A large-diameter TBM tunneling control parameter real-time prediction method

Abstract

The application discloses a kind of super diameter TBM driving control parameter real-time prediction method, belong to tunnel boring machine construction technical field. Including: obtaining the torque of current driving cycle, thrust, penetration, speed and the rock mass integrity coefficient and uniaxial compressive strength obtained by advanced geological prediction in front, constructs real-time driving state vector;Real-time driving state vector is input time sequence characteristic extraction network, generates the feature code containing historical driving mode;According to rock mass integrity coefficient and uniaxial compressive strength, the weight vector of corresponding surrounding rock grade is generated by gate grading unit, and the feature code is weighted using weight vector, to obtain the adaptive characteristics of surrounding rock;Surrounding rock adaptive characteristics are input multi-output prediction head, and the torque prediction value, thrust prediction value and penetration prediction value of next time step are output simultaneously.

Description

Technical Field

[0001] This invention relates to the field of tunnel boring machine (TBM) construction technology, specifically a method for real-time prediction of control parameters for ultra-large diameter TBM tunneling. Background Technology

[0002] Large-diameter TBMs require real-time adjustments to control parameters such as cutterhead torque, thrust, and penetration depth based on surrounding rock conditions during tunneling to ensure both efficiency and equipment safety. Existing methods for predicting tunneling control parameters largely rely on operator experience or empirical models based on fixed geological conditions, making it difficult to maintain accurate predictions in long-distance tunnels where the strength and integrity of the surrounding rock dynamically change. Some data-driven methods train deep neural networks using historical tunneling data, but these methods only use the TBM's operating parameters as input, neglecting the utilization of the rock mass characteristics ahead. This leads to a sharp increase in prediction bias near transitional rock sections or anomalous geological formations.

[0003] Existing technical solutions suffer from two main drawbacks. Firstly, the rock mass integrity coefficient and uniaxial compressive strength information obtained from advanced geological forecasting are typically only used as auxiliary references and fail to be deeply integrated with tunneling state parameters in the prediction model. Consequently, the model cannot adaptively adjust its internal feature representation according to different surrounding rock grades. When surrounding rock conditions change significantly, a single, fixed network parameter is insufficient to characterize the differentiated mapping relationship between tunneling parameters under different surrounding rock grades, leading to lagging or inaccurate prediction results. Secondly, to achieve continuous model optimization during tunneling, conventional online update methods require a complete backpropagation of all network connection weights, incurring high computational costs that are difficult to complete within the TBM real-time control cycle. Furthermore, indiscriminately updating all parameters easily introduces historical pattern disturbances unrelated to current geological conditions, reducing the model's stability and generalization ability. How to enable the prediction model to dynamically adapt to changing surrounding rock conditions and achieve online model evolution in a lightweight manner is a problem that urgently needs to be solved. Summary of the Invention

[0004] This invention provides a real-time prediction method for tunneling control parameters of ultra-large diameter TBMs, aiming to achieve adaptive prediction of tunneling control parameters based on the surrounding rock conditions ahead, and to complete incremental model updates with low computational cost after obtaining actual parameters, so that the prediction model can maintain high accuracy and real-time performance in complex and variable geological environments.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a real-time prediction method for control parameters of ultra-large diameter TBM tunneling, the method comprising the following contents.

[0006] The current tunneling cycle's cutterhead torque, thrust, penetration depth, and rotational speed, along with the rock mass integrity coefficient and uniaxial compressive strength obtained through advanced geological prediction, are used to construct a real-time tunneling state vector. Preferably, the rock mass integrity coefficient and uniaxial compressive strength sequence obtained from the advanced geological prediction are subjected to Kalman filtering, and the current measured drilling parameters are integrated to output the filtered rock mass integrity coefficient and uniaxial compressive strength. This, along with the cutterhead torque, thrust, penetration depth, and rotational speed, constitutes the real-time tunneling state vector, thereby eliminating observational noise in the geological prediction and improving the reliability of the state vector.

[0007] The real-time tunneling state vector is input into a temporal feature extraction network to generate feature codes containing historical tunneling patterns. As a technical solution of this invention, the temporal feature extraction network consists of a temporal convolutional module and a bidirectional gated recurrent unit connected in series. The temporal convolutional module uses multi-scale one-dimensional convolutional kernels to extract local tunneling patterns across different time spans. These multi-scale one-dimensional convolutional kernels include kernels of lengths three, five, and seven. Each kernel processes the input sequence in parallel, and the outputs are batch normalized and activated before being fused through a concatenation operation to form a multi-scale local feature representation, thereby capturing the variation patterns of tunneling parameters at different time scales. The bidirectional gated recurrent unit encodes the convolutional output sequence, generating feature codes that simultaneously contain forward and backward temporal dependencies, enabling the network to fully understand the dynamic evolution characteristics of the tunneling process.

[0008] Based on the rock mass integrity coefficient and uniaxial compressive strength in the real-time tunneling state vector, a weight vector corresponding to the surrounding rock grade is generated through a gating classification unit. This weight vector is then used to weight the feature encoding to obtain adaptive features of the surrounding rock. The gating classification unit contains a surrounding rock grade inference layer and a weight generation layer. The surrounding rock grade inference layer inputs the rock mass integrity coefficient and uniaxial compressive strength into their respective membership functions to obtain the membership degree for each preset surrounding rock grade. It then calculates the activation intensity of each rule using fuzzy logic operators and generates a surrounding rock grade confidence vector through weighted averaging and defuzzification. The weight generation layer uses the surrounding rock grade confidence vector as input and generates a weight vector corresponding to the surrounding rock grade through a learnable multilayer perceptron. This method utilizes geological information to dynamically adjust the importance of temporal features, enabling the model to autonomously focus on key historical patterns based on surrounding rock conditions.

[0009] The adaptive features of the surrounding rock are input into the multi-output prediction head, which simultaneously outputs the predicted values ​​of cutterhead torque, thrust, and penetration for the next time step. The multi-output prediction head comprises a shared fully connected layer and three parallel branch fully connected layers. The shared fully connected layer performs a nonlinear transformation on the adaptive features of the surrounding rock to obtain shared features. The three branch fully connected layers take the shared features as input and output the predicted values ​​of cutterhead torque, thrust, and penetration, respectively. This structure combining shared and branch features can simultaneously predict multiple tunneling parameters, preserving the coupling relationship between parameters while also considering their individual specificities.

[0010] Once the actual tunneling parameters for the current cycle are obtained, an incremental parameter update rule is adopted based on the prediction residuals. Only the connection weights related to residual propagation in the temporal feature extraction network and the gated hierarchical unit are updated. Specifically, the residual vectors between the predicted cutterhead torque, thrust, and penetration depth and their corresponding actual values ​​are calculated. These residual vectors are then used to backpropagate through the temporal feature extraction network and the gated hierarchical unit, calculating the gradient of each layer's connection weights. According to a preset gradient threshold, connection weights with gradient magnitudes lower than the threshold are reset to zero, while only connection weights with gradient magnitudes higher than the threshold are retained and updated. Preferably, after each backpropagation, the absolute gradient value of each connection weight is calculated. Weights with absolute gradient values ​​less than an adaptive threshold are marked as frozen and do not participate in the current parameter update. The adaptive threshold is the standard deviation of the absolute gradient value of the current layer. This update strategy only corrects key connections that contribute significantly to the error propagation path, significantly reducing the computational overhead of online learning while maintaining the model's ability to remember historical data, enabling rapid adaptive adjustment of the model under complex geological conditions.

[0011] As a technical solution of the present invention, the parameters of the temporal feature extraction network are obtained in the following way: pre-training is performed using historical tunneling cycle data in a self-supervised contrastive learning manner. Positive sample pairs are feature codes for adjacent tunneling cycles, and negative sample pairs are feature codes for tunneling cycles with a time interval greater than a preset threshold. A contrastive loss function is used to update the parameters of the temporal convolutional module and the bidirectional gated recurrent unit. Through contrastive learning pre-training, the model can extract temporally consistent tunneling state representations from a large amount of unlabeled tunneling data, providing a good initialization foundation for subsequent real-time prediction and incremental updates.

[0012] The above method organically integrates surrounding rock condition reasoning, multi-scale temporal coding, multi-task synchronous prediction, and sparse incremental update. It can continuously and reliably predict the key control parameters of ultra-large diameter TBMs as the tunneling process progresses, responds promptly to changes in rock mass, and has low computational resource consumption, which is conducive to online deployment in the tunnel site environment.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0014] A gated grading unit generates a weight vector corresponding to the surrounding rock grade based on the rock mass integrity coefficient and uniaxial compressive strength. This weight vector is then used to weight the feature encoding output by the temporal feature extraction network to obtain adaptive features for the surrounding rock. The gated grading unit contains a surrounding rock grade inference layer and a weight generation layer. The surrounding rock grade inference layer uses fuzzy inference rules to map rock mass parameters to a surrounding rock grade confidence vector, and then a learnable multilayer perceptron generates the weight vector. This structure allows the same temporal feature encoding to generate different feature activation modes under different surrounding rock conditions. The prediction functions for torque, thrust, and penetration change dynamically with the surrounding rock grade, maintaining low prediction errors even in areas with alternating soft and hard rocks or fractured zones.

[0015] An incremental parameter update rule based on prediction residuals is adopted, updating only the connection weights related to residual propagation in the temporal feature extraction network and gated hierarchical units. After obtaining the actual tunneling parameters of the current cycle, the residual vector of each prediction value is calculated and backpropagated. According to a preset gradient threshold, the gradient of weights below the threshold is set to zero, and only the connection weights above the threshold are updated. By using the standard deviation of the absolute value of the gradient of the current layer as an adaptive threshold, connections that significantly contribute to reducing prediction residuals can be automatically identified, and the remaining irrelevant weights are frozen. This method does not require a complete traversal of all network parameters, significantly reducing the computational cost of a single update. It can be completed within the time interval between two tunneling cycles, and does not introduce gradient noise unrelated to the current geological environment. The model's parameters evolve smoothly during continuous tunneling, and the long-term prediction stability is improved. Multiple output prediction heads simultaneously provide predicted values ​​for torque, thrust, and penetration, sharing the adaptive feature representation of the surrounding rock, further reducing redundant computation. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of a method for real-time prediction of control parameters for ultra-large diameter TBM tunneling;

[0018] Figure 2 This is a schematic diagram of the temporal feature extraction network structure;

[0019] Figure 3 These are curves comparing the feature responses of the temporal convolution module with different kernel lengths;

[0020] Figure 4 These are Gaussian membership function curves for five preset surrounding rock grades of rock mass integrity coefficient;

[0021] Figure 5 This is a schematic diagram of the connection weight gradient absolute value sorting and adaptive threshold;

[0022] Figure 6 This is a comparison curve of the rock mass integrity coefficient advanced geological prediction and the Kalman filter result;

[0023] Figure 7 It is the curve showing the change in information noise contrast estimation loss during the contrastive learning training process. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] See Figure 1 This invention provides a real-time prediction method for control parameters of ultra-large diameter TBM tunneling, comprising: acquiring the cutterhead torque, thrust, penetration depth, rotational speed of the current tunneling cycle, as well as the rock mass integrity coefficient and uniaxial compressive strength obtained through advanced geological prediction, and constructing a real-time tunneling state vector. The real-time tunneling state vector is input into a temporal feature extraction network to generate feature codes containing historical tunneling patterns. Based on the rock mass integrity coefficient and uniaxial compressive strength in the real-time tunneling state vector, a weight vector corresponding to the surrounding rock grade is generated through a gating and grading unit. The feature codes are weighted using the weight vector to obtain adaptive features of the surrounding rock. The adaptive features of the surrounding rock are input into a multi-output prediction head, which synchronously outputs the predicted values ​​of cutterhead torque, thrust, and penetration depth for the next time step. After obtaining the actual tunneling parameters of the current cycle, based on the prediction residuals, an incremental parameter update rule is adopted to update only the connection weights related to residual propagation in the temporal feature extraction network and the gating and grading unit.

[0026] Example 1:

[0027] In specific implementation, please refer to Figure 2The temporal feature extraction network consists of a temporal convolutional module and a bidirectional gated recurrent unit connected in series. The real-time tunneling state vector is a multi-dimensional vector containing six components: cutterhead torque, thrust, penetration depth, rotational speed, rock mass integrity coefficient, and uniaxial compressive strength. The real-time tunneling state vectors of a predetermined number of consecutive tunneling cycles, including the current tunneling cycle, are arranged in chronological order to form an input sequence of the predetermined length. This input sequence is then fed into the temporal convolutional module.

[0028] The temporal convolution module contains three parallel one-dimensional convolutional layers. The first one-dimensional convolutional layer has a kernel length of three, the second has a kernel length of five, and the third has a kernel length of seven. Each of the three one-dimensional convolutional layers performs a one-dimensional convolution operation on the input sequence in the temporal dimension. The one-dimensional convolutional kernel of length three captures short-range local tunneling patterns between three adjacent tunneling cycles, the one-dimensional convolutional kernel of length five captures medium-range local tunneling patterns between five adjacent tunneling cycles, and the one-dimensional convolutional kernel of length seven captures long-range local tunneling patterns between seven adjacent tunneling cycles. The output of each one-dimensional convolutional layer is processed by a batch normalization layer and an activation function layer, with the activation function being a modified linear unit (MRU). After processing by the batch normalization layer and the MRU, the three feature maps are concatenated in the channel dimension to form a concatenated feature map that integrates short-range, medium-range, and long-range temporal scale information. This concatenated feature map serves as a multi-scale local feature representation. Multi-scale local feature representation preserves the sequence order in the time dimension, forming a multi-scale local feature representation sequence.

[0029] The bidirectional gated recurrent unit (BRN) receives the multi-scale local feature representation sequence output by the temporal convolution module. The BRN consists of a forward-gated recurrent unit (PGU) and a backward-gated recurrent unit (GRU). The forward-gated RNU processes the multi-scale local feature representation sequence sequentially along the forward temporal direction, updating the forward hidden state at each time step. The backward-gated RNU processes the multi-scale local feature representation sequence sequentially along the reverse temporal direction, updating the backward hidden state at each time step. After all time steps have been processed, the forward hidden state of the forward-gated RNU at the last time step is extracted, and the backward hidden state of the backward-gated RNU at the first time step is also extracted. The forward and backward hidden states are concatenated to obtain a concatenated vector, which serves as the feature encoding containing both forward and backward temporal dependencies.

[0030] See Figure 3In the figure, the horizontal axis represents the tunneling cycle number, ranging from 0 to 600, and the vertical axis represents the feature value, ranging from approximately 63 to 103. The figure shows the feature curves extracted by three temporal convolutional modules for different convolutional kernel lengths (3, 5, and 7, respectively). The three curves have highly consistent shapes, exhibiting obvious periodic fluctuations. The feature value rises and falls periodically with the tunneling cycle number, with alternating peaks and troughs. The highest feature value is close to 102, and the lowest is about 65, indicating that different convolutional kernel lengths can effectively capture the temporal changes in the tunneling state.

[0031] Specifically, the curve (solid line) of the temporal convolutional module with a kernel length of 3 shows more detailed fluctuations in short-range local feature capture, and the curve is slightly more delicate; the curve (dashed line) of the temporal convolutional module with a kernel length of 5 has a moderate smoothness and can capture local tunneling patterns at a medium scale; the curve (dotted line) of the temporal convolutional module with a kernel length of 7 is smoother and reflects the changing trend of long-range local tunneling patterns. The three curves basically overlap within the entire tunneling cycle sequence, with no obvious deviation, indicating that the temporal convolutional module adopts a parallel multi-scale convolutional kernel design, which can effectively fuse multi-scale temporal feature information, taking into account the extraction of short-range, medium-range, and long-range local tunneling patterns, and is conducive to improving the expressive power of the temporal feature extraction network.

[0032] Example 2:

[0033] In practice, the gating and grading unit contains a surrounding rock grade inference layer and a weight generation layer, with the output of the surrounding rock grade inference layer connected to the input of the weight generation layer.

[0034] The surrounding rock grade inference layer receives two scalar values: the rock mass integrity coefficient and the uniaxial compressive strength, extracted from the real-time tunneling state vector. The surrounding rock grade inference layer predefines five preset surrounding rock grades, denoted as Surrounding Rock Grade 1, Surrounding Rock Grade 2, Surrounding Rock Grade 3, Surrounding Rock Grade 4, and Surrounding Rock Grade 5. The surrounding rock quality corresponding to these five preset grades increases sequentially from extremely fragmented to extremely intact.

[0035] Within the rock mass integrity coefficient inference layer, five membership functions are established. Each membership function corresponds to a preset rock mass integrity coefficient. The membership functions for the rock mass integrity coefficient adopt a Gaussian membership function, and the expression is as follows:

[0036]

[0037] in, This represents the input rock mass integrity coefficient, with a value ranging from 0 to 1; For the index of preset surrounding rock grade, ; Indicates the rock mass integrity coefficient For the first The membership degree of a preset surrounding rock grade, with a value range from 0 to 1; The rock mass integrity coefficient corresponds to the first The five center values ​​of the Gaussian membership function for the preset surrounding rock grades are 0.1, 0.3, 0.5, 0.7, and 0.9, respectively. The basis for these values ​​is that the domain of the rock mass integrity coefficient is divided into five intervals at equal intervals, and the center of each interval is used as the membership center of the corresponding surrounding rock grade. The rock mass integrity coefficient corresponds to the first The width parameters of the Gaussian membership functions for each preset surrounding rock grade are all set to 0.12. This value ensures that the overlap between adjacent membership functions at the boundary is approximately 0.5.

[0038] Within the rock mass classification inference layer, five membership functions are established for uniaxial compressive strength. Each membership function corresponds to a preset rock mass classification. The membership function for uniaxial compressive strength adopts a Gaussian membership function, and its expression is as follows:

[0039]

[0040] in, This represents the input uniaxial compressive strength, in megapascals (MPa), and its value range is a real number greater than 0. For the index of preset surrounding rock grade, ; Indicates uniaxial compressive strength For the first The membership degree of a preset surrounding rock grade, with a value range from 0 to 1; Indicates the uniaxial compressive strength corresponding to the first The five central values ​​of the Gaussian membership function for the preset surrounding rock grades are 5 MPa, 25 MPa, 60 MPa, 100 MPa and 150 MPa respectively. The basis for these values ​​is to divide the uniaxial compressive strength range of common rocks from extremely soft rocks to extremely hard rocks into five grades according to logarithmic intervals. Indicates the uniaxial compressive strength corresponding to the first The Gaussian membership function width parameters are preset for each surrounding rock grade. The five width parameters are 3, 8, 15, 20, and 30 respectively. The values ​​are based on the estimated statistical standard deviation of the uniaxial compressive strength variation range for each grade.

[0041] The surrounding rock grade inference layer defines five fuzzy inference rules, each corresponding to a preset surrounding rock grade. The fuzzy inference rule is stated as follows: If the rock mass integrity coefficient belongs to the first... The preset surrounding rock grade and the uniaxial compressive strength belongs to the first... If there is a preset surrounding rock grade, then the surrounding rock grade is the first one. There are several preset surrounding rock grades. The surrounding rock grade inference layer uses a fuzzy AND operator to calculate the activation intensity of each fuzzy inference rule; the fuzzy AND operator performs a minimum value operation. Activation strength of fuzzy inference rules The calculation method is as follows:

[0042]

[0043] in, This indicates taking the smaller of the two values.

[0044] The surrounding rock grade inference layer defuzzifies the five fuzzy inference rules by weighted averaging the activation intensities, generating a surrounding rock grade confidence vector. The surrounding rock grade confidence vector is a five-dimensional vector, where the first... The elements are denoted as The calculation method is as follows:

[0045]

[0046] in, Indicates the first The activation strength of a fuzzy inference rule. For the summation index, ; denominator This represents the sum of the activation intensities of all five fuzzy inference rules. Rock mass level confidence vector. Each element in the set has a value between 0 and 1, and the sum of all elements is 1.

[0047] The weight generation layer receives the rock grade confidence vector output by the rock grade inference layer. The weight generation layer is a learnable multilayer perceptron, consisting of an input layer, a hidden layer, and an output layer. The input layer contains five neurons, receiving the five elements of the rock grade confidence vector. The hidden layer contains sixteen neurons, using a modified linear unit (MLU) as its activation function. The output layer contains the same number of neurons as the feature encoding dimension, and its activation function is a sigmoid function, compressing the output values ​​to the 0-1 range. All connection weights of the multilayer perceptron are learned and updated during the overall training of the prediction model using backpropagation, with an initial learning rate of 0.001. An adaptive moment estimation optimizer is used. The vector output by the weight generation layer is the weight vector corresponding to the rock grade, with the same dimension as the feature encoding. The weight vector is element-wise multiplied with the feature encoding to obtain the adaptive rock grade feature.

[0048] See Figure 4 The curve in the figure represents the rock mass integrity coefficient of the surrounding rock grade in Example 2. A schematic diagram of the five established Gaussian membership functions. The horizontal axis represents the rock mass integrity coefficient. The value ranges from 0 to 1; the vertical axis represents the membership degree value of the corresponding membership function, also ranging from 0 to 1. The figure contains five curves, each corresponding to one of the five preset rock mass grades. All curves exhibit a Gaussian distribution, with their center values ​​staggered, covering the entire range of the rock mass integrity coefficient.

[0049] Specifically, the membership function center value corresponding to rock grade one (red solid line) is approximately 0.1, and the membership degree is in... The peak value of 1 indicates the rock mass grade when its integrity is extremely fragmented. The central value for grade two (purple dashed line) is approximately 0.3, for grade three (brown dotted line) approximately 0.5, for grade four (pink dotted line) approximately 0.7, and for grade five (gray dotted line) approximately 0.9. All curves have a peak value of 1, shifting sequentially to the right. The widths of all membership functions remain consistent, exhibiting relatively smooth overlap. The membership value at the overlap is approximately 0.5, ensuring that when the rock mass integrity coefficient is in the boundary region between two grades, it can simultaneously and effectively reflect the membership relationship between two adjacent grades.

[0050] Example 3:

[0051] In practice, after obtaining the actual cutterhead torque, actual thrust, and actual penetration depth collected at the end of the current tunneling cycle, the actual cutterhead torque is recorded as... The actual thrust is recorded as The actual penetration depth is recorded as The predicted value of the cutter head torque previously output by the multi-output predictor head is denoted as... The thrust prediction value is denoted as The penetration prediction value is denoted as The predicted residual vector is calculated, and the residual vector consists of three residual components, namely the torque residual. Thrust residual Penetration residual .

[0052] Constructing a mean squared error loss function based on residual vectors , The calculation method is as follows Using the loss function Backpropagation is performed on the temporal feature extraction network and the gated hierarchical unit. The backpropagation path passes through the multi-output prediction head to the gated hierarchical unit and the temporal feature extraction network. During backpropagation, all connection weights of the multi-output prediction head remain frozen, and the gradients of the connection weights in the multi-output prediction head are not calculated. Gradient information is only transmitted to the trainable weight generation layer connection weights in the gated hierarchical unit, and the trainable temporal convolutional module kernel weights and the gated connection weights of the bidirectional gated recurrent unit in the temporal feature extraction network. The center value of the Gaussian membership function in the surrounding rock grade inference layer is determined. , , , All values ​​are preset fixed values ​​and do not participate in gradient calculation; the parameters of the surrounding rock grade inference layer are not updated.

[0053] After a single backpropagation calculation, the gradient values ​​of each trainable connection weight in the temporal feature extraction network and the gating hierarchical unit are obtained. For the ... The first layer Let the absolute value of the gradient of each connection weight be denoted as . Regarding the first Calculate the standard deviation of the absolute values ​​of the gradients of all trainable connection weights in each layer. As an adaptive threshold. Standard deviation The calculation formula is:

[0054]

[0055] in, Indicates the first The total number of trainable connection weights in a layer; Indicates the first Layer The absolute value of the gradient of each connection weight; Indicates the first Layer The absolute value of the gradient of each connection weight. For the summation index, The range of values ​​is to ; Indicates the first The arithmetic mean of the absolute values ​​of the gradients of all trainable connection weights in the layer.

[0056] In obtaining the first Adaptive threshold of the layer After that, the first absolute value of gradient in layer Less than the adaptive threshold The connection weights are marked as frozen, and the gradients corresponding to the frozen connection weights are forcibly set to zero. The frozen connection weights do not change during this parameter update. The ... absolute value of gradient in layer Greater than or equal to the adaptive threshold The connection weights are marked as unfrozen, retaining their original calculated gradient values. A stochastic gradient descent optimizer is used to update the unfrozen connection weights, with a learning rate of 0.0005. The update rule is: the new value of the unfrozen connection weight equals the current value minus the product of the learning rate and the corresponding gradient. All frozen connection weights that were set to zero retain their original values ​​in this update. After each update, the updated connection weight values ​​are stored for forward computation in the next tunneling cycle. After each tunneling cycle is completed and the actual tunneling parameters are obtained, the above residual calculation, backpropagation, gradient absolute value statistics, adaptive threshold calculation, and selective weight update process are performed once, realizing incremental parameter updates for the connection weights of the temporal feature extraction network and the gated hierarchical unit during the tunneling process.

[0057] See Figure 5 In the graph, the horizontal axis represents the weight index, sorted in descending order of absolute gradient value, while the vertical axis represents the absolute gradient value of the corresponding weight. The blue solid line curve reflects the trend of the weights' absolute gradient values ​​from largest to smallest, with the initial weights having larger absolute gradient values ​​that gradually decrease, exhibiting a monotonically decreasing trend. The red dashed line represents the adaptive threshold. The weight index is located at a fixed horizontal line of approximately 0.02 on the vertical axis. The blue curve intersects the red dashed line at approximately 300, indicating that the absolute gradient value of the first 300 weights is greater than or equal to the adaptive threshold, and they belong to the unfrozen state. The absolute gradient value of weights with an index greater than 300 is lower than the threshold, and they are classified as frozen.

[0058] Example 4:

[0059] In practice, advanced geological prediction methods include seismic wave reflection detection, ground-penetrating radar detection, or borehole camera detection. Advanced geological prediction obtains a sequence of rock mass integrity coefficients and a sequence of uniaxial compressive strengths within a predetermined range ahead of the tunnel face along the tunnel excavation direction. The rock mass integrity coefficient sequence contains multiple predicted rock mass integrity coefficient values ​​arranged at fixed sampling intervals along the tunnel axis, and the uniaxial compressive strength sequence contains multiple predicted uniaxial compressive strength values ​​arranged at the same fixed sampling intervals along the tunnel axis. The fixed sampling interval is set to 0.5 meters.

[0060] Kalman filtering was applied to the rock mass integrity coefficient sequence and the uniaxial compressive strength sequence, respectively. The state vector of the Kalman filter contains two components: the filtered rock mass integrity coefficient and the filtered uniaxial compressive strength. In the state equation of the Kalman filter, the state transition matrix is ​​set as a second-order identity matrix, specifically:

[0061]

[0062] in, Represents the state transition matrix, matrix elements and For fixed values, This indicates that the current state is the same as the previous state. This indicates that there is no mutual transfer relationship between the rock mass integrity coefficient component and the uniaxial compressive strength component. This value is based on the assumption that the rock mass integrity coefficient and uniaxial compressive strength do not change within a 0.5-meter interval between adjacent sampling points.

[0063] The observation vector of the Kalman filter consists of the predicted values ​​of rock mass integrity coefficient and uniaxial compressive strength obtained from advanced geological forecasting. The observation matrix of the Kalman filter is set as a second-order identity matrix. The process noise covariance matrix of the Kalman filter is set as a diagonal matrix, with diagonal elements of 0.0001 and 0.01, where 0.0001 corresponds to the process noise variance of the rock mass integrity coefficient and 0.01 corresponds to the process noise variance of the uniaxial compressive strength. This value is based on the fact that the fluctuation range of the predicted results is small after the rock mass integrity coefficient is normalized to the range of 0 to 1, and the fluctuation range of the predicted results of the uniaxial compressive strength in the megapascal range is about 0.1 megapascals. The observation noise covariance matrix of the Kalman filter is set as a diagonal matrix. The diagonal elements are determined according to the technical specifications of the advanced geological prediction equipment, and are set to 0.001 and 0.1 respectively. 0.001 corresponds to the observation noise variance of the rock mass integrity coefficient, and 0.1 corresponds to the observation noise variance of the uniaxial compressive strength. The basis for this value is that the detection accuracy of the rock mass integrity coefficient by the seismic wave reflection method for advanced geological prediction is about 0.03, and the detection accuracy of the uniaxial compressive strength is about 0.3 MPa. The square of these values ​​is taken as the observation noise variance.

[0064] Kalman filtering performs one prediction step and one update step at each sampling point. The prediction step uses the filtered state and process noise covariance matrix of the previous sampling point to calculate the prior state estimate and prior error covariance matrix for the current sampling point. The update step uses the observed values, observation noise covariance matrix, and observation matrix of the current sampling point to calculate the Kalman gain, and uses the Kalman gain to correct the prior state estimate, obtaining the posterior state estimate and corresponding posterior error covariance matrix for the current sampling point. The first component of the posterior state estimate is used as the filtered rock mass integrity coefficient, and the second component is used as the filtered uniaxial compressive strength.

[0065] The method for integrating the currently measured drilling parameters is as follows: the cutterhead torque, thrust, penetration depth, and rotational speed measured by sensors installed on the TBM during the current tunneling cycle are combined with the filtered rock integrity coefficient and filtered uniaxial compressive strength at the tunnel face location after Kalman filtering. These six values—cutterhead torque, thrust, penetration depth, rotational speed, filtered rock integrity coefficient, and filtered uniaxial compressive strength—are arranged in a predetermined order to form a six-dimensional vector, which is the real-time tunneling state vector. The arrangement order of the real-time tunneling state vector is: cutterhead torque value in the first dimension, thrust value in the second dimension, penetration depth value in the third dimension, rotational speed value in the fourth dimension, filtered rock integrity coefficient value in the fifth dimension, and filtered uniaxial compressive strength value in the sixth dimension.

[0066] See Figure 6 In the figure, the horizontal axis represents the distance along the tunnel axis, ranging from 0 meters to 300 meters, and the vertical axis represents the rock mass integrity coefficient, ranging from approximately 0.3 to 0.85. The figure shows two curves: the blue dashed line represents the predicted rock mass integrity coefficient sequence obtained from advanced geological prediction, and the red solid line represents the corresponding rock mass integrity coefficient sequence after Kalman filtering. The blue dashed line curve exhibits significant high-frequency fluctuations, reflecting certain measurement noise and local fluctuations in the original data from the advanced geological prediction, especially noticeable near sampling points along the tunnel axis. The red solid line curve is smoother than the blue dashed line, with significantly reduced fluctuation amplitude, indicating that Kalman filtering effectively suppresses random noise in the advanced geological prediction data, highlighting the true trend of rock mass integrity coefficient changes.

[0067] Within the range of 0 to 300 meters from the axis, the filtered rock mass integrity coefficient curve exhibits a slow, periodic change. Peaks appear approximately at 15, 65, 115, 170, 210, and 265 meters, corresponding to local maximum values ​​of the rock mass integrity coefficient, fluctuating between 0.7 and 0.8. Troughs are distributed at approximately 40, 90, 140, 190, 240, and 290 meters, with integrity coefficients decreasing to the range of 0.35 to 0.45, indicating a significant spatial distribution variation in rock mass integrity. The filtered curve effectively preserves the spatial trend of the rock mass integrity coefficient while reducing the interference of local anomalous fluctuations on prediction.

[0068] Example 5:

[0069] In practice, the multi-output prediction head comprises a shared fully connected layer and three parallel branch fully connected layers. The input to the shared fully connected layer is the adaptive feature vector of the surrounding rock, the dimension of which is denoted as... A shared fully connected layer contains one fully connected network, and the number of neurons in the shared fully connected layer is set to... The shared fully connected layer uses a modified linear unit as its activation function. The shared fully connected layer performs a nonlinear transformation on the adaptive feature vector of the surrounding rock, outputting a vector with dimension 1. The shared eigenvectors. The calculation process of the nonlinear transformation is to multiply the connection weight matrix of the shared fully connected layer with the adaptive eigenvectors of the surrounding rock, add the bias vector, and then output it through the modified linear unit activation function.

[0070] The output of the shared fully connected layer is simultaneously connected to three branch fully connected layers. The first branch fully connected layer is for toolhead torque prediction, the second for thrust prediction, and the third for penetration prediction. These three branch fully connected layers receive the shared feature vector in parallel and perform independent computations. Each branch fully connected layer contains a fully connected network layer, and the number of neurons in each branch fully connected layer is set to a specific value. Each fully connected layer in each branch uses a modified linear unit (MRU) as its activation function. The output of each fully connected layer is connected to an output neuron node, which does not use an activation function and outputs a linear prediction value. The output neuron node connected to the first fully connected layer outputs the predicted cutterhead torque, the output neuron node connected to the second fully connected layer outputs the predicted thrust, and the output neuron node connected to the third fully connected layer outputs the predicted penetration. These three prediction values ​​are generated synchronously during the same forward computation.

[0071] The parameters of the temporal feature extraction network were obtained through self-supervised contrastive learning pre-training. Self-supervised contrastive learning pre-training used historical tunneling cycle data to construct training samples, which came from completed tunneling sections of ultra-large diameter TBMs. Each historical tunneling cycle data point contained six parameter values: cutterhead torque, thrust, penetration depth, rotational speed, rock integrity coefficient, and uniaxial compressive strength within a single tunneling cycle. Each historical tunneling cycle data point was constructed as a historical tunneling state vector, constructed in the same way as the real-time tunneling state vector. The historical tunneling state vectors of a predetermined number of consecutive tunneling cycles were arranged chronologically into a historical input sequence, which was then fed into the temporal convolutional module and bidirectional gated recurrent unit to be pre-trained, extracting the corresponding historical feature codes. The historical feature codes were extracted using the same method as the feature codes.

[0072] In self-supervised contrastive learning pre-training, positive sample pairs are defined as feature codes of adjacent tunneling cycles, meaning that the historical feature codes corresponding to two temporally consecutive tunneling cycles constitute a positive sample pair. Negative sample pairs are defined as feature codes of tunneling cycles with a time interval greater than a preset threshold, which is set to ten tunneling cycles. This threshold is based on the fact that in ultra-large diameter TBM tunneling, the surrounding rock conditions and tunneling conditions change significantly after ten tunneling cycles, and the two types of samples do not constitute a homogeneous pattern. The samples within a training batch are organized as follows: an anchor point historical feature code is selected, and the feature codes of adjacent tunneling cycles corresponding to the anchor point historical feature code are used as positive samples. The feature codes of other time intervals within the same batch that meet the negative sample pair condition are used as negative samples.

[0073] Self-supervised contrastive learning pre-training uses a contrastive loss function to update the parameters of the temporal convolutional module and the bidirectional gated recurrent unit. The information-noise contrastive estimation loss function is selected. The calculation method is as follows:

[0074]

[0075] in, This represents the historical feature encoding vector of the anchor point; This represents the historical feature encoding vector of the adjacent tunneling cycle that forms a positive sample pair with the historical feature encoding vector of the anchor point; This represents the set of negative samples consisting of all historical feature encoding vectors that form negative sample pairs with the historical feature encoding vector of the anchor point. Represents the set of negative samples The first in A negative sample historical feature encoding vector; This represents the cosine similarity function, used to measure the degree of similarity between two vectors, with a value between -1 and 1. This represents the temperature coefficient, which is set to 0.1. This value is based on the fact that in self-supervised contrastive learning, a temperature coefficient of 0.1 can form a moderate gradient discrimination between positive and negative samples, avoiding the gradient from being too concentrated on a few extremely difficult samples. Represented by natural constant An exponential function with base 0.

[0076] In the self-supervised contrastive learning pre-training phase, an adaptive moment estimation optimizer was used with a learning rate of 0.001, 200 training epochs, and a batch size of 64. In each training epoch, all historical tunneling cycle data were divided into batches, and the contrastive loss function was calculated batch by batch. This loss function was then used for backpropagation and parameter updates of all convolutional kernel weights in the temporal convolutional module and all gated connection weights in the bidirectional gated recurrent units. After training, the parameters of the temporal convolutional module and the bidirectional gated recurrent units were saved and used as the initial parameters for the temporal feature extraction network, subsequently used in the prediction process based on the real-time tunneling state vector and the incremental parameter update process. The center value and width parameters of the membership function in the rock grade inference layer, as well as the connection weights of all fully connected layers in the multi-output prediction head, were not trained during the pre-training phase and were assigned values ​​using random initialization after pre-training.

[0077] See Figure 7 In the figure, the horizontal axis represents the number of training steps, ranging from 0 to 1000, and the vertical axis represents the information-noise contrastive estimation loss, ranging from approximately 0.15 to 0.35. The curve reflects the changing trend of the loss function during the pre-training process of the temporal feature extraction network through self-supervised contrastive learning in Example 5. The overall curve shows relatively stable fluctuations as the number of training steps increases, without a significant continuous downward trend. The loss value is mainly concentrated between 0.200 and 0.325, indicating that the model maintains a certain contrast discrimination ability during training.

[0078] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for real-time prediction of control parameters for ultra-large diameter TBM tunneling, characterized in that, The method includes: The current tunneling cycle's cutterhead torque, thrust, penetration depth, rotational speed, as well as the rock mass integrity coefficient and uniaxial compressive strength obtained through advanced geological prediction, are used to construct a real-time tunneling state vector. The real-time tunneling state vector is input into a temporal feature extraction network to generate feature codes containing historical tunneling patterns. Based on the rock mass integrity coefficient and uniaxial compressive strength in the real-time tunneling state vector, a weight vector corresponding to the surrounding rock grade is generated through a gating classification unit. The feature encoding is weighted using the weight vector to obtain the adaptive features of the surrounding rock. The adaptive features of the surrounding rock are input into the multi-output prediction head, which simultaneously outputs the predicted values ​​of the cutterhead torque, thrust, and penetration depth for the next time step. Once the actual tunneling parameters for the current cycle are obtained, based on the predicted residuals, an incremental parameter update rule is adopted to update only the connection weights related to residual propagation in the temporal feature extraction network and the gating hierarchical unit.

2. The method for real-time prediction of control parameters for ultra-large diameter TBM tunneling according to claim 1, characterized in that, The step of inputting the real-time tunneling state vector into a temporal feature extraction network to generate feature codes containing historical tunneling patterns is specifically as follows: The temporal feature extraction network is composed of a temporal convolutional module and a bidirectional gated recurrent unit connected in series. The temporal convolution module uses multi-scale one-dimensional convolution kernels to extract local tunneling patterns across different time spans; The bidirectional gated recurrent unit encodes the convolutional output sequence to generate feature codes that include forward and backward temporal dependencies.

3. The method for real-time prediction of control parameters for ultra-large diameter TBM tunneling according to claim 1, characterized in that, Based on the rock mass integrity coefficient and uniaxial compressive strength in the real-time tunneling state vector, a weight vector corresponding to the surrounding rock grade is generated through a gated grading unit, specifically as follows: The gated classification unit contains a surrounding rock grade inference layer and a weight generation layer. The surrounding rock grade inference layer outputs a surrounding rock grade confidence vector based on the input rock mass integrity coefficient and uniaxial compressive strength using fuzzy inference rules. The weight generation layer takes the surrounding rock grade confidence vector as input and generates the weight vector corresponding to different surrounding rock grades through a learnable multilayer perceptron.

4. The method for real-time prediction of control parameters for ultra-large diameter TBM tunneling according to claim 1, characterized in that, Once the actual tunneling parameters for the current cycle are obtained, based on the predicted residuals, an incremental parameter update rule is adopted to update only the connection weights related to residual propagation in the temporal feature extraction network and the gating hierarchical unit. Specifically: Calculate the residual vectors between the predicted values ​​of the cutterhead torque, thrust, and penetration and their corresponding actual values. The residual vector is used to backpropagate the temporal feature extraction network and the gated hierarchical unit to calculate the gradient of the connection weights of each layer; Based on a preset gradient threshold, connection weights with gradient magnitudes lower than the threshold are reset to zero, while only connection weights with gradient magnitudes higher than the threshold are retained and updated.

5. The method for real-time prediction of control parameters for ultra-large diameter TBM tunneling according to claim 2, characterized in that, The temporal convolution module uses multi-scale one-dimensional convolution kernels to extract local mining patterns across different time spans. The multi-scale one-dimensional convolution kernels include one-dimensional convolution kernels with lengths of three, five, and seven. Each convolution kernel processes the input sequence in parallel, and the output is batch normalized and activated by a concatenation operation to form a multi-scale local feature representation.

6. The method for real-time prediction of control parameters for ultra-large diameter TBM tunneling according to claim 3, characterized in that, The surrounding rock grade inference layer outputs a surrounding rock grade confidence vector based on the input rock mass integrity coefficient and uniaxial compressive strength using fuzzy inference rules, specifically: The rock mass integrity coefficient and uniaxial compressive strength are respectively input into their respective membership functions to obtain the membership degree for each preset surrounding rock grade; The activation intensity of each rule is calculated by fuzzing and operators, and then the surrounding rock grade confidence vector is generated by weighted average defuzzification.

7. The method for real-time prediction of control parameters for ultra-large diameter TBM tunneling according to claim 4, characterized in that, The step of resetting connection weights with gradient magnitudes lower than a preset gradient threshold to zero, and retaining and updating only connection weights with gradient magnitudes higher than the gradient threshold, specifically involves: After each backpropagation, the absolute value of the gradient of each connection weight in the temporal feature extraction network and the gated hierarchical unit is calculated; Weights whose absolute gradient value is less than the adaptive threshold are marked as frozen and will not participate in this parameter update. The adaptive threshold is taken as the standard deviation of the absolute value of the gradient of the current layer.

8. The method for real-time prediction of control parameters for ultra-large diameter TBM tunneling according to claim 1, characterized in that, The process involves acquiring the cutterhead torque, thrust, penetration depth, rotational speed of the current tunneling cycle, as well as the rock mass integrity coefficient and uniaxial compressive strength obtained through advanced geological prediction, to construct a real-time tunneling state vector. Specifically: The rock mass integrity coefficient and uniaxial compressive strength sequence obtained from the advanced geological prediction are subjected to Kalman filtering, and the current measured drilling parameters are fused together to output the filtered rock mass integrity coefficient and uniaxial compressive strength, which, together with the cutterhead torque, thrust, penetration depth, and rotational speed, constitute the real-time tunneling state vector.

9. The method for real-time prediction of control parameters for ultra-large diameter TBM tunneling according to claim 1, characterized in that, The process of inputting the adaptive features of the surrounding rock into the multi-output prediction head and simultaneously outputting the predicted values ​​of the cutterhead torque, thrust, and penetration depth for the next time step is as follows: The multi-output prediction head comprises a shared fully connected layer and three parallel branch fully connected layers; The shared fully connected layer performs a nonlinear transformation on the adaptive characteristics of the surrounding rock to obtain shared characteristics; The three fully connected branches take the shared features as input and output the predicted values ​​of the cutterhead torque, thrust, and penetration.

10. The method for real-time prediction of control parameters for ultra-large diameter TBM tunneling according to claim 2, characterized in that, The parameters of the temporal feature extraction network are obtained in the following way: Using historical tunneling cycle data, pre-training is performed using a self-supervised contrastive learning approach. Positive sample pairs are feature codes of adjacent tunneling cycles, and negative sample pairs are feature codes of tunneling cycles with a time interval greater than a preset threshold. The parameters of the temporal convolutional module and the bidirectional gated recurrent unit are updated using a contrastive loss function.

Images