Pipe jacking over-excavation prediction and slurry concentration adjustment method based on LSTM and supervised learning optimization
By combining LSTM with supervised learning optimization, high-precision prediction of multi-source data and mud concentration adjustment in pipe jacking construction are achieved, solving the problems of ground disturbance and surface subsidence caused by over-excavation in pipe jacking construction, and improving the stability and safety of construction.
Patent Information
- Application Number
- CN202511683245.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-17
- Publication Date
- 2026-02-17
AI Technical Summary
Over-excavation in existing pipe jacking construction causes ground disturbance and surface subsidence. Traditional prediction and control methods lack adaptability and intelligent decision-making closed loops, making it difficult to integrate multi-source sensor data and optimize model parameters.
A method based on LSTM and supervised learning optimization is adopted, which combines multi-source data to predict over-excavation of pipe jacking. Through error correction and parameter optimization, a built-in reinforcement learning agent is used to adjust the mud concentration, thereby achieving high-precision prediction and control.
It improves the prediction accuracy and control effect of pipe jacking construction, reduces over-excavation rate, reduces the risk of surface subsidence, and enhances construction stability and efficiency.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to a pipe jacking overbreak prediction and mud concentration adjustment method, in particular to a pipe jacking overbreak prediction and mud concentration adjustment method based on LSTM and supervised learning optimization. BACKGROUND
[0002] During pipe jacking construction, overbreak may cause stratum disturbance, ground settlement and voids around the pipe, which poses a serious threat to construction safety and engineering quality. Traditional prediction and control methods mainly rely on empirical formulas or manual adjustment of mud concentration, lacking self-adaptive ability to complex nonlinear geological conditions and real-time construction conditions. With the development of artificial intelligence technology, time series prediction models based on deep learning have been applied to tunnel and pipe jacking engineering, but existing methods have the following shortcomings:
[0003] 1. The model is single and difficult to integrate multi-source sensor data (such as soil pressure, jacking force, grouting pressure, mud flow, concentration, etc.);
[0004] 2. Prediction and control are separated, lacking intelligent decision-making loop for mud concentration adjustment;
[0005] 3. Model parameter optimization relies on manual debugging and lacks generalization ability.
[0006] Therefore, there is an urgent need for a pipe jacking overbreak prediction and mud concentration adjustment method based on multi-source data driving, deep learning and intelligent optimization. SUMMARY
[0007] The application proposes a pipe jacking overbreak prediction and mud concentration adjustment method based on LSTM and supervised learning optimization, which realizes high-precision prediction of overbreak risk in pipe jacking construction and reduces overbreak rate and improves construction stability.
[0008] Technical scheme: The application comprises the following steps:
[0009] Data acquisition and preprocessing: Collect multi-source time series data and static geological data at the construction site, and preprocess the multi-source time series data;
[0010] Time series prediction: built-in LSTM, model the preprocessed multi-source time series data, and output preliminary overbreak volume or overbreak probability prediction results;
[0011] Error correction: built-in supervised learning model, model the residual error of time series prediction, and add the correction value to the preliminary prediction result to obtain the final prediction result;
[0012] Parameter optimization: use Bayesian optimization algorithm to automatically optimize the hyperparameters of LSTM or supervised learning model;
[0013] Decision control: The built-in reinforcement learning agent outputs mud concentration regulation strategies based on the system state.
[0014] The hidden state update expression of the LSTM is:
[0015]
[0016] in, This represents the cell state at the previous moment. , , These are the activation vectors for the input, forget, and output gates, respectively. This represents the candidate cell state.
[0017] The error correction specifically includes:
[0018] Generate preliminary predictions using LSTM Calculate the preliminary prediction results of LSTM. The residual between the actual value and the true value;
[0019] Using a supervised learning model with residuals as the objective. The output of the supervised learning model is used as a correction value, which is added to the initial prediction result to obtain the final corrected prediction result:
[0020]
[0021] in, This represents the input feature vector. These are the parameters for the supervised learning model.
[0022] The input feature vector It includes one or more of the following: hidden state of LSTM, geocoding, and statistical features.
[0023] When correcting the error, alternating training or cascaded training is used: first train the LSTM to fix its parameters, and then train the residual model; or joint training is used.
[0024] When optimizing the parameters, the desired improvement in EI is used as the acquisition function to balance exploring regions with high uncertainty and utilizing regions with low mean. Its expression is:
[0025]
[0026] in, The optimal target value observed so far. The CDF is a standard normal distribution.
[0027] When optimizing the parameters, ordered encoding is used for discrete or integer parameters, or a hybrid kernel is used in the surrogate model; the TPE with a tree structure is used to replace the Gaussian process in the high-dimensional space; and long-term indicators of the control end are incorporated into the optimization objective.
[0028] The decision control is modeled as a Markov decision process, with the state including short-term predictions from LSTM, current and historical sensor observations, and the most recent control action; the action is the mud concentration setting; and the reward is a combination of over-excavation risk, mud cost, and change cost.
[0029] The reward is a weighted composite reward:
[0030]
[0031] in, The over-excavation probability is predicted based on the current model. This represents the actual mud concentration. This is a nominal or optimal concentration reference value. The weights are adjustable.
[0032] A pipe jacking over-excavation prediction and mud concentration adjustment system based on LSTM and supervised learning optimization includes:
[0033] Data acquisition and preprocessing module: Acquires multi-source time-series data and static geological data from the construction site, and preprocesses the multi-source time-series data;
[0034] Time series prediction module: Built-in LSTM, which models the preprocessed multi-source time series data and outputs preliminary over-excavation amount or over-excavation probability prediction results;
[0035] Error correction module: It has a built-in supervised learning model to model the residuals of the time series prediction module and adds the correction value to the preliminary prediction result to obtain the final prediction result;
[0036] Parameter optimization module: Uses Bayesian optimization algorithm to automatically optimize the hyperparameters of LSTM or supervised learning models;
[0037] Decision control module: It has a built-in reinforcement learning agent that outputs mud concentration regulation strategies based on the system status;
[0038] System integration and deployment module: Integrates and deploys the above modules on edge computing nodes or the cloud, providing real-time inference, control decision-making and human-computer interaction functions.
[0039] Beneficial effects: This invention has the following advantages:
[0040] (1) Significantly improved prediction accuracy: This invention organically combines temporal deep learning, supervised residual correction, multi-source fusion, Bayesian optimization and reinforcement learning. The encoding of historical dynamic information by LSTM, plus the compensation of static or few-sample bias by supervised learning, can reduce prediction bias and improve recall / precision.
[0041] (2) The control effect is more optimized in the long term: Reinforcement learning aims to accumulate control performance, which can not only suppress the probability of immediate over-excavation, but also achieve a balance between mud consumption and long-term settlement risk;
[0042] (3) Improved engineering adaptability and automation: Multi-source fusion and automatic Bayesian hyperparameter search enable the model to adapt to different geological and construction conditions, reducing manual parameter tuning;
[0043] (4) System security and interpretability: Through weight visualization, policy constraints and rollback mechanisms, the system provides sufficient human supervision and policy transparency while ensuring automation, thereby reducing on-site risks.
[0044] Therefore, this invention can reduce the over-excavation rate, reduce the risk of surface subsidence, optimize mud consumption, and improve construction efficiency and safety, and has significant engineering application value and promotion prospects. Attached Figure Description
[0045] Figure 1 This is a flowchart of the present invention;
[0046] Figure 2 This is a schematic diagram of the LSTM structure of the present invention;
[0047] Figure 3 This is a Bayesian optimization example diagram of the present invention. Detailed Implementation
[0048] The invention will now be further described with reference to the accompanying drawings.
[0049] Example 1
[0050] like Figure 1 As shown, the method for predicting over-excavation and adjusting mud concentration based on LSTM and supervised learning optimization in this embodiment includes the following steps:
[0051] S1. Data Acquisition and Preprocessing: Collect multi-source time-series data and static geological data from the construction site, and preprocess the multi-source time-series data.
[0052] Multi-source time-series data were collected from the construction site, including jacking force, earth pressure, grouting pressure, mud concentration, mud discharge volume, surface settlement, and geological parameters. Outliers were removed through interpolation and filtering, and normalization was applied. , Let j be the value of the j-th feature at time t. The mean, The standard deviation is used to map variables of different dimensions to the same scale.
[0053] S2, Temporal Prediction: Built-in Long Short-Term Memory Network, such as Figure 2 As shown, the preprocessed multi-source time series data is modeled to output preliminary prediction results of over-excavation volume or over-excavation probability.
[0054] This embodiment chooses LSTM as the backbone because of its natural ability to model temporal dependencies: the forget gate can remove historical information that is no longer important, the input gate controls the writing of new information, and the output gate determines the contribution of the current memory to subsequent predictions. For pipe jacking over-excavation prediction, there is a multi-scale coupling effect in physics: "early grouting - short-term stratum response - long-term settlement". LSTM can encode these temporal causes and effects in the form of hidden states and cellular states, adapting to engineering processes with hysteresis effects.
[0055] In practical architecture, this invention recommends a multi-layer LSTM (1–3 layers), with each layer employing appropriate dropout and LayerNorm to prevent overfitting and training instability. The output layer can be set as a regression head (directly predicting overmining) or a probabilistic head (predicting future overmining probability). The loss function is chosen based on the task, selecting either MSE (regression) or cross-entropy (classification), supplemented by L2 regularization and early stopping strategies. During training, batches should be constructed in chronological order to avoid label leakage caused by disrupting the time sequence. The hidden state update of LSTM can be summarized as follows:
[0056]
[0057] in, This represents the cell state at the previous moment. , , These are the activation vectors for the input, forget, and output gates, respectively. This represents the candidate cell state.
[0058] Training techniques include: an appropriate learning rate, the Adam optimizer, batch normalization or layer normalization, and gradient pruning to avoid gradient explosion; for long sequences, segmented backpropagation can be used to reduce memory overhead. To improve model reliability, time-series cross-validation is recommended during the training phase instead of random splitting.
[0059] S3. Error Correction: The built-in supervised learning model models the residuals of the time series prediction module and adds the correction value to the preliminary prediction result to obtain the final prediction result.
[0060] While LSTM excels at temporal pattern recognition, it can introduce systematic biases when dealing with static geological features or extreme limited sample conditions. For example, the same instantaneous earth pressure can have different implications for over-excavation risk under different geological conditions. To address this, this invention introduces a supervised learning module to model and correct the residuals of the LSTM, thus combining the advantages of both models—LSTM excels at learning dynamic causality from sequences, while supervised learning (such as random forests and gradient boosting trees) is robust to high-dimensional heterogeneous static features and nonlinear interactions.
[0061] First, use LSTM to generate preliminary predictions. Calculate the preliminary prediction results of LSTM. The residual between the true value and the actual value; using the residual as the target, a supervised learning model is employed. Among them, the input feature vector It includes one or more of the following: LSTM hidden states, geocoding, and statistical features; the output of the supervised learning model is used as a correction value and added to the preliminary prediction result to obtain the final corrected prediction result:
[0062]
[0063] in, This represents the input feature vector. To supervise the learning of model parameters, this strategy separates the complex temporal modeling task from the well-interpretable static / semi-static feature fitting, avoiding overfitting or underfitting of a single model on high-dimensional heterogeneous data. In engineering practice, alternating or cascaded training can be used: first train the LSTM to fix its parameters, then train the residual model; or joint training (end-to-end fine-tuning) can be used to further improve performance. Tree-based methods are preferred for the residual model due to their high tolerance for missing and outliers, and the output of feature importance is helpful for engineering interpretation. During evaluation, the prediction error distribution should be monitored simultaneously to confirm whether residual correction introduces new biases.
[0064] S4. Parameter Optimization: Bayesian optimization algorithm is used, such as... Figure 3 As shown, the hyperparameters of long short-term memory networks or supervised learning models are automatically optimized.
[0065] Model performance is highly sensitive to hyperparameters (such as the number of LSTM layers, number of units per layer, time window length, learning rate, residual model complexity, etc.), while artificial grids or random search are inefficient and costly in high-dimensional mixed spaces. This invention employs Bayesian optimization (based on Gaussian processes or other surrogate models) to intelligently search for hyperparameters, minimizing validation set error or comprehensively controlling performance indicators. The core idea of Bayesian optimization is to approximate the objective function with a probabilistic model and select the next evaluation point based on uncertainty information, rapidly converging to good parameters through a small number of model training iterations. The surrogate model then searches for candidate hyperparameters... Give the predicted mean with standard deviation Subsequently, this invention employs the expected improvement (EI) as the acquisition function to balance exploration (high uncertainty region) and utilization (low mean region), and its expression is:
[0066]
[0067] in, The optimal target value observed so far. The CDF is a standard normal distribution. This function tends to explore when there is high surrogate uncertainty and tends to exploit when the mean is significantly better than the current optimum.
[0068] For discrete or integer parameters (such as the number of layers), ordered encoding or hybrid kernels can be used in the surrogate model; for high-dimensional spaces, a tree-structured TPE can be used to replace the Gaussian process; and long-term indicators of the control end (such as the cumulative reward or overmining rate of RL) can be incorporated into the optimization objective to achieve "training for control" oriented hyperparameter selection. To reduce online costs, Bayesian optimization can evaluate multiple candidates in parallel in offline simulation or historical playback environments, and finally deploy them after verification in a small-scale field experiment.
[0069] S5. Decision Control: Built-in reinforcement learning agent outputs mud concentration control strategy based on system status.
[0070] Mud concentration regulation is a sequential decision problem: current concentration adjustments will affect over-excavation risk and settlement over the next few minutes to hours through grouting seepage and formation mechanics, thus requiring consideration of long-term cumulative effects. This problem is modeled as a Markov Decision Process (MDP), where the state includes short-term predictions from an LSTM (the probability or amount of over-excavation in the next few steps), current and historical sensor observations, and recent control actions; the action is the mud concentration setting (continuous or discrete); and the reward is a combination of over-excavation risk, mud cost, and the cost of change.
[0071] Reward design is crucial to the success of reinforcement learning. If the reward is solely based on a negative immediate over-digging probability, the strategy may aggressively increase mud concentration to mitigate risk while ignoring costs; conversely, if the reward is solely based on concentration deviation, it may lead to tolerance of over-digging. Therefore, this invention employs a weighted composite reward:
[0072]
[0073] in, The over-excavation probability is predicted based on the current model. This represents the actual mud concentration. This is a nominal or optimal concentration reference value. The weighting is adjustable. This reward implies a balance between safety and economics: it penalizes both high risk and excessive deviation from nominal concentrations.
[0074] For reinforcement learning agents in the decision control module, SAC or DDPG are recommended for continuous action spaces due to their superior performance in sample efficiency and stability; Double DQN can be used for discrete action spaces. Training should prioritize offline training in a high-fidelity simulation environment or on real historical trajectories, followed by online fine-tuning with a low exploration rate (e.g., low noise). A dual-track parallel validation period should be implemented between policy output and human suggestions to ensure safety. To address policy drift and model errors, rule-based protection (such as single-step maximum concentration change limits, recovery thresholds, and manual takeover buttons) should be added to the control loop to ensure rapid rollback in cases of policy instability or abnormal behavior.
[0075] Example 2
[0076] The LSTM-based and supervised learning-optimized pipe jacking over-excavation prediction and mud concentration adjustment system of this embodiment includes:
[0077] Data acquisition and preprocessing module: Acquires multi-source time-series data and static geological data from the construction site, and preprocesses the multi-source time-series data;
[0078] Time series prediction module: It has a built-in long short-term memory network to model the preprocessed multi-source time series data and output preliminary over-mining amount or over-mining probability prediction results.
[0079] Error correction module: It has a built-in supervised learning model to model the residuals of the time series prediction module and adds the correction value to the preliminary prediction result to obtain the final prediction result;
[0080] Parameter optimization module: Uses Bayesian optimization algorithm to automatically optimize the hyperparameters of long short-term memory networks or supervised learning models;
[0081] Decision control module: It has a built-in reinforcement learning agent that outputs mud concentration regulation strategies based on the system status;
[0082] System integration and deployment module: Integrates and deploys the above modules on edge computing nodes or the cloud, providing real-time inference, control decision-making and human-computer interaction functions.
[0083] The system implementation of this invention includes: an edge data acquisition layer (on-site PLC / RTU and sensors), a data transmission and preprocessing layer (data buffering, time alignment, anomaly detection), a model inference layer (deploying LSTM+ supervised models and fusion modules at the edge or in the cloud), an optimization decision layer (Bayesian optimization service and reinforcement learning strategy engine), and a display and control layer (visual interface, alarm and automatic control interface). To meet the real-time and reliability requirements of construction sites, key inference and control modules can be deployed on edge computing nodes, and necessary summaries can be reported to the cloud for offline training and large-scale optimization.
[0084] In terms of real-time performance, the system should support end-to-end latency control from sensor data acquisition to policy output within the range of several seconds to tens of seconds, depending on the sampling frequency and control accuracy requirements. To reduce latency, lightweight network structures, model quantization, and distillation techniques can be employed; local threshold logic can be implemented for key thresholds (such as alarms for over-digging) to ensure immediate response.
[0085] The safety mechanisms include: (1) redundant sensors and data verification. If a sensor is abnormal, it will automatically be downgraded and prompted on the interface; (2) the strategy output is limited by engineering rules (maximum / minimum concentration, single change range); (3) automatic decision bypass mechanism: when the strategy confidence is low or communication is abnormal, switch to empirical rules or manual control mode; (4) model version management and rollback, which makes it easy to restore to the historical stable version when the model performance deteriorates.
[0086] The human-machine interface design should emphasize risk interpretability: displaying over-excavation probability curves for several future steps, the historical changes and expected benefits of mud concentration in the strategy recommendations, and providing causal explanations for the strategy recommendations (e.g., "Current earth pressure increases + grouting pressure decreases → risk increases, it is recommended to increase mud concentration by 0.5%)), so that construction personnel can quickly judge and make decisions. The offline component is responsible for periodic retraining, Bayesian hyperparameter search, and long-term strategy optimization using reinforcement learning, ensuring that the system continuously adapts and updates itself as construction progresses and geological information becomes available.
Claims
1. A method for predicting over-excavation and adjusting mud concentration in pipe jacking based on LSTM and supervised learning optimization, characterized in that, Includes the following steps: Data acquisition and preprocessing: Collect multi-source time-series data and static geological data from the construction site, and preprocess the multi-source time-series data; Time series prediction: Built-in LSTM is used to model the preprocessed multi-source time series data and output preliminary prediction results of over-mining amount or over-mining probability. Error correction: The built-in supervised learning model models the residuals of time series predictions and adds the correction value to the preliminary prediction result to obtain the final prediction result; Parameter optimization: The Bayesian optimization algorithm is used to automatically optimize the hyperparameters of the LSTM or supervised learning model; Decision control: The built-in reinforcement learning agent outputs mud concentration regulation strategies based on the system state.
2. The method for predicting over-excavation and adjusting mud concentration in pipe jacking based on LSTM and supervised learning optimization according to claim 1, characterized in that, The hidden state update expression of the LSTM is: in, This represents the cell state at the previous moment. , , These are the activation vectors for the input, forget, and output gates, respectively. This represents the candidate cell state.
3. The method for predicting over-excavation and adjusting mud concentration in pipe jacking based on LSTM and supervised learning optimization according to claim 1, characterized in that, The error correction specifically includes: Generate preliminary predictions using LSTM Calculate the preliminary prediction results of LSTM. The residual between the actual value and the true value; Using a supervised learning model with residuals as the objective. The output of the supervised learning model is used as a correction value, which is added to the initial prediction result to obtain the final corrected prediction result: in, This represents the input feature vector. For supervised learning model parameters.
4. The method for predicting over-excavation and adjusting mud concentration in pipe jacking based on LSTM and supervised learning optimization according to claim 3, characterized in that, The input feature vector It includes one or more of the following: hidden state of LSTM, geocoding, and statistical features.
5. The method for predicting over-excavation and adjusting mud concentration in pipe jacking based on LSTM and supervised learning optimization according to claim 1, characterized in that, When correcting errors, alternating training, cascaded training, or joint training are employed.
6. The method for predicting over-excavation and adjusting mud concentration in pipe jacking based on LSTM and supervised learning optimization according to claim 1, characterized in that, When optimizing the parameters, the desired improvement in EI is used as the acquisition function to balance exploring regions with high uncertainty and utilizing regions with low mean. Its expression is: in, The optimal target value observed so far. The CDF is a standard normal distribution.
7. The method for predicting over-excavation and adjusting mud concentration in pipe jacking based on LSTM and supervised learning optimization according to claim 1 or 6, characterized in that, When optimizing the parameters, ordered encoding is used for discrete or integer parameters, or a hybrid kernel is used in the surrogate model; the TPE with a tree structure is used to replace the Gaussian process in the high-dimensional space; and long-term indicators of the control end are incorporated into the optimization objective.
8. The method for predicting over-excavation and adjusting mud concentration in pipe jacking based on LSTM and supervised learning optimization according to claim 1, characterized in that, The decision control is modeled as a Markov decision process, with the state including short-term predictions from LSTM, current and historical sensor observations, and the most recent control action; the action is the mud concentration setting; and the reward is a combination of over-excavation risk, mud cost, and change cost.
9. The method for predicting over-excavation and adjusting mud concentration in pipe jacking based on LSTM and supervised learning optimization according to claim 8, characterized in that, The reward is a weighted composite reward: in, The over-excavation probability is predicted based on the current model. This represents the actual mud concentration. This is a nominal or optimal concentration reference value. The weights are adjustable.
10. A pipe jacking over-excavation prediction and mud concentration adjustment system based on LSTM and supervised learning optimization, the system being used to implement the pipe jacking over-excavation prediction and mud concentration adjustment method based on LSTM and supervised learning optimization as described in any one of claims 1 to 9, characterized in that, include: Data acquisition and preprocessing module: Acquires multi-source time-series data and static geological data from the construction site, and preprocesses the multi-source time-series data; Time series prediction module: It integrates LSTM to model the preprocessed multi-source time series data and output preliminary over-excavation amount or over-excavation probability prediction results. Error correction module: It has a built-in supervised learning model to model the residuals of the time series prediction module and adds the correction value to the preliminary prediction result to obtain the final prediction result; Parameter optimization module: Uses Bayesian optimization algorithm to automatically optimize the hyperparameters of LSTM or supervised learning models; Decision control module: It has a built-in reinforcement learning agent that outputs mud concentration regulation strategies based on the system status; System integration and deployment module: Integrates and deploys the above modules on edge computing nodes or the cloud, providing real-time inference, control decision-making and human-computer interaction functions.