A TBM jamming risk prediction method based on a Transformer
By using a Transformer-based multi-task learning model and expert system, combined with expert experience and adaptive adjustment, synchronous prediction of TBM card machine risk was achieved, solving the problem of insufficient model generalization ability in existing technologies and improving the accuracy and adaptability of early warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGAN UNIV
- Filing Date
- 2026-05-18
- Publication Date
- 2026-06-12
AI Technical Summary
Existing TBM jamming risk identification methods suffer from problems such as strong subjectivity, insufficient single-parameter threshold alarms, insufficient extraction capabilities of traditional machine learning algorithms, limited model generalization ability, and poor engineering adaptability, making it difficult to achieve synchronous and accurate prediction of jamming risks of cutter head and shield.
A Transformer-based multi-task learning model is adopted, combined with an expert system for data annotation and feature extraction, to construct a shared feature extraction layer for the risks of the cutter head and shield. Feature learning is performed through a multi-head attention module and a feedforward neural network module, and adaptive adjustments are made to adapt to different tunnel working conditions. A continuous sliding window is used for comprehensive judgment.
It enables simultaneous and accurate prediction of risks associated with both the cutter head and the shield, improving the accuracy, real-time performance, and reliability of early warnings, and enhancing the model's adaptability and prediction accuracy in complex geological environments.
Smart Images
Figure CN122196704A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent early warning technology for tunnel excavation construction, and in particular to a TBM (Tunnel Boring Machine) risk prediction method based on Transformer. Background Technology
[0002] Tunnel boring machines (TBMs), with their advantages of high mechanization, high construction efficiency, and high-quality tunnel completion, have become core equipment for the construction of long tunnels, hydraulic tunnels, and traffic tunnels. However, under complex geological conditions, the TBM tunneling process is susceptible to adverse factors such as fault fracture zones, weak interlayers, high ground stress, groundwater, and uneven surrounding rock, which can easily induce machine jamming accidents such as cutterhead jamming and shield jamming. Machine jamming accidents not only cause equipment downtime, project delays, and huge economic losses, but may also trigger secondary disasters such as surrounding rock instability and collapse, seriously threatening the lives of construction personnel and the progress of project construction.
[0003] Existing methods for identifying TBM jamming risks suffer from the following drawbacks: First, they rely on the experience of construction personnel, resulting in high subjectivity, low standardization, and difficulty in achieving continuous and automated monitoring. Second, they employ single-parameter threshold alarms, relying solely on a few indicators such as thrust and cutterhead torque, failing to fully reflect the multi-parameter coupling characteristics and temporal evolution of jamming risks. Third, traditional machine learning algorithms lack the ability to extract long-term dependencies and complex nonlinear coupling relationships, making it difficult to meet the prediction accuracy requirements in complex geological environments. Fourth, most existing methods model cutterhead jamming and shield jamming risks separately, failing to fully utilize the shared features and task coupling relationships between the two types of risks, thus limiting the model's generalization ability. Fifth, the geological conditions, construction parameters, and equipment status of different tunnels vary significantly, making single models less adaptable to different projects and hindering cross-project application. Therefore, there is an urgent need to develop an intelligent early warning method for TBM jamming risks that integrates expert experience, possesses strong temporal modeling capabilities, supports simultaneous prediction of multiple risks, and has strong adaptability to various working conditions. Summary of the Invention
[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a Transformer-based method for predicting TBM jamming risks. This method enables simultaneous and accurate prediction of risks related to jamming of the cutter head and the shield, solving the problems of weak temporal feature extraction capabilities, lack of joint prediction of multiple risks, and poor engineering adaptability in existing technologies. It improves the accuracy, real-time performance, and reliability of TBM jamming risk warning.
[0005] To achieve the above objectives, the present invention provides the following solution:
[0006] A Transformer-based method for predicting the risk of TBM card machines includes the following steps:
[0007] S1: Collect tunneling parameter data during the TBM construction process, and preprocess the tunneling parameter data to obtain time-series sample data;
[0008] S2: Based on the expert system, the preprocessed time series sample data is classified and labeled according to preset rules and thresholds to classify the risk of TBM card machine;
[0009] S3: Construct a multi-task learning model based on Transformer, and use labeled time-series sample data to supervise the training of the multi-task learning model to obtain the TBM card machine risk prediction model.
[0010] S4: The real-time collected TBM tunneling parameters are preprocessed and input into the TBM jamming risk prediction model. The model outputs the jamming risk level of the cutterhead and the jamming risk level of the shield at the corresponding time. Based on the prediction results of the continuous sliding window, a comprehensive judgment is made to trigger the real-time warning of the corresponding level.
[0011] Furthermore, the tunneling parameter data includes:
[0012] The data includes cutterhead torque, penetration depth, cutterhead rotation speed, total propulsion force, and formation response parameters. Simultaneously, the mean reflection intensity, mean regional reflection intensity, and geological risk level characteristics from the TSP forecast results are collected to characterize changes in the surrounding rock structure and potential risks.
[0013] Furthermore, the formation response parameters include surrounding rock type, formation water-bearing capacity, rock mass integrity parameters, and local geological anomalies.
[0014] Furthermore, the preprocessing specifically includes:
[0015] After data cleaning, outlier removal, missing value completion, and normalization of the tunneling parameter data, time-series sample data is constructed based on the processed tunneling parameter data and a preset sliding window.
[0016] Furthermore, the classification and labeling of the expert system are specifically as follows:
[0017] Based on historical construction data, engineering geological conditions, on-site construction experience, and preset risk threshold rules, judgment logics for cutterhead jamming risk and shield jamming risk were established respectively. TBM jamming risk was divided into cutterhead jamming risk and shield jamming risk, and both types of risk were divided into three levels: low risk, medium risk, and high risk. Among them, cutterhead jamming risk was judged based on cutterhead torque, penetration depth, propulsion speed and their coupling relationship, while shield jamming risk was judged based on total propulsion force, formation response parameters and their changing trends.
[0018] Furthermore, the TBM card machine risk prediction model includes a shared feature extraction layer, a card disc risk prediction branch, and a card shield risk prediction branch:
[0019] The shared feature extraction layer uses a Transformer encoding structure to extract features. The Transformer encoding structure includes an input mapping layer, a position embedding layer, and a Transformer encoding unit. Each Transformer encoding unit includes a multi-head attention module, a feedforward neural network module, a residual connection module, and a layer normalization module.
[0020] The extracted features are input into the risk task prediction branches for the card blade disk and the card shield.
[0021] Both the tool disk risk prediction branch and the shield risk prediction branch contain a fully connected layer, a ReLU activation function, a Dropout layer, and a Softmax classification layer, which respectively output the low-risk, medium-risk, and high-risk probability distributions corresponding to the tool disk risk and the shield risk.
[0022] Furthermore, step S3 also includes constructing an adaptive risk prediction framework for different tunnel conditions: based on the geological conditions, surrounding rock characteristics, construction parameter distribution characteristics, and historical risk sample characteristics of different tunnels, at least one of the model parameters, risk judgment threshold, and task loss weights is adaptively adjusted. The adjustment methods include independent training for each tunnel, dynamic configuration of task weights, incremental learning, and transfer optimization.
[0023] Furthermore, the real-time early warning specifically refers to:
[0024] Consistency analysis is performed on the risk prediction results of multiple consecutive sliding windows. When a preset number of consecutive sliding windows output the same risk level, a warning signal of the corresponding level is triggered. Among them, low risk outputs a normal construction prompt, medium risk outputs enhanced monitoring and warning, and high risk outputs a shutdown and handling alarm.
[0025] The present invention discloses the following technical effects:
[0026] This invention introduces an expert system to perform rule-based risk labeling on construction data, combining engineering experience with data-driven methods, and taking into account the rationality, interpretability and engineering applicability of model training;
[0027] This invention divides the risk of TBM jamming into blade jamming risk and shield jamming risk, and performs synchronous prediction based on a multi-task learning model, which is beneficial to make full use of the shared information between the two types of risks and improve the overall predictive ability of the model.
[0028] This invention uses Transformer to model TBM multivariate time series data, which can effectively extract long-term time-dependent features and multi-parameter coupling features, and improve the accuracy of identifying machine jamming risks in complex geological environments;
[0029] This invention can adaptively adjust model parameters, risk thresholds, or task weights for different tunnel working conditions, thereby improving the model's adaptability and generalization ability under different geological conditions and construction environments.
[0030] This invention uses multiple consecutive sliding time windows for comprehensive judgment, which can reduce the impact of instantaneous abnormal fluctuations on the early warning results, improve the stability of early warning and reduce the false alarm rate, making it more suitable for real-time risk early warning at construction sites. Attached Figure Description
[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0032] Figure 1 This is a schematic diagram of the overall process of the Transformer-based TBM card machine risk prediction method of the present invention;
[0033] Figure 2 This is a schematic diagram of the multi-task learning model structure based on Transformer of the present invention;
[0034] Figure 3 This is a schematic diagram of the risk level labeling process based on an expert system in this invention;
[0035] Figure 4 This is a schematic diagram of the real-time risk prediction and continuous sliding window early warning determination process in this invention. Detailed Implementation
[0036] 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, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0038] like Figures 1-4As shown, this embodiment of the invention provides a Transformer-based TBM card machine risk prediction method, including:
[0039] Tunneling parameter data acquisition and preprocessing, including:
[0040] During TBM construction, tunneling parameter data is acquired through equipment control systems, sensor monitoring systems, and construction information acquisition systems. This tunneling parameter data includes, but is not limited to, thrust, cutterhead torque, penetration depth, advance speed, ground response parameters, and the corresponding time-series variation data of these parameters.
[0041] Specifically, formation response parameters may include surrounding rock type, formation water-bearing capacity, parameters related to rock mass integrity, local geological anomalies, or monitoring parameters related to surrounding rock response. These parameters can be continuously collected at set time intervals to form a multivariate time-series dataset.
[0042] Simultaneously, the mean reflection intensity, regional mean reflection intensity, and geological risk level characteristics of the TSP (Tunnel Seismic Prediction) forecast results were collected.
[0043] The collected raw data is preprocessed to improve the stability of subsequent model training and real-time prediction. The preprocessing includes:
[0044] (1) Data cleaning: Delete invalid data caused by sensor failure, communication abnormality or human error;
[0045] (2) Outlier removal or correction: Identify abrupt changes that exceed the normal operating range of the equipment and process them using threshold correction, neighborhood smoothing or interpolation.
[0046] (3) Missing value completion: missing data are completed using forward filling, backward filling, mean compensation or interpolation methods;
[0047] (4) Normalization: Standardize or normalize parameters with different dimensions to eliminate the impact of differences in the dimensions of different parameters on model training;
[0048] (5) Construction of time series samples: Slice the continuous time series data according to the preset time window length and sliding step size to construct the time series sample data input to the model. The sliding window length should be able to cover the typical time range corresponding to the formation and evolution of TBM card machine risk. The sliding step size can be set according to the real-time warning requirements to ensure the integrity of time series features while taking into account computational efficiency and warning timeliness.
[0049] Risk level labeling using expert systems includes:
[0050] Risk levels were assigned to time-series sample data using a pre-built expert system. This expert system, based on historical construction data, engineering geological conditions, construction experience, and pre-defined risk assessment rules and thresholds, categorized the risks of cutterhead jamming and shield jamming. TBM jamming risks were classified into two types: cutterhead jamming and shield jamming; each type was further divided into three levels: low risk, medium risk, and high risk.
[0051] Specifically, the risk of cutterhead jamming can be determined by analyzing cutterhead torque, penetration depth, and their interrelationships; the risk of shield jamming can be determined by analyzing thrust, formation response parameters, and related trends. When geological conditions deteriorate, key construction parameters continuously deviate from normal ranges, and gradually approach the characteristics of historical jamming events, it can be classified as high risk; when parameters show some anomalies but have not yet reached severe instability or strong jamming signs, it can be classified as medium risk; when parameters are generally stable, construction status is normal, and there is no obvious risk evolution trend, it can be classified as low risk. Through these methods, cutterhead jamming risk labels and shield jamming risk labels are formed respectively, thereby constructing a supervised sample set for multi-task learning.
[0052] Building and training a multi-task learning model using Transformer includes:
[0053] After completing sample labeling, a Transformer-based multi-task learning model is constructed. This model includes a shared feature extraction layer and task output layers corresponding to the risks of the blade disk and shield respectively. The shared feature extraction layer employs a Transformer encoding structure to extract long-term temporal dependencies, dynamically changing features, and multi-parameter coupling features from the temporal samples. The Transformer encoding structure sequentially includes an input mapping layer, a position embedding layer, and at least two stacked Transformer encoding units.
[0054] The temporal sample data is first converted into a feature vector of uniform dimension through the input mapping layer, and then the positional encoding is superimposed through the position embedding layer to preserve the temporal order information. Subsequently, it is input into the Transformer encoding unit for deep feature extraction. Each Transformer encoding unit includes at least a multi-head attention module, a feedforward neural network module, a residual connection module, and a normalization module.
[0055] The multi-head attention module is used to learn the correlation between different time positions to enhance the model's ability to represent long-distance dependencies and local mutation information; the feedforward neural network module is used to improve the ability to express nonlinear features; the residual connection module and the normalization module are used to alleviate the gradient vanishing problem in the training process of deep networks and improve the model training stability and convergence efficiency.
[0056] After the shared features are extracted, the extracted features are input into the output layers for the card-operated disk risk task and the card-operated shield risk task, respectively. Each task output layer can employ at least one fully connected network and a Softmax classification function to output the probability distributions of three risk levels: low, medium, and high. The prediction results for the card-operated disk risk level and the card-operated shield risk level are obtained based on the principle of maximizing probability.
[0057] During the model training phase, time-series samples labeled by an expert system were used for supervised training of the multi-task learning model. During training, the risk classification losses for the cutter head and shield were calculated separately and then weighted and summed according to preset weights to form the model's total loss function. Through iterative optimization of the total loss function, the trained TBM cutter head risk prediction model was obtained.
[0058] An adaptive risk prediction framework is used to output real-time risk prediction and early warning information for different tunnel working conditions, including:
[0059] Considering the differences in surrounding rock type, ground stress level, groundwater conditions, construction parameter distribution and historical risk sample characteristics among different tunnels, this embodiment further constructs corresponding risk prediction frameworks for different tunnel conditions.
[0060] Based on the geological conditions, surrounding rock characteristics, statistical characteristics of construction parameters, or distribution characteristics of historical risk samples of different tunnels, at least one of the following can be adaptively adjusted: model parameters can be trained separately for different tunnels; the loss weights of different tasks can be adjusted on the basis of a unified network structure; different risk judgment thresholds can be set according to differences in working conditions; or the model can be incrementally updated or optimized by combining historical samples. By establishing an adaptive risk prediction framework under different working conditions, the applicability of the model to complex engineering scenarios can be improved, and the accuracy and generalization ability of risk identification among different tunnels can be enhanced.
[0061] After the model training is completed and deployed to the construction site, the TBM tunneling parameters are collected in real time and preprocessed to generate real-time sliding window sample inputs to the model. The model synchronously outputs the current risk level of cutterhead jamming and shield jamming.
[0062] The risk assessment of tool holders employs a combination of multi-indicator comprehensive judgment and risk level correction. The overall assessment model can be expressed as follows:
[0063]
[0064]
[0065] in, The revised risk level for the tool holder; This is the adjustment amount for the risk level; This indicates the number of core indicators that meet the criteria for identifying a higher risk level.
[0066]
[0067] In the formula, Risk level of tool disc jamming; , , These represent the average values of the cutterhead torque, penetration depth, and cutterhead rotation speed within a certain time window, respectively. , and This is a risk level mapping function built based on engineering experience.
[0068] The card shield risk assessment adopts a method that combines multi-indicator comprehensive judgment with risk level correction. The overall assessment model can be expressed as:
[0069]
[0070]
[0071] In the formula, The revised card shield risk level; Risk level of card shield; This is the adjustment amount for the risk level.
[0072]
[0073] In the formula, T These represent the average values of total thrust, cutterhead torque, and penetration within a certain time window, respectively. , and This is a risk level mapping function built based on engineering experience.
[0074] For low-risk output alerts, the system indicates that the construction status is basically normal; for medium-risk output alerts, the system reminds on-site personnel to strengthen monitoring and parameter adjustments; for high-risk output alarms, the system reminds personnel to take timely measures such as stopping the machine for inspection, adjusting the attitude, optimizing parameters, or handling geological issues.
[0075] To improve the stability of early warnings and reduce false alarm rates, this embodiment further employs a comprehensive judgment based on the prediction results of multiple consecutive sliding time windows. When multiple consecutive sliding time windows output the same risk level, a corresponding level of early warning signal is triggered. When three consecutive sliding time windows output a high-risk level, the system triggers a high-risk early warning; if only a single window shows an anomaly and subsequent windows do not continuously show the same level, a formal early warning is not triggered. This method can effectively suppress false alarms caused by short-term fluctuations, local noise, or accidental anomalies. Through the above steps, this embodiment achieves simultaneous prediction and real-time intelligent early warning of the risks of cutterhead jamming and shield jamming during TBM construction.
[0076] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0077] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A Transformer-based method for predicting the risk of a TBM (Transformer Machine) card reader, characterized in that, Includes the following steps: S1: Collect tunneling parameter data during the TBM construction process, and preprocess the tunneling parameter data to obtain time-series sample data; S2: Based on the expert system, the preprocessed time series sample data is classified and labeled according to preset rules and thresholds to identify the risk of TBM card machines; S3: Construct a multi-task learning model based on Transformer, and use labeled time-series sample data to supervise the training of the multi-task learning model to obtain the TBM card machine risk prediction model; S4: The real-time collected TBM tunneling parameters are preprocessed and input into the TBM jamming risk prediction model. The model outputs the jamming risk level of the cutterhead and the jamming risk level of the shield at the corresponding time. Based on the prediction results of the continuous sliding window, a comprehensive judgment is made to trigger the corresponding level of real-time early warning.
2. The TBM card machine risk prediction method based on Transformer according to claim 1, characterized in that, The tunneling parameter data includes: The data includes cutterhead torque, penetration depth, cutterhead rotation speed, total propulsion force, and formation response parameters; simultaneously, the mean reflection intensity, mean regional reflection intensity, and geological risk level characteristics from the TSP forecast results are collected.
3. The TBM card machine risk prediction method based on Transformer according to claim 2, characterized in that, The formation response parameters include surrounding rock type, formation water-bearing capacity, rock mass integrity parameters, and local geological anomalies.
4. The TBM card machine risk prediction method based on Transformer according to claim 1, characterized in that, The preprocessing specifically includes: After data cleaning, outlier removal, missing value completion, and normalization of the tunneling parameter data, time-series sample data is constructed based on the processed tunneling parameter data and a preset sliding window.
5. The TBM card machine risk prediction method based on Transformer according to claim 1, characterized in that, The classification and labeling of the expert system are specifically as follows: Based on historical construction data, engineering geological conditions, on-site construction experience, and preset risk threshold rules, judgment logics for cutterhead jamming risk and shield jamming risk were established respectively. TBM jamming risk was divided into cutterhead jamming risk and shield jamming risk, and both types of risk were divided into three levels: low risk, medium risk, and high risk. Among them, cutterhead jamming risk was judged based on cutterhead torque, penetration depth, propulsion speed and their coupling relationship, while shield jamming risk was judged based on total propulsion force, formation response parameters and their changing trends.
6. The TBM card machine risk prediction method based on Transformer according to claim 1, characterized in that, The TBM jamming risk prediction model includes a shared feature extraction layer, a jamming disc risk prediction branch, and a jamming shield risk prediction branch. The shared feature extraction layer uses a Transformer encoding structure to extract features. The Transformer encoding structure includes an input mapping layer, a location embedding layer, and a Transformer encoding unit. Each Transformer encoding unit includes a multi-head attention module, a feedforward neural network module, a residual connection module, and a layer normalization module. Both the tool disk risk prediction branch and the shield risk prediction branch contain a fully connected layer, a ReLU activation function, a Dropout layer, and a Softmax classification layer, which respectively output the low-risk, medium-risk, and high-risk probability distributions corresponding to the tool disk risk and the shield risk.
7. The TBM card machine risk prediction method based on Transformer according to claim 1, characterized in that, Step S3 also includes constructing an adaptive risk prediction framework for different tunnel conditions: based on the geological conditions, surrounding rock characteristics, construction parameter distribution characteristics and historical risk sample characteristics of different tunnels, at least one of the model parameters, risk judgment thresholds and task loss weights is adaptively adjusted. The adjustment methods include independent training for each tunnel, dynamic configuration of task weights, incremental learning and transfer optimization.
8. The TBM card machine risk prediction method based on Transformer according to claim 1, characterized in that, The real-time early warning specifically refers to: Consistency analysis is performed on the risk prediction results of multiple consecutive sliding windows. When a preset number of consecutive sliding windows output the same risk level, a warning signal of the corresponding level is triggered. Low-risk outputs are indicated by normal construction prompts, medium-risk outputs by enhanced monitoring and early warning, and high-risk outputs by shutdown and emergency response alarms.