Tunnel TBM path adjustment real-time planning method based on intelligent algorithm

Abstract

The invention provides a tunnel TBM path adjustment real-time planning method based on an intelligent algorithm. The tunnel TBM path adjustment real-time planning method comprises five steps of data acquisition, data calculation, deviation judgment and decision making, neural network prediction and correction strategy generation. The data acquisition comprises tunnel section data acquisition and attitude parameters of the shield tunneling machine; the data calculation comprises point cloud denoising, point cloud registration, data fusion, axis reconstruction, deviation index calculation and result output; the deviation judgment and decision-making decides whether deviation correction is carried out or not; neural network prediction is based on LSTM + CNN hybrid model processing time-space sequence data, and future offset is predicted; and the deviation correction strategy converts a prediction result into an executable control instruction through reinforcement learning. According to the method, real-time path planning and deviation rectification are carried out through combination of multi-modal data fusion and a hybrid intelligent algorithm, three-dimensional laser scanning and TBM data, the tunneling path is measured and adjusted in real time after tunneling is completed each time, the problem of path planning lag of a traditional method under complex working conditions is solved, and the tunnel construction precision and efficiency are improved.

Description

Technical Field

[0001] This invention belongs to the field of tunnel construction technology, specifically relating to a real-time planning method for tunnel TBM path adjustment based on intelligent algorithms. Background Technology

[0002] With the rapid development of transportation infrastructure, underground transportation is booming. Every major city needs to build subways and tunnels that traverse mountains, all of which require drilling to achieve completion. The traditional tunneling method is drill-and-blast, which is relatively slow. Its disadvantages become even more apparent when dealing with projects with complex geological conditions or hard surrounding rock.

[0003] Currently, TBM (Tunnel Boring Machine) methods are generally used for drilling ultra-long tunnels. However, when using drilling rigs, if interference is not eliminated in time, it may cause the tunneling route to deviate from the designed route, resulting in irreparable consequences. Path adjustment relies on manual analysis and monitoring data, which suffers from poor real-time performance and low correction accuracy. While existing intelligent algorithms can predict deviations, they lack dynamic optimization of equipment parameters, making it difficult to achieve coordinated control of path smoothness and cumulative deviation. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time planning method for tunnel boring machine (TBM) path adjustment based on intelligent algorithms. By using multimodal data fusion and hybrid intelligent algorithms, the tunneling route is measured and adjusted in real time after each tunneling operation, thus solving the problem of path planning lag in complex working conditions using traditional methods.

[0005] Therefore, the technical solution provided by the present invention is as follows:

[0006] A real-time planning method for tunnel TBM path adjustment based on intelligent algorithms includes the following steps:

[0007] Step 1) Data acquisition: Obtain cross-sectional data after each tunneling operation through cross-sectional scanning, and transmit the cross-sectional data and the attitude parameters of the tunnel boring machine to the edge computing unit.

[0008] Step 2) Data processing: point cloud denoising, point cloud registration and data fusion, then axis reconstruction, calculation of deviation index and output of results;

[0009] Among them, the deviation indicators include lateral offset Δd, longitudinal cumulative error Δl, and cross-sectional ellipticity e;

[0010] Step 3) Deviation Judgment and Decision: Set a deviation threshold according to the actual situation of the project, compare the deviation index with the deviation threshold, and if the deviation is within the threshold range, continue tunneling according to the original route;

[0011] If the deviation is greater than the deviation threshold, proceed to step 4);

[0012] Step 4) Neural Network Prediction: Based on the LSTM+CNN hybrid model, the input historical deviation sequence, TBM running parameters and point cloud spatial features are processed, and the expected offset and recommended correction angle for the next N advances are output through the model structure and training.

[0013] The historical deviation sequence is formed by taking the lateral offset Δd, longitudinal cumulative error Δl, cross-sectional ellipticity e, pitch angle change Δpitch, and roll angle change Δroll of the most recent M tunneling rings, constructing a time series matrix, normalizing it, and arranging it in ring order.

[0014] The TBM operating parameters are obtained by extracting the sliding average and standard deviation of the cutterhead torque τ, feed speed v, and feed force f of the current ring and the nearest M ring as feature vectors;

[0015] Step 5) Reinforcement learning to generate correction strategies: Based on the prediction results, with the goal of minimizing cumulative deviation and path smoothness, reinforcement learning is used to generate correction strategies and output control commands for the tool head steering angle and feed pressure adjustment values.

[0016] Furthermore, the specific process of step 1) is as follows:

[0017] The accelerometer on the 3D laser scanner installed at the tail of the tunnel boring machine (TBM) identifies the TBM's travel status. When it is detected that the TBM has finished one advance and is no longer moving, the 3D laser scanner scans the tunnel that has been excavated and uploads five attitude parameters of the TBM, namely pitch angle, roll angle, thrust, cutterhead torque and thrust speed, to the edge computing unit.

[0018] Furthermore, the point cloud denoising described in step 2) is based on the statistical distance of neighboring points. By using statistical filtering, flying points and dust noise generated during the scanning process are filtered out, so that the average distance between each point in the point cloud and its K nearest neighbors does not exceed the threshold. If it does, it is removed.

[0019] The threshold is obtained by the following formula:

[0020] D max =μ+std×σ

[0021]

[0022] In the formula, μ is the average distance between the point to be judged and its K nearest neighbors; σ is the standard deviation of the distance; std is a user-defined standard deviation coefficient used to control the influence of the standard deviation of the distance on the distance threshold; D i Let i be the distance between point i and its nearest neighbor.

[0023] Furthermore, the specific process of point cloud registration described in step 2) is as follows:

[0024] The point cloud from this scan is precisely registered with the previous scan to determine the actual position and pose changes of the TBM in the global coordinate system. The ICP algorithm is used for registration, with the goal of minimizing the distance between corresponding points. The objective function is shown below:

[0025]

[0026] Where R is the rotation matrix, T is the translation vector, n is the number of effective corresponding point pairs, and P i It is a point cloud to be registered, q i It is a reference point cloud, and R and T are solved iteratively.

[0027] Furthermore, data fusion involves adding a unified timestamp to the registered point cloud data and the TBM attitude and operating parameters collected at the same time, and then associating them with the corresponding tunneling mileage.

[0028] Furthermore, the specific process for axis reconstruction and deviation index calculation in step 2) is as follows:

[0029] (1) Obtain the design axis of this mileage section from the design documents;

[0030] (2) Extract the actual centerline of the current section from the registered point cloud data. Through section fitting, slice the point cloud on multiple cross sections perpendicular to the tunneling direction, and then perform ellipse fitting to obtain the actual axis.

[0031] (3) Align the design axis with the actual axis and calculate the deviation index;

[0032] (4) The lateral offset Δd is obtained by calculating the distance between the actual center point and the design center point in the direction perpendicular to the design axis. The longitudinal cumulative error Δl is obtained by calculating the total length difference between the actual axis and the design axis in the direction along the design axis from the starting point to the current mileage. The ellipticity e of the section is obtained by calculating the ratio of the major axis to the minor axis of the fitted ellipse.

[0033] Furthermore, the output of step 2) includes the lateral offset Δd, longitudinal cumulative error Δl, cross-sectional ellipticity e, and its corresponding tunnel mileage position L of the current ring.

[0034] Furthermore, the cloud spatial features mentioned in step 4) involve dividing the registered point cloud into a 0.1m×0.1m×0.1m voxel grid, calculating the point density value within each grid, and generating a three-dimensional tensor as the input to the CNN.

[0035] Furthermore, the model structure and training described in step 4) include input branching and feature fusion z. t Output layer, training data, and deployment;

[0036] The input branch processes the bias and TBM parameter sequence through an LSTM branch, and the CNN branch processes the point cloud features; the computation process is divided into forget gate f. t Input gate i t New candidate values Cell state update C t Output gate o t and hidden state output h t The formula is shown below:

[0037] f t =σ(W f ·[h t-1 x t ]+b f )

[0038] i t =σ(W i ·[h t-1 x t ]+b i )

[0039]

[0040] o t =σ(W o ·[h t-1 x t ]+b o )

[0041] h t =σ t ☉tanhC t

[0042] Where x t Let h be the input vector of the t-th ring, containing Δd, Δl, e, Δpitch, Δroll, τ, v, f; t This is the final output temporal feature vector;

[0043] Feature fusion z t The temporal feature vector h output by LSTM t The spatial feature vector v output by the CNN t The formula for splicing and merging is as follows:

[0044] z t =[h t ;v t ]

[0045] The output layer is a fully connected layer that outputs the expected lateral offset sequence and the expected longitudinal cumulative error increment sequence for the next N advances.

[0046] Training data requires a large amount of historical engineering data, including scan data, bias data, and TBM parameters, for supervised learning. The deviation between the model's predictions and the true values ​​is evaluated using a loss function. The MSE loss function is used, as shown below:

[0047]

[0048] Where y true y pred These are the horizontal offset sequence and the vertical error increment sequence for the future N rings, respectively.

[0049] Its deployment involves placing the trained model on edge computing units or industrial control computers with stronger computing power.

[0050] Furthermore, the reinforcement learning described in step 5) includes constructing a reinforcement learning state space, defining an action space and a reward function, training a policy network using the PPO algorithm, generating and outputting control instructions, and establishing a security protection mechanism.

[0051] The process of constructing the reinforcement learning state space involves fusing real-time bias metrics, TBM operating parameters, and the predicted future N-ring offsets from the neural network; the state space is expressed by the following equation:

[0052] s t =[Δd t , Δl t e t ,τ t v t f t pitch t ,roll t Δd {pred,t+1：t+N} , Δl {pred,t+1：t+N} ,...]

[0053] In the formula, Δd t For the current lateral offset, Δl t e represents the current longitudinal cumulative error. t τ represents the current ellipticity. t v is the torque of the cutter head. t To increase speed, f t For propulsion, pitch t For pitch angle, roll t Let Δd be the roll angle. {pred，t+1：t+N} , Δl {pred,t+1：t+N} This is the predicted future N-ring offset sequence.

[0054] Furthermore, the motion space and reward function output the cutterhead steering angle adjustment and propulsion pressure adjustment values, where the motion space is expressed by the following formula:

[0055] at =[Δa t ,Δp t ]

[0056] In the formula, Δa t Δp is the adjustment amount of the cutter head rotation angle. t To advance the pressure adjustment value;

[0057] The reward function is expressed as follows:

[0058] r t =-w1*(|Δd t |+|Δl t |)-w2*(|Δd {pred，t+1} |+…+|Δd {pred,t+N} |+|Δl {pred，t+1) |+…+|Δl {pred,t+N} |)-w3*(|Δa t |+|Δp t |)-w4*I(e t >e threshold )+w5*(|Δd {t-1} |-|Δd t |)+(|Δl {t-1} |-|Δl t |))

[0059] In the formula, w1, w2, w3, w4, and w5 are weighting coefficients, which are positive real numbers, I() is an indicator function, and e threshold This is the ellipticity safety threshold.

[0060] The reward function aims to minimize cumulative bias and path smoothness, and includes a bias penalty term, an action amplitude penalty term, and a bias reduction reward term. Furthermore, it employs the PPO algorithm to train the policy network.

[0061] In the TBM digital twin environment, a pre-trained policy network is deployed, and its parameters are fine-tuned online using actual tunneling data. Its PPO algorithm, as a policy gradient method, incorporates importance sampling and pruning mechanisms in its objective function for stable training.

[0062] Furthermore, it generates and outputs control commands, and generates the cutter head rotation angle and feed pressure setpoints based on the strategy network output actions; the commands are transmitted to the TBM control system through industrial communication protocols.

[0063] Furthermore, its safety protection mechanism automatically switches to manual operation mode and alarms when the action exceeds the equipment's safety range or when multiple consecutive correction cycles fail.

[0064] The beneficial effects of this invention are:

[0065] The present invention provides a real-time planning method for TBM path adjustment based on intelligent algorithms. Through multimodal data fusion and hybrid intelligent algorithms, and by combining 3D laser scanning with TBM data, it performs real-time path planning and correction. After each tunneling operation, the tunneling route is measured and adjusted in real time, solving the problem of path planning lag in complex working conditions in traditional methods. This improves the accuracy and efficiency of tunnel construction, overcomes the "impossible triangle" of accuracy, efficiency and safety in TBM construction, and provides a highly reliable, low-cost intelligent solution for long tunnel projects. It has a fast response speed and saves manpower and material resources. Attached Figure Description

[0066] Figure 1 This is a flowchart of the method of the present invention;

[0067] Figure 2 This is a schematic diagram of the installation position of the shield machine 3D laser scanner frame according to the present invention;

[0068] Figure 3 This is a schematic diagram of the three-dimensional laser scanner frame structure of the present invention;

[0069] In the image: 1. Tunnel boring machine; 2. Tail stair railing; 3. 3D laser scanner frame; 4. Accelerometer; 5. Support frame; 6. Installation platform; 7. Support frame. Detailed Implementation

[0070] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification.

[0071] Exemplary embodiments of the present invention will now be described with reference to the accompanying drawings. However, the present invention may be embodied in many different forms and is not limited to the embodiments described herein. These embodiments are provided to fully and completely disclose the invention and to fully convey its scope to those skilled in the art. The terminology used in the exemplary embodiments illustrated in the drawings is not intended to limit the invention.

[0072] Unless otherwise stated, the terms used herein (including technical terms) have their common meaning as understood by one of ordinary skill in the art. Furthermore, it is understood that terms defined in commonly used dictionaries should be understood to have a meaning consistent with the context of their relevant field, and not to be interpreted as having an idealized or overly formal meaning.

[0073] Example 1

[0074] This embodiment provides a real-time planning method for tunnel TBM path adjustment based on intelligent algorithms, such as... Figure 1 As shown, it includes the following steps:

[0075] Step 1) Data acquisition: Obtain cross-sectional data after each tunneling operation through cross-sectional scanning, and transmit the cross-sectional data and the attitude parameters of the tunnel boring machine to the edge computing unit.

[0076] Step 2) Data processing: point cloud denoising, point cloud registration and data fusion, then axis reconstruction, calculation of deviation index and output of results;

[0077] Among them, the deviation indicators include lateral offset Δd, longitudinal cumulative error Δl, and cross-sectional ellipticity e;

[0078] Step 3) Deviation Judgment and Decision: Set a deviation threshold according to the actual situation of the project, compare the deviation index with the deviation threshold, and if the deviation is within the threshold range, continue tunneling according to the original route;

[0079] If the deviation is greater than the deviation threshold, proceed to step 4);

[0080] Step 4) Neural Network Prediction: Based on the LSTM+CNN hybrid model, the input historical deviation sequence, TBM running parameters and point cloud spatial features are processed, and the expected offset and recommended correction angle for the next N advances are output through the model structure and training.

[0081] The historical deviation sequence is formed by taking the lateral offset Δd, longitudinal cumulative error Δl, cross-sectional ellipticity e, pitch angle change Δpitch, and roll angle change Δroll of the most recent M tunneling rings, constructing a time series matrix, normalizing it, and arranging it in ring order.

[0082] The TBM operating parameters are obtained by extracting the sliding average and standard deviation of the cutterhead torque τ, feed speed v, and feed force f of the current ring and the nearest M ring as feature vectors;

[0083] Step 5) Reinforcement learning to generate correction strategies: Based on the prediction results, with the goal of minimizing cumulative deviation and path smoothness, reinforcement learning is used to generate correction strategies and output control commands for the tool head steering angle and feed pressure adjustment values.

[0084] Example 2

[0085] Based on Example 1, this example provides a real-time planning method for tunnel TBM path adjustment based on intelligent algorithms. The specific process of step 1) is as follows:

[0086] The accelerometer on the 3D laser scanner installed at the tail of the tunnel boring machine (TBM) identifies the TBM's travel status. When it is detected that the TBM has finished one advance and is no longer moving, the 3D laser scanner scans the tunnel that has been excavated and uploads five attitude parameters of the TBM, namely pitch angle, roll angle, thrust, cutterhead torque and thrust speed, to the edge computing unit.

[0087] like Figure 2 As shown, the 3D laser scanner is mounted on the tail stair railing 2 of the tunnel boring machine 1 via a 3D laser scanner frame 3 to maintain its stability. Figure 3 As shown, the 3D laser scanner frame 3 includes a support frame 5, a mounting platform 6, and a support frame 7. The support frame 5 is connected to the rear stair railing 2. The lower end of the support frame 7 is fixed to the support frame 5, and the upper end of the support frame 7 is fixed to the mounting platform 6. The support frame 7 is composed of multiple support rods forming a pyramidal structure. The 3D laser scanner is mounted on the mounting platform 6, and the acceleration sensor is mounted on the 3D laser scanner.

[0088] Example 3

[0089] Based on Example 1, this example provides a real-time planning method for tunnel TBM path adjustment based on intelligent algorithms, the specific steps of which are as follows:

[0090] Step 1: After each tunneling operation, a 3D laser scanner mounted on the tunnel boring machine is used to collect data to obtain cross-sectional data after tunneling. At the same time, the attitude parameters of the tunnel boring machine are transmitted to the edge computing unit to complete the data collection work. Furthermore, the attitude parameters are the pitch angle, roll angle, thrust, cutterhead torque and thrust speed of the tunnel boring machine.

[0091] Step 2: After obtaining the scan data, extract key indicators: lateral offset, longitudinal cumulative error, and cross-sectional ellipticity. Process the data to obtain the axis, and then compare it with the design axis to obtain the deviation.

[0092] Step 3: After the deviation calculation is completed, set the deviation threshold according to the actual situation of the project. If it is greater than the threshold, proceed to step 4. If it is less than the threshold, continue tunneling according to the original route without changing the route.

[0093] Step 4: Perform neural network prediction. Input: historical deviation sequence, TBM operating parameters. Output: expected offset and recommended correction angle for the next N advances.

[0094] Step 5: Based on the prediction results, with the goal of minimizing the cumulative deviation and path smoothness, use reinforcement learning to generate a correction strategy and output control commands: cutterhead rotation angle and propulsion pressure adjustment value.

[0095] Accelerometers are used to identify the travel status of the tunnel boring machine. When the accelerometer detects that the tunnel boring machine has stopped traveling after one advance, the 3D laser scanner starts to work and scans the tunnel that has been excavated. At the same time, it uploads five attitude parameters of the tunnel boring machine, namely pitch angle, roll angle, thrust, cutterhead torque and thrust speed, to the edge box (edge ​​computing unit).

[0096] Step 2 involves the calculation of the scanned data. First, data preprocessing is performed, including point cloud denoising, point cloud registration, and data fusion. Then, axis reconstruction, deviation index calculation, and result output are performed.

[0097] Furthermore, point cloud denoising involves filtering out flying points and dust noise generated during the scanning process using statistical filtering. Statistical filtering is based on neighborhood distance statistics. The point cloud should exhibit a relatively uniform distribution, meaning the average distance between each point and its K nearest neighbors should not exceed a threshold. Points exceeding this threshold are considered outliers and removed. The threshold is determined using the following formula:

[0098] D max =μ+std×σ

[0099]

[0100] μ is the average distance between the point to be judged and its K nearest neighbors; σ is the standard deviation of the distance; std is a user-defined standard deviation coefficient used to control the influence of the standard deviation of the distance on the distance threshold; D i Let i be the distance between point i and its nearest neighbor.

[0101] Furthermore, the point cloud registration involves accurately registering the point cloud from the current scan with the previous scan to determine the actual position and pose changes of the TBM in the global coordinate system. The ICP algorithm is used for registration, with the goal of minimizing the distance between corresponding points. The objective function is shown below:

[0102]

[0103] Where R is the rotation matrix, T is the translation vector, n is the number of effective corresponding point pairs, and P i It is a point cloud to be registered, q i It is a reference point cloud, and R and T are solved iteratively.

[0104] Furthermore, the data fusion involves adding a unified timestamp to the registered point cloud data and the TBM attitude and operating parameters collected at the same time, and then associating them with the corresponding tunneling mileage.

[0105] Furthermore, the axis reconstruction includes the extraction of the design axis, the actual axis, and axis alignment. The design axis is the theoretical tunnel centerline for that mileage section obtained from the design documents. The actual axis is extracted from the registered point cloud data to obtain the actual centerline of the current section. Through section fitting, the point cloud is sliced ​​on multiple cross-sections perpendicular to the excavation direction, and then ellipse fitting is performed to obtain the actual tunnel axis. The axis alignment involves aligning the extracted actual axis with the design axis in space.

[0106] Furthermore, the deviation indices include lateral offset, longitudinal cumulative error, and cross-sectional ellipticity calculations. Lateral offset is the distance between the actual center point and the design center point on the calculated cross-section, perpendicular to the design axis. Longitudinal cumulative error is the total length difference between the actual axis and the design axis along the design axis from the starting point to the current mileage. Cross-sectional ellipticity is the ratio of the major axis to the minor axis of the fitted ellipse on key cross-sections, used to evaluate whether the tunnel cross-section has experienced compression deformation or irregular shape due to over-excavation / under-excavation.

[0107] The output includes the lateral offset Δd of the current ring, the longitudinal cumulative error Δl, the cross-sectional ellipticity e, and the corresponding tunnel mileage position L.

[0108] Step 3 is deviation judgment and decision-making. After the deviation calculation is completed, a deviation threshold is set according to the actual situation of the project. If the deviation is greater than the threshold, proceed to step 4. If the deviation is less than the threshold, continue tunneling according to the original route without changing the route.

[0109] Step 4 involves neural network prediction. The LSTM+CNN hybrid model is well-suited for processing spatiotemporal sequence data and predicting future offsets. This step is divided into input data preprocessing and model structure and training.

[0110] Furthermore, its input data preprocessing includes historical deviation sequences, TBM operating parameters, and point cloud spatial features. Even further, the historical deviation sequence of the input data consists of the lateral offset Δd, longitudinal cumulative error Δl, cross-sectional ellipticity e, and attitude angle changes (i.e., pitch angle change Δpitch and roll angle change Δroll) of the most recent M tunneling rings, forming a time series matrix, which is then normalized and arranged in ring order.

[0111] Furthermore, its TBM operating parameters are derived by extracting the sliding average and standard deviation of the cutterhead torque τ, propulsion speed v, and propulsion force f of the current ring and the nearest M ring as feature vectors. Its point cloud spatial features are obtained by dividing the registered point cloud into a 0.1m × 0.1m × 0.1m voxel grid, statistically analyzing the point density values ​​within each grid, and generating a 3D tensor as input to the CNN.

[0112] Furthermore, its model structure and training include input branches, feature fusion, output layers, training data, and deployment. Even further, its input branches process bias and TBM parameter sequences through an LSTM branch, and point cloud features through a CNN branch; its computation process is divided into forgetting gates. t Input gate i t New candidate values Cell state update C t Output gate o t and hidden state output ht The formula is shown below:

[0113] f t =σ(W f ·[h t-1 x t ]+b f )

[0114] i t =σ(W i ·[h t-1 x t ]+b i )

[0115]

[0116] o t =σ(W o ·[h t-1 x t ]+b o )

[0117] h t =o t ☉tanhC t

[0118] Where x t Let h be the input vector of the t-th ring, containing Δd, Δl, e, Δpitch, Δroll, τ, v, f; t This is the final output temporal feature vector.

[0119] Furthermore, its features are fused into a temporal feature vector h output by LSTM. t The spatial feature vector v output by the CNN t The formula for splicing and merging is as follows:

[0120] z t =[h t ;v t ]

[0121] Furthermore, its output layer is a fully connected layer that outputs the expected lateral offset sequence and the expected longitudinal cumulative error increment sequence for the next N advances.

[0122] Furthermore, its training data requires a large amount of historical engineering data, including scan data, bias data, and TBM parameters, for supervised learning. The loss function used is MSE, as shown below:

[0123]

[0124] Where y true y predThese are the horizontal offset sequence and the vertical error increment sequence for the future N rings, respectively.

[0125] Its deployment involves placing the trained model on edge computing units or industrial control computers with stronger computing power.

[0126] Step 5 involves generating a correction policy using reinforcement learning. To address the optimal decision-making problem in dynamic environments, this step transforms the prediction results into executable control instructions through reinforcement learning. This includes constructing the reinforcement learning state space, defining the action space and reward function, training the policy network using the PPO algorithm, generating and outputting control instructions, and implementing a safety protection mechanism.

[0127] Furthermore, the reinforcement learning state space is constructed by fusing real-time bias indicators, TBM operating parameters, and the expected offset of future N-rings predicted by the neural network. The state space is expressed by the following equation:

[0128] s t =[Δd t Δl t e t , τ t ,v t f t pitch t ,roll t Δd {pred,t+1：t+N} ,Δl {pred,t+1：t+N} ,…]

[0129] Where Δd t For the current lateral offset, Δl t e represents the current longitudinal cumulative error. t τ represents the current ellipticity. t v is the torque of the cutter head. t To increase speed, f t For propulsion, pitch t For pitch angle, roll t Let Δd be the roll angle. {pred，t+1：t+N} , Δl (pred，t+1：t+N} This is the predicted future N-ring offset sequence.

[0130] Furthermore, it defines the motion space and reward function: the motion space outputs the cutterhead steering angle adjustment and the propulsion pressure adjustment value, where the motion space is expressed by the following formula:

[0131] α t =[Δa t , Δp t ]

[0132] Where Δa t Δp is the adjustment amount of the cutter head rotation angle. t To advance the pressure adjustment value.

[0133] The reward function is expressed as follows:

[0134] r t =-w1*(|Δd t |+|Δl t |)-w2*(|Δd {pred,，t+1} |+…+|Δd {pred，t+N} |+|Δl {pred,t+1} |+…+|Δl {pred，t+N} |)-w3*(|Δa t |+|Δp t |)-w4*I(e t >e threshold )+w5*((|Δd {t-1} |-|Δd t |)+(|Δl {t-1} |-|Δl t |))

[0135] Where w1, w2, w3, w4, and w5 are weighting coefficients, which are positive real numbers, I() is the indicator function, and e threshold This is the ellipticity safety threshold.

[0136] The reward function aims to minimize cumulative deviation and path smoothness, and includes a deviation penalty term, an action amplitude penalty term, and a deviation reduction reward term. Furthermore, it employs the PPO algorithm to train the policy network, pre-training the network in a TBM digital twin environment, and then fine-tuning the network parameters online using actual tunneling data after deployment. The PPO algorithm, as a policy gradient method, incorporates importance sampling and pruning mechanisms in its objective function for stable training. Further, it generates and outputs control commands, generating cutterhead steering angle and feed pressure setpoints based on the policy network's output actions; these commands are transmitted to the TBM control system via an industrial communication protocol. Finally, its safety protection mechanism automatically switches to manual operation mode and triggers an alarm when actions exceed the equipment's safety range or when multiple consecutive correction cycles fail.

[0137] The above examples are merely illustrative of the present invention and do not constitute a limitation on the scope of protection of the present invention. All designs that are the same as or similar to the present invention are within the scope of protection of the present invention.

Claims

1. A real-time planning method for tunnel TBM path adjustment based on intelligent algorithms, characterized in that: Includes the following steps: Step 1) Data acquisition: Obtain cross-sectional data after each tunneling operation through cross-sectional scanning, and transmit the cross-sectional data and the attitude parameters of the tunnel boring machine to the edge computing unit. Step 2) Data processing: point cloud denoising, point cloud registration and data fusion, then axis reconstruction, calculation of deviation index and output of results; Among them, the deviation indicators include lateral offset Δd, longitudinal cumulative error Δl, and cross-sectional ellipticity e; Step 3) Deviation Judgment and Decision: Set a deviation threshold according to the actual situation of the project, compare the deviation index with the deviation threshold, and if the deviation is within the threshold range, continue tunneling according to the original route; If the deviation is greater than the deviation threshold, proceed to step 4); Step 4) Neural Network Prediction: Based on the LSTM+CNN hybrid model, the input historical deviation sequence, TBM running parameters and point cloud spatial features are processed, and the expected offset and recommended correction angle for the next N advances are output through the model structure and training. The historical deviation sequence is formed by taking the lateral offset Δd, longitudinal cumulative error Δl, cross-sectional ellipticity e, pitch angle change Δpitch, and roll angle change Δroll of the most recent M tunneling rings, constructing a time series matrix, normalizing it, and arranging it in ring order. The TBM operating parameters are obtained by extracting the sliding average and standard deviation of the cutterhead torque τ, feed speed v, and feed force f of the current ring and the nearest M ring as feature vectors; Step 5) Reinforcement learning to generate correction strategies: Based on the prediction results, with the goal of minimizing cumulative deviation and path smoothness, reinforcement learning is used to generate correction strategies and output control commands for the tool head steering angle and feed pressure adjustment values.

2. The real-time planning method for tunnel TBM path adjustment based on intelligent algorithms according to claim 1, characterized in that: The specific process of step 1) is as follows: The accelerometer on the 3D laser scanner installed at the tail of the tunnel boring machine (TBM) identifies the TBM's travel status. When it is detected that the TBM has finished one advance and is no longer moving, the 3D laser scanner scans the tunnel that has been excavated and uploads five attitude parameters of the TBM, namely pitch angle, roll angle, thrust, cutterhead torque and thrust speed, to the edge computing unit.

3. The real-time planning method for tunnel TBM path adjustment based on intelligent algorithms according to claim 1, characterized in that: The point cloud denoising described in step 2) is based on the statistical distance of neighboring points. By using statistical filtering, flying points and dust noise generated during the scanning process are filtered out, so that the average distance between each point in the point cloud and its K nearest neighbors does not exceed the threshold. If it does, it is removed. The threshold is obtained by the following formula: D max = μ + std × σ In the formula, μ is the average distance between the point to be judged and its K nearest neighbors; σ is the standard deviation of the distance; std is a user-defined standard deviation coefficient used to control the influence of the standard deviation of the distance on the distance threshold; D i Let i be the distance between point i and its nearest neighbor.

4. The real-time planning method for tunnel TBM path adjustment based on intelligent algorithms according to claim 1, characterized in that: The specific process of point cloud registration described in step 2) is as follows: The point cloud from this scan is precisely registered with the previous scan to determine the actual position and pose changes of the TBM in the global coordinate system. The ICP algorithm is used for registration, with the goal of minimizing the distance between corresponding points. The objective function is shown below: Where R is the rotation matrix, T is the translation vector, n is the number of effective corresponding point pairs, and P i It is a point cloud to be registered, q i It is a reference point cloud, and R and T are solved iteratively.

5. The real-time planning method for tunnel TBM path adjustment based on intelligent algorithms according to claim 1, characterized in that: The specific process for axis reconstruction and deviation index calculation in step 2) is as follows: (1) Obtain the design axis of this mileage section from the design documents; (2) Extract the actual centerline of the current section from the registered point cloud data. Through section fitting, slice the point cloud on multiple cross sections perpendicular to the tunneling direction, and then perform ellipse fitting to obtain the actual axis. (3) Align the design axis with the actual axis and calculate the deviation index; (4) The lateral offset Δd is obtained by calculating the distance between the actual center point and the design center point in the direction perpendicular to the design axis. The longitudinal cumulative error Δl is obtained by calculating the total length difference between the actual axis and the design axis in the direction along the design axis from the starting point to the current mileage. The ellipticity e of the section is obtained by calculating the ratio of the major axis to the minor axis of the fitted ellipse.

6. The real-time planning method for tunnel TBM path adjustment based on intelligent algorithms according to claim 1, characterized in that: The output of step 2) includes the lateral offset Δd, longitudinal cumulative error Δl, cross-sectional ellipticity e, and the corresponding tunnel mileage position L of the current ring.

7. The real-time planning method for tunnel TBM path adjustment based on intelligent algorithms according to claim 1, characterized in that: The cloud spatial features mentioned in step 4) involve dividing the registered point cloud into a 0.1m×0.1m×0.1m voxel grid, calculating the point density value in each grid, and generating a three-dimensional tensor as the input to the CNN.

8. The real-time planning method for tunnel TBM path adjustment based on intelligent algorithms according to claim 1, characterized in that: The model structure and training described in step 4) include input branching and feature fusion z. t Output layer, training data, and deployment; The input branch processes the bias and TBM parameter sequence through the LSTM branch, and the CNN branch processes the point cloud features; Will Its calculation process is divided into forgetting gate f t Input gate i t New candidate values Cell state update C t Output gate o t and hidden state output h t The formula is shown below: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t ⊙tanhC t Where x t Let h be the input vector of the t-th ring, containing Δd, Δl, e, Δpitch, Δroll, τ, v, f; t This is the final output temporal feature vector; Feature fusion z t The temporal feature vector h output by LSTM t The spatial feature vector v output by the CNN t The formula for splicing and merging is as follows: from t =[h t ;in t ] The output layer is a fully connected layer that outputs the expected lateral offset sequence and the expected longitudinal cumulative error increment sequence for the next N advances.

9. The real-time planning method for tunnel TBM path adjustment based on intelligent algorithms according to claim 1, characterized in that: The reinforcement learning described in step 5) includes constructing a reinforcement learning state space, defining an action space and a reward function, training a policy network using the PPO algorithm, generating and outputting control commands, and establishing a security protection mechanism. The process of constructing the reinforcement learning state space involves integrating real-time deviation indicators, TBM operating parameters, and the expected future N-ring offsets predicted by the neural network. The state space is expressed by the following equation: s t =[Δd t ,Δl t ,e t ,τ t ,v t ,f t ,pitch t ,roll t ,Δd {pred,t+1：t+N} ,Δl {pred，t+1：t+N} ,…] In the formula, Δd t For the current lateral offset, Δl t e represents the current longitudinal cumulative error. t τ represents the current ellipticity. t v is the torque of the cutter head. t To increase speed, f t For propulsion, pitch t For pitch angle, roll t Let Δd be the roll angle. {pred，t+1：t+N} , Δl {pred，t+1：t+N} This is the predicted future N-ring offset sequence.

10. A real-time planning method for tunnel TBM path adjustment based on intelligent algorithms according to claim 9, characterized in that: The motion space and reward function output the cutterhead rotation angle adjustment and the propulsion pressure adjustment value in the motion space, where the motion space is expressed by the following formula: a t =[D a t ,Δp t ] In the formula, Δa t Δp is the adjustment amount of the cutter head rotation angle. t To advance the pressure adjustment value; The reward function is expressed as follows: r t =-w1*(|Δd t |+|Δl t |)-w2*(|Δd {pred，t+1} |+…+|Δd {pred，t+N} |+|Δl {pred，t+1} |+…+|Δl {pred，t+N} |)-w a *(|Δa t |+|Δp t |)-w4*I(e t >e threshold )+w5*((|Δd {t-1} |*|Δd t |)+(|Δl {t-1} |-|Δl t |)) In the formula, w1, w2, w3, w4, and w5 are weighting coefficients, which are positive real numbers, and I... O e is an indicator function threshold This is the ellipticity safety threshold.

Images