LSTM-based shield synchronous grouting slurry time series deformation prediction method and system

Abstract

The application discloses a shield synchronous grouting slurry time sequence deformation prediction method and system based on LSTM. The method comprises the following steps: collecting time sequence data of the whole process of slurry consolidation deformation; dividing samples and predicting the maximum deformation of each sample based on static influence parameters after preprocessing; constructing a time sequence feature sequence fusing static parameters, time characteristics and the maximum deformation, and obtaining a modeling data set through sample conversion and normalization; constructing and training an LSTM time sequence prediction model; inputting the historical feature sequence of a sample to be predicted into the model to obtain a deformation ratio prediction result, and multiplying the deformation ratio prediction result by the maximum deformation to obtain an actual deformation prediction result. The application sets the prediction target as the deformation ratio instead of the absolute deformation, eliminates the scale difference under different working conditions, improves the generalization ability and the prediction accuracy, replaces repeated physical experiments with data driving, and provides support for real-time decision-making of shield grouting.

Description

Technical Field

[0001] This invention relates to the field of shield tunnel construction technology, and in particular to a method and system for predicting the temporal deformation of grout during shield synchronous grouting based on LSTM. Background Technology

[0002] Shield tunneling, an automated construction method that involves excavating tunnels and simultaneously lining segments within the ground, effectively minimizes the impact on surface traffic and underground pipelines while ensuring the stability of the excavation face and surrounding rock strata. Therefore, it is widely used in urban underground engineering construction. However, the tunneling process inevitably disturbs the surrounding soft soil layers, creating annular gaps between the shield shell and the assembled segments. Improper handling can lead to ground subsidence and even damage to nearby buildings. To control ground deformation, simultaneous grouting is commonly used in engineering projects. This involves injecting grout material simultaneously with the formation of the shield tail gap to promptly fill the void and suppress ground subsidence. Therefore, the complete phase transition behavior of the grout from liquid injection to solid-state stability directly determines the grouting control effect and the development process of ground subsidence.

[0003] Currently, research on the performance of synchronous grouting slurries mainly relies on traditional physical experimental methods, measuring slurry fluidity, filling capacity, and consolidation strength through laboratory tests. However, these methods generally suffer from long experimental cycles and high costs. Furthermore, the loading methods, boundary constraints, and environmental factors under laboratory conditions often fail to fully replicate the real-world working conditions of tunnel engineering, resulting in significant limitations in understanding the entire phase transition process of the slurry. More importantly, existing research largely focuses on the rheological properties and filling capacity of the slurry in its liquid phase, or only on its final strength and total deformation after complete consolidation, while paying insufficient attention to the phase transition process from liquid to solid, especially the temporal characteristics of deformation development over time during consolidation. In fact, these temporal characteristics not only reveal the dynamic mechanism of slurry consolidation but also directly relate to the temporal evolution of ground settlement and the real-time adjustment strategies of grouting technology, making them one of the core pieces of information for accurately controlling the impact of shield tunneling construction.

[0004] Due to a lack of systematic understanding of the deformation law of grout consolidation time, current engineering practice mainly relies on repeated experiments and empirical proportions for trial-and-error control. This approach cannot accurately describe the nonlinear evolution process of grout deformation, nor can it accurately predict and control ground settlement in real time during construction. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the problem that the nonlinear evolution law of deformation is difficult to accurately describe due to the neglect of the temporal characteristics of slurry consolidation in the prior art, which restricts the ability of real-time decision-making and risk control in engineering, while overcoming the limitations of traditional physical experimental methods that are time-consuming and labor-intensive.

[0006] To address the aforementioned technical problems, this invention provides a method for predicting the temporal deformation of grout during synchronous grouting in tunnel boring machines based on LSTM, comprising: S1: Time-series data acquisition is performed on the entire process of consolidation deformation of the synchronous grouting slurry in the shield tunnel to obtain the original time-series deformation dataset; S2: Preprocess the original time-series deformation dataset to obtain an effective time-series deformation dataset; based on the static influence parameters of slurry consolidation, divide the effective time-series deformation dataset into samples and obtain the predicted value of the maximum deformation of the entire consolidation process for each sample. S3: Construct a time-series feature sequence based on the effective time-series deformation dataset, the static influence parameters, and the predicted maximum deformation value; perform sample transformation and feature normalization processing on the time-series feature sequence to obtain the modeling feature dataset. S4: Construct an LSTM time series prediction model, and use the modeling feature dataset to train and optimize the LSTM time series prediction model to obtain a trained time series prediction model; S5: Input the historical time-series feature sequence of the sample to be predicted into the trained time-series prediction model to obtain the corresponding deformation ratio prediction result. Based on the deformation ratio prediction result and the maximum deformation prediction value of the sample to be predicted, obtain the slurry time-series deformation prediction result of the sample to be predicted.

[0007] In one embodiment of the present invention, the method for constructing the LSTM time series prediction model in step S4 is as follows: A time-series prediction network based on LSTM is constructed. The time-series prediction network includes an input layer, at least one long short-term memory network layer, a dropout layer, and a fully connected output layer. The input of the time-series prediction network is a historical time-series feature sequence, and the output is the slurry deformation ratio at the target time step. The slurry deformation ratio is the ratio of the slurry deformation amount to the predicted value of the maximum deformation amount.

[0008] In one embodiment of the present invention, in step S4, the method for training and optimizing the LSTM time series prediction model using the modeling feature dataset to obtain the trained time series prediction model is as follows: The modeling feature dataset is divided into a training set and an independent test set according to a preset ratio. K-fold cross-validation is performed on the training set, which is then divided into K subsets. One subset is used as the validation set, and the remaining subsets are used for model training and parameter tuning. The key hyperparameters of the LSTM time series prediction model are optimized with the goal of minimizing the mean absolute error of K-fold cross-validation. The optimal hyperparameter combination that minimizes the model's prediction error is determined. After configuring the network structure of the LSTM time series prediction model based on the optimal hyperparameter combination, the model is iteratively trained using the complete training set. After training, the model's prediction performance is comprehensively verified using an independent test set to obtain the trained time series prediction model.

[0009] In one embodiment of the present invention, the method for collecting time-series data on the entire consolidation deformation process of the shield tunnel synchronous grouting slurry in step S1 is as follows: Various grouts were prepared to simulate different formation conditions and confining pressures. Consolidation deformation tests were conducted on each grout combination. The entire process of the grout from the start of pressure to the point of deformation stabilization was continuously monitored, and the original time-series deformation dataset was collected.

[0010] In one embodiment of the present invention, in step S2, the static influence parameters of slurry consolidation include slurry mix ratio parameters, formation type, and confining pressure value.

[0011] In one embodiment of the present invention, in step S2, the method for dividing the effective time-series deformation dataset into samples based on the static influence parameters of slurry consolidation and obtaining the predicted value of the maximum deformation during the entire consolidation process for each sample is as follows: Using a fixed set of static influence parameters as the unique partitioning unit, continuous temporal deformation data with the same static influence parameters in the effective temporal deformation dataset are divided into independent samples, thus completing the sample partitioning of the entire dataset. Using the static influence parameters corresponding to each independent sample as input features, a pre-trained random forest regression model is used to predict the maximum deformation value of the sample during the entire process of slurry injection from liquid to solid stability.

[0012] In one embodiment of the present invention, in step S3, the method for constructing a time-series feature sequence based on the effective temporal deformation dataset, the static influence parameter, and the predicted maximum deformation value is as follows: For each monitoring time point of each independent sample, a comprehensive feature vector is constructed that integrates static influence parameters, time-derived features, and the predicted value of maximum deformation. All comprehensive feature vectors are arranged in chronological order of monitoring time to form a continuous time-series feature sequence. The time-derived features include at least one of time, time square, and time cube.

[0013] In one embodiment of the present invention, the method for sample transformation of the time-series feature sequence in step S3 is as follows: Using the time-series feature sequence of each independent sample arranged in chronological order of monitoring time as the basic unit, a sliding window is used to progressively truncate the time-series feature sequence along the time axis. For each segment of continuous historical time-series feature sequence, the slurry deformation ratio at the target prediction time point corresponding to this segment of continuous historical time-series feature sequence is simultaneously matched. The matched historical time-series feature sequences and slurry deformation ratios are converted into a set of supervised learning samples until the supervised learning sample conversion of all time-series feature sequences is completed. The slurry deformation ratio is the ratio of the actual slurry deformation at the corresponding time step to the predicted maximum deformation value of the entire consolidation process corresponding to that sample.

[0014] In one embodiment of the present invention, the method for obtaining the predicted result of the slurry time-series deformation of the sample to be predicted based on the predicted deformation ratio and the predicted maximum deformation value of the sample to be predicted in step S5 is as follows: The deformation ratio prediction result of the target time step output by the trained time-series prediction model is multiplied by the maximum deformation prediction value of the entire consolidation process corresponding to the sample to be predicted, and the actual deformation prediction result of the slurry at the target time step of the sample to be predicted is obtained, thus completing the prediction of the full time-series deformation of the slurry.

[0015] This invention also provides an LSTM-based system for predicting the temporal deformation of grout during synchronous grouting in tunnel boring machines, comprising: The data acquisition module is used to collect time-series data on the entire process of consolidation deformation of the synchronous grouting slurry in the shield tunnel, and obtain the original time-series deformation dataset. The preprocessing and maximum deformation prediction module is used to preprocess the original time-series deformation dataset to obtain an effective time-series deformation dataset; based on the static influence parameters of slurry consolidation, the effective time-series deformation dataset is divided into samples, and the predicted value of the maximum deformation of the entire consolidation process corresponding to each sample is obtained. The feature construction module is used to construct a time-series feature sequence based on the effective time-series deformation dataset, the static influence parameters, and the predicted value of the maximum deformation, and to perform sample transformation and feature normalization processing on the time-series feature sequence to obtain the modeling feature dataset; The model training and optimization module is used to construct an LSTM time series prediction model. The LSTM time series prediction model is trained and optimized using the modeling feature dataset to obtain a trained time series prediction model. The prediction module is used to input the historical time-series feature sequence of the sample to be predicted into the trained time-series prediction model to obtain the corresponding deformation ratio prediction result. Based on the deformation ratio prediction result and the maximum deformation prediction value of the sample to be predicted, the slurry time-series deformation prediction result of the sample to be predicted is obtained.

[0016] The technical solution of the present invention has the following advantages compared with the prior art: This invention introduces the maximum deformation prediction value as a benchmark feature and sets the prediction target as the deformation ratio rather than the absolute deformation amount, effectively eliminating scale differences between samples under different working conditions. This allows the temporal prediction network to focus more on learning the temporal evolution of deformation rather than absolute values, significantly improving the model's prediction accuracy and generalization ability under multiple working conditions. Simultaneously, by replacing a large number of repetitive physical experiments with a data-driven model, it can complete the prediction of the entire slurry deformation process under different working conditions within seconds and quickly perform multi-parameter sensitivity analysis, saving time and economic costs. Furthermore, the two-layer temporal prediction network structure constructed in this invention can effectively capture long-term dependencies in the slurry deformation process, further improving prediction accuracy. Attached Figure Description

[0017] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the LSTM-based method for predicting the temporal deformation of grout during synchronous grouting in tunnel boring machines, as provided in this embodiment of the invention. Figure 2 This is a diagram of the final trained time-series prediction model architecture in an embodiment of the present invention; Figure 3 This is a comparison chart of the predicted deformation time history curve and the measured deformation fitting curve under two different working conditions in the embodiments of the present invention; Figure 4 This is a comparison chart of the measured and predicted values ​​of the time series prediction model in this embodiment of the invention on two new datasets. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0020] Example 1: like Figure 1 As shown, the present invention provides a method for predicting the temporal deformation of grout in shield tunneling synchronous grouting based on LSTM, comprising: S1: Time-series data acquisition is performed on the entire process of consolidation deformation of the synchronous grouting slurry in the shield tunnel to obtain the original time-series deformation dataset; S2: Preprocess the original time-series deformation dataset to obtain an effective time-series deformation dataset; based on the static influence parameters of slurry consolidation, divide the effective time-series deformation dataset into samples and obtain the predicted value of the maximum deformation of the entire consolidation process for each sample. S3: Construct a time-series feature sequence based on the effective time-series deformation dataset, the static influence parameters, and the predicted maximum deformation value; perform sample transformation and feature normalization processing on the time-series feature sequence to obtain the modeling feature dataset. S4: Construct an LSTM time series prediction model, and use the modeling feature dataset to train and optimize the LSTM time series prediction model to obtain a trained time series prediction model; S5: Input the historical time-series feature sequence of the sample to be predicted into the trained time-series prediction model to obtain the corresponding deformation ratio prediction result. Based on the deformation ratio prediction result and the maximum deformation prediction value of the sample to be predicted, obtain the slurry time-series deformation prediction result of the sample to be predicted.

[0021] This invention collects and preprocesses time-series data on the entire process of slurry consolidation and deformation. Based on static influence parameters, samples are divided and the predicted maximum deformation value of each sample is obtained. Then, a time-series feature sequence integrating static parameters, time features, and the predicted maximum deformation value is constructed. After sample transformation and normalization, a modeling feature dataset is obtained. An LSTM time-series prediction model is constructed for training and optimization. The historical time-series feature sequence of the sample to be predicted is input into the model to obtain the deformation ratio prediction result, which is then multiplied by the predicted maximum deformation value to obtain the actual slurry time-series deformation prediction result. This invention introduces the predicted maximum deformation value as a benchmark feature and sets the prediction target as the deformation ratio rather than the absolute deformation amount, effectively eliminating the scale difference between samples under different working conditions. This allows the time-series prediction network to focus more on learning the temporal evolution law of deformation, significantly improving the prediction accuracy and generalization ability of the model under multiple working conditions. At the same time, it replaces a large number of repetitive physical experiments with data-driven methods, and can complete the prediction of the entire process of grout deformation under different working conditions within seconds, saving time and economic costs. It provides efficient and accurate technical support for real-time decision-making and ground settlement risk control of shield tunnel synchronous grouting.

[0022] In this embodiment of the invention, in step S1, time-series data is collected on the entire process of consolidation deformation of the shield tunnel synchronous grouting slurry to obtain the original time-series deformation dataset. Through systematic experimental design and a dedicated consolidation deformation testing device, full-cycle time-series deformation data of the slurry from liquid injection to solid stability covering multiple working conditions are obtained, providing a high-quality data foundation for subsequent feature engineering and model training.

[0023] Design a four-factor, three-level orthogonal experiment to generate representative slurry mix proportions. The four factors are denoted as follows: A is the water-cement ratio (W / B), B is the bentonite to solids ratio (B / S), C is the bentonite to water ratio (Ben / W), and D is the cement to fly ash ratio (C / F).

[0024] The levels of each factor are as follows: the three levels of factor A (water-cement ratio) are 0.6, 1.0, and 1.4; the three levels of factor B (bentonite to solids ratio) are 0.5, 1.0, and 1.5; the three levels of factor C (bentonite to water ratio) are 0.1, 0.2, and 0.3; and the three levels of factor D (cement to fly ash ratio) are 0, 0.2, and 0.4. The corresponding relationships of the above factor levels are summarized in Table 1.

[0025] Table 1:

[0026] Based on a four-factor, three-level orthogonal experimental design, nine different slurry mix proportions were developed. The factor level combinations and the actual mass (in g) of each component per 1 kg of slurry for each scheme are shown in Table 2. For cement quality, For fly ash quality, For the quality of bentonite, For sand quality, The mass is the water mass. The mass of each component is calculated according to the following formula: ; Where A is the water-cement ratio, B is the bentonite-to-solids ratio, C is the bentonite-to-water ratio, and D is the cement-to-fly ash ratio. Based on the above ratios and the corresponding A, B, C, and D values ​​for each group of schemes in Table 1, the actual mass of each material in 1 kg of slurry was calculated, and the results are shown in Table 2.

[0027] Table 2:

[0028] In Table 2, A value of 0 indicates that no cement is used in this mix proportion, and only fly ash is used as the cementing material. This also falls within the slurry mix proportion range covered by this invention. Accurately weigh the cement, fly ash, bentonite, sand, and water according to the mass of each component listed in Table 2, and mix them thoroughly to obtain the slurry sample to be tested.

[0029] Furthermore, the prepared slurry was injected into a self-developed consolidation deformation device for testing.

[0030] In this embodiment of the invention, the device has a replaceable cushion structure at its bottom, with three types of cushion layers—sand, silt, and clay—laid out to simulate different geological formations. The device applies a set confining pressure to the slurry surface via an external air pressure control system, selecting three confining pressure levels: 100 kPa, 200 kPa, and 300 kPa. Before the test, the slurry is poured into the device container, ensuring the liquid level is horizontal. At the start of the test, the air pressure control system is activated to apply the preset confining pressure, and simultaneously, a laser displacement sensor is activated to continuously monitor the change in slurry level height over time. The data acquisition frequency can be set according to actual needs; in this embodiment, it is set to record the liquid level height value once per second.

[0031] For each slurry mix design in Table 2, tests were conducted sequentially under three formation conditions and three confining pressure conditions. That is: 9 slurry mix designs × 3 formation conditions × 3 confining pressure conditions = 81 independent experiments. Recording for each experiment began the moment the confining pressure was applied and continued until the slurry level changed by less than a preset threshold over a continuous period. At this point, the slurry was considered to have reached a stable solid state, and data acquisition was stopped.

[0032] Thus, each set of experiments yielded a complete time-deformation curve from the onset of compression to the point of deformation stabilization, with the curve containing deformation data at multiple time steps. Summarizing all 81 sets of experimental data, a raw time-series deformation dataset was constructed. This dataset contains over 1800 time-series data points, comprehensively covering the full-time behavior of grout consolidation deformation under different grout mix ratios, formation conditions, and confining pressures.

[0033] It should be understood that the specific parameters given in this embodiment (such as factor levels, confining pressure values, formation types, sampling frequencies, etc.) are merely illustrative examples and not intended to limit the scope of protection of this invention. Any method based on the same principle that uses orthogonal experimental design combined with a consolidation deformation simulation device to obtain slurry time-series deformation data falls within the scope of protection of this invention.

[0034] The original temporal deformation dataset obtained by the above method will be used for preprocessing, feature engineering and model training in subsequent steps, so as to achieve efficient and accurate prediction of slurry temporal deformation.

[0035] The collected raw time-series deformation data may contain outliers caused by transient equipment interference, sensor noise, or external vibrations. To ensure the validity of the data and the reliability of subsequent modeling, step S2 involves checking the raw monitoring data. When deformation data at a certain time point is found to deviate significantly from the trend of changes before and after that time point and does not conform to the physical laws of slurry consolidation deformation, the outlier is removed. Adjacent data points can be interpolated using linear interpolation to ensure the continuity of the time-series data. After data cleaning, a valid time-series deformation dataset is obtained.

[0036] Furthermore, in the consolidation deformation test of the grout injected synchronously into the shield tunnel, each set of experiments was uniquely determined by a fixed set of static working condition parameters. These static working condition parameters include grout mix proportions, stratum type, and confining pressure. The grout mix proportions include the water-cement ratio, bentonite-to-solids ratio, bentonite-to-water ratio, and cement-to-fly ash ratio; the stratum type is divided into sand, silt, and clay; and the confining pressure is set at three levels: 100 kPa, 200 kPa, and 300 kPa. Since the orthogonal experiment in this embodiment generated nine grout mix proportion schemes, and each proportion was tested under three strata and three confining pressures, a total of eighty-one different combinations of static working condition parameters were formed.

[0037] Using a fixed set of static working parameters as the unique dividing unit, continuous time-series deformation data with the same static working parameters in the effective time-series deformation dataset are divided into independent samples. In other words, all monitoring data from the same slurry mix ratio, the same formation type, and the same confining pressure conditions constitute one sample, and each sample corresponds to a complete time-deformation curve from the start of slurry induction to deformation stabilization. Following this method, the entire effective time-series deformation dataset is divided into eighty-one independent samples.

[0038] Furthermore, for each independent sample, it is necessary to obtain the predicted value of the maximum deformation of that sample during the entire consolidation process of the slurry. This value is the theoretical extreme value predicted by the regression model based on static working condition parameters, denoted as... The predicted maximum deformation value represents the potential deformation capacity of the slurry under specific working conditions.

[0039] Preferably, in this embodiment of the invention, a random forest regression model is used to achieve this prediction function. Random forest is an ensemble learning algorithm based on multiple decision trees, which can effectively handle regression problems with high-dimensional input features and small sample data, while also having strong noise resistance and overfitting resistance. In the specific implementation, the number of decision trees in the random forest is set to five hundred to ensure sufficient prediction accuracy and stability.

[0040] Training a random forest model requires sample data. Based on eighty-one pre-defined independent samples, for each sample, its static working condition parameters are used as the input feature vector, and the final deformation at which deformation stabilizes in the actual monitoring curve of that sample is used as the output label, thus constructing a training dataset. The input feature vector contains seven dimensions: water-cement ratio, bentonite to solids ratio, bentonite to water ratio, cement to fly ash ratio, stratum type code, and confining pressure value. The stratum type is quantified using discrete coding: sand is coded as one, silt as two, and clay as three. These input features and their corresponding final deformation labels are then fed into the random forest model for training. The model learns the mapping relationship between static parameters and final deformation from a large number of samples, gradually fitting a regression function that can predict the maximum deformation based on the static parameters.

[0041] Once the random forest model is trained, it can be used to predict the maximum deformation of other samples. For any independent sample, simply input its static load parameters into the trained random forest model, and the model can quickly output the predicted maximum deformation value for that sample. No additional physical tests are required; the process can be completed in milliseconds based solely on the slurry mix ratio, formation type, and confining pressure.

[0042] Through data preprocessing, sample partitioning, and acquisition of the maximum deformation prediction value, the original time-series deformation dataset was organized into a series of independent samples with static working condition parameters and maximum deformation prediction values.

[0043] Furthermore, in step S3, after completing the sample division and obtaining the predicted value of the maximum deformation during the entire consolidation process of each sample, feature engineering is performed on each independent sample to transform the original time-deformation data into a feature sequence suitable for the input of the time-series prediction model.

[0044] For each independent sample, the original data is a continuous time-deformation curve from the start of slurry compression to the point of deformation stabilization. To fully explore the inherent laws in the slurry deformation process, a multi-dimensional comprehensive feature vector is constructed for each monitoring time point on the curve.

[0045] Specifically, the feature vector at each time point contains ten dimensions. Four slurry mix proportions from the static influence parameters are incorporated into the feature vector: water-cement ratio, bentonite-to-solids ratio, bentonite-to-water ratio, and cement-to-fly ash ratio. These parameters determine the physicochemical properties of the slurry and are fundamental factors influencing deformation behavior. Secondly, the geological strata type corresponding to that time point is added to the feature vector in a discrete encoding form: sand is coded as 1, silt as 2, and clay as 3. The confining pressure value is directly used as a feature dimension in its numerical form.

[0046] Considering that slurry consolidation deformation is a typical nonlinear process, with the deformation rate initially rapid and then slowing down over time, this pattern cannot be fully described using only a linear time variable. Therefore, this embodiment introduces time-derived features into the feature vector, specifically including the time value at the current monitoring time point, the square of time, and the cube of time. These three time features together constitute a polynomial extension of the time dimension, enabling the model to better fit the nonlinear trajectory of the deformation curve from rapid change to slow convergence.

[0047] Furthermore, in this embodiment of the invention, the predicted maximum deformation value obtained by the random forest model for this sample is used as the last dimension of the feature vector. This feature provides the model with a potential scale benchmark for slurry deformation under the current working condition, enabling subsequent time-series prediction models to perceive the differences in deformation capacity between different working conditions, thereby focusing more on learning the relative evolution law of deformation rather than absolute values.

[0048] By arranging the ten-dimensional feature vectors corresponding to each time point in chronological order of monitoring, a continuous temporal feature sequence for the sample is formed. This sequence fully preserves the dynamic information, static parameters, and scale reference information throughout the entire slurry consolidation process.

[0049] Furthermore, after constructing the temporal feature sequence, it needs to be converted into a supervised learning sample format suitable for training and prediction of temporal prediction networks. Since temporal prediction networks such as LSTM typically require a fixed-length historical information segment as input and a predicted value at a future time as output, a sliding window method is used to cut and reassemble the feature sequence.

[0050] The historical time step T is set to 4, meaning that the feature sequences of four consecutive time steps are used as input. Specifically, starting from the beginning of the temporal feature sequence of each independent sample, the feature vectors of the 1st to 4th time steps are extracted sequentially and combined into an input sequence. Simultaneously, the deformation ratio of the 5th time step immediately following is used as the output label corresponding to this input sequence. The deformation ratio is defined as the ratio of the actual deformation of the slurry at that time step to the predicted maximum deformation value of that sample.

[0051] The sliding window is moved forward one time step, and the feature sequences from time steps 2 to 5 are extracted as input, while the deformation ratio from time step 6 is used as output. This process is repeated until the window slides to the end of the sequence and no complete input-output pair can be formed. Through the above operations, the continuous temporal feature sequence of each independent sample is transformed into multiple supervised learning samples. Each sample contains a fixed-length historical temporal feature input sequence and a target deformation ratio label for training or prediction.

[0052] Because the physical meanings of each dimension in the feature vector constructed in this embodiment differ, their dimensions and numerical ranges vary significantly. For example, the confining pressure value is on the order of hundreds, the time value may be on the order of thousands, while the slurry mix ratio parameter and formation coding are usually within a smaller range of 0 to 3. If the unprocessed features are directly input into the model, the features with larger values ​​will dominate during training, causing the model to ignore features with smaller values ​​but important physical meanings, affecting the model's convergence speed and prediction accuracy. To solve this problem, this embodiment performs maximum / minimum value normalization on all features before inputting the supervised learning samples into the model. The normalization formula is: , in, These are the original eigenvalues. The minimum value of this feature across the entire dataset. This represents the maximum value of the feature across the entire dataset. Through this linear transformation, the numerical range of each feature dimension is mapped to... Within the closed interval, the difference in dimensions is eliminated, so that each feature has equal importance weight in model training.

[0053] After feature vector construction, sliding window sample transformation, and extremum normalization, a structured modeling feature dataset was finally formed, containing thousands of "feature-label" sample pairs. Each sample pair consists of a fixed-length historical time series feature sequence and a corresponding deformation ratio label.

[0054] Furthermore, in step S4, a time-series prediction model with a long short-term memory network as its core is constructed to learn the evolution of the slurry deformation ratio over time.

[0055] In this embodiment of the invention, the constructed temporal prediction network sequentially includes an input layer, at least one long short-term memory (LSTM) network layer, a dropout layer, and a fully connected output layer. The network input is a historical temporal feature sequence, i.e., a comprehensive feature vector of the previous several time steps; the network output is the slurry deformation ratio at the target time step. This deformation ratio is defined as the ratio of the slurry deformation amount to the predicted maximum deformation amount corresponding to the sample, and its value ranges from 0 to 1.

[0056] Since the current state of the slurry during consolidation and deformation is highly dependent on the historical state, traditional single LSTM layers may be inefficient in capturing this long-term dependency.

[0057] To address this, a two-layer LSTM network is constructed. The first LSTM layer is used to initially extract features and dependencies from the time series. The second LSTM layer performs higher-level abstraction and refinement based on the features extracted by the first layer, to better represent the deformation pattern of the slurry from rapid deformation to slow stabilization throughout the entire process. A dropout layer is placed after both the first and second LSTM layers. During model training, the dropout layers randomly block the outputs of some neurons according to a set probability, preventing them from participating in the forward and backward propagation of the current round, thereby effectively preventing overfitting and enhancing the model's generalization ability. After the two LSTM layers and the dropout layers, a fully connected layer is connected to map the high-dimensional features extracted by the LSTM to the final output value, i.e., the deformation ratio prediction result at the current time step.

[0058] Furthermore, after constructing the aforementioned two-layer LSTM network structure, it needs to be trained and optimized using the modeling feature dataset. In this embodiment, the modeling feature dataset is divided into a training set and an independent test set according to a preset ratio. The training set is used for learning model parameters, and the independent test set is used for validating the final model performance. Typically, 80% of the data is used as the training set, and 20% as the independent test set.

[0059] On the training set, K-fold cross-validation is used to train the model and tune its hyperparameters. Specifically, the training set is divided into K subsets, where K is 5 in this embodiment. One subset is used as the validation set, and the remaining K-1 subsets are used for model training and parameter tuning. After each round of training, the mean absolute error of the model on the validation set is calculated. After K rounds of training, K validation errors are obtained, and their average value is used as the model performance evaluation metric for that hyperparameter combination.

[0060] Since the predictive performance of LSTM models is highly sensitive to the selection of hyperparameters, including the number of hidden units in each LSTM layer and the dropout rate in dropout layers, this embodiment employs a Bayesian optimization algorithm to automatically optimize key hyperparameters. Bayesian optimization uses minimizing the mean absolute error of K-fold cross-validation as its direct objective function.

[0061] Several sets of hyperparameter combinations are randomly initialized within the preset hyperparameter search space, and the mean absolute error of K-fold cross-validation is calculated for each set of hyperparameters.

[0062] Based on existing evaluation results, a Gaussian process surrogate model is constructed to fit the distribution of the objective function. The expected value of the acquisition function is used to balance exploration and utilization in the parameter space, iteratively selecting the next most promising hyperparameter combination for evaluation. For each new hyperparameter combination evaluated, the surrogate model is updated, and the next iteration continues. After several iterations, the optimal hyperparameter combination that minimizes the verification error is obtained.

[0063] After obtaining the optimal hyperparameter combination, the number of hidden units and dropout rate of the LSTM network are configured according to these optimal parameters, and then the model is iteratively trained using all training set samples. The model training uses the Adam optimizer with an initial learning rate of 0.001 and a piecewise decay strategy to optimize the convergence process. The batch size is set to 32, and the maximum training epochs are set to 600. During training, the model continuously updates the network weights through backpropagation, gradually learning the mapping relationship from historical time-series feature sequences to future deformation ratios.

[0064] After training, the model was comprehensively validated using an independent test set. Historical time-series feature sequences from the test set were input into the model to obtain the deformation ratio prediction result. This result was then multiplied by the predicted maximum deformation value corresponding to that sample to reconstruct the actual deformation. The model's predictive performance on the test set was evaluated using metrics such as the coefficient of determination, root mean square error, and mean absolute error. Experimental results show that the model constructed in this embodiment exhibits excellent fitting accuracy on both the training and test sets, and also demonstrates good generalization ability for unseen combinations of working conditions.

[0065] Furthermore, in this embodiment of the invention, the trained time-series prediction model is used to predict the time-series deformation of the slurry in the sample to be predicted.

[0066] For any given working condition to be predicted, the static influence parameters and historical time-series deformation data of the initial stage are obtained. This data can be acquired in real time using laser displacement sensors or pressure sensors installed at the tunnel boring machine (TBM) construction site, or by retrieving existing historical monitoring records from the engineering database. After obtaining the raw data, it needs to be preprocessed using the same methods as the training data, including outlier removal and missing value filling, to ensure data quality.

[0067] Furthermore, according to the processing flow of steps S1 to S3 mentioned above, feature construction and normalization operations are performed on the data of the sample to be predicted.

[0068] Following the method in step S3, a ten-dimensional comprehensive feature vector is constructed for each monitoring time point in the historical time series. The feature vector includes four slurry mix ratio parameters, formation type code, confining pressure value, current time value, time squared, time cubed, and the predicted maximum deformation value. These feature vectors are arranged chronologically according to the monitoring time to form a continuous time-series feature sequence. To maintain consistency with training, a sliding window method is used to extract historical time-series feature sequences. The historical time step is set to 4, meaning that the feature sequences of the four most recent consecutive time steps are selected as the model input. Simultaneously, maximum / minimum value normalization is performed on each dimension of the input feature sequence.

[0069] like Figure 2 As shown, the normalized historical time-series feature sequence is input into the trained LSTM time-series prediction model. The model undergoes forward propagation through the input layer, the first LSTM layer, the first dropout layer, the second LSTM layer, the second dropout layer, and a fully connected layer, finally outputting the predicted deformation ratio for the next time step, denoted as . The deformation ratio is a value between 0 and 1, representing the proportion of the slurry deformation at the predicted time to the predicted maximum deformation value of that sample. Since the model learns the general pattern of the deformation ratio evolution over time in a large number of samples during the training phase, rather than the absolute deformation, this ratio has good generalization ability and can adapt to the differences in deformation morphology under different working conditions.

[0070] To obtain the true absolute amount of deformation, the deformation ratio prediction results output by the model will be used. The predicted maximum deformation value of this sample Multiplying these two components yields the predicted actual deformation of the slurry at the target time step for the sample to be predicted. The formula is as follows: , The product result is expressed in length units (such as millimeters) and can be directly used for engineering analysis and decision-making.

[0071] If it is necessary to predict the slurry deformation at multiple future time steps, a recursive prediction method can be used: The deformation ratio predicted at the current time step (or the actual deformation after reconstruction) is used as part of the historical information for subsequent predictions. A comprehensive feature vector containing the latest information is reconstructed and combined with the feature sequences from the previous three time steps to form a new input sequence of length 4, used to predict the deformation ratio at the next time step. This process is iterated to gradually generate the deformation prediction curve for the entire future time series. For example, after obtaining the predicted value for the 5th time step, the feature sequences from the 2nd to 5th time steps are used as input to predict the 6th time step; then the 3rd to 6th time steps are used to predict the 7th time step, and so on, until the required full-process prediction is completed.

[0072] This embodiment effectively eliminates the impact of differences in the absolute scale of deformation under different working conditions on model performance by setting the prediction target as the deformation ratio rather than the absolute deformation amount. This allows the LSTM network to focus on learning the temporal evolution shape and rate of change of the deformation curve. Because the maximum deformation prediction value... The actual deformation amount, which has been accurately given by the random forest model based on static parameters, includes both working condition-specific information and inherits the high-precision fitting ability of the LSTM model to time series patterns.

[0073] The predicted results can be output to the display terminal of the tunnel boring machine (TBM) construction control system via a computer interface, or directly transmitted to the automatic grouting parameter adjustment module. For example, if it is predicted that the grout deformation rate will be too fast or the final deformation will exceed the safety threshold in the near future, the system can issue an early warning signal and suggest that construction personnel adjust the grouting pressure, change the grout mix ratio, or adjust the grouting speed, thereby achieving proactive control over the risk of ground settlement.

[0074] To verify the prediction accuracy and generalization ability of the method of the present invention, the training process and performance evaluation of the model are further described in detail.

[0075] After constructing the modeling feature dataset, the complete dataset is randomly divided into a training set and an independent test set according to a preset ratio, with the training set accounting for 80% and the independent test set accounting for 20%. Five-fold cross-validation is performed on the training set, which involves dividing the training set into five equal subsets, using one subset as the validation set, and using the remaining four subsets for model training and parameter tuning.

[0076] Since the predictive performance of LSTM models is highly sensitive to the selection of hyperparameters, including the number of hidden units in each LSTM layer and the dropout rate, this embodiment employs a Bayesian optimization algorithm to automatically optimize key hyperparameters. Bayesian optimization uses minimizing the mean absolute error (MAO) of 5-fold cross-validation as its direct objective function. First, several sets of hyperparameter combinations are randomly initialized within a predefined hyperparameter search space, and the MAO is calculated for each set. Then, a Gaussian process surrogate model is constructed based on the existing evaluation results to fit the distribution of the objective function. The expected improvement acquisition function is used to balance exploration and utilization in the parameter space, iteratively selecting the next most promising hyperparameter combination for evaluation. Each time a new hyperparameter combination is evaluated, the surrogate model is updated, and the next iteration continues until the optimal hyperparameter combination that minimizes the validation error is found. Using this optimal hyperparameter configuration, the LSTM network structure is configured, the final model is trained on all training set samples, and performance validation is performed on an independent test set.

[0077] Furthermore, to evaluate the predictive performance of the model, three complementary metrics are used: the coefficient of determination (R²), the root mean square error (RMSE), and the mean absolute error (MAE). The formulas for each metric are as follows: , , , in, The total number of sample data. For monitoring data, For predictive data, To monitor the average value of the data, the predicted deformation ratio from the model output is compared with the corresponding predicted maximum deformation value. After multiplying to obtain the actual deformation value, all predicted indicators remain at a high level.

[0078] To verify the prediction accuracy and generalization ability of the method of the present invention, two independent validation datasets were used for model validation. Both datasets were completely independent new working condition data. For each dataset, the static features of the samples (formation type, confining pressure, water-cement ratio W / B, bentonite to solid ratio B / S, bentonite to water ratio Ben / W, cement to fly ash ratio C / F) were extracted.

[0079] Load the pre-trained model, import the new validation dataset, group the samples according to static features such as formation, confining pressure, and slurry mix ratio, and match the maximum deformation of each sample. The data is then serialized. For each sample, the measured deformation values ​​are used as initial conditions for the first four time steps. From the fifth time step onwards, a normalized feature vector (containing time polynomials, static parameters, etc.) is input through a sliding window. The deformation ratio is predicted by the LSTM model and then multiplied by Once the absolute deformation prediction value is obtained, the full-time-series prediction is completed by rolling point by point. For example... Figure 3 As shown, the predicted deformation time history curves of the model under two different working conditions exhibit a high degree of consistency with the measured fitted curves, demonstrating that the model can accurately fit the nonlinear trajectory of deformation development over time. Furthermore, as... Figure 4 As shown in the graph comparing the measured and predicted values ​​of the model's deformation on two new datasets, the data points are closely distributed near the diagonal, indicating that the model maintains high accuracy on external data, demonstrating its robustness and applicability.

[0080] Example 2: Based on the same inventive concept as Embodiment 1, this invention also provides an LSTM-based shield tunnel synchronous grouting slurry time-series deformation prediction system, used to implement the steps of the LSTM-based shield tunnel synchronous grouting slurry time-series deformation prediction method described in Embodiment 1, including the following modules: The data acquisition module is used to collect time-series data on the entire process of consolidation deformation of the synchronous grouting slurry in the shield tunnel, and obtain the original time-series deformation dataset. The preprocessing and maximum deformation prediction module is used to preprocess the original time-series deformation dataset to obtain an effective time-series deformation dataset; based on the static influence parameters of slurry consolidation, the effective time-series deformation dataset is divided into samples, and the predicted value of the maximum deformation of the entire consolidation process corresponding to each sample is obtained. The feature construction module is used to construct a time-series feature sequence based on the effective time-series deformation dataset, the static influence parameters, and the predicted value of the maximum deformation, and to perform sample transformation and feature normalization processing on the time-series feature sequence to obtain the modeling feature dataset; The model training and optimization module is used to construct an LSTM time series prediction model. The LSTM time series prediction model is trained and optimized using the modeling feature dataset to obtain a trained time series prediction model. The prediction module is used to input the historical time-series feature sequence of the sample to be predicted into the trained time-series prediction model to obtain the corresponding deformation ratio prediction result. Based on the deformation ratio prediction result and the maximum deformation prediction value of the sample to be predicted, the slurry time-series deformation prediction result of the sample to be predicted is obtained.

[0081] The data acquisition module, preprocessing and maximum deformation prediction module, feature construction module and perception, model training and optimization module and prediction module of the LSTM-based shield synchronous grouting slurry temporal deformation prediction system proposed in this embodiment are respectively used to implement steps S1, S2, S3, S4 and S5 in the LSTM-based shield synchronous grouting slurry temporal deformation prediction method in Embodiment 1. To avoid redundancy, they will not be described again here.

[0082] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0083] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0084] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0085] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0086] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for predicting the temporal deformation of grout during synchronous grouting in tunnel boring machines based on LSTM, characterized in that, include: S1: Time-series data acquisition is performed on the entire process of consolidation deformation of the synchronous grouting slurry in the shield tunnel to obtain the original time-series deformation dataset; S2: Preprocess the original temporal deformation dataset to obtain an effective temporal deformation dataset; Based on the static influence parameters of slurry consolidation, the effective time-series deformation dataset is divided into samples, and the predicted value of the maximum deformation during the entire consolidation process corresponding to each sample is obtained. S3: Construct a time-series feature sequence based on the effective time-series deformation dataset, the static influence parameters, and the predicted maximum deformation value; perform sample transformation and feature normalization processing on the time-series feature sequence to obtain the modeling feature dataset. S4: Construct an LSTM time series prediction model, and use the modeling feature dataset to train and optimize the LSTM time series prediction model to obtain a trained time series prediction model; S5: Input the historical time-series feature sequence of the sample to be predicted into the trained time-series prediction model to obtain the corresponding deformation ratio prediction result. Based on the deformation ratio prediction result and the maximum deformation prediction value of the sample to be predicted, obtain the slurry time-series deformation prediction result of the sample to be predicted.

2. The method for predicting the temporal deformation of grout in shield tunneling based on LSTM according to claim 1, characterized in that: In step S4, the method for constructing the LSTM time series prediction model is as follows: A time-series prediction network based on LSTM is constructed. The time-series prediction network includes an input layer, at least one long short-term memory network layer, a dropout layer, and a fully connected output layer. The input of the time-series prediction network is a historical time-series feature sequence, and the output is the slurry deformation ratio at the target time step. The slurry deformation ratio is the ratio of the slurry deformation amount to the predicted value of the maximum deformation amount.

3. The method for predicting the temporal deformation of grout in shield tunneling based on LSTM according to claim 1, characterized in that: In step S4, the method for training and optimizing the LSTM time series prediction model using the modeling feature dataset to obtain the trained time series prediction model is as follows: The modeling feature dataset is divided into a training set and an independent test set according to a preset ratio. K-fold cross-validation is performed on the training set, which is then divided into K subsets. One subset is used as the validation set, and the remaining subsets are used for model training and parameter tuning. The key hyperparameters of the LSTM time series prediction model are optimized with the goal of minimizing the mean absolute error of K-fold cross-validation. The optimal hyperparameter combination that minimizes the model's prediction error is determined. After configuring the network structure of the LSTM time series prediction model based on the optimal hyperparameter combination, the model is iteratively trained using the complete training set. After training, the model's prediction performance is comprehensively verified using an independent test set to obtain the trained time series prediction model.

4. The method for predicting the temporal deformation of grout in shield tunneling based on LSTM according to claim 1, characterized in that: In step S1, the method for collecting time-series data on the entire process of consolidation deformation of the shield tunneling synchronous grout is as follows: Various grouts were prepared to simulate different formation conditions and confining pressures. Consolidation deformation tests were conducted on each grout combination. The entire process of the grout from the start of pressure to the point of deformation stabilization was continuously monitored, and the original time-series deformation dataset was collected.

5. The method for predicting the temporal deformation of grout in shield tunneling based on LSTM according to claim 1, characterized in that: In step S2, the static influence parameters of slurry consolidation include slurry mix ratio parameters, formation type, and confining pressure value.

6. The method for predicting the temporal deformation of grout in shield tunneling synchronous grouting based on LSTM according to claim 1 or 5, characterized in that: In step S2, the method for dividing the effective time-series deformation dataset into samples based on the static influence parameters of slurry consolidation and obtaining the predicted maximum deformation value of each sample during the entire consolidation process is as follows: Using a fixed set of static influence parameters as the unique partitioning unit, continuous temporal deformation data with the same static influence parameters in the effective temporal deformation dataset are divided into independent samples, thus completing the sample partitioning of the entire dataset. Using the static influence parameters corresponding to each independent sample as input features, a pre-trained random forest regression model is used to predict the maximum deformation value of the sample during the entire process of slurry injection from liquid to solid stability.

7. The method for predicting the temporal deformation of grout in shield tunneling based on LSTM according to claim 1, characterized in that: In step S3, the method for constructing a time-series feature sequence based on the effective temporal deformation dataset, the static influence parameters, and the predicted maximum deformation value is as follows: For each monitoring time point of each independent sample, a comprehensive feature vector is constructed that integrates static influence parameters, time-derived features, and the predicted value of maximum deformation. All comprehensive feature vectors are arranged in chronological order of monitoring time to form a continuous time-series feature sequence. The time-derived features include at least one of time, time square, and time cube.

8. The method for predicting the temporal deformation of grout in shield tunneling based on LSTM according to claim 1, characterized in that: In step S3, the method for sample transformation of the time-series feature sequence is as follows: Using the time-series feature sequence of each independent sample arranged in chronological order of monitoring time as the basic unit, a sliding window is used to progressively truncate the time-series feature sequence along the time axis. For each segment of continuous historical time-series feature sequence, the slurry deformation ratio at the target prediction time point corresponding to this segment of continuous historical time-series feature sequence is simultaneously matched. The matched historical time-series feature sequences and slurry deformation ratios are converted into a set of supervised learning samples until the supervised learning sample conversion of all time-series feature sequences is completed. The slurry deformation ratio is the ratio of the actual slurry deformation at the corresponding time step to the predicted maximum deformation value of the entire consolidation process corresponding to that sample.

9. The method for predicting the temporal deformation of grout in shield tunneling based on LSTM according to claim 1, characterized in that: In step S5, the method for obtaining the predicted slurry time-series deformation of the sample to be predicted based on the predicted deformation ratio and the predicted maximum deformation value of the sample to be predicted is as follows: The deformation ratio prediction result of the target time step output by the trained time-series prediction model is multiplied by the maximum deformation prediction value of the entire consolidation process corresponding to the sample to be predicted, and the actual deformation prediction result of the slurry at the target time step of the sample to be predicted is obtained, thus completing the prediction of the full time-series deformation of the slurry.

10. A time-series deformation prediction system for shield tunnel synchronous grouting based on LSTM, characterized in that, include: The data acquisition module is used to collect time-series data on the entire process of consolidation deformation of the synchronous grouting slurry in the shield tunnel, and obtain the original time-series deformation dataset. The preprocessing and maximum deformation prediction module is used to preprocess the original temporal deformation dataset to obtain an effective temporal deformation dataset; Based on the static influence parameters of slurry consolidation, the effective time-series deformation dataset is divided into samples, and the predicted value of the maximum deformation during the entire consolidation process corresponding to each sample is obtained. The feature construction module is used to construct a time-series feature sequence based on the effective time-series deformation dataset, the static influence parameters, and the predicted value of the maximum deformation, and to perform sample transformation and feature normalization processing on the time-series feature sequence to obtain the modeling feature dataset; The model training and optimization module is used to construct an LSTM time series prediction model. The LSTM time series prediction model is trained and optimized using the modeling feature dataset to obtain a trained time series prediction model. The prediction module is used to input the historical time-series feature sequence of the sample to be predicted into the trained time-series prediction model to obtain the corresponding deformation ratio prediction result. Based on the deformation ratio prediction result and the maximum deformation prediction value of the sample to be predicted, the slurry time-series deformation prediction result of the sample to be predicted is obtained.

Images