Intelligent adaptation method for synchronous double slurry of shield

Abstract

The application relates to an intelligent adaptation method for synchronous double slurry of a shield, which comprises the following steps: randomly extracting a plurality of groups of proportioning data in the material proportioning range of the double slurry, and constructing an initial sampling database; performing performance testing on all proportioning schemes in the initial sampling database, and recording the performance indexes of each group of proportioning; systematically surveying and quantitatively grading the construction conditions influencing the proportioning design, and obtaining a construction condition combination; performing correlation analysis on the performance indexes of each group of proportioning and the construction condition combination, establishing a benchmark proportioning model according to the correlation analysis result, combining expert experience and engineering prior knowledge, and performing constraint in combination with the benchmark proportioning model; training a multi-target optimization model, and obtaining a trained multi-target optimization model; inputting a plurality of groups of proportioning parameters into the trained multi-target optimization model, outputting an adaptation result, and realizing intelligent proportioning recommendation. The application realizes intelligent and accurate adaptation of slurry proportioning, and effectively guarantees the durability and environmental protection of the engineering.

Description

Technical Field

[0001] This invention relates to the field of underground tunnel construction technology, and in particular to an intelligent adaptation method for synchronous dual-liquid grouting in shield tunneling. Background Technology

[0002] Shield tunneling, as an advanced mechanized tunnel construction technology, has been widely applied to various underground projects such as urban subways and cross-river / sea tunnels. During shield tunneling, timely and effective filling of the shield tail gap is a crucial step in controlling ground deformation, preventing segment floating, and ensuring the long-term stability of the tunnel structure. Synchronous grouting, as the mainstream technology in tunnel construction, focuses on preparing a grout with good fluidity, pumpability, early strength, later stability, and durability. Cement-water glass two-component grouts have become the mainstream synchronous grouting material for underwater shield tunneling due to their controllable setting time, high early strength, and strong resistance to groundwater dilution. However, when facing complex and variable geological conditions (such as high water pressure and corrosive environments), insufficient durability often arises, easily leading to risks such as excessive long-term surface settlement and segment leakage. The existing technical bottlenecks of underwater shield tunneling synchronous two-component grouting are as follows:

[0003] 1. Current mix design relies heavily on limited experimental data and engineers' experience. There is a lack of systematic correlation between slurry performance (such as durability and environmental friendliness) and complex and variable construction conditions (such as the degree of stratum weathering, groundwater pressure, and tunneling slope), resulting in poor material mix adaptability and difficulty in achieving dynamic optimization.

[0004] 2. Existing technologies primarily focus on the short-term working performance and early mechanical strength of slurries, with insufficient research on their performance degradation under harsh environments such as sulfate corrosion and wet-dry cycles during long-term service. Furthermore, there is a lack of mandatory requirements and systematic control over environmental indicators such as heavy metal leaching, chloride ion content, and pH in slurry materials, which is inconsistent with current green and sustainable engineering construction concepts.

[0005] 3. Faced with massive amounts of material combinations, performance data, and complex field conditions, traditional methods struggle to effectively uncover the intrinsic relationships between the data. There is a lack of an intelligent system that integrates materials science, geotechnical engineering, and artificial intelligence technologies to construct a database of material properties and operating conditions, and to perform intelligent prediction and adaptive recommendations through machine learning algorithms.

[0006] Therefore, developing an intelligent adaptation method for shield tunneling synchronous dual-liquid grout that can be accurately adapted to multi-dimensional construction conditions through an intelligent system has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent adaptation method for synchronous dual-liquid grout in shield tunneling, which aims to solve the problems of insufficient long-term durability of existing grouts and their disconnect from complex and variable construction conditions, while taking into account green and environmental protection requirements and making up for the lack of data-driven intelligent decision-making capabilities.

[0008] To achieve the above objectives, the present invention provides the following solution:

[0009] A smart adaptation method for simultaneous dual-liquid slurry tunneling includes:

[0010] The Latin hypercube sampling method was used to randomly select several sets of mixing ratio data within the material ratio range of the two-liquid slurry to construct an initial sampling database.

[0011] Performance tests were conducted on all the formulation schemes in the initial sampling database, and the performance indicators of each formulation were recorded.

[0012] A systematic investigation and quantitative classification of construction conditions affecting mix design were conducted to obtain combinations of construction conditions.

[0013] The performance indicators of each mix ratio are correlated with the combination of construction conditions. Based on the correlation analysis results, combined with expert experience and prior engineering knowledge, a benchmark mix ratio model applicable to construction conditions is established.

[0014] The benchmark mix design model is used as a constraint, and the multi-objective optimization model is trained according to the material mix parameters and the corresponding construction conditions to obtain the trained multi-objective optimization model. The multi-objective optimization model is constructed based on the integrated model of the enhancement and improvement model and the recurrent neural network.

[0015] Several sets of matching parameters are selected and input into the multi-objective optimization model after training, and the adaptation results are output to achieve intelligent matching recommendation.

[0016] Optionally, the Latin hypercube sampling method may be used to randomly select several sets of proportion data within the material proportion range, including:

[0017] The range of values ​​for each raw material parameter in the two-component slurry is evenly divided into several equal parts. A sample value is randomly selected from each equal part. Several sample values ​​are then randomly paired to obtain several sets of mixing ratio data.

[0018] Optionally, the performance tests include: long-term durability and environmental performance;

[0019] The performance indicators include mass loss rate, compressive strength, compressive strength loss rate, heavy metal leaching concentration, chloride ion content, and pH value.

[0020] Optionally, the systematic investigation and quantitative classification of construction conditions affecting the mix design includes: classifying the degree of rock weathering, groundwater pressure and tunneling slope, and arranging the classification results to obtain the combination of construction conditions.

[0021] Optionally, constraining the model by combining it with the baseline mix design, and training the multi-objective optimization model based on the material mix design parameters and the corresponding construction condition combination, includes:

[0022] S1. Initialize the training sample weight distribution and recurrent neural network parameters of the enhancement model;

[0023] S2. Input several sets of the proportion parameters into the recurrent neural network. The recurrent neural network performs forward propagation calculation according to the time step and finally outputs the probability distribution result of the proportion adaptability under each construction condition.

[0024] S3. Input the probability distribution result into the enhancement model, learn a weak classifier based on the current weight distribution, calculate the classification error rate and weight coefficient, update the weight distribution of the training samples, repeat S3, and finally obtain several weak classifiers. The construction condition combination is used as the true fitting label of the sample, and the classification error rate is calculated based on the true fitting label of the sample.

[0025] S4. Combine several weak classifiers linearly according to their weight coefficients to obtain a combined classifier;

[0026] S5. Based on the combined classifier, output the classification results of suitable or unsuitable matching ratios, and verify the results based on the benchmark matching ratio model.

[0027] Optionally, the parameters for initializing the recurrent neural network include: the input layer-hidden layer weight matrix, the previous time step hidden layer-current hidden layer weight matrix, the hidden layer-output layer weight matrix, and the hidden layer bias vector.

[0028] Optionally, update the training sample weight distribution D. m+1 include:

[0029] D m+1 =(w (m+1), 1,···, w (m+1), i ,···, w (m+1), N );

[0030] ;

[0031] Where m is the preset number of weak classifiers, N is the total number of training samples, and w (m+1),i w represents the weight of the (m+1)th weak classifier for the i-th training sample. m，iZ represents the weights of the m-th weak classifier for the i-th training sample. m It is the normalization factor, α m For the m-th weak classifier The classification error rate, y i To truly fit the labels to the samples, x i Let i be the i-th training sample.

[0032] Optionally, after implementing intelligent mix recommendation, the following can be included: dynamically adjusting the recommended mix ratio by acquiring real-time monitoring data from the construction site.

[0033] The present invention also provides a shield tunneling synchronous dual-liquid slurry, which is composed of liquid A and liquid B. The raw materials and corresponding weight parts of liquid A are: 200-300 parts of cement, 0-100 parts of granulated blast furnace slag powder, 0-50 parts of nano silica, 30-40 parts of bentonite, 3-4 parts of stabilizer, and 800-850 parts of water; liquid B is water glass, with a weight part of 50-60 parts.

[0034] Optionally, the granulated blast furnace slag powder is S95 grade granulated blast furnace slag powder; the cement is PO42.5 cement; the bentonite is high-viscosity sodium-based bentonite for civil engineering; and the nano-silica has a specific surface area of ​​150-400 m². 2 / g, with a silica content greater than 95%; the bentonite is a high-viscosity sodium-based bentonite for civil engineering; the nano-silica has a specific surface area of ​​150-400 m². 2 / g; The water glass is industrial liquid sodium silicate with a modulus of 3.1-3.4.

[0035] The beneficial effects of this invention are as follows: By introducing green and active materials such as granulated blast furnace slag powder and nano-silica into the traditional cement-water glass system, this invention significantly improves the long-term durability of the grout and effectively ensures the long-term service stability of the tunnel structure. By constructing an intelligent database system integrating material proportions, multi-dimensional construction conditions, and long-term performance, and utilizing an integrated algorithm of AdaBoost and Recurrent Neural Networks (RNN), an intelligent mapping and recommendation mechanism from complex working conditions to the optimal proportion is established. This method changes the traditional proportioning mode that relies on experience and trial and error, achieving dynamic and precise adaptation of grout performance to working conditions. This not only significantly improves the reliability and efficiency of grouting projects but also provides key technical support for the digital and intelligent construction of tunnel engineering. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart of a smart adaptation method for synchronous dual-liquid slurry in tunnel boring machines, according to an embodiment of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] like Figure 1 As shown in the figure, this embodiment proposes a smart adaptation method for simultaneous dual-liquid slurry tunneling, including:

[0041] The Latin hypercube sampling method was used to randomly select several sets of mixing ratio data within the material ratio range of the two-liquid slurry to construct an initial sampling database.

[0042] Performance tests were conducted on all the formulation schemes in the initial sampling database, and the performance indicators of each formulation were recorded.

[0043] A systematic investigation and quantitative classification of construction conditions affecting mix design were conducted to obtain combinations of construction conditions.

[0044] The performance indicators of each mix ratio are correlated with the construction conditions. Based on the correlation analysis results, combined with expert experience and prior engineering knowledge, a benchmark mix ratio model for the applicable construction conditions is established.

[0045] The benchmark mix design model is used as a constraint, and the multi-objective optimization model is trained based on the material mix parameters and the corresponding construction conditions to obtain the trained multi-objective optimization model. The multi-objective optimization model is constructed based on the integrated model of the reinforcement and improvement model and the recurrent neural network.

[0046] Several sets of matching parameters are selected and input into the multi-objective optimization model after training, and the adaptation results are output to achieve intelligent matching recommendation.

[0047] Furthermore, the Latin hypercube sampling method was used to randomly select several sets of proportion data within the material proportion range, including:

[0048] The range of values ​​for each raw material parameter in the two-component slurry is evenly divided into several equal parts. A sample value is randomly selected from each equal part. Several sample values ​​are then randomly paired to obtain several sets of mixing ratio data.

[0049] Further performance testing includes: long-term durability and environmental performance;

[0050] Performance indicators include mass loss rate, compressive strength, compressive strength loss rate, heavy metal leaching concentration, chloride ion content, and pH value.

[0051] Specifically, long-term durability performance includes: resistance to long-term sulfate erosion at different ages (28d, 90d, 360d, 720d), resistance to sulfate wet-dry cycles, and performance indicators including mass loss rate, compressive strength, and compressive strength loss rate; environmental performance includes: heavy metal leaching concentration, chloride ion content, and pH value.

[0052] Furthermore, a systematic investigation and quantitative classification of construction conditions affecting the mix design is carried out, including classifying the degree of rock weathering, groundwater pressure and tunneling slope, and then deduplicating and arranging the classification results to obtain the combination of construction conditions.

[0053] Specifically, before conducting multi-dimensional construction condition correlation analysis, it is necessary to systematically investigate and quantitatively classify the construction conditions affecting the mix design, construct a classification system based on rock weathering degree, groundwater pressure, and tunnel boring machine slope, integrate and optimize the three classification systems, and perform deduplication and arrangement to finally establish a multi-dimensional construction condition classification system. The specific classification method is as follows:

[0054] (a) Construction of a rock weathering degree classification system: Based on the exploration data obtained from geological drilling, rock quality index determination, wave velocity testing and point load test, and in accordance with established standards, the rock mass in the tunnel crossing area is classified into the following grades according to the degree of weathering: fully weathered (Grade V), strongly weathered (Grade IV), moderately weathered (Grade III), slightly weathered (Grade II) and unweathered (Grade I).

[0055] (b) Construction of groundwater pressure classification system: Based on hydrogeological exploration, borehole water level observation and seepage test, the static water pressure at the tunnel body is predicted and classified according to the pressure value range: low pressure (<0.1MPa), medium pressure (0.1-0.6MPa), high pressure (0.6-2.0MPa) and ultra-high pressure (>2.0MPa);

[0056] (c) Construction of the tunnel boring machine excavation slope classification system: Based on the longitudinal profile design of the tunnel line and combined with the real-time measurement data of the tunnel boring machine attitude, the excavation slope is divided into: steep downhill (<-3%), gentle downhill (-3%~0%), horizontal slope (approximately 0%), gentle uphill (0%~+3%) and steep uphill (>+3%).

[0057] Furthermore, constraints are applied by combining the baseline mix design model, and the multi-objective optimization model is trained based on the material mix design parameters and corresponding construction conditions, including:

[0058] S1. Initialize the training sample weight distribution and recurrent neural network parameters of the reinforcement boosting model;

[0059] S2. Based on the input of several sets of mix proportion parameters into the recurrent neural network, the recurrent neural network performs forward propagation calculation according to the time step, and finally outputs the probability distribution result of mix proportion adaptability under each construction condition.

[0060] S3. Input the probability distribution results into the enhancement model, learn the weak classifier based on the current weight distribution, calculate the classification error rate and weight coefficients, update the weight distribution of the training samples, repeat S3, and finally obtain several weak classifiers. Among them, the combination of construction conditions is used as the true fitting label of the sample, and the classification error rate is calculated based on the true fitting label of the sample.

[0061] S4. Combine several weak classifiers linearly according to their weight coefficients to obtain a combined classifier;

[0062] S5. Based on the combined classifier, output the classification results of suitable or unsuitable matching ratios, and verify the results according to the benchmark matching ratio model.

[0063] Specifically, model training includes:

[0064] Model training preprocessing: Material ratio parameters are used as input parameters, construction condition combination parameters are used as output parameters, and performance indicators are used as the basis for judging prediction accuracy.

[0065] Intelligent model training: Using the benchmark mix design model as a prior constraint, a multi-objective optimization model is trained through an integrated algorithm combining AdaBoost and Recurrent Neural Network. This establishes a mapping relationship between multi-dimensional construction conditions and performance indicators, enabling intelligent mix design recommendations for the long-term stability of grouting under different construction conditions.

[0066] The ensemble algorithm of AdaBoost combined with Recurrent Neural Network (RNN) is implemented as follows:

[0067] Step 1: Algorithm Input and Parameter Initialization

[0068] a) The input data includes several sets of material proportioning parameters, multi-age durability and environmental performance index data, and multi-dimensional construction condition classification data. The construction condition classification data and performance index data are organized into an RNN input sequence {x} according to time series. t (t=1,2,…,T, where T is the length of the time series data), the material proportioning parameters and corresponding performance indicators constitute the training sample set of AdaBoost;

[0069] b) Initialize the AdaBoost training sample weight distribution D1=(w 1,1 ,···,w 1,i ,···,w 1,N ), w 1,i =1 / N, i=1,2,…,N, where w 1,i =1 / N, where N is the total number of training samples, i=1,2,…,N; Initialize RNN network parameters: Input layer-hidden layer weight matrix U (dimension d) h ×d x d h Let d be the dimension of the hidden layer. x For x t (Feature dimension), previous time step hidden layer - current hidden layer weight matrix W (dimension d) h ×d h Hidden layer-output layer weight matrix V (dimension d) y ×d h d y (output dimension), hidden layer bias vector b h (Dimension d) h ×1), Output layer bias vector b y (Dimension d) y ×1), and set the initial hidden state h0 to d. h A vector of all zeros × 1;

[0070] Step 2: RNN Temporal Feature Extraction Module

[0071] a) Perform RNN forward propagation computation at time steps t=1 to T, and calculate the hidden state h at each time step. t Through formula h t =f(U·x t +W·h t-1 +b h The calculation is performed, where the activation function f(·) is the tanh function, to achieve the fusion of input features and historical time series features;

[0072] b) Output at each time step t Through formula t =g(V·h t +b y The calculation is performed, where the activation function g(·) is the softmax function, which maps the hidden layer output to the probability distribution of the mix ratio adaptability under each construction condition. The result of this probability distribution is used as the input feature of the weak classifier of AdaBoost.

[0073] Step 3: Iteratively train the AdaBoost module:

[0074] a) For iteration numbers M = 1, 2, ..., m (where m is the preset number of weak classifiers), based on the current weight distribution D... m Using the probability distribution features output by the RNN as input, the m-th weak classifier G is learned. m (x): X→{-1,+1}, where X is the feature space, and the output -1 indicates that the matching is not suitable, and +1 indicates that the matching is suitable.

[0075] b) Calculate the m-th weak classifier G m The classification error rate ε of (x) m =Σ(i=1 to N)w m,i ·I(G m (x i )≠y i ), where w m,i Let y be the weight of the m-th weak classifier for the i-th training sample. i For the sample to be truly fitted with labels, I(·) is the indicator function, when G m (x i )≠y i When I(·) = 1, otherwise I(·) = 0;

[0076] c) Calculate the m-th weak classifier G m The weighting coefficient α of (x) m =(1 / 2) ln((1-ε m ) / ε m ), where ln is the natural logarithm, and ε m ∈(0,0.5);

[0077] d) Update the training sample weight distribution D m+1 =(w (m+1), 1,···, w (m+1), i ,···, w (m+1), N The updated formula is: ;

[0078] Among them, Z m Normalization factor , ensure D m+1The probability distribution is represented by m, where m is the preset number of weak classifiers, N is the total number of training samples, and w is the number of training samples. (m+1),i w represents the weight of the (m+1)th weak classifier for the i-th training sample. m，i Let α be the weight of the m-th weak classifier for the i-th training sample. m For the m-th weak classifier G m The classification error rate of (x), y i To truly fit the labels to the samples, x i Let i be the i-th training sample.

[0079] Step 4: Integrated Model Output

[0080] a) Linearly combine the m weak classifiers according to their weight coefficients to obtain the combined classifier G'(x) = Σ (m = 1 to m)α m ·G m (x);

[0081] b) Output the final classification result through the sign function sign(G'(x)), that is, the output of the ensemble algorithm is G(x)=sign(G'(x)). Combined with the benchmark mix ratio model, establish the mapping relationship between multi-dimensional construction conditions, performance indicators and material ratios to achieve intelligent mix ratio recommendation.

[0082] Furthermore, achieving intelligent mix design recommendation includes: dynamically adjusting the recommended mix design based on real-time monitoring data from the construction site.

[0083] Specifically, the method in this embodiment dynamically adjusts the recommended mix ratio based on real-time monitoring data from the construction site using a real-time optimization module. The monitoring data includes, but is not limited to, the amount of segment uplift and the amount of surface subsidence.

[0084] The method in this embodiment also continuously incorporates new construction mix design data and its effect evaluation into the database through an incremental learning module, thereby achieving continuous optimization and updating of the model.

[0085] This embodiment also provides a shield tunneling synchronous dual-liquid slurry, which consists of liquid A and liquid B. The raw materials and corresponding weight parts of liquid A are: 200-300 parts of cement, 0-100 parts of granulated blast furnace slag powder, 0-50 parts of nano silica, 30-40 parts of bentonite, 3-4 parts of stabilizer, and 800-850 parts of water; liquid B is water glass, with a weight part of 50-60 parts.

[0086] Furthermore, the granulated blast furnace slag powder is S95 grade granulated blast furnace slag powder; the cement is PO42.5 cement; the bentonite is high-viscosity sodium-based bentonite for civil engineering; and the nano-silica has a specific surface area of ​​150-400 m². 2 / g; silica content greater than 95%; water glass is industrial liquid sodium silicate with a modulus of 3.1-3.4.

[0087] Example 1:

[0088] This embodiment is applied to a shield tunnel project, traversing strata mainly composed of moderately weathered sandstone with localized strongly weathered interlayers. Groundwater is abundant and contains a small amount of sulfate. The shield tunneling gradient alternates between gentle uphill and horizontal slopes along the route. To address the problems of insufficient durability, difficulty in meeting environmental standards, and poor mix proportion compatibility in synchronous grouting for shield tunneling in this project, an intelligent adaptation method for synchronous dual-liquid grouting is adopted, as detailed below.

[0089] (I) Construction of the sampling database (Latin hypercube sampling):

[0090] Parameter range determination: Define the value range of the 7 input parameters: cement: 200~300 parts; granulated blast furnace slag powder: 0~100 parts; nano silica: 0~50 parts; bentonite: 30~40 parts; stabilizer: 3~4 parts; water: 800~850 parts; water glass: 50~60 parts.

[0091] Interval division and sampling: The value interval of each parameter is evenly divided into 20 equal parts (number of sampling groups = 20 groups, to ensure that the sample covers the entire interval). One sample value is randomly selected in each interval to ensure that each interval is sampled only once. Partial sampling results are shown in Table 1 below:

[0092] Table 1

[0093]

[0094] Sample pairing: The 20 sample values ​​generated by each of the 7 parameters are randomly paired to form 20 sets of non-repeating matching schemes, which constitute the initial sampling database.

[0095] (II) Performance Testing and Feature Extraction (Standardized Indoor Tests):

[0096] All 20 mix design schemes were subjected to standardized tests in accordance with GB / T 50082-2024 "Standard for Test Methods of Long-Term Performance and Durability of Ordinary Concrete", GB / T 30810-2014 "Determination of Leachable Heavy Metals in Cement Mortar", and HJ 557-2010 "Leaching Toxicity of Solid Waste - Horizontal Oscillation Method". The test results are shown in Table 2 (taking core group 3 as an example):

[0097] Table 2

[0098]

[0099] Note: The leaching concentration of heavy metals in all groups met the limit requirements of GB 5085.3-2007 "Identification Standard for Hazardous Waste - Leaching Toxicity Identification", and the pH value met the environmental protection control standard for shield tunneling grouting (8.5~11.0).

[0100] (III) Investigation and Classification of Construction Condition Parameters:

[0101] Rock weathering classification: Thirty rock core samples were obtained through geological drilling. Rock quality index (RQD), wave velocity test and point load test were conducted. According to the "Engineering Rock Mass Classification Standard" GB / T 50218-2014, the rock mass in the tunnel crossing area was determined to be mainly moderately weathered (Class III), with a local 100m section being strongly weathered (Class IV).

[0102] Groundwater pressure classification: Water level observation and seepage test were conducted through 3 hydrogeological boreholes. The static water pressure at the tunnel body was calculated to be 1.2 MPa, which is classified as high pressure according to the classification standard.

[0103] Excavation slope classification: Based on the longitudinal profile design of the tunnel line and combined with the real-time measurement data of the shield machine's attitude (monitored by the laser guidance system), the excavation slope range is determined to be +0.5‰ to +2.8‰, which belongs to the gentle uphill slope (0‰ to +3‰).

[0104] Multi-dimensional working condition combinations: Integrating the above classification results and removing duplicates, the core working condition combinations for this project are formed: Level III weathering + high pressure + gentle slope uphill, and Level IV weathering + high pressure + gentle slope uphill.

[0105] (iv) Multi-dimensional construction condition correlation analysis:

[0106] Pearson correlation analysis was used to correlate the performance data of 20 mix proportions with 2 core operating condition combinations. The results are as follows:

[0107] Condition 1 (Level III weathering + high pressure + gentle slope uphill): It has the highest correlation with the performance data of sampling group 3 (correlation coefficient 0.92). The 720d compressive strength of this group is 3.58 MPa and the mass loss rate is 0.71%, which meets the long-term durability requirements.

[0108] Condition 2 (Level IV weathering + high pressure + gentle slope uphill): It has the highest correlation with sampling group 8 (270 parts cement, 70 parts granulated blast furnace slag powder, 35 parts nano silica, 37 parts bentonite, 3.7 parts stabilizer, 830 parts water, 57 parts water glass) (correlation coefficient 0.89). Its 720d compressive strength is 3.36 MPa and its mass loss rate is 0.85%, which is suitable for the grouting requirements of strongly weathered strata.

[0109] (v) Construction of benchmark proportion parameters:

[0110] Based on the experience of shield tunneling grouting engineering experts and prior knowledge of similar projects, a benchmark mix design model was established for the above two sets of core working conditions:

[0111] Standard mix proportions for working condition 1: 240-260 parts cement, 40-60 parts granulated blast furnace slag powder, 20-30 parts nano silica, 34-36 parts bentonite, 3.4-3.6 parts stabilizer, 815-825 parts water, and 54-56 parts water glass;

[0112] Standard mix proportions for working condition 2: 260-280 parts cement, 60-80 parts slag powder, 36-38 parts bentonite, 3.6-3.8 parts stabilizer, 825-835 parts water, and 56-58 parts water glass;

[0113] The constraints of the baseline ratio are: 720d sulfate erosion resistance mass loss rate ≤1.0%, compressive strength ≥3.2MPa, and heavy metal leaching concentration ≤0.01 mg / L.

[0114] (vi) Model training preprocessing:

[0115] Input parameters: Seven material parameters (cement, granulated blast furnace slag powder, nano silica, bentonite, stabilizer, water, and water glass) from 20 different mix proportions are selected as model inputs, with dimension d. x =7;

[0116] Output parameters: Select two sets of core operating condition combinations as output labels, dimension d y =2;

[0117] Performance constraints: 720d compressive strength, mass loss rate, and heavy metal leaching concentration are used as the criteria for judging the accuracy of prediction. Error thresholds are set as follows: compressive strength prediction error ≤ ±0.15MPa, mass loss rate prediction error ≤ ±0.1%.

[0118] (vii) Intelligent Model Training (AdaBoost + RNN Ensemble Algorithm):

[0119] 1. Algorithm environment and parameter settings:

[0120] Software environment: Python 3.9, TensorFlow 2.8 framework;

[0121] RNN parameters: Hidden layer dimension d h =64, time series length T=4 (corresponding to 4 ages: 28d, 90d, 360d, 720d), input layer-hidden layer weight matrix U (64×7), hidden layer recurrent weight matrix W (64×64), hidden layer-output layer weight matrix V (2×64), bias vector b h (64×1), b y(2×1), activation functions f(·)=tanh, g(·)=softmax;

[0122] AdaBoost parameters: number of weak classifiers m=10 (using decision stumps as weak classifiers), number of training samples N=20, initial weights w 1,i =1 / 20=0.05 (i=1~20).

[0123] 2. Model training steps:

[0124] (1) RNN temporal feature extraction:

[0125] Initialization: The initial hidden state h0 is a 64×1 vector of all zeros;

[0126] Time step calculation (taking sampling group 3 as an example):

[0127] t=1(28d):x1=[250,50,35,17.5,3.5,820,55], h1=tanh (U·x1+W·h0+b h =tanh(64×7 matrix × 7×1 vector + 64×64 matrix × 64×1 zero vector + 64×1 vector) = [-0.12, 0.35, ..., 0.28] (64-dimensional vector), y1 = softmax(V·h1 + b) y = [0.89, 0.11] (The probability of working condition 1 is 0.89, and the probability of working condition 2 is 0.11).

[0128] t=2 (56d): x2=[250,50,35,17.5,3.5,820,55] (material parameters remain unchanged, performance indicators are updated over time), h2=tanh (U·x2+W·h1+b h )=[0.08,0.42,...,0.33], y2=softmax (V·h2+b y = [0.91, 0.09];

[0129] t=3 (360 d): h3=tanh(U·x3+W·h2+b h )=[0.05,0.45,...,0.36], y3=softmax(V·h3+b y = [0.93, 0.07];

[0130] t=4 (720d): h4=tanh(U·x4+W·h3+b h )=[0.03,0.48,...,0.39], y4=softmax(V·h4+b y= [0.95, 0.05];

[0131] (2) AdaBoost iterative training:

[0132] First iteration (M=1):

[0133] Based on the initial weight distribution D1 (all 0.05), the probability features output by the RNN are used as the input features of the weak classifier of AdaBoost to train the weak classifier G1(x). The classification results are: 18 groups are correctly classified and 2 groups are misclassified.

[0134] Classification error rate ε1 = Σ (i = 1 ~ 20) w 1,i ·I(G1(xi)≠y i ) = 0.05 × 2 = 0.1;

[0135] The coefficient of the weak classifier is α1 = (1 / 2)ln((1-0.1) / 0.1) = (1 / 2)ln(9)≈1.0986;

[0136] Update weight distribution D2: w 2,i =(w 1,i ·exp(-α1·y i ·G1(x i ))) / Z1, where Z1=Σ(i=1~20)w 1,i ·exp(-α1·y i ·G1(x i The weights of misclassified samples are updated to 0.05×exp(1.0986) / 0.816≈0.135, and the weights of correctly classified samples are updated to 0.05×exp(-1.0986) / 0.816≈0.022.

[0137] Iterations 2-10: Repeat the above steps. After each iteration, the weights of misclassified samples are increased, and the weights of correctly classified samples are decreased, ultimately resulting in 10 weak classifiers G1(x)~G 10 (x);

[0138] Integrated output: G'(x) = α1G1(x) + α2G2(x) + ... + α 10 G 10 (x), the final classifier G(x)=sign(G'(x)), when G(x)=+1, outputs the fit ratio, and when G(x)=-1, outputs the unfit and not recommended combination.

[0139] (3) Model validation:

[0140] The results of the verification using 5 sets of new mix proportion data (not used in training) showed that, compared with the benchmark mix proportion model established by the experience of shield grouting engineering experts and prior knowledge of similar projects in (5), the compressive strength prediction error was ≤ ±1.2MPa, the mass loss rate prediction error was ≤ ±0.08%, the adaptation accuracy reached 98%, and the preset error threshold was met.

[0141] (viii) Real-time optimization and incremental learning:

[0142] Real-time optimization: The construction site uses an automated monitoring system to collect data on segment uplift (monitoring value 0.8~1.2mm, threshold ≤3mm) and ground settlement (monitoring value 1.0~1.5mm, threshold ≤5mm) in real time. When the segment uplift exceeds 1.2mm, the model automatically adjusts the mix ratio: the amount of cement is increased by 5 parts and the amount of bentonite is increased by 1 part to ensure the early strength of the grout is improved.

[0143] Incremental learning: During the construction process, 8 sets of actual application ratios and performance data were added (such as the actual ratio used under working condition 1: 255 parts cement, 55 parts granulated blast furnace slag powder, 27.5 parts nano silica, 35.5 parts bentonite, 3.5 parts stabilizer, 822 parts water, 55.5 parts water glass, 360-day compressive strength 3.62 MPa, mass loss rate 0.68%). These were incorporated into the database, and the model was retrained, which continuously improved the model's adaptability, and the accuracy of the ratio prediction reached 99% in the later stage.

[0144] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A smart adaptation method for simultaneous dual-liquid slurry in shield tunneling, characterized in that, include: The Latin hypercube sampling method was used to randomly select several sets of mixing ratio data within the material ratio range of the two-liquid slurry to construct an initial sampling database. Performance tests were conducted on all the formulation schemes in the initial sampling database, and the performance indicators of each formulation were recorded. A systematic investigation and quantitative classification of construction conditions affecting mix design were conducted to obtain combinations of construction conditions. The performance indicators of each mix ratio are correlated with the combination of construction conditions. Based on the correlation analysis results, combined with expert experience and prior engineering knowledge, a benchmark mix ratio model applicable to construction conditions is established. The benchmark mix design model is used as a constraint, and the multi-objective optimization model is trained according to the material mix parameters and the corresponding construction conditions to obtain the trained multi-objective optimization model. The multi-objective optimization model is constructed based on the integrated model of the enhancement and improvement model and the recurrent neural network. Several sets of matching parameters are selected and input into the multi-objective optimization model after training, and the adaptation results are output to achieve intelligent matching recommendation.

2. The intelligent adaptation method for synchronous dual-liquid slurry tunneling according to claim 1, characterized in that, The Latin hypercube sampling method was used to randomly select several sets of material proportion data within the material proportion range, including: The range of values ​​for each raw material parameter in the two-component slurry is evenly divided into several equal parts. A sample value is randomly selected from each equal part. Several sample values ​​are then randomly paired to obtain several sets of mixing ratio data.

3. The intelligent adaptation method for synchronous dual-liquid slurry tunneling according to claim 1, characterized in that, The performance tests include: long-term durability and environmental performance; The performance indicators include mass loss rate, compressive strength, compressive strength loss rate, heavy metal leaching concentration, chloride ion content, and pH value.

4. The intelligent adaptation method for synchronous dual-liquid slurry tunneling according to claim 1, characterized in that, The systematic investigation and quantitative classification of construction conditions affecting mix design includes classifying the degree of rock weathering, groundwater pressure and tunneling slope, and then deduplicating and combining the classification results to obtain the combination of construction conditions.

5. The intelligent adaptation method for synchronous dual-liquid slurry tunneling according to claim 1, characterized in that, The process of constraining the model by combining the baseline mix design with the training of the multi-objective optimization model based on the material mix design parameters and the corresponding construction conditions includes: S1. Initialize the training sample weight distribution and recurrent neural network parameters of the enhancement model; S2. Input several sets of the proportion parameters into the recurrent neural network. The recurrent neural network performs forward propagation calculation according to the time step and finally outputs the probability distribution result of the proportion adaptability under each construction condition. S3. Input the probability distribution result into the enhancement model, learn a weak classifier based on the current weight distribution, calculate the classification error rate and weight coefficient, update the weight distribution of the training samples, repeat S3, and finally obtain several weak classifiers. The construction condition combination is used as the true fitting label of the sample, and the classification error rate is calculated based on the true fitting label of the sample. S4. Combine several weak classifiers linearly according to their weight coefficients to obtain a combined classifier; S5. Based on the combined classifier, output the classification results of suitable or unsuitable matching ratios, and verify the results based on the benchmark matching ratio model.

6. The intelligent adaptation method for synchronous dual-liquid slurry tunneling according to claim 5, characterized in that, The parameters for initializing a recurrent neural network include: the input layer-hidden layer weight matrix, the previous time step hidden layer-current hidden layer weight matrix, the hidden layer-output layer weight matrix, and the hidden layer bias vector.

7. The intelligent adaptation method for synchronous dual-liquid slurry tunneling according to claim 5, characterized in that, Update the training sample weight distribution D m+1 include: D m+1 =(w (m+1),1 ,···, w (m+1), i ,···, w (m+1), N ); ； Where m is the preset number of weak classifiers, N is the total number of training samples, and w (m+1),i w represents the weight of the (m+1)th weak classifier for the i-th training sample. m，i Z represents the weights of the m-th weak classifier for the i-th training sample. m It is the normalization factor, α m For the m-th weak classifier The classification error rate, y i To truly fit the labels to the samples, x i Let i be the i-th training sample.

8. The intelligent adaptation method for synchronous dual-liquid slurry tunneling according to claim 1, characterized in that, After implementing intelligent mix design recommendation, the following steps are taken: obtain real-time monitoring data from the construction site and dynamically adjust the recommended mix design.

9. A shield tunneling synchronous dual-liquid slurry applicable to the intelligent adaptation method according to any one of claims 1-8, characterized in that, It consists of liquid A and liquid B. The raw materials and corresponding weight parts of liquid A are: 200-300 parts of cement, 0-100 parts of granulated blast furnace slag powder, 0-50 parts of nano silica, 30-40 parts of bentonite, 3-4 parts of stabilizer, and 800-850 parts of water; liquid B is water glass, with a weight part of 50-60.

10. The shield tunneling synchronous dual-liquid slurry according to claim 9, characterized in that, The granulated blast furnace slag powder is S95 grade granulated blast furnace slag powder; the cement is PO42.5 cement; the bentonite is high-viscosity sodium-based bentonite for civil engineering; the nano-silica has a specific surface area of ​​150-400 m². 2 / g, with a silica content greater than 95%; the water glass is industrial liquid sodium silicate with a modulus of 3.1-3.4.

Images