A Cement-Water Glass Gelatin Strength Assessment Method and System Based on Big Data

Abstract

This invention discloses a method and system for evaluating the strength of cement-water glass gel based on big data. The method includes data acquisition, feature mapping, latent variable estimation, hybrid model training, interpretable output, and strength evaluation. This invention relates to the field of intelligent material evaluation technology, specifically a method and system for evaluating the strength of cement-water glass gel based on big data. This scheme transforms persistent coherent information into a continuously persistent image with physical weights and discretizes it into vectors. It then constructs differentiable microscopic features by combining spectral peak features, porosity, and particle size. Using generalized Avramyan dynamics as a framework, the microscopic features are mapped to kinetic parameters, and temperature-humidity coupling is introduced. Combined with extended Kalman filtering, online estimation of gel reactivity and uncertainty quantification are achieved. A hybrid model combining physical priors and neural network correction is constructed, integrating physical constraints and data-driven training, thus improving the accuracy, physical consistency, and generalization ability of the evaluation.

Description

Technical Field

[0001] This invention relates to the field of intelligent material evaluation technology, specifically to a method and system for evaluating the strength of cement-water glass gel based on big data. Background Technology

[0002] With the rapid development of big data, the Internet of Things, and artificial intelligence, materials science and engineering are evolving towards high-throughput characterization, online sensing, and digital twins. The improvement of numerous characterization methods and computing power has enabled the simultaneous acquisition of multimodal large-sample data on the formulation, environment, microstructure, and mechanical properties of cement-water glass gel, a multi-scale dynamic system. This is driving the transformation from empirical experiments to data-driven and physically coupled real-time prediction, monitoring, and closed-loop control. At the same time, the industry's requirements for interpretability, uncertainty quantification, and engineering deployability are also increasing.

[0003] Traditional methods mainly rely on offline experimental design, manual feature engineering, or single physical models, which have several shortcomings: feature extraction often ignores the continuous differentiable representations of topology and spectroscopy, making it difficult to use for end-to-end learning; pure physical or empirical models are difficult to incorporate complex microscopic representations and heterogeneity between samples, easily leading to underfitting or loss of generalization ability; while purely data-driven methods are flexible, they lack physical constraints, are prone to non-physical interpretations, and usually cannot provide reliable uncertainty estimates; in addition, traditional processes are not suitable for real-time estimation of unobservable states and maintaining stable predictive capabilities under sensor data drift or formulation changes, thus failing to meet the needs of real-time evaluation, interpretable decision-making, and reliability assurance in modern production. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method and system for evaluating the strength of cement-water glass gel based on big data. Addressing the problems of discontinuous and non-differentiable features, and difficulty in associating microscopic topological information with physical formulations, traditional feature extraction methods convert persistent cohomological information into a continuously persistent image with physical weights and discretize it into a vector. Simultaneously, physical quantitative indicators of peak area and half-maximum width are extracted from the diffraction pattern to construct a microscopic feature vector coupled with the formulation and pore size. Furthermore, addressing the problems of traditional kinetic modeling and empirical fitting methods, such as difficulty in real-time estimation of gel reactivity, difficulty in coupling microscopic features with environmental disturbances, and lack of uncertainty quantification, this solution uses a generalized... Using dynamics as the physical framework, the microscopic feature vectors are mapped to dynamic parameters and temperature-humidity coupling is introduced. Combined with extended Kalman filtering, the latent variable gelation degree is advanced a priori and updated by observation. To address the problems of underfitting in traditional pure physical models and loss of physical constraints in pure data-driven models, this scheme constructs a hybrid predictor consisting of a physical prior strength curve and a correction term output by a differentiable neural network. It is trained by jointly using data fitting error, physical residual consistency penalty and monotonicity constraint, so that the prediction follows both dynamic physical consistency and has flexible correction capability driven by data.

[0005] The technical solution adopted in this invention is as follows: a method for evaluating the strength of cement-water glass gel based on big data, which includes the following steps:

[0006] Step S1: Data acquisition, collecting experimental formula records, time-series characteristic data, microscopic characterization data, and intensity labels at corresponding times from historical samples;

[0007] Step S2: Feature mapping, extract differentiable topological and frequency domain features from micro-representation data, obtain continuous persistent images with physical weights and discretize them to construct micro-feature vectors;

[0008] Step S3: Latent variable estimation. A low-dimensional state-space model with the degree of gel reaction as the latent variable is established. The kinetic parameters are mapped using microscopic and temporal characteristic data, and prediction and updating are performed through extended Kalman filtering to obtain virtual measurements.

[0009] Step S4: Hybrid model training. Construct a hybrid predictor consisting of physical priors and data-driven correction terms. The physical priors are given by microscopic feature vectors and virtual measurements. Fitting error, physical residuals, and monotonicity penalties are added during joint training.

[0010] Step S5: Interpretability output provides confidence intervals and local feature contributions for the prediction results. Deep ensemble and heterogeneous variance regression are used to simultaneously model the mean and data noise, and then local linear approximation is used to make explicit feature contributions.

[0011] Step S6: Intensity assessment. Perform feature mapping and latent variable estimation on the sample to be assessed. Input the micro-feature vector, time-series feature data and virtual measurement into the hybrid model, and output the intensity prediction value, confidence interval and local feature contribution in real time.

[0012] Further, in step S1, the data acquisition includes collecting historical experimental and formulation data, time-series characteristic data, microscopic characterization data, and strength label data of cement-water glass gel. The experimental and formulation data includes: unique identifier, water-cement ratio, silica mass fraction, sodium oxide mass fraction, admixture type and mass fraction, initial mixing temperature, and stirring time. The time-series characteristic data includes: temperature, humidity, solution pH, and conductivity. The microscopic characterization data includes: electron micrographs, X-ray diffraction patterns, Fourier transform infrared spectroscopy, porosity, and median particle diameter. The strength label data refers to the compressive strength of the cement-water glass gel.

[0013] Further, in step S2, the feature mapping constructs differentiable topological and frequency domain feature maps for the microscopic characterization data to obtain continuous persistent image vectors, while simultaneously extracting physical quantification indicators of peak position and peak intensity, specifically including the following steps:

[0014] Step S21: Construct a morphological subset by extracting a binary sublayer from the electron micrograph and constructing an upstream filtering sequence to generate a morphological subset;

[0015] Step S22: Obtain the persistent image, calculate the persistent coherence information generated by morphological filtering, and only take 0 and 1 dimensions: obtain the persistent image;

[0016] Step S23: Persistent image transformation, performing a physically weighted transformation on the persistent image;

[0017] Step S24: Construct micro-feature vectors by discretizing the persistent image after physical weight transformation on a fixed grid to obtain vectorized features; then extract spectral energy features from the Fourier infrared spectrum; and construct the final micro-feature vectors.

[0018] Further, in step S3, the latent variable estimation involves establishing a low-dimensional, physics-driven state-space model, treating the degree of gelation as a latent variable that cannot be directly and completely observed, and using temporal and microscopic features, performing posterior estimation of the latent variable through extended Kalman filtering to obtain the estimated virtual measurement. Specifically, this includes the following steps:

[0019] Step S31: Construct a dynamic model, using a generalized model with temperature and humidity coupling. The model maps microscopic features to dynamic constants through parameters;

[0020] Step S32: Discretization, perform forward Euler discretization on the latent variables, and construct the observation vector at the same time;

[0021] Step S33: Extended Kalman Filtering. The extended Kalman filter is used to estimate the latent variables posteriorly. First, the latent variables posteriorly are estimated at the current time by combining the time step and dynamic rate information. Then, the covariance matrix is ​​calculated. Finally, the overall update is performed.

[0022] Further, in step S4, the hybrid model training constructs a physically regularized hybrid predictor, taking microscopic feature vectors, temporal features, and dummy measurements as inputs to predict the target compression strength, specifically including the following steps:

[0023] Step S41: Construct a prior physical model of intensity, taking the parameters determined by the microscopic eigenvectors and the time-varying virtual measurement as inputs, and outputting a predicted intensity curve over time based on physical experience.

[0024] Step S42: Define a data-driven correction term, which is a differentiable neural network that accepts microscopic features, temporal features and virtual measurements as inputs and outputs a correction amount for physical priors;

[0025] Step S43: Final prediction. Add the physical prior prediction to the data-driven correction term to obtain the final mixed model strength prediction value, which is used as the model output during training and inference.

[0026] Step S44: Training loss, which includes data fitting error, physical residual penalty, and monotonicity penalty.

[0027] Further, in step S5, the interpretability output provides confidence intervals and local feature contributions for the prediction results. It employs deep ensemble and heterogeneous variance regression to simultaneously model the mean and data noise, and then uses local linear approximation to make explicit feature contributions. Specifically, this includes the following steps:

[0028] Step S51: Model integration, integration The mean and variance outputs of each model are ensembled through independent training. The total variance is obtained from different initialization models with the same structure; the total variance is decomposed into two parts: observable noise and parameter uncertainty.

[0029] Step S52: Generate confidence intervals, providing confidence intervals based on the approximate Gaussian assumption. The width of the confidence interval is determined by the total standard deviation.

[0030] Step S53: Generate contribution values ​​and estimate feature contributions for the input vector;

[0031] Step S54: Uncertainty decomposition, the total variance is decomposed into two parts. The first part is the process uncertainty that propagates from the uncertainty of the dynamic hidden variables to the intensity, and the second part is the observation uncertainty introduced by the data-driven part and observation noise.

[0032] Step S55: Chain propagation, based on the chain rule, maps the extended Kalman filter posterior covariance to the uncertainty of intensity prediction through the physical prior model.

[0033] Further, in step S6, the strength assessment involves acquiring the cement-water glass gel data to be assessed, inputting it into a hybrid model, using the hybrid model to predict the strength of the cement-water glass gel in real time as the assessment strength, and generating an interpretable output.

[0034] The cement-water glass gel strength evaluation system based on big data provided by this invention includes a data acquisition module, a feature mapping module, a latent variable estimation module, a hybrid model training module, an interpretable output module, and a strength evaluation module.

[0035] The data acquisition module collects historical experimental and formulation data, time-series characteristic data, microscopic characterization data, and strength label data of cement-water glass gel, and sends the data to the feature mapping module, latent variable estimation module, hybrid model training module, interpretable output module, and strength assessment module.

[0036] The feature mapping module receives data sent by the data acquisition module, constructs differentiable topological and frequency domain feature mappings for the microscopic characterization data, obtains continuous persistent image vectors, extracts physical quantification indicators of peak position and peak intensity, and sends the data to the latent variable estimation module.

[0037] The latent variable estimation module receives data sent by the data acquisition module and the feature mapping module, establishes a low-dimensional physical-driven state-space model, treats the degree of gelation as a latent variable that cannot be directly and completely observed, uses temporal and micro-features, performs posterior estimation of the latent variable through extended Kalman filtering to obtain the estimated virtual measurement, and sends the data to the hybrid model training module.

[0038] The hybrid model training module receives data from the data acquisition module and the latent variable estimation module, constructs a physically regularized hybrid predictor, takes micro-feature vectors, temporal features and virtual measurements as inputs, predicts the target compression strength, and sends the data to the interpretable output module.

[0039] The interpretable output module receives data sent by the data acquisition module and the hybrid model training module, provides confidence intervals and local feature contributions for the prediction results, uses deep ensemble and heterogeneous variance regression to simultaneously model the mean and data noise, then uses local linear approximation to make explicit feature contributions, and sends the data to the intensity assessment module.

[0040] The strength assessment module receives data from the data acquisition module and the interpretability output module, obtains the cement-water glass gel data to be assessed, inputs it into the hybrid model, uses the hybrid model to predict the strength of the cement-water glass gel in real time as the assessment strength, and generates interpretable output.

[0041] The beneficial effects achieved by the present invention using the above solution are as follows:

[0042] (1) In view of the problems of discontinuous and non-differentiable features and difficulty in associating microscopic topological information with physical formulation in traditional feature extraction methods, this scheme transforms persistent cohomology information into a continuous persistent image with physical weights and discretizes it into a vector. At the same time, it extracts physical quantitative indicators of peak area and half peak width from the diffraction pattern and constructs a microscopic feature vector coupled with formulation and pore size, thereby obtaining a differentiable and physically-aware feature table that can be used for end-to-end training.

[0043] (2) To address the problems of traditional kinetic modeling and empirical fitting methods, such as difficulty in real-time estimation of gel reactivity, difficulty in coupling microscopic features with environmental disturbances, and lack of uncertainty quantification, this scheme uses a generalized... Using dynamics as the physical framework, microscopic feature vectors are mapped to dynamic parameters and temperature-humidity coupling is introduced. Combined with extended Kalman filtering, the latent variable gelation degree is advanced a priori and updated by observation. This not only realizes real-time virtual measurement estimation, but also provides posterior covariance for uncertainty propagation.

[0044] (3) To address the problems of traditional pure physics models being prone to underfitting and pure data-driven models being prone to losing physical constraints, this scheme constructs a hybrid predictor consisting of a physical prior strength curve and a correction term output by a differentiable neural network. It uses data fitting error, physical residual consistency penalty and monotonicity constraint for training, so that the prediction follows both dynamic physical consistency and has flexible correction capability driven by data, thereby improving generalization and suppressing non-physical understanding. Attached Figure Description

[0045] Figure 1 A schematic diagram illustrating the cement-water glass gel strength evaluation method based on big data provided by this invention;

[0046] Figure 2 A schematic diagram of the cement-water glass gel strength evaluation system based on big data provided by the present invention;

[0047] Figure 3 This is a schematic diagram of step S2;

[0048] Figure 4 This is a schematic diagram of step S3;

[0049] Figure 5 This is a schematic diagram of step S4;

[0050] Figure 6 This is a schematic diagram of step S5.

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0052] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0053] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0054] Example 1, see Figure 1 The present invention provides a method for evaluating the strength of cement-water glass gel based on big data, which includes the following steps:

[0055] Step S1: Data acquisition, collecting experimental formula records, time-series characteristic data, microscopic characterization data, and intensity labels at corresponding times from historical samples;

[0056] Step S2: Feature mapping, extract differentiable topological and frequency domain features from microscopic images and diffraction patterns, obtain continuous persistent images with physical weights and discretize them to construct microscopic feature vectors;

[0057] Step S3: Latent variable estimation. A low-dimensional state-space model with the degree of gel reaction as the latent variable is established. The kinetic parameters are mapped using microscopic and temporal characteristic data, and prediction and updating are performed through extended Kalman filtering to obtain virtual measurements.

[0058] Step S4: Hybrid model training. Construct a hybrid predictor consisting of physical priors and data-driven correction terms. The physical priors are given by microscopic feature vectors and virtual measurements. Fitting error, physical residuals, and monotonicity penalties are added during joint training.

[0059] Step S5: Interpretability output provides confidence intervals and local feature contributions for the prediction results. Deep ensemble and heterogeneous variance regression are used to simultaneously model the mean and data noise, and then local linear approximation is used to make explicit feature contributions.

[0060] Step S6: Intensity assessment. Perform feature mapping and latent variable estimation on the sample to be assessed. Input the micro-feature vector, time-series feature data and virtual measurement into the hybrid model, and output the intensity prediction value, confidence interval and local feature contribution in real time.

[0061] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the data acquisition includes collecting historical experimental and formulation data, time-series characteristic data, microscopic characterization data, and strength label data of cement-water glass gel. The experimental and formulation data includes: unique identifiers. Water-to-glue ratio Mass fraction of silicon dioxide Sodium oxide mass fraction Types and mass fractions of admixtures Initial mixing temperature Stirring time The time-series feature data includes: temperature ,humidity pH of the solution conductivity The microscopic characterization data includes: electron micrographs. X-ray diffraction pattern Fourier transform infrared spectrum Porosity Median diameter of particles The strength label data refers to the compressive strength of cement-water glass gel. .

[0062] Example 3, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S2, the feature mapping constructs a differentiable topological and frequency domain feature mapping for the microscopic characterization data to obtain a continuous persistent image vector. At the same time, it extracts physical quantitative indicators of peak position and peak intensity from the diffraction spectrum. Specifically, it includes the following steps:

[0063] Step S21: Construct a morphological subset, extract a binary sublayer from the electron micrograph, and construct an upstream filtering sequence. The threshold set is... Generate morphological subsets , means as follows:

[0064] ;

[0065] in, This represents the grayscale value of pixel p in an electron micrograph. Indicates by The threshold sequence obtained by the method Represents a binary subset of thresholds;

[0066] Step S22: Obtain the persistent image, calculate the persistent coherence information generated by morphological filtering, and only take 0-dimensional and 1-dimensional data: the resulting persistent image is represented as follows:

[0067] ;

[0068] in, This represents the set of persistent homology points obtained by morphological filtering. The threshold pair representing the k-th topological feature. Indicates the duration threshold, Indicates the termination threshold;

[0069] Step S23: Persistent image transformation, performing a physically weighted transformation on the persistent image, represented as: ;in, This represents the persistent image after undergoing a physical weight transformation. Represents a two-dimensional Gaussian kernel function. Denotes the Gaussian kernel covariance matrix. Indicates the transpose symbol. This represents an exponential function with the natural constant as its base. The physical weighting function is defined as follows: ;in, This represents linear normalization, normalized to... , , and Represents the physical weighted hyperparameters;

[0070] Step S24: Construct micro-feature vectors, and... Vectorized features are obtained by discretization on a fixed grid G. Then, spectral energy characteristics, including peak area, are extracted from the Fourier transform infrared spectrum. With half-peak full width Construct spectral vectors ; Obtain the final microscopic feature vector ;in, This indicates the concatenation of vectors. This represents a discretized persistent image vector of length p; and They represent the first Peak area With half the full width of the peak; This represents the total number of peaks in the Fourier transform infrared spectrum. and They represent the first The peak area and half-peak width, This represents the final microscopic eigenvector.

[0071] By performing the above operations, this scheme addresses the problems of discontinuous and non-differentiable features, and difficulty in associating microscopic topological information with physical formulations in traditional feature extraction methods. It transforms persistent cohomological information into a continuous persistent image with physical weights and discretizes it into a vector. At the same time, it extracts physical quantitative indicators of peak area and half-peak width from the diffraction pattern and constructs a microscopic feature vector coupled with formulation and pore size, thereby obtaining a differentiable, physically-aware feature table that can be used for end-to-end training.

[0072] Example 4, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S3, the latent variable estimation establishes a low-dimensional physical-driven state-space model, treating the degree of gelation as a latent variable that cannot be directly and completely observed. Utilizing time series characteristics and microscopic features The latent variables are estimated posteriorly using extended Kalman filtering to obtain the estimated virtual measurement. Specifically, it includes the following steps:

[0073] Step S31: Construct a dynamic model, using a generalized model with temperature and humidity coupling. The model maps microscopic features to dynamic constants through parameters, as follows:

[0074] ;

[0075] in, The degree of gel reaction is represented by a hidden variable, with an initial value of [value missing]. ; and These represent the reaction rate constant and the reaction order, respectively, which are mapped from microscopic features to the environment; and This represents the learnable dynamic weight parameters. and These represent the linear sensitivity coefficients of temperature and relative humidity to the rate, respectively. Indicates reference temperature. Indicates reference humidity. Indicates the initial reaction constant;

[0076] Step S32: Discretization. Perform forward Euler discretization on the latent variables and construct the observation vector, as follows:

[0077] ;

[0078] in, Indicates the discrete time step. Indicates at time Discretized latent variables, Gaussian white noise represents the random perturbation in the dynamic model. Represents the observation vector. This represents the reaction rate constant at time t. Indicates the reaction order at time t;

[0079] Step S33: Extended Kalman Filtering. The latent variables are estimated posteriorly using extended Kalman filtering to obtain the virtual measurement. ;

[0080] First, combine the time step In addition to the dynamic rate information, we can advance to obtain the posterior estimated latent variables at the current time: ;

[0081] in, Let represent the latent variables of the prior estimate at time t. Indicates at time The latent variable estimated posteriorly is the filter output from the previous time step. ; Indicates time The reaction rate constant, Indicates time The reaction stage;

[0082] Then calculate the covariance matrix: ;

[0083] in, Let represent the prior covariance matrix, and let represent the variance of the prior covariance matrix. Uncertainty, This represents the posterior covariance of the previous time step. Represents the process noise covariance. Let Jacobian matrix represent the state transition matrix, specifically the derivative of the state update mapping with respect to the state at point [0, 1]. The evaluation at this point is expressed as: ;

[0084] Finally, perform a complete update: ;

[0085] in, This represents the Kalman gain, used to map observation information into a weight matrix for state updates. The Jacobian matrix representing the observation mapping with respect to the state is taken at the prediction point. Location: ; Represents the observation noise covariance matrix; Let represent the posterior estimate at time t, where is the time t. The real-time virtual measurement estimate incorporates observation information; Indicates time The actual observed vector, This represents an observational mapping based on the predicted state and microscopic features; Indicates the posterior covariance. Represents the identity matrix.

[0086] By performing the above operations, this scheme addresses the problems of traditional kinetic modeling and empirical fitting methods, such as difficulty in real-time estimation of gel reactivity, difficulty in coupling microscopic features with environmental disturbances, and lack of uncertainty quantification. It utilizes a generalized... Using dynamics as the physical framework, microscopic feature vectors are mapped to dynamic parameters and temperature-humidity coupling is introduced. Combined with extended Kalman filtering, the latent variable gelation degree is advanced a priori and updated by observation. This not only realizes real-time virtual measurement estimation, but also provides posterior covariance for uncertainty propagation.

[0087] Example 5, see Figure 1 and Figure 5 This embodiment is based on the above embodiment. In step S4, the hybrid model training constructs a physically regularized hybrid predictor, which incorporates the micro-feature vectors. Time series characteristics and virtual measurement As input, predict the target compressive strength. Specifically, it includes the following steps:

[0088] Step S41: Construct a priori physical model of intensity, using microscopic feature vectors Determined parameters and time-varying virtual measurements As input, the output is a predicted intensity curve over time based on physical experience, as shown below:

[0089] ;

[0090] in, This represents the compressive strength predicted a priori by physics. This represents the set of parameters of the physical prior model. Represents the microscopic eigenvectors The final intensity upper bound function obtained by mapping, This represents a small network with a three-layer perceptron structure. The scale parameter, which represents the microscopic characteristics, controls how quickly the intensity increases with reactivity. The power exponent, determined by microscopic characteristics, controls... The nonlinear delay effect;

[0091] Step S42: Define a data-driven correction term, which is a differentiable neural network that accepts microscopic features, temporal features, and dummy measurements as inputs, and outputs a correction amount for the physical prior. ;

[0092] in, This represents the intensity correction term of the neural network output. This represents a parameterized neural network mapper, with the parameter set as follows: , Represents a time-series feature data vector;

[0093] Step S43: Final prediction. The physical prior prediction is added to the data-driven correction term to obtain the final mixed model strength prediction value, which is used as the model output during training and inference. ;in, This represents the intensity prediction by the model at time t;

[0094] Step S44: Training loss, model loss includes data fitting error, physical residual penalty, and monotonicity penalty: ;

[0095] in, Indicates the overall training loss. Represents the number of training samples. Indicates the training sample index. Denotes the set of observation times for sample i. Represents the weighting coefficients at sample time points. The model represents the samples At any moment The prediction Indicates sample At any moment The measured compressive strength, Indicates the weight of the physical residual term. Indicates the monotonicity penalty weight. Indicates short-run differences, Set to 0.5 hours. This means that a penalty is only applied when the forecast shows a decline; the loss is a one-sided constraint, and the physical residual is... Defined as the consistency residual between the intensity model and the dynamic variables: .

[0096] By performing the above operations, this scheme addresses the problems of underfitting in traditional pure physics models and losing physical constraints in pure data-driven models. It constructs a hybrid predictor consisting of a physical prior strength curve and a correction term output by a differentiable neural network. The predictor is trained by combining data fitting error, physical residual consistency penalty, and monotonicity constraint. This makes the prediction both follow the consistency of dynamic physics and has the flexible correction capability of data-driven methods, thereby improving generalization and suppressing non-physical interpretations.

[0097] Example 6, see Figure 1 and Figure 6 This embodiment is based on the above embodiment. In step S5, the interpretability output provides confidence intervals and local feature contributions for the prediction results. It uses deep ensemble and heterogeneous variance regression to simultaneously model the mean and data noise, and then uses local linear approximation to make explicit feature contributions. Specifically, it includes the following steps:

[0098] Step S51: Model integration, integrating the models The mean and variance outputs, integrated from independently trained datasets. Obtained from different initialization models with the same structure, each model outputs: , The total variance is decomposed into two parts: observable noise and parameter uncertainty, as shown below:

[0099] ;

[0100] in, Indicates the number of ensemble models. Let represent the mean prediction of the m-th model at time t. Let represent the variance of the m-th model output, and represent the observable noise estimate; This indicates the predicted mean after integration. This represents the total variance after integration;

[0101] Step S52: Generate confidence intervals, providing confidence intervals based on the approximate Gaussian assumption. The width of the confidence interval is determined by the total standard deviation, as shown below:

[0102] ;

[0103] in, Indicates confidence level The closed interval, Indicates the standard quantile. The total standard deviation;

[0104] Step S53: Generate contribution values ​​for the input vector. The estimated feature contribution is:

[0105] ;

[0106] in, Indicates the input number Each component at time... Contribution to prediction This represents a combined input vector, including micro-features, temporal features, and virtual measurements; Representing vectors The One component; This represents the baseline value of the o-th feature, which is the mean of the training set; The coefficients represent the coefficients of the local linear regression, which are obtained by weighted fitting of the linear model at the prediction points. The kernel function used is the Gaussian kernel function.

[0107] Step S54: Uncertainty decomposition. The total variance is decomposed into two parts. The first part is the process uncertainty propagated from the uncertainty of the dynamic latent variables to the intensity. The second part is the observation uncertainty introduced by the data-driven part and observation noise, as shown below:

[0108] ;

[0109] in, The variance component representing the process model originates from the extended Kalman filter's application to the latent variables. The propagation of uncertainty This indicates the observation uncertainty in the second part;

[0110] Step S55: Chain propagation, based on the chain rule, extends the posterior covariance of the extended Kalman filter. The uncertainty mapped to the intensity prediction through the physical prior model is expressed as follows:

[0111] ;

[0112] in, Indicates the current moment, using the current time. and Regarding physical priors The partial derivative of .

[0113] Example 7, see Figure 1 This embodiment is based on the above embodiment. In step S6, the strength assessment involves obtaining the cement-water glass gel data to be assessed, inputting it into the hybrid model, using the hybrid model to predict the strength of the cement-water glass gel in real time as the assessment strength, and generating interpretable output.

[0114] Example 8, see Figure 1 and Figure 2Based on the above embodiments, the cement-water glass gel strength evaluation system based on big data provided by the present invention includes a data acquisition module, a feature mapping module, a latent variable estimation module, a hybrid model training module, an interpretable output module, and a strength evaluation module.

[0115] The data acquisition module collects historical experimental and formulation data, time-series characteristic data, microscopic characterization data, and strength label data of cement-water glass gel, and sends the data to the feature mapping module, latent variable estimation module, hybrid model training module, interpretable output module, and strength assessment module.

[0116] The feature mapping module receives data sent by the data acquisition module, constructs differentiable topological and frequency domain feature mappings for the microscopic characterization data, obtains continuous persistent image vectors, extracts physical quantitative indicators of peak position and peak intensity from the diffraction spectrum, and sends the data to the latent variable estimation module.

[0117] The latent variable estimation module receives data sent by the data acquisition module and the feature mapping module, establishes a low-dimensional physical-driven state-space model, treats the degree of gelation as a latent variable that cannot be directly and completely observed, uses temporal and micro-features, performs posterior estimation of the latent variable through extended Kalman filtering to obtain the estimated virtual measurement, and sends the data to the hybrid model training module.

[0118] The hybrid model training module receives data from the data acquisition module and the latent variable estimation module, constructs a physically regularized hybrid predictor, takes micro-feature vectors, temporal features and virtual measurements as inputs, predicts the target compression strength, and sends the data to the interpretable output module.

[0119] The interpretable output module receives data sent by the data acquisition module and the hybrid model training module, provides confidence intervals and local feature contributions for the prediction results, uses deep ensemble and heterogeneous variance regression to simultaneously model the mean and data noise, then uses local linear approximation to make explicit feature contributions, and sends the data to the intensity assessment module.

[0120] The strength assessment module receives data from the data acquisition module and the interpretability output module, obtains the cement-water glass gel data to be assessed, inputs it into the hybrid model, uses the hybrid model to predict the strength of the cement-water glass gel in real time as the assessment strength, and generates interpretable output.

[0121] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0123] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for evaluating the strength of cement-water glass gel based on big data, characterized in that, The method includes the following steps: Step S1: Data acquisition, collecting experimental formula records, time-series characteristic data, microscopic characterization data, and intensity labels at corresponding times from historical samples; Step S2: Feature mapping, extract differentiable topological and frequency domain features from micro-representation data, obtain continuous persistent images with physical weights and discretize them to construct micro-feature vectors; Step S3: Latent variable estimation. A low-dimensional state-space model with the degree of gel reaction as the latent variable is established. The kinetic parameters are mapped using microscopic and temporal characteristic data, and prediction and updating are performed through extended Kalman filtering to obtain virtual measurements. Step S4: Hybrid model training. Construct a hybrid predictor consisting of physical priors and data-driven correction terms. The physical priors are given by microscopic feature vectors and virtual measurements. Fitting error, physical residuals, and monotonicity penalties are added during joint training. Step S5: Interpretability output provides confidence intervals and local feature contributions for the prediction results. Deep ensemble and heterogeneous variance regression are used to simultaneously model the mean and data noise, and then local linear approximation is used to make explicit feature contributions. Step S6: Intensity assessment. Perform feature mapping and latent variable estimation on the sample to be assessed. Input the micro-feature vector, time-series feature data and virtual measurement into the hybrid model, and output the intensity prediction value, confidence interval and local feature contribution in real time.

2. The method for evaluating the strength of cement-water glass gel based on big data according to claim 1, characterized in that: In step S1, the data acquisition includes collecting historical experimental and formulation data, time-series characteristic data, microscopic characterization data, and strength label data for cement-water glass gel. The experimental and formulation data includes: unique identifier, water-cement ratio, silica mass fraction, sodium oxide mass fraction, type and mass fraction of admixtures, initial mixing temperature, and stirring time. The time-series characteristic data includes: temperature, humidity, solution pH, and conductivity. The microscopic characterization data includes: electron micrographs, X-ray diffraction patterns, Fourier transform infrared spectroscopy, porosity, and median particle diameter. The strength label data refers to the compressive strength of the cement-water glass gel.

3. The method for evaluating the strength of cement-water glass gel based on big data according to claim 1, characterized in that: In step S2, the feature mapping constructs differentiable topological and frequency domain feature maps for the microscopic representation data to obtain continuous persistent image vectors, while simultaneously extracting physical quantification indicators of peak position and peak intensity. Specifically, this includes the following steps: Step S21: Construct a morphological subset by extracting a binary sublayer from the electron micrograph and constructing an upstream filtering sequence to generate a morphological subset; Step S22: Obtain the persistent image, calculate the persistent coherence information generated by morphological filtering, and only take 0 and 1 dimensions: obtain the persistent image; Step S23: Persistent image transformation, performing a physically weighted transformation on the persistent image; Step S24: Construct micro-feature vectors by discretizing the persistent image after physical weight transformation on a fixed grid to obtain vectorized features; then extract spectral energy features from the Fourier infrared spectrum; and construct the final micro-feature vectors.

4. The method for evaluating the strength of cement-water glass gel based on big data according to claim 1, characterized in that: In step S3, the latent variable estimation involves establishing a low-dimensional, physically driven state-space model, treating the degree of gelation as a latent variable that cannot be directly and completely observed, and using temporal and microscopic features, performing posterior estimation of the latent variable through extended Kalman filtering to obtain the estimated virtual measurement. Specifically, this includes the following steps: Step S31: Construct a dynamic model, using a generalized model with temperature and humidity coupling. The model maps microscopic features to dynamic constants through parameters; Step S32: Discretization, perform forward Euler discretization on the latent variables, and construct the observation vector at the same time; Step S33: Extended Kalman Filtering. The extended Kalman filter is used to estimate the latent variables posteriorly. First, the latent variables posteriorly are estimated at the current time by combining the time step and dynamic rate information. Then, the covariance matrix is ​​calculated. Finally, the overall update is performed.

5. The method for evaluating the strength of cement-water glass gel based on big data according to claim 1, characterized in that: In step S4, the hybrid model training constructs a physically regularized hybrid predictor, taking microscopic feature vectors, temporal features, and virtual measurements as inputs to predict the target compression strength. This specifically includes the following steps: Step S41: Construct a prior physical model of intensity, taking the parameters determined by the microscopic eigenvectors and the time-varying virtual measurement as inputs, and outputting a predicted intensity curve over time based on physical experience. Step S42: Define a data-driven correction term, which is a differentiable neural network that accepts microscopic features, temporal features and virtual measurements as inputs and outputs a correction amount for physical priors; Step S43: Final prediction. Add the physical prior prediction to the data-driven correction term to obtain the final mixed model strength prediction value, which is used as the model output during training and inference. Step S44: Training loss, which includes data fitting error, physical residual penalty, and monotonicity penalty.

6. The method for evaluating the strength of cement-water glass gel based on big data according to claim 1, characterized in that: In step S5, the interpretability output provides confidence intervals and local feature contributions for the prediction results. It employs deep ensemble and heterogeneous variance regression to simultaneously model the mean and data noise, and then uses local linear approximation to make explicit feature contributions. Specifically, it includes the following steps: Step S51: Model integration, integration The mean and variance outputs of each model are ensembled through independent training. The total variance is obtained from different initialization models with the same structure; the total variance is decomposed into two parts: observable noise and parameter uncertainty. Step S52: Generate confidence intervals, providing confidence intervals based on the approximate Gaussian assumption. The width of the confidence interval is determined by the total standard deviation. Step S53: Generate contribution values ​​and estimate feature contributions for the input vector; Step S54: Uncertainty decomposition, the total variance is decomposed into two parts. The first part is the process uncertainty that propagates from the uncertainty of the dynamic hidden variables to the intensity, and the second part is the observation uncertainty introduced by the data-driven part and observation noise. Step S55: Chain propagation, based on the chain rule, maps the extended Kalman filter posterior covariance to the uncertainty of intensity prediction through the physical prior model.

7. The method for evaluating the strength of cement-water glass gel based on big data according to claim 1, characterized in that: In step S6, the strength assessment involves acquiring the cement-water glass gel data to be assessed, inputting it into a hybrid model, using the hybrid model to predict the strength of the cement-water glass gel in real time as the assessment strength, and generating interpretable output.

8. A cement-water glass gel strength assessment system based on big data, used to implement the cement-water glass gel strength assessment method based on big data as described in any one of claims 1-7, characterized in that: It includes a data acquisition module, a feature mapping module, a latent variable estimation module, a hybrid model training module, an interpretable output module, and an intensity evaluation module.

9. The cement-water glass gel strength evaluation system based on big data according to claim 8, characterized in that: The data acquisition module collects historical experimental and formulation data, time-series characteristic data, microscopic characterization data, and strength label data of cement-water glass gel, and sends the data to the feature mapping module, latent variable estimation module, hybrid model training module, interpretable output module, and strength assessment module. The feature mapping module receives data sent by the data acquisition module, constructs differentiable topological and frequency domain feature mappings for the microscopic characterization data, obtains continuous persistent image vectors, extracts physical quantification indicators of peak position and peak intensity, and sends the data to the latent variable estimation module. The latent variable estimation module receives data sent by the data acquisition module and the feature mapping module, establishes a low-dimensional physical-driven state-space model, treats the degree of gelation as a latent variable that cannot be directly and completely observed, uses temporal and micro-features, performs posterior estimation of the latent state through extended Kalman filtering to obtain the estimated virtual measurement, and sends the data to the hybrid model training module. The hybrid model training module receives data from the data acquisition module and the latent variable estimation module, constructs a physically regularized hybrid predictor, takes micro-feature vectors, temporal features and virtual measurements as inputs, predicts the target compression strength, and sends the data to the interpretable output module. The interpretable output module receives data sent by the data acquisition module and the hybrid model training module, provides confidence intervals and local feature contributions for the prediction results, uses deep ensemble and heterogeneous variance regression to simultaneously model the mean and data noise, then uses local linear approximation to make explicit feature contributions, and sends the data to the intensity assessment module. The strength assessment module receives data from the data acquisition module and the interpretability output module, obtains the cement-water glass gel data to be assessed, inputs it into the hybrid model, uses the hybrid model to predict the strength of the cement-water glass gel in real time as the assessment strength, and generates interpretable output.

Images