Shield whole process intelligent control method and system based on multi-task model coupling

Abstract

This invention discloses a shield tunneling full-process intelligent control method and system based on multi-task model coupling, belonging to the interdisciplinary technical field of industrial intelligent control systems and tunnel engineering digital twins. The method includes the following steps: S1, standardized data acquisition and cross-stage storage; S2, construction of a multi-stage coupled prediction model and parameter output; S3, multi-objective collaborative optimization and adaptive control decision generation; S4, control command execution and real-time monitoring feedback closed loop. This invention employs the above-mentioned shield tunneling full-process intelligent control method based on multi-task model coupling to achieve automatic acquisition, cross-stage integration, and real-time sharing of multi-source heterogeneous data; improve prediction accuracy and generalization ability; solve problems such as poor process coordination and insufficient control precision; avoid settlement risks; and achieve precise and intelligent management and control.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary technical field of industrial intelligent control systems and digital twins for tunnel engineering, specifically involving an intelligent control method for the entire shield tunneling process based on multi-task model coupling. Background Technology

[0002] The rapid expansion of subway mileage and scale has placed higher demands on the safety of tunnel construction and the control of its impact on the surrounding environment. Due to its significant advantages such as high construction speed, minimal disturbance to the surrounding environment, and high degree of mechanization, the shield tunneling method is widely used in subway tunnel construction.

[0003] However, existing technologies have the following shortcomings: For example, TPC software only covers shield tunneling PLC data and segment quality data, lacking correlation with geological information; some early warning platforms rely on manual data uploads, leading to delays and error risks, and various data types cannot be integrated and shared across dimensions, forming "data silos," resulting in a lack of comprehensive support for construction decisions and difficulty in coping with dynamic changes in complex working conditions; many models focus on single-stage modeling, or only use formulas to determine the excavation face state, or only rely on machine learning to optimize tunneling parameters, failing to form a complete prediction system encompassing geological conditions, attitude, and settlement. Furthermore, many models rely on empirical parameters or single data-driven approaches, without integrating physical mechanisms, making them prone to prediction deviations when geological conditions change abruptly, and exhibiting weak generalization ability; most are "segmented management," lacking coordination between processes such as advancement, assembly, and settlement control. Conventional shield tunneling requires stopping to assemble segments, and frequent start-stop operations can easily lead to settlement and equipment failure. Simultaneously, most systems rely on manual parameter adjustments or only achieve single-objective optimization, failing to dynamically adapt to multiple objectives such as advancement speed and earth pressure balance, resulting in insufficient control precision in complex geological formations.

[0004] Therefore, a new method is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent control method for the entire shield tunneling process based on multi-task model coupling. This method realizes automatic acquisition, cross-stage integration and real-time sharing of multi-source heterogeneous data; improves prediction accuracy and generalization ability; solves problems such as poor process coordination and insufficient control accuracy; avoids settlement risk; and achieves precise and intelligent management and control.

[0006] To achieve the above objectives, this invention provides an intelligent control method for the entire shield tunneling process based on multi-task model coupling, comprising the following steps: S1. Store data in a standardized data format, relying on the B / S dynamic system architecture and TCP / IP network configuration to achieve automatic data collection and standardized storage; S2. Construct a multi-stage coupled prediction model that includes a front chamber earth pressure prediction model, a shield attitude prediction model, and a surface settlement prediction model. The front chamber earth pressure prediction model is based on two-dimensional coupled Markov chain to simulate the stratum distribution, combined with Bayesian machine learning to complete the soil parameters, and calculates the earth chamber pressure setpoint through soil mechanics formula. The shield tunnel attitude prediction model adopts an improved Transformer architecture, embedding a finite element proxy model to form a physical information fusion model, and outputs attitude prediction results. The surface subsidence prediction model uses a random forest model with ground loss data, combined with the ground loss rate identified by YOLOv8, to output the final subsidence prediction value. S3. Define a multi-objective collaborative optimization engine, construct a system containing loss functions and equipment safety constraints; obtain a Pareto front candidate solution set through a genetic algorithm solver, and output the optimized control parameters. S4. The optimized control parameters from S3 are transmitted to the tunnel boring machine execution system to drive the equipment to adjust the construction status; and the real-time collected data is fed back to the multi-stage coupled prediction model and the multi-objective collaborative optimization engine to form a closed-loop control.

[0007] Preferably, the standardized data format used to store data in S1 specifically refers to: The system uses "Project Name" as the primary key to link multi-source heterogeneous data, including project overview, construction parameters, tunnel boring machine attitude, surface settlement, and stratum parameters. Each data table is associated through the primary key "Project Name" and subkeys, including ring number and date. A hybrid storage mode of MySQL and Excel is adopted, with MySQL serving as the backend database to support efficient data management and real-time access, and Excel used for intelligent auxiliary decision-making and suggestion interaction.

[0008] Preferably, S2 employs a hybrid model architecture combining convolutional neural networks and long short-term memory networks to achieve earth pressure prediction, which includes the following sub-steps: S201. Organize and preprocess the multi-source heterogeneous data collected during the tunnel boring machine construction process; S202. Based on the physical characteristics of tunnel construction parameters, the engineering features are divided into four categories: deviation, velocity, grouting pressure, and zoned pressure. A grouped convolution strategy is used to process the four categories of features independently. S203. Utilize the long short-term memory characteristics of the long short-term memory network to dynamically capture long and short-term dependencies in time-series data; S204. By integrating the spatial features extracted by the convolutional neural network with the temporal features learned by the long short-term memory network through the fully connected layer, the output node is designed according to the engineering requirements for predicting the earth pressure in the shield tunneling front chamber, and the prediction result of the earth pressure in the front chamber is finally output.

[0009] Preferably, the improved Transformer architecture for the shield tunnel attitude prediction model in S2 includes: Let the input data dimension be m×1, the first... i The input parameters for each sample are a i ={ a 1 ,a 2 ,...,a m}, the position vector p= { p 1 ,p 2 ,...,p m Add the input parameter to the input parameter to obtain the new input parameter. b i ={ b 1 ,b 2 ,...,b m The calculation formula is: ; The input parameters are passed to an encoder module containing M encoder units. Each unit consists of a self-attention layer and a feedforward neural network layer. Residual connections and layer normalization are performed after each layer. The formula for calculating the self-attention layer is: ; ; ; in, For query matrix; The key matrix; It is a value matrix; The input vector; For query vector; The key vector; It is a value vector; The result of the self-attention mechanism is represented as follows: ; Where softmax is the normalization function; The encoded data is fed into a decoder module containing N decoder units. Each decoder unit sequentially includes a self-attention layer, an encoder-decoder attention layer, and a feedforward neural network layer. The encoder-decoder attention layer reuses the self-attention layer. After each layer, residual joins and layer normalization are performed. After the data is processed by the decoder, initial predicted values ​​are generated through a feedforward layer and residual-normalization operations. The error between the initial predicted values ​​and the measured values ​​is calculated, and the backpropagation algorithm is used for iterative updates. , , and feedforward layer weights.

[0010] Preferably, in S2, the physical information fusion model of the shield attitude prediction model is constructed in the following way: An embedded finite element numerical simulation proxy model is used, which consists of an input layer, a hidden layer, and an output layer. The input parameters are processed by the finite element proxy model to obtain the physical prediction results. ; The weighted fusion is calculated using the following formula: ; in, Weights for physical information; The prediction results for the Transformer model; For fusion prediction results; By comparison The deviation from the measured value is dynamically adjusted. .

[0011] Preferably, in S2, the YOLOv8 identification of formation loss rate by the surface subsidence prediction model includes: The formula for calculating the formation loss rate is as follows: ; in, Formation loss rate; This represents the actual volume of the construction waste. To design the excavation volume.

[0012] Preferably, S2 also includes: To collect earth pressure; This is for attitude deviation; For tunnel burial depth; This refers to the water level depth. As a spatiotemporal identifier, soil pressure is collected and correlated with tunnel depth, water level depth, and spatiotemporal identifier information to form a feature vector. ; In the random forest model, a formation loss rate loss is introduced, and the total loss function is expressed as: ; in, Total loss; This is the mean square error loss; The final settlement prediction value is obtained by calculation using the following formula: ; in, This represents the final predicted settlement value. These are the predicted values ​​from the random forest model.

[0013] Preferably, the loss functions in S3 include the propulsion velocity loss function, the earth pressure loss function, the attitude loss function, and the settlement loss function; specifically: Propulsion speed loss function Represented as: ; in, For control parameter vectors; The tunnel boring machine's advance speed; The target speed for tunnel boring machine (TBM) advancement; For coefficients; Earth pressure loss function Represented as: ; in, This is the starting time of the integration; The time length of the integration; In order to be in Under its influence, at all times The actual earth pressure value; For a moment The corresponding target earth pressure value; For coefficients; In order to be in The average value of earth pressure over the time interval under the action of the earth; Shield attitude loss function Represented as: ; in, To correspond to the positional deviations of the shield in the x-direction (horizontal transverse direction) and y-direction, as well as the angle of the shield axis. ; for Dimension weight coefficients; In order to be in Under the action, the tunnel boring machine is The dimensional attitude deviation value; For the shield tunneling k The maximum allowable pose deviation value for a given dimension; For coefficients; The gradient of the shield tunnel attitude deviation; Surface subsidence loss function Represented as: ; in, The number of surface subsidence monitoring points; For coefficients; In order to be in u Under the action, the first Surface subsidence values ​​at each monitoring point; For coefficients; This is the safe threshold for land subsidence.

[0014] Preferably, the equipment safety constraints in S3 include torque safety constraints, cylinder pressure difference constraints, grouting volume lower limit constraints, and axis continuity constraints; specifically: Torque safety constraints Represented as: ; in, In order to be in u The actual torque value of the tunnel boring machine under action; This represents the maximum allowable torque value for the tunnel boring machine. Hydraulic cylinder pressure difference constraint Represented as: ; in, This refers to the pressure value of the hydraulic cylinder on the left side of the tunnel boring machine. This refers to the pressure value of the hydraulic cylinder on the right side of the tunnel boring machine. This represents the average pressure of the left and right hydraulic cylinders. This represents the maximum permissible value of the relative proportion of the pressure difference between the left and right hydraulic cylinders. Grouting volume lower limit constraint Represented as: ; in, This is the minimum allowable value for the grouting volume; In order to be in u The actual grouting volume under the action; Axis continuity constraint Represented as: ; in, The coordinates represent the length of the axis along the tunnel boring machine's advance direction; This represents the maximum permissible value for the curvature of the axis.

[0015] This invention also provides an intelligent control system for the entire shield tunneling process based on multi-task model coupling, including: A multi-stage coupled prediction module is connected to the data acquisition and storage module and is used to construct a multi-stage coupled prediction model that includes a front chamber earth pressure prediction model, a shield attitude prediction model and a surface settlement prediction model. The front chamber earth pressure prediction model is based on two-dimensional coupled Markov chain to simulate the stratum distribution, combined with Bayesian machine learning to complete the soil parameters, and calculates the earth chamber pressure setpoint through soil mechanics formula. The shield tunnel attitude prediction model adopts an improved Transformer architecture, embedding a finite element proxy model to form a physical information fusion model, and outputs attitude prediction results. The surface subsidence prediction model uses a random forest model with ground loss data, combined with the ground loss rate identified by YOLOv8, to output the final subsidence prediction value. A multi-objective optimization decision module, connected to the multi-stage coupled prediction module, is used to define a multi-objective collaborative optimization engine, construct a loss function and equipment safety constraints; obtain a Pareto front candidate solution set through a genetic algorithm solver, and output the optimized control parameters. The control execution and feedback module is connected to the multi-objective optimization decision module. It is used to transmit the optimized control parameters to the tunnel boring machine execution system to drive the equipment to adjust the construction state; and to feed back the real-time collected data to the multi-stage coupled prediction model and the multi-objective collaborative optimization engine to form a closed-loop control.

[0016] Therefore, the present invention adopts the above-mentioned intelligent control method for the entire shield tunneling process based on multi-task model coupling. Compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) This invention adopts a unified and standardized data format, uses “project name” as the main key to string together multi-source heterogeneous data such as construction parameters, geological distribution, and shield attitude, and combines a hybrid storage mode of MySQL and Excel to overcome the “data silo” problem of traditional systems, which is characterized by single-type data processing, reliance on manual data upload and the risk of delay and error, and the inability of multi-source heterogeneous data to be integrated and shared across dimensions. This invention achieves automatic data collection, cross-stage integration and real-time sharing. (2) This invention constructs a multi-stage coupled model of “stratum simulation-attitude prediction-settlement calculation”, uses a two-dimensional coupled Markov chain to simulate stratum distribution, improves the Transformer architecture to integrate physical information to predict attitude, and forms a full-process prediction system of “physical mechanism + data-driven” dual engine integration. It overcomes the problem of traditional technology’s single-stage modeling and failure to form a full-chain prediction system of stratum-attitude-settlement, and significantly improves the model’s generalization ability and full-chain prediction accuracy. (3) This invention establishes a full-process adaptive strategy covering propulsion, attitude and settlement. By defining loss functions and safety constraints through a multi-objective optimization mathematical model, it dynamically adapts among multiple objectives. This overcomes the problems of traditional systems' "segmented management" which leads to a lack of coordination in the process, the need to stop the machine to assemble the segments causing settlement and equipment failure, and the reliance on manual parameter adjustment or optimization of only a single objective, resulting in insufficient control accuracy in complex strata. It avoids the settlement risk caused by frequent start-stop operations and achieves precise control with "intelligent as the main method and manual as the auxiliary method".

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a diagram of the front chamber earth pressure prediction model architecture for an embodiment of the intelligent control method for the entire shield tunneling process based on multi-task model coupling of the present invention. Figure 2 This is a schematic diagram of the Transformer algorithm for physical information fusion in an embodiment of the intelligent control method for the entire shield tunneling process based on multi-task model coupling of the present invention. Figure 3 This is a flowchart illustrating the construction and training process of the surface settlement machine learning model in an embodiment of the intelligent control method for the entire shield tunneling process based on multi-task model coupling of the present invention. Figure 4 This is a diagram illustrating the intelligent control process architecture for shield tunneling based on a multi-task model coupling intelligent control method for the entire shield tunneling process according to the present invention. Detailed Implementation

[0019] 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. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.

[0020] Example 1 like Figures 1-4 As shown, the intelligent control method for the entire shield tunneling process based on multi-task model coupling of the present invention includes: S1. MySQL relational database is used as the core storage medium, configured as a [Server Machine] server type (medium memory usage, supporting server application operation) and a [Multifunction Database] purpose (optimizing the performance of InnoDB and MyISAM dual engines). TCP / IP network connection is enabled, and data security and operability are ensured through permission management and a graphical interface. Excel is used as an auxiliary interaction medium to support data aggregation and uploading by all participants. Using "Project Name" as the primary key to connect all data tables, a multi-dimensional standardized data table structure is defined, including: Project Overview Table: Stores basic information such as "Project Name", "Shield Type", "Shield Manufacturer", and "Shield Outer Diameter"; Construction Parameter Table: Using “Ring Number”, “Construction Date”, and “Construction Time” as subkeys, it records real-time construction data such as earth pressure, total thrust, cutterhead parameters, attitude parameters, and grouting parameters; Tunnel Boring Machine Attitude Table: It stores attitude data such as roll angle, slope angle, and cut / tail deviation by linking construction parameters through subkeys; Surface Settlement Monitoring Table: Using "Date" as the subkey, record the axial and bilateral settlement data of the cross-sections within ±50 rings before and after the current ring number; Stratigraphic Parameter Table: Using "ring number" as the subkey, it binds the geological data such as soil layer number, water content, density, and mechanical parameters at the corresponding location; Surrounding environment data table: including information such as the location relationship of surrounding structures (overpass / underpass / parallel relationship, minimum clearance) and soil reinforcement methods; After each data source is entered into an Excel file in a uniform format, it is uploaded to the shield tunnel construction management and control platform. The data format and logical consistency are checked through the association verification of primary keys and subkeys, thus eliminating "data silos". S2. Construct a multi-stage coupled prediction model, which includes a front chamber earth pressure prediction model, a shield attitude prediction model, and a surface settlement prediction model. The earth pressure prediction model employs a hybrid model architecture combining convolutional neural networks and long short-term memory networks to predict earth pressure, and includes the following sub-steps: S201. Organize and preprocess the multi-source heterogeneous data collected during the tunnel boring machine construction process; S202. Based on the physical characteristics of tunnel construction parameters, the engineering features are divided into four categories: deviation, velocity, grouting pressure, and zoned pressure. A grouped convolution strategy is used to process the four categories of features independently. S203. Utilize the long short-term memory characteristics of the long short-term memory network to dynamically capture long and short-term dependencies in time-series data; S204. By integrating the spatial features extracted by the convolutional neural network with the temporal features learned by the long short-term memory network through the fully connected layer, the output node is designed according to the engineering requirements for predicting the earth pressure in the shield tunneling front chamber, and the prediction result of the earth pressure in the front chamber is finally output.

[0021] The soil parameter completion program relies on the Shanghai soft clay database to construct a multivariate probability distribution model and uses kernel density estimation (KDE) to characterize the correlation between parameters. Using Bayesian machine learning, it first defines parameter dependencies (e.g., the at-rest lateral pressure coefficient K0 depends on the internal friction angle and water content), and constructs a directed acyclic graph (DAG) based on this. Then, it estimates the conditional probability table (CPT) using maximum a posteriori probability (MAP) to form a prior knowledge network. Finally, combined with the established database probability distribution model, it establishes a probability distribution model suitable for specific sites to complete the missing soil parameters such as the internal friction angle and at-rest lateral pressure coefficient in the geological survey report. The coupling of the stratigraphic distribution simulation program and the soil parameter prediction program first determines the soil type at the current location through coupled Markov chain stratigraphic reconstruction and matches it with the corresponding sub-distribution in the Shanghai soft clay database; inputting some parameters revealed in the geological survey report, the program completes the missing parameters such as the internal friction angle and the static lateral pressure coefficient through a Bayesian model to form a complete set of soil parameters; and using classical soil mechanics formulas, the program calculates the self-weight stress and horizontal lateral pressure values ​​of the soil in front of the shield to form the shield soil chamber pressure setpoint. The shield tunneling attitude prediction model adopts an improved Transformer architecture, embedding a finite element proxy model to form a physical information fusion model, specifically: The Transformer architecture was selected to build a data-driven prediction model, which was optimized for shield tunneling attitude prediction tasks. Let the input data dimension be m×1, the first... i The input parameters for each sample are a i ={ a 1 ,a 2 ,...,a m}, the position vector p= { p 1 ,p 2 ,...,p m Add the input parameter to the input parameter to obtain the new input parameter. b i ={ b 1 ,b 2 ,...,b m}, incorporating parameter temporal and spatial correlation information, the calculation formula is as follows: ; The input parameters are passed to an encoder module containing M encoder units. Each unit consists of a self-attention layer and a feedforward neural network layer. Residual connections and layer normalization are performed after each layer. The formula for calculating the self-attention layer is: ; ; ; in, For query matrix; The key matrix; It is a value matrix; The input vector; For query vector; The key vector; It is a value vector; In this matrix, the first i Line 1 j Column elements Indicates the first j The input parameter for the first i The influence weights of each input parameter. Value vector. This indicates the influence of the initial value of each input parameter on the prediction result; The final result of the self-attention mechanism is achieved by using the weight matrix. AND value vector Multiply and normalize to obtain: ; Where softmax is the normalization function; The encoded data enters a decoder module containing N decoder units. Each unit sequentially contains a self-attention layer, an encoder-decoder attention layer, and a self-attention layer that directly reuses the encoder-decoder attention layer. The feedforward neural network layer is also equipped with residual connections and layer normalization; After the data is processed by the decoder, predicted values ​​are generated through a feedforward layer and residual-normalization operations. The error between the predicted and measured values ​​is calculated, and the backpropagation algorithm is used for iterative updates. , , and feedforward layer weights; To improve the generalization performance of the model, a finite element method (FEM) numerical simulation surrogate model is embedded to form a fusion prediction system. The input parameters are processed by the input layer, hidden layer, and output layer of the FEM surrogate model, which is composed of a multilayer perceptron (MLP), to obtain the physical prediction results. Weight fusion is performed, and the calculation formula is as follows: ; in, Weights for physical information; The prediction results for the Transformer model; For fusion prediction results; By comparing the deviations between the fused predicted values ​​and the measured values, dynamic adjustments are made. To balance physical constraints with data-driven prediction accuracy.

[0022] The surface subsidence prediction model employs a random forest model with formation loss data, combined with formation loss rates identified by YOLOv8, specifically: Multi-source heterogeneous data were collected, including earth pressure, attitude deviation, formation parameters (measured / predicted), other construction parameters, and formation loss rate (acquired through image recognition). A depth camera was deployed to collect images of excavated soil from the tunnel conveyor belt. The YOLOv8 instance segmentation algorithm was used to identify the excavated soil images and calculate the actual volume of the excavated soil. The formation loss rate was then calculated based on the actual volume of the excavated soil, using the following formula: ; in, Formation loss rate; This represents the actual volume of the construction waste. To design the excavation volume; All input parameters are associated with spatiotemporal identifiers such as construction ring number and borehole number to form a standardized feature vector. ;in, To collect earth pressure; This is for attitude deviation; For tunnel burial depth; This refers to the water level depth. As a spatiotemporal identifier; In the Random Forest (RF) model, a formation loss rate loss is introduced, and the total loss function is expressed as: ; in, Total loss; This is the mean square error loss; The final settlement prediction value is obtained by calculation using the following formula: ; in, This represents the final predicted settlement value. These are the predicted values ​​from the RF model. S3. Define a multi-objective collaborative optimization engine, with loss functions including propulsion velocity loss function, earth pressure loss function, attitude loss function, and settlement loss function; specifically: Propulsion speed loss function Represented as: ; in, For control parameter vectors; The tunnel boring machine's advance speed; The target speed for tunnel boring machine (TBM) advancement; For coefficients; Earth pressure loss function Represented as: ; in, This is the starting time of the integration; The time length of the integration; In order to be in Under its influence, at all times The actual earth pressure value; For a moment The corresponding target earth pressure value; For coefficients; In order to be in The average value of earth pressure over the time interval under the action of the earth; Shield attitude loss function Represented as: ; in, To correspond to the positional deviations of the tunnel boring machine (TBM) in the x-direction (horizontal transverse) and y-direction (horizontal longitudinal) and the angle of the TBM axis. This is used to describe the deviation of the shield tunneling attitude in different dimensions; for The weight coefficients of the dimensions are used to adjust the importance of pose deviations in different dimensions in the loss function. The larger the weight, the more significant the impact of pose deviations in that dimension on the loss function. In order to be in Under the action, the tunnel boring machine is The dimensional attitude deviation value; For the shield tunneling k The maximum allowable pose deviation value for a given dimension; For coefficients; The gradient of the shield tunnel attitude deviation; Surface subsidence loss function Represented as: ; in, The number of surface subsidence monitoring points; For coefficients; In order to be in u Under the action, the first Surface subsidence values ​​at each monitoring point; For coefficients; The safe threshold for surface subsidence; Equipment safety constraints include torque safety constraints, cylinder pressure difference constraints, grouting volume lower limit constraints, and axis continuity constraints; specifically: Torque safety constraints Represented as: ; in, In order to be in u The actual torque value of the tunnel boring machine under action; This represents the maximum allowable torque value for the tunnel boring machine. Hydraulic cylinder pressure difference constraint Represented as: ; in, This refers to the pressure value of the hydraulic cylinder on the left side of the tunnel boring machine. This refers to the pressure value of the hydraulic cylinder on the right side of the tunnel boring machine. This represents the average pressure of the left and right hydraulic cylinders. This represents the maximum permissible value of the relative proportion of the pressure difference between the left and right hydraulic cylinders. Grouting volume lower limit constraint Represented as: ; in, This is the minimum allowable value for the grouting volume; In order to be in u The actual grouting volume under the action; Axis continuity constraint Represented as: ; in, The coordinates represent the length of the axis along the tunnel boring machine's advance direction; This represents the maximum permissible value for the curvature of the axis. The Pareto front candidate solution set is obtained by solving a set of solutions using a genetic algorithm (GA), and then optimized control parameter vectors are generated by combining fuzzy preference and engineering constraint filtering. u ; S4. Optimize the control parameters from S3. u The data is transmitted to the tunnel boring machine's execution system to control the operating status of equipment such as the cutterhead, jacks, and grouting equipment. Data such as actual soil pressure, attitude deviation, and surface settlement during construction are collected by devices such as depth cameras and sensors, and fed back to the database in the S1 standard format. The measured data is input into the prediction module S2 to update the model parameters, and simultaneously fed back to the optimization engine S3 to resolve the optimal control strategy, thus realizing closed-loop adaptive control throughout the entire process.

[0023] Therefore, the present invention adopts the above-mentioned intelligent control method for the entire shield tunneling process based on multi-task model coupling. This method realizes automatic acquisition, cross-stage integration and real-time sharing of multi-source heterogeneous data; improves prediction accuracy and generalization ability; solves problems such as poor process coordination and insufficient control accuracy; avoids settlement risk; and achieves precise and intelligent management and control.

[0024] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A shield tunneling full-process intelligent control method based on multi-task model coupling, characterized in that, Includes the following steps: S1. Store data in a standardized data format, relying on the B / S dynamic system architecture and TCP / IP network configuration to achieve automatic data collection and standardized storage; S2. Construct a multi-stage coupled prediction model that includes a front chamber earth pressure prediction model, a shield attitude prediction model, and a surface settlement prediction model. The fore-cabin earth pressure prediction model adopts a hybrid model architecture of convolutional neural network and long short-term memory network to output the fore-cabin earth pressure prediction results. The shield tunnel attitude prediction model adopts an improved Transformer architecture, embedding a finite element proxy model to form a physical information fusion model, and outputs attitude prediction results. The surface subsidence prediction model uses a random forest model with stratum loss data, combined with the stratum loss rate identified by YOLOv8, to output the final subsidence prediction value. S3. Define a multi-objective collaborative optimization engine, construct a system containing loss functions and equipment safety constraints; obtain a Pareto front candidate solution set through a genetic algorithm solver, and output the optimized control parameters. S4. The optimized control parameters from S3 are transmitted to the tunnel boring machine execution system to drive the equipment to adjust the construction status; and the real-time collected data is fed back to the multi-stage coupled prediction model and the multi-objective collaborative optimization engine to form a closed-loop control.

2. The intelligent control method for the entire shield tunneling process based on multi-task model coupling according to claim 1, characterized in that, The standardized data format described in S1 specifically stores the data as follows: The system uses "Project Name" as the primary key to link multi-source heterogeneous data, including project overview, construction parameters, tunnel boring machine attitude, surface settlement, and stratum parameters. Each data table is associated through the primary key "Project Name" and subkeys, including ring number and date. A hybrid storage mode of MySQL and Excel is adopted, with MySQL serving as the backend database to support efficient data management and real-time access, and Excel used for intelligent auxiliary decision-making and suggestion interaction.

3. The intelligent control method for the entire shield tunneling process based on multi-task model coupling according to claim 2, characterized in that, S2 employs a hybrid model architecture combining convolutional neural networks and long short-term memory networks to predict earth pressure, comprising the following sub-steps: S201. Organize and preprocess the multi-source heterogeneous data collected during the tunnel boring machine construction process; S202. Based on the physical characteristics of tunnel construction parameters, the engineering features are divided into four categories: deviation, velocity, grouting pressure, and zoned pressure. A grouped convolution strategy is used to process the four categories of features independently. S203. Utilize the long short-term memory characteristics of the long short-term memory network to dynamically capture long and short-term dependencies in time-series data; S204. By integrating the spatial features extracted by the convolutional neural network with the temporal features learned by the long short-term memory network through a fully connected layer, the prediction result of the front chamber earth pressure is finally output.

4. The intelligent control method for the entire shield tunneling process based on multi-task model coupling according to claim 3, characterized in that, The improved Transformer architecture for the shield tunneling attitude prediction model in S2 includes: Let the input data dimension be m×1, the first... i The input parameters for each sample are a i ={ a 1 ,a 2 ,...,a m }, the position vector p= { p 1 , p 2 ,...,p m Add the input parameter to the input parameter to obtain the new input parameter. b i ={ b 1 ,b 2 ,...,b m The calculation formula is: ； The input parameters are passed to an encoder module containing M encoder units. Each unit consists of a self-attention layer and a feedforward neural network layer. Residual connections and layer normalization are performed after each layer. The formula for calculating the self-attention layer is: ； ； ； in, For query matrix; The key matrix; It is a value matrix; The input vector; For query vector; The key vector; It is a value vector; The result of the self-attention mechanism is represented as follows: ； Where softmax is the normalization function; The encoded data is fed into a decoder module containing N decoder units. Each decoder unit sequentially includes a self-attention layer, an encoder-decoder attention layer, and a feedforward neural network layer. The encoder-decoder attention layer reuses the self-attention layer. After each layer, residual joins and layer normalization are performed. After the data is processed by the decoder, initial predicted values ​​are generated through a feedforward layer and residual-normalization operations. The error between the initial predicted values ​​and the measured values ​​is calculated, and the backpropagation algorithm is used for iterative updates. , , and feedforward layer weights.

5. The intelligent control method for the entire shield tunneling process based on multi-task model coupling according to claim 4, characterized in that, In S2, the physical information fusion model of the shield tunnel attitude prediction model is constructed in the following way: An embedded finite element numerical simulation proxy model is used, which consists of an input layer, a hidden layer, and an output layer. The input parameters are processed by the finite element proxy model to obtain the physical prediction results. ; The weighted fusion is calculated using the following formula: ； in, Weights for physical information; The prediction results for the Transformer model; For fusion prediction results; By comparison The deviation from the measured value is dynamically adjusted. .

6. The intelligent control method for the entire shield tunneling process based on multi-task model coupling according to claim 5, characterized in that, In S2, the YOLOv8 identification of the formation loss rate in the surface subsidence prediction model includes: The formula for calculating the formation loss rate is as follows: ； in, Formation loss rate; This represents the actual volume of the construction waste. To design the excavation volume.

7. The intelligent control method for the entire shield tunneling process based on multi-task model coupling according to claim 6, characterized in that, S2 also includes: To collect earth pressure; This is for attitude deviation; For tunnel burial depth; This refers to the water level depth. For spatiotemporal identification; collect earth pressure and tunnel depth, water level depth, and spatiotemporal identification information and correlate them to form a feature vector. ; In the random forest model, a formation loss rate loss is introduced, and the total loss function is expressed as: ； in, Total loss; This is the mean square error loss; The final settlement prediction value is obtained by calculation using the following formula: ； in, This represents the final predicted settlement value. These are the predicted values ​​from the random forest model.

8. The intelligent control method for the entire shield tunneling process based on multi-task model coupling according to claim 7, characterized in that, The loss functions in S3 include the propulsion velocity loss function, earth pressure loss function, attitude loss function, and settlement loss function; specifically: Propulsion speed loss function Represented as: ； in, For control parameter vectors; The tunnel boring machine's advance speed; The target speed for tunnel boring machine (TBM) advancement; For coefficients; Earth pressure loss function Represented as: ； in, This is the starting time of the integration; The time length of the integration; In order to be in Under its influence, at all times The actual earth pressure value; For a moment The corresponding target earth pressure value; For coefficients; In order to be in The average value of earth pressure over the time interval under the action of the earth; Shield attitude loss function Represented as: ； in, To correspond to the positional deviations of the shield in the x-direction (horizontal transverse direction) and y-direction, as well as the angle of the shield axis. ; for Dimension weight coefficients; In order to be in Under the action, the tunnel boring machine is The dimensional attitude deviation value; For the shield tunneling k The maximum allowable pose deviation value for a given dimension; For coefficients; The gradient of the shield tunnel attitude deviation; Surface subsidence loss function Represented as: ； in, The number of surface subsidence monitoring points; For coefficients; In order to be in u Under the action, the first Surface subsidence values ​​at each monitoring point; For coefficients; This is the safe threshold for land subsidence.

9. The intelligent control method for the entire shield tunneling process based on multi-task model coupling according to claim 8, characterized in that, Equipment safety constraints in S3 include torque safety constraints, cylinder pressure difference constraints, grouting volume lower limit constraints, and axis continuity constraints; specifically: Torque safety constraints Represented as: ； in, In order to be in u The actual torque value of the tunnel boring machine under action; This is the maximum allowable torque value for the tunnel boring machine; Hydraulic cylinder pressure difference constraint Represented as: ； in, This refers to the pressure value of the hydraulic cylinder on the left side of the tunnel boring machine. This refers to the pressure value of the hydraulic cylinder on the right side of the tunnel boring machine. This represents the average pressure of the left and right hydraulic cylinders. This represents the maximum permissible value of the relative proportion of the pressure difference between the left and right hydraulic cylinders. Grouting volume lower limit constraint Represented as: ； in, This is the minimum allowable value for the grouting volume; In order to be in u The actual grouting volume under the action; Axis continuity constraint Represented as: ； in, The coordinates represent the length of the axis along the tunnel boring machine's advance direction; This represents the maximum permissible value for the curvature of the axis.

10. A shield tunneling full-process intelligent control system based on multi-task model coupling, applied to the shield tunneling full-process intelligent control method based on multi-task model coupling as described in any one of claims 1-9, characterized in that, include: The data acquisition and storage module is used to store data in a standardized data format. It relies on the B / S dynamic system architecture and TCP / IP network configuration to achieve automatic data acquisition and standardized storage. A multi-stage coupled prediction module is connected to the data acquisition and storage module and is used to construct a multi-stage coupled prediction model that includes a front chamber earth pressure prediction model, a shield attitude prediction model and a surface settlement prediction model. The front chamber earth pressure prediction model is based on two-dimensional coupled Markov chain to simulate the stratum distribution, combined with Bayesian machine learning to complete the soil parameters, and calculates the earth chamber pressure setpoint through soil mechanics formula. The shield tunnel attitude prediction model adopts an improved Transformer architecture, embedding a finite element proxy model to form a physical information fusion model, and outputs attitude prediction results. The surface subsidence prediction model uses a random forest model with stratum loss data, combined with the stratum loss rate identified by YOLOv8, to output the final subsidence prediction value. A multi-objective optimization decision module, connected to the multi-stage coupled prediction module, is used to define a multi-objective collaborative optimization engine, construct a loss function and equipment safety constraints; obtain a Pareto front candidate solution set through a genetic algorithm solver, and output the optimized control parameters. The control execution and feedback module is connected to the multi-objective optimization decision module. It is used to transmit the optimized control parameters to the tunnel boring machine execution system to drive the equipment to adjust the construction state; and to feed back the real-time collected data to the multi-stage coupled prediction model and the multi-objective collaborative optimization engine to form a closed-loop control.

Images