Cement hydration heat prediction method and system based on non-weight sharing and variational mechanism

Abstract

The application provides a cement hydration heat prediction method and system based on a non-weight sharing and variational mechanism, and relates to the technical field of cement hydration heat prediction.The method comprises the following steps: obtaining hydration heat time series data and attribute parameters, generating an enhanced sequence through preprocessing, feature fusion and time coding, constructing a sample through a sliding window, inputting a prediction model composed of a non-weight sharing convolution module, a variational convolution module and a full connection module, and obtaining a prediction result.The non-weight sharing convolution module dynamically generates a convolution kernel to adapt to the feature changes in each hydration stage.The variational convolution module introduces probability regularization and KL constraint to enhance the robustness of the model.The application solves the problems of weak dynamic law capturing ability of non-stationary time series and poor generalization ability under small sample high noise in the traditional method, and realizes more accurate, stable and generalization-capable prediction of cement hydration heat.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cement hydration heat prediction, and particularly relates to a cement hydration heat prediction method and system based on non-weight sharing and variational mechanism. BACKGROUND

[0002] As the core cementitious material of modern infrastructure construction, the hydration process of cement is not a simple single dissolution or single chemical reaction, but a complex physical and chemical process covering dissolution, nucleation, crystal growth and diffusion control in multiple phases and multiple stages. Hydration heat evolution time series data records the thermodynamic fingerprint of this process in detail, from the chemical silence of the induction period, the explosive growth of the acceleration period to the diffusion control of the deceleration period, accurately reflecting the construction process of cement strength. Precise prediction of the whole process of cement hydration heat has high engineering value and social benefits, especially in the aspects of mass concrete temperature control and crack prevention and low-carbon material research and development. In dam, bridge pier and nuclear power engineering, if the accumulation of hydration heat cannot be predicted and controlled in advance, irreversible temperature cracks will occur in the core of the structure, which will seriously threaten the safety of the project. By using a prediction model to replace the traditional long-period calorimetric experiment, the research and development cycle of new low-clinker and high-dosage auxiliary cementitious material cement formulations can be greatly shortened, thereby accelerating the research and development of low-carbon materials.

[0003] However, the precise prediction of cement hydration heat faces more complex challenges than the prediction of general chemical reaction heat. Unlike the monotonic decay or regular exponential change of conventional inorganic salt dissolution or simple exothermic reaction, cement hydration reaction has significant nonlinear, non-stationary and multimodal characteristics. The hydration rates of C3S, C2S, C3A and C4AF, the four minerals in cement clinker, differ greatly and compete with each other, resulting in a complex wave peak pattern of hydration heat curve, such as the main peak and the possible ettringite conversion secondary peak after the main peak. This phase change type mutation of "induction period-acceleration period" causes the reaction kinetics mechanism to shift essentially on the time axis, and the data rule of the previous stage cannot be directly deduced to the next stage, further increasing the difficulty of cement hydration heat prediction.

[0004] Current prediction methods often rely on semi-empirical models based on physicochemical mechanisms (such as Arrhenius's law or Avrami's equation) that are idealized single-component assumptions or simplistic nucleation and growth theories, making it difficult to accurately describe the complex mineral competition and time-dependent reactions in modern multi-component cementitious systems. Furthermore, the parameters of these models are highly dependent on expensive experimental calibration and lack generalization ability for new formulations. Data-driven methods based on traditional machine learning or standard deep learning (such as CNNs and RNNs), while avoiding complex mechanism derivations, generally employ static weight-sharing mechanisms in their core architecture. They use fixed convolutional kernels or weight parameters for feature extraction across the entire timeline, forcibly fitting the same set of feature patterns to various reaction stages with drastically different physical mechanisms. This makes it difficult for the model to capture the nonlinear dynamics of key inflection points, thus affecting prediction accuracy. Moreover, due to the long and costly testing cycles of hydration heat, the scarcity of training sample data, and susceptibility to environmental noise, existing deterministic models are prone to overfitting under small sample conditions. Furthermore, their robustness and predictive stability often decrease significantly when faced with formulation fine-tuning or environmental fluctuations in actual working conditions. Summary of the Invention

[0005] To address at least one of the technical problems mentioned above, this invention constructs a time-series data prediction model that includes a non-weight-sharing convolution module, a variational convolution module, and a fully connected module. This model solves the shortcomings of traditional models, such as difficulty in flexibly capturing the dynamic evolution of non-stationary time-series data during cement hydration, insufficient robustness, and susceptibility to overfitting.

[0006] To achieve the above objectives, the first aspect of the present invention provides a method for predicting the heat of hydration of cement based on non-weight sharing and variational mechanisms, comprising:

[0007] Obtain historical time-series data of hydration heat and cement property parameters of cement samples; wherein, the cement property parameters include any one or more of the following: chemical composition, physical properties, water-cement ratio, and reaction temperature;

[0008] The historical time series data of hydration heat and cement property parameters are preprocessed and feature fused to generate an enhanced feature sequence containing material properties and time stage information.

[0009] The enhanced feature sequence is processed using a sliding window to construct model input samples, where each input sample corresponds to one or more future time steps of hydration heat value as a prediction label;

[0010] The input samples are fed into a hydration heat prediction model that sequentially includes a non-weight-shared convolution module, a variational convolution module, and a fully connected module to obtain prediction results. Specifically, the non-weight-shared convolution module dynamically generates convolution kernel parameters based on the input samples and uses these dynamic kernels to extract features from the input data, outputting time-sensitive features. The variational convolution module performs probabilistic encoding on the time-sensitive features, introducing randomness through reparameterization techniques, and outputs a probabilistically regularized feature representation. The fully connected module integrates and performs nonlinear mapping on the probabilistically regularized feature representation, outputting the final hydration heat prediction value.

[0011] Furthermore, the preprocessing and feature fusion include:

[0012] The historical time series data of hydration heat and cement property parameters were standardized respectively.

[0013] The standardized time-series data is concatenated with attribute parameters to form fused data;

[0014] Temporal location codes are embedded into the fused data to generate the enhanced feature sequence.

[0015] Furthermore, the operation of the non-weight-sharing convolution module to extract time-sensitive features specifically includes:

[0016] Dynamically changing convolution kernel parameters are generated based on the input data through convolution operations;

[0017] Expand the input data into local sequence blocks along the time dimension;

[0018] The expanded input data is then used to perform window-by-window calculations with the dynamically generated convolution kernel parameters to obtain time-sensitive features.

[0019] Furthermore, the variational convolution module performs probabilistic encoding and introduces randomness, specifically including:

[0020] Perform two independent convolution operations on the input features, and output the mean parameters of the latent variable distributions respectively. And log-variance parameter The standard deviation is calculated using the logarithmic variance parameter. ;

[0021] Random noise sampled from a standard normal distribution ;

[0022] Calculate latent variable representations based on reparameterization techniques. .

[0023] Furthermore, the training process of the hydration heat prediction model includes:

[0024] Prepare a training dataset containing input samples and their corresponding predicted labels;

[0025] The training dataset is input into the hydration heat prediction model for forward propagation, and the prediction loss is calculated.

[0026] The predicted loss includes Divergence loss, using The divergence constrains the distribution of latent variables in the variational convolution module. The formula for calculating the divergence constraint term is:

[0027] ;

[0028] Based on the predicted loss, the model parameters are updated using the backpropagation algorithm until the model converges.

[0029] Furthermore, the method for generating the embedded time position code is as follows: [The text abruptly ends here, so the translation stops.] Each time step has a position encoding vector. The Wei and Di +1 dimension is calculated using the following formula:

[0030] ;

[0031] ;

[0032] in, For dimensional indexing, The total dimension for location encoding.

[0033] A second aspect of the present invention provides a cement hydration heat prediction system based on non-weight sharing and variational mechanisms, comprising:

[0034] The data acquisition unit is used to acquire the historical time-series data of the heat of hydration of cement samples and cement property parameters; wherein, the cement property parameters include any one or more of the following: chemical composition, physical properties, water-cement ratio, and reaction temperature;

[0035] The feature enhancement unit is used to preprocess and fuse the hydration heat history time series data and cement property parameters to generate an enhanced feature sequence containing material properties and time stage information.

[0036] The sample construction unit is used to process the enhanced feature sequence using a sliding window to construct model input samples, wherein each input sample corresponds to one or more future time steps of hydration heat value as a prediction label.

[0037] The prediction unit is used to input the input samples into a hydration heat prediction model that sequentially includes a non-weight-shared convolution module, a variational convolution module, and a fully connected module to obtain the prediction result. Specifically, the non-weight-shared convolution module dynamically generates convolution kernel parameters based on the input samples and uses these dynamic convolution kernels to extract features from the input data, outputting time-sensitive features. The variational convolution module performs probabilistic encoding on the time-sensitive features, introducing randomness through reparameterization techniques, and outputs a probabilistically regularized feature representation. The fully connected module integrates and performs nonlinear mapping on the probabilistically regularized feature representation, outputting the final predicted hydration heat value.

[0038] A fifth aspect of the present invention provides a computer program product comprising software code, wherein the program in the software code performs the steps of the cement hydration heat prediction method based on non-weight sharing and variational mechanism as described in the first aspect of the present invention.

[0039] Compared with existing technologies, the cement hydration heat prediction method and system provided by this invention based on non-weight sharing and variational mechanisms have the following beneficial effects:

[0040] (1) This invention is based on a non-weight-sharing convolution module, which dynamically generates convolution kernel parameters that change with time steps based on the input data, replacing the traditional static convolution kernel. This module enables the feature extractor to adapt to time, adaptively employing different feature extraction modes for different stages of the hydration process (such as the stable induction period and the rapidly changing acceleration period). This mechanism breaks the implicit assumption of temporal stationarity in traditional models, thereby achieving precise capture of key turning points and stage features in hydration heat time series data, significantly improving the timeliness and accuracy of prediction.

[0041] (2) This invention, based on a variational convolution module, transforms deterministic weight learning into learning the probability distribution of the weights. Random sampling (reparameterization technique) is introduced during feature extraction, shifting from a fixed numerical mapping to inferring the probability distribution of the weights, which is equivalent to implicit data augmentation in the feature space. This mechanism enhances the model's tolerance to noise in the training data and small fluctuations in the input, effectively suppressing overfitting. Furthermore, this module optimizes the fully connected structure in traditional variational operations, introducing a probability distribution within the "sliding window" convolution operation to avoid disrupting the inherent temporal topology of the hydration heat data, thereby improving the model's generalization ability while maintaining its robustness.

[0042] (3) In order to achieve a stable mapping from deep features that have undergone dynamic and probabilistic processing to the final hydration heat prediction value, this invention is based on a fully connected module. It receives high-order abstract features processed by the aforementioned module and maps the high-dimensional and complex feature representation to a low-dimensional prediction output space through multi-layer nonlinear transformation and integration. This design ensures the efficient use of feature information and the reliable output of the final prediction result, and completes the end-to-end modeling from complex time series patterns to specific quantitative prediction values. Attached Figure Description

[0043] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0044] Figure 1 This is a flowchart of the cement hydration heat prediction method based on non-weight sharing and variational mechanism provided in Embodiment 1 of the present invention;

[0045] Figure 2 This is a diagram of the prediction model architecture provided in Embodiment 1 of the present invention;

[0046] Figure 3 This is a schematic diagram of the prediction model training provided in Embodiment 1 of the present invention;

[0047] Figure 4 This is an architecture diagram of the non-weight-sharing convolution module provided in Embodiment 1 of the present invention;

[0048] Figure 5 This is a diagram of the variational convolution module architecture provided in Embodiment 1 of the present invention;

[0049] Figure 6 This is a diagram of the fully connected module architecture provided in Embodiment 1 of the present invention;

[0050] Figure 7 This is an architecture diagram of the cement hydration heat prediction system based on non-weight sharing and variational mechanism provided in Embodiment 2 of the present invention. Detailed Implementation

[0051] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0052] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0053] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0054] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0055] Example 1

[0056] like Figure 1 This embodiment provides a method for predicting the heat of hydration of cement based on non-weight sharing and variational mechanisms, including:

[0057] Obtain historical time-series data of hydration heat and cement property parameters of cement samples; wherein, the cement property parameters include any one or more of the following: chemical composition, physical properties, water-cement ratio, and reaction temperature;

[0058] The historical time series data of hydration heat and cement property parameters are preprocessed and feature fused to generate an enhanced feature sequence containing material properties and time stage information.

[0059] The enhanced feature sequence is processed using a sliding window to construct model input samples, where each input sample corresponds to one or more future time steps of hydration heat value as a prediction label;

[0060] like Figure 2 The input samples are fed into a hydration heat prediction model that sequentially includes a non-weight-shared convolution module, a variational convolution module, and a fully connected module to obtain prediction results. Specifically, the non-weight-shared convolution module dynamically generates convolution kernel parameters based on the input samples and uses these dynamic kernels to extract features from the input data, outputting time-sensitive features. The variational convolution module performs probabilistic encoding on the time-sensitive features, introducing randomness through reparameterization techniques, and outputs a probabilistically regularized feature representation. The fully connected module integrates and performs nonlinear mapping on the probabilistically regularized feature representation, outputting the final hydration heat prediction value.

[0061] This method constructs and applies a hydration heat prediction model composed of a non-weight-shared convolutional module, a variational convolutional module, and a fully connected module. The model structure is designed to address the non-stationary and non-linear characteristics of cement hydration heat. Through dynamic generation of convolutional kernels, probabilistic feature extraction, and multi-layer mapping, it achieves accurate prediction of complex hydration processes. First, the non-weight-shared convolutional module dynamically adapts to the feature extraction requirements of different hydration stages, capturing temporal dynamic changes. Second, the variational convolutional module introduces probabilistic regularization, improving the model's robustness to noise and small samples. Finally, the fully connected module integrates features and outputs predicted values, achieving stable end-to-end prediction.

[0062] Specifically, the preprocessing and feature fusion include:

[0063] The historical time series data of hydration heat and cement property parameters were standardized respectively.

[0064] The standardized time-series data is concatenated with attribute parameters to form fused data;

[0065] Temporal location codes are embedded into the fused data to generate the enhanced feature sequence.

[0066] Standardization eliminates the influence of different dimensions, splices and fuses multi-source features, and temporal location encoding introduces stage information of the hydration process, enabling the model to simultaneously perceive material properties and reaction process, providing rich and structured input for subsequent dynamic feature extraction.

[0067] Specifically, such as Figure 3 The training process of the hydration heat prediction model includes:

[0068] Prepare a training dataset containing input samples and their corresponding predicted labels;

[0069] The training dataset is input into the hydration heat prediction model for forward propagation, and the prediction loss is calculated.

[0070] The predicted loss includes Divergence loss, using The divergence constrains the distribution of latent variables in the variational convolution module. The formula for calculating the divergence constraint term is:

[0071] ;

[0072] Based on the predicted loss, the model parameters are updated using the backpropagation algorithm until the model converges.

[0073] The KL divergence measures the difference between the learned distribution and the standard normal distribution. By minimizing this loss, the feature distribution becomes smoother and more compact, avoiding overfitting to training noise, thereby enhancing the generalization ability and stability of the model when facing new data.

[0074] Specifically, as Figure 4 , the operation of the non-weight-sharing convolution module for extracting time-sensitive features specifically includes:

[0075] Generating dynamically changing convolution kernel parameters through convolution operations based on the input data;

[0076] Unfolding the input data in the time dimension into local sequence blocks;

[0077] Performing window-by-window calculations on the unfolded input data and the dynamically generated convolution kernel parameters to obtain time-sensitive features.

[0078] Traditional convolution uses a fixed kernel and is difficult to adapt to the feature differences in different stages of the hydration process (such as the induction period and the acceleration period); while this module generates convolution kernels in real time based on the input data, enabling the feature extractor to "change with time", thereby more precisely capturing the temporal dynamics and solving the technical problem of the insufficient fitting ability of traditional models for non-stationary time series.

[0079] Specifically, as Figure 5 , the operation of the variational convolution module for probability encoding and introducing randomness specifically includes:

[0080] Performing two independent convolution operations on the input features, respectively outputting the mean parameter of the latent variable distribution and the logarithmic variance parameter , and calculating the standard deviation through the logarithmic variance parameter;

[0081] Sampling random noise from the standard normal distribution;

[0082] Calculating the latent variable representation based on the reparameterization trick.

[0083] Predicting the mean and variance respectively through two independent convolutions, introducing random noise to generate the latent variable representation, forcing the model to learn a smooth feature distribution rather than memorizing noisy data; and then making the distribution tend to the standard normal through the KL divergence constraint, thereby enhancing the robustness of the model to input fluctuations and noise and effectively alleviating the problem of overfitting in small samples.

[0084] Specifically, the method for generating the embedded time position encoding is: for the th time step in the fused data, the th dimension of its position encoding vector and the The +1 dimension is calculated by the following formula:

[0085] ;

[0086] ;

[0087] where, is the dimension index, is the total dimension of the position encoding.

[0088] The position encoding enables the model to perceive the order and phase information of each time step in the input sequence, enhances the model's ability to model the time dependence of the hydration process, and thus more accurately identifies the change rules of key stages such as the induction period and the acceleration period.

[0089] In a specific embodiment, the method for predicting the heat of hydration of cement based on non-weight sharing and variational mechanism includes: obtaining groups of historical experimental data and property data of the heat of hydration of cement paste under different working conditions. Among them, the property data includes component contents such as parameters such as calculated mass fraction, specific surface area, residue on sieve, water-cement ratio, reaction temperature, etc. In order to provide sufficient data for the model, it is monitored for 24 hours in total, the sampling interval is set to 1 minute, and each group of sequences contains 1440 time points. The original data is cleaned and standardized to meet the requirements of k-fold cross-validation; in each fold of the experiment, for the cement property data in the training set, it is standardized on the cement dimension, and the mean (mean) and variance (std) data are saved; subsequently, the mean and variance in the training set are used to standardize the cement property data in the test set, so as to eliminate the differences in the dimensions and numerical scales of different cement properties, enable the model to perform stable learning and fair comparison based on a consistent data distribution during the training and testing phases, and at the same time avoid information leakage in the test set.

[0090] In the process of constructing the time series, in order to introduce the stage physical information of the hydration reaction (such as the induction period, the acceleration period, the deceleration period), the position encoding of the sequence time information is added to the property data. Among them, the position encoding dimension is set, and the position encoding corresponding to the th time step is , so as to enhance the model's absolute perception ability of the reaction process.

[0091]

[0092]

[0093] The standardized cement property data is fused with time location coding and hydration heat time series data to generate an enhanced feature sequence containing material properties and reaction stage information.

[0094] In the sample construction phase, in each round of experiments, based on the training and test sets partitioned during standardization, the corresponding enhanced feature sequences are separated, and the sliding window method is used to construct the model input samples. Specifically, a fixed-length sliding window is set. That is, using 180 consecutive time steps as an input sample window, starting from the initial time step of the enhanced feature sequence, multiple window sub-sequences are continuously extracted as input samples. Each input sample corresponds to the hydration heat value of the next future time step in time sequence as the prediction label. For the first sample , recorded as Each time step includes attribute features and 10-dimensional temporal location encoding information for that operating condition. The sample corresponds to the ground truth label. Defined as the hydration heat release rate sequence one time step after the end of this window.

[0095] Therefore, the input dimension of each sample is [35, 180], and the size of R is [1].

[0096] In each experiment, based on the above standardization and feature fusion process, input samples and labels for the training set and test set are constructed respectively.

[0097] A leave-one-out cross-validation strategy was employed to partition the sample set. Specifically, in each round of validation, one complete set of hydration heat time-series data from all experimental batches was selected as the test set, while the data from all other conditions were used as the training set. This process was repeated until all conditions had been validated at least once. Throughout the partitioning and training process, the physical time-causal order of the hydration reaction was strictly followed to ensure that the test data represented entirely new and unknown conditions for the model. This prevented data leakage across batches and fully guaranteed the model's generalization performance and reliability for predicting different cement formulations under small sample conditions.

[0098] During the model training phase, 100 samples are randomly selected from the training set to form a batch (batch size = 100) for training. For example... Figure 2 The overall structure of the prediction model includes two layers of non-weight-shared convolutional modules, one layer of variational convolutional modules, and two layers of fully connected modules.

[0099] The core idea of ​​the non-weight-sharing convolution module is to dynamically change the convolution kernel parameters as the time series slides, thereby enhancing the model's ability to perceive the differences in features across different time slices. Specifically, for example... Figure 4The non-weight-shared convolution module first uses the input data and ordinary convolution operations to generate convolution kernel parameters that vary with the sequence, and then performs shape transformation on the generated convolution kernel parameters to match subsequent computational needs. Subsequently, it uses the input data and the generated non-weight-shared convolution kernel parameters again for computation to achieve dynamic convolution operations.

[0100] By dynamically generating convolution kernel parameters through the first-layer convolution, the convolution kernel is no longer fixed but becomes a function of the input sequence. Furthermore, by expanding the input data and performing window-by-window computation with the dynamic kernel, the convolution operation parameters adaptively change with each time step. This solves the technical problem of different mechanisms at different stages of the hydration process, enabling the model to analyze data from different stages such as the induction and acceleration phases using different feature extractors, thereby accurately capturing dynamic time-varying patterns.

[0101] Existing technologies mostly employ standard convolutional neural networks (CNNs), whose inherent weight-sharing mechanism ignores the physical fact that the cement hydration reaction is controlled by different chemical kinetic mechanisms at different stages (induction, acceleration, and deceleration phases), resulting in limited accuracy when processing non-stationary time-series data. This invention innovatively introduces a non-weight-sharing convolutional module, which dynamically generates and applies convolutional kernel parameters unique to the time step of the input data. This "time-dependent" adaptive feature extraction mechanism enables the model to accurately identify subtle evolutions during the induction phase, sensitively capture dramatic increases during the acceleration phase, and accurately predict the peak value and phase of the main exothermic peak, significantly improving the model's fitting accuracy for complex nonlinear dynamic processes.

[0102] A batch of samples of size [batchsize, 35, 180] is used as the input to the first non-weight-shared convolutional module. The output channels are set to 64, the kernel size to 3, the dilation rate to 1, no padding (padding=0), and the stride to 1. First, a normal one-dimensional convolution operation with padding of 1 is performed on the input data to obtain an intermediate result of size [batchsize, 300, 180]. This result is then processed by batch normalization and activation function, and then subjected to a one-dimensional convolution operation without padding to generate an output of size [batchsize, 64×35×3, 178]. Next, this output is reshaped to the corresponding size, which is used as the non-weight-shared convolutional kernel of the current module. The input data of the current layer is unfolded through a sliding window operation to generate local subsequence blocks in the time dimension, so that each window corresponds to a set of continuous time segments. Subsequently, the expanded input is multiplied element-wise by the dynamically generated convolution kernel and summed along the window dimension to complete the dynamic convolution calculation for this layer.

[0103] A batch of samples of size [batchsize, 64, 178] is used as the input to the second-layer non-weight-shared convolutional module. The output channel number is set to 32, the kernel size to 5, the dilation rate to 1, no padding (padding=0), and the stride to 1. First, a normal one-dimensional convolution operation with padding of 1 is performed on the input data to obtain an intermediate result of size [batchsize, 300, 178]. This result is then processed by batch normalization and activation function, and then subjected to a one-dimensional convolution operation without padding to generate an output of size [batchsize, 32×64×5, 174]. Next, this output is reshaped to the corresponding size, which is used as the non-weight-shared convolutional kernel of the current module. The input data of the current layer is unfolded through a sliding window operation to generate local subsequence blocks in the time dimension, so that each window corresponds to a set of continuous time segments. Subsequently, the expanded input is multiplied element-wise by the dynamically generated convolution kernel and summed along the window dimension to complete the dynamic convolution calculation for this layer.

[0104] The output of the non-weight-sharing convolutional module is used as the input of the first-layer variational convolutional module. The variational convolutional module has 32 input channels, 8 output channels, a kernel size of 5, a dilation rate of 1, no padding (padding=0), and a stride of 1. Figure 5 In the variational convolution module, the input data is first processed through two independent one-dimensional convolution operations to obtain the mean parameters of the latent variable distribution. With log-variance parameter The standard deviation is calculated using the logarithmic variance parameter. Subsequently, latent variable representations are generated using reparameterization techniques, with a size of [batchsize, 8, 170], where, This is random noise sampled from a standard normal distribution.

[0105] ;

[0106] The mean of the latent variable distribution is predicted by performing two independent convolutions on the same input. and logarithmic variance Instead of directly outputting a definite value, feature representations are sampled from this distribution during forward propagation using reparameterization techniques. The core effect of this mechanism is the introduction of randomness, forcing the module to learn a smooth feature distribution that covers the uncertainty of the data, rather than memorizing the noise of the training samples. This effectively alleviates the overfitting problem under small sample sizes and enhances the model's generalization ability. Traditional deterministic neural networks are prone to overfitting due to the high cost of acquiring cement hydration heat experimental data, limited sample size, and susceptibility to experimental noise, leading to prediction failures when faced with new formulas or new working conditions. This invention constructs a variational convolution module in the feature extraction layer, transforming deterministic weight learning into probabilistic distribution learning. By introducing random Gaussian perturbations during forward propagation, the model is essentially performing implicit data augmentation in the feature space. This mechanism not only effectively suppresses the interference of experimental noise on model training but also gives the model tolerance to small fluctuations in input data, improving the model's generalization ability when facing fluctuating working conditions in industrial settings.

[0107] latent variables obtained The output of the variational convolution module is processed by batch normalization and activation functions to obtain the final output of the module.

[0108] Among them, the distribution error is minimized using KL divergence in the latent space. Constraints are imposed to ensure the hidden space. It can retain as much information as possible while aligning it with a Gaussian distribution, and its KL divergence constraint method is consistent with that in VAE, i.e. ,in Represents the logarithmic variance. This represents the mean. The standard deviation is represented by . A KL divergence term is added to the training loss to measure and minimize the difference between the latent variable distribution learned by the module and the standard normal distribution. This constraint prevents the variational module from learning meaningless or overly complex feature distributions, guiding it to learn more compact and generalizable feature representations, thus improving the model's stability in scenarios with small samples and high noise levels.

[0109] The input from the previous layer is subjected to dimensionality compression and reshaping operations to ensure a two-dimensional representation of [batchsize, 1360], which is then used as the input to the fully connected network module. For example... Figure 6 A fully connected network consists of two layers of fully connected modules:

[0110] Fully connected module 1: 1360 input neurons and 128 output neurons;

[0111] Fully connected module 2: It has 128 input neurons and 1 output neuron. A batch normalization layer and a non-linear activation function are added after fully connected module 1 to enhance the model's expressive power and stability. After layer-by-layer mapping and non-linear transformation by the fully connected module, the model finally outputs a vector group P of size [batchsize, 1]. This vector group represents the predicted result of subsequent cement hydration heat at the sample time.

[0112] Example 2

[0113] like Figure 7 As shown, this embodiment provides a cement hydration heat prediction system based on non-weight sharing and variational mechanisms, including:

[0114] The data acquisition unit is used to acquire the historical time-series data of the heat of hydration of cement samples and cement property parameters; wherein, the cement property parameters include any one or more of the following: chemical composition, physical properties, water-cement ratio, and reaction temperature;

[0115] The feature enhancement unit is used to preprocess and fuse the hydration heat history time series data and cement property parameters to generate an enhanced feature sequence containing material properties and time stage information.

[0116] The sample construction unit is used to process the enhanced feature sequence using a sliding window to construct model input samples, wherein each input sample corresponds to one or more future time steps of hydration heat value as a prediction label.

[0117] The prediction unit is used to input the input samples into a hydration heat prediction model that sequentially includes a non-weight-shared convolution module, a variational convolution module, and a fully connected module to obtain the prediction result. Specifically, the non-weight-shared convolution module dynamically generates convolution kernel parameters based on the input samples and uses these dynamic convolution kernels to extract features from the input data, outputting time-sensitive features. The variational convolution module performs probabilistic encoding on the time-sensitive features, introducing randomness through reparameterization techniques, and outputs a probabilistically regularized feature representation. The fully connected module integrates and performs nonlinear mapping on the probabilistically regularized feature representation, outputting the final predicted hydration heat value.

[0118] Example 3

[0119] Embodiment 3 of the present invention provides an electronic device.

[0120] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements the steps in the cement hydration heat prediction method based on non-weight sharing and variational mechanism as described in Embodiment 1 of the present invention.

[0121] The detailed steps are the same as those of the cement hydration heat prediction method based on non-weight sharing and variational mechanism provided in Example 1, and will not be repeated here.

[0122] Example 4

[0123] Embodiment 4 of the present invention provides a computer-readable storage medium.

[0124] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the cement hydration heat prediction method based on non-weight sharing and variational mechanism as described in Embodiment 1 of the present invention.

[0125] The detailed steps are the same as those of the cement hydration heat prediction method based on non-weight sharing and variational mechanism provided in Example 1, and will not be repeated here.

[0126] Example 5

[0127] Embodiment 5 of the present invention provides a computer program product.

[0128] A computer program product includes software code, wherein the program in the software code performs the steps of the cement hydration heat prediction method based on non-weight sharing and variational mechanism as described in Embodiment 1 of the present invention.

[0129] The detailed steps are the same as those of the cement hydration heat prediction method based on non-weight sharing and variational mechanism provided in Example 1, and will not be repeated here.

[0130] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0131] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0132] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0133] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0134] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0135] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0136] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A method for predicting cement hydration heat based on non-weight sharing and variational mechanisms, characterized in that, include: Obtain historical time-series data of hydration heat and cement property parameters of cement samples; wherein, the cement property parameters include any one or more of the following: chemical composition, physical properties, water-cement ratio, and reaction temperature; The historical time series data of hydration heat and cement property parameters are preprocessed and feature fused to generate an enhanced feature sequence containing material properties and time stage information. The enhanced feature sequence is processed using a sliding window to construct model input samples, where each input sample corresponds to one or more future time steps of hydration heat value as a prediction label; The input samples are fed into a hydration heat prediction model that sequentially includes a non-weight-shared convolution module, a variational convolution module, and a fully connected module to obtain the prediction result. Specifically, the non-weight-shared convolution module dynamically generates convolution kernel parameters based on the input samples and uses these dynamic kernels to extract features from the input data, outputting time-sensitive features. The variational convolution module performs probabilistic encoding on the time-sensitive features, introducing randomness through reparameterization techniques, and outputs a probabilistically regularized feature representation. The fully connected module integrates and performs nonlinear mapping on the probabilistically regularized feature representation, outputting the final hydration heat prediction value. The operation of the non-weight-sharing convolution module to extract time-sensitive features specifically includes: Dynamically changing convolution kernel parameters are generated based on the input data through convolution operations; Expand the input data into local sequence blocks along the time dimension; The expanded input data is then used to perform window-by-window calculations with the dynamically generated convolutional kernel parameters to obtain time-sensitive features. The variational convolution module performs probabilistic encoding and introduces randomness, specifically including: Perform two independent convolution operations on the input features, and output the mean parameters of the latent variable distributions respectively. And log-variance parameter The standard deviation is calculated using the logarithmic variance parameter. ; Random noise sampled from a standard normal distribution ; Calculate latent variable representations based on reparameterization techniques. ; The training process of the hydration heat prediction model includes: Prepare a training dataset containing input samples and their corresponding predicted labels; The training dataset is input into the hydration heat prediction model for forward propagation, and the prediction loss is calculated. The predicted loss includes Divergence loss, using The divergence constrains the latent variable distribution of the variational convolution module. The formula for calculating the divergence constraint term is: ； Based on the predicted loss, the model parameters are updated using the backpropagation algorithm until the model converges.

2. The method as described in claim 1, characterized in that, The preprocessing and feature fusion include: The historical time series data of hydration heat and cement property parameters were standardized respectively. The standardized time-series data is concatenated with attribute parameters to form fused data; Temporal location codes are embedded into the fused data to generate the enhanced feature sequence.

3. The method as described in claim 2, characterized in that, The method for generating the embedded time position code is as follows: For the fused data, the first... Each time step has a position encoding vector. The Wei and Di +1 dimension is calculated using the following formula: ； ； in, For dimensional indexing, The total dimension for location encoding.

4. A cement hydration heat prediction system based on non-weight sharing and variational mechanism, characterized in that, include: The data acquisition unit is used to acquire the historical time-series data of the heat of hydration of cement samples and cement property parameters; wherein, the cement property parameters include any one or more of the following: chemical composition, physical properties, water-cement ratio, and reaction temperature; The feature enhancement unit is used to preprocess and fuse the hydration heat history time series data and cement property parameters to generate an enhanced feature sequence containing material properties and time stage information. The sample construction unit is used to process the enhanced feature sequence using a sliding window to construct model input samples, wherein each input sample corresponds to one or more future time steps of hydration heat value as a prediction label. The prediction unit is used to input the input samples into a hydration heat prediction model that sequentially includes a non-weight-shared convolution module, a variational convolution module, and a fully connected module to obtain the prediction result. Specifically, the non-weight-shared convolution module dynamically generates convolution kernel parameters based on the input samples and uses these dynamic convolution kernels to extract features from the input data, outputting time-sensitive features. The variational convolution module performs probabilistic encoding on the time-sensitive features, introducing randomness through reparameterization techniques, and outputs a probabilistically regularized feature representation. The fully connected module integrates and performs nonlinear mapping on the probabilistically regularized feature representation, outputting the final predicted hydration heat value. The operation of the non-weight-sharing convolution module to extract time-sensitive features specifically includes: Dynamically changing convolution kernel parameters are generated based on the input data through convolution operations; Expand the input data into local sequence blocks along the time dimension; The expanded input data is then used to perform window-by-window calculations with the dynamically generated convolutional kernel parameters to obtain time-sensitive features. The variational convolution module performs probabilistic encoding and introduces randomness, specifically including: Perform two independent convolution operations on the input features, and output the mean parameters of the latent variable distributions respectively. And log-variance parameter The standard deviation is calculated using the logarithmic variance parameter. ; Random noise sampled from a standard normal distribution ; Calculate latent variable representations based on reparameterization techniques. ; The training process of the hydration heat prediction model includes: Prepare a training dataset containing input samples and their corresponding predicted labels; The training dataset is input into the hydration heat prediction model for forward propagation, and the prediction loss is calculated. The predicted loss includes Divergence loss, using The divergence constrains the latent variable distribution of the variational convolution module. The formula for calculating the divergence constraint term is: ； Based on the predicted loss, the model parameters are updated using the backpropagation algorithm until the model converges.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the cement hydration heat prediction method based on non-weight sharing and variational mechanism as described in any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the cement hydration heat prediction method based on non-weight sharing and variational mechanism as described in any one of claims 1 to 3.

7. A computer program product, comprising software code, characterized in that, The program in the software code executes the steps of the cement hydration heat prediction method based on non-weight sharing and variational mechanism as described in any one of claims 1 to 3.

Images