Shield soil bin pressure intelligent prediction method and prediction system

Abstract

The invention provides a shield soil bin pressure intelligent prediction method and prediction system. The prediction method comprises the steps that basic data used for predicting the shield soil bin pressure are acquired, the basic data comprise forward data and reverse data, and the basic data are preprocessed; constructing a bidirectional prediction data set, wherein the bidirectional prediction data set comprises a forward prediction data set and a reverse prediction data set; constructing a hybrid deep learning model, and training the hybrid deep learning model to obtain a bidirectional prediction model; importing the bidirectional prediction model into an industrial personal computer; the industrial personal computer is connected to a shield tunneling machine PLC, and a shield tunneling soil bin pressure prediction value is comprehensively obtained by combining forward prediction and reverse prediction. According to the prediction method provided by the invention, the mixed deep learning prediction model is constructed, the bidirectional prediction model is obtained through training, the change trend of the pressure of the shield soil bin is accurately predicted and captured from the positive and negative aspects, the pressure value of the soil bin in the shield region can be efficiently and accurately predicted, and the safety of shield construction is guaranteed.

Description

Technical Field

[0001] The invention relates to the technical field of tunnel shield construction, and in particular to an intelligent prediction method and prediction system for shield soil bin pressure. Background Art

[0002] In the prediction of shield bunker pressure, the existing shield bunker pressure prediction model shows high prediction accuracy. For example, Lin Chunjin et al. (2021) established a bunker pressure prediction model based on BP neural network (PSO-BP) optimized by particle swarm algorithm. The model takes the main tunneling parameters of the shield as input and the bunker pressure as output. The predicted mean absolute percentage error (MAPE) on the test set is 6.28%, and the determination coefficient (R2) is 0.8313, which can accurately predict the bunker pressure. Liu et al. (2022) proposed a shield machine multi-point bunker pressure change prediction method based on hybrid deep learning. Discrete wavelet transform (DWT), one-dimensional convolutional neural network (1DCNN) and long short-term memory network (LSTM) are used to construct an intelligent prediction model to achieve scientific prediction of the trend of multi-point bunker pressure changes.

[0003] However, these methods are data-driven prediction models that are highly dependent on a large amount of high-quality data. Their effectiveness also depends to a large extent on a large amount of high-quality data. Uneven data quality or insufficient data will greatly limit the prediction performance of the model. In the case of low data quality or insufficient data, the prediction performance of the model may be greatly affected, and the model interpretation bias may occur.

[0004] Moreover, the generalization ability of existing models under different geological conditions and construction environments is limited, which affects the possibility of their widespread application. The "black box" characteristics of data-driven models make the prediction results difficult to interpret, affecting the transparency and credibility of the models. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides an intelligent prediction method and system for shield soil bin pressure, constructs a hybrid deep learning prediction model based on DWT-PSO-GA-LSTM, and trains a bidirectional prediction model to accurately predict and capture the changing trend of shield soil bin pressure from both positive and negative aspects, and can efficiently and accurately predict the soil bin pressure value in the shield area, thereby ensuring the safety of shield construction.

[0006] The technical solution of the present invention is as follows:

[0007] The present invention first provides a shield soil bin pressure intelligent prediction method, comprising the following steps:

[0008] Acquire basic data for predicting shield soil bin pressure, the basic data including forward data and reverse data, and pre-process the basic data;

[0009] Based on the preprocessed basic data, construct a bidirectional prediction data set, wherein the bidirectional prediction data set includes a forward prediction data set and a reverse prediction data set;

[0010] Based on the bidirectional prediction data set, a hybrid deep learning model is constructed, and the hybrid deep learning model is trained to obtain a bidirectional prediction model;

[0011] Importing the bidirectional prediction model into an industrial computer;

[0012] The industrial computer is connected to the PLC of the shield machine, and the bidirectional prediction model is connected to the PLC software program of the shield machine. The soil bin pressure value is forward predicted through the forward data, and the soil bin pressure value is reversely predicted in real time through the reverse data. The predicted value of the shield soil bin pressure is comprehensively obtained by combining the forward prediction and the reverse prediction.

[0013] In some feasible embodiments, the forward data include formation parameters, geometric parameters and groundwater parameters of the shield tunneling range, and the reverse data include shield tunneling parameters.

[0014] In some practicable embodiments, preprocessing the basic data includes:

[0015] (1) The basic data is normalized using the following formula (1):

[0016]

[0017] Among them, x is the normalized basic data, x0 is the initial basic data, and x max is the maximum value of the basic data, x min is the minimum value of the basic data;

[0018] (2) Use feature engineering to process the normalized basic data.

[0019] In some achievable embodiments, processing the normalized basic data using feature engineering includes:

[0020] (2.1) Perform one-hot encoding on the normalized basic data and convert the feature data with multiple categories into a binary matrix representation, as shown in the following formula (2):

[0021]

[0022] Among them, Encoded i is the code of the i-th category, and category is the actual category of the sample;

[0023] (2.2) The matrix is ​​regularized, where L1 regularization is used for feature selection, as shown in equation (3), and L2 regularization is used for data denoising, as shown in equation (4):

[0024]

[0025] Among them, λ is the regularization strength, w i is the feature weight after processing;

[0026] (2.3) Perform feature data interaction and create new features through feature data interaction.

[0027] In some achievable embodiments, constructing a forward prediction data set based on the preprocessed forward data includes:

[0028] Each segment ring is regarded as a unit cell and divided in the longitudinal section of the shield tunnel;

[0029] Collect and organize physical and mechanical parameters under different geological conditions, including density, elastic modulus and friction angle;

[0030] Record the formation parameters, geometric parameters and groundwater parameters of each cell, the formation parameters include geological type, length, proportion and physical and mechanical parameters of different geological soil bodies, the geometric parameters include cover soil thickness and excavation diameter, and the groundwater parameters include groundwater type and water level height;

[0031] A forward prediction dataset is constructed based on the recorded formation parameters, geometric parameters, and groundwater parameters.

[0032] In some feasible embodiments, in the construction of the hybrid deep learning model, the formation parameters, geometric parameters, groundwater parameters and shield parameters are used as input data, and the output data includes two aspects: instantaneous value prediction and control value prediction. Each aspect includes the mean soil bin pressure term, upper and lower gradient terms, left and right gradient terms, local heterogeneous terms and upper soil bin pressure value.

[0033] In some achievable embodiments, in constructing the hybrid deep learning model, a time series is assigned to the preprocessed basic data, the preprocessed basic data is converted into three-dimensional time series data, and a time series data set is constructed.

[0034] In some achievable embodiments, constructing a time series dataset includes:

[0035] The data of formation parameters, geometric parameters, and groundwater parameters from time 1 to n are used as the first time series input data of the entire bidirectional prediction data set;

[0036] The instantaneous soil bin pressure mean term P a, upper and lower gradient terms P b , left and right gradient terms P c , local heterogeneity term P d and the upper soil bin pressure value P1, and the soil bin pressure mean value of the control value Upper and lower gradient terms P b * , left and right gradient terms Local heterogeneity And the upper soil bin pressure value P1 * The data from time 1 to time n is used as the first time series output data of the entire bidirectional prediction data set;

[0037] Set the length of each time series of input data to n-5, and the length of each time series of output data to 5, that is, use the data of the first n-5 moments as input data, and the data of the last 5 moments as output data, and slide the window in sequence to obtain input data and output data;

[0038] The data from time 2 to n+1 is used as the second time series of the entire bidirectional prediction dataset. The above steps are repeated to construct the time series. Each time series input and output window slides backward by one unit until no new time series can be generated, thus forming a complete bidirectional prediction dataset required by the hybrid deep learning model.

[0039] In some feasible embodiments, the hybrid deep learning model is constructed by a DWT model, a PSO-GA model, and a LSTM model, wherein the DWT model is used to reduce the noise of the input data, the PSO-GA model is used to quickly extract the features of the input data, and the LSTM model is used to determine the changing trend of the pressure in the entire closed soil bin.

[0040] The present invention also provides a shield soil bin pressure intelligent prediction system, comprising an industrial computer, wherein the industrial computer is connected to the shield machine PLC and is used to execute the above prediction method to predict the shield soil bin pressure value.

[0041] The beneficial effects of the present invention compared with the prior art are as follows: the present invention provides a shield soil bin pressure intelligent prediction method and prediction system, including a shield soil bin pressure intelligent prediction method and a shield soil bin pressure intelligent prediction system based on an embedded industrial computer, constructing a hybrid deep learning prediction model based on DWT-PSO-GA-LSTM, and combining software and hardware to achieve effective connection between algorithm prediction and shield machine control, and successfully applied to shield construction sites. The specific advantages include at least one or more of the following:

[0042] (1) The prediction method provided by the present invention constructs a hybrid deep learning prediction model and trains a bidirectional prediction model to accurately predict and capture the changing trend of the shield soil bin pressure from both positive and negative aspects. It can efficiently and accurately predict the soil bin pressure value in the shield area. In this way, the main control program of the shield machine can be updated according to the comprehensive prediction value of the soil bin pressure, and the excavation parameters of the shield machine can be adjusted to ensure the safety of the shield construction.

[0043] (2) The prediction method provided by the present invention provides a comprehensive and reliable data basis for the training, verification and testing of hybrid deep learning models by inputting forward data (formation parameters, geometric parameters and groundwater parameters, etc.) and reverse data (shield parameters).

[0044] (3) The prediction method provided by the present invention effectively organizes the original construction data through normalization and feature engineering preprocessing data, providing a more reliable input data set for the subsequent artificial intelligence model, thereby improving the accuracy and reliability of the prediction results at the source of the data.

[0045] (4) The prediction method provided by the present invention constructs a hybrid deep learning model composed of DWT, PSO-GA, and LSTM, which can reduce the noise of the input data, thereby reducing the interference of the noise on the multi-dimensional soil bin pressure spatial field. It has low requirements on the quality of basic data, can quickly extract the features of the input data, and improve the training accuracy of the prediction model, thereby accurately predicting the value of the soil bin pressure and judging the changing trend of the pressure in the entire closed soil bin.

[0046] (5) Compared with the traditional manual control stage, the prediction method provided by the present invention has a significantly better control effect on the soil bin pressure in the intelligent prediction control stage, which meets the actual construction requirements on site.

[0047] (6) The prediction method provided by the present invention is combined with hardware to achieve effective connection between algorithm prediction and shield machine control through the combination of software and hardware, and is successfully applied to shield construction sites.

[0048] It should be understood that the implementation of any embodiment of the present invention does not mean that multiple or all of the above-mentioned beneficial effects must be possessed or achieved at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0050] The structures, proportions, sizes, etc. illustrated in this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with the technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantial technical significance. Any structural modification, change in proportion or adjustment in size shall still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0051] Figure 1 It is a schematic diagram of the overall process of a shield soil bin pressure intelligent prediction method of the present invention;

[0052] Figure 2 A schematic diagram of single-ring stratigraphic feature division according to an embodiment of the present invention;

[0053] Figure 3 A schematic diagram of a method for constructing a time series data set according to an embodiment of the present invention;

[0054] Figure 4 A hybrid deep learning model framework diagram of an embodiment of the present invention;

[0055] Figure 5 A schematic diagram of an LSTM neural network training process according to an embodiment of the present invention;

[0056] Figure 6 A structural diagram of an industrial computer device according to an embodiment of the present invention;

[0057] Figure 7 A parameter diagram of an industrial computer according to an embodiment of the present invention;

[0058] Figure 8 A flowchart of intelligent prediction of shield soil bin pressure according to an embodiment of the present invention;

[0059] Fig. 9 This is a diagram of the on-site application of intelligent prediction of shield soil bin pressure according to an embodiment of the present invention.

[0060] The same or corresponding symbols in the drawings indicate the same or corresponding parts. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clearly understood, the embodiments of the present invention are further described in detail below in conjunction with the embodiments and drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0062] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0063] It should be understood that the terms "include / comprise", "consist of..." or any other variations are intended to cover non-exclusive inclusion, so that a product, device, process or method that includes a series of elements includes not only those elements, but also may include other elements not explicitly listed when necessary, or also includes elements inherent to such product, device, process or method. In the absence of more restrictions, the elements defined by the sentence "include / comprise...", "consist of..." do not exclude the presence of other identical elements in the product, device, process or method that includes the elements.

[0064] It is also necessary to understand that terms such as "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside" and "outside" to indicate directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device, component or structure must have a specific direction, be constructed or operate in a specific direction, and should not be understood as a limitation on the present invention.

[0065] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0066] The implementation of the present invention is described in detail below in conjunction with preferred embodiments.

[0067] The present invention provides a shield soil bin pressure intelligent prediction method, see Figure 1The flowchart shown includes the following steps: S10, obtaining basic data for predicting shield soil bin pressure, the basic data including forward data and reverse data, and preprocessing the basic data; S20, constructing a bidirectional prediction data set based on the preprocessed basic data, the bidirectional prediction data set including a forward prediction data set and a reverse prediction data set; S30, constructing a hybrid deep learning model based on the bidirectional prediction data set, training the hybrid deep learning model, and obtaining a bidirectional prediction model; S40, importing the bidirectional prediction model into an industrial computer; S50, connecting the industrial computer to the shield machine PLC, connecting the bidirectional prediction model to the shield machine PLC software program, forward predicting the soil bin pressure value through the forward data, and reversely predicting the soil bin pressure value in real time through the reverse data, combining the forward prediction and the reverse prediction to comprehensively obtain the predicted shield soil bin pressure value, thereby realizing intelligent prediction of the shield soil bin pressure. The prediction method provided by the present invention constructs a hybrid deep learning prediction model and trains a bidirectional prediction model. On the one hand, the soil bin pressure value is predicted from the forward direction through the pre-acquired forward data such as formation parameters, geometric parameters and groundwater parameters. On the other hand, the soil bin pressure value is predicted from the reverse direction in real time through the real-time captured shield parameters as reverse data. A comprehensive prediction value of the soil bin pressure is given by combining the forward prediction and the reverse prediction soil bin pressure value results. The changing trend of the shield soil bin pressure can be captured from the positive accurate prediction, and the soil bin pressure value in the shield area can be predicted efficiently and accurately. In this way, the shield machine main control program can be updated according to the comprehensive prediction value of the soil bin pressure, and the shield machine excavation parameters can be adjusted to ensure the safety of the shield construction.

[0068] The method steps of the present invention are introduced below with a specific embodiment.

[0069] In step S10, basic data for predicting shield soil bin pressure is obtained, the basic data including forward data and reverse data, and the basic data is preprocessed.

[0070] First of all, it should be noted that the basic data in the present invention includes two aspects, namely forward data and reverse data. Among them, the forward data includes the formation parameters, geometric parameters and groundwater parameters of the shield crossing range, and the reverse data includes shield parameters. The formation parameters include geological type, length, proportion and physical and mechanical parameters of different geological soils, geometric parameters include cover thickness and excavation diameter, and groundwater parameters include groundwater type and water level height; shield parameters are shield excavation parameters that affect the shield soil bin pressure, specifically including excavation speed v, cutter head speed w, cutter head torque T, total thrust F and screw conveyor speed n. The forward data are all pre-set, and the reverse data are captured in real time during the shield excavation process. The subsequent soil bin pressure is comprehensively predicted by the forward and reverse parameters. The basic data parameters are shown in Table 1 below.

[0071] Table 1 Basic data parameter table

[0072]

[0073] When using artificial intelligence methods to predict shield construction parameters, sufficient and effective data preprocessing is a key step to ensure model performance. Considering the temporal and spatial characteristics of shield construction information parameters, special processing of shield construction parameter time series and traversal geological information is required to more accurately capture the dynamic changes of soil bin pressure. By normalizing and scaling the data, the original construction data is effectively organized, providing a more reliable input data set for subsequent artificial intelligence models, thereby improving the accuracy and reliability of prediction results at the source of the data.

[0074] In some embodiments, data preprocessing includes the following steps:

[0075] (1) Data normalization:

[0076] In order to eliminate the analytical bias caused by differences in dimensions and units of different evaluation indicators, data normalization is required to ensure the comparability of indicators and make each indicator of the same magnitude for easy comprehensive evaluation, while retaining the original characteristics of the data and optimizing analysis efficiency.

[0077] The predicted data is normalized using the following formula (1):

[0078]

[0079] Among them, x is the normalized basic data, x0 is the initial basic data, and x max is the maximum value of the basic data, x min is the minimum value of the basic data;

[0080] (2) Use feature engineering to process the normalized basic data:

[0081] Feature engineering is a crucial stage in data preprocessing, which aims to improve the performance of machine learning models by constructing and selecting features. Its goal is to optimize the dataset to make it more suitable for model training, thereby improving the accuracy, generalization ability and interpretability of the model. In the overall framework of data preprocessing, feature engineering plays a role in shaping data representation.

[0082] The normalized prediction data is feature scaled to form feature data with multiple categories. The purpose of feature scaling is to adjust the numerical ranges of different features to a similar scale to prevent certain features from having too much influence on the weight of the model. This part is handled by data normalization.

[0083] In some embodiments, feature engineering is used to process the normalized prediction data in the following form:

[0084] (2.1) Perform one-hot encoding on the normalized basic data. One-hot encoding converts feature data with multiple categories into a binary matrix, as shown in the following formula (2):

[0085]

[0086] Among them, Encoded i is the code of the i-th category, and category is the actual category of the sample.

[0087] (2.2) The matrix is ​​regularized, where L1 regularization is used for feature selection, as shown in equation (3), and L2 regularization is used for data denoising, as shown in equation (4):

[0088]

[0089] Among them, λ is the regularization strength, w i is the feature weight after processing.

[0090] (2.3) Perform feature data interaction and use simple mathematical operations to create new features through feature data interaction.

[0091] It should be noted that the interaction of feature data does not generate new feature data, but new features. The new features are generated based on the relationship and combination effect of the original feature data. For example, two sets of feature data are respectively the excavation speed and the cutterhead speed. The interaction of these two sets of feature data (excavation speed ÷ cutterhead speed) can generate a new feature, namely "penetration degree".

[0092] In summary, feature engineering is a key link in data preprocessing. By cleverly selecting, creating, and processing features, it helps to improve model performance, make it better adapt to the complex structure of the data set, and provide stronger support for the training and prediction of machine learning models.

[0093] In step S20, a bidirectional prediction data set is constructed based on the preprocessed basic data. The bidirectional prediction data set includes a forward prediction data set and a reverse prediction data set, and both include at least a training set and a test set.

[0094] In some embodiments, for the forward prediction dataset, based on the preprocessed forward data, the forward prediction dataset is constructed in the following manner:

[0095] First, each segment ring with a width of 1.6 m is used as a unit cell to divide the longitudinal section of the shield tunnel;

[0096] Then, the physical and mechanical parameters under different geological conditions are collected and sorted out, including density, elastic modulus and friction angle;

[0097] Then record the stratigraphic parameters, geometric parameters and groundwater parameters of each unit cell stratum. The stratigraphic parameters include geological type, length, proportion and physical and mechanical parameters of different geological soil bodies. The geometric parameters include cover soil thickness and excavation diameter. The groundwater parameters include groundwater type and water level. See Table 1. The geological types include silty clay, silty fine sand, clay silt, etc. The length and proportion refer to the characteristics of a certain stratum in each unit cell. The stratigraphic length and stratigraphic proportion of each stratum can be combined Figure 2 As can be seen, geological type, length and proportion are used as data sources for integral calculation to calculate the stratum proportion data, that is, the proportion of each stratum in the stratum parameters in the input data of Table 1. In Table 1, the input data includes four parts: stratum parameters, geometric parameters, groundwater parameters, and shield parameters. For the stratum parameters, the input data is the proportion of each stratum, and this proportion data is obtained through integral calculation.

[0098] It should be noted that the formation parameters contain two dimensions. The formation proportion data is only one dimension of the formation parameters, and the other dimension is the physical and mechanical parameters of different soil bodies, as shown in Table 2 below.

[0099] Table 2 Physical and mechanical parameters of different soils

[0100]

[0101] Finally, these parameters will be used as input data features for the hybrid deep learning model to construct a forward prediction dataset based on the collected formation parameters, geometric parameters, and groundwater parameters.

[0102] The goal of the entire forward prediction dataset is to provide a reliable basis for training, validation, and testing deep learning models.

[0103] For the reverse prediction dataset, there is no need to construct it like the forward prediction dataset. The reverse prediction dataset can be directly obtained and will not be elaborated here.

[0104] In addition, both the forward prediction data set and the reverse prediction data set include at least a training set and a test set. The training set is a data set used by the machine learning model for training and learning, and is used to train the parameters of the model. The test set is a data set used to evaluate the final performance of the model. It can also include a validation set, which is a data set used to evaluate the performance of the model, to adjust the parameters of the model during the training process, improve the performance of the model, and avoid overfitting or underfitting of the model. Generally speaking, the training set accounts for a larger proportion, usually 60%-80% of the total data set, and the validation set or test set accounts for a smaller proportion, usually 10%-20% of the total data set.

[0105] In step S30, a hybrid deep learning model is constructed based on the bidirectional prediction data set, and the hybrid deep learning model is trained to obtain a bidirectional prediction model.

[0106] In building a hybrid deep learning model, see Table 1 and combine Figure 2 As shown in the figure, the formation parameters are used as the input data of the model, including: the proportion of silty clay is X1, the proportion of fine sand is X2, and the proportion of clay silt is X3.

[0107] The input data also includes geometric parameters, including cover thickness X4 and excavation diameter X5;

[0108] The input data also include groundwater parameters, including groundwater X6 and confined water X7.

[0109] The input data also includes shield parameters, including excavation speed X8, cutter head speed X9, cutter head torque X 10 , total thrust X 11 , screw conveyor speed X 12 .

[0110] The above input data are input into the model, and the soil bin pressure mean term, upper and lower gradient terms, left and right gradient terms, local heterogeneous terms and upper soil bin pressure value are taken as output data.

[0111] In some embodiments, the output data considers two aspects: instantaneous value prediction and control value prediction. Each aspect includes the soil bin pressure mean item, upper and lower gradient items, left and right gradient items, local heterogeneous items and upper soil bin pressure value. The data parameters are shown in Table 1.

[0112] The instantaneous value includes the mean value of the soil bin pressure P a , upper and lower gradient terms P b , left and right gradient terms P c and the local heterogeneous term P d And the upper soil bin pressure value P1. The instantaneous value is the predicted value of the soil bin pressure during single-ring shield excavation. Single-ring excavation refers to the process of shield excavating a segment ring under specific geological conditions. In this process, the prediction of instantaneous soil bin pressure is very critical to preventing the collapse of the face and maintaining the stability of the soil bin. It can timely reflect the geological changes and abnormal conditions in front of the face, which is of great significance for preventing ground subsidence and controlling the shield posture adjustment.

[0113] The control value includes the mean value of soil bin pressure Up and down gradient terms Left and right gradient terms Local heterogeneity And the upper soil bin pressure value P1 *The control value is the predicted value of the soil bin pressure during multi-ring shield tunneling. The multi-ring tunneling process focuses on the tunneling effect of multiple consecutive segment rings and the long-term control of the soil bin pressure. At this level, the control value prediction of the soil bin pressure is more conducive to the intelligent control of shield parameters and the planning of the construction strategy of the entire tunneling section, ensuring the stability of the soil bin pressure during the construction process and reducing ground settlement and the impact on surrounding buildings.

[0114] The present invention combines the prediction of instantaneous value and control value, and can fully understand the law of soil bin pressure change from both microscopic and macroscopic levels, thereby improving the accuracy and reliability of the overall prediction of soil bin pressure.

[0115] In building the hybrid deep learning model, the hybrid deep learning model is trained using the training set, and the obtained bidirectional prediction model is tested using the test set to obtain the test results. After testing the test set, the performance of the model in the actual application scenario is simulated to verify its generalization ability on unknown data. By performing performance evaluation on the test set, the accuracy, robustness and reliability of the model can be objectively evaluated to ensure that it can achieve good results in practical applications.

[0116] Furthermore, in order to continuously predict the physical characteristics of the soil bin pressure spatial field and meet the input data requirements of the hybrid deep learning model, the preprocessed basic data must be converted into three-dimensional time series data. Therefore, in constructing the hybrid deep learning model, the preprocessed basic data is assigned a time series, the preprocessed basic data is converted into three-dimensional time series data, and a time series data set is constructed.

[0117] Specifically, the sliding window method is used to construct the time series data set, including:

[0118] The data of formation parameters X1, X2, X3, geometric parameters X4, X5, and groundwater parameters X6, X7 from time 1 to time n are used as the first time series input data of the entire bidirectional prediction data set;

[0119] The instantaneous soil bin pressure mean term P a , upper and lower gradient terms P b , left and right gradient terms P c , local heterogeneity term P d and the upper soil bin pressure value P1, and the soil bin pressure mean value of the control value Up and down gradient terms Left and right gradient terms Local heterogeneity And the upper soil bin pressure value P1 * The data from time 1 to time n is used as the first time series output data of the entire bidirectional prediction data set;

[0120] Set the length of each time series of input data to n-5, and the length of each time series of output data to 5, that is, use the data of the first n-5 moments as input data, and the data of the last 5 moments as output data, and slide the window in sequence to obtain input and output data;

[0121] The data from time 2 to n+1 is used as the second time series of the entire bidirectional prediction dataset. The above steps are repeated to construct the time series. Each time series input and output window slides backward by one unit until no new time series can be generated, thus forming a complete bidirectional prediction dataset required by the hybrid deep learning model.

[0122] For details on how to construct time series data, please refer to Figure 3 As shown. Moments can be understood as assigning a time series to each set of data. In this embodiment, the length of each time series is 20, where the length of the input data series is 15 (1-15) and the length of the output data series is 5 (16-20).

[0123] The first time series takes the data from time 1 to 15 as input, and the soil bin pressure characteristics and pressure values ​​corresponding to time 16 to 20 are output; the second series inputs the data from time 2 to 16, and outputs the relevant characteristics and pressure values ​​from time 17 to 21; and so on, each sequence input and output window slides back one unit until all data are covered, forming a complete data set required by the hybrid deep learning model. Based on the above analysis, the prediction model expression can be expressed by formula (5):

[0124]

[0125] Where:

[0126] k - current moment;

[0127] i - spatial distribution characteristics of soil bin pressure field, which are mean term i = a; upper and lower gradient term i = b; left and right gradient term i = c; local heterogeneity term i = d; upper soil bin pressure value i = e;

[0128] h - training time domain;

[0129] p-prediction time domain, and h≥p must be satisfied.

[0130] In the present invention, the training time domain and the prediction time domain are distributed as h=15, p=5; {P i (k+h),P i (k+h+1),L,P i (k+h+p-1)} is the soil bin pressure field value in the p prediction domain; P i (k),P i (k+1),L,P i(k+h-1) is the soil bin pressure field value in the h training domain; X 1i ~X 7i They are the input parameters of h in the training domain, and their specific meanings are shown in Table 1.

[0131] 70% of the generated prediction dataset is used as the training set of the hybrid deep learning model, 20% is used as the validation set of the hybrid deep learning model, and 10% is used as the test set of the hybrid deep learning model.

[0132] After completing the construction of the time series data set, the present invention constructs a hybrid deep learning model consisting of DWT, PSO-GA, and LSTM, such as Figure 4 As shown in the figure, the DWT model is used to reduce the noise of the input data, the PSO-GA model is used to quickly extract the features of the input data, and the LSTM model is used to determine the changing trend of the pressure in the entire closed soil bin.

[0133] (1) About DWT algorithm and data denoising

[0134] The shield construction information parameters have the characteristics of huge data volume and strong continuity. At the same time, the current on-site construction is mainly controlled by manual operation, which inevitably causes abnormal fluctuations in information parameters. Such noisy shield construction information parameters will have a great impact on the performance of prediction results.

[0135] In view of this, the present invention introduces a discrete wavelet transform denoising (DWT) processing method to perform denoising on noisy shield construction information parameter data. This method can overcome the shortcomings of the fixed window of the traditional denoising method, can reduce the noise of the input data, has low requirements on the quality of the basic data, and has the advantages of high efficiency and good stability for processing non-stationary and nonlinear signals. The DWT denoising method is based on multi-level decomposition of the signal. Its essence is to purify useful data information through decomposition and reconstruction, while suppressing the noise part in the data, so as to achieve the purpose of data denoising. The method can be divided into the following three main steps:

[0136] Step 1: Signal decomposition. For a one-dimensional signal x(n), the calculation expression of its DWT is shown in formula (6):

[0137]

[0138] Where:

[0139] ψ a,b (n) - wavelet function;

[0140] a-scale;

[0141] b - translation parameter.

[0142] Through multi-level wavelet transform, the signal is decomposed into approximate coefficient A and detail coefficient D. Among them, the approximate coefficient A is the information obtained by the signal through the low-pass filter during the wavelet transform process, which reflects the general outline or trend of the signal and represents the low-frequency component of the signal, that is, the flat or slowly changing part of the signal, and retains the main information and structure in the signal, which is very important for capturing the basic form of the signal. The detail coefficient D is the information obtained by the signal through the high-pass filter during the wavelet transform process, which reflects the detailed characteristics of the signal and represents the rapidly changing content in the signal. It is very critical for analyzing and recovering the high-frequency information, mutation points and details of the signal.

[0143] Step 2: Noise reduction processing. Threshold processing is performed on the detail coefficient D. The present invention uses a soft threshold processing method. The specific expression of soft threshold processing is shown in formula (7):

[0144] S(x)=sign(x)·(|x|-λ) + (7)

[0145] Where:

[0146] sign(x) - the sign function of x;

[0147] (x) + - positive value function;

[0148] λ - threshold value.

[0149] Step 3: Signal reconstruction. Through inverse wavelet transform, the processed approximate coefficient A and the detail coefficient D processed by threshold are combined into the denoised signal, and its calculation expression is shown in formula (8):

[0150]

[0151] DWT is an ideal tool for signal time-frequency analysis and data denoising. It inherits and develops the positioning idea of ​​short-time Fourier transform (STFT) while overcoming the disadvantage that the window size does not change with frequency. The DWT algorithm called Mallat is used, which is a powerful tool for establishing the relationship between DWT and multi-resolution analysis based on spatial decomposition. Time series signals can be decomposed by the Mallat algorithm, and the calculation formula is shown in equation (9).

[0152] x(t)=A n (t)+D n (t)+D n_1 (t)+L+D1(t) (9)

[0153] Where:

[0154] A n (t) - approximation of the original signal x(t);

[0155] D n (t) - Detailed parameter values ​​related to noise information in the nonlinear decomposition.

[0156] Through multi-resolution, x(t) is decomposed into n (t), D n (t), D n_1 (t)…and D1(t).

[0157] (2) About PSO structural parameter setting

[0158] J.Kennedy and R.Eberthart proposed a bionic optimization method called particle swarm optimization (PSO) in 1995. This is an evolutionary computing technology based on swarm intelligence. The particle swarm algorithm is designed to use individuals called "particles" and groups called "swarms" that are updated from iteration to iteration to search for the global optimal solution. Each particle represents a potential solution in the search space, which adjusts its position and speed according to the local optimum and the global optimum. As a simple and effective random search algorithm, the PSO algorithm can obtain better results than the gradient descent method and the exact penalty function method when solving some highly nonlinear optimization problems.

[0159] In the particle swarm optimization algorithm, the structural parameter setting represents the adjustment of the algorithm's hyperparameters and specific structures to more effectively adapt to specific problems. These parameters play a vital role in improving algorithm performance and convergence. Common structural parameters involve the setting of the number of particles, inertia weight, learning factor, etc.

[0160] The number of particles N represents the number of candidate solutions in the search space. A larger number of particles can improve the ability of global search, but it will also increase the computational cost. In this study, N is set to 100. The inertia weight ω controls the movement speed and direction of the particles, affects the inertia of the particle movement, and is used to balance the global search and local search. Its calculation expression is shown in formula (10):

[0161]

[0162] Where:

[0163] ω(t+1)-inertia weight at iteration (t+1);

[0164] ω max - Maximum value of inertia weight;

[0165] ω min - minimum value of inertia weight;

[0166] t-current iteration number;

[0167] max-iter - Maximum number of iterations.

[0168] In this study, ω max =0.9,ω min =0.2, set to max -iter =2700 times.

[0169] Formula (10) shows that the inertia weight changes linearly with the number of iterations.

[0170] The learning factor is used to control the position update of particles based on individual experience and group experience. The individual learning factor controls the search ability of each particle, and the social learning factor controls the degree to which the particle is affected by group information. Usually c1 represents the individual learning factor, which represents the degree of deviation of the particle from the individual optimal position; c2 represents the social learning factor, which represents the degree of deviation of the particle from the global optimal position. When the maximum number of iterations is reached or the change value of the objective function is less than 0.001, the algorithm converges.

[0171] (3) About GA structural parameter setting

[0172] Genetic algorithm (GA) is an algorithm based on computer simulation of biological systems. It draws inspiration from the biological evolution mechanism in nature and realizes the ability of random global search. The optimization process of this algorithm embodies an efficient and parallel global search method. It seeks the best solution to the problem through the automatic acquisition and accumulation of search space knowledge and adaptive control of the search process. In the genetic algorithm, a fixed-size population consists of multiple individuals, each of which represents a possible solution, and the fitness function is used to evaluate the performance of these individuals. Individuals with poor fitness will be eliminated, and new populations will be formed through operations such as selection, crossover, and mutation. This process is repeated until the optimal solution is obtained. In the genetic algorithm, the adaptability of the algorithm can be optimized by adjusting hyperparameters and specific structural parameters such as population size, crossover rate, mutation rate, etc., which is crucial for algorithm performance and convergence.

[0173] The present invention selects the population size N = 100; the crossover rate is set to P c =0.8, the above parameter settings can ensure sufficient crossover and promote gene mixing. The mutation rate is set to P m =0.2, which can increase the probability of mutation and ensure the diversity of the population. Roulette selection method and multi-point crossover mutation operation are used. When the maximum number of iterations is reached or the change value of the objective function is less than 0.001, the convergence ends.

[0174] (4) About LSTM structure parameter setting

[0175] The LSTM model consists of 4 layers, with 32 neurons in each layer. The model structure is connected to the maximum pooling layer, adding a decay layer and a dense layer, and implementing random and cyclic dropout strategies in the middle layer to reduce overfitting and improve prediction accuracy. The learning rate and the maximum number of cycles are core parameters, and the optimal values ​​are determined through testing. The testing process is as follows:

[0176] In the selection of learning rate test, the prediction of soil bin pressure mean term is taken as an example, and the LSTM model is used for training. The error changes when the learning rate is 0.1, 0.05, 0.01, 0.005, and 0.0001 are studied. The training process is as follows Figure 5 As shown in Table 3, the maximum number of iterations is 2700. The specific test results are shown in Table 3.

[0177] It can be seen from Table 3 that as the learning rate decreases, the model performance shows a trend of first increasing and then decreasing. When the learning rate is set to 0.001, a very poor prediction effect is achieved. When the test learning rate is 0.01, the five performance evaluation indicators all achieve extreme values. Therefore, the learning rate in the present invention is set to 0.01.

[0178] Table 3 Comparison of LSTM neural network learning rate settings

[0179]

[0180]

[0181] In the selection of the maximum number of cycles, the soil bin pressure mean item prediction is also taken as an example. The LSTM model is used for training, and the learning rate is 0.01. The maximum number of cycles of 100, 200, 300, 400, and 500 are selected for testing. The test results are shown in Table 4.

[0182] Table 4 Comparison of the maximum number of cycles of LSTM neural network

[0183]

[0184] As can be seen from Table 4, when the maximum number of cycles changes, the overall error index does not change much. However, the actual training time increases significantly, and when the maximum number of cycles is 300, the performance evaluation index reaches an extreme value. Therefore, the maximum number of cycles is set to 300 in the research of the present invention.

[0185] Through analysis and testing, the parameter settings of the LSTM model were determined, as shown in Table 5.

[0186] Table 5 LSTM hyperparameter settings

[0187]

[0188] It is worth noting that the learning rate update means that the learning rate is updated once the number of cycles reaches 50, and the learning rate in the iteration process is gradually reduced by multiplying it by the learning factor.

[0189] The present invention constructs a hybrid deep learning model composed of DWT, PSO-GA, and LSTM, which can accurately predict the soil bin pressure mean term, upper and lower gradient terms, left and right gradient terms, local heterogeneous terms, and upper soil bin pressure value, which of course includes both instantaneous value and control value, so as to judge the changing trend of the pressure in the entire closed soil bin.

[0190] In step S40, the bidirectional prediction model is imported into the industrial computer.

[0191] See also Figure 6 , Figure 7 Industrial computer, or industrial control computer, is a computer system designed for industrial control. It uses a bus structure and integrates hardware components such as computer motherboard, CPU, hard disk, memory, peripherals and interfaces, and is equipped with an operating system, control network and protocol. It has powerful computing power and a friendly human-computer interface. Industrial computers are widely used in various industries. They are the product of the combination of computer technology and automatic control technology. They are used to achieve precise, reliable and efficient automatic control and data processing functions.

[0192] In this embodiment, the bidirectional prediction model is imported into the industrial computer according to the following steps:

[0193] First, write the DWT-PSO-GA-LSTM hybrid deep learning model (prediction model) code in Matlab software and test it;

[0194] Then, import it into the embedded industrial computer, that is, import the written and tested code into the industrial computer so that it can be integrated into the shield machine edge controller (PLC) system operation control.

[0195] In step S50, the industrial computer is connected to the PLC of the shield machine, and the bidirectional prediction model is connected to the PLC software program of the shield machine. The soil bin pressure value is predicted forwardly through the forward data, and the soil bin pressure value is predicted in real time through the reverse data. The predicted value of the shield soil bin pressure is obtained by combining the forward prediction and the reverse prediction.

[0196] See also Figure 8 , Fig. 9In this embodiment, first, the industrial computer is connected to the shield machine PLC to realize the orderly connection between the algorithm and the shield machine software program. It should be noted here that the shield machine system and the industrial computer program interface need to match. Then, read the data for real-time prediction. This step includes two data sources, one is the soil bin pressure value predicted by the forward prediction of geometric parameters, geological parameters and groundwater parameters; the other is the soil bin pressure value predicted by the reverse real-time prediction of the shield parameters captured in real time. The comprehensive prediction value of the soil bin pressure is given by combining the forward prediction and reverse prediction of the soil bin pressure values. Finally, the control program runs. According to the comprehensive prediction value of the soil bin pressure, the shield machine main control program is updated, the shield machine excavation parameters are adjusted, and finally the intelligent prediction of the shield soil bin pressure is realized.

[0197] The present invention further provides a shield soil bin pressure intelligent prediction system for executing the aforementioned method. The system can be implemented in various ways, including software, hardware, firmware or any combination thereof. One or more units in the system can be implemented using software and / or firmware, such as machine executable instructions stored on a storage medium. In addition to or as an alternative to machine executable instructions, some or all of the units can be at least partially implemented by one or more hardware logic components.

[0198] The following is the result of a specific embodiment of a shield soil bin pressure intelligent prediction method and prediction system of the present invention.

[0199] The present invention carried out field tests in a certain section of a subway line under construction in Beijing, and conducted statistical analysis on the soil bin pressure on rings 1 to 394 of the right-line shield tunnel. The soil bin pressure values ​​in the manual control stage and the intelligent prediction stage were considered respectively, and five evaluation indicators including maximum value, minimum value, extreme value, average value and standard deviation were selected to comprehensively and objectively analyze the advantages of intelligent prediction over traditional manual control. The results are shown in Table 6.

[0200] Table 6 Analysis of intelligent prediction results of soil bin pressure

[0201]

[0202] By analyzing Table 6, it can be seen that compared with the traditional manual control stage, the soil bin pressure control effect is obviously better in the intelligent prediction stage (testing stage and application stage), and the range and standard deviation values ​​are smaller than those of traditional manual control, which meets the actual construction requirements on site.

[0203] It is easy for those skilled in the art to understand that, under the premise of no conflict, the above-mentioned preferred solutions can be freely combined and superimposed.

[0204] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A shield soil bunker pressure intelligent prediction method, characterized in that: The steps include: Acquire basic data for predicting shield soil bin pressure, the basic data including forward data and reverse data, and pre-process the basic data; Based on the preprocessed basic data, construct a bidirectional prediction data set, wherein the bidirectional prediction data set includes a forward prediction data set and a reverse prediction data set; Based on the bidirectional prediction data set, a hybrid deep learning model is constructed, and the hybrid deep learning model is trained to obtain a bidirectional prediction model; Importing the bidirectional prediction model into an industrial computer; The industrial computer is connected to the PLC of the shield machine, and the bidirectional prediction model is connected to the PLC software program of the shield machine. The soil bin pressure value is forward predicted through the forward data, and the soil bin pressure value is reversely predicted in real time through the reverse data. The predicted value of the shield soil bin pressure is comprehensively obtained by combining the forward prediction and the reverse prediction.

2. The prediction method according to claim 1, characterized in that: The forward data include the formation parameters, geometric parameters and groundwater parameters of the shield tunneling range, and the reverse data include the shield tunneling parameters.

3. The prediction method according to claim 1, characterized in that: Preprocessing the basic data includes: (1) The basic data is normalized using the following formula (1): Among them, x is the normalized basic data, x0 is the initial basic data, and x max is the maximum value of the basic data, x min is the minimum value of the basic data; (2) Use feature engineering to process the normalized basic data.

4. The prediction method according to claim 3, characterized in that: Using feature engineering, the normalized basic data is processed including: (2.1) Perform one-hot encoding on the normalized basic data and convert the feature data with multiple categories into a binary matrix representation, as shown in the following formula (2): Among them, Encoded i is the code of the i-th category, and category is the actual category of the sample; (2.2) The matrix is ​​regularized, where L1 regularization is used for feature selection, as shown in equation (3), and L2 regularization is used for data denoising, as shown in equation (4): Among them, λ is the regularization strength, w i is the feature weight after processing; (2.3) Perform feature data interaction and create new features through feature data interaction.

5. The prediction method according to claim 1, characterized in that: Based on the preprocessed forward data, constructing a forward prediction data set includes: Each segment ring is regarded as a unit cell and divided in the longitudinal section of the shield tunnel; Collect and organize physical and mechanical parameters under different geological conditions, including density, elastic modulus and friction angle; Record the formation parameters, geometric parameters and groundwater parameters of each cell, the formation parameters include geological type, length, proportion and physical and mechanical parameters of different geological soil bodies, the geometric parameters include cover soil thickness and excavation diameter, and the groundwater parameters include groundwater type and water level height; A forward prediction dataset is constructed based on the recorded formation parameters, geometric parameters, and groundwater parameters.

6. The prediction method according to claim 2, characterized in that: In constructing the hybrid deep learning model, the formation parameters, geometric parameters, groundwater parameters and shield parameters are used as input data, and the output data includes two aspects: instantaneous value prediction and control value prediction. Each aspect includes the soil bin pressure mean term, upper and lower gradient terms, left and right gradient terms, local heterogeneous terms and upper soil bin pressure value.

7. The prediction method according to claim 6, characterized in that: In constructing the hybrid deep learning model, a time series is assigned to the preprocessed basic data, the preprocessed basic data is converted into three-dimensional time series data, and a time series data set is constructed.

8. The prediction method according to claim 7, characterized in that: The constructing of the time series dataset comprises: The data of formation parameters, geometric parameters, and groundwater parameters from time 1 to n are used as the first time series input data of the entire bidirectional prediction data set; The instantaneous soil bin pressure mean term P a , upper and lower gradient terms P b , left and right gradient terms P c , local heterogeneity term P d and the upper soil bin pressure value P1, and the soil bin pressure mean value of the control value Up and down gradient terms Left and right gradient terms Local heterogeneity And the upper soil bin pressure value P1 * The data from time 1 to time n is used as the first time series output data of the entire bidirectional prediction data set; Set the length of each time series of input data to n-5, and the length of each time series of output data to 5, that is, use the data of the first n-5 moments as input data, and the data of the last 5 moments as output data, and slide the window in sequence to obtain input data and output data; The data from time 2 to n+1 is used as the second time series of the entire bidirectional prediction dataset. The above steps are repeated to construct the time series. Each time series input and output window slides backward by one unit until no new time series can be generated, thus forming a complete bidirectional prediction dataset required by the hybrid deep learning model.

9. The prediction method according to claim 1, characterized in that: The hybrid deep learning model is constructed by a DWT model, a PSO-GA model, and a LSTM model, wherein the DWT model is used to reduce the noise of the input data, the PSO-GA model is used to quickly extract the features of the input data, and the LSTM model is used to determine the changing trend of the pressure in the entire closed soil bin.

10. An intelligent prediction system for shield soil bunker pressure, characterized in that: It comprises an industrial computer, which is connected to the PLC of the shield machine and is used to execute the prediction method according to any one of claims 1 to 9 to predict the pressure value of the shield soil bin.

Images