Curing agent feeding control method of nomadic flow state stabilized soil production line

Abstract

The invention provides a curing agent feeding control method for a nomadic type flow state solidified soil production line, and belongs to the technical field of production control. The curing agent feeding control method comprises the steps that firstly, real-time feeding data of the production line are collected, and a feeding stable component and a feeding fluctuation component are obtained through time sequence singular value decomposition; establishing a fluctuation relation matrix to quantify influence factors, calculating a feeding compensation matrix to perform parameter correction, establishing a preliminary feeding control model by adopting deep learning, generating candidate parameters through Latin hypercube sampling and performing clustering analysis, selecting representative parameters to perform experimental verification, and finally continuously optimizing model parameters by adopting an online learning algorithm. The accurate control of curing agent feeding is realized, so that the quality stability of the solidified soil is ensured, and the technical problem that curing agent feeding parameters of a nomadic type flow state solidified soil production line are difficult to accurately regulate and control in real time in the prior art is solved.

Description

Technical Field

[0001] The invention belongs to the technical field of production control, and in particular relates to a curing agent feeding control method for a nomadic fluidized solidified soil production line. Background Art

[0002] As an important engineering material, fluidized solidified soil is widely used in foundation treatment, roadbed filling, underground engineering and other fields. The nomadic fluidized solidified soil production line is flexible and can use local materials. It can effectively reduce engineering costs and improve construction efficiency. In the production process of fluidized solidified soil, the feeding control of the curing agent directly affects the product quality and engineering performance. Traditional curing agent feeding control methods mainly include quantitative feeding, proportional adjustment, online weighing and other technical means. Quantitative feeding realizes the addition of curing agent by pre-setting the feeding amount, and proportional adjustment controls the amount of curing agent in proportion to the soil material conveying amount. Online weighing uses a weighing sensor to monitor the feeding weight in real time and make adjustments.

[0003] However, in practical applications, these traditional technologies have many shortcomings. First, the properties of soil materials have significant spatial variability. Parameters such as moisture content and density change with the construction location. The traditional fixed parameter feeding method is difficult to adapt to such changes. Secondly, environmental factors such as temperature and humidity affect the activity and fluidity of the curing agent, which in turn affects the feeding accuracy. Thirdly, there are interference factors such as vibration and material accumulation during the operation of the equipment, which cause fluctuations in the feeding system. In addition, fluctuations in the quality of the curing agent itself will also affect the final curing effect. The combined effect of these factors makes it difficult for traditional feeding control methods to achieve precise regulation, and often problems such as feeding amount deviation and uneven mixing occur, affecting the strength and durability of the cured soil.

[0004] At present, in response to the above problems, the industry has adopted a variety of improvement measures, such as adding buffer bins, optimizing the structure of spiral feeders, and adopting multi-point online monitoring. However, these measures still do not fundamentally solve the problem of real-time and precise control of feeding parameters. Especially in a nomadic construction environment, due to the constantly changing working environment and diverse sources of soil materials, the traditional fixed parameter control method is more difficult to meet engineering needs. In other words, there is a technical problem in the prior art that the feeding parameters of the curing agent of the nomadic fluidized solidified soil production line are difficult to accurately control in real time. Summary of the invention

[0005] In view of this, the present invention provides a curing agent feeding control method for a nomadic fluidized solidified soil production line, which can solve the technical problem in the prior art that the curing agent feeding parameters of the nomadic fluidized solidified soil production line are difficult to accurately control in real time.

[0006] The present invention is implemented as follows: The present invention provides a curing agent feeding control method for a nomadic fluidized solidified soil production line, comprising the following steps: collecting real-time feeding data, curing agent quality data, ambient temperature data, and soil moisture content data of a curing agent feeding device in the nomadic fluidized solidified soil production line; performing singular value decomposition on the real-time feeding data to obtain feeding stable component data and feeding fluctuation component data; constructing time series feature vectors of the curing agent quality data, ambient temperature data, soil moisture content data, and the feeding fluctuation component data, and using the least squares method to fit a fluctuation relationship matrix; calculating a feeding compensation coefficient based on the fluctuation relationship matrix and establishing a feeding compensation matrix; correcting the real-time feeding data according to the feeding compensation matrix; collecting the corrected feeding data and the corresponding solidified soil strength data to establish a feeding strength mapping data set; using a deep neural network to train the feeding intensity mapping data set to obtain a feeding control model, and applying the feeding control model to the curing agent feeding device of the nomadic fluidized solidified soil production line for real-time feeding control.

[0007] Among them, the real-time feeding data includes the curing agent feeding weight, feeding rate, and screw feeder speed per unit time; the curing agent quality data includes the curing agent fineness and curing agent activity; the soil moisture content data includes the soil moisture content curve and soil temperature.

[0008] Among them, the feeding trajectory matrix of the real-time feeding data is constructed based on the time series. When the feeding trajectory matrix is ​​subjected to singular value decomposition, the left and right singular vectors corresponding to the largest k singular values ​​are selected to reconstruct the feeding stable component data, wherein the k value is determined by calculating the cumulative contribution rate of the singular values. When the cumulative contribution rate reaches 85%, the k value is determined, and the components corresponding to the remaining singular values ​​constitute the feeding fluctuation component data.

[0009] Among them, the feeding compensation matrix includes a curing agent feeding weight compensation coefficient, a feeding rate compensation coefficient, and a screw feeder speed compensation coefficient. The value range of the curing agent feeding weight compensation coefficient is 0.85 to 1.15, the value range of the feeding rate compensation coefficient is 0.9 to 1.1, and the value range of the screw feeder speed compensation coefficient is 0.95 to 1.05.

[0010] The feed strength mapping data set includes strength values ​​at 3 days, 7 days and 28 days, and the solidified soil strength data includes unconfined compressive strength and elastic modulus of the solidified soil.

[0011] Among them, the deep neural network adopts a multi-layer perceptron structure, including an input layer, 3 hidden layers and an output layer. The number of nodes in the hidden layer is 2 times, 1.5 times and 1 times the input dimension respectively, and the activation function adopts a rectified linear unit function.

[0012] The training of the deep neural network uses a back-propagation algorithm for parameter optimization, the learning rate is set to 0.001, the batch size is set to 32, the number of training rounds is set to 1000 rounds, and the proportion of the validation set is 20%.

[0013] Among them, the Latin hypercube sampling method is used to generate 1000 groups of candidate feeding parameters, and the candidate feeding parameters are input into the preliminary feeding control model to obtain theoretical strength data; the K-means clustering algorithm is used to perform cluster analysis on the theoretical strength data, and the candidate range of the number of clusters is 3 to 10.

[0014] Among them, an online learning algorithm is used to optimize the preliminary feeding control model. The online learning algorithm uses a stochastic gradient descent method to update the model parameters. The initial learning rate is set to 0.01, the learning rate decay coefficient after each iteration is 0.95, the weight coefficient of the historical data is 0.8, and the weight coefficient of the new data is 0.2.

[0015] Wherein, the feeding control model updates the control parameters every 30 seconds.

[0016] Compared with the prior art, the present invention provides a curing agent feeding control method for a nomadic fluidized solidified soil production line. The curing agent feeding control method proposed in the present invention realizes dynamic optimization and precise control of feeding parameters by constructing a feeding data analysis model based on a time series. The method first uses the time series singular value decomposition technology to decompose the feeding data into stable components and fluctuating components, effectively identifying the inherent stability characteristics and external interference factors of the system. By establishing a fluctuation relationship matrix, the degree of effect of each influencing factor on the feeding accuracy is accurately quantified, providing a theoretical basis for parameter compensation.

[0017] On this basis, the present invention uses deep learning technology to establish a feeding control model, which can adaptively learn and update control parameters to effectively cope with the impact of environmental changes and material fluctuations. Through the online learning algorithm, the model can continuously optimize the control strategy based on real-time feedback to ensure that the feeding parameters are always kept in the optimal state. Especially in complex nomadic construction environments, this method shows good adaptability and robustness.

[0018] The present invention solves the problem that it is difficult to accurately control the curing agent feeding parameters in real time in traditional technologies, which is mainly reflected in the following aspects: it realizes adaptive adjustment of the feeding parameters, overcoming the limitations of traditional fixed parameter control; it establishes a complete parameter optimization mechanism to improve control accuracy; it has online learning capabilities and can continuously optimize the control effect. These characteristics make the present invention particularly suitable for the actual application scenarios of nomadic fluidized solidified soil production lines. In summary, the present invention solves the technical problem that it is difficult to accurately control the curing agent feeding parameters of nomadic fluidized solidified soil production lines in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0021] like Figure 1 FIG. 1 is a flow chart of a curing agent feeding control method for a nomadic fluidized solidified soil production line provided by the present invention. The method comprises the following steps:

[0022] S01, collecting real-time feeding data, curing agent quality data, ambient temperature data, and soil moisture data of the curing agent feeding device in the nomadic fluidized solidified soil production line; wherein the real-time feeding data includes the curing agent feeding weight per unit time, feeding rate, and screw feeder speed; the curing agent quality data includes curing agent fineness and curing agent activity; the soil moisture data includes soil moisture curve and soil temperature;

[0023] S02, constructing a feeding trajectory matrix of the real-time feeding data based on the time series, performing singular value decomposition on the feeding trajectory matrix, and obtaining feeding stable component data and feeding fluctuation component data; wherein the feeding stable component data represents the reference feeding parameters of the curing agent feeding process, and the feeding fluctuation component data represents the disturbance parameters of the curing agent feeding process;

[0024] S03, constructing a time series feature vector of the curing agent quality data, ambient temperature data, soil moisture content data and the feeding fluctuation component data, and using the least squares method to fit to obtain a fluctuation relationship matrix; wherein the fluctuation relationship matrix is ​​used to characterize the influence of various parameters on the curing agent feeding fluctuation;

[0025] S04, calculating the feeding compensation coefficient based on the fluctuation relationship matrix, and establishing a feeding compensation matrix; wherein the feeding compensation matrix includes a curing agent feeding weight compensation coefficient, a feeding rate compensation coefficient, and a screw feeder speed compensation coefficient;

[0026] S05, correcting the real-time feeding data according to the feeding compensation matrix to obtain corrected feeding data; wherein the corrected feeding data is equal to the product of the original feeding data and the corresponding compensation coefficient;

[0027] S06, collecting multiple groups of the corrected feeding data and corresponding solidified soil strength data, and establishing a feeding strength mapping data set; wherein the solidified soil strength data includes the unconfined compressive strength and elastic modulus of the solidified soil;

[0028] S07, using a deep neural network to train the feeding intensity mapping data set to obtain a preliminary feeding control model; wherein the deep neural network adopts a multi-layer perceptron structure;

[0029] S08, generating multiple groups of candidate feeding parameters using a Latin hypercube sampling method, inputting the candidate feeding parameters into the preliminary feeding control model, and obtaining theoretical strength data;

[0030] S09, performing cluster analysis on the theoretical strength data using a K-means clustering algorithm to obtain a plurality of cluster centers; wherein the cluster centers represent representative feeding parameter combinations;

[0031] S10, selecting the feeding parameters corresponding to the plurality of cluster centers for experimental verification, and recording the experimental result data; wherein the experimental result data includes the solidified soil strength, solidification time, and solidification uniformity;

[0032] S11, using an online learning algorithm to optimize the preliminary feeding control model according to the experimental result data to obtain an optimized feeding control model; wherein the online learning algorithm uses a stochastic gradient descent method to update model parameters;

[0033] S12. Applying the optimized feeding control model to the curing agent feeding device of the nomadic fluidized solidified soil production line to perform real-time feeding control; wherein the feeding control model updates the control parameters once every preset time.

[0034] The specific implementation of the above steps is described in detail below. The specific implementation of step S01 is to use sensors and data acquisition devices to monitor and collect various parameters of the curing agent feeding device in the nomadic fluidized solidified soil production line in real time. For real-time feeding data, a weight sensor is used to collect the weight of curing agent feeding per second, with a value range of 0 to 100 kilograms per second, and a speed sensor is used to collect the speed data of the screw feeder, with a value range of 0 to 1500 revolutions per minute; for curing agent quality data, a laser particle size analyzer is used to measure the fineness of the curing agent, where the fineness is expressed as specific surface area, with a value range of 300 to 450 square meters per kilogram, and a cement activity meter is used to measure the activity of the curing agent, with a value range of 85% to 98%; for ambient temperature data, a temperature sensor is used to measure the ambient temperature, with a value range of minus 20 to 40 degrees Celsius; for soil moisture content data, a moisture meter is used to measure the moisture content of the soil, with a value range of 15% to 35%, and an infrared thermometer is used to measure the soil temperature, with a value range of 5 to 35 degrees Celsius. All collected data are recorded and stored at a frequency of once per second. The purpose of this step is to obtain key parameter data in the curing agent feeding process and provide basic data support for subsequent data analysis and model construction.

[0035] The specific implementation method of step S02 is to first arrange the collected real-time feeding data in chronological order, construct an m×n-dimensional feeding trajectory matrix, where m represents the number of sampling time points, and n represents the number of feeding parameters, including curing agent feeding weight, feeding rate and screw feeder speed; then perform singular value decomposition on the matrix, decompose it into the product of three matrices, where the singular values ​​of the diagonal matrix are arranged from large to small, and select the left and right singular vectors corresponding to the largest k singular values ​​to reconstruct the feeding stable component data, where the k value is determined by calculating the cumulative contribution rate of the singular value, and the k value is determined when the cumulative contribution rate reaches 85%, and the components corresponding to the remaining singular values ​​constitute the feeding fluctuation component data; the feeding stable component data reflects the main characteristics and trends of the curing agent feeding process, and the feeding fluctuation component data reflects the disturbance and fluctuation in the feeding process. The function of this step is to decompose the feeding data into stable components and fluctuation components, which are convenient for separate analysis and control.

[0036] The specific implementation method of step S03 is to merge the curing agent quality data, environmental temperature data, soil moisture data and feeding fluctuation component data according to the time correspondence to construct a comprehensive feature vector; for the curing agent quality data, it includes two feature dimensions of fineness and activity; for the environmental temperature data, it includes one feature dimension of temperature value; for the soil moisture data, it includes two feature dimensions of moisture content and temperature; for the feeding fluctuation component data, it includes three feature dimensions of feeding weight fluctuation, rate fluctuation and speed fluctuation; the least squares method is used to calculate the correlation coefficient between each feature dimension and the feeding fluctuation component, and construct a p×q-dimensional fluctuation relationship matrix, where p represents the number of influencing factors, q represents the number of feeding parameters, and each element in the matrix represents the contribution of the corresponding influencing factor to the feeding fluctuation. The purpose of this step is to quantify the influence of various factors on feeding fluctuation and provide a basis for subsequent compensation control.

[0037] The specific implementation method of step S04 is to calculate the feeding compensation coefficient based on the fluctuation relationship matrix. First, the weight coefficient of each influencing factor is calculated, and the numerical value in the fluctuation relationship matrix is ​​converted into a weight value between 0 and 1 by a normalization processing method; then, according to the current working state, including parameters such as ambient temperature, curing agent quality and soil moisture content, the corresponding comprehensive influencing factor is calculated; finally, the weight coefficient is multiplied by the comprehensive influencing factor to obtain the feeding compensation coefficient, wherein the curing agent feeding weight compensation coefficient has a value range of 0.85 to 1.15, the feeding rate compensation coefficient has a value range of 0.9 to 1.1, and the screw feeder speed compensation coefficient has a value range of 0.95 to 1.05; these compensation coefficients constitute a compensation matrix for real-time adjustment of feeding parameters. The role of this step is to establish a dynamic compensation mechanism to adapt to changes in various working conditions.

[0038] The specific implementation method of step S05 is to multiply the compensation coefficient in the feeding compensation matrix with the real-time feeding data. For the curing agent feeding weight, the original feeding weight is multiplied by the weight compensation coefficient to obtain the corrected feeding weight; for the feeding rate, the original feeding rate is multiplied by the rate compensation coefficient to obtain the corrected feeding rate; for the screw feeder speed, the original speed is multiplied by the speed compensation coefficient to obtain the corrected speed; the corrected data will be sent to the feeding device as an actual control instruction for execution, and the corrected data will be recorded for subsequent analysis and optimization. The purpose of this step is to realize real-time compensation adjustment of feeding parameters and improve the accuracy and stability of feeding.

[0039] The specific implementation method of step S06 is to collect multiple groups of corrected feeding data during the production process, and at the same time collect corresponding solidified soil strength data, wherein the test of unconfined compressive strength is carried out according to the standard curing age, including the strength values ​​of 3 days, 7 days, and 28 days, and the test of elastic modulus is determined by stress-strain method; the feeding data and strength data are sorted into data pairs according to the corresponding relationship, and a feeding strength mapping data set is constructed, which contains feeding parameters and corresponding strength parameters, and each group of data contains complete feeding process parameters and strength test results. The role of this step is to establish the corresponding relationship between feeding parameters and solidified soil strength, and provide data support for subsequent model training.

[0040] The specific implementation method of step S07 is to use a deep neural network to train the feed strength mapping data set. The multi-layer perceptron structure used includes an input layer, multiple hidden layers and an output layer. The number of nodes in the input layer is equal to the dimension of the feed parameter, and the number of nodes in the output layer is equal to the dimension of the strength parameter; the hidden layer adopts a 3-layer structure, and the number of nodes in each layer is 2 times, 1.5 times and 1 times the input dimension respectively; the activation function adopts a corrected linear unit function; the back propagation algorithm is used for parameter optimization, the learning rate is set to 0.001, the batch size is set to 32, and the number of training rounds is set to 1000 rounds; the early stopping method is used to prevent overfitting during the training process, the proportion of the validation set is 20%, and the training is stopped when the loss function of the validation set does not decrease for 10 consecutive rounds. The purpose of this step is to build a preliminary model that can predict the strength of solidified soil.

[0041] The specific implementation method of step S08 is to use the Latin hypercube sampling method to generate multiple sets of candidate parameters within the feasible domain of the feeding parameters, the number of sampling points is set to 1000, and each sampling point contains a complete feeding parameter combination; these candidate parameters are input into the trained preliminary feeding control model to obtain the corresponding theoretical strength prediction values, including the compressive strength and elastic modulus at different ages; these theoretical strength data will be used for subsequent cluster analysis to find representative feeding parameter combinations. The role of this step is to obtain a large amount of strength data through model prediction, providing a basis for parameter optimization.

[0042] The specific implementation method of step S09 is to use the K-means clustering algorithm to perform cluster analysis on the theoretical intensity data. First, the optimal number of clusters is determined by the silhouette coefficient method, and the candidate range of the number of clusters is 3 to 10; then the intensity data is standardized, and the Euclidean distance is used as the distance metric; the cluster center is iteratively calculated, and the iteration is stopped when the position change of the center point is less than 0.001 or the maximum number of iterations is 100; finally, multiple cluster centers are obtained, each cluster center represents a set of characteristic intensity values, and the corresponding feeding parameter combination can be obtained by reverse search. The purpose of this step is to find a representative feeding parameter combination and reduce the workload of experimental verification.

[0043] The specific implementation method of step S10 is to experimentally verify the representative feeding parameters obtained by cluster analysis, and repeat the test three times for each group of parameters to ensure the reliability of the results; strictly control various parameters during the experiment, record the strength development process of the solidified soil, including the compressive strength at different ages; record the time from feeding to initial setting as the solidification time data; use ultrasonic detection method to measure the wave velocity of the solidified soil specimen to evaluate the uniformity of solidification; compare the experimental results with the theoretical prediction values, calculate the prediction error, and provide a basis for model optimization. The role of this step is to verify the accuracy of the model prediction and obtain real experimental data.

[0044] The specific implementation method of step S11 is to optimize the preliminary feeding control model using an online learning algorithm. First, the experimentally verified data is sorted into training samples in chronological order; the model parameters are updated using the stochastic gradient descent method, and the learning rate adopts an adaptive adjustment strategy. The initial learning rate is set to 0.01, and the learning rate decay coefficient is 0.95 after each iteration; each time a new set of experimental data is received, the parameters are updated once, and the influence of historical data and new data are considered during the update. The weight coefficient of historical data is 0.8, and the weight coefficient of new data is 0.2; through multiple iterations of optimization, the prediction results of the model are closer to the actual situation. The purpose of this step is to improve the prediction accuracy and adaptability of the model.

[0045] The specific implementation method of step S12 is to apply the optimized feeding control model to the actual production process. The model receives real-time working condition parameters as input, including ambient temperature, curing agent quality, soil moisture content, etc.; based on these input parameters, the model calculates the optimal feeding parameter combination, including curing agent feeding weight, feeding rate and screw feeder speed; the control system updates the control parameters every 30 seconds to ensure that the feeding process can adapt to changes in working conditions; at the same time, the control effect is recorded, including feeding accuracy, solidified soil performance and other indicators, to provide a basis for subsequent model maintenance and optimization. The role of this step is to realize intelligent control of the curing agent feeding process and improve production efficiency and product quality.

[0046] The singular value decomposition of the feeding trajectory matrix in step S02 is specifically expressed as follows:

[0047] X=USV T ;

[0048] Where X is an m×n-dimensional feeding trajectory matrix; U is an m×m-dimensional left singular matrix; S is an m×n-dimensional diagonal matrix with singular values ​​on the diagonal; V is an n×n-dimensional right singular matrix; m is the number of sampling time points; n is the number of feeding parameters; and T represents the transpose of the matrix.

[0049] The reconstruction of the stable component data of feeding is specifically expressed as follows:

[0050]

[0051] In the formula, X s is the stable component matrix after reconstruction; k is the number of selected singular values; σ i is the i-th singular value; u i is the i-th column of the left singular matrix; v i is the i-th column of the right singular matrix.

[0052] The calculation of the feeding fluctuation component data is specifically expressed as follows:

[0053] X f =XX s ;

[0054] In the formula, X f is the fluctuation component matrix; X is the original feeding trajectory matrix; X s is the stable component matrix.

[0055] The construction of the time series feature vector in step S03 is specifically expressed as follows:

[0056] F=[α1f1, α2f2, α3f3, α4f4, α5f5];

[0057] Where F is the comprehensive feature vector; f1 is the fineness feature of the curing agent; f2 is the activity feature of the curing agent; f3 is the ambient temperature feature; f4 is the moisture content feature of the soil material; f5 is the temperature feature of the soil material; α1, α2, α3, α4, α5 are the weight coefficients of each feature, and they satisfy

[0058] The least squares fitting of the volatility relationship matrix is ​​specifically expressed as follows:

[0059] R=F T X f ;

[0060] Where R is the fluctuation relationship matrix; F is the comprehensive eigenvector; X f is the fluctuation component matrix.

[0061] The calculation of the feeding compensation coefficient in step S04 is specifically expressed as follows:

[0062]

[0063] In the formula, C w , C v , C sare the curing agent feeding weight compensation coefficient, feeding rate compensation coefficient and screw feeder speed compensation coefficient respectively; β1, β2, β3, β4, β5, β6 are unknown coefficients; r ij is the element in the i-th row and j-th column of the fluctuation relationship matrix; f i is the i-th eigenvalue; p is the feature dimension; γ1, γ2, γ3 are error terms.

[0064] The deep neural network training process in step S07 is specifically expressed as follows:

[0065] h l =σ(W l h l-1 +b l );

[0066] In the formula, h l is the output of the lth layer; W l is the weight matrix of the lth layer; h l-1 is the output of the l-1th layer; b l is the bias term of the lth layer; σ is the rectified linear unit activation function.

[0067] The loss function is specifically expressed as follows:

[0068]

[0069] Where L is the loss function; n is the number of samples; y i is the true value of the i-th sample; is the predicted value of the i-th sample; λ is the regularization coefficient; is the L2 norm of the weight matrix of the lth layer.

[0070] The K-means clustering in step S09 is specifically expressed as follows:

[0071]

[0072] In the formula, J is the clustering objective function; k is the number of clusters; n is the number of samples; x i is the i-th sample point; μ j is the j-th cluster center; ∥·∥ represents the Euclidean distance.

[0073] The online learning update in step S11 is specifically expressed as follows:

[0074]

[0075] Where W t+1 is the updated model parameter; W t is the current model parameter; η t is the learning rate of the tth iteration; L tis the loss function of the tth iteration; Represents the gradient operator.

[0076] The adaptive adjustment of the learning rate is specifically expressed as follows:

[0077] η t =η0λ t ;

[0078] Where η t is the learning rate of the tth iteration; η0 is the initial learning rate, which is 0.01; λ is the learning rate attenuation coefficient, which is 0.95; t is the number of iterations.

[0079] The weight calculation of parameter update is specifically expressed as follows:

[0080] W new =ω h W old +ω n W exp ;

[0081] Where W new is the updated parameter; W old is the historical parameter; W exp is the parameter obtained from the new experimental data; ω h is the historical data weight coefficient, which is 0.8; n is the new data weight coefficient, and its value is 0.2.

[0082] Among them, for the construction of the feeding trajectory matrix X, its matrix form is as follows:

[0083]

[0084] In the formula, w ij represents the weight of curing agent at the i-th time point; v ij represents the feeding rate at the i-th time point; s ij represents the speed of the screw feeder at the i-th time point; m represents the number of sampling time points.

[0085] The specific composition of the comprehensive feature vector F is as follows:

[0086]

[0087] In the formula, f ij It represents the value of the i-th feature at the j-th time point; each column in the matrix corresponds to the fineness of the curing agent, the activity of the curing agent, the ambient temperature, the moisture content of the soil and the temperature of the soil.

[0088] The specific form of the fluctuation relationship matrix R is as follows:

[0089]

[0090] In the formula, r ij Represents the influence coefficient of the i-th feature on the j-th feeding parameter.

[0091] The calculation of the feed compensation coefficient takes into account the coupling effect of multiple influencing factors, among which the β coefficient is obtained through historical data regression analysis. The specific steps are as follows: first, collect at least 100 sets of historical production data, including operating parameters and corresponding optimal compensation coefficients; then use the multivariate linear regression method to solve the β coefficient; finally, confirm the effectiveness of the model through cross-validation.

[0092] The structural design of the deep neural network is based on the following considerations: the number of nodes in the input layer is the same as the dimension of the feeding parameter, and the actual parameter value is used as the input; the hidden layer is designed with a decreasing number of nodes, which is conducive to feature extraction and dimensionality reduction; the number of nodes in the output layer is the same as the dimension of the intensity parameter. l The initialization uses the He initialization method:

[0093]

[0094] Where n l represents the number of nodes in the lth layer; randn represents the standard normal distribution random number generation function.

[0095] The iterative process of the K-means clustering algorithm includes the following steps: 1. Randomly select k initial cluster centers; 2. Calculate the Euclidean distance from each sample point to each cluster center; 3. Divide the sample points into the nearest cluster center; 4. Recalculate the center point of each class; 5. Repeat steps 2 to 4 until convergence. The calculation formula of the Euclidean distance is:

[0096]

[0097] Where, d ij represents the distance from the i-th sample point to the j-th cluster center; x ik represents the k-th eigenvalue of the i-th sample; μ jk Represents the k-th dimension coordinate of the j-th cluster center.

[0098] In online learning algorithms, the loss function L t The design takes into account two aspects: prediction error and model complexity:

[0099] L t =α1L mse +α2L reg ;

[0100] Where, L mse is the mean square error loss term; L regis the regularization term; α1, α2 are weight coefficients, and the optimal value is determined by cross-validation.

[0101]

[0102] This loss function design can prevent model overfitting and improve generalization ability while ensuring prediction accuracy. The regularization coefficient λ is determined by the grid search method with a search range of [0.001, 0.1] and a step size of 0.001.

[0103] The effects brought by the above calculation process include: separating the feeding fluctuation through singular value decomposition to facilitate targeted control; adopting a multi-level compensation mechanism to consider the coupling of various influencing factors; introducing an online learning mechanism to enable the model to continuously adapt to changes in working conditions; and reducing the workload of experimental verification and improving optimization efficiency through cluster analysis. These innovations make this method more adaptable and stable than traditional control methods.

[0104] Specifically, the principle of the present invention is: the technical principle of the present invention is based on theories such as system identification, adaptive control and machine learning. First, the time series singular value decomposition method is used to analyze the feeding data. The core idea of ​​this method is to convert the time series data into a matrix form and extract the main features and noise components in the data through singular value decomposition. In the present invention, this decomposition can effectively distinguish the inherent characteristics and external interference of the feeding system, providing a basis for subsequent parameter compensation.

[0105] Secondly, the establishment of the fluctuation relationship matrix adopts the least squares fitting principle, and an accurate mathematical model is constructed by analyzing the relationship between each influencing factor and the feeding fluctuation. The advantage of this method is that it can quantitatively describe the influence of each factor and avoid the subjectivity of traditional empirical control. The design of the compensation matrix is ​​based on feedback control theory, and the feeding parameters are dynamically adjusted by real-time calculation of the compensation coefficient.

[0106] The application of deep neural networks provides the present invention with powerful learning capabilities. The multi-layer perceptron structure can capture the nonlinear relationship between the feeding parameters and the properties of the solidified soil, while the stochastic gradient descent algorithm ensures that the model can be efficiently learned and updated. The use of the Latin hypercube sampling method improves the coverage efficiency of the parameter space, and K-means clustering helps identify the most representative parameter combinations.

[0107] The introduction of online learning algorithms enables the system to continuously accumulate experience and optimize control strategies. This continuous learning mechanism is particularly suitable for the characteristics of nomadic production lines because it can adapt to changes in the environment and conditions. The entire control process forms a closed-loop system, which achieves precise control of feeding parameters through continuous data collection, analysis, optimization and feedback.

[0108] A specific embodiment 1 of the present invention is provided below, and the specific implementation method of each step in this embodiment 1 is described in detail as follows.

[0109] The specific implementation method of step S01 is to use a variety of sensors and data acquisition devices to monitor and collect various parameters of the curing agent feeding device in the nomadic fluidized solidified soil production line in real time, and collect the curing agent feeding weight data w(t) per unit time through a weighing sensor. The sampling frequency is once per second, and the data acquisition expression is: w(t) = w b (t)+δ w (t), where w b (t) is the reference feed weight, δ w (t) is the weight fluctuation term; the speed data s(t) of the screw feeder is collected by the speed sensor, the sampling frequency is once per second, and the data collection expression is: s(t) = s b (t)+δ s (t), where s b (t) is the reference speed, δ s (t) is the speed fluctuation term; the feeding rate v(t) is calculated by the time derivative of the curing agent feeding weight: For the quality data of the curing agent, a laser particle size analyzer is used to measure the fineness f1(t) of the curing agent, and the particle size distribution is described by the log-normal distribution model: Where x is the particle size, μ and σ are distribution parameters; the activity of the curing agent f2(t) is measured by a cement activity meter; for the ambient temperature data, a temperature sensor is used to measure the ambient temperature f3(t); for the soil moisture content data, a moisture meter is used to measure the soil moisture content f4(t), and an infrared thermometer is used to measure the soil temperature f5(t).

[0110] The specific implementation method of step S02 is to first arrange the collected real-time feeding data in chronological order to construct a feeding trajectory matrix X, whose matrix form is: Where w ij represents the weight of curing agent at the i-th time point, v ij represents the feeding rate at the i-th time point, s ij represents the screw feeder speed at the i-th time point, and m represents the number of sampling time points; then the matrix is ​​subjected to singular value decomposition: X = USV T , where U is an m×m dimensional left singular matrix, S is an m×n dimensional diagonal matrix, the elements on the diagonal are singular values, and V is an n×n dimensional right singular matrix; the left and right singular vectors corresponding to the largest k singular values ​​are selected to reconstruct the stable component data of the feed: Where σ i is the i-th singular value, ui is the i-th column of the left singular matrix, v i is the i-th column of the right singular matrix; the calculation of the feeding fluctuation component data is: X f =XX s .

[0111] The specific implementation of step S03 is to combine the curing agent quality data, the ambient temperature data, the soil moisture content data and the feeding fluctuation component data according to the time correspondence to construct a comprehensive feature vector F: Where f ij Represents the value of the i-th feature at the j-th time point; the least squares method is used to calculate the fluctuation relationship matrix R: R = F T X f , its matrix form is: Where r ij Represents the influence coefficient of the i-th feature on the j-th feeding parameter.

[0112] The specific implementation of step S04 is to calculate the feed compensation coefficient based on the fluctuation relationship matrix, and the calculation formula of the curing agent feed weight compensation coefficient is: The calculation formula of the feeding rate compensation coefficient is: The calculation formula of the screw feeder speed compensation coefficient is: Where β1, β2, β3, β4, β5, β6 are unknown coefficients obtained through historical data regression analysis, r ij is the element in the i-th row and j-th column of the fluctuation relationship matrix, f i is the i-th eigenvalue, p is the characteristic dimension, γ1, γ2, and γ3 are error terms; the calculation of the compensation coefficient takes into account the coupling effect of multiple influencing factors, and its parameters are determined by the multivariate linear regression method.

[0113] The specific implementation method of step S05 is to multiply the compensation coefficient in the feeding compensation matrix with the real-time feeding data, and the corrected curing agent feeding weight calculation formula is: new (t) = C w w(t), the modified feed rate calculation formula is: v new (t) = C v ·v(t), the corrected screw feeder speed calculation formula is: s new (t) = C s ·s(t), where w new (t), v new (t), s new(t) are the corrected feed weight, feed rate and screw feeder speed, respectively; w(t), v(t) and s(t) are the original feed weight, feed rate and screw feeder speed, respectively.

[0114] The specific implementation method of step S06 is to collect multiple sets of corrected feeding data during the production process and correspondingly collect solidified soil strength data, the solidified soil strength data including the unconfined compressive strength σ c (t) and elastic modulus E(t), where the unconfined compressive strength test is carried out according to the standard curing age, including the strength values ​​at 3 days, 7 days, and 28 days. The strength development law is described by the exponential function model: σ c (t) = σ c∞ (1-e -kt ), where σ c∞ is the ultimate strength value, k is the strength development rate coefficient, and t is the age; the elastic modulus is determined by the stress-strain method, and the relationship between the elastic modulus and the compressive strength is described by the power function model: E(t) = ασ c (t) β , where α and β are fitting coefficients; the feeding data and intensity data are sorted into data pairs according to the corresponding relationship, and the feeding intensity mapping data set is constructed.

[0115] The specific implementation method of step S07 is to use a deep neural network to train the feeding intensity mapping data set. The network structure uses a multi-layer perceptron, and the output calculation formula of each layer is: l =σ(W l h l-1 +b l ), where h l is the output of the lth layer, W l is the weight matrix of the lth layer, h l-1 is the output of the l-1th layer, b l is the bias term of the lth layer, σ is the modified linear unit activation function; the weight matrix is ​​initialized using the He initialization method: Where n l represents the number of nodes in the lth layer, randn represents the standard normal distribution random number generation function; the design of the loss function takes into account the prediction error and model complexity: Where n is the number of samples, y i is the true value of the i-th sample, is the predicted value of the i-th sample, λ is the regularization coefficient, is the L2 norm of the weight matrix of the lth layer.

[0116] The specific implementation method of step S08 is to use the Latin hypercube sampling method to generate candidate parameters within the feasible domain of the feeding parameters. The sampling process adopts a stratified random sampling strategy. For the i-th parameter, its sampling interval [a i , b i ] is evenly divided into n subintervals, and a sample point is randomly generated in each subinterval: Where x ij is the sampling value of the i-th parameter in the j-th subinterval, r ij is a random number in the interval [0, 1]; these candidate parameters are input into the trained preliminary feeding control model to obtain the theoretical strength data: y pred =f(x, W, b), where f represents the neural network model, x is the input parameter, and W and b are the weight and bias parameters of the model respectively.

[0117] The specific implementation method of step S09 is to use the K-means clustering algorithm to perform cluster analysis on the theoretical intensity data. First, the Euclidean distance between sample points is calculated: Where d ij represents the distance from the i-th sample point to the j-th cluster center, x ik represents the k-th eigenvalue of the i-th sample, μ jk Represents the k-th dimension coordinate of the j-th cluster center; the clustering objective function is: The objective function value is minimized through iterative optimization; the new cluster center is obtained by calculating the mean of all points in the cluster: Where n j is the number of samples in the jth class.

[0118] The specific implementation method of step S10 is to experimentally verify the representative feeding parameters obtained by cluster analysis. The test method of the solidified soil strength is carried out in accordance with the standard specification, including the preparation, curing and strength test of the specimen; the solidification time is determined by measuring the time from feeding to initial setting, and the determination of the initial setting time is determined by a Vicat instrument: Where t init is the initial setting time, t0 is the start measurement time, Δt is the measurement interval, h, h0, h1 are the measurement height, initial height and end height respectively; the curing uniformity is evaluated by ultrasonic testing method, and the wave velocity calculation formula is: Where l is the length of the specimen and t is the sound wave propagation time.

[0119] The specific implementation method of step S11 is to use an online learning algorithm to optimize the feeding control model. The design of the loss function takes into account the prediction error and model complexity: L t =α1L mse +α2L reg , where L mse is the mean square error loss term: L reg is the regularization term: α1, α2 are weight coefficients; the parameter update adopts the stochastic gradient descent method: Where η t is the learning rate, and its adjustment formula is: t =η0λ t ; The weight calculation formula for parameter update is: W new =ω h W old +ω h W exp .

[0120] The specific implementation method of step S12 is to apply the optimized feeding control model to the actual production process. The control system updates the parameters every 30 seconds, and the calculation of the control output adopts a combination of feedforward and feedback: u(t) = u f (t)+u b (t), where u f (t) is the feedforward control quantity: u f (t) = f(x(t)), f is the neural network model, x(t) is the current working condition parameter; u b (t) is the feedback control quantity: Where e(t) is the control error, k p , k i , k d are the proportional, integral, and differential coefficients respectively.

[0121] In order to better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: When a certain unit is carrying out high-speed railway roadbed construction, it is necessary to process a large amount of fluidized solidified soil, and the curing agent feeding control method of the nomadic fluidized solidified soil production line of the present invention is implemented. First, the production line is equipped with equipment, including data acquisition equipment such as weight sensor, speed sensor, laser particle size analyzer, cement activity meter, temperature sensor, moisture meter and infrared thermometer. The model and measurement range of the data acquisition equipment are shown in Table 1.

[0122]

[0123]

[0124] In the actual production process, the real-time operation data of the curing agent feeding device is first collected. The sampling time is 8 hours, the sampling interval is 1 second, and a total of 28,800 data points are obtained. The original data collected at each time point include the feeding weight w(t), feeding rate v(t) and screw feeder speed s(t). The data collection results are shown in Table 2.

[0125]

[0126] The relationship between the weight of the curing agent and the time is expressed by the following formula: w(t) = w b (t)+δ w (t), where the reference feed weight w b The setting value of (t) is 45 kg / s. The weight fluctuation term δ w The standard deviation of (t) is 0.35 kg / s. The control of the screw feeder speed adopts the following expression: s(t) = s b (t)+δ s (t), where the reference speed s b (t) is set to 850 rpm, and the speed fluctuation term δ s The standard deviation of (t) is 12.3 rpm. The feed rate is calculated by the weight change rate:

[0127] According to the collected data, the feeding trajectory matrix X is constructed, and its specific form is:

[0128] The eigenvalue analysis results obtained by singular value decomposition are shown in Table 3.

[0129] Eigenvalue number Eigenvalue Contribution rate / % Cumulative contribution rate / % 1 2856.5 65.3 65.3 2 428.3 20.3 85.6 3 65.2 14.4 100.0

[0130] The calculated statistical results of the fluctuation component data are shown in Table 4.

[0131] parameter Mean Standard Deviation Maximum Minimum Weight Fluctuation 0.82 0.35 1.86 0.12 Rate Fluctuation 0.15 0.06 0.32 0.02 Speed ​​fluctuation 25.6 12.3 58.9 3.5

[0132] Based on the fluctuation component data and operating parameters, a comprehensive feature vector F is constructed, and the fluctuation relationship matrix R is obtained by least squares fitting. The specific form is:

[0133] The feed compensation coefficient is calculated according to the fluctuation relationship matrix, and the calculation results of the compensation coefficient are shown in Table 5.

[0134] Compensation coefficient Mean Range of variation Standard Deviation Weight compensation 1.025 0.886 to 1.142 0.086 Rate Compensation 0.985 0.912 to 1.065 0.052 Speed ​​compensation 0.992 0.962 to 1.038 0.025

[0135] The compensation coefficient is used to correct the real-time feeding data. The correction process uses the following calculation formula: w new (t) = C w w(t), v new (t) = C v v(t),s new (t) = C s ·s(t). The statistical results of the modified feeding parameters are shown in Table 6.

[0136] parameter Standard deviation before correction Corrected standard deviation Improvement rate / % Feed weight 0.35 0.12 65.7 Feeding rate 0.06 0.025 58.3 Speed 12.3 5.8 52.8

[0137] The strength test was carried out to verify the revised feeding parameters. The test adopted standard curing conditions and the strength development law was described by an exponential function model: σ c (t) = σ c∞ (1-e -kt ), where the ultimate strength value σ c∞ The strength development rate coefficient k is 9.5 MPa and 0.12. The relationship between elastic modulus and compressive strength adopts the power function model: E(t) = ασ c (t) β , where α is 0.56 and β is 0.85. The test results are shown in Table 7.

[0138] Age / day Compressive strength / MPa Elastic modulus / GPa Coefficient of variation / % 3 2.85 1.25 8.5 7 4.62 2.18 7.8 28 8.95 4.35 6.2

[0139] A deep neural network model is constructed based on the experimental data, and the output calculation of each layer of the network is calculated using the following formula: l =σ(W l h l-1 +b l ), where the activation function σ adopts the rectified linear unit function. The network structure parameters are shown in Table 8.

[0140] Network Layer Number of nodes Activation Function Parameter quantity Input Layer 8 none 0 Hidden layer 1 16 ReLU 144 Hidden Layer 2 12 ReLU 204 Hidden layer 3 8 ReLU 104 Output Layer 2 Linear 18

[0141] The model training adopts batch stochastic gradient descent method, and the loss function is: The regularization coefficient λ is 0.001. The model prediction error obtained after 1000 rounds of training is shown in Table 9.

[0142] index Training set error / % Validation set error / % Test set error / % Compressive strength 4.8 5.2 5.5 Elastic modulus 5.2 5.6 5.9

[0143] The Latin hypercube sampling method was used to generate 1000 sets of candidate feeding parameters, and the parameter sampling adopted a stratified random sampling strategy: The corresponding intensity data is obtained through model prediction and analyzed using the K-means clustering algorithm. The clustering objective function is: The cluster analysis results are shown in Table 10.

[0144] Serial number Feed weight Feeding rate Speed Predicted intensity / MPa Measured strength / MPa Relative error / % 1 45.6 2.3 855 8.92 8.85 0.79 2 42.8 2.1 825 8.56 8.42 1.66 3 48.5 2.5 885 9.15 9.08 0.77 4 44.2 2.2 845 8.78 8.65 1.50 5 46.8 2.4 865 8.96 8.88 0.90

[0145] The model is optimized using an online learning algorithm based on experimental data, and the parameters are updated using the stochastic gradient descent method: The adaptive adjustment of learning rate adopts exponential decay strategy: η t =η0λ t , where the initial learning rate η0 is 0.01 and the decay coefficient λ is 0.95. The model parameters are updated using a weighted average method: W new =ωh W old +ω n W exp , where the historical data weight coefficient ω n is 0.8, the new data weight coefficient ω n The optimized model was used in actual production for 3 months, and the production operation data statistics are shown in Table 11.

[0146]

[0147]

[0148] The product quality test results are shown in Table 12.

[0149] Quality indicators Pass rate / % Coefficient of variation / % Stability improvement / % strength 98.5 6.8 35.6 Curing time 97.8 8.2 32.4 Uniformity 98.2 7.5 38.5

[0150] The traditional curing agent feeding control method has the following main problems: 1. The fixed ratio is used for feeding control, which cannot adapt to the changes in the properties of raw materials and environmental conditions, resulting in large fluctuations in product quality. The coefficient of variation of the strength of the cured soil is usually above 15%. 2. The parameters are adjusted based on manual experience, which is not timely enough and is easily affected by subjective factors. The product qualification rate is generally around 85%. 3. It is impossible to achieve real-time optimization of process parameters, resulting in low production efficiency, low raw material utilization, and high production costs.

[0151] The present invention adopts a data-driven intelligent control method, and through real-time monitoring and adaptive adjustment, it has achieved significant progress in the following aspects: 1. A feeding fluctuation separation method based on singular value decomposition has been established, which realizes the accurate identification of the stable component and the fluctuating component of the feeding process, and provides a reliable basis for parameter optimization. Specifically, the standard deviation of the feeding parameters has been reduced by more than 52%, and the feeding accuracy has been significantly improved. 2. A multi-level compensation mechanism has been constructed, which comprehensively considers multiple influencing factors such as the quality of the curing agent, the ambient temperature, and the moisture content of the soil. The calculation of the compensation coefficient is more accurate and comprehensive, so that the coefficient of variation of the product quality is reduced to less than 8%. 3. The method of combining deep neural network and online learning is adopted, and the model has strong adaptive ability and generalization performance, the prediction error is controlled within 6%, and the product qualification rate is increased to more than 98%. 4. The process parameters are optimized by cluster analysis and experimental verification, which reduces the workload of manual experiments, improves the efficiency of parameter optimization, and reduces the number of experiments by about 60%. 5. The intelligent and automated control of the production process is realized, manual operations are reduced, production efficiency is improved, and the utilization rate of raw materials is increased by 15%, reducing production costs.

[0152] In general, the present invention effectively solves the problems existing in the traditional curing agent feeding control method through innovative data processing methods and intelligent control strategies, achieves a comprehensive improvement in product quality, production efficiency and economic benefits, and has important practical significance for improving the automation and intelligence level of fluidized solidified soil production.

[0153] It should be noted that the variables involved in the present invention are explained in detail as shown in Table 13 below.

[0154] Table 13 Variable explanation table

[0155]

[0156]

[0157] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A curing agent feeding control method for a nomadic fluidized solidified soil production line, characterized in that: The following steps are involved: The real-time feeding data, curing agent quality data, ambient temperature data, and soil moisture content data of the curing agent feeding device in the nomadic fluidized solidified soil production line are collected; singular value decomposition is performed on the real-time feeding data to obtain feeding stable component data and feeding fluctuation component data; time series feature vectors of the curing agent quality data, ambient temperature data, soil moisture content data, and the feeding fluctuation component data are constructed, and a fluctuation relationship matrix is ​​obtained by least squares fitting; a feeding compensation coefficient is calculated based on the fluctuation relationship matrix and a feeding compensation matrix is ​​established; the real-time feeding data is corrected according to the feeding compensation matrix; the corrected feeding data and the corresponding solidified soil strength data are collected to establish a feeding strength mapping data set; a deep neural network is used to train the feeding strength mapping data set to obtain a feeding control model, and the feeding control model is applied to the curing agent feeding device of the nomadic fluidized solidified soil production line for real-time feeding control.

2. The curing agent feeding control method of the nomadic fluidized solidified soil production line according to claim 1 is characterized in that: The real-time feeding data includes the feeding weight of the curing agent per unit time, the feeding rate, and the speed of the screw feeder; the curing agent quality data includes the fineness of the curing agent and the activity of the curing agent; the soil moisture content data includes the soil moisture content curve and the soil temperature.

3. The curing agent feeding control method of the nomadic fluidized solidified soil production line according to claim 1 is characterized in that: The feeding trajectory matrix of the real-time feeding data is constructed based on the time series. When the feeding trajectory matrix is ​​subjected to singular value decomposition, the left and right singular vectors corresponding to the largest k singular values ​​are selected to reconstruct the feeding stable component data, wherein the k value is determined by calculating the cumulative contribution rate of the singular values. When the cumulative contribution rate reaches 85%, the k value is determined, and the components corresponding to the remaining singular values ​​constitute the feeding fluctuation component data.

4. The curing agent feeding control method of the nomadic fluidized solidified soil production line according to claim 1 is characterized in that: The feeding compensation matrix includes a curing agent feeding weight compensation coefficient, a feeding rate compensation coefficient, and a screw feeder speed compensation coefficient. The curing agent feeding weight compensation coefficient ranges from 0.85 to 1.15, the feeding rate compensation coefficient ranges from 0.9 to 1.1, and the screw feeder speed compensation coefficient ranges from 0.95 to 1.

05.

5. The curing agent feeding control method of the nomadic fluidized solidified soil production line according to claim 1 is characterized in that: The feed strength mapping data set includes strength values ​​at 3 days, 7 days, and 28 days of age, and the solidified soil strength data includes unconfined compressive strength and elastic modulus of the solidified soil.

6. The curing agent feeding control method of the nomadic fluidized solidified soil production line according to claim 1 is characterized in that: The deep neural network adopts a multi-layer perceptron structure, including an input layer, three hidden layers and an output layer. The number of nodes in the hidden layer is 2 times, 1.5 times and 1 times the input dimension respectively, and the activation function adopts a rectified linear unit function.

7. The curing agent feeding control method of the nomadic fluidized solidified soil production line according to claim 6 is characterized in that: The deep neural network training adopts the back propagation algorithm for parameter optimization, the learning rate is set to 0.001, the batch size is set to 32, the number of training rounds is set to 1000 rounds, and the proportion of the validation set is 20%.

8. The curing agent feeding control method of the nomadic fluidized solidified soil production line according to claim 1 is characterized in that: A Latin hypercube sampling method is used to generate 1000 groups of candidate feeding parameters, and the candidate feeding parameters are input into the preliminary feeding control model to obtain theoretical strength data; a K-means clustering algorithm is used to perform cluster analysis on the theoretical strength data, and the candidate range of the number of clusters is 3 to 10.

9. The curing agent feeding control method of the nomadic fluidized solidified soil production line according to claim 1 is characterized in that: An online learning algorithm is used to optimize the preliminary feeding control model. The online learning algorithm uses a stochastic gradient descent method to update model parameters. The initial learning rate is set to 0.01, the learning rate decay coefficient after each iteration is 0.95, the weight coefficient of the historical data is 0.8, and the weight coefficient of the new data is 0.

2.

10. The curing agent feeding control method of the nomadic fluidized solidified soil production line according to claim 1 is characterized in that: The feeding control model updates the control parameters every 30 seconds.

Images