, can simultaneously consider both past and future information, effectively capturing temporal features in the data. Furthermore, the introduction of a fuzzy processing layer further reduces noise and uncertainty in the data, improving the model's robustness and predictive accuracy.

[0101] 3.3 Avoid local optimization and improve the confidence of the estimated BP value:

[0102] To prevent the model from falling into a local optimal solution, the present invention adds a backtracking layer between the "encoding layer and the decoding layer." The backtracking layer uses the least squares method to perform residual backtracking, compensating the current prediction based on historical prediction errors, thereby generating a more accurate intermediate vector as input to the decoding layer. This design enables the model to dynamically adjust the prediction strategy, avoiding premature convergence to a local optimal solution during training. At the same time, the application of the DS evidence theory algorithm also provides a higher confidence level for the inferred BP value, because it can integrate multiple pieces of evidence and calculate a comprehensive confidence level, thereby more comprehensively evaluating the uncertainty of the prediction results.

[0103] 3.4 Supports multi-variable data driving and input / output sequences of unequal length:

[0104] By constructing a bidirectional Seq-Bi-GRU-DS fuzzy neural network, this paper naturally supports the problem of unequal input and output sequence lengths. The design of the encoding and decoding layers enables the model to flexibly handle input sequences of varying lengths and generate output sequences of corresponding lengths. Furthermore, the multi-variable data-driven design enables the model to comprehensively consider the impact of various tunneling parameters on tunneling speed, improving prediction accuracy and generalization capabilities.

[0105] 3.5 Further improve the control effect of TBM:

[0106] Based on the advance speed predicted by the aforementioned model, the present invention uses a PID controller system to precisely control the TBM. Based on the difference and direction between the predicted and actual values, the PID controller calculates an electrical signal to control the TBM's advance speed and adjusts the advance parameters in real time to achieve the desired advance speed. This prediction-based control strategy not only improves advance efficiency but also enhances the stability and safety of the tunneling process.

[0107] Secondly, intelligent control system for cantilever tunneling:

[0108] The system includes a processor and a memory connected to the processor. The memory stores program instructions. When the program instructions are executed by the processor, the processor executes the intelligent control method described above. After the control electrical signal u(t) is obtained, it is input into the TBM control system, and the TBM control system controls the TBM's actuator to perform corresponding operations.

[0109] Compared with the prior art, the present invention has the following beneficial effects:

[0110] 1. Improving Prediction Accuracy: This method comprehensively considers multiple tunneling parameters, such as total thrust, cutterhead power, and torque, and uses heat encoding operations to construct a rich input sequence, providing comprehensive information for the model, thereby improving the accuracy of tunneling speed prediction. The bidirectional Seq-Bi-GRU neural network can capture the temporal dependencies and long-term memory effects in the data. Compared with traditional BPNN neural networks, it has stronger generalization capabilities and prediction accuracy when processing complex sequence data. A fuzzy processing layer reduces data noise, and the DS evidence theory algorithm integrates multiple pieces of evidence to calculate the joint trust, further improving the confidence and reliability of the prediction results.

[0111] Second, avoiding local optimization and improving robustness: The backtracking layer of the present invention uses the least squares method to perform residual backtracking, compensating the current prediction based on historical prediction errors. This prevents the model from falling into local optimal solutions and improves the robustness and stability of the prediction. Fuzzy operations are used to process the hidden state sequence output by the encoding layer, reducing uncertainty and noise in the data and enhancing the model's ability to handle abnormal data.

[0112] 3. Precise Control of TBM Advance Speed: This invention uses a PID controller to adjust advance parameters in real time based on the difference between predicted and actual advance speeds, achieving precise control of the TBM's advance speed and improving both efficiency and safety. The PID controller dynamically adjusts control strategies based on the predicted results, enabling the TBM to maintain a stable advance speed under varying geological conditions.

[0113] 4. Support for unequal input / output sequence lengths: The combination of the Seq2Seq architecture and the bidirectional Bi-GRU neural network adopted in this invention naturally supports the problem of unequal input and output sequence lengths, enabling the model to flexibly handle tunneling parameter sequences and tunneling speed sequences of different lengths.

[0114] 5. Improved Construction Efficiency: This invention reduces unnecessary downtime and energy consumption during excavation by accurately predicting and controlling excavation speed, thereby improving excavation efficiency. Accurate prediction and control help workers avoid unexpected situations during excavation, such as machine jams and overexcavation, reducing excavation costs and safety risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0115] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0116] Figure 1 Schematic diagram of the method flow of the present invention;

[0117] Figure 2 Schematic diagram of the composition of the input sequence S of the present invention;

[0118] Figure 3 Schematic diagram of the coding layer architecture of the present invention;

[0119] Figure 4 Schematic diagram of the GRU unit of the present invention;

[0120] Figure 5 This is a schematic diagram of the fuzzy processing layer architecture of the present invention;

[0121] Figure 6 This is a schematic diagram of the backtracking layer architecture of the present invention;

[0122] Figure 7 Schematic diagram of the decoding layer architecture of the present invention;

[0123] Figure 8 Schematic diagram of the data-driven approach of the DS evidence theory algorithm of the present invention;

[0124] Figure 9 It is a schematic diagram of the system composition of the present invention;

[0125] Figure 10 This is a comparison diagram of the test examples of the present invention;

[0126] Figure 11 This is a schematic diagram of the finite element site model of the test example of the present invention;

[0127] Figure 12 Schematic diagram for comparison of the first finite element simulation of the test example of the present invention;

[0128] Figure 13 Schematic diagram for comparison of the second finite element simulation of the test example of the present invention. DETAILED DESCRIPTION

[0129] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. The following description sets forth many specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0130] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0131] Example 1: Figure 1 As shown, this embodiment discloses an intelligent control method for cantilever tunneling, which specifically includes the following steps S1 to S3.

[0132] In this embodiment, if Figure 2 As shown, regarding step S1, data engineering: This step aims to collect and process key parameters in the TBM excavation process to build a model that supports multi-independent variable data driving and thus achieve precise control. At the current time step t, the independent variables X(t) of the TBM excavation parameters are collected through the TBM's own sensor system or control system, including the total thrust G composed by heat encoding operation. t (unit kN), cutter head power C p (Unit: kW), cutter head torque C t (unit kN·m), cutterhead penetration C h (Unit mm·r -1 ), cutter head speed C s (Unit min·r -1 ), Field Penetration FPI (unit kN·cutter -1 ), torque penetration index TPI [unit (kN·m)·(mm·r -1 ) -1 ] and the current tunneling speed PR' (unit m·h -1 ); let the future excavation speed PR be the dependent variable Y; the independent variable X constitutes the input sequence S;

[0133] Specifically, in step S100, construct the independent variable X(t): at the current time step t, the independent variable X(t) collected by the TBM's own sensor system or control system is:

[0134] X(t)=[G t ,Cp ,C t ,C h ,C s ,FPI,TPI,PR′];

[0135] Specifically, step S101, constructing an input sequence S: constructing the independent variable X(t) into an input sequence S, which is represented as a set of independent variables at a series of time steps: S=[X(1),X(2),…,X(t),…,X(T)];

[0136] Where T represents the total time step of the input sequence. The independent variable X(t) at each time step t contains the eight tunneling parameters mentioned above.

[0137] As you can see, by collecting multiple independent variables closely related to tunneling efficiency, this step constructs a comprehensive input feature space, providing a rich source of information for the model. This helps the model more accurately capture the complex relationships within the tunneling process, improving prediction accuracy. Arranging the independent variables in time-step order to form the input sequence S allows the model to learn the temporal dependencies within the tunneling process. This dependency is crucial for predicting future tunneling speed PR, as it reflects the changing trends in tunneling status over time. A longer input sequence S contains more historical information, helping the model capture the long-term memory effects of the tunneling process. This helps the model consider states further into the past when making predictions, improving prediction robustness.

[0138] It's important to note that in practical applications, the length of the input sequence S (i.e., the time step T) can be determined based on specific needs and data availability. Longer input sequences contain more historical information, helping the model more accurately capture the temporal dependencies and long-term memory effects of the data. However, this also increases the model's complexity and computational effort. Therefore, a trade-off must be struck between model accuracy and computational efficiency.

[0139] In this embodiment, step S2 aims to build a neural network model that can handle the problem of unequal lengths between input and output sequences while improving prediction confidence. To this end, we combine the Seq2Seq model with a bidirectional Bi-GRU neural network, introduce a fuzzy processing layer and the DS evidence theory algorithm, and construct a bidirectional Seq-Bi-GRU-DS fuzzy neural network.

[0140] In this scheme, regarding the construction of bidirectional Seq-Bi-GRU neural network:

[0141] (1) Encoding layer: The encoding layer receives the input sequence S and calculates the candidate hidden states for the forward and backward directions through multiple groups of GRU (gated recurrent units) in the forward and backward directions, respectively. These candidate hidden states take into account both the historical and future information in the sequence, forming a hidden state sequence H that contains rich temporal features.

[0142] (2) Decoding layer: The decoding layer makes predictions based on the hidden state sequence H output by the encoding layer. Similarly, multiple groups of forward and backward GRU units are used to calculate the candidate hidden states and form a set of prediction sequences S' of the dependent variable Y.

[0143] (3) Fuzzy processing layer: A fuzzy processing layer is introduced between the encoding layer and the decoding layer to perform fuzzy processing on the hidden state sequence H output by the encoding layer. Fuzzy processing can reduce noise and uncertainty in the data and generate a more robust modified hidden state sequence H'. This step helps improve the model's ability to handle complex data and improve the accuracy of predictions.

[0144] (4) Lookback layer and residual lookback: A lookback layer is added between the encoding layer and the decoding layer to correct the hidden state sequence H' of the previous time step. t-1 Perform residual backtracking. Calculate the historical prediction error using the least squares method and use this error to compensate for the corrected hidden state at the previous time step, generating an intermediate vector C. This step can use historical information to correct the current prediction, further improving the accuracy of the prediction.

[0145] (5) Application of the DS Evidence Theory Algorithm: In the decoding layer, the DS Evidence Theory algorithm is used to calculate the confidence of different dependent variables Y in the prediction sequence S'. DS Evidence Theory can integrate multiple pieces of evidence (i.e., the hidden states of the decoding layer) and calculate a comprehensive confidence level. Based on the confidence level, the most credible final dependent variable Y' is extracted from the prediction sequence S'. This step ensures that the prediction results output by the model have a high degree of confidence and reliability.

[0146] In general, this step implements the "encoding layer-decoding layer" processing by constructing a bidirectional Seq-Bi-GRU neural network, effectively supporting the problem of unequal lengths between input and output sequences. This enables the model to more flexibly handle complex data in the actual excavation process, improving the model's applicability and generalization capabilities. The fuzzy processing layer can reduce noise and uncertainty in the data and generate a more robust modified hidden state sequence. At the same time, the application of the DS evidence theory algorithm can fuse multiple pieces of evidence and calculate a comprehensive trust level, further improving the confidence level of the prediction. This makes the prediction results output by the model more reliable and helps guide actual excavation operations. The introduction of the backtracking layer enables the model to use historical prediction errors to correct current predictions. This step can reduce the prediction bias caused by the accumulation of historical errors and improve the accuracy and stability of the prediction.

[0147] Specifically, such as Figures 3 and 4 As shown, regarding step S200, encoding layer processing: the encoding layer is responsible for converting the input sequence S into a hidden state sequence H, providing a basis for subsequent calculations. This step achieves comprehensive encoding of the input sequence through forward and backward GRU units.

[0148] For each time step t, the forward GRU unit receives the input sequence S and the hidden state of the previous time step , calculate the forward candidate hidden state for the current time step and the forward hidden state ;

[0149] Then similarly, for each time step t (from T to 1), the backward GRU unit receives the input sequence S and the hidden state of the next time step , calculate the backward candidate hidden state of the current time step and the backward hidden state ;

[0150] Finally, for each time step t, the forward hidden state and the backward hidden state Merge to form the final hidden state h(t).

[0151] Specifically, regarding the forward calculation, backward calculation and their connection in step S200:

[0152] (1) Forward hidden state Sequence: For each time step t, the forward GRU unit receives the current element in the input sequence S and the hidden state of the previous time step, and calculates the forward candidate hidden state and forward hidden state of the current time step through the internal data-driven mechanism (including update gate, reset gate, calculation of candidate hidden state and actual hidden state): ;

[0153] (2) Backward hidden state :Similar to the forward calculation, but in the opposite direction. For each time step t (from T to 1), the backward GRU unit receives the current element in the input sequence S and the hidden state of the next time step, performs the same internal data-driven calculation, and obtains the backward candidate hidden state and the backward hidden state of the current time step: ;

[0154] (3) Hidden state sequence H: ;

[0155] in, , [·;·] represents the vector concatenation operation. T is the length of the input sequence.

[0156] Furthermore, the internal data of the GRU unit is driven by:

[0157] (1) Update gate: Calculate the update gate output z for the current time step t , which determines the current hidden state h t How much of the previous time step's hidden state h should be retained t-1 , and how many candidate hidden states of the current time step should be received : ;

[0158] in, represents the sigmoid activation function, W z and U z is the weight matrix, b z is the bias vector.

[0159] (2) Reset gate: Calculate the reset gate output r of the current time step t , which determines how much the hidden state ht-1 of the previous time step should be forgotten at the current time step: ;

[0160] Among them, W r and U r is the weight matrix, b r is the bias vector.

[0161] (3) Candidate hidden states: ;

[0162] Among them, tanh represents the hyperbolic tangent activation function, ⊙ represents element multiplication (Hadamard product), W h and U h is the weight matrix, b h is the bias vector.

[0163] (4) Actual hidden state: According to the output z of the update gate t , the hidden state h of the previous time step t-1 and the candidate hidden state at the current time step Perform linear combination to get the actual hidden state h t :

[0164] ;

[0165] As can be understood, through the forward and backward GRU units, the encoding layer is able to fully capture the temporal features of the input sequence S, achieving multi-independent data-driven implementation. This enables the model to more accurately understand the complex relationship between the input and output sequences, providing a solid foundation for subsequent prediction and control. The combination of the bidirectional encoding mechanism (forward and backward) in this step allows the model to simultaneously consider historical and future information, resulting in a more accurate inference of the BP value (or other relevant variables). This comprehensive information fusion helps improve prediction confidence, enabling the model to make more reliable decisions in complex tunneling environments. The update and reset gate mechanisms within the GRU units allow the model to dynamically adjust the degree of hidden state updates and forgetting. This flexibility helps the model avoid local optima during training, allowing it to find a more optimal global solution.

[0166] Specifically, such as Figure 5 As shown, regarding step S201, the fuzzy processing layer processes: in the intelligent control method for cantilever tunneling, the fuzzy processing layer further extracts and transforms information by fuzzy processing the hidden state sequence H output by the encoding layer to improve the confidence and accuracy of the inferred BP value.

[0167] (1) Define N fuzzy rules, each rule consists of a premise part (fuzzy set A i ) and the conclusion part (linear transformation function F i ). The fuzzy set Ai describes the fuzzy state that the hidden state h(t) may belong to, and the linear transformation function F i It defines how to transform h(t) to obtain the modified hidden state h'(t) when h(t) belongs to this fuzzy state. A fuzzy set is a set of "IF-THEN" rule statements:

[0168] ;

[0169] Among them, A i is a fuzzy set, F i is a linear transformation function.

[0170] (2) Fuzzy set A i It is a fuzzy set defined on the hidden state space, which is a Gaussian fuzzy set:

[0171] ;

[0172] Among them, μ i and are the mean and standard deviation of the Gaussian fuzzy set respectively.

[0173] (3) Linear transformation: F i (h(t)) = W i h(t) + b i ;

[0174] Among them, W i is the weight matrix, b i is the bias vector.

[0175] The linear transformation is a scaling, translation, or rotation operation on the original hidden state, depending on the weight matrix W. i and the bias vector b i The value of .

[0176] (4) Fuzzy operation processing: For each time step t and each fuzzy rule R i , calculate h(t) belongs to fuzzy set A i The membership degree α i (t): ;

[0177] Membership degree α i (t) represents the hidden state h(t) and the fuzzy set A i The degree of similarity or matching between them. i When (t) is close to 1, it means that h(t) belongs to the fuzzy set A to a large extent. i ; When αi(t) is close to 0, it means h(t) is consistent with the fuzzy set A i There is a big difference.

[0178] Then, using the membership degree α i (t) linear transformation F i (h(t)) is weighted averaged to obtain the corrected hidden state ; Modify the hidden state sequence H' = [h'(1), h'(2), ..., h'(T)].

[0179] The modified hidden state sequence H' is a further refinement and transformation of the original hidden state sequence H. It contains more fuzzy information and nonlinear features, providing a richer and more accurate information basis for subsequent prediction or control modules.

[0180] As can be understood, the fuzzy processing layer, by introducing fuzzy sets and linear transformation functions, provides a more detailed and comprehensive description of the hidden states. This description captures the fuzzy relationships and nonlinear characteristics between hidden states, thereby increasing the confidence level of the estimated BP value (or other related variables). Furthermore, the weighted average processing method reduces the error and uncertainty that may be introduced by a single rule. The fuzzy processing layer can process hidden state sequences H of arbitrary length from the encoding layer, and each fuzzy rule can be defined and adjusted according to actual needs. This enables the model to flexibly adapt to the multi-variable data-driven requirements of different tunneling environments and task requirements. The fuzzy processing layer, by introducing multiple fuzzy rules and weighted averaging, increases the diversity and robustness of the model. This design helps the model avoid local optimal solutions during training, thereby finding a better global solution. The modified hidden state sequence H' contains more fuzzy information and nonlinear characteristics, which are crucial for guiding TBM (tunnel boring machine) tunneling operations accurately and stably. Therefore, the introduction of the fuzzy processing layer helps improve TBM control effectiveness, making the tunneling process more efficient and safe.

[0181] Specifically, such as Figure 6 As shown, regarding step S202, backtracking layer processing: This step aims to support multi-variable data driving and improve the confidence of the estimated BP value by calculating residuals, backtracking fitting, generating intermediate vectors, and combining the vectors with the encoding layer output. It includes:

[0182] (1) Calculate the residual: Let E t−1 is the prediction error of the previous time step, that is, the actual value and predicted value The differences between: ;

[0183] The residual is the difference between the actual value and the predicted value, reflecting the accuracy of the model's prediction. During tunneling, due to the influence of various factors such as geological conditions and tunneling parameters, actual tunneling results often deviate from the predicted values. Therefore, calculating the residual is the basis for evaluating the model's prediction performance and conducting subsequent retrospective processing. By calculating the residual, deviations in the model's predictions can be promptly identified, providing an accurate data foundation for subsequent retrospective processing. Furthermore, the magnitude of the residual can be used as a key indicator for evaluating model performance.

[0184] (2) Least squares backtracking: Use the least squares method to fit the residuals and obtain the residual backtracking matrix W e , such that: ;

[0185] Where W is the backtracking matrix, ∥⋅∥ 2Is the square of the 2-norm of the vector. In the backtracking layer processing, the residual backtracking matrix W e It is used to describe the relationship between the residual and the modified hidden state sequence, so as to realize the backtracking processing of the residual.

[0186] Among them, solving W e Available through:

[0187] Among them, ┳ represents the transpose operation.

[0188] It is understandable that the least squares backtracking method can accurately fit the relationship between the residual and the modified hidden state sequence, thereby achieving precise backtracking of the residual. This helps improve the model's adaptability to complex changes during the tunneling process and avoid falling into local optimal solutions.

[0189] (3) Using the obtained residual backtracking matrix W e and the modified hidden state sequence H′ t−1 , generate the intermediate vector C: This vector contains information about the residual backpropagation and serves as input to the decoding layer, helping it better understand and process the complex changes during tunneling. The generation of the intermediate vector C provides the decoding layer with rich residual backpropagation information. This helps the decoding layer fully consider the impact of historical residuals when generating predictions, thereby improving the accuracy and confidence of the predictions.

[0190] (4) Take the intermediate vector C as the input of the decoding layer: Combine the intermediate vector C with the encoding layer output H of the current time step t Combined as the input of the decoding layer. Here, the combination method can be concatenation method: H′′ t =f(H t ,C);

[0191] Where f represents the splicing operation function: H′′t=[H t ;C];

[0192] It can be understood that by combining the intermediate vector C with the coding layer output H t Combining this as the input to the decoding layer fully leverages historical residual backtracking information and the encoding layer output information at the current time step, improving the accuracy and confidence of the predictions generated by the decoding layer. This combination also supports multi-independent variable data-driven models with unequal input and output sequence lengths, enabling the model to more flexibly handle complex changes during tunneling.

[0193] Specifically, such as Figure 7As shown in step S203, decoding layer processing: The decoding layer processing is a key step following the encoding layer and the backtracking layer. It is responsible for generating a prediction sequence for the dependent variable Y (such as tunneling speed, tunneling force, etc.) based on the hidden state information provided by the encoding layer and the residual backtracking information generated by the backtracking layer.

[0194] (1) The computation of the forward and backward GRU units in the decoding layer is similar to that of the encoding layer, but the difference is that the decoding layer generates predictions, while the encoding layer extracts features from the input sequence. The forward GRU unit processes information from left to right in chronological order, while the backward GRU unit processes information from right to left. The combination of the two can capture the bidirectional dependencies in the time series:

[0195] ;

[0196] ;

[0197] ;

[0198] in, represents the candidate hidden state computed by the forward GRU unit at time step t. → indicates that this is a forward (i.e., from left to right, or in chronological order) computation. h′ is the representation of the hidden state. f represents the output of the forward GRU. t is the index of the time step. GRU f Represents the data-driven operation of the forward GRU unit (see step S200 for details). represents the candidate hidden state calculated by the forward GRU unit at time step t−1. It is the hidden state of the previous time step of the current time step t and is used to pass information in the forward GRU. Represents the candidate hidden state calculated by the backward GRU unit at time step t. b Represents the data-driven operation of the backward GRU unit (see step S200 for details). Represents the candidate hidden state calculated by the backward GRU unit at time step t+1.

[0199] It can be understood that by combining forward and backward GRU units, the decoding layer is able to capture the bidirectional dependencies in the time series, thereby generating more accurate predictions. This bidirectional processing mechanism helps improve the model's ability to predict complex changes in the tunneling process.

[0200] (2) Use hidden state representation h' t And the intermediate vector C generates the predicted sequence S' of the dependent variable Y, and a fully connected layer NN can be used: ;

[0201] S' is a sequence containing the predicted values ​​of multiple dependent variables Y: S' = [Y'1, Y'2, ..., Y' N ];

[0202] Preferably, the fully connected layer is usually represented as a combination of a linear transformation plus a nonlinear activation function. In the decoding layer, the fully connected layer receives the hidden state representation h′ t and intermediate vector C as input, and outputs the predicted sequence S': S'=σ(W[h' t ;C]+b);

[0203] Where: W is the weight matrix, which is determined by the dimension of the input features (i.e., the concatenated hidden state representation and the intermediate vector) and the dimension of the output prediction sequence. [h′ t ;C] represents the hidden state h′ t and the intermediate vector C are concatenated into a long vector. b is the bias vector, which has the same dimension as the output prediction sequence. σ(⋅) is a nonlinear activation function, such as ReLU, sigmoid, or softmax, used to increase the nonlinear expression capability of the model.

[0204] It can be understood that the weight matrix W and the bias vector b jointly perform a linear transformation on the input features. The purpose of this step is to map the input features to a new feature space in order to better capture the linear relationship between input and output. The role of the nonlinear activation function σ(⋅) is to introduce nonlinear expression capabilities to the model. By concatenating the hidden state representation h′ t The fully connected layer can fuse features from different sources by combining the intermediate vector C. This fusion helps the model capture the complex dependencies in the input sequence, thereby improving the accuracy of prediction.

[0205] It can be understood that by combining the hidden state representation and the intermediate vector, the decoding layer can account for the impact of multiple independent variables (such as excavation parameters and geological conditions) on excavation performance, achieving multi-independent data-driven implementation. The intermediate vector C contains information about residual backpropagation, which helps the decoding layer consider the impact of historical residuals when generating predictions, thereby improving the confidence of the predictions. Because the decoding layer processes based on time steps, it can handle cases where the input and output sequences are of unequal length, making the model more flexible and adaptable to complex changes in the excavation process. By combining forward and backward GRU units and residual backpropagation information, the decoding layer can more comprehensively capture dependencies in the time series, thereby avoiding being trapped in local optimal solutions.

[0206] Further, such as Figure 8 As shown, the DS evidence theory algorithm is used to calculate the confidence of different dependent variables Y in the prediction sequence S': that is, for each dependent variable Y' i, calculate its confidence Bel(Y' i ):

[0207] (1) Identification framework Θ and its proposition: First, we need to define an identification framework Θ, which includes all possible tunneling speed ranges. By dividing the tunneling speed into several intervals, each interval represents a proposition Y i , indicating the possibility that the excavation speed falls within this interval:

[0208] Let Θ = {Y1, Y2, ..., Y n}, where Y i The proposition that the excavation speed falls within the i-th interval.

[0209] (2) Execute the basic probability assignment mechanism: a method of assigning confidence to each proposition based on the available evidence. Here, the basic probability can be assigned based on the information about the tunneling speed provided by the TBM sensor data, that is, calculating each proposition Y i The basic probability distribution m(Y i ). This basic probability is expressed as the likelihood or confidence that the tunneling speed falls within the interval:

[0210] Based on the TBM’s sensor data or other available evidence, we assign a basic probability m(Y i ), indicating the confidence that the tunneling speed falls within this interval. Considering the subset structure, we need to calculate the i The sum of the basic probability distributions of all subsets of , to obtain the confidence Bel(Y i When there are multiple sources of evidence (such as TBM sensor data and external detection data), we use Dempster's combination rule to combine the basic probability distributions of these evidence sources to obtain a comprehensive basic probability distribution.

[0211] Considering the existence of subset structure, and our essential idea is to calculate the confidence more accurately, we need to consider all the sets containing Y i The basic probability distribution of the subset of ′. Let P(Y i ) represents proposition Y i The set of all subsets of (including Y i itself), then the confidence Bel(Yi ′ ) is calculated as:

[0212] ;

[0213] Furthermore, based on this idea, we can further expand on the framework of the DS evidence theory algorithm: not only is the information about tunneling speed provided by TBM sensor data, but we can also further obtain two sources of evidence based on external detection of TBM tunneling speed (such as underground radar). Then, we need to use the Dempster combination rule to allow us to combine the basic probability distributions from different evidence sources to obtain a new, comprehensive basic probability distribution. This comprehensive basic probability distribution can then be used to calculate the confidence level:

[0214] ;

[0215] Among them, m(Y i ) is the combined basic probability distribution. m1(A) and m2(B) are the basic probability distributions of the two evidence sources (the TBM sensor data and the external detection of the TBM's tunneling speed). A and B are propositions in the identification framework Θ, which come from the two evidence sources respectively. K is a normalization constant used to ensure that the sum of the combined basic probability distributions is 1:

[0216] ;

[0217] Among them, Ø represents the empty set; represents the sum of the products of the basic probability assignments of all pairs of propositions that have no intersection.

[0218] (3) Once the basic probability distribution m(Y i ), we can use it to calculate the confidence Bel(Y i ′). If the predicted value Y i ′ directly corresponds to a proposition Y i (depends on the number of propositions, that is, if conditions permit, the more propositions the better), then the confidence Bel(Y i ′) can be simply represented by m(Y i ) gives: Bel(Y' i ) = m(Y i );

[0219] If the predicted value Y i ′ does not correspond to a proposition Y i , then it must also be in Y i-1 and Y i+1 In the interval, we can determine whether it belongs to Y according to its trend. i-1 or Y i+1 .

[0220] Furthermore, by combining information from different sources, we can reduce the potential errors or uncertainties inherent in a single source of evidence, thereby improving the accuracy of advance rate predictions. In complex or uncertain environments, a single source of evidence may not provide sufficiently reliable information. By combining multiple sources of evidence, we can make the system more robust and better able to cope with various uncertainties and changes. Accurate advance rate predictions are crucial for TBM operation and maintenance. By combining information from multiple sources, we can provide decision makers with more comprehensive and accurate data support, helping them make more informed decisions.

[0221] (4) Based on the confidence level, extract the most credible final dependent variable Y', that is, the dependent variable with the highest confidence level: ;

[0222] It can be understood that by selecting the predicted value with the highest confidence as the final dependent variable, we can ensure that control decisions are based on the most reliable information, thereby improving the control effectiveness of the TBM. Furthermore, this method can automatically adapt to different excavation conditions and environmental changes, enhancing the system's adaptability and stability. In summary, using the DS evidence theory algorithm to calculate the confidence levels of different dependent variables Y in the prediction sequence S' is a key step in the intelligent control method for cantilever tunneling. By combining information from multiple evidence sources, calculating confidence levels, and selecting the most reliable predicted value, we can more accurately assess the uncertainty of the excavation speed, provide decision makers with more comprehensive and accurate data support, and further improve the control effectiveness of the TBM.

[0223] In this embodiment, step S3 involves executing control. This step uses a PID (Proportional-Integral-Derivative) controller system to perform precise calculations based on the difference and direction between the final dependent variable Y' (i.e., the target advance rate) and the current advance rate PR'. This generates an electrical signal to control the advance rate of the TBM (Tunnel Boring Machine), thereby providing real-time control of the TBM.

[0224] Based on the difference and direction between the final dependent variable Y' and the current advance speed PR', a PID controller system is used to calculate and derive an electrical signal to control the TBM's advance speed and thereby control the TBM.

[0225] Error e(t) = Y' set -PR';

[0226] Integral term: I(t) = I(t-1) + e(t)·△t; where I(t-1) is the integral term value at the previous sampling time, e(t) is the current error, and △t is the time interval (time gain).

[0227] Differential term: ; The differential term can respond to possible changes in the system in advance and improve the response speed and stability of the control.

[0228] Proportional term: P(t) = K p e(t); The proportional term can respond quickly to error changes, but it may cause system oscillations and therefore needs to be used in combination with other control terms.

[0229] Control signal u(t) = P(t) +K i I(t) + K d D(t);

[0230] Among them, K p is the proportional gain; K i is the integral gain; K d is the differential gain; t is the time gain. The set value Y' is the target tunneling speed; I(t) is the integral term value at the current moment; e(t-1) is the error at the previous sampling moment; P(t) and D(t) are the differential term and proportional term values ​​at the current moment.

[0231] It's clear that by comprehensively considering the current error, the accumulation of past errors, and the changing trend of future errors, the PID controller can more precisely control the TBM's advance speed, reduce errors, and improve control accuracy. The integral term eliminates static errors in the system, improving its stability; the differential term predicts and suppresses system oscillations, further enhancing its stability.

[0232] Example 2: This example further provides a training method for the bidirectional Seq-Bi-GRU-DS fuzzy neural network as described in Example 1. The training parameters include:

[0233] (1) Reset the gate weight matrix W r and U r , and the bias vector b r ;

[0234] (2) Update the gate weight matrix W z and U z , and the bias vector b z ;

[0235] (3) Weight matrix W of candidate hidden states h and U h , and the bias vector b h ;

[0236] (4) Linear transformation weight matrix W i and the bias vector b i ;

[0237] (5) The weight matrix W and bias vector b of the fully connected layer.

[0238] Training methods include:

[0239] P1, create a data set D = [(X1, Y1), (X2, Y2), ..., (X n ,Y n )];

[0240] Among them, (X i ,Y i ) is the i-th training sample, n is the number of samples, where X i is the historical value of the i-th independent variable as described in Example 1, Y i It corresponds to X i The value of the dependent variable.

[0241] The dataset D is then divided into a training set and a test set, and the training set is used to perform subsequent training steps.

[0242] P2. Execute steps S200 to S203 as described in the first embodiment.

[0243] P3. Calculate loss: Use a suitable loss function (such as mean square error, cross entropy, etc.) to calculate the error between the predicted sequence S′ and the target sequence Y to obtain the loss value L. Based on the loss value L, calculate the gradient of the loss function with respect to each parameter using the chain rule.

[0244] For the GRU units in the encoding layer and decoding layer, the gradients of the update gate, reset gate, candidate hidden state, and actual hidden state with respect to the weight matrix and bias vector are calculated.

[0245] For the blurring layer, the gradients of the weight matrix and bias vector of the linear transformation are calculated.

[0246] For the regression layer, the gradient of the residual regression matrix is ​​calculated.

[0247] For fully connected layers, the gradients of the weight matrix and bias vector are calculated.

[0248] P4. Update parameters: Use the gradient descent algorithm (SGD) to update the values ​​of all weight matrices and bias vectors. For each parameter, update it according to the learning rate and the calculated gradient.

[0249] P5. Iterative training: Repeat the steps of forward propagation, loss calculation, backpropagation, and parameter update. Adjust the batch size based on the performance on the test set to optimize the model performance until the preset number of training rounds is reached or the loss value converges to a satisfactory level.

[0250] Furthermore, the Python execution program of this bidirectional Seq - Bi - GRU - DS fuzzy neural network is as follows:

[0251] import torch

[0252] import torch.nn as nn

[0253] import torch.optim as optim

[0254] class BiGRUCell(nn.Module):

[0255] def __init__(self, input_size, hidden_size):

[0256] super(BiGRUCell, self).__init__()

[0257] self.hidden_size = hidden_size

[0258] self.reset_gate = nn.Linear(input_size + hidden_size, hidden_size)

[0259] self.update_gate = nn.Linear(input_size + hidden_size,hidden_size)

[0260] self.candidate_state = nn.Linear(input_size + hidden_size,hidden_size)

[0261] self.sigmoid = nn.Sigmoid()

[0262] self.tanh = nn.Tanh()

[0263] def forward(self, x, hidden):

[0264] r_t = self.sigmoid(self.reset_gate(torch.cat([x, hidden], dim = 1)))

[0265] z_t = self.sigmoid(self.update_gate(torch.cat([x, hidden],dim=1)))

[0266] h_hat_t = self.tanh(self.candidate_state(torch.cat([x, r_t *hidden], dim=1)))

[0267] h_t = (1 - z_t) * hidden + z_t * h_hat_t

[0268] return h_t

[0269] class FuzzyLayer(nn.Module):

[0270] def __init__(self, hidden_size, num_rules):

[0271] super(FuzzyLayer, self).__init__()

[0272] self.hidden_size = hidden_size

[0273] self.num_rules = num_rules

[0274] self.linear_transforms = nn.ModuleList([nn.Linear(hidden_size, hidden_size) for _ in range(num_rules)])

[0275] self.means = nn.Parameter(torch.randn(num_rules, hidden_size))

[0276] self.stds = nn.Parameter(torch.rand(num_rules, hidden_size))

[0277] def forward(self, H):

[0278] T, hidden_size = H.size()

[0279] H_prime = torch.zeros_like(H)

[0280] for t in range(T):

[0281] memberships = torch.exp(-torch.sum((H[t] - self.means) **2 / (2 * self.stds ** 2), dim=1))

[0282] weighted_sum = torch.zeros(hidden_size)

[0283] for i in range(self.num_rules):

[0284] weighted_sum += memberships[i] * self.linear_transforms[i](H[t])

[0285] H_prime[t] = weighted_sum / torch.sum(memberships)

[0286] return H_prime

[0287] class Seq2SeqBiGRUDS(nn.Module):

[0288] def __init__(self, input_size, hidden_size, output_size, num_rules):

[0289] super(Seq2SeqBiGRUDS, self).__init__()

[0290] self.encoder_forward = BiGRUCell(input_size, hidden_size)

[0291] self.encoder_backward = BiGRUCell(input_size, hidden_size)

[0292] self.fuzzy_layer = FuzzyLayer(hidden_size * 2, num_rules)

[0293] self.decoder_forward = BiGRUCell(hidden_size * 2, hidden_size)

[0294] self.decoder_backward = BiGRUCell(hidden_size * 2, hidden_size)

[0295] self.fc = nn.Linear(hidden_size * 2, output_size)

[0296] def forward(self, S):

[0297] T = S.size(1)

[0298] h_forward = torch.zeros(S.size(0), self.encoder_forward.hidden_size).to(S.device)

[0299] h_backward = torch.zeros(S.size(0), self.encoder_backward.hidden_size).to(S.device)

[0300] H_forward = []

[0301] H_backward = []

[0302] for t in range(T):

[0303] h_forward = self.encoder_forward(S[:, t], h_forward)

[0304] H_forward.append(h_forward)

[0305] for t in reversed(range(T)):

[0306] h_backward = self.encoder_backward(S[:, t], h_backward)

[0307] H_backward.append(h_backward)

[0308] H = torch.stack([torch.cat([H_forward[t], H_backward[T - 1 - t]], dim = 1) for t in range(T)], dim = 1)

[0309] H_prime = self.fuzzy_layer(H)

[0310] C = torch.zeros_like(H_prime[:, 0])

[0311] decoder_outputs = []

[0312] for t in range(T):

[0313] h_decoder_forward = self.decoder_forward(torch.cat([H_prime[:, t], C], dim = 1), h_forward)

[0314] h_decoder_backward = self.decoder_backward(torch.cat([H_prime[:, t], C], dim = 1), h_backward)

[0315] h_decoder = torch.cat([h_decoder_forward, h_decoder_backward], dim = 1)

[0316] decoder_outputs.append(self.fc(h_decoder))

[0317] return torch.stack(decoder_outputs, dim = 1)

[0318] The training module is as follows:

[0319] def train_model(model, train_loader, criterion, optimizer, num_epochs):

[0320] for epoch in range(num_epochs):

[0321] model.train()

[0322] total_loss = 0

[0323] for S, Y in train_loader:

[0324] S, Y = S.to(device), Y.to(device)

[0325] optimizer.zero_grad()

[0326] outputs = model(S)

[0327] loss = criterion(outputs, Y)

[0328] loss.backward()

[0329] optimizer.step()

[0330] total_loss += loss.item()

[0331] print(f'Epoch [{epoch+1} / {num_epochs}], Loss: {total_loss / len(train_loader):.4f}')

[0332] # Parameter settings

[0333] input_size = 10

[0334] hidden_size = 50

[0335] output_size = 1

[0336] num_rules = 5

[0337] num_epochs = 20

[0338] learning_rate = 0.001

[0339] # Initialize the model, loss function, and optimizer

[0340] model = Seq2SeqBiGRUDS(input_size, hidden_size, output_size, num_rules).to(device)

[0341] criterion = nn.MSELoss()

[0342] optimizer = optim.Adam(model.parameters(), lr=learning_rate)

[0343] # Train the model

[0344] train_model(model, train_loader, criterion, optimizer, num_epochs)

[0345] In the above program, the forward and backward Bi-GRU units are used to process the input sequence S, respectively, generating the forward hidden state sequence H_forward and the backward hidden state sequence H_backward. The forward and backward hidden states are concatenated to form the final hidden state sequence H.

[0346] Fuzzy processing layer: Apply fuzzy rules to the hidden state sequence H and calculate the membership of the hidden state h(t) at each time step t to each fuzzy set. Use the membership to perform a weighted average of the linear transformation results to generate a modified hidden state sequence H'.

[0347] Backtracking layer: Calculate the prediction error of the previous time step and use the least squares method to fit the residual to obtain the residual backtracking matrix W e . Using W e And the modified hidden state sequence H' generates the intermediate vector C.

[0348] Decoding layer: Use forward and backward Bi-GRU units to process the modified hidden state sequence H' and the intermediate vector C to generate the hidden state of the decoding layer. The hidden state of the decoding layer is converted into the predicted sequence S' of the dependent variable Y through the fully connected layer.

[0349] The training process uses the mean squared error (MSE) as the loss function and updates the model parameters through the Adam optimizer. In each epoch, the training data is forward propagated, the loss is calculated, and then backpropagated and the parameters are updated.

[0350] Example 3: In Example 1, when a longer input sequence is present in step S101, it may contain more historical information, helping the model more accurately capture the temporal dependencies and long-term memory effects of the data; however, this also increases the model's complexity and computational effort. Therefore, a trade-off must be made between model accuracy and computational efficiency. To this end, this example provides an optimization solution for step S101, namely, introducing a dynamic adjustment mechanism to dynamically adjust the length of the input sequence S (i.e., the time step T):

[0351] P1. Calculate the energy function e ij : ;

[0352] Among them, q j k represents the query vector of the jth element (i.e., independent variable) in the input sequence S; i Represents the key vector of the i-th element in the input sequence S; d k is the key vector k i ; ⋅ represents the vector dot product operation.

[0353] The dot product attention is calculated by computing the query vector q j and key vector k i The larger the dot product result, the more similar the two vectors are.

[0354] P2. Calculate the weight factor p ij : ;

[0355] Among them, the weight factor p ij represents the attention weight of the i-th element in the input sequence S to the j-th element in the output sequence; T is the total length of the input sequence. k is the index.

[0356] P3, according to the weight factor p ij The size of the input elements that have little impact on the output is selectively ignored, thereby shortening the actual processing length of the input sequence. If a threshold is set, all weight factors p less than the threshold are ignored. ij The element (i.e., independent variable) in the input sequence S corresponding to is removed.

[0357] It's understandable that by ignoring input elements with minimal impact on the output, the model can reduce the amount of data it needs to process, thereby lowering computational complexity and improving efficiency. Dynamic adjustment ensures that the model can still capture key information when processing shortened input sequences, thereby maintaining model accuracy. This is because the dynamic adjustment mechanism dynamically adjusts attention weights, highlighting and helping us further eliminate input elements that have a significant impact on the output. Because the dynamic adjustment mechanism can dynamically focus on different parts of the input sequence, it makes the model more adaptable to input sequences of varying lengths. This allows the model to maintain good performance when processing input data with different time steps T.

[0358] Example 4: In step S201 of Example 1, A iis a fuzzy set. This embodiment further defines a geological condition fuzzy set for the specific working environment of the TBM, including the following three types:

[0359] (1) Soft rock fuzzy set A soft (h(x)): Soft rock fuzzy set is used to describe the situation where the geological conditions encountered during tunneling are relatively soft and the rock strength is low. It is defined as a Gaussian fuzzy set, where the mean μ soft Represents the typical soft rock strength value, standard deviation σ soft Reflects the fluctuation range of rock strength:

[0360] ;

[0361] Among them, the mean μ soft Based on historical excavation data or geological exploration results, a typical soft rock strength value is determined as the mean. This value represents the central trend of the soft rock strength. Standard deviation σ soft This value reflects the fluctuation range of soft rock strength, that is, the degree of dispersion of rock strength around the mean. A larger standard deviation indicates a wider fluctuation range of rock strength. The Gaussian function is used as the membership function for the fuzzy set because it is smooth, continuous, and differentiable, effectively describing the fuzziness of rock strength within the soft rock range.

[0362] (2) Hard rock fuzzy set A hard (h(x)): The hard rock fuzzy set is used to describe the hard geological conditions encountered during tunneling, where the rock strength is high. It is also defined as a Gaussian fuzzy set, but the mean and standard deviation are adjusted to reflect the characteristics of hard rock:

[0363] ;

[0364] Among them, the mean μ hard Based on historical excavation data or geological exploration results, a typical hard rock strength value is determined as the mean. This value represents the central trend of hard rock strength and is usually larger than the mean value of soft rock. Standard deviation σ hard This reflects the fluctuation range of hard rock strength. Because hard rock strength is generally stable, its standard deviation may be relatively small. A Gaussian function is also used as the membership function of the fuzzy set to describe the fuzziness of rock strength within the hard rock range.

[0365] (3) Mixed rock fuzzy set A mixed (h(x)): This is used to describe the complex geological conditions and the mixture of soft and hard rocks during tunneling. Due to the complex geological conditions of mixed rock, a single Gaussian fuzzy set may not accurately describe its characteristics. Therefore, this embodiment considers combining multiple Gaussian fuzzy sets or using more complex fuzzy set shapes (such as trapezoidal fuzzy sets) to define:

[0366] ;

[0367] Among them, A i (x) represents the i-th Gaussian fuzzy set (which can be a soft rock fuzzy set or a hard rock fuzzy set), w i represents the weight of the i-th fuzzy set, and .

[0368] Or, using the trapezoidal fuzzy set definition:

[0369] ;

[0370] Among them, a, b, c, and d are the four parameters of the trapezoidal fuzzy set, representing the starting value of the mixed rock strength range, the transition value between soft rock and hard rock, the typical value of hard rock, and the end value of the mixed rock strength range. By combining multiple Gaussian fuzzy sets, the geological conditions of mixed rocks can be described more flexibly. Each Gaussian fuzzy set can represent rocks of different strength ranges in the mixed rock, and the weight w i This reflects the relative importance of these different intensity ranges in migmatites.

[0371] A trapezoidal fuzzy set is a simpler fuzzy set shape that uses four parameters to define a trapezoidal region. Among mixed rock fuzzy sets, the trapezoidal fuzzy set can well describe the transition of rock strength from soft rock to hard rock, as well as the starting and ending values ​​of the mixed rock strength range.

[0372] Through the above definitions, geological condition fuzzy sets can be used to more accurately describe and handle different geological conditions during the excavation process, thereby improving the intelligent control effect of cantilever tunneling.

[0373] Example 5: In step S203 of Example 1, calculate each proposition Y i The basic probability distribution m(Yi) reflects the probability distribution of proposition Y i The method is to:

[0374] P1. Define the constraints of m(Yi):

[0375] (1) The basic probability assignment for the empty set is 0: m(∅) = 0. This means there is no belief in the empty set (i.e., impossible events).

[0376] (2) The sum of the basic probability distributions of all propositions is 1: ∑ m(Yi) = 1, where Y i Iterate over all propositions in the identification frame Θ. This means that the sum of the confidences of all possible answers is 1.

[0377] P2. Specific allocation: For each proposition Y i ,m(Yi ) is assigned a value based on the available evidence. Sensor data is analyzed and processed to assess the likelihood that the advancement rate falls within the interval represented by the proposition. This involves collecting sensor data from the TBM and preprocessing it to extract information related to advancement rate. Based on the selected features, an allocation mechanism is established to calculate the probability of each proposition Y using statistical methods. i The basic probability distribution m(Y i ):

[0378] (1) For each proposition Y i , calculate the frequency of tunneling speed falling within this interval in the sensor data. This can be achieved by counting the number of data points falling within this interval and dividing it by the total number of data points.

[0379] (2) Basic probability distribution: m(Yi) = frequency(Yi)

[0380] Among them, frequency (Yi) represents the frequency at which the tunneling speed falls within the interval represented by proposition Yi.

[0381] (3) Normalization: Since frequency estimation may result in the sum of the basic probability distributions of all propositions not being 1, normalization is required. This is achieved by dividing the basic probability distribution of each proposition by the sum of the basic probability distributions of all propositions:

[0382] m'(Y i ) = m(Y i ) / ∑ m(Yj) ;

[0383] Among them, m'(Y i ) represents the normalized basic probability distribution, ∑m(Y j ) represents the sum of the basic probability distributions of all propositions.

[0384] For example, the following sensor data for the tunneling speed are provided (unit: m / min):

[0385] 12, 18, 25, 10, 20, 30, 15, 22, 10, 28

[0386] The tunneling speed is divided into the following three intervals and the corresponding propositions are defined:

[0387] Y1: [0, 15];

[0388] Y2: (15, 25];

[0389] Y3: (25, 35];

[0390] Now calculate the basic probability distribution of each proposition according to the above algorithm:

[0391] The data points that fall within the Y1 interval are: 12, 10, 10, a total of 3, with a frequency of 3 / 10;

[0392] The data points that fall within the Y2 interval are: 18, 20, 15, 22, a total of 4, with a frequency of 4 / 10;

[0393] The data points that fall within the Y3 interval are: 25, 30, and 28, a total of 3, with a frequency of 3 / 10;

[0394] Basic probability distribution:

[0395] m(Y1) = 3 / 10;

[0396] m(Y2) = 4 / 10;

[0397] m(Y3) = 3 / 10;

[0398] Normalization processing:

[0399] m'(Y1) = (3 / 10) / (3 / 10 + 4 / 10 + 3 / 10) = 3 / 10;

[0400] m'(Y2) = (4 / 10) / (3 / 10 + 4 / 10 + 3 / 10) = 4 / 10;

[0401] m'(Y3) = (3 / 10) / (3 / 10 + 4 / 10 + 3 / 10) = 3 / 10;

[0402] Example 6: This example further discloses an intelligent control system for cantilever tunneling as the hardware carrier of all the above examples, such as Figure 9 As shown, the system includes a processor and a memory connected to the processor. The memory stores program instructions. When the program instructions are executed by the processor, the processor executes the intelligent control method described above. After the control electrical signal u(t) is obtained, it is input into the TBM control system, and the TBM control system controls the TBM's actuator to perform corresponding operations.

[0403] It's important to note that the TBM control system drives the TBM's actuators (such as the thrust cylinder, cutterhead motor, and steering system) based on received control signals, enabling precise tunneling operations. Intelligent control systems also incorporate a feedback mechanism that continuously monitors the TBM's actual operating status through sensors and feeds this information back to the processor, creating a closed-loop control system that further improves control accuracy and stability. Modern intelligent control systems often also support remote monitoring, allowing operators to monitor the tunneling process remotely via the internet from a safe location, adjust control strategies promptly, and even intervene remotely when necessary.

[0404] Test example:

[0405] 1. Experimental purpose:

[0406] This experiment aims to verify the superiority of the bidirectional Seq-Bi-GRU-DS fuzzy neural network in intelligent control of tunnel excavation by comparing the prediction accuracy and soil stress effect of the bidirectional Seq-Bi-GRU-DS fuzzy neural network (experimental group) and the traditional BPNN neural network (control group) when used in conjunction with a TBM control system for tunnel excavation.

[0407] (II) Experimental and control group settings:

[0408] 2.1 Experimental Group

[0409] Control method: the bidirectional Seq-Bi-GRU-DS fuzzy neural network described in Examples 1 to 5;

[0410] Its characteristics include supporting unequal lengths of input and output sequences, being able to accurately describe the entire confidence interval distribution, and optimizing the evaluation of the PR value of the tunneling speed.

[0411] 2.2 Control group:

[0412] Control method: traditional BPNN neural network;

[0413] Its characteristic is that it does not support unequal lengths of input and output sequences, and its application effect is limited in complex and changeable construction environments.

[0414] (3) Experimental methods:

[0415] Simulation environment: Create shield tunneling scenarios in ANSYS finite element simulation software such as Figure 10 As shown, the shield (8600mm diameter) is traversing a full-face sandy, gravelly, and muddy stratum (soil pressure range: 180-200kP). The shield is configured with a 440-ring cutterhead torque of 2200-2800kN, a soil chamber pressure of 1.2-1.5bar, and a propulsion speed of 50-70mm / min. The shield thrust is 10,000kN to 30,000kN.

[0416] Both the experimental and control groups performed predictions according to the above parameters, outputting the predicted mileage, which was then compared with the simulated mileage to determine the accuracy of the predictions.

[0417] After the predicted mileage comparison process was completed, control parameters for the experimental and control groups were configured in the simulation scenario to control the TBMs. The stress effects of the soil layers in each group were simulated, including stress analysis of the upper and lower soil layers and excavation stress analysis.

[0418] (IV) Experimental results:

[0419] 4.1 Prediction Accuracy

[0420] like Figure 10 As shown in the figure, the prediction accuracy of the experimental group is significantly better than that of the control group. The experimental group can more accurately predict the advancement distance of the shield machine, which is more consistent with the simulation results.

[0421] 4.2 Stress effect:

[0422] Stress analysis of upper and lower soil layers ( Figure 12 ): The stress generated by the experimental group (Part B) was significantly smaller than that of the control group (Part A), indicating that the experimental group performed better in controlling soil stress.

[0423] Tunneling stress analysis ( Figure 13 ): Similarly, the stress generated by the experimental group (Part B) is also significantly smaller than that of the control group (Part A), further verifying the stress control effect of the experimental group during the excavation process.

[0424] (V) Results Analysis

[0425] The reason why the experimental group performed better than the control group in terms of prediction accuracy and stress control is mainly due to the following advantages of the bidirectional Seq-Bi-GRU-DS fuzzy neural network:

[0426] The bidirectional Seq-Bi-GRU-DS fuzzy neural network accurately describes the entire confidence interval distribution in advance, allowing the model to more accurately estimate the advance rate (PR) during statistical analysis. Based on this accurately estimated PR, the model can indirectly optimize soil stress and reduce soil disturbance during excavation.

[0427] Unlike traditional BPNN neural networks, the bidirectional Seq-Bi-GRU-DS fuzzy neural network supports unequal lengths of input and output sequences, which enables the model to more effectively handle differences in input data dimensions and quantity caused by changes in geological conditions, adjustments to tunnel boring machine operating strategies, etc.

[0428] All of the above embodiments merely represent implementation methods of the present invention in practical applications. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the scope of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the appended claims.

[0429] For those skilled in the art, it can be further appreciated that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0430] At the same time, those skilled in the art will understand that all or part of the processes in all the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

Claims

1. An intelligent control method for cantilever tunneling, characterized in that: Including at the current time step t, perform the following steps: S1, Data Engineering: Collect independent variables X(t) of TBM excavation parameters, including total thrust G t , Cutter head power C p , cutter head torque C t , Cutter penetration C h , Cutter head speed C s , field penetration index FPI, torque penetration index TPI and current advance speed PR'; let the future advance speed PR be the dependent variable Y; the independent variable X constitutes the input sequence S; S2, bidirectional Seq-Bi-GRU fuzzy neural network operation: the encoding layer receives the input sequence S, calculates the candidate hidden states in the forward and backward directions through multiple groups of forward and backward GRU units, and forms a hidden state sequence H, which is input into the fuzzy processing layer for fuzzy operation processing, and generates a modified hidden state sequence H' to the decoding layer; based on the modified hidden state sequence H', the decoding layer calculates the candidate hidden states in the forward and backward directions through multiple groups of forward and backward GRU units to form a predicted sequence S' of the dependent variable Y; the DS evidence theory algorithm is used to calculate the confidence of different dependent variables Y in the predicted sequence S', and the most credible final dependent variable Y' is extracted; S3, execute control: Based on the difference and direction between the final dependent variable Y' and the current tunneling speed PR', the PID controller calculates the control electrical signal u(t) that controls the TBM tunneling speed at the current time step t and controls the TBM.

2. The intelligent control method according to claim 1, characterized in that: In S1, the independent variable X(t)=[G t ,C p ,C t ,C h ,C s ,FPI,TPI,PR′]; The input sequence S=[X(1),X(2),…,X(t),…,X(T)]; Where T represents the total number of time steps in the input sequence.

3. The intelligent control method according to claim 1, characterized in that: In S2, the data-driven process of the bidirectional Seq-Bi-GRU fuzzy neural network includes: Encoding layer processing: For each time step t, the forward GRU unit receives the input sequence S and the hidden state of the previous time step , calculate the forward candidate hidden state for the current time step and the forward hidden state ; For each time step t, the backward GRU unit receives the input sequence S and the hidden state of the next time step , calculate the backward candidate hidden state of the current time step and the backward hidden state ; For each time step t, the forward hidden state and the backward hidden state Merge to form the final hidden state h(t); Fuzzy processing layer processing: at each time step t and each fuzzy rule R i , calculate h(t) belongs to fuzzy set A i The membership degree α i (t); using membership α i (t) linear transformation F i (h(t)) is weighted averaged to obtain a modified hidden state and a modified hidden state sequence H' = [h'(1), h'(2), ..., h'(T)]; where T is the length of the input sequence S; Decoding layer processing: including the calculation of the forward GRU unit, the calculation of the backward GRU and the decoding operation composed of the splicing operation, using the hidden state representation h' t The prediction sequence S' of the dependent variable Y is generated by the intermediate vector C; the confidence of different dependent variables Y in the prediction sequence S' is calculated using the DS evidence theory algorithm, and the most credible final dependent variable Y' is extracted.

4. The intelligent control method according to claim 3, characterized in that: A backtracking layer is also included between the encoding layer and the decoding layer, and the backtracking layer is used to modify the hidden state sequence H' of the previous time step. t-1 , perform residual backtracking through the least squares method, use historical prediction errors for compensation, and generate an intermediate vector C; use the intermediate vector C as the input of the decoding layer; The decoding layer calculates the candidate hidden states of the forward and backward directions based on the intermediate vector C and the modified hidden state sequence H' through multiple groups of forward and backward GRU units to form a predicted sequence S' of the dependent variable Y.

5. The intelligent control method according to claim 3 or 4, characterized in that: In the encoding layer, the forward hidden state : ; The backward hidden state : ; The hidden state sequence H: ; in, , [·;·] represents vector concatenation operation.

6. The intelligent control method according to claim 3 or 4, characterized in that: In the fuzzy processing layer: based on N fuzzy sets A i and the linear transformation function F i Fuzzy rules R i , for each time step t and each of the fuzzy rules R i , calculate the final hidden state h(t) belongs to the fuzzy set A i The membership degree α i (t): ; Then, using the membership degree α i (t) linear transformation F i (h(t)) is weighted averaged to obtain the corrected hidden state .

7. The intelligent control method according to claim 4, characterized in that: The data-driven process of the backtracking layer includes: Let E t−1 is the prediction error of the previous time step, that is, the actual value and predicted value The differences between: ; Use the least squares method to fit the residuals and get the residual backtracking matrix W e , such that: ； Where W is the backtracking matrix, ∥⋅∥ 2 is the square of the 2-norm of the vector; Using the obtained residual backtracking matrix W e and the modified hidden state sequence H′ t−1 , generate the intermediate vector C: ； Combine the intermediate vector C with the encoding layer output H of the current time step t Combined as input to the decoding layer: H′′ t =f(H t ,C); Where f represents the splicing operation function: H′′t=[H t ;C].

8. The intelligent control method according to claim 4, characterized in that: In the decoding layer, the data-driven process of the DS evidence theory algorithm includes for each dependent variable Y' i , calculate its confidence Bel(Y' i ): The identification framework Θ divides the tunneling speed into several intervals, each of which represents a proposition: Let Θ = {Y1, Y2, ..., Y n }, where Y i The proposition that the excavation speed falls within the i-th interval is the dependent variable Y' i ; Based on the information about the tunneling speed provided by the TBM sensor data, the basic probability distribution m(Y i ); let P(Y i ) represents proposition Y i The set of all subsets of, the confidence Bel(Yi ′ ) is calculated as: ; Select the confidence level Bel(Yi ′ ) corresponds to the basic probability distribution m(Y i ), find the corresponding proposition Y i ; Then extract the most credible final dependent variable Y', that is, the dependent variable with the highest confidence: .

9. The intelligent control method according to claim 3 or 4, characterized in that: In S3, the data driving method of the PID controller is: Error e(t) = Y' set -PR'; Integral term: I(t) = I(t-1) + e(t)·△t; Differential term: ; Proportional term: P(t) = K p ·e(t); Control signal u(t) = P(t) +K i I(t) + K d D(t); Among them, K p is the proportional gain; K i is the integral gain; K d is the differential gain; t is the time gain; The set value Y' is the target tunneling speed; I(t) is the integral term value at the current moment; I(t-1) is the integral term value at the previous sampling moment; e(t-1) is the error at the previous sampling moment; P(t) and D(t) are the differential term and proportional term values ​​at the current moment.

10. An intelligent control system for cantilever tunnel boring, characterized by: The system includes a processor and a memory connected to the processor, wherein program instructions are stored in the memory, and when the program instructions are executed by the processor, the processor executes the intelligent control method according to any one of claims 1 to 9; After the control electrical signal u(t) is obtained, it is input into the TBM control system, and the TBM control system controls the TBM's actuators to perform corresponding operations.

Images