Method and device for improving strength of flow-state solidified soil based on multivariate analysis

Abstract

The invention relates to the technical field of multivariable analysis, and discloses a method and a device for improving the strength of flow-state solidified soil based on multivariable analysis, and the method comprises the following steps: collecting a material proportion parameter matrix, a stirring parameter matrix and an environment monitoring data matrix in flow-state solidified soil equipment, and executing feature extraction and feature integration; obtaining a target feature vector; performing intensity prediction of the time sequence on the target feature vector to obtain an intensity prediction value of the flow-state solidified soil; sensitivity analysis is carried out based on the intensity predicted value, and an intensity improvement potential index, a parameter adjustment stability index and an energy consumption influence index are obtained; according to the strength improvement potential index, the parameter adjustment stability index and the energy consumption influence index, adaptive parameter optimization is carried out through discount inverse reinforcement learning, and a target control parameter set is obtained. And the strength and the quality uniformity of the flow-state solidified soil are effectively improved while the operation stability is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of multivariate analysis, and particularly to a method and device for improving the strength of fluid-solidified soil based on multivariate analysis. Background Art

[0002] As an important engineering material, the strength prediction and control of fluid-solidified soil are crucial for engineering quality and safety. However, the existing strength prediction methods for fluid-solidified soil have obvious limitations. Usually, they are only based on simple linear or shallow models, and it is difficult to accurately capture the complex non-linear relationship between material properties and control parameters. When facing various changing factors in engineering practice, these methods show problems of poor adaptability and low accuracy. Especially for the precise quantification of the influence of key control parameters such as water-cement ratio, stirring speed, and curing agent dosage on the final strength, the existing methods are difficult to provide sufficient theoretical support and practical guidance.

[0003] Traditional strength prediction and control methods for fluid-solidified soil mainly rely on historical proportioning data, ignoring the significant influence of external factors such as environmental temperature, humidity, and fluctuations in raw material properties on strength indicators. This single data source processing mode leads to the lack of environmental adaptability of the prediction model, unstable performance under different construction conditions, and large prediction deviations. At the same time, existing control optimization methods often regard fluid-solidified soil equipment as a "black box", lacking in-depth understanding and modeling of its internal mechanism, and unable to achieve precise parameter regulation and strength prediction. Summary of the Invention

[0004] The present invention provides a method and device for improving the strength of fluid-solidified soil based on multivariate analysis. The present invention can effectively capture the complex non-linear characteristics of fluid-solidified soil parameters, and while ensuring operation stability, effectively improve the strength and quality uniformity of fluid-solidified soil.

[0005] In a first aspect, the present invention provides a method for improving the strength of fluid-solidified soil based on multivariate analysis. The method for improving the strength of fluid-solidified soil based on multivariate analysis includes: Collect the material proportioning parameter matrix, stirring parameter matrix, and environmental monitoring data matrix in the fluid-solidified soil equipment, and perform feature extraction and feature integration to obtain a target feature vector; Perform strength prediction of the time series on the target feature vector to obtain the strength prediction value of the fluid-solidified soil; Perform sensitivity analysis based on the strength prediction value to obtain a strength improvement potential index, a parameter adjustment stability index, and an energy consumption impact index; According to the strength improvement potential index, the parameter adjustment stability index, and the energy consumption impact index, perform adaptive parameter optimization through discounted inverse reinforcement learning to obtain a set of target control parameters.

[0006] In a second aspect, the present invention provides a device for enhancing the strength of fluid-solidified soil based on multivariate analysis. The device for enhancing the strength of fluid-solidified soil based on multivariate analysis includes: A data acquisition module, configured to acquire a material ratio parameter matrix, a mixing parameter matrix, and an environmental monitoring data matrix in a fluid-solidified soil device, and perform feature extraction and feature integration to obtain a target feature vector; A strength prediction module, configured to perform strength prediction of a time series on the target feature vector to obtain a strength prediction value of the fluid-solidified soil; A sensitivity analysis module, configured to perform sensitivity analysis based on the strength prediction value to obtain a strength enhancement potential index, a parameter adjustment stability index, and an energy consumption impact index; A parameter optimization module, configured to perform adaptive parameter optimization through discounted inverse reinforcement learning according to the strength enhancement potential index, the parameter adjustment stability index, and the energy consumption impact index to obtain a set of target control parameters.

[0007] In the technical solution provided by the present invention, through multi-source data acquisition and preprocessing technology, three types of key data, namely material ratio, mixing parameters, and environmental monitoring, are systematically collected, and the data is precisely processed to provide a comprehensive and reliable data basis for subsequent analysis, effectively solving the problem of single data source in traditional methods. Deep convolutional neural networks are used for feature extraction, and a dedicated network structure is designed for different types of parameters, which can effectively capture the complex non-linear features of fluid-solidified soil parameters, greatly improving the accuracy and comprehensiveness of feature expression. The introduction of low-rank multi-modal fusion and bottleneck transformer structure explores the correlation between different modal features through tensor decomposition and self-attention mechanism, solves the problem of multi-source heterogeneous data fusion, and realizes the efficient modeling of complex relationships between parameters. Combining an improved time series prediction algorithm, the strength values at multiple time scales are predicted simultaneously, and uncertainty estimation is provided, enhancing the reliability and practicality of the prediction results. The sensitivity analysis method based on the Lagrangian function accurately quantifies the impact of control parameter adjustment on strength, and the discounted inverse reinforcement learning mechanism realizes the adaptive optimization of control parameters, effectively improving the final strength and quality uniformity of fluid-solidified soil while ensuring operation stability, and solving the problem of difficult optimization control in traditional methods under unknown model conditions. Description of the Drawings

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0009] Figure 1Schematic diagram of the steps of the method for improving the strength of fluid-solidified soil based on multivariate analysis in the embodiments of the present invention; Figure 2 Schematic diagram of the structure of the device for improving the strength of fluid-solidified soil based on multivariate analysis in the embodiments of the present invention. Detailed implementation manners

[0010] The embodiments of the present invention provide a method and a device for improving the strength of fluid-solidified soil based on multivariate analysis. Terms such as "first", "second", "third", "fourth", etc. (if any) in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "comprising" or "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0011] For ease of understanding, the specific process of the embodiments of the present invention will be described below. Please refer to Figure 1 One embodiment of the method for improving the strength of fluid-solidified soil based on multivariate analysis in the embodiments of the present invention includes: Step S1: Collect the material ratio parameter matrix, mixing parameter matrix and environmental monitoring data matrix in the fluid-solidified soil equipment, and perform feature extraction and feature integration to obtain a target feature vector; It can be understood that the execution subject of the present invention can be a device for improving the strength of fluid-solidified soil based on multivariate analysis, or a terminal or a server. Specifically, it is not limited here. The embodiments of the present invention will be described by taking the server as the execution subject as an example.

[0012] Specifically, collect the material ratio data. The strength of the fluidized solidified soil is related to the material ratio. Key ratio parameters, such as the water-cement ratio, the dosage of the curing agent, the admixture content, and the cement dosage, are collected through material ratio sensors. These material ratio parameters directly affect the chemical reaction and structure formation of the soil, thus affecting its ultimate strength. The sensor forms the original material ratio data set with the collected data, and each sample in it records the actual ratio of various materials. Similarly, collect the mixing parameters, which have a direct impact on the mixing process of the fluidized solidified soil. Through the mixing parameter sensor, key parameters during the mixing process, such as the mixing speed, mixing time, and mixing power, are monitored in real time. These parameters will affect the uniformity of the soil mass and the final curing effect, and thus determine the magnitude of its strength. The collected mixing parameter data forms the original mixing parameter data set, which contains the operation information of each mixing stage. The environmental temperature, relative humidity, and atmospheric pressure, etc., are collected in real time through the environmental monitoring sensor to form the original environmental monitoring data set, recording the real-time data under different environmental conditions. These external conditions will act together with the material ratio and mixing parameters to affect the final curing effect. Interpolate the missing data in the original material ratio data set, the original mixing parameter data set, and the original environmental monitoring data set. For continuous data with no more than three missing points, the cubic spline interpolation method is used for filling, effectively using the surrounding data for smooth filling to maintain the continuity of the data. For the case where the continuous missing points exceed three, a predictive filling algorithm based on the historical data pattern is used to perform intelligent filling according to the existing historical data trend. After the interpolation process, the target data set is obtained. Normalize the target material ratio data set, the target mixing parameter data set, and the target environmental monitoring data set respectively to eliminate the dimensional differences between different parameters, make them have the same weight, and thus avoid some parameters having too much influence on the analysis result due to large dimensions. The normalization process converts all numerical ranges to the interval [0,1], ensuring that parameters with different dimensions have a unified scale, and obtaining the material ratio parameter matrix, the mixing parameter matrix, and the environmental monitoring data matrix. Extract features from the material ratio parameter matrix, the mixing parameter matrix, and the environmental monitoring data matrix respectively, and extract the most representative features for predicting the strength of the fluidized solidified soil from the high-dimensional data. Through deep learning methods such as convolutional neural networks, deep features with strong expression ability are extracted from each data matrix. These deep features reveal the non-linear influence of the material ratio, the mixing process, and the environmental conditions on the strength of the fluidized solidified soil. Integrate the features of the material ratio parameter matrix, the mixing parameter matrix, and the environmental monitoring data matrix to construct the target feature vector.

[0013] The material ratio parameter matrix is input into a three - layer convolutional structure for convolutional operation and ReLU activation. The material ratio parameter matrix contains the ratio data of the fluidized solidified soil at each time point, and these ratios play a decisive role in the strength of the soil. To fully exploit the potential of these data, the material ratio parameter matrix is input into a three - layer convolutional neural network structure for convolutional operation. The convolutional layer captures the local features of the data by using convolutional kernels of different sizes, and at the same time, performs a non - linear transformation on the result of convolution through an activation function (such as ReLU), enabling the network to capture more complex patterns and features. After the convolutional operation, a material ratio feature vector is obtained, representing the deep - level features of the key information in the material ratio parameter matrix. The mixing parameter matrix is input into a two - layer convolutional structure for convolutional operation and Leaky ReLU activation. Mixing parameters, such as mixing speed, mixing time, and mixing power, have a great impact on the strength of the fluidized solidified soil. For the mixing parameter matrix, a two - layer convolutional neural network structure is adopted. The size and stride of the convolutional kernel are still set to adapt to the characteristics of the data, and features are extracted through convolutional operation and the Leaky ReLU activation function. The Leaky ReLU activation function retains some information when dealing with negative numbers, which can avoid the problem of dead neurons, thus ensuring more effective feature extraction. After the convolutional operation, a mixing parameter feature vector is obtained. At the same time, a different strategy is adopted for the feature extraction of the environmental monitoring data matrix because environmental monitoring data (such as temperature, humidity, and atmospheric pressure) has long - term trends and short - term fluctuations. To better capture these features at different scales, the environmental monitoring data matrix is input into a two - channel convolutional structure for processing. The first channel uses a larger convolutional kernel to capture long - term trend features, and the second channel uses a smaller convolutional kernel to capture short - term fluctuation features. After the two channels extract different types of feature information respectively, the features of these two channels are merged through a connection layer to form a comprehensive environmental monitoring feature vector. Feature integration of the material ratio feature vector, mixing parameter feature vector, and environmental monitoring feature vector is performed through low - rank multi - modal fusion and a bottleneck transformer. The low - rank multi - modal fusion technology projects these three feature vectors into a shared feature space through methods such as tensor decomposition to capture the inherent correlation between them. In this process, the computational complexity is effectively reduced in a low - rank manner while maintaining the interaction information between multi - modal features. The fused feature vector is input into the bottleneck transformer for processing. The bottleneck transformer is a deep - learning model based on the self - attention mechanism. Through this model, the long - range dependence relationship between features is effectively captured, and the input features are weighted and combined to further strengthen the feature representation. Through these operations, the target feature vector is finally obtained.

[0014] The material ratio feature vector, stirring parameter feature vector, and environmental monitoring feature vector are respectively mapped through a linear projection matrix. The linear projection operation maps each feature vector into a common feature space, thereby realizing the unified representation of different modality features. These projection operations respectively obtain the material ratio projection vector, stirring parameter projection vector, and environmental monitoring projection vector. Through this mapping, the features originally from different data sources are transformed into vectors in the shared feature space, and then can be effectively fused. Based on the material ratio projection vector, stirring parameter projection vector, and environmental monitoring projection vector, three modality interaction tensors are constructed. The modality interaction tensor captures the complex interaction between the material ratio, stirring process, and environmental conditions by modeling the mutual relationship between the projection vectors. By constructing the modality interaction tensor, the potential connections between different data sources are fully explored to ensure that various types of information can act together in the high-dimensional space. For the three modality interaction tensors, Tucker decomposition processing is performed. Tucker decomposition is an efficient tensor decomposition method that represents a complex high-dimensional tensor as the product of multiple low-dimensional factors. Through Tucker decomposition, the complexity of the tensor is effectively reduced, and important feature information can be retained. In this process, a low-rank representation tensor is extracted to capture the non-linear interaction relationship between different modality features. The core tensor in the low-rank representation tensor is reconstructed into a vector form to obtain the fused feature vector. The fused feature vector is input into a bottleneck transformer for feature integration. The bottleneck transformer is a model based on the self-attention mechanism and has strong feature modeling capabilities. The bottleneck transformer contains four encoder layers, and each encoder layer includes an 8-head self-attention mechanism and a feed-forward neural network. The self-attention mechanism can effectively calculate the weight relationship between different features, enabling the model to capture long-range feature dependencies. Through this mechanism, the bottleneck transformer automatically adjusts the influence weights between various features, thereby more accurately fusing multiple feature information and further improving the performance of the model. The last layer of the bottleneck transformer further compresses the feature representation by reducing the feature dimension to half of the original dimension in the middle layer, forcing the model to learn a more compact feature representation. This bottleneck design enables the model to retain the most important information during the dimension reduction process, avoid overfitting, and improve the generalization ability of the model. Through this integration process, the target feature vector is finally obtained.

[0015] Step S2: Perform the intensity prediction of the time series on the target feature vector to obtain the strength prediction value of the fluid-state solidified soil; Specifically, the target feature vector is input into a neural network structure composed of a three-layer bidirectional long short-term memory network for temporal feature extraction. The bidirectional long short-term memory network is a deep learning model suitable for processing time series data, capable of capturing both forward and backward dependencies in time series data. Each layer of bidirectional LSTM cells learns from both the forward and backward directions of the time series, enabling the network to more comprehensively understand the long-term dependencies in the data. At each layer of the network, the LSTM cells maintain long-term dependency information through a memory mechanism and effectively filter out key information by combining gating mechanisms, thereby improving the model's ability to model complex time series data. In this way, the input target feature vector, after being processed by the three-layer bidirectional LSTM network, generates output features containing rich temporal information, which represent the strength change trend of the fluid-solidified soil. Analyze the time dependence relationship of the target feature vector through the time convolution module. The core function of the time convolution module is to capture the dependencies at different time scales in time series data. The time convolution module uses multiple one-dimensional convolutional kernels, which are respectively used to capture short-term, medium-term, and long-term time dependencies. By using convolutional kernels of different sizes, the module captures features at different time scales in the time series, ensuring that the model can fully understand short-term fluctuations and long-term trends. At the same time, to enhance the model's attention to key features, the time convolution features are weighted and fused through an attention mechanism. The attention mechanism calculates the importance weights of the features at each time step, enabling the network to automatically focus on the time points that have the most impact on the final prediction. Perform a residual connection on the output features of the bidirectional long short-term memory network and the fused time convolution features to alleviate the problem of gradient disappearance or gradient explosion in deep networks and help the model better learn the relationship between low-level and high-level features. By directly adding the output features of the bidirectional LSTM and the fused time convolution features to form residual features, the model can better combine the two types of feature information while retaining the independence of each type of feature. Input the residual features into a multi-task learning architecture for final strength prediction. The multi-task learning architecture simultaneously solves multiple related prediction tasks through a shared encoder and multiple task-specific decoder structures. In this architecture, the shared encoder extracts general temporal features from the input residual features, while the three task-specific decoders perform strength predictions for different time scales (7 days, 14 days, and 28 days) respectively. This multi-task learning method can improve the training efficiency and prediction ability of the model by sharing information, and at the same time enable the model to learn from each other among multiple tasks, enhancing its generalization ability. For each task-specific decoder, the model is trained by optimizing the mean squared error loss function corresponding to the time scale, and finally obtains the strength prediction values of the fluid-solidified soil at three different time points (7 days, 14 days, and 28 days). In the multi-task learning framework, an uncertainty estimation mechanism is introduced to provide corresponding uncertainty estimates for each strength prediction value.This uncertainty estimation helps to evaluate the reliability of the prediction results and provides a reference for decision-making in practical applications. Especially when dealing with the strength prediction of fluid-solidified soil with high variability, uncertainty estimation is particularly important.

[0016] Step S3: Conduct a sensitivity analysis based on the strength prediction value to obtain the strength improvement potential index, parameter adjustment stability index, and energy consumption impact index. Specifically, the control optimization problem of fluidized soil solidification equipment is modeled as a collaborative solution framework of the main problem and sub-problems, where the goal of the main problem is to maximize the strength function of fluidized soil solidification. This objective function depends on multiple control parameters, such as water-cement ratio, mixing speed, and curing agent dosage. These parameters directly affect the strength of the soil. The sub-problems are used to evaluate the expected strength and sensitivity of each control parameter after adjustment, that is, by fine-tuning each parameter, evaluate its influence on the strength of fluidized soil solidification and its changing trend. The main problem and sub-problems work together to ensure that the optimization process effectively finds the optimal control strategy for maximizing strength. After the model framework is determined, vector conversion is performed based on the strength prediction value of fluidized soil solidification and the corresponding uncertainty estimate value, and these values ​​are converted into control parameter vectors. The vector contains the current state and uncertainty of each control parameter. The Lagrangian function is constructed to describe the relationship between the control parameters and the strength of fluidized soil solidification. The Lagrangian function is in the form of the difference between the maximum strength function of fluidized soil solidification and a series of constraints, which include the physical and technical limitations of each control parameter. In this way, the effective constraints on the range of control parameters are maintained during the optimization process, and the practical feasibility of the optimization process is ensured. The second-order derivative of the Lagrangian function with respect to the control parameters is obtained to obtain the first parameter sensitivity matrix that reflects the second-order sensitivity of the control parameter changes to the strength. This matrix provides detailed information on the interaction between the control parameters and their effects on the strength, especially the second-order sensitivity of the strength changes when the control parameters change slightly. By calculating the sensitivity matrix, it is possible to identify which control parameters are most sensitive to the change in strength, thereby providing guidance for subsequent optimization. In order to improve the accuracy and efficiency of the optimization, the first parameter sensitivity matrix is ​​iteratively optimized so that the sensitivity matrix can be continuously adjusted and improved in multiple calculations, making each round of calculation more accurate and stable. After iterative optimization, the second parameter sensitivity matrix is ​​obtained. This matrix further refines the sensitivity evaluation of the control parameters and can identify the most sensitive direction of parameter adjustment by calculating its eigenvalues ​​and eigenvectors. The eigenvector with the largest eigenvalue represents the most sensitive direction in the parameter adjustment process, that is, adjusting the control parameters along this direction will have the greatest impact on the strength of the fluidized solidified soil. By calculating these eigenvectors, a set of parameter sensitive directions is obtained, which helps determine which control parameters need to be adjusted in order to achieve the best strength improvement effect. Based on the parameter sensitive direction set, a pre-evaluation index system for control parameter optimization is constructed, which includes three key indicators: strength improvement potential index, parameter adjustment stability index and energy consumption impact index. The strength improvement potential index quantifies the optimization potential of each parameter by calculating the expected improvement effect of parameter adjustment on strength. The parameter adjustment stability index evaluates the robustness of different control parameter adjustments. The smaller the stability index, the more stable the adjustment scheme is under different conditions.The energy consumption impact index is used to quantify the impact of control parameter adjustment on the system energy consumption, ensuring that not only the intensity is improved during the optimization process, but also the energy efficiency factor is considered to reduce energy consumption. By comprehensively evaluating these indicators, a pre-evaluation result of parameter adjustment is formed.

[0017] Step S4: According to the intensity improvement potential index, the parameter adjustment stability index, and the energy consumption impact index, adaptive parameter optimization is performed through discounted inverse reinforcement learning to obtain a set of target control parameters.

[0018] Specifically, based on the strength improvement potential index, parameter adjustment stability index, and energy consumption impact index, the state space and action space are constructed. In the state space, it includes the current physical state variables and environmental condition variables of the fluid-solidified soil. The physical state variables include factors such as the current strength, temperature, and humidity of the fluid-solidified soil, while the environmental condition variables include changes in external environments such as atmospheric pressure and relative humidity. These state variables affect the solidification process of the soil mass and the final strength performance, so they are considered in the state space. The action space includes key control parameters such as the water-cement ratio, mixing speed, and curing agent dosage. These control parameters directly determine the solidification characteristics and final strength of the soil mass. In this framework, the state-action pair set forms the basic element in reinforcement learning. Each state corresponds to one or more possible actions, and the characteristics of the fluid-solidified soil are adjusted through these actions. A state transition model is established through the state-action pair set. The state transition model is used to describe the process of the system transitioning from one state to another after taking a certain action. This model captures the mutual relationships and change rules among various factors in the system through the simulation of the fluid-solidified soil equipment and environmental conditions. The state transition model provides the dynamic change rules of the system for reinforcement learning, ensuring that the learning algorithm can take appropriate actions in different states. A parameterized reward function is constructed. The reward function reflects the contribution degree of each action to the change of the system state. In this function, factors such as strength improvement potential, stability, and energy consumption are comprehensively considered to ensure that each action can move towards the optimal goal. To solve the parameterized reward function, a method of minimizing the feature expectation difference between the expert trajectory and the optimal policy trajectory is adopted. The expert trajectory represents the decision-making process taken under known expert experience, while the optimal policy trajectory is the best decision-making path obtained through reinforcement learning. By comparing the differences between the two, the weights in the reward function are continuously adjusted to minimize their feature expectation difference. This process helps the model better simulate the behavior of experts and gradually adjusts the learning strategy, enabling the reinforcement learning algorithm to converge to the best control strategy during the optimization process. Through this step, an optimized weight vector is obtained, representing the relative importance of each factor in the optimal control strategy. Based on the optimized weight vector, the discounted inverse reinforcement learning algorithm is executed to optimize the control strategy. In discounted inverse reinforcement learning, a discount factor is introduced to balance the relationship between short-term rewards and long-term rewards in the multi-step decision-making process. Through this process, the algorithm comprehensively considers the influence between the current decision and future rewards, thereby optimizing the overall control strategy. To ensure the smoothness and stability of the obtained control strategy, a policy regularization term is introduced into the optimized control strategy. The role of the regularization term is to avoid excessive fluctuations during the training process by restricting the change amplitude of the policy, ensuring that the finally obtained control strategy can smoothly adjust each control parameter. Through this smoothing process, the finally obtained control strategy is more stable and adapts to the requirements under different environmental conditions.Through the discounted inverse reinforcement learning algorithm, a set of target control parameters is finally obtained. These parameter sets are the optimization results for different scenarios and can achieve the best effect of improving the strength of fluid-solidified soil in practical applications. The adjustment of each control parameter (such as water-cement ratio, mixing speed, and curing agent dosage) has been optimized to ensure the best balance among strength, stability, and energy efficiency.

[0019] In the embodiment of the present invention, through the multi-source data acquisition and preprocessing technology, three types of key data, namely material ratio, mixing parameters, and environmental monitoring, are systematically collected, and the data is accurately processed to provide a comprehensive and reliable data basis for subsequent analysis, effectively solving the problem of single data source in traditional methods. The deep convolutional neural network is used for feature extraction, and a special network structure is designed for different types of parameters, which can effectively capture the complex non-linear features of fluid-solidified soil parameters and greatly improve the accuracy and comprehensiveness of feature expression. The low-rank multi-modal fusion and bottleneck transformer structure are introduced. By tensor decomposition and self-attention mechanism, the correlation between different modal features is explored, solving the problem of multi-source heterogeneous data fusion and realizing the efficient modeling of complex relationships between parameters. Combining with the improved time series prediction algorithm, the strength values at multiple time scales are predicted simultaneously, and uncertainty estimation is provided, enhancing the reliability and practicality of the prediction results. Based on the sensitivity analysis method of the Lagrangian function, the influence of control parameter adjustment on strength is accurately quantified. The discounted inverse reinforcement learning mechanism realizes the adaptive optimization of control parameters, effectively improving the final strength and quality uniformity of fluid-solidified soil while ensuring operation stability, and solving the problem that it is difficult to optimize control under unknown model conditions in traditional methods.

[0020] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Collect the water-cement ratio, curing agent dosage, admixture dosage, and cement dosage through the material ratio sensor to obtain the original material ratio data set; Collect the mixing speed, mixing time, and mixing power through the mixing parameter sensor to obtain the original mixing parameter data set; Collect the environmental temperature, relative humidity, and atmospheric pressure through the environmental monitoring sensor to obtain the original environmental monitoring data set; Interpolate the missing data in the original material ratio data set, original mixing parameter data set, and original environmental monitoring data set to obtain the target material ratio data set, target mixing parameter data set, and target environmental monitoring data set; Normalize the target material ratio data set, target mixing parameter data set, and target environmental monitoring data set respectively to obtain the material ratio parameter matrix, mixing parameter matrix, and environmental monitoring data matrix; Feature extraction and feature integration are respectively performed on the material ratio parameter matrix, the mixing parameter matrix, and the environmental monitoring data matrix to obtain the target feature vector.

[0021] Specifically, the water-cement ratio, the dosage of curing agent, the admixture content, and the cement dosage are collected through the material ratio sensor to obtain the original material ratio data set. The material ratio sensor is installed at the key nodes of the fluidized soil-solidified soil production equipment and can monitor the material ratio in real time. For example, in the monitoring of the water-cement ratio ( ), the sensor measures the actual usage of water and cement through a flow meter, and then calculates the current water-cement ratio; for the dosage of curing agent ( ), the admixture content ( ), and the cement dosage ( ), the sensor directly obtains the actual data of the feeding through a mass sensor to form a multi-dimensional original material ratio data set . Among them, represents the number of data sampling times, is the sampling time. At the same time, the mixing speed, mixing time, and mixing power are collected through the mixing parameter sensor to obtain the original mixing parameter data set. The mixing process has a significant impact on the uniformity and strength of the fluidized soil-solidified soil. The mixing parameter sensor can monitor the operating state of the mixing device in real time. For example, the mixing speed is obtained through a rotational speed sensor, the mixing time is recorded through a time recording device, and the mixing power is measured by a power meter to measure the actual power consumption of the motor, obtaining the original mixing parameter data set . These data can reflect the energy input and rheological properties during the mixing process and play an important role in predicting the strength of the fluidized soil-solidified soil. The environmental temperature, relative humidity, and atmospheric pressure are collected through the environmental monitoring sensor to obtain the original environmental monitoring data set. During the production and construction process of the fluidized soil-solidified soil, the change of environmental conditions will also have an impact on the final material properties. For example, temperature ( ), humidity ( ), and pressure ( ) will all affect the hydration reaction rate and curing effect of the material. The environmental monitoring sensor collects these data through a thermometer, a hygrometer, and a barometer respectively to form the original environmental monitoring data set These external condition data provide a comprehensive description of the material curing environment, which can help the model better understand the impact of the environment on material properties. Imputation is performed on the missing data in the original material ratio dataset, the original mixing parameter dataset, and the original environmental monitoring dataset to ensure data integrity and continuity. When the number of consecutive missing points does not exceed 3, cubic spline interpolation is used. By constructing a smooth interpolation function, the positions of the missing data are predicted. When the number of consecutive missing points exceeds 3, a predictive filling algorithm based on historical data patterns is used. For example, by constructing an autoregressive moving average model and using the time series characteristics of historical data for predictive filling. After imputation, the target material ratio dataset, the target mixing parameter dataset, and the target environmental monitoring dataset are obtained respectively. Normalization is performed on the target material ratio dataset, the target mixing parameter dataset, and the target environmental monitoring dataset to eliminate the dimensional differences between different data dimensions and make all data within the same scale interval. The normalization formula is:

[0022] where is the normalized data, and are the minimum and maximum values in the dataset respectively. Through normalization, the data range is limited between [0, 1]. After normalization, the data is converted into matrix form to obtain the material ratio parameter matrix , the mixing parameter matrix and the environmental monitoring data matrix , where . Each row in these matrices represents the sampled data at a time point, and each column represents a specific parameter. The feature extraction process is implemented through a convolutional neural network. Each matrix extracts deep features through different convolutional structures. For example, for the material ratio matrix, a three-layer convolutional structure is used, and the material ratio feature vector is extracted through specific convolutional kernels and activation functions; the mixing parameter matrix and the environmental monitoring data matrix also obtain feature vectors and respectively through different convolutional structures. To fuse these features, a low-rank multi-modal fusion method is adopted to map the feature vectors , and to a shared feature space and achieve feature integration through tensor decomposition technology. The fused feature vector is represented by the formula:

[0023] where , and is the weight matrix for feature fusion, which is dynamically adjusted through model training to achieve the best feature fusion effect. The finally obtained target feature vector contains the key information in multi-source data and can comprehensively reflect the parameter characteristics in the production process of fluid-solidified soil.

[0024] In a specific embodiment, the process of performing steps to extract and integrate features from the material ratio parameter matrix, mixing parameter matrix, and environmental monitoring data matrix to obtain the target feature vector may specifically include the following steps: Input the material ratio parameter matrix into a three-layer convolutional structure for convolutional operation and ReLU activation to obtain the material ratio feature vector; Input the mixing parameter matrix into a two-layer convolutional structure for convolutional operation and Leaky ReLU activation to obtain the mixing parameter feature vector; Input the environmental monitoring data matrix into a two-channel convolutional structure for convolutional operation to capture long-term trend features and short-term fluctuation features, obtain two-channel features, and perform connection layer merging on the two-channel features to obtain the environmental monitoring feature vector; Perform feature integration on the material ratio feature vector, mixing parameter feature vector, and environmental monitoring feature vector through low-rank multi-modal fusion and bottleneck transformer to obtain the target feature vector.

[0025] Specifically, input the material ratio parameter matrix into a three-layer convolutional structure for convolutional operation and ReLU activation to extract the material ratio feature vector . In the convolutional network, the first-layer convolutional operation uses number of convolution kernels, and the convolutional operation is implemented through the following formula:

[0026] where, is the first-layer convolutional kernel weight matrix, is the bias vector, represents the convolutional operation, is the ReLU activation function, is the output feature map of the first layer. The second layer of convolution uses number of convolution kernels to continue feature extraction:

[0027] Similarly, the third-layer convolutional operation uses number of convolution kernels to obtain the final material ratio feature vector :

[0028] Among them, and are the weight matrix and bias vector of the third-layer convolution respectively. Through such multi-layer convolution operations, the model can gradually extract high-dimensional feature information in the material ratio parameters and effectively capture complex patterns in the data. At the same time, for the stirring parameter matrix , a two-layer convolution structure is used for convolution operation and Leaky ReLU activation to obtain the stirring parameter feature vector . The activation function in the stirring parameter feature extraction uses Leaky ReLU. This activation function still retains a certain output in the negative value interval to avoid the problem of neuron inactivation. The first-layer convolution operation uses number of convolution kernels, and the formula is as follows:

[0029] Among them, is the Leaky ReLU activation function, and are the weight matrix and bias of the first-layer convolution respectively. The second-layer convolution operation uses number of convolution kernels to obtain the stirring parameter feature vector:

[0030] This feature extraction method can effectively identify the potential impact of changes in stirring parameters on the strength of fluidized solidified soil. For example, how small changes in stirring speed and power affect the final material properties. At the same time, the environmental monitoring data matrix is input into a two-channel convolution structure. Different convolution kernels are used to capture long-term trend features and short-term fluctuation features respectively, and two-channel features are obtained. These two-channel features are merged through a connection layer to obtain the environmental monitoring feature vector . The first channel uses number of convolution kernels to extract long-term trend features:

[0031] The second channel uses number of convolution kernels, which are specifically used to capture short-term fluctuation features:

[0032] The two-channel features are merged through a connection operation to obtain the comprehensive environmental monitoring feature vector . After completing the feature extraction, the material ratio feature vector , the stirring parameter feature vector and the environmental monitoring feature vector Input into the low-rank multimodal fusion module, and the final target feature vector is obtained through feature integration . Map these three feature vectors to the same feature space through linear projection:

[0033] where 、 and are projection vectors, 、 and are the linear projection matrices respectively. Construct the three-modal interaction tensor , and obtain the low-rank representation tensor through Tucker decomposition. Tucker decomposition represents the interaction tensor as the product of the core tensor and the factor matrices:

[0034] where, represents the mode product of the tensor and the matrix, is the factor matrix, is the rank parameter. Reconstruct the core tensor of the low-rank representation into the fusion feature vector , and input it into the bottleneck transformer to further integrate features. The bottleneck transformer extracts deep features through the multi-head self-attention mechanism and the feed-forward neural network, and finally compresses the feature dimension to 1 / 2 of the original dimension to obtain the final target feature vector .

[0035] In a specific embodiment, the process of performing the steps to integrate the features of the material ratio feature vector, the stirring parameter feature vector, and the environmental monitoring feature vector through low-rank multimodal fusion and the bottleneck transformer to obtain the target feature vector may specifically include the following steps: Map the material ratio feature vector, the stirring parameter feature vector, and the environmental monitoring feature vector respectively through the linear projection matrix to obtain the material ratio projection vector, the stirring parameter projection vector, and the environmental monitoring projection vector; Construct the corresponding three-modal interaction tensors based on the material ratio projection vector, the stirring parameter projection vector, and the environmental monitoring projection vector; Perform Tucker decomposition on the three-modal interaction tensors to obtain the low-rank representation tensor, and reconstruct the core tensor in the low-rank representation tensor into a vector form to obtain the fusion feature vector; Input the fusion feature vector into the bottleneck transformer for feature integration. The bottleneck transformer contains 4 encoder layers, and each encoder layer has an 8-head self-attention mechanism and a feed-forward neural network to obtain the bottleneck transformer output features; The output features of the bottleneck transformer are reduced to 1 / 2 of the original dimension in the middle layer to obtain the target feature vector.

[0036] Specifically, the feature vectors from different sources are mapped through a linear projection matrix into a unified feature space for subsequent multimodal fusion processing. Material ratio feature vector , stirring parameter feature vector and environmental monitoring feature vector are respectively linearly transformed through the linear projection matrices , and to obtain the material ratio projection vector , stirring parameter projection vector and environmental monitoring projection vector , and the calculation formula is:

[0037] where represents the dimension of the shared feature space. This dimension setting can reduce the computational complexity and establish a unified feature scale between different modalities, making the subsequent tensor decomposition operation more efficient. Based on the projection vectors, three-modal interaction tensors are constructed to capture the high-order interaction relationships between multimodal features. Interaction tensor is constructed through the outer product operation, and each element represents the correlation between the three feature dimensions. The specific calculation formula is:

[0038] where , , are respectively the values of the material ratio, stirring parameter, and environmental monitoring projection vectors on the , , dimensions. Interaction tensor effectively encodes the ternary relationship between multimodal features in this way and can reflect the potential impact of different feature combinations on the strength of fluid-solidified soil. To reduce the computational complexity of the high-dimensional tensor, the three-modal interaction tensors are subjected to Tucker decomposition. Tucker decomposition is a multilinear algebra method that represents a high-dimensional tensor as the product of a core tensor and a set of factor matrices. The specific formula is:

[0039] where represents the mode product of the tensor and the matrix. , , are the factor matrices of three modalities respectively. is the rank parameter of the tensor, which controls the compactness of the core tensor. Tucker decomposition compresses the original high-dimensional interaction tensor into a more manageable low-rank representation by reducing the dimension, while retaining the key information in the multi-modal features. The core tensor in the low-rank representation tensor is reconstructed into a vector form to obtain the fused feature vector . The reconstruction process unfolds the core tensor along different modalities into one-dimensional vectors to achieve an effective mapping from high-dimensional features to low-dimensional features. This mapping not only preserves the high-order interaction relationships between multi-modal features, but also simplifies the feature representation into a form that is more easily processed by the model. To improve the quality of the feature representation, the fused feature vector is input into the bottleneck transformer for feature integration. The bottleneck transformer contains four encoder layers, and each encoder layer consists of an eight-head self-attention mechanism and a feed-forward neural network. The self-attention mechanism can calculate the correlation between each feature in the feature vector and other features. By calculating the attention weight matrix , the calculation formula of the self-attention mechanism is:

[0040] where , and are the query, key, and value matrices respectively, , , are trainable weight matrices, is the dimension of the attention head. The multi-head attention mechanism concatenates the attention outputs of different heads and performs a non-linear transformation through a feed-forward neural network, enabling the feature vector to fully capture long-range feature dependencies. The bottleneck transformer introduces a feature dimension compression mechanism in each encoder layer, and by reducing the feature dimension to half of the original dimension in the intermediate layer, it achieves a compact processing of the feature representation. Through the dimension compression matrix and the recovery matrix for linear transformation:

[0041] where is the output feature of the bottleneck transformer, is the final target feature vector. Dimension compression not only improves the computational efficiency of the model, but also forces the feature representation to be more compact, enabling the model to learn more critical feature information, thereby improving the performance and stability of the fluid-solidified soil strength prediction model.

[0042] In a specific embodiment, the process of executing step S2 may specifically include the following steps: Input the target feature vector into a neural network structure composed of a three-layer bidirectional long short-term memory network for temporal feature extraction to obtain the output features of the bidirectional long short-term memory network; Perform time-dependence analysis on the target feature vector through a temporal convolution module to obtain temporal convolution features, and perform weighted fusion of the temporal convolution features through an attention mechanism to obtain fused temporal convolution features; Perform residual connection on the output features of the bidirectional long short-term memory network and the fused temporal convolution features to obtain residual features; Input the residual features into a multi-task learning architecture, and simultaneously predict the strength values of the fluid-solidified soil at 7 days, 14 days, and 28 days through the shared encoder and three task-specific decoder structures in the multi-task learning architecture to obtain the strength prediction values of the fluid-solidified soil and the corresponding uncertainty estimates.

[0043] Specifically, input the target feature vector into a neural network structure composed of a three-layer bidirectional long short-term memory network (BiLSTM) for temporal feature extraction to obtain the output features of the bidirectional long short-term memory network. . In this process, the BiLSTM network can simultaneously capture the forward and backward dependencies in the time series data, enabling the model to comprehensively understand the changing trends and dynamic characteristics of the features in the time dimension. For each layer of BiLSTM, the update formulas for the hidden state and the memory cell are as follows:

[0044]

[0045] where , , represent the activation values of the output gate, forget gate, and input gate respectively, © is the element-wise product operation, is the hyperbolic tangent activation function, is the candidate memory state. Through these gating mechanisms, the BiLSTM network effectively selectively remembers and forgets historical information, ensuring that the model still maintains stable learning ability when processing long sequence data. Perform time-dependence analysis on the target feature vector through a temporal convolution module to obtain temporal convolution features . The temporal convolution module uses convolution kernels of different sizes (such as 3, 5, and 7) to capture short-term, medium-term, and long-term dependencies in the data. The core calculation formula for the temporal convolution operation is:

[0046] Among them, Conv1D represents a one-dimensional convolution operation, is the convolution kernel weight matrix, is the bias vector, and is the number of convolution kernels. In the temporal convolution module, multi-scale temporal feature information is captured simultaneously through convolution kernels of different scales, ensuring that the model can effectively model various temporal patterns. To further enhance the expressive power of the temporal convolution features, an attention mechanism is introduced to perform weighted fusion on the temporal convolution features to obtain fused temporal convolution features . The core of the attention mechanism lies in calculating the importance of each time step in the features. By constructing query, key, and value matrices , , , the attention weight matrix is calculated and the features are weighted and summed. The specific formula is:

[0047] Among them, , , , , and are learnable weight matrices, and the softmax function ensures that the attention weights are normalized to a probability distribution at each time step. Through this weighted fusion method, the model automatically focuses on the time points in the features that are more critical for intensity prediction, improving the model's ability to capture complex temporal dependencies. The output features of the bidirectional long short-term memory network and the fused temporal convolution features are subjected to residual connection to obtain residual features . Residual connection not only retains the information in the original temporal features but also alleviates the vanishing gradient problem in deep networks through direct addition operations. The calculation formula for the residual features is:

[0048] Through this operation, while the model performs deep feature extraction, it can still maintain the original information of the shallow features, thereby improving the stability of prediction and the generalization ability of the model. The residual features are input into a multi-task learning architecture for multi-time-scale prediction of the strength of fluid-solidified soil. The multi-task learning architecture consists of a shared encoder and three task-specific decoders, and can simultaneously predict the strength values at 7 days, 14 days, and 28 days , and . The shared encoder converts the residual features into general temporal features , and then input them into the task-specific decoders respectively. Each decoder makes predictions for a specific time scale, and the prediction formula is:

[0049] where is the weight matrix of the decoder, is the bias term. This multi-task prediction method not only improves the training efficiency of the model but also can enhance the accuracy of predictions at each time scale through collaborative learning between tasks. In addition to predicting the strength value, an uncertainty estimate value is provided to evaluate the reliability of the prediction result. The model adds an uncertainty branch to the output layer of each decoder, and realizes uncertainty quantification by calculating the variance of the prediction result. The formula for uncertainty estimation is:

[0050] where and are the weights and biases of the uncertainty estimation. Through the exponential mapping on the logarithmic scale, it is ensured that the uncertainty is always positive. This method can help the model to provide a confidence interval while making prediction outputs, providing a basis for risk assessment in decision-making in practical applications. The model simultaneously outputs the strength prediction values , , of the fluid-solidified soil at different time scales, as well as the corresponding uncertainty estimation values , , .

[0051] In a specific embodiment, the process of executing step S3 may specifically include the following steps: Model the fluid-solidified soil equipment control optimization problem as a collaborative solution framework of the main problem and sub-problems. The main problem is defined as maximizing the fluid-solidified soil strength function, and the sub-problems are responsible for evaluating the expected strength and sensitivity after parameter adjustment; Based on the collaborative solution framework, perform vector conversion on the strength prediction value of the fluid-solidified soil and the corresponding uncertainty estimation value to obtain the control parameter vector; Based on the control parameter vector and the set of constraint conditions, construct the Lagrangian function, and take the second derivative of the Lagrangian function with respect to the control parameters to form the first parameter sensitivity matrix reflecting the second-order sensitivity of the control parameter change to the strength; Iteratively optimize the first parameter sensitivity matrix to obtain the second parameter sensitivity matrix, and calculate the eigenvalues and eigenvectors of the second parameter sensitivity matrix. The eigenvector with the largest eigenvalue indicates the most sensitive direction of parameter adjustment, and obtain the set of parameter sensitive directions; Construct a pre - evaluation index system for optimizing control parameters based on the set of parameter - sensitive directions, and obtain the strength improvement potential index, parameter adjustment stability index, and energy consumption impact index.

[0052] Specifically, model the control optimization problem of the fluid - state solidified soil equipment as a collaborative solution framework of the main problem and sub - problems. With maximizing the strength function of the fluid - state solidified soil as the core goal, simultaneously evaluate the expected strength and sensitivity after adjusting each control parameter. In this collaborative solution framework, the main problem defines the objective function, that is, maximizing the strength of the fluid - state solidified soil, while the sub - problems focus on analyzing the specific impact of parameter changes on the strength, providing precise direction guidance for the optimization process. In the main problem, the strength function of the fluid - state solidified soil is expressed as , where is the control parameter vector, including key process parameters such as water - cement ratio, mixing speed, and curing agent dosage. The goal of maximizing the strength of the fluid - state solidified soil is expressed as:

[0053] where, are the inequality constraint conditions, are the equality constraint conditions, and these constraint conditions reflect the physical, technological, and equipment limitations. For example, the water - cement ratio must be within a certain range, and the mixing speed and curing agent dosage are also restricted by the equipment capacity. In the sub - problems, evaluate the expected improvement effect and sensitivity of each control parameter adjustment on the strength. By vector - transforming the strength prediction value of the fluid - state solidified soil and its corresponding uncertainty estimate value to obtain a comprehensive control parameter vector , the transformation formula of the control parameter vector is:

[0054] where, is a very small positive number used to avoid the denominator being zero. This transformation process enables the control parameters to not only consider the predicted strength value but also introduce the uncertainty of the prediction result, thus being able to more accurately reflect the expected strength change after adjusting the control parameters. Based on the control parameter vector and the set of constraint conditions, construct the Lagrangian function , and its expression is:

[0055] where, and They are the Lagrange multiplier vectors corresponding to the inequality and equality constraints respectively. The Lagrangian function combines the objective function and the constraint conditions. By solving the optimal solution of the Lagrangian function, both the objective optimization and the constraint conditions are satisfied. To evaluate the second-order sensitivity of the fluidized solidified soil strength to the change of control parameters, the Lagrangian function is differentiated twice with respect to the control parameter to obtain the first parameter sensitivity matrix . This Hessian matrix can reveal the mutual influence among control parameters and the sensitivity of parameter changes to strength by calculating the second-order partial derivative matrix. To improve the accuracy of sensitivity analysis, the first parameter sensitivity matrix is iteratively optimized to obtain the second parameter sensitivity matrix . In the iterative process, the BFGS algorithm in the quasi-Newton method is adopted, and the specific update formula is:

[0056] where is the parameter update vector, is the gradient change, is the Hessian matrix at the -th iteration. Through this iterative formula, the sensitivity matrix approaches the true second-order sensitivity matrix continuously, providing more accurate sensitivity analysis results. Calculate the eigenvalues and eigenvectors of the second parameter sensitivity matrix . The eigenvector with the largest eigenvalue indicates the most sensitive direction of control parameter adjustment, that is, adjusting the parameters along the direction will be able to produce the most significant strength improvement effect. The set of parameter sensitive directions is composed of these eigenvectors, providing clear direction guidance for the next control parameter optimization. Based on the set of parameter sensitive directions, a pre-evaluation index system for control parameter optimization is constructed, including the strength improvement potential index , the parameter adjustment stability index and the energy consumption impact index . The strength improvement potential index is calculated by the following formula:

[0057] where is the parameter adjustment amount, is the gradient of the strength function, quantifies the expected improvement effect of parameter adjustment on strength. The parameter adjustment stability index evaluates the robustness of parameter adjustment, and the calculation formula is:

[0058] Smaller values indicate a more stable adjustment effect, indicating that when the control parameters change, the strength of the fluid-solidified soil will not fluctuate violently. Energy consumption impact index is used to evaluate the impact of control parameter adjustment on the system energy consumption and is calculated by a linear weighted model as:

[0059] where is the energy consumption weight vector, represents the absolute change in parameter adjustment. Energy consumption impact index reflects the energy cost that needs to be paid during the optimization process and helps to achieve a balance between strength improvement and energy efficiency. Through this collaborative solution framework, the optimal combination of control parameters is found, and a multi-dimensional balance is achieved among strength improvement, stability, and energy consumption.

[0060] In a specific embodiment, the process of executing step S4 may specifically include the following steps: Construct a state space and an action space based on the strength improvement potential index, parameter adjustment stability index, and energy consumption impact index. The state space includes the current physical state variables and environmental condition variables of the fluid-solidified soil, and the action space includes the water-cement ratio, mixing speed, and curing agent dosage, to obtain a set of state-action pairs; Establish a state transition model through the set of state-action pairs and construct a parameterized reward function for the state transition model; Solve the parameterized reward function by minimizing the characteristic expectation difference between the expert trajectory and the optimal policy trajectory to obtain an optimized weight vector; Execute the discounted inverse reinforcement learning algorithm based on the optimized weight vector to obtain an optimized control policy, introduce a policy regularization term in the optimized control policy and perform smoothing processing to obtain a set of target control parameters.

[0061] Specifically, the state space includes the current physical state variables and environmental condition variables of the fluid-solidified soil, such as the current strength , water-cement ratio , mixing speed , curing agent dosage , environmental temperature , relative humidity , and atmospheric pressure etc. These state variables constitute a state vector:

[0062] where represents the dimension of the state variables. In the state space, each state Represents the specific state of the system at time These states not only describe the current physical properties of the fluid-solidified soil but also consider the influence of the external environment, which helps to comprehensively evaluate the adjustment effect of the control parameters. The action space is composed of control parameters, including the adjustment amount of the water-cement ratio , the adjustment amount of the stirring speed , and the adjustment amount of the curing agent dosage . These actions define the set of executable operations in the system. The action vector is represented as:

[0063] where is the action dimension. In the reinforcement learning framework, each action represents the specific control strategy taken in the state . These strategies, by affecting the control parameters, cause the system to change in the desired strength improvement direction. Based on the state space and the action space, the state-action pair set is obtained, and a state transition model is established through these state-action pairs. The state transition model is used to describe the probability distribution of the system transferring from the current state to the next state after executing the action . The state transition model is represented by the following formula:

[0064] where is the state transition function, and is the process noise, representing the influence of uncertain factors in the system. The state transition model is fitted through historical data and simulation experiments, which can accurately predict the state change after adjusting the control parameters, ensuring that the reinforcement learning model can reasonably select the optimal action based on the current state. Based on the state transition model, a parameterized reward function is constructed. The form of the reward function is a linearly weighted feature representation:

[0065] where is the feature vector of the state-action pair, and is the weight vector to be solved. In this reward function, the feature vector contains information such as strength improvement potential, stability, and energy consumption impact. For example, it is represented as:

[0066] where is the strength improvement potential index, is the parameter adjustment stability index, is the energy consumption impact index. This feature-based reward function can achieve the dynamic balance of different features in the decision-making process through the learning of the weight vector , so as to ensure that the control strategy of the system can find the optimal solution under multiple objectives. To solve the weight vector in the parameterized reward function, the weight is optimized by minimizing the feature expectation difference between the expert trajectory and the optimal policy trajectory. The expert trajectory is a sequence of state-action pairs based on historical data or expert experience, and the optimal policy trajectory is the trajectory generated by the current policy. The solution of the weight vector is achieved through the following minimization objective function:

[0067] where, represents the feature expectation. By comparing the feature expectations of the expert trajectory and the optimal trajectory, the learned reward function is made as close as possible to the reward distribution of the expert behavior, thus ensuring the rationality and effectiveness of the policy. After obtaining the optimized weight vector , the discounted inverse reinforcement learning algorithm is executed based on this weight vector. Specifically, the discount factor [0, 1) is used to adjust the trade-off between long-term and short-term rewards, and the discounted cumulative reward function is:

[0068] By maximizing the cumulative reward function, the reinforcement learning algorithm can dynamically generate the optimal control strategy to ensure that the actions selected in different states can all maximize the strength improvement effect of the fluid-solidified soil. To avoid the control strategy being too aggressive or unstable, a policy regularization term is introduced into the optimal control strategy and smoothed. The policy regularization term realizes the smoothness of the policy by restricting the amplitude of the action change at adjacent moments, and the regularization form is:

[0069] where, is the regularization coefficient, which controls the smoothness of the policy change. The introduction of the regularization term helps to maintain the stability of the system during the control process and avoid excessive fluctuations in the performance of the fluid-solidified soil due to excessive parameter adjustment. Through the discounted inverse reinforcement learning algorithm and the policy regularization method, the target control parameter set is obtained. These control parameters can dynamically respond to the changes in the system state and achieve the adaptive optimization of the fluid-solidified soil equipment under different environmental conditions, so as to ensure the stability of the operation and the optimal energy efficiency while maintaining the strength improvement.

[0070] The method for improving the strength of fluid-solidified soil based on multivariate analysis in the embodiments of the present invention has been described above. Next, the device for improving the strength of fluid-solidified soil based on multivariate analysis in the embodiments of the present invention will be described. Please refer to Figure 2 , an embodiment of the device for improving the strength of fluid-solidified soil based on multivariate analysis in the embodiments of the present invention includes: A collection module, configured to collect a material ratio parameter matrix, a stirring parameter matrix, and an environmental monitoring data matrix in a fluid-solidified soil device, and perform feature extraction and feature integration to obtain a target feature vector; A strength prediction module, configured to perform strength prediction of a time series on the target feature vector to obtain a strength prediction value of the fluid-solidified soil; A sensitivity analysis module, configured to perform sensitivity analysis based on the strength prediction value to obtain a strength improvement potential index, a parameter adjustment stability index, and an energy consumption impact index; A parameter optimization module, configured to perform adaptive parameter optimization through discounted inverse reinforcement learning according to the strength improvement potential index, the parameter adjustment stability index, and the energy consumption impact index to obtain a set of target control parameters.

[0071] Through the collaborative cooperation of the above-mentioned various components, through multi-source data collection and preprocessing technologies, three types of key data, namely material ratio, stirring parameters, and environmental monitoring, are systematically collected, and the data is accurately processed, providing a comprehensive and reliable data basis for subsequent analysis, effectively solving the problem of single data source in traditional methods. Deep convolutional neural networks are used for feature extraction, and a dedicated network structure is designed for different types of parameters, which can effectively capture the complex non-linear features of fluid-solidified soil parameters, greatly improving the accuracy and comprehensiveness of feature expression. The introduction of low-rank multi-modal fusion and bottleneck transformer structures explores the correlation between different modal features through tensor decomposition and self-attention mechanisms, solves the problem of multi-source heterogeneous data fusion, and realizes the efficient modeling of complex relationships between parameters. Combining an improved time series prediction algorithm, the strength values at multiple time scales are predicted simultaneously, and uncertainty estimation is provided, enhancing the reliability and practicality of the prediction results. The sensitivity analysis method based on the Lagrangian function accurately quantifies the impact of control parameter adjustment on strength, and the discounted inverse reinforcement learning mechanism realizes the adaptive optimization of control parameters, effectively improving the final strength and quality uniformity of fluid-solidified soil while ensuring operation stability, and solving the problem that it is difficult to optimize control under unknown model conditions in traditional methods.

[0072] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described system, system, and unit can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0073] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0074] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for improving the strength of fluidized solidified soil based on multivariate analysis, characterized in that: include: Collect material ratio parameter matrix, mixing parameter matrix and environmental monitoring data matrix in fluidized soil solidification equipment, perform feature extraction and feature integration, and obtain target feature vector; Performing time series strength prediction on the target characteristic vector to obtain a strength prediction value of the fluidized solidified soil; Perform sensitivity analysis based on the strength prediction value to obtain strength improvement potential index, parameter adjustment stability index and energy consumption impact index; According to the intensity improvement potential index, the parameter adjustment stability index and the energy consumption impact index, adaptive parameter optimization is performed through discounted inverse reinforcement learning to obtain a target control parameter set.

2. The method for improving the strength of fluidized solidified soil based on multivariate analysis according to claim 1, characterized in that: The material ratio parameter matrix, the mixing parameter matrix and the environmental monitoring data matrix in the fluidized soil solidification equipment are collected, and feature extraction and feature integration are performed to obtain the target feature vector, including: The material ratio sensor is used to collect the water-cement ratio, curing agent dosage, admixture dosage and cement dosage to obtain the original material ratio data set; The stirring parameter sensor is used to collect stirring speed, stirring time and stirring power to obtain an original stirring parameter data set; The environmental monitoring sensor collects the ambient temperature, relative humidity and atmospheric pressure to obtain the original environmental monitoring data set; Interpolating missing data in the original material ratio dataset, the original mixing parameter dataset, and the original environmental monitoring dataset to obtain a target material ratio dataset, a target mixing parameter dataset, and a target environmental monitoring dataset; Normalizing the target material ratio data set, the target mixing parameter data set, and the target environment monitoring data set to obtain a material ratio parameter matrix, a mixing parameter matrix, and an environment monitoring data matrix; Feature extraction and feature integration are performed on the material ratio parameter matrix, the stirring parameter matrix and the environmental monitoring data matrix respectively to obtain a target feature vector.

3. The method for improving the strength of fluidized solidified soil based on multivariate analysis according to claim 2, characterized in that: The feature extraction and feature integration of the material ratio parameter matrix, the stirring parameter matrix and the environmental monitoring data matrix are performed respectively to obtain a target feature vector, including: Input the material ratio parameter matrix into a three-layer convolution structure for convolution operation and ReLU activation to obtain a material ratio feature vector; Input the stirring parameter matrix into a two-layer convolution structure for convolution operation and Leaky ReLU activation to obtain a stirring parameter feature vector; Input the environmental monitoring data matrix into a dual-channel convolution structure to perform convolution operation to capture long-term trend characteristics and short-term fluctuation characteristics, obtain two channel characteristics, and perform connection layer merging on the two channel characteristics to obtain an environmental monitoring feature vector; The material ratio feature vector, the stirring parameter feature vector and the environmental monitoring feature vector are integrated through low-rank multimodal fusion and bottleneck transformer to obtain a target feature vector.

4. The method for improving the strength of fluidized solidified soil based on multivariate analysis according to claim 3, characterized in that: The method integrates the material ratio feature vector, the stirring parameter feature vector and the environmental monitoring feature vector through low-rank multimodal fusion and bottleneck transformer to obtain a target feature vector, including: The material ratio feature vector, the stirring parameter feature vector and the environmental monitoring feature vector are respectively mapped by a linear projection matrix to obtain a material ratio projection vector, a stirring parameter projection vector and an environmental monitoring projection vector; Constructing corresponding three modal interaction tensors based on the material ratio projection vector, the stirring parameter projection vector and the environmental monitoring projection vector; Performing Tucker decomposition processing on the three modal interaction tensors to obtain a low-rank representation tensor, and reconstructing the core tensor in the low-rank representation tensor into a vector form to obtain a fused feature vector; Inputting the fused feature vector into a bottleneck transformer for feature integration, wherein the bottleneck transformer comprises four encoder layers, each encoder layer having an eight-head self-attention mechanism and a feedforward neural network, and obtaining the output feature of the bottleneck transformer; The bottleneck transformer output features are output by reducing the feature dimension to 1 / 2 of the original dimension in the intermediate layer to obtain the target feature vector.

5. The method for improving the strength of fluidized solidified soil based on multivariate analysis according to claim 1, characterized in that: The performing time series strength prediction on the target characteristic vector to obtain the strength prediction value of the fluidized solidified soil includes: Inputting the target feature vector into a neural network structure consisting of a three-layer bidirectional long short-term memory network to extract temporal features, and obtaining a bidirectional long short-term memory network output feature; Performing a time dependency analysis on the target feature vector through a time convolution module to obtain a time convolution feature, and performing a weighted fusion of the time convolution feature using an attention mechanism to obtain a fused time convolution feature; Performing a residual connection on the bidirectional long short-term memory network output feature and the fused temporal convolution feature to obtain a residual feature; The residual features are input into a multi-task learning architecture. Through the shared encoder and three task-specific decoder structures in the multi-task learning architecture, the strength values ​​of the fluidized solidified soil after 7 days, 14 days and 28 days are predicted simultaneously to obtain the strength prediction value of the fluidized solidified soil and the corresponding uncertainty estimation value.

6. The method for improving the strength of fluidized solidified soil based on multivariate analysis according to claim 5, characterized in that: The sensitivity analysis based on the strength prediction value is performed to obtain the strength improvement potential index, parameter adjustment stability index and energy consumption impact index, including: The fluidized soil solidification equipment control optimization problem is modeled as a collaborative solution framework of the main problem and sub-problems. The main problem is defined as maximizing the fluidized soil solidification strength function, and the sub-problems are responsible for evaluating the expected strength and sensitivity after parameter adjustment. Based on the collaborative solution framework, vector transformation is performed on the strength prediction value of the fluidized solidified soil and the corresponding uncertainty estimation value to obtain a control parameter vector; Constructing a Lagrangian function based on the control parameter vector and the constraint condition set, and calculating the second-order derivative of the Lagrangian function with respect to the control parameter to form a first parameter sensitivity matrix reflecting the second-order sensitivity of the control parameter change to the intensity; Iteratively optimize the first parameter sensitivity matrix to obtain a second parameter sensitivity matrix, and calculate eigenvalues ​​and eigenvectors for the second parameter sensitivity matrix, wherein the eigenvector with the largest eigenvalue indicates the most sensitive direction of parameter adjustment, thereby obtaining a set of parameter sensitive directions; Based on the parameter sensitive direction set, a pre-evaluation index system for control parameter optimization is constructed to obtain strength improvement potential index, parameter adjustment stability index and energy consumption impact index.

7. The method for improving the strength of fluidized solidified soil based on multivariate analysis according to claim 1, characterized in that: The adaptive parameter optimization is performed through discounted inverse reinforcement learning according to the strength improvement potential index, the parameter adjustment stability index and the energy consumption impact index to obtain a target control parameter set, including: According to the strength improvement potential index, the parameter adjustment stability index and the energy consumption impact index, a state space and an action space are constructed, wherein the state space includes the current physical state variables and environmental condition variables of the fluidized solidified soil, and the action space includes the water-cement ratio, the stirring speed and the amount of the curing agent, to obtain a state-action pair set; Establishing a state transition model through the state-action pair set, and constructing a parameterized reward function for the state transition model; Solving the parameterized reward function by minimizing the expected difference in features between the expert trajectory and the optimal strategy trajectory to obtain an optimized weight vector; A discounted inverse reinforcement learning algorithm is executed based on the optimized weight vector to obtain an optimal control strategy, a strategy regularization term is introduced into the optimal control strategy and smoothing is performed to obtain a target control parameter set.

8. A fluidized solidified soil strength enhancement device based on multivariate analysis, characterized in that: Used to perform the method for improving the strength of fluidized solidified soil based on multivariate analysis according to any one of claims 1 to 7, the fluidized solidified soil strength improvement device based on multivariate analysis comprises: The acquisition module is used to collect the material ratio parameter matrix, mixing parameter matrix and environmental monitoring data matrix in the fluidized soil solidification equipment, and perform feature extraction and feature integration to obtain the target feature vector; A strength prediction module, used for performing time series strength prediction on the target feature vector to obtain a strength prediction value of the fluidized solidified soil; A sensitivity analysis module, used to perform sensitivity analysis based on the strength prediction value to obtain a strength improvement potential index, a parameter adjustment stability index, and an energy consumption impact index; The parameter optimization module is used to perform adaptive parameter optimization through discounted inverse reinforcement learning according to the strength improvement potential index, the parameter adjustment stability index and the energy consumption impact index to obtain a target control parameter set.

Images