Intelligent analysis method and system for cement material proportioning

Abstract

The application discloses a cement material proportioning intelligent analysis method and system, which can be executed through an intelligent chip and belongs to the technical field of computers. Through the introduction of dynamic quality risk assessment and collaborative enhancement processing, real-time quantification and improvement of cement production multi-source heterogeneous data quality are realized. Based on the risk level, the prediction modeling strategy is dynamically adjusted, and multi-objective optimization is performed by fully considering the prediction uncertainty, so that a more robust recommended proportioning scheme is generated. In addition, through the closed-loop feedback mechanism, the system parameters are collaboratively adjusted, ensuring the continuous self-adaptation and optimization of the entire analysis chain, effectively improving the reliability and stability of the cement material proportioning under complex working conditions.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an intelligent analysis method and system for cement material proportioning. Background Technology

[0002] Optimizing cement material proportions is a core aspect of ensuring engineering performance and economic efficiency. Current methods mainly rely on static mechanistic models or data-driven models based on historical data for prediction and optimization. However, cement production data is characterized by its multi-source, heterogeneous, and dynamic nature. Fluctuations in raw materials and changes in operating conditions often lead to real-time fluctuations in the quality of input data, directly affecting the reliability of prediction models.

[0003] Therefore, in the context of cement production where data quality changes dynamically, how to build an intelligent system that can assess data quality risks in real time and dynamically and adaptively adjust the entire chain analysis strategy (including data processing, model fusion, and optimization decision-making) accordingly, so as to fundamentally improve the robustness of mix optimization under complex working conditions, has become a key technical problem that urgently needs to be solved. Summary of the Invention

[0004] This application provides a method and system for intelligent analysis of cement material proportioning, which can improve the robustness of proportioning optimization under complex working conditions. The technical solution is as follows: On the one hand, a method for intelligent analysis of cement material proportioning is provided, the method comprising: Dynamic quality risk assessment and collaborative enhancement processing are performed on multi-source heterogeneous data of cement production to generate a physical information enhancement dataset and dynamic data quality risk value. The dynamic data quality risk value is associated with a predefined risk level. The multi-source heterogeneous data includes material chemical composition, proportioning settings, process parameters and microstructure image data. Based on the risk level associated with the dynamic data quality risk value, a target modeling strategy is selected from multiple preset hybrid prediction modeling strategies. The target modeling strategy is then used to perform hybrid prediction modeling of the physical information augmentation dataset, which combines mechanism and data fusion, to obtain the matching performance prediction results and the corresponding uncertainty interval. Based on the predicted performance of the proportions and the uncertainty interval, and according to the optimization objective set corresponding to the risk level, multi-objective optimization is performed to generate a recommended proportion scheme; Based on the deviation between actual production feedback and the predicted performance of the formula ratio, the parameters of the dynamic quality risk assessment, the thermodynamic simulation boundary conditions of the synergistic enhancement treatment, the fusion weights of the hybrid prediction modeling, and the criteria for the risk level are synergistically adjusted. The actual production feedback is the actual performance result obtained by producing using the recommended formula ratio. Attached Figure Description

[0005] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0006] Figure 1 This is a flowchart of an intelligent analysis method for cement material proportioning provided in an embodiment of this application; Figure 2 This is a flowchart of another intelligent analysis method for cement material proportioning provided in the embodiments of this application. Detailed Implementation

[0007] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0008] In this application, the terms "first," "second," etc., are used to distinguish identical or similar items with essentially the same function. It should be understood that there is no logical or temporal dependency between "first," "second," and "nth," nor are there any restrictions on quantity or execution order.

[0009] For ease of understanding, the following explains some key terms in this embodiment: Intelligent analysis method for cement material proportioning: This method aims to achieve intelligent recommendation of material proportions in the cement production process by integrating data processing, predictive modeling, and optimization decision-making, so as to adapt to the dynamically changing production environment.

[0010] Multi-source heterogeneous data refers to data of various types and formats originating from the cement production process, such as material chemical composition, proportioning settings, process parameters, and microstructure image data. These data collectively describe various aspects of cement production.

[0011] Dynamic quality risk assessment refers to the continuous and adaptive evaluation of the quality status of real-time cement production data, quantifying the reliability level of data at different time points and from different sources, and converting it into actionable risk values.

[0012] Collaborative enhancement processing: This refers to targeted preprocessing and information supplementation of raw multi-source heterogeneous data based on the data quality risk assessment results, in order to improve the integrity, accuracy and usability of the data, thereby generating a more informative physical information enhanced dataset.

[0013] Physical information augmentation dataset: refers to a dataset that has undergone collaborative augmentation processing. It not only contains the original data information, but also incorporates additional information obtained through physical mechanism simulation or feature extraction, so as to more comprehensively characterize the properties of cement materials.

[0014] Dynamic data quality risk value: refers to a quantitative indicator obtained through dynamic quality risk assessment, used to characterize the current data quality level and potential risks. This risk value is associated with a predefined risk level.

[0015] Risk level: refers to the discrete category divided according to the dynamic data quality risk value, such as low risk, medium risk, high risk, etc. Different risk levels indicate different reliability levels of data quality and guide the selection of subsequent processing strategies.

[0016] Hybrid predictive modeling strategy: refers to a set of pre-defined methods for fusion predictive modeling of mechanisms and data. Each strategy may correspond to different model structures, fusion mechanisms or parameter configurations to adapt to different data quality and prediction needs.

[0017] Target modeling strategy: refers to the predictive modeling method that is most suitable for the current data quality status, selected from multiple pre-set hybrid predictive modeling strategies based on the risk level associated with the current dynamic data quality risk value.

[0018] Hybrid predictive modeling that combines mechanism and data: This refers to a predictive method that combines the advantages of physical mechanism models and data-driven models. By effectively integrating the two, it can achieve more accurate and robust prediction of matching performance under different data quality conditions.

[0019] Mix proportion performance prediction results: refers to the predicted values ​​of performance indicators (such as strength, setting time, etc.) that a specific cement material mix proportion scheme may achieve in actual production, obtained through mixed prediction modeling.

[0020] Uncertainty range: refers to a range around the predicted performance of the matching ratio, used to quantify the possible fluctuations or errors in the predicted values, reflecting the inherent uncertainties of data quality and the model.

[0021] Optimization objective set: refers to the multiple performance and cost objectives that need to be considered simultaneously when generating recommended mix proportions, such as maximizing cement strength, minimizing production costs, and minimizing environmental impact.

[0022] Multi-objective optimization: refers to the process of finding a set of solutions that can simultaneously satisfy multiple conflicting optimization objectives without improving any single objective without harming the other objectives.

[0023] Recommended mix design: refers to the cement material mix design obtained through a multi-objective optimization process, which is considered the optimal or second-best under current production conditions.

[0024] Actual production feedback: refers to the real performance data and production results obtained through monitoring and testing after the recommended proportioning scheme is applied to actual cement production.

[0025] Collaborative adjustment: This refers to the coordinated and adaptive correction and optimization of multiple internal components of the system (such as risk assessment parameters, data augmentation boundary conditions, model fusion weights, and risk level criteria) based on the deviation between actual production feedback and performance prediction results, in order to continuously improve the accuracy and robustness of the entire system.

[0026] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0027] Traditional cement material proportioning optimization methods primarily rely on static mechanistic models or data-driven models based on historical data for prediction and optimization. However, cement production data is characterized by its multi-source, heterogeneous, and dynamic nature. Fluctuations in raw materials and changes in operating conditions often lead to real-time fluctuations in input data quality, directly affecting the reliability of the prediction model. Existing data quality assessment technologies are mostly based on static thresholds or offline analysis, failing to dynamically quantify the risks of real-time production data streams. Even when using hybrid modeling that integrates mechanistic and data approaches, the fusion architecture is often fixed or pre-set, unable to dynamically adjust the model's reliable sources based on data quality risks. Furthermore, downstream multi-objective optimization processes typically treat prediction results as deterministic values, failing to fully consider the prediction uncertainties caused by data quality fluctuations. This results in poor anti-interference capabilities of the optimization scheme in actual production, and performance and cost objectives are prone to deviation.

[0028] Based on this, this application proposes an intelligent analysis method for cement material proportioning, see [link to relevant documentation]. Figure 1 This includes the following steps.

[0029] 101. Perform dynamic quality risk assessment and collaborative enhancement processing on multi-source heterogeneous data of cement production to generate physical information enhancement dataset and dynamic data quality risk value. The dynamic data quality risk value is associated with a predefined risk level. The multi-source heterogeneous data includes material chemical composition, proportioning settings, process parameters and microstructure image data.

[0030] 102. Based on the risk level associated with dynamic data quality risk value, select a target modeling strategy from multiple pre-set hybrid prediction modeling strategies, and use the target modeling strategy to perform hybrid prediction modeling of mechanism and data fusion on the physical information augmentation dataset to obtain the matching performance prediction results and the corresponding uncertainty interval.

[0031] 103. Based on the performance prediction results and uncertainty range of the ratio, and according to the optimization objective set corresponding to the risk level, perform multi-objective optimization to generate a recommended ratio scheme.

[0032] 104. Based on the deviation between actual production feedback and performance prediction results, the parameters of dynamic quality risk assessment, the thermodynamic simulation boundary conditions of synergistic enhancement processing, the fusion weights of hybrid prediction modeling, and the criteria for risk level are synergistically adjusted. Actual production feedback is the actual performance result obtained by producing using the recommended proportion scheme.

[0033] This embodiment achieves real-time quantification and improvement of the quality of multi-source heterogeneous data in cement production by introducing dynamic quality risk assessment and collaborative enhancement processing. The predictive modeling strategy is dynamically adjusted based on risk levels, and multi-objective optimization is performed to fully consider prediction uncertainties, thereby generating a more robust recommended mix proportion scheme. Furthermore, a closed-loop feedback mechanism is used to collaboratively adjust system parameters, ensuring continuous adaptation and optimization throughout the entire analysis chain, effectively improving the reliability and stability of cement material proportioning under complex operating conditions.

[0034] This application further proposes dynamic quality risk assessment and collaborative enhancement processing for multi-source heterogeneous data in cement production, generating a physical information enhancement dataset and dynamic data quality risk values. (See [link to relevant documentation]). Figure 2 The process includes the following steps.

[0035] 201. Perform multi-dimensional credibility quantification on multi-source heterogeneous data to obtain dynamic data quality risk values ​​and associated risk levels. This multi-dimensional credibility quantification includes coupled calculation of the basic credibility of the data source, instantaneous anomaly, and time-series logical consistency.

[0036] 202. Based on risk level, determine the target operation mode from multiple predefined data augmentation operation modes, which are associated with different risk levels.

[0037] 203. Perform data augmentation operations corresponding to the target operation mode on the multi-source heterogeneous data. The data augmentation operations include generating virtual performance data associated with the material ratios in the multi-source heterogeneous data and extracting quantitative features from the microstructure images in the multi-source heterogeneous data.

[0038] 204. The virtual performance data, quantitative features, and multi-source heterogeneous data after quality compensation are fused to generate a physical information augmentation dataset.

[0039] Among them, multi-dimensional credibility quantification is performed on multi-source heterogeneous data to obtain dynamic data quality risk values ​​and associated risk levels. The aim is to conduct real-time and comprehensive credibility assessment of cement production data from different sources and in different formats, so as to quantify its quality risks.

[0040] This multi-dimensional trustworthiness metric involves coupled calculations of the data source's fundamental trustworthiness, transient anomalies, and time-series logical consistency. This coupled calculation aims to comprehensively consider the inherent reliability of the data source, the degree of data anomaly at a specific moment, and the logical consistency of the data over time, thereby providing a comprehensive and dynamic data quality risk assessment.

[0041] Based on risk levels, a target data augmentation mode is determined from multiple predefined data augmentation operation modes. This aims to adaptively select the most appropriate data augmentation strategy based on the currently assessed data quality risk level. This ensures that the intensity and type of data augmentation are matched to the actual quality of the data.

[0042] These predefined data augmentation operation modes are associated with different risk levels, aiming to provide the system with a flexible and configurable data processing strategy to address data quality issues of varying degrees. This correlation allows the data augmentation process to intelligently adjust based on fluctuations in data quality.

[0043] Performing data augmentation operations corresponding to the target operation mode on multi-source heterogeneous data aims to apply selected data augmentation strategies to the original multi-source heterogeneous data to improve its quality and usability. This operation is a key step in data preprocessing, ensuring that subsequent models receive high-quality input.

[0044] This data augmentation operation involves generating virtual performance data associated with material proportions from multi-source heterogeneous data. It aims to supplement or expand existing datasets through synthesis, especially when real performance data is scarce or when actual test data for certain material proportion combinations is lacking. This helps fill data gaps and improve the model's generalization ability.

[0045] This data augmentation also includes extracting quantitative features from microstructural images of multi-source heterogeneous data, aiming to transform the qualitative information contained in cement microstructural images into quantitative numerical features that can be directly utilized by machine learning models. This helps models to gain a deeper understanding of the microscopic mechanisms between material proportions and final performance.

[0046] By fusing virtual performance data, quantitative features, and quality-compensated multi-source heterogeneous data, a physically enhanced dataset is generated. The aim is to integrate all enhanced and quality-compensated data into a unified and richer dataset. This provides high-quality, multi-dimensional, and physically-informed input for subsequent hybrid predictive modeling based on mechanism and data fusion.

[0047] The above technical solutions improve the accuracy and reliability of subsequent hybrid predictive modeling that integrates mechanisms and data, laying a solid foundation for intelligent analysis of cement material proportions and effectively addressing the challenges of dynamic changes in cement production data.

[0048] This application further proposes a multi-dimensional reliability measurement method for multi-source heterogeneous data to obtain dynamic data quality risk values ​​and associated risk levels. This process includes the following steps: Obtain the basic attribute information, real-time monitoring sequences, and historical data records of each data source in this multi-source heterogeneous data.

[0049] The basic attribute information refers to the inherent characteristics of the data source, such as the model of the data acquisition equipment, accuracy class, calibration date, data format, supplier reputation, and data generation frequency. This information can be obtained by consulting the equipment manual, system configuration documents, or data source registry. For example, for a temperature sensor, its basic attribute information might include its measurement range, accuracy class, and last calibration time. Another way to obtain this information is through the data source's metadata management system, which stores and maintains static descriptive information for all data sources. The real-time monitoring sequence refers to an ordered set of data generated by the data source within a continuous time period, reflecting the current state or process. For example, real-time sensor data such as temperature, pressure, and flow rate collected during cement production, or real-time video stream data from the production line. These sequence data are typically collected and transmitted in real-time through industrial control systems (ICS), distributed control systems (DCS), or Internet of Things (IoT) platforms. Another way to obtain this information is by subscribing to real-time data streams through message queues (such as Kafka) and organizing them by timestamps. The historical data record refers to a set of data accumulated and stored by the data source over a past period. For example, production batch data from the past few months or years, material composition analysis reports, and equipment maintenance logs. These records are typically stored in relational databases (such as SQL Server), non-relational databases (such as MongoDB), or data lakes (such as HDFS). Another way to access them is through a data warehouse system, which cleans, integrates, and archives historical data to facilitate querying and analysis.

[0050] Based on this basic attribute information, the basic credibility of each data source is determined; based on this real-time monitoring sequence, the instantaneous anomaly is determined; based on this historical data record, the temporal logic consistency is verified and a temporal logic consistency score is obtained.

[0051] The basic reliability score is a quantitative assessment of the inherent reliability of the data source. For example, it can be assigned a value based on attributes such as the data source's accuracy level, calibration status, and vendor rating, using preset scoring rules or expert systems. For instance, a recently calibrated high-precision sensor might receive a higher basic reliability score, while an older, uncalibrated sensor would receive a lower score. Another method is based on the type of data source; for example, laboratory analysis data typically has a higher basic reliability score than online sensor data. The transient anomaly score measures the degree to which real-time data points deviate from normal behavior patterns. For example, statistical methods (such as Z-score and IQR) can be used to detect whether real-time data points exceed a certain confidence interval of their historical distribution, or machine learning models (such as Isolation Forest and One-Class SVM) can be used to identify anomalous patterns. Another method is to set a hard threshold; when a real-time data value exceeds preset physical or engineering limitations, it is considered an anomaly. The temporal logical consistency score is an indicator that assesses whether a data sequence conforms to expected patterns or physical constraints over time. For example, it can be determined by examining the trend, periodicity, or causal relationship of the data sequence with other relevant variables. For example, the cooling rate of cement clinker should be negatively correlated with the cooling air volume. If an abnormal trend appears in historical data, the time-series logical consistency score will decrease. Another verification method is to use physical models or empirical formulas to substitute historical data into the model for prediction and compare the deviation between the predicted and actual values. The smaller the deviation, the higher the consistency score.

[0052] The basic credibility, transient anomaly, and time-series logical consistency scores are weighted and fused to obtain the dynamic data quality risk value for each data source.

[0053] Herein, the weighted fusion is a method of integrating credibility quantization results of different dimensions into a single index. For example, a linear weighted summation manner may be adopted, that is: dynamic data quality risk value = w1 × basic credibility + w2 × instantaneous anomaly degree + w3 × time series logic consistency score, where w1, w2 and w3 are preset weight coefficients with a sum of 1. These weights may be set according to expert experience, or optimized by training a machine learning model (such as a regression model) with historical data. Another fusion manner is to adopt a fuzzy logic inference system, taking the quantization results of all dimensions as fuzzy inputs, and obtaining the final risk value through a fuzzy rule base and a defuzzification process. The dynamic data quality risk value is a comprehensive quantitative index that reflects the current quality status of the data source in real time. This value is usually a continuous numerical value, for example, between 0 and 1, where 0 indicates extremely high data quality, and 1 indicates extremely low data quality. This value changes dynamically and can be updated in real time as basic attributes, real-time monitoring and historical records change. For example, when a certain sensor has an instantaneous anomaly, its dynamic data quality risk value will increase immediately.

[0054] The dynamic data quality risk value of each data source is compared with a preset threshold to determine the risk level. The preset threshold is a boundary value for dividing the continuous dynamic data quality risk value into discrete risk levels.

[0055] For example, three thresholds T1 and T2 may be set to divide the risk value into three levels: "low risk" (risk value < T1), "medium risk" (T1 ≤ risk value < T2) and "high risk" (risk value ≥ T2). These thresholds may be set according to industry standards, production experience or historical data analysis results. Another setting manner is to adopt an adaptive threshold, for example, dynamically adjusting the threshold based on the percentile of the historical risk value distribution. The risk level is a classification label for data quality risk, which is used to simplify the risk management and decision-making process. For example, it may be divided into discrete levels such as "low risk", "medium risk" and "high risk", or more detailed levels such as "extremely low", "relatively low", "medium", "relatively high" and "extremely high". These levels are directly associated with subsequent data processing, model selection and optimization strategies. For example, a high risk level may trigger more stringent data enhancement operations or more conservative predictive modeling strategies.

[0056] Through the above technical solutions, the dynamic and multi-dimensional risk assessment mechanism improves the robustness and adaptability of the entire intelligent cement material ratio analysis system to data quality fluctuations, and ensures the reliability of the ratio optimization scheme under complex working conditions.

[0057] The present application further proposes a method for performing a data enhancement operation according to a target operation mode, which specifically includes the following steps: The system acquires definitions for multiple predefined data augmentation operation modes associated with different risk levels. These predefined data augmentation operation modes are a set of data processing strategies pre-set during the system design phase based on a deep understanding of the characteristics and risks of cement production data, targeting different data quality risk levels. Each operation mode encapsulates specific data augmentation logic, execution parameters, and data range. For example, the definitions of these modes can be stored in configurable metadata files, such as XML or JSON format, or managed in the form of structured database tables, which record in detail the operation type, algorithm parameter configuration, and data processing boundary conditions corresponding to each risk level. These predefined modes include at least a first operation mode associated with low risk levels and a second operation mode associated with high risk levels. The first operation mode is specifically designed for scenarios with low data quality risk, defining a first set of execution parameters and a first data range. The first set of execution parameters typically involves algorithm configurations with low computational complexity; for example, it can use statistical interpolation or simple regression models to generate virtual data, or perform basic image feature extraction (such as edge detection or grayscale histogram analysis). The first data range is limited to the normal fluctuation range of historical data; for example, it only covers the 95% confidence interval of validated production parameters. In contrast, the second operating mode targets scenarios with higher data quality risks, defining a second set of execution parameters and a second data range. The computational complexity of the second set of execution parameters is significantly higher than that of the first set. For example, it may employ virtual data generated based on complex physical simulation models or deep learning models, or perform advanced image feature extraction (such as feature learning from convolutional neural networks or 3D structure reconstruction). The second data range is broader than the first, meaning it can explore a wider data space; for example, it can extend to the 99.9% confidence interval of historical data, and even include extreme parameter combinations or theoretical physical limits that have not yet been validated in actual production. This differentiated mode definition ensures that data augmentation operations can be customized according to the actual data quality risks, avoiding a "one-size-fits-all" strategy, thereby improving the targeting and efficiency of data processing.

[0058] Based on whether the target operating mode is the first or second operating mode, the system selects the corresponding first set of execution parameters and first data range, or the second set of execution parameters and second data range. This step enables dynamic adaptive adjustment of the data augmentation strategy. Specifically, the system determines the target operating mode to be used based on the previously assessed dynamic data quality risk value. If the dynamic data quality risk value indicates a low risk level, the system automatically selects and loads the first set of execution parameters and first data range defined by the first operating mode. Conversely, if the dynamic data quality risk value indicates a high risk level, the system selects and loads the second set of execution parameters and second data range defined by the second operating mode. This selection process can be implemented through conditional logic (e.g., using if-else statements in software programs) or by consulting a preset mapping table. For example, when the risk level is "low," the system reads the parameter set and data range associated with the "first operating mode" from the configuration library. When the risk level is "high," it reads the parameter set and data range associated with the "second operating mode." In this way, data augmentation operations can dynamically adjust their computing resource allocation and the breadth of data exploration based on the real-time risk level, ensuring the accuracy of processing and the rationality of resource utilization.

[0059] Using the selected execution parameters and data range, data augmentation operations are performed on multi-source heterogeneous data to generate virtual performance data and extract quantitative features. In this step, the system applies the aforementioned selected execution parameters and data range, matched to the current risk level, to the actual data processing. For example, if the first set of execution parameters and the first data range are selected, the data augmentation operation will be performed with lower computational complexity within a narrower data space, possibly generating virtual performance data through simple interpolation algorithms and extracting basic texture features from microstructure images. If the second set of execution parameters and the second data range are selected, the data augmentation operation will be performed with higher computational complexity within a broader data space, possibly generating richer virtual performance data through complex physical simulations or deep learning models, and extracting more advanced and semantically rich quantitative features from microstructure images, such as porosity distribution and crystal morphology parameters. The generation of virtual performance data can include simulation calculations based on thermodynamic principles, machine learning predictions based on historical data patterns, or rule inference combined with expert knowledge. Extraction of quantitative features can utilize advanced image processing algorithms, such as deep learning-based image segmentation and feature recognition techniques, or traditional digital image processing methods. This step is the final step in the data augmentation process, designed to provide a more comprehensive and reliable dataset of physically augmented information for subsequent hybrid predictive modeling.

[0060] By employing the above technical solutions, when data quality risk is low, parameters with lower computational complexity and a narrower data range are used, avoiding unnecessary waste of computational resources. Conversely, when data quality risk is high, parameters with higher computational complexity and a wider data range are used to ensure more comprehensive and in-depth data exploration and enhancement when there is significant uncertainty in data quality, thus compensating for the deficiencies of the original data. This dynamic adaptive adjustment mechanism enables data augmentation operations to be performed efficiently and comprehensively under different risk scenarios, thereby improving the quality and reliability of the physical information augmentation dataset. This provides a more solid foundation for subsequent hybrid predictive modeling of mechanism and data fusion, ultimately improving the accuracy and robustness of matching performance prediction results.

[0061] This application proposes an intelligent analysis method for cement material proportioning. It employs a target modeling strategy to perform hybrid predictive modeling by fusing mechanistic and data aspects of a physically-informed (PI) dataset, yielding proportioning performance prediction results and corresponding uncertainty intervals. Specifically, this process includes: determining rules for feature separation of the PI dataset, rules for assigning fusion weights to the mechanistic and data-driven prediction outputs, and rules for calculating the uncertainty interval based on the target modeling strategy. According to the feature separation rules, the PI dataset is separated into a mechanistic constraint feature subset and a data-driven feature subset. These subsets are then input into corresponding processing channels to obtain mechanistic prediction outputs and data-driven prediction outputs, respectively. According to the fusion weight allocation rules, fusion weights are dynamically assigned to the mechanistic prediction outputs and data-driven prediction outputs based on a dynamic data quality risk value. The assigned fusion weights are used to perform a weighted fusion of the mechanistic prediction outputs and data-driven prediction outputs to obtain a fused prediction distribution. According to the uncertainty interval calculation rules, the uncertainty interval corresponding to the proportioning performance prediction results is determined based on the fused prediction distribution and the dynamic data quality risk value.

[0062] The rules for feature separation, fusion weights, and uncertainty interval calculation, based on the target modeling strategy, aim to provide a set of customized guiding principles for the subsequent hybrid predictive modeling process. Since the target modeling strategy is selected from multiple pre-defined strategies based on dynamic data quality risk values, the rules determined here are dynamic and adaptable to the current data quality status. The feature separation rules define how to divide the features in the physically augmented dataset into mechanistic-constrained features and data-driven features.

[0063] Based on feature separation rules, the physical information augmentation dataset is separated into a mechanistic constraint feature subset and a data-driven feature subset. These subsets are then input into their respective processing channels to obtain mechanistic prediction outputs and data-driven prediction outputs, respectively. This step is fundamental to hybrid prediction modeling. It divides the input data into two categories according to pre-defined feature separation rules and sends them to dedicated processing channels. This ensures that features of different properties are processed by the most suitable model or method, thereby improving the accuracy and robustness of the predictions.

[0064] Based on the fusion weight allocation rules, fusion weights are dynamically assigned to the mechanistic prediction output and the data-driven prediction output based on the dynamic data quality risk value. This step is one of the core innovations of this method, as it enables the adaptiveness of prediction model fusion. Through the dynamic data quality risk value, the system can assess the reliability of the current input data in real time and adjust the relative importance of the mechanistic model and the data-driven model in the final prediction accordingly, thereby maintaining the accuracy and stability of the prediction even when data quality fluctuates.

[0065] The mechanistic prediction output and the data-driven prediction output are weighted and fused using assigned fusion weights to obtain the fused prediction distribution. This step integrates the prediction results from two independently processed channels to form a more comprehensive and robust prediction. Weighted fusion fully leverages the stability of the mechanistic model and the flexibility of the data-driven model, while also considering their relative reliability under current data quality conditions.

[0066] Based on the uncertainty interval calculation rules, and using the fused prediction distribution and dynamic data quality risk value, the uncertainty interval corresponding to the matching performance prediction result is determined. This step aims to provide a quantitative reliability measure for the final matching performance prediction result. By combining the fused prediction distribution and dynamic data quality risk value, the system can generate a more accurate and informative uncertainty interval, which is crucial for subsequent optimization decisions.

[0067] The above technical solution, by determining feature separation rules, fusion weight allocation rules, and uncertainty interval calculation rules based on the target modeling strategy, ensures that the entire prediction process can be customized according to the current data quality risk. Based on the feature separation rules, the physically augmented dataset is separated into a mechanistic constraint feature subset and a data-driven feature subset, and input into corresponding processing channels respectively. This allows features of different properties to be optimally processed, avoiding the bias that may be caused by a single model processing heterogeneous data. More importantly, based on the fusion weight allocation rules, fusion weights are dynamically allocated between the mechanistic prediction output and the data-driven prediction output based on the dynamic data quality risk value. This makes model fusion no longer static and fixed, but can respond in real time to fluctuations in data quality. When the data quality risk is high, the system tends to rely more on the stability of the physical mechanism model. When the data quality risk is low, the flexibility and accuracy of the data-driven model can be fully utilized, thereby improving the robustness and accuracy of the prediction results in complex and ever-changing production environments. By using assigned fusion weights to weight and fuse the predicted outputs, a fused predicted distribution is obtained. Based on the uncertainty interval calculation rules, the uncertainty interval corresponding to the performance prediction result is determined according to the fused predicted distribution and the dynamic data quality risk value. This not only provides a point estimate of the prediction result, but also provides its reliability range, providing more comprehensive information for subsequent optimization decisions. It effectively avoids the risk of treating the prediction result as a certain value, thereby improving the anti-interference ability and performance stability of the recommended ratio scheme in actual production.

[0068] This application further proposes a step that, based on feature separation rules, separates the physical information augmentation dataset into a mechanism-constrained feature subset and a data-driven feature subset, and then inputs the mechanism-constrained feature subset and the data-driven feature subset into corresponding processing channels for processing to obtain mechanism prediction output and data-driven prediction output, respectively. The specific steps include: Based on feature separation rules, this method identifies and extracts a subset of mechanistic constraints from a physical information augmentation dataset. This subset consists of deterministic physical quantities based on cement chemistry theory, and a subset of data-driven features containing quantitative feature parameters extracted from microstructure images. The feature separation rules are a set of predefined criteria or algorithms used to systematically decompose complex physical information augmentation datasets into independent subsets suitable for different processing paths (mechanism-driven or data-driven). For example, the feature separation rules could identify deterministic physical quantities such as cement chemical composition (e.g., C3S, C2S, C3A, C4AF content), water-cement ratio, specific surface area, and theoretical heat of hydration as mechanistic constraint feature subsets through keyword matching or data type identification. Simultaneously, quantitative feature parameters extracted from microstructure images, such as pore size distribution, connectivity, fractal dimension of hydration products, and particle shape descriptors, would be identified as data-driven feature subsets. Alternatively, the feature separation rules could employ machine learning classifiers or clustering algorithms, learning through pre-training to distinguish between features primarily governed by physical laws and empirical features primarily based on images. This physical information augmentation dataset is a comprehensive dataset that has undergone quality assessment and enhancement. It integrates multiple data sources, including material chemical composition, proportioning settings, process parameters, and microstructure image data, and may also contain virtual performance data, aiming to provide rich and reliable input for subsequent modeling. The mechanism-constrained feature subset is another subset of the physical information augmentation dataset. Its features can be directly explained and predicted by established physical or chemical laws (especially cement chemistry theory). For example, it may include clinker mineral composition (such as the content of tricalcium silicate, dicalcium silicate, tricalcium aluminate, and tetracalcium aluminoferrite), heat of hydration, and theoretical strength development. The data-driven feature subset is another subset of the physical information augmentation dataset. Its features are less correlated with explicit physical laws but are rich in empirical information, especially quantitative feature parameters extracted from microstructure images. For example, it may include image texture features (such as Haralick features, local binary patterns), shape features (such as aspect ratio, roundness), and statistical features (such as average strength, variance).

[0069] Based on the target modeling strategy, a corresponding theoretical calculation model or simplified physical model is configured for the mechanistic constraint channel. A subset of mechanistic constraint features is input into the mechanistic constraint channel, processed by the model to obtain initial mechanistic prediction values. This target modeling strategy is selected from multiple pre-defined hybrid prediction modeling strategies based on dynamic data quality risk values. It defines feature separation rules, fusion weight allocation rules, and uncertainty interval calculation rules. The mechanistic constraint channel is a processing path specifically designed to handle features governed by physical laws, employing either a theoretical calculation model or a simplified physical model. For example, the mechanistic constraint channel can be a software module implementing thermodynamic equilibrium calculations, hydration kinetic models, or mass balance equations. Alternatively, it can be a simulation environment running finite element analysis (FEA) or computational fluid dynamics (CFD) to simulate simplified physical scenarios. The theoretical calculation model or simplified physical model is a mathematical representation of the physical phenomenon, ranging from first-principles equations to approximate models. For example, the model can be a set of differential equations describing reaction kinetics, a thermodynamic model calculating phase stability, or a simplified constitutive model for material properties. Another approach is to use an empirical correlation derived from physical principles, or a reduced-order model that captures key physical behaviors with less computational resources. The initial mechanistic prediction is the raw output of the theoretical computational model or the simplified physical model before calibration or correction.

[0070] Based on the calibration rules defined by the target modeling strategy, the initial mechanism prediction value is compared with historical calibration data obtained from the physical information augmentation dataset. The initial mechanism prediction value is then corrected based on the comparison results to obtain the mechanism prediction output. The calibration rules are predefined criteria and procedures used to adjust the model output based on historical observation data. For example, the calibration rules can employ statistical methods such as linear regression, Kalman filtering, or Bayesian updates to correct systematic biases or drift using historical data. Alternatively, the calibration rules can use expert-defined correction factors or lookup tables to map the initial prediction bias to specific adjustments based on known historical differences. The historical calibration data consists of validated and considered reliable past performance data and corresponding input parameters; for example, it could be a database containing measured properties (such as strength and setting time) of previously produced cement batches, along with corresponding material composition and process parameters. The mechanism prediction output is the calibrated, final mechanism prediction result, reflecting the physical principles and historical validation.

[0071] A subset of data-driven features is input into a data-driven channel, where a nonlinear feature transformation and mapping are performed by a machine learning model specified by the target modeling strategy to obtain a data-driven prediction output. This data-driven channel is a processing path specifically designed to handle features processed through experience and data-driven models (typically machine learning algorithms). For example, this data-driven channel can be a software module hosting various machine learning models (such as neural networks, support vector machines, and random forests). Alternatively, the data-driven channel can be a distributed computing framework used to train and deploy deep learning models on large datasets. This machine learning model is an algorithm that learns patterns from data without explicit programming. For example, this model can be deep neural networks (DNNs), convolutional neural networks (CNNs) for image features, recurrent neural networks (RNNs) for sequence data, or ensemble methods such as gradient boosting. Another implementation is that the model can be support vector machines (SVMs), Gaussian process regression (GPR), or a decision tree-based model. This nonlinear feature transformation and mapping is the process by which a machine learning model learns the complex nonlinear relationship between input features and output predictions. For example, it can be achieved through hidden layers in neural networks, kernel functions in support vector machines, or complex decision boundaries in tree-based models. The data-driven prediction output is the prediction result generated by the machine learning model based on a subset of data-driven features.

[0072] The above technical solutions ensure the accurate execution and dynamic adaptation of feature separation, effectively address data quality fluctuations, and thus improve the robustness and accuracy of hybrid predictive modeling.

[0073] This application further proposes, based on the target modeling strategy, to configure a corresponding theoretical calculation model or simplified physical model for the mechanistic constraint channel. Specifically, if the target modeling strategy is a first modeling strategy associated with a high-risk level, then a simplified physical model that has undergone order reduction is configured for the mechanistic constraint channel. If the target modeling strategy is a second modeling strategy associated with a low-risk level, then a refined theoretical calculation model containing a complete subset of reactions is configured for the mechanistic constraint channel.

[0074] The target modeling strategy is selected from multiple pre-defined hybrid predictive modeling strategies based on the risk level associated with dynamic data quality risk values. It encapsulates the overall guidelines for how to perform hybrid predictive modeling that integrates mechanistic and data fusion at a specific risk level, including feature separation rules, fusion weight allocation rules, and uncertainty interval calculation rules. It also defines the model configuration rules for the mechanistic constraint channel. For example, the target modeling strategy can be a predefined configuration template containing instructions on which mechanistic model should be used for different risk levels. Alternatively, it can be a decision tree or rule engine that dynamically determines the type of mechanistic model based on the input risk level. The mechanistic constraint channel is a processing path within the hybrid predictive modeling framework, specifically designed to handle deterministic knowledge based on physical laws, chemical reaction mechanisms, or engineering principles. Its role is to process a subset of mechanistic constraint features from the physically augmented dataset using theoretical calculation models or simplified physical models to generate mechanistic predictive outputs. This channel ensures that the prediction results conform to basic physicochemical laws, providing interpretability and stability for the model. Theoretical calculation models are mathematical models built upon mature theories and formulas in fields such as cement chemistry and materials science. They can accurately describe the relationship between material composition, reaction processes, and performance. For example, they can be models that predict the mineral composition of cement clinker based on methods such as Bogue calculations and Rietveld refinement, or models that predict the development of cement strength based on hydration kinetics theory. These models typically contain a complete subset of reactions and detailed physicochemical parameters, providing high-precision prediction results, but their computational complexity is relatively high. Simplified physical models, on the other hand, reduce the computational complexity of theoretical calculation models by methods such as order reduction, parameter simplification, or the introduction of assumptions. For example, they can be simplified by ignoring minor reactions, using empirical coefficients to replace complex mechanism parameters, or using linearization approximations. These models significantly reduce computational resources and time consumption while maintaining certain physical meaning, making them suitable for scenarios with high real-time requirements or significant uncertainties in data quality.

[0075] When the dynamic data quality risk value is high, indicating significant issues with the reliability, completeness, or consistency of the input data, the system selects the first modeling strategy associated with the high risk level. This strategy aims to address poor data quality by reducing the model's reliance on accurate data, minimizing prediction errors caused by data noise or missing data, and avoiding overfitting. In this case, configuring a simplified physical model after order reduction means selecting a model that is more computationally efficient and has relatively lower requirements for the accuracy of the input data. For example, empirical formulas based on a few key parameters can be used, or complex thermodynamic models can be linearized or approximated using polynomials. This type of model can quickly provide prediction results and reduce the computational burden and uncertainty introduced by data quality issues, thereby improving the system's robustness to fluctuations in data quality.

[0076] When the dynamic data quality risk value is low, indicating high input data quality and reliability, the system will select a second modeling strategy associated with the low risk level. This strategy aims to fully leverage the advantages of high-quality data, pursuing higher prediction accuracy and deeper mechanistic insights. In this case, configuring a refined theoretical calculation model containing a complete subset of reactions means adopting a model that can comprehensively reflect the complex physicochemical processes of cement materials. For example, a detailed hydration reaction kinetic model can be used, considering factors such as the interaction of multiple mineral phases and the evolution of pore structure. This model can provide more accurate and physically meaningful prediction results, maximizing the value of high-quality data and thus improving the accuracy of mix proportion performance prediction.

[0077] The above technical solution solves the limitation that traditional fixed model configuration cannot adapt to the dynamic changes in cement production data, ensures the efficiency and accuracy of hybrid predictive modeling under different data quality conditions, and thus improves the adaptability and reliability of the entire intelligent analysis method for cement material proportioning.

[0078] This application further proposes a method for dynamically allocating fusion weights between the mechanism-driven prediction output and the data-driven prediction output based on a dynamic data quality risk value and a fusion weight allocation rule. Specifically, the method includes: determining the form and parameters of the basic function based on the fusion weight allocation rule to obtain a weight allocation function; determining a relative reliability measure of the data-driven prediction output under the current input conditions based on the dynamic data quality risk value, wherein this relative reliability measure is negatively correlated with the data quality level represented by the dynamic data quality risk value; inputting this relative reliability measure into the weight allocation function to generate a first fusion weight for the data-driven prediction and a second fusion weight for the mechanism-driven prediction, wherein the second fusion weight is complementary to the first fusion weight.

[0079] The fusion weighting rule defines how to determine the relative contributions of mechanistic prediction output and data-driven prediction output in the final fusion result based on input conditions (e.g., dynamic data quality risk values). This rule can be a pre-defined set of logic or algorithms to guide the design and parameter selection of the weighting function. For example, the rule could be a set of conditional statements specifying that when the risk value is below a certain threshold, the data-driven prediction has a higher weight, and when the risk value is above a certain threshold, the mechanistic prediction has a higher weight. Alternatively, the rule could be a mathematical model selection strategy that chooses different weighting function types (e.g., linear, nonlinear, step function, etc.) and their parameters based on the risk level.

[0080] The basic function form refers to the type of mathematical expression used to construct the weight allocation function, such as a linear function, a sigmoid function, an exponential function, or a polynomial function. The parameters of this basic function are the adjustable variables within these function forms, used to fine-tune the function's behavior; for example, the slope and intercept in a linear function, or the gain and offset in a sigmoid function. For instance, the basic function form could be a sigmoid function, whose parameters include the center point of the function (corresponding to a risk value or reliability metric) and the slope (determining the rate of weight change). Alternatively, the basic function form could be a piecewise linear function, whose parameters include the start and end points of each segment, as well as the slope and intercept of each segment.

[0081] The weight allocation function is a mathematical mapping that takes an input (e.g., a relative reliability metric) and outputs one or more fused weights. This function is constructed based on a predefined fused weight allocation rule, combined with the form and parameters of a fundamental function. For example, the weight allocation function could be a sigmoid function, mapping the relative reliability metric between 0 and 1 to a first fused weight between 0 and 1, where the steepness and center point of the function curve can be adjusted by the fundamental function parameters. Alternatively, the weight allocation function could be a lookup table or an interpolation function, interpolating the relative reliability metric at predefined discrete points to obtain the corresponding first fused weight.

[0082] The dynamic data quality risk value is a quantitative indicator that measures the reliability, completeness, and accuracy of multi-source heterogeneous data on cement production in a real-time stream. It reflects the potential risk that current data input may lead to prediction bias. This risk value, as an input for weight allocation, directly affects the relative reliability of data-driven predictions, and thus the fusion ratio of mechanism and data prediction. A high risk value indicates poor data quality and reduced reliability of data-driven predictions. Conversely, a low risk value indicates good data quality and high reliability of data-driven predictions.

[0083] The relative reliability measure of the data-driven prediction output under the current input conditions is a quantitative indicator representing the degree of trustworthiness of the data-driven prediction output under the current dynamic data quality risk value. It transforms the abstract risk value into a more intuitive reliability score that can be used for weighted calculations, and is negatively correlated with the data quality level represented by the dynamic data quality risk value. For example, the dynamic data quality risk value can be normalized (e.g., mapped to the interval between 0 and 1), and then the normalized risk value can be subtracted from 1 to obtain the relative reliability measure. Alternatively, an inverse proportional function or an exponential decay function can be designed to take the dynamic data quality risk value as input and output a reliability score between 0 and 1, where a higher risk value results in a lower reliability score.

[0084] The first fusion weight is assigned to the data-driven prediction output, representing its proportion of contribution to the final fusion result. The second fusion weight is assigned to the mechanistic prediction output, representing its proportion of contribution to the final fusion result. This complementarity means that the sum of the two is 1 (or some fixed constant), ensuring that the total contribution remains consistent. Once the first fusion weight is calculated using the weight allocation function, the second fusion weight can be obtained directly by subtracting the first fusion weight from 1.

[0085] The above technical solution ensures the overall consistency of the fusion results, avoids weight conflicts, and enables the intelligent balancing of the contribution ratios of mechanism and data prediction based on real-time data quality risks. When data quality is good, the weight of data-driven prediction is increased accordingly, fully utilizing its ability to learn from complex data. When data quality is poor, the weight of mechanism prediction is increased accordingly, leveraging its robustness based on physical principles to compensate for the shortcomings of data-driven prediction. This dynamic and adaptive weight allocation mechanism improves the robustness and accuracy of hybrid predictive modeling in the face of varying cement production data quality, making the mix proportion performance prediction results more reliable, thus providing a more solid foundation for subsequent optimization decisions.

[0086] This application further proposes a method for determining the uncertainty interval corresponding to the performance prediction result based on the uncertainty interval calculation rules, the fused prediction distribution, and the dynamic data quality risk value. The method includes: determining the calculation method for the basic confidence interval and the modulation function for the interval adjustment coefficient based on the uncertainty interval calculation rules; determining the basic confidence interval according to the selected calculation method based on the dispersion of the fused prediction distribution; inputting the dynamic data quality risk value into the modulation function to convert the dynamic data quality risk value into an interval adjustment coefficient; and modulating the width of the basic confidence interval using the interval adjustment coefficient to obtain the uncertainty interval.

[0087] The system, based on the rules for calculating uncertainty intervals, determines the calculation method for the basic confidence interval and the modulation function for the interval adjustment coefficient. This aims to provide a clear guiding framework for subsequent interval calculations and adjustments, ensuring the standardization, consistency, and adaptability of the calculation process. Specifically, multiple basic confidence interval calculation methods can be preset, such as those based on statistical principles (e.g., standard deviation, quantiles) or Bayesian posterior distributions, and multiple modulation functions, such as linear, exponential, or piecewise functions. The system can determine which method and function to use by configuring parameters or selecting an algorithm based on specific application scenarios, data characteristics, or preset strategies. Alternatively, the system can use machine learning models, such as neural networks, to learn the relationship between uncertainty intervals and dynamic data quality risk values ​​in historical data, thereby adaptively generating or optimizing the calculation method and modulation function for the basic confidence interval.

[0088] Based on the dispersion of the fused predicted distribution, a baseline confidence interval is determined according to the selected calculation method. Its function is to quantify the inherent variability or uncertainty of the model's prediction results, providing a core, unadjusted benchmark interval for subsequent risk adjustment. For example, if the selected calculation method is based on statistical principles, the system can calculate the standard deviation or variance of the fused predicted distribution, and then calculate the corresponding confidence interval based on a preset confidence level (e.g., 95%) and statistical distribution (e.g., normal distribution). Alternatively, if the selected calculation method is based on a nonparametric method, the system can directly determine the upper and lower limits of the baseline confidence interval based on the quantiles of the fused predicted distribution (e.g., the 2.5% and 97.5% quantiles).

[0089] The dynamic data quality risk value is input into a modulation function to transform it into an interval adjustment coefficient. The aim is to quantify the impact of external data quality risk into a coefficient that can be used to adjust the width of the uncertainty interval, thus achieving a dynamic mapping between the risk value and uncertainty. For example, the modulation function can be a simple linear function, such as coefficient = 1 + k × risk value, where k is a preset scaling factor; the higher the risk value, the larger the coefficient. Alternatively, the modulation function can be a non-linear function, such as an exponential or logarithmic function, to more precisely capture the impact of different risk levels on the width of the uncertainty interval—for example, adjusting smaller values ​​for low risk and larger values ​​for high risk.

[0090] The uncertainty interval is obtained by modulating the width of the basic confidence interval using an interval adjustment coefficient. Its function is to combine the model's inherent predictive uncertainty with external data quality risks, generating a dynamically adjusted uncertainty interval that comprehensively reflects the current predictive reliability. Specifically, the half-width of the basic confidence interval (i.e., half the interval width) can be multiplied by the interval adjustment coefficient to obtain the modulated half-width. Then, a new uncertainty interval is constructed centered on the prediction result. Alternatively, the interval adjustment coefficient can be directly applied to the upper and lower limits of the basic confidence interval. For example, the new upper limit = prediction result + (basic upper limit - prediction result) × coefficient, and the new lower limit = prediction result - (prediction result - basic lower limit) × coefficient.

[0091] The aforementioned technical solution enables the uncertainty interval to respond in real-time to changes in data quality risk values, accurately capturing the additional uncertainty introduced by data fluctuations. When data quality is poor (higher risk), the uncertainty interval widens accordingly, transmitting a higher level of uncertainty signal to subsequent optimization decision-making processes, prompting the system to adopt a more conservative optimization strategy. Conversely, when data quality is high (lower risk), the uncertainty interval narrows, allowing for more precise optimization. This directly solves the problem of treating predicted results as deterministic values, improving the anti-interference capability and robustness of recommended ratio schemes in actual production, thereby preventing deviations between performance and cost targets and ensuring the reliability of optimization decisions.

[0092] This application further proposes an intelligent analysis method for cement material proportioning. Based on the predicted performance of the proportioning and the uncertainty interval, and according to the optimization objective set corresponding to the risk level, it performs multi-objective optimization to generate a recommended proportioning scheme. The method includes: selecting a target optimization strategy from multiple pre-set optimization strategies based on the risk level and the corresponding optimization objective set. This target optimization strategy defines constraints and search space rules that match the risk level and the optimization objective set. Based on the search space rules defined by the target optimization strategy, the feasible parameter space for this optimization is determined. Within this feasible parameter space, based on the constraints defined by the target optimization strategy, and integrating at least one performance objective and at least one cost objective defined by the optimization objective set, a multi-objective optimization model is constructed. These constraints include the utilization of the uncertainty interval. The multi-objective optimization model is solved to obtain a Pareto non-dominated solution set. According to pre-set decision rules, a recommended proportioning scheme is determined from the Pareto non-dominated solution set.

[0093] Specifically, this risk level is a predefined level associated with the dynamic data quality risk value after dynamic quality risk assessment and collaborative enhancement processing of multi-source heterogeneous data on cement production. It is used to characterize the reliability level of the current input data.

[0094] When determining the feasible parameter space for this optimization, this method relies on the search space rules defined by the selected objective optimization strategy. For example, if the search space rules indicate that a wider range of parameters can be explored under low-risk conditions, the feasible parameter space will expand accordingly. If a more conservative parameter range needs to be limited under high-risk conditions, the feasible parameter space will shrink accordingly. One implementation is to parse the search space rules, filter matching data similar to the current risk level from historical data, and construct the feasible parameter space based on the distribution of these data. Another implementation is to directly select a feasible parameter space from a predefined parameter space template based on the risk level.

[0095] The performance objective could be the compressive strength or setting time of cement, while the cost objective could be raw material cost or energy cost. This multi-objective optimization model is a mathematical model that integrates these objective functions and constraints. The constraints include the utilization of uncertainty intervals, meaning that the uncertainty of the prediction results is no longer simply ignored but is considered an important factor in the optimization process.

[0096] This method solves the multi-objective optimization model to obtain a Pareto non-dominated solution set. This Pareto non-dominated solution set is a set of solutions where any one solution cannot improve any objective without sacrificing at least one other objective. These solutions represent the optimal trade-offs between the different objectives.

[0097] This method determines the recommended allocation scheme from the Pareto non-dominated solution set based on preset decision rules. These preset decision rules can be set according to actual production needs and preferences. For example, a comprehensive score can be calculated for each solution in the Pareto solution set based on cost-effectiveness, performance stability, or the importance of specific production objectives, and the solution with the highest score can be selected as the recommended allocation scheme. Alternatively, an interactive decision support system can be used, allowing users to manually select the allocation scheme that best suits their preferences based on a visualized Pareto front.

[0098] Through the above technical solution, this application can adaptively adjust the optimization strategy, search space, and constraints based on dynamically changing data quality risks, thereby constructing a more robust multi-objective optimization model. Incorporating the prediction uncertainty range into the optimization constraints ensures that the generated recommended mix proportions not only consider performance and cost objectives but also fully account for the prediction risks caused by data quality fluctuations, effectively preventing deviations from performance and cost objectives due to data uncertainty in actual production. This dynamic adaptive optimization mechanism enhances the anti-interference capability and practical application effect of cement material proportioning under complex working conditions, ensuring the reliability and stability of the optimization scheme.

[0099] This application further proposes a method for determining the feasible parameter space for this optimization based on the search space rules defined by the target optimization strategy. This method includes: parsing the search space rules defined by the target optimization strategy to obtain historical data filtering sub-rules and parameter space construction sub-rules; based on the historical data filtering sub-rules, selecting a subset of reference data from historical proportioning data and corresponding dynamic data quality risk value records; based on the proportioning parameter distribution in the reference data subset, applying parameter space construction sub-rules to generate a candidate parameter space; and taking the intersection of the candidate parameter space and the full parameter space defined by the physical allowable range of material proportioning parameters to obtain the feasible parameter space for this optimization.

[0100] Specifically, the target optimization strategy refers to the specific optimization method or algorithm selected during multi-objective optimization based on the current data quality risk level and the set of optimization objectives. It can be one of several pre-defined strategies, each potentially corresponding to different optimization algorithms, constraint handling methods, or search mechanisms. For example, when the data quality risk is high, a conservative strategy with a smaller search range might be chosen. When the risk is low, a more aggressive strategy with a broader search range might be selected. This can be implemented through a strategy selection module that retrieves and activates the corresponding optimization strategy configuration from a predefined strategy library based on the input risk level and set of optimization objectives. Alternatively, it can be achieved through a rule-based expert system that dynamically generates or combines the components of the optimization strategy based on the current system state (such as risk level and performance objective priority).

[0101] The search space rules refer to a series of criteria or algorithms used to define the range and distribution characteristics of decision variables (i.e., material proportioning parameters) in an optimization problem. These rules guide the optimizer to find the optimal solution, ensuring the effectiveness and efficiency of the search. They can be implemented as structured data formats (such as JSON or XML), containing upper and lower limits for parameters, discretization step size, and statistical descriptions based on historical data distribution (such as mean, variance, and cluster centers). Alternatively, they can exist as executable code modules that, when called, dynamically calculate and return the boundaries or distribution characteristics of the search space based on the input conditions.

[0102] This historical data filtering sub-rule refers to the criteria used to select a subset of reliable data relevant to the current optimization task from a large amount of historical matching data. Its purpose is to ensure that the data used to construct the search space is representative and credible. This can be implemented by filtering based on a time window, such as selecting only data from the most recent N days or N weeks; or by filtering based on a data quality threshold, such as retaining only data with a dynamic data quality risk value below a certain threshold; or by filtering based on similarity measures, such as selecting historical data similar to the current production conditions or target matching parameters in certain key characteristics; or by using expert experience rules, such as excluding data recorded during specific abnormal events.

[0103] This parameter space construction sub-rule refers to a specific algorithm or method used to construct or define a candidate parameter space based on the distribution characteristics of a selected subset of reference data. Its purpose is to transform historical experience into an effective search region. This can be implemented using statistical methods, such as calculating the mean and standard deviation of each matching parameter in the reference dataset, and then defining a hyperrectangular or hyperellipsoidal space centered on the mean and with the standard deviation as its radius. Alternatively, it can be based on machine learning methods, such as performing cluster analysis on the reference data and defining the cluster centers and their surrounding areas as the candidate space. Or, density estimation methods can be used to select high-density regions as candidate spaces.

[0104] Historical material proportioning data refers to the collection of material proportioning schemes actually used in past production processes, along with their corresponding production conditions and performance results. This data forms the basis for learning and optimizing material proportions. It can be stored in a relational database, containing fields such as proportioning parameters, production date, and batch number. Alternatively, it can be stored in a data lake or distributed file system in unstructured or semi-structured form, such as CSV or Parquet files.

[0105] This dynamic data quality risk value record refers to a dynamic assessment result reflecting the data quality level, associated with historical matching data. These risk values ​​help the system identify and exclude low-quality data. It can be implemented by storing it in the database along with historical matching data as an attribute field for each data record, or by storing it separately in a time-series database and associating it with historical matching data via timestamps or batch numbers.

[0106] This reference data subset refers to the set of data samples selected from all historical data after processing by historical data filtering sub-rules, which are considered to have high quality and relevance. This subset will be used for subsequent candidate parameter space construction. It can be implemented by using a database query language (such as SQL) combined with filtering conditions to extract data rows that meet the criteria from the historical database. Alternatively, it can be implemented by loading historical data using data processing scripts (such as Python Pandas) and applying filtering functions to generate a new data frame.

[0107] The distribution of these proportioning parameters refers to the statistical characteristics of each material proportioning parameter in the reference data subset, such as their mean, median, variance, skewness, kurtosis, and the correlation between them. This can be achieved by using statistical analysis tools (such as NumPy and SciPy) to calculate the descriptive statistics and covariance matrix of each parameter, or by using visualization tools (such as Matplotlib and Seaborn) to plot histograms and scatter plots of the parameters for a visual analysis of their distribution characteristics.

[0108] The candidate parameter space refers to a preliminary region that may contain the optimal solution, generated by constructing sub-rules from the parameter space based on the distribution characteristics of a subset of the reference data. This space is typically smaller than the full parameter space but larger than the final feasible parameter space. It can be implemented as a hyperrectangle, whose boundaries are determined by the minimum and maximum values ​​of the parameters in the reference dataset, or by the mean plus or minus a certain factor of the standard deviation. Alternatively, it can be defined as a hyperellipsoid, centered on the parameter mean, with its shape determined by the covariance matrix, better reflecting the correlation between parameters.

[0109] The physical allowable range of this material proportioning parameter refers to the actual physical boundaries that the material proportioning parameter can take, determined according to physical laws, chemical reaction principles, production process limitations, or safety regulations. For example, the addition ratio of a certain material cannot be negative, nor can it exceed 100%. This can be implemented by storing it as a hard-coded constant or as a parameter in a configuration file, such as min_cement_ratio=0.2, max_water_cement_ratio=0.5. Alternatively, it can be integrated with a production process expert system or simulation model to dynamically obtain or verify the physical allowable range of the parameter.

[0110] The full parameter space refers to the multidimensional space comprised of all possible combinations of material proportioning parameters within their respective physical limits. This is the theoretically complete set of all possible proportioning schemes. It can be implemented as a Cartesian product of the physical limits of each parameter, forming a multidimensional hyperrectangle. Alternatively, in some cases, physical constraints may exist, making the full parameter space not a simple hyperrectangle, but a complex region defined by a series of inequalities.

[0111] The feasible parameter space refers to the region within the full parameter space that simultaneously satisfies the definition of the candidate parameter space. This space is the actual region for the final multi-objective optimization search. It considers historical data experience, adheres to physical constraints, and can adapt to changes in data quality risks through dynamic adjustment. It can be implemented by using the intersection operation of set theory to compare the boundary of the candidate parameter space with the boundary of the full parameter space, and taking the overlapping portion as the feasible parameter space. Alternatively, in the optimization algorithm, the definition of the feasible parameter space can be used as one of the constraints, guiding the optimizer to search within the region that satisfies these constraints.

[0112] The above technical solutions can improve the robustness and practical application effect of the ratio optimization scheme, so that the optimization process can still accurately and effectively find the optimal ratio scheme when facing dynamically changing data quality.

[0113] This application further proposes a method for generating a candidate parameter space based on the matching parameter distribution in the reference data subset and applying sub-rules to the parameter space construction. The method includes: determining whether the reference data subset has high-risk or low-risk characteristics based on the statistical characteristics of the dynamic data quality risk values. If it is determined to have high-risk characteristics, the matching parameter distribution is subjected to a first type of processing using the parameter space construction sub-rules. This first type of processing includes: calculating the mean vector and covariance matrix of the matching parameter distribution, and constructing a multidimensional ellipsoidal space centered on the mean vector and bounded by multiple standard deviation ranges determined by the covariance matrix, as the candidate parameter space. If it is determined to have low-risk characteristics, the matching parameter distribution is subjected to a second type of processing using the parameter space construction sub-rules. This second type of processing includes: performing cluster analysis on the matching parameter distribution to identify at least one core data cluster, and constructing a convex hull space or multidimensional bounding box space that can encompass all identified core data clusters, as the candidate parameter space.

[0114] Based on the statistical characteristics of the dynamic data quality risk values ​​of the reference data subset, it is determined whether the reference data subset has high-risk or low-risk characteristics. This step aims to assess the overall reliability level of the data subset by analyzing the dynamic data quality risk values ​​associated with it, thereby selecting an appropriate strategy for the subsequent construction of the parameter space. Specifically, the average or median of all dynamic data quality risk values ​​in the reference data subset can be calculated and compared with a preset risk threshold. If it exceeds the threshold, it is judged as high-risk; otherwise, it is judged as low-risk. Alternatively, the proportion of samples in the reference data subset with risk values ​​higher than a certain threshold can be statistically analyzed. If this proportion exceeds a preset percentage, it is judged as high-risk.

[0115] If the data is determined to have high-risk characteristics, the first type of processing is applied to the matching parameter distribution using the sub-rules constructed from the parameter space. This first type of processing aims to construct a relatively conservative and robust candidate parameter space using statistical methods when there is high uncertainty in data quality, thereby reducing the negative impact of noisy data on the optimization process. This first type of processing includes: calculating the mean vector and covariance matrix of the matching parameter distribution, and constructing a multidimensional ellipsoidal space centered on the mean vector and bounded by multiple standard deviation ranges determined by the covariance matrix, as the candidate parameter space. For example, the standard maximum likelihood estimation method can be used to calculate the mean vector and covariance matrix, and then, based on the properties of the multivariate normal distribution, a chi-square distribution critical value corresponding to a confidence level (such as 95% or 99%) can be selected to determine the ellipsoid boundary. Furthermore, to enhance robustness to potential outliers in high-risk data, robust statistical methods (such as minimum covariance determinant MCD estimation) can also be used to calculate the mean vector and covariance matrix, making the constructed ellipsoidal space less sensitive to outliers.

[0116] If the distribution is determined to have low-risk characteristics, the second type of processing is applied to the distribution by constructing sub-rules in the parameter space. This second type of processing aims to more precisely identify reliable and dense regions in the data when the data quality is high, thereby constructing a more accurate and potentially broader candidate parameter space to fully explore optimization potential. This second type of processing includes: performing cluster analysis on the distribution to identify at least one core data cluster, and constructing a convex hull space or multidimensional bounding box space that can encompass all identified core data clusters as the candidate parameter space. For example, clustering algorithms such as K-Means and DBSCAN can be used to cluster the distribution, and then core data clusters can be identified based on the density of the clustering results, the number of samples, or the closeness to historical optimal solutions. After identifying the core data clusters, the convex hulls of all data points in these core data clusters can be calculated to form a polygonal or polyhedral space as the candidate parameter space. Alternatively, the maximum and minimum values ​​of each core data cluster in each ratio parameter dimension can be calculated, and then the union of the maximum and minimum values ​​of all core data clusters can be taken to construct a multidimensional bounding box space that can contain all core data clusters.

[0117] The above technical solution enables more precise capture of dense areas of high-quality data, thereby generating a more refined and representative candidate parameter space. This fully utilizes the information contained in the high-quality data, providing a broader and more effective search range for multi-objective optimization. This risk-driven dynamic processing mechanism allows the generation of the candidate parameter space to effectively address the dynamic changes in cement production data, improving the reliability and anti-interference capabilities of subsequent multi-objective optimization processes. It ensures that more accurate and robust recommended mix proportions can be generated under different data quality conditions.

[0118] This application further proposes to construct a multi-objective optimization model within the feasible parameter space, based on the constraints defined by the objective optimization strategy, and integrating at least one performance objective and at least one cost objective defined by the optimization objective set. The specific steps include: The optimization objective set is used to extract the performance and cost objectives to be optimized, and these objectives are then constructed as performance objective functions and cost objective functions, respectively. The feasible parameter space defines the effective range of material proportioning parameters that the optimization algorithm can search, ensuring that the generated proportioning scheme is physically feasible and meets production requirements. For example, this feasible parameter space can be defined by preset hard upper and lower limits for material proportioning parameters, setting the content range of each component based on industry standards or safety regulations. Alternatively, the feasible parameter space can be dynamically delineated based on historical production data and expert experience, using statistical analysis or machine learning methods (such as boundary detection and convex hull algorithms) to define a more refined region that conforms to actual production conditions. The constraints defined by the objective optimization strategy specify various restrictions that must be met during the optimization process to ensure that the generated proportioning scheme meets requirements in terms of performance, cost, and production process. For example, these constraints can be hard constraints, such as specific performance indicators of cement (such as compressive strength and setting time) must meet minimum or maximum requirements. Alternatively, these constraints can be resource or process constraints, such as supply limitations of specific raw materials, capacity limitations of production equipment, or requirements for mixing uniformity. The optimization objective set defines at least one performance objective and at least one cost objective, clarifying multiple conflicting or interrelated optimization directions that need to be considered simultaneously in the optimization process to achieve a comprehensive optimum. For example, the performance objective could be the mechanical properties of cement, such as compressive strength and flexural strength at different ages. Alternatively, it could be the durability or environmental performance of cement, such as resistance to carbonation, heat of hydration, and carbon emissions. The cost objective could be the raw material procurement cost per unit of cement product. Alternatively, it could be the energy consumption cost or waste disposal cost during the production process. This multi-objective optimization model is a mathematical model designed to simultaneously optimize two or more conflicting objective functions. It typically does not have a single optimal solution but rather yields a set of Pareto optimal solutions. For example, this multi-objective optimization model can employ evolutionary algorithm-based methods, such as Non-Dominated Sorting Genetic Algorithm-II (NSGA-II) or Multi-Objective Particle Swarm Optimization (MOPSO), to approximate the Pareto front through iterative search. Alternatively, this multi-objective optimization model can employ mathematical programming methods, such as the weighted sum method or the ε-constraint method, to transform the multi-objective problem into a single-objective problem or a series of single-objective problems for solution. This method parses the performance and cost objectives to be optimized, aiming to identify the specific performance and cost indicators that need to be focused on in this optimization task from a predefined set of optimization objectives. For example, these can be manually selected and configured through a user interface or configuration file to explicitly specify the performance objectives (such as 28-day compressive strength) and cost objectives (such as material costs) to be considered in this optimization. Alternatively, the system can automatically select the most relevant performance and cost objectives from a preset objective library based on the current production batch, product type, or risk level.This structure comprises performance and cost objective functions, aiming to transform abstract performance and cost objectives into mathematical expressions that can be processed by optimization algorithms. These functions take material proportioning parameters as input and output corresponding performance or cost values. For example, the performance objective function could be a mathematical expression of the proportioning performance prediction results obtained based on hybrid predictive modeling.

[0119] Based on the constraints defined by the target optimization strategy, the width of the uncertainty interval is transformed into a limitation on the fluctuation range of the performance objective function value. This transformation aims to quantify the uncertainty of the prediction result into a constraint on the optimization objective function, thereby ensuring that the risks arising from data quality fluctuations are considered during the optimization process and improving the robustness of the optimization scheme. For example, the performance objective function value f_performance(x) can be added to or subtracted from a multiple of the half-width of the uncertainty interval as a new constraint, requiring f_performance(x) - k × uncertainty_half_width(x) >= target_min_performance, where k is a safety factor. Alternatively, the width of the uncertainty interval itself can be used as an additional constraint, requiring it not to exceed a preset threshold, for example, uncertainty_width(x) <= max_allowed_uncertainty, to ensure the reliability of the prediction.

[0120] Using material proportioning parameters as decision variables, a multi-objective optimization model is constructed within the feasible parameter space, simultaneously incorporating performance and cost objective functions and subject to one or more constraints. These material proportioning parameters, as decision variables, aim to clarify the parameters that the optimization algorithm can adjust during the search process; these parameters directly determine the final cement formulation. For example, the material proportioning parameter can refer to the mass percentage or relative proportion of each major component in the cement (such as clinker, limestone, fly ash, slag, gypsum, etc.). Alternatively, the material proportioning parameter can also be the absolute mass or volume of each component, adjustable while satisfying the total mass or total volume constraints, or it can consider the type and amount of specific additives.

[0121] Through the above technical solutions, even when data quality fluctuates, the system can prioritize the proportioning schemes with more stable prediction results and lower uncertainty, thereby avoiding deviations in actual performance caused by prediction uncertainty. Using material proportioning parameters as decision variables, a multi-objective optimization model is constructed within the feasible parameter space, incorporating performance objective functions, cost objective functions, and uncertainty constraints. This ensures that the generated recommended proportioning schemes meet performance and cost requirements while possessing stronger anti-interference capabilities and robustness, improving the reliability of proportioning optimization under complex operating conditions.

[0122] This application further proposes a method for determining recommended proportioning schemes from a Pareto non-dominated solution set based on preset decision rules. The method includes: obtaining a decision preference configuration corresponding to a risk level, which defines weighting coefficients for multiple attributes used to evaluate solutions; for each solution in the Pareto non-dominated solution set, using these weighting coefficients to weight and synthesize the predicted proportioning performance value, the material cost value, and the corresponding prediction uncertainty interval width of each solution to calculate the comprehensive utility value of each solution; comparing the comprehensive utility values ​​of all solutions in the Pareto non-dominated solution set, selecting the solution with the highest comprehensive utility value, and outputting the material proportioning parameters corresponding to the solution with the highest comprehensive utility value as the recommended proportioning scheme.

[0123] This involves obtaining a decision preference configuration corresponding to the risk level. This configuration defines the weighting coefficients of multiple attributes used to evaluate the solution, aiming to dynamically adjust the focus of the decision based on the current data quality risk level. The decision preference configuration can be a predefined mapping table that associates different risk levels (e.g., low, medium, high risk) with a specific set of weighting coefficients. For example, in a low-risk scenario, there might be more emphasis on performance and cost optimization, while in a high-risk scenario, greater weight might be given to predictive uncertainty to ensure the robustness of the solution. Another implementation approach is that the decision preference configuration can consist of a set of parameterized functions that take the risk level as input and dynamically calculate the weighting coefficients of each attribute. For example, as the risk level increases, the weighting coefficients related to uncertainty might be adjusted using an increasing function, while the weighting coefficients related to performance or cost might be adjusted using a decreasing function.

[0124] For each solution in the Pareto non-dominated solution set, a weighted summation is performed on the predicted performance value, material cost value, and corresponding prediction uncertainty interval width of each solution using weighting coefficients. The goal is to integrate multiple conflicting optimization objectives and prediction uncertainties into a single, quantifiable index for unified comparison. One implementation method is a linear weighted summation, where the comprehensive utility value equals the sum of the products of each attribute value (such as the predicted performance value, material cost value, and prediction uncertainty interval width) and its corresponding weighting coefficient. For example, comprehensive utility value = W_performance × performance value + W_cost × cost value - W_uncertainty × uncertainty interval width (where cost and uncertainty are usually the objectives to be minimized, so their weighting coefficients may be negative or reversed before calculation). Another implementation method is a weighted product method, where the comprehensive utility value equals the product of powers of each attribute value, where the power is determined by the corresponding weighting coefficient. This method is more suitable when there are non-linear relationships between attributes.

[0125] This approach compares the overall utility values ​​of all solutions in the Pareto non-dominated solution set, selects the solution with the highest overall utility value, and outputs the material proportioning parameters corresponding to this solution as the recommended proportioning scheme. The goal is to determine the optimal final recommended scheme from numerous feasible schemes that aligns with the current risk appetite. One implementation is to directly iterate through all solutions in the Pareto non-dominated solution set, calculate the overall utility value of each solution, and simply select the solution with the highest overall utility value as the recommended scheme. Another implementation involves introducing additional decision rules to refine the selection when multiple solutions have very close or identical highest overall utility values. For example, prioritizing the scheme with the lowest cost or the smallest uncertainty interval width can further optimize the decision.

[0126] The above technical solutions can ensure that the recommended proportioning scheme has stronger anti-interference ability in actual production, reduce the risk of deviation from performance and cost targets, and thus improve the reliability and practicality of proportioning optimization.

[0127] This application proposes an intelligent analysis method for cement material proportioning, which further includes: based on the deviation between actual production feedback and the predicted performance of the proportioning, coordinating the adjustment of parameters for dynamic quality risk assessment, thermodynamic simulation boundary conditions for synergistic enhancement processing, fusion weights for hybrid prediction modeling, and criteria for risk level. Specifically, the coordinating adjustment process includes: comparing and analyzing the sequence of deviations with the sequence of uncertainty intervals to identify systematic prediction deviations exceeding the uncertainty interval range. The causes of these systematic prediction deviations are categorized into one or more of three categories: data quality, physical mechanism, or model fusion, yielding a causal classification result. Based on this causal classification result, a first type of parameter is generated for updating the evaluation weights in the dynamic quality risk assessment, a second type of parameter is generated for correcting the thermodynamic simulation boundary conditions in the synergistic enhancement processing, and a third type of parameter is generated for fine-tuning the fusion weight allocation strategy in the hybrid prediction modeling. Using parameter iteration, the first type of parameter, the second type of parameter, and the third type of parameter are applied to their respective processes to complete the coordinating adjustment.

[0128] Specifically, by comparing the sequence of this deviation with the sequence within the uncertainty range, systematic prediction biases exceeding the uncertainty range are identified. The aim is to filter out significant prediction errors that are not random fluctuations but have persistence or trends from the differences between actual production feedback and prediction results. This ensures that subsequent adjustments target systemic problems rather than random noise.

[0129] The causes of this systematic prediction bias are categorized into one or more of three categories: data quality, physical mechanism, or model fusion. The purpose of this categorization is to accurately pinpoint the root cause of the systematic prediction bias.

[0130] Based on the causal classification results, a first type of parameter is generated to update the evaluation weight in the dynamic quality risk assessment, a second type of parameter is generated to correct the thermodynamic simulation boundary conditions in the synergistic enhancement treatment, and a third type of parameter is generated to fine-tune the fusion weight allocation strategy in the hybrid prediction model. The aim is to generate targeted adjustment parameters based on the identified causes.

[0131] By using parameter iteration, the first type of parameter, the second type of parameter, and the third type of parameter are applied to the corresponding processes to complete the coordinated adjustment. The purpose is to apply the generated adjustment parameters to the system in a stable and controlled manner, avoiding system instability caused by a one-time large adjustment.

[0132] Through the above technical solutions, the entire intelligent analysis system for cement material proportioning can continuously learn and optimize from actual production, dynamically adapt to complex factors such as raw material fluctuations and changes in operating conditions, thereby improving the accuracy of proportioning performance prediction, the reliability of optimization schemes, and the anti-interference ability in actual production, ultimately achieving better performance and cost targets, and solving the problem of persistent prediction deviations and the impact on the reliability of optimization schemes in existing technologies.

[0133] This application further proposes a comparative analysis of the bias sequence and the uncertainty interval sequence to identify systematic prediction biases exceeding the uncertainty interval range. Specifically, this includes: determining the ratio sequence between the absolute value sequence of the bias and the half-width sequence of the uncertainty interval; statistically analyzing the proportion of elements with values ​​greater than 1 in the ratio sequence; if this proportion exceeds a first proportion threshold, a systematic bias determination is triggered; and performing a statistical trend test on the ratio sequence. If the test results indicate that the ratio sequence exhibits a monotonically increasing or decreasing trend, and the statistical significance level of this trend is below a significance threshold, then a systematic prediction bias is determined to exist.

[0134] The step of determining the ratio between the absolute value sequence of the deviation and the half-width sequence of the uncertainty interval aims to quantify the magnitude of the prediction deviation relative to its own uncertainty range. The absolute value sequence of the deviation represents the absolute value of the difference between the actual production feedback and the predicted performance, reflecting the degree of deviation between the prediction model and the actual situation. The half-width sequence of the uncertainty interval represents half the width of the confidence interval of the prediction result, characterizing the reliability range of the prediction result. By calculating the ratio between these two, a standardized, dimensionless index can be obtained to measure whether the prediction deviation exceeds the reliability range estimated by the model itself. Specifically, a direct division method can be used: ratio = absolute value of deviation / half-width value. Alternatively, a normalization process can be used, for example, ratio = (absolute value of deviation - half-width value) / half-width value, where a ratio greater than 0 indicates that the uncertainty interval has been exceeded.

[0135] The proportion of elements with values ​​greater than 1 in the ratio sequence is counted. If this proportion exceeds a first proportion threshold, a systematic bias determination is triggered. This step is used to initially determine the frequency with which prediction deviations exceed the uncertainty interval. Elements with values ​​greater than 1 in the ratio sequence indicate that the absolute value of the prediction deviation has exceeded the half-width of the uncertainty interval, meaning the actual value falls outside the prediction confidence interval. By counting the proportion of these elements exceeding the interval, the stability of the prediction model over a period of time can be assessed. If this proportion exceeds a preset "first proportion threshold," it indicates that the model's prediction results frequently deviate from its own estimated reliability range, potentially indicating a systematic problem, thus triggering further systematic bias determination. This first proportion threshold can be a fixed percentage, for example, set to 20% or 30%, meaning that if more than 20% or 30% of the predicted values ​​exceed their uncertainty interval within a certain observation period, a determination is triggered. Alternatively, this first proportion threshold can be dynamically adjusted based on historical data or the tolerance of a specific application scenario; for example, a dynamic threshold can be determined based on the distribution of the proportion of deviations under historical normal operating conditions.

[0136] A statistical trend test is performed on the ratio sequence. If the test results indicate that the ratio sequence exhibits a monotonically increasing or decreasing trend, and the statistical significance level of this trend is below a significance threshold, then a systematic prediction bias is identified. This step aims to identify whether the prediction bias exhibits persistent and directional changes to distinguish between random fluctuations and systematic problems. A monotonically increasing or decreasing trend indicates that the prediction bias is not a random fluctuation, but rather increases or decreases continuously with changes in time or certain conditions, which is usually a manifestation of a systematic problem. Statistical trend tests (e.g., Mann-Kendall test, linear regression slope significance test, etc.) can quantify the existence and strength of this trend. Furthermore, by determining whether the statistical significance level of this trend is below a preset "significance threshold" (e.g., p-value less than 0.05), it can be ensured that the identified trend is not accidental, but a statistically significant true trend. For example, the non-parametric Mann-Kendall test can be used to detect whether the ratio sequence has a monotonically increasing trend; this test does not require the data to follow a specific distribution. Another approach is to use linear regression analysis, with the ratio series as the dependent variable and the time series as the independent variable, and then test the significance of the regression coefficients (slopes). This significance threshold is usually set at 0.05 or 0.01 to determine whether the observed trend is statistically reliable.

[0137] The above technical solutions ensure the accuracy and reliability of the judgment of systematic prediction deviations, and provide a solid foundation for the subsequent coordinated adjustment of dynamic quality risk assessment parameters, collaborative enhancement treatment boundary conditions, and hybrid prediction modeling fusion weights. This enhances the robustness and adaptability of the entire intelligent analysis method for cement material proportioning under complex working conditions.

[0138] This application further proposes a method for classifying the causes of systematic prediction bias into one or more of three categories: data quality, physical mechanism, or model fusion, to obtain the causal classification result. The specific steps include: obtaining the deviation sequence corresponding to the systematic prediction bias, and obtaining the material proportioning parameter sequence, dynamic data quality risk value sequence, and historical fusion weights of the hybrid prediction model that are temporally associated with the deviation sequence. Determining the first correlation strength between the deviation sequence and the material proportioning parameter sequence, and the second correlation strength between the deviation sequence and the dynamic data quality risk value sequence. Determining the statistical distribution differences of the deviation sequence under different value ranges of the historical fusion weights. Based on the first correlation strength, the second correlation strength, and the statistical distribution differences, and according to a preset causal discrimination rule, classifying the causes of systematic prediction bias into one or more of the categories of data quality, physical mechanism, or model fusion, to obtain the causal classification result.

[0139] The system acquires the deviation sequence corresponding to the systematic prediction deviation, and retrieves the material proportioning parameter sequence, dynamic data quality risk value sequence, and historical fusion weight history of the hybrid predictive modeling that are temporally associated with the deviation sequence. This aims to provide a comprehensive and time-aligned data foundation for subsequent causal analysis. Specifically, the system can be configured with a data acquisition module to extract, in real time, the deviation sequence, material proportioning parameter sequence, dynamic data quality risk value sequence, and historical fusion weight history of the hybrid predictive modeling corresponding to the time period of the systematic prediction deviation from the production database, sensor network, or historical data storage. These data sequences are aligned and correlated using timestamps to ensure data synchronization. Alternatively, a streaming data processing framework can be used to subscribe to relevant data streams. When a systematic prediction deviation is detected, a historical data backtracking mechanism is triggered to retrieve all relevant sequence data within a specified time window from the data lake or data warehouse, and preprocessing is performed to ensure consistency in data format and temporal granularity.

[0140] The purpose of determining the primary correlation strength between the deviation series and the material proportioning parameter series, and the secondary correlation strength between the deviation series and the dynamic data quality risk value series, is to quantify the statistical relationship between predicted deviations and potential influencing factors (such as material proportioning parameters and data quality risk). This helps identify which factors are significantly associated with the deviation. For example, statistical methods such as Pearson correlation coefficient, Spearman's rank correlation coefficient, or mutual information can be used to calculate the primary correlation strength between the deviation series and the material proportioning parameter series, and the secondary correlation strength between the deviation series and the dynamic data quality risk value series. These correlation indices can reflect the degree of linear or non-linear dependence between the two series. Furthermore, Granger causality tests or vector autoregression (VAR) models can be used to analyze the lead-lag relationship and causal strength between the deviation series and the material proportioning parameter series and the dynamic data quality risk value series, thereby revealing their dynamic associations more deeply and quantifying the correlation strength.

[0141] Determining the statistical distribution differences of the deviation sequence across different value segments of the fusion weight history serves to analyze whether the distribution characteristics of the prediction deviation change with variations in the model fusion weights, thereby determining whether the model fusion strategy is a potential cause of systematic bias. Specifically, the fusion weight history can be divided into several discrete value segments, such as low-weight, medium-weight, and high-weight segments. Then, the statistical distribution of the deviation sequence within each segment, such as mean, variance, skewness, and kurtosis, is calculated, and statistical methods such as chi-square test, Kolmogorov-Smirnov test, or t-test are used to assess whether there are significant differences between these distributions. Another approach is to use nonparametric density estimation methods, such as kernel density estimation, to estimate the probability density function of the deviation sequence across different fusion weight segments, and quantify the statistical distribution differences by comparing the overlap or distance of these probability density functions, such as the Wasserstein distance.

[0142] Based on the strength of the first correlation, the strength of the second correlation, and the statistical distribution difference, and according to predefined causal discrimination rules, the causes of systematic prediction bias are categorized into one or more of the following categories: data quality, physical mechanism, or model fusion, resulting in a causal classification result. This step synthesizes the aforementioned analysis results and objectively identifies the root cause of systematic prediction bias using a predefined set of logic or models. The predefined causal discrimination rules can be a set of threshold-based logical judgments. For example, if the strength of the first correlation exceeds a certain predefined threshold, the physical mechanism category is tended to be added to the candidate causal set. If the strength of the second correlation exceeds another predefined threshold, the data quality category is tended to be added to the candidate causal set. If the significance level of the statistical distribution difference is below a certain predefined threshold, the model fusion category is tended to be added to the candidate causal set. These rules can be combined to identify multiple causes. In addition, a classification model based on decision trees, support vector machines (SVMs) or neural networks can be constructed, using the strength of the first correlation, the strength of the second correlation, and the statistical distribution difference as input features. By training historical data (including systematic prediction bias cases with labeled causes) to learn the cause discrimination rules, systematic prediction bias can be automatically classified into one or more of the data quality, physical mechanism, or model fusion categories.

[0143] Through the above technical solutions, this application can collect multi-dimensional data sequences based on time correlation, avoiding isolated analysis and providing a complete and real-time data foundation for cause identification, ensuring that the analysis covers the entire production cycle. By quantifying the relationship between deviation and material ratio, it helps identify physical mechanism-related causes and captures the direct impact of ratio changes on deviation. Simultaneously, by correlating deviation with data quality risks, it identifies data quality-related causes, reflecting the contribution of data fluctuations to deviation. By analyzing the distribution changes of deviation under different fusion weights, it identifies model fusion-related causes and reveals the systematic impact of model strategy adjustments on deviation. Finally, by integrating multiple quantitative indicators and preset rules, it achieves objective, multi-category cause classification, avoiding subjective misjudgments and ensuring comprehensive and reliable classification results. This provides an accurate basis for subsequent parameter adjustments, thereby improving the accuracy and comprehensiveness of systematically predicting deviation cause classification, and further enhancing the effectiveness of subsequent parameter adjustments and the overall robustness of the system.

[0144] This application further proposes classifying the causes of this systematic prediction bias into one or more of three categories: data quality, physical mechanism, or model fusion, to obtain the cause classification result. Specifically, this includes: determining whether the strength of the first correlation exceeds a first correlation threshold; if so, adding the physical mechanism category to the candidate cause set. Determining whether the strength of the second correlation exceeds a second correlation threshold; if so, adding the data quality category to the candidate cause set. Determining whether the significance level of the statistical distribution difference is lower than a fusion significance threshold; if so, adding the model fusion category to the candidate cause set. All categories in the candidate cause set are then identified as the cause classification result.

[0145] The first correlation strength quantifies the degree of association between the systematic prediction deviation sequence and the material proportioning parameter sequence. For example, it can be calculated using statistical indicators such as Pearson correlation coefficient and Spearman rank correlation coefficient, or assessed using methods such as mutual information and Granger causality tests. The first correlation threshold is a pre-defined critical value used to determine whether the first correlation strength reaches a significant level sufficient to indicate a problem with the physical mechanism. This threshold can be set based on domain expert experience, such as 0.7 or 0.8, or optimized through statistical analysis of historical data or training of machine learning models. The physical mechanism category indicates the type of cause that primarily stems from an improperly set physical mechanism model or its parameters. The candidate cause set is a dynamically constructed set used to temporarily store the cause categories identified during the current judgment process that may lead to systematic prediction deviations.

[0146] The second correlation strength quantifies the degree of association between the systematic prediction bias sequence and the dynamic data quality risk value sequence. Its calculation method can be similar to the first correlation strength, for example, using Pearson correlation coefficient, Spearman rank correlation coefficient, or time series cross-correlation analysis. The second correlation threshold is a pre-defined critical value used to determine whether the second correlation strength reaches a significant level sufficient to indicate a data quality problem. This threshold can also be set empirically or calibrated using data-driven methods. The data quality category indicates the type of cause of the systematic prediction bias, primarily stemming from input data quality problems (such as inaccurate, incomplete, or anomaly-related data).

[0147] The significance level of this statistical distribution difference is used to assess whether there are significant differences in the statistical distribution of the systematic prediction bias sequence across different ranges of the historical fusion weights in the mixed prediction model. For example, non-parametric tests such as the Kolmogorov-Smirnov test or the Mann-Whitney U test can be used to compare whether the bias distributions in different ranges originate from the same population. Alternatively, analysis of variance (ANOVA) or t-tests can be used to assess differences in mean or variance. This significance level is usually expressed as a p-value. The fusion significance threshold is a pre-defined critical p-value used to determine whether the statistical distribution difference is statistically significant. Common statistical significance levels include 0.05 or 0.01, and their specific values ​​can be adjusted according to the tolerance for false positives and false negatives. The model fusion category indicates the type of cause of systematic prediction bias, primarily stemming from inappropriate mechanisms and data fusion strategies (such as fusion weight allocation).

[0148] The step of identifying all categories in the candidate cause set as the cause classification result involves summarizing all possible cause categories identified through the aforementioned threshold judgment. Specifically, as long as any judgment condition (such as the first correlation strength exceeding the first correlation threshold) is met, the corresponding cause category (such as the physical mechanism category) is added to the candidate cause set. Ultimately, all categories contained in the candidate cause set together constitute the cause classification result of this systematic prediction bias.

[0149] The above technical solutions enhance the robustness and adaptability of the intelligent analysis method for cement material proportioning, enabling the system to more accurately understand the sources of deviations and perform targeted optimizations. This ensures the effectiveness and reliability of the proportioning optimization scheme in the complex and ever-changing cement production environment.

[0150] This application further proposes, based on the causal classification result, to generate a first type of parameter for updating the evaluation weights in the dynamic quality risk assessment, a second type of parameter for correcting the thermodynamic simulation boundary conditions in the synergistic enhancement treatment, and a third type of parameter for fine-tuning the fusion weight allocation strategy in the hybrid predictive modeling. Specifically, if the causal classification result includes the data quality category, then based on the dynamic data quality risk value sequence corresponding to the systematic prediction deviation, the baseline value of the evaluation weight is updated through regression analysis to obtain the first type of parameter. If the causal classification result includes the physical mechanism category, then based on the systematic prediction deviation sequence and the corresponding material ratio parameter sequence, the correction amount of the thermodynamic simulation boundary conditions is fitted to obtain the second type of parameter. If the causal classification result includes the model fusion category, then based on the distribution differences of the systematic prediction deviation in different segments of the fusion weight historical record, the compensation coefficient of the fusion weight allocation strategy is determined to obtain the third type of parameter.

[0151] The causal classification result includes the data quality category, which refers to the determination, through analysis of the causes of systematic prediction deviations, that the deviations mainly originate from the quality problems of the input data, such as sensor failures, data transmission errors, or insufficient capture of fluctuations in the properties of raw materials. The dynamic data quality risk value sequence corresponding to this systematic prediction deviation is a set of dynamic data quality risk values ​​arranged chronologically according to each time point where the systematic prediction deviation occurs. This sequence reflects the real-time status of data quality when the deviation occurs. Regression analysis is a statistical method used to establish a mathematical relationship model between one or more independent and dependent variables. Here, it can be used to quantify the correlation between the dynamic data quality risk value sequence and the systematic prediction deviation, thereby guiding the adjustment of evaluation weights. For example, a linear regression model can be used to determine the linear relationship between risk values ​​and deviations, or a nonlinear regression model such as support vector regression (SVR) can be used to capture more complex nonlinear dependencies. The benchmark value of this evaluation weight is the initial or default parameter used in the dynamic quality risk assessment process to measure the relative importance of different data quality dimensions (such as the basic credibility of the data source, transient anomalies, and temporal logical consistency). Updating these benchmark values ​​means dynamically adjusting the weighting of each dimension in the risk assessment based on the actual impact of data quality issues on prediction bias, so that the risk assessment results more accurately reflect the true state of data quality. This first type of parameter is specifically designed to adjust the dynamic quality risk assessment process, and it directly affects how subsequent data quality risk values ​​are calculated.

[0152] The causal classification result includes this physical mechanism category, which refers to the determination, through analysis of the causes of systematic prediction deviations, that the deviations mainly stem from the inaccurate description of the actual production process by the physical mechanism model. Examples include changes in material properties, drift in production environment parameters, or the simplified physical model failing to fully capture complex reaction processes. This systematic prediction deviation sequence and its corresponding material proportioning parameter sequence are presented. The former is the systematic prediction deviation values ​​arranged in chronological order, and the latter is the sequence corresponding to the material proportioning parameters used when these deviations occurred. By analyzing these two sequences, the systematic error patterns exhibited by the physical mechanism model under specific proportioning conditions can be revealed. Fitting is a process of finding functional relationships between data points using mathematical methods. Here, it is used to establish a mapping relationship between systematic prediction deviations and material proportioning parameters, thereby inferring how the thermodynamic simulation boundary conditions should be adjusted. For example, polynomial fitting can be used to approximate the relationship between deviations and proportioning parameters, or nonlinear function fitting methods such as Gaussian process regression or neural networks can be used to capture more complex, nonlinear deviation-proportioning parameter relationships, thus obtaining more accurate boundary condition corrections. The correction amount for the thermodynamic simulation boundary conditions involves adjusting key input parameters (such as reaction temperature, pressure, and component activity coefficients) used in the co-enhanced processing. For example, if the model is found to systematically underestimate the intensity under high calcium-to-silicon ratio conditions, it may be necessary to correct the activation energy or equilibrium constant of a certain reaction in the simulation. This second type of parameter is specifically used to adjust the thermodynamic simulation process in the co-enhanced processing, and it directly affects the generation of virtual performance data and the construction of the physical information augmentation dataset.

[0153] The causal classification result includes the model fusion category, which means that by analyzing the causes of systematic prediction bias, it is determined that the bias mainly stems from problems in the fusion method between mechanistic prediction output and data-driven prediction output. For example, improper fusion weight allocation may lead to the model fusion failing to effectively combine the advantages of both in certain situations. The distribution difference of this systematic prediction bias across different historical fusion weight intervals refers to analyzing whether there are significant differences in the statistical characteristics (such as mean, variance, skewness, etc.) of the systematic prediction bias under different fusion weight settings (i.e., the respective contributions of the mechanistic model and the data-driven model). For example, if it is found that the bias is greater when the data-driven model weight is too high, it indicates that the fusion strategy may over-rely on the data-driven model. Determining the compensation coefficient for this fusion weight allocation strategy means introducing a correction factor to fine-tune the weight allocation based on the original fusion weight allocation strategy. For example, if a large bias is found within a certain weight interval, a compensation coefficient can be introduced to slightly increase the weight of the mechanistic model or slightly decrease the weight of the data-driven model within that interval to correct the bias caused by improper fusion. The third type of parameter is specifically used to adjust the fusion weight allocation strategy in hybrid predictive modeling, and it directly affects the weighted fusion process of mechanistic prediction output and data-driven prediction output.

[0154] Through the above technical solutions and the refined parameter generation mechanism, the collaborative adjustment process can more accurately locate and solve the root cause of systematic prediction deviations, thereby improving the prediction accuracy of the intelligent analysis method for cement material proportions and the anti-interference ability of the optimization scheme, and ensuring the robustness of proportion optimization under complex working conditions.

[0155] This application further proposes an iterative parameter approach, applying the first, second, and third types of parameters to their respective processes to achieve coordinated adjustment. This coordinated adjustment includes: sequentially updating the first, second, and third types of parameters as values ​​according to a preset update order, and associating them with the corresponding dynamic quality risk assessment, coordinated enhancement processing, and hybrid predictive modeling processes. For each type of parameter, in the next execution of the process associated with that parameter, a progressive update is performed using linear interpolation of the old and new parameters, where the update step size of this linear interpolation is adaptively adjusted based on the strength of the systemic prediction bias. After all three types of parameters have completed this progressive update, a complete proportion performance prediction and optimization process is triggered to obtain new proportion performance prediction results and the corresponding new uncertainty interval. The new deviation between the new actual production feedback and the new proportion performance prediction results is calculated, and it is verified whether the new deviation falls within the new uncertainty interval. If it does, the coordinated adjustment is confirmed to be complete.

[0156] This approach employs parameter iteration, applying first, second, and third types of parameters to their respective processes to achieve coordinated adjustments. This strategy, using small, rapid steps to gradually approach the target state, aims to avoid instability caused by large, one-time parameter changes. Its purpose is to ensure stable system operation during adjustments and to allow for continuous monitoring and feedback of the adjustments' effects. Specifically, it can be implemented in two ways: sequential iteration, where different types of parameters are updated in a pre-defined logical order, with the impact of each update evaluated before moving on to the next; and parallel iteration and synchronization, where parameter update calculations are performed simultaneously in different processing modules, but data synchronization and consistency checks are conducted at key nodes to ensure overall coordination.

[0157] Following a preset update order, the first type of parameter, the second type of parameter, and the third type of parameter are sequentially updated as values, and associated with the corresponding dynamic quality risk assessment, collaborative enhancement processing, and hybrid predictive modeling process. This preset update order aims to provide a structured framework for multi-parameter collaborative adjustment, ensuring the logic and controllability of the adjustment process. By clearly defining the order of parameter updates, potential conflicts or mutual interference between different parameter adjustments can be effectively avoided, thereby improving adjustment efficiency and stability. Associating the updated values ​​with the corresponding processes ensures the targeted and effective nature of parameter adjustments. For example, a fixed update order can be set, such as first adjusting parameters related to data quality, then adjusting parameters related to physical mechanisms, and finally adjusting parameters related to model fusion. Alternatively, the update order can be dynamically determined based on the classification of the causes of systematic prediction bias, prioritizing the adjustment of parameters related to the main causes to accelerate problem resolution.

[0158] For each type of parameter, in the next execution of the process associated with that parameter, a gradual update is performed using linear interpolation of the old and new parameters. The update step size of this linear interpolation is adaptively adjusted based on the strength of the systematic prediction bias. Gradual updates, through linear interpolation, allow parameters to smoothly transition from old to new values, rather than abruptly, reducing the risk of system oscillations or instability. The adaptively adjusted update step size adds intelligence to the adjustment process, enabling it to flexibly adjust the update speed according to the severity of the systematic prediction bias. Specifically, an initial interpolation ratio can be set, for example, 10% for new parameters and 90% for old parameters. Then, this ratio is dynamically adjusted through a modulation function based on the strength of the systematic prediction bias; the greater the bias, the higher the weight of the new parameter. Alternatively, the strength of the systematic prediction bias can be directly mapped to a learning rate. This learning rate, multiplied by the difference between the old and new parameters, serves as the increment for this update, thus achieving adaptive adjustment of the step size.

[0159] After all three types of parameters have completed the incremental update, a complete proportioning performance prediction and optimization process is triggered to obtain new proportioning performance prediction results and corresponding new uncertainty intervals. This step aims to perform a holistic evaluation of the system after a round of collaborative adjustments. By triggering a complete prediction and optimization process, the overall effect of all parameter adjustments can be comprehensively verified, ensuring that the system can operate stably and accurately under the new parameter configuration. Obtaining new prediction results and uncertainty intervals provides a benchmark for subsequent verification. Implementation methods can include: automated process triggering, where the system automatically starts a complete proportioning performance prediction and optimization process when the incremental update status flags of all parameters are set to "complete". Alternatively, periodic batch processing can be used, where parameter updates and overall process triggering are performed at predetermined time intervals (e.g., at the end of each shift or day) to adapt to the production rhythm.

[0160] The new deviation between the new actual production feedback and the new performance prediction result is calculated, and it is verified whether the new deviation falls within the new uncertainty interval. If it does, the coordinated adjustment is confirmed to be complete. This is the final verification step in the coordinated adjustment process, and its core function is to quantitatively evaluate the effectiveness of the adjustment and determine whether the system has recovered to an acceptable performance level. By comparing the new deviation with the new uncertainty interval, it is possible to objectively determine whether the reliability of the prediction model has been improved. Specifically, the absolute value comparison method can be used, that is, calculating the absolute value of the new deviation and determining whether it is less than or equal to the half-width of the new uncertainty interval. Alternatively, a statistical significance test can be used. For example, if the new uncertainty interval represents a confidence interval, the new deviation is tested to see if it falls within that confidence interval. If it does, the adjustment is considered effective, and the coordinated adjustment is confirmed to be complete.

[0161] Through the above technical solutions and mechanisms, this application can continuously optimize the parameters of core links such as dynamic quality risk assessment, collaborative enhancement processing, and hybrid prediction modeling, thereby improving the robustness and reliability of the entire intelligent analysis method for cement material proportioning under complex working conditions. This makes the recommended proportioning scheme more resistant to interference in actual production and reduces the deviation between performance and cost targets.

[0162] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0163] This application also provides a cement material proportioning intelligent analysis system, the system comprising: The processing module is used to perform dynamic quality risk assessment and collaborative enhancement processing on multi-source heterogeneous data of cement production, generate physical information enhancement dataset and dynamic data quality risk value. The dynamic data quality risk value is associated with a predefined risk level. The multi-source heterogeneous data includes material chemical composition, proportioning settings, process parameters and microstructure image data.

[0164] The strategy selection module is used to select a target modeling strategy from multiple preset hybrid prediction modeling strategies based on the risk level associated with the dynamic data quality risk value. The target modeling strategy is then used to perform hybrid prediction modeling of the physical information augmentation dataset by combining mechanism and data fusion, so as to obtain the matching performance prediction results and the corresponding uncertainty interval.

[0165] The generation module is used to generate a recommended ratio scheme by performing multi-objective optimization based on the performance prediction results of the ratio and the uncertainty interval, and according to the optimization objective set corresponding to the risk level.

[0166] The adjustment module is used to coordinately adjust the parameters of the dynamic quality risk assessment, the thermodynamic simulation boundary conditions of the synergistic enhancement treatment, the fusion weights of the hybrid prediction model, and the criteria for the risk level based on the deviation between the actual production feedback and the performance prediction results of the ratio. The actual production feedback is the actual performance result obtained by producing using the recommended ratio scheme.

[0167] It should be noted that the intelligent cement material proportioning analysis system provided in the above embodiments is only illustrated by the division of the above functional modules when analyzing the proportions. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer equipment can be divided into different functional modules to complete all or part of the functions described above. In addition, the intelligent cement material proportioning analysis system and the intelligent cement material proportioning analysis method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be repeated here.

[0168] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for intelligent analysis of cement material proportioning, characterized in that, The method includes: Multi-dimensional reliability quantification is performed on multi-source heterogeneous data of cement production to obtain dynamic data quality risk values ​​and associated risk levels. This multi-dimensional reliability quantification includes coupled calculations of the basic reliability of the data source, instantaneous anomalies, and time-series logical consistency. The multi-source heterogeneous data includes material chemical composition, proportion settings, process parameters, and microstructure image data. Based on the risk levels, a target operation mode is determined from multiple predefined data augmentation operation modes, each associated with a different risk level. Data augmentation operations corresponding to the target operation mode are performed on the multi-source heterogeneous data. These operations include generating virtual performance data associated with the material proportions in the multi-source heterogeneous data and extracting quantitative features from the microstructure images in the multi-source heterogeneous data. The virtual performance data, the quantitative features, and the quality-compensated multi-source heterogeneous data are then fused to generate a physical information augmented dataset. Based on the risk level associated with the dynamic data quality risk value, a target modeling strategy is selected from multiple preset hybrid prediction modeling strategies. The target modeling strategy is then used to perform hybrid prediction modeling of the physical information augmentation dataset, which combines mechanism and data fusion, to obtain the matching performance prediction results and the corresponding uncertainty interval. Based on the predicted performance of the proportions and the uncertainty interval, and according to the optimization objective set corresponding to the risk level, multi-objective optimization is performed to generate a recommended proportion scheme; Based on the deviation between actual production feedback and the predicted performance of the formula ratio, the parameters of dynamic quality risk assessment, the thermodynamic simulation boundary conditions of the synergistic enhancement treatment, the fusion weights of the hybrid prediction modeling, and the criteria for the risk level are synergistically adjusted. The actual production feedback is the actual performance result obtained by producing using the recommended formula ratio scheme.

2. The method according to claim 1, characterized in that, The step of performing the data augmentation operation corresponding to the target operation mode on the multi-source heterogeneous data includes: The definitions of multiple predefined data augmentation operation modes associated with different risk levels are obtained. The multiple predefined data augmentation operation modes include at least a first operation mode associated with low risk level and a second operation mode associated with high risk level. The first operation mode defines a first set of execution parameters and a first data range. The second operation mode defines a second set of execution parameters and a second data range. The computational complexity of the second set of execution parameters is higher than that of the first set of execution parameters. The second data range is wider than that of the first data range. Based on whether the target operation mode is the first operation mode or the second operation mode, select the corresponding first set of execution parameters and the first data range or the second set of execution parameters and the second data range; Using the selected execution parameters and data range, the data augmentation operation is performed on the multi-source heterogeneous data to generate the virtual performance data and extract the quantitative features.

3. The method according to claim 1, characterized in that, The method employs the target modeling strategy to perform hybrid predictive modeling of the physical information augmentation dataset, combining mechanism and data fusion, to obtain the performance prediction results and corresponding uncertainty intervals, including: Based on the target modeling strategy, rules for feature separation of the physical information augmentation dataset, rules for assigning fusion weights to the mechanism and data prediction output, and rules for calculating the uncertainty interval are determined. Based on the feature separation rules, the physical information augmentation dataset is separated into a mechanism-constrained feature subset and a data-driven feature subset. The mechanism-constrained feature subset and the data-driven feature subset are then input into their respective processing channels for processing to obtain mechanism prediction output and data-driven prediction output. According to the fusion weight allocation rule, the fusion weights are dynamically allocated between the mechanism prediction output and the data-driven prediction output based on the dynamic data quality risk value; The mechanism prediction output and the data-driven prediction output are weighted and fused using the assigned fusion weights to obtain the fused prediction distribution; Based on the uncertainty interval calculation rules, and based on the fused prediction distribution and the dynamic data quality risk value, the uncertainty interval corresponding to the matching performance prediction result is determined.

4. The method according to claim 3, characterized in that, According to the feature separation rule, the physical information augmentation dataset is separated into a mechanism-constrained feature subset and a data-driven feature subset. The mechanism-constrained feature subset and the data-driven feature subset are then input into corresponding processing channels for processing to obtain mechanism prediction output and data-driven prediction output, including: Based on the feature separation rules, the mechanism-constrained feature subset consisting of deterministic physical quantities based on cement chemistry theory and the data-driven feature subset containing quantitative feature parameters extracted from the microstructure image are identified and extracted from the physical information-enhanced dataset. Based on the target modeling strategy, a corresponding theoretical calculation model or simplified physical model is configured for the mechanism constraint channel, and the mechanism constraint feature subset is input into the mechanism constraint channel. The model is then used to process the data to obtain the initial mechanism prediction value. Based on the calibration rules defined by the target modeling strategy, the initial mechanism prediction value is compared with the historical calibration data obtained from the physical information augmentation dataset, and the initial mechanism prediction value is corrected according to the comparison results to obtain the mechanism prediction output. The data-driven feature subset is input into the data-driven channel, and nonlinear feature transformation and mapping are performed through the machine learning model specified by the target modeling strategy to obtain the data-driven prediction output.

5. The method according to claim 3, characterized in that, The determination of the uncertainty interval corresponding to the matching performance prediction result based on the uncertainty interval calculation rules, the fused prediction distribution, and the dynamic data quality risk value includes: Based on the aforementioned uncertainty interval calculation rules, the calculation method for the basic confidence interval and the modulation function for the interval adjustment coefficient are determined. Based on the degree of dispersion of the fused predicted distribution, the basic confidence interval is determined according to the selected calculation method; The dynamic data quality risk value is input into the modulation function to convert the dynamic data quality risk value into an interval adjustment coefficient; The uncertainty interval is obtained by modulating the width of the basic confidence interval using the interval adjustment coefficient.

6. The method according to claim 1, characterized in that, The step of generating a recommended proportioning scheme by performing multi-objective optimization based on the performance prediction results and the uncertainty interval, and according to the optimization objective set corresponding to the risk level, includes: Based on the risk level and the corresponding set of optimization objectives, a target optimization strategy is selected from multiple preset optimization strategies. The target optimization strategy defines the constraints and search space rules that match the risk level and the set of optimization objectives. Based on the search space rules defined by the target optimization strategy, determine the feasible parameter space for this optimization; Within the feasible parameter space, based on the constraints defined by the objective optimization strategy, and integrating at least one performance objective and at least one cost objective defined by the optimization objective set, a multi-objective optimization model is constructed, wherein the constraints include the utilization of the uncertainty interval; Solving the multi-objective optimization model yields the Pareto non-dominated solution set; Based on preset decision rules, the recommended ratio scheme is determined from the Pareto non-dominated solution set.

7. The method according to claim 1, characterized in that, The method of coordinating adjustments to the parameters of dynamic quality risk assessment, the thermodynamic simulation boundary conditions of the synergistic enhancement processing, the fusion weights of the hybrid prediction modeling, and the criteria for the risk level based on the deviation between actual production feedback and the predicted performance of the proportions includes: By comparing and analyzing the sequence of the deviation with the sequence of the uncertainty interval, systematic prediction deviations that exceed the range of the uncertainty interval are identified. The causes of the systematic prediction bias are classified into one or more of three categories: data quality, physical mechanism, or model fusion, to obtain the cause classification results; Based on the cause classification results, a first type of parameter is generated for updating the evaluation weight in the dynamic quality risk assessment, a second type of parameter is generated for correcting the thermodynamic simulation boundary conditions in the collaborative enhancement processing, and a third type of parameter is generated for fine-tuning the fusion weight allocation strategy in the hybrid prediction modeling. By using parameter iteration, the first type of parameters, the second type of parameters, and the third type of parameters are applied to the corresponding processes to complete the coordinated adjustment.

8. The method according to claim 7, characterized in that, The generation of a first type of parameter based on the causal classification result, used to update the evaluation weights in the dynamic quality risk assessment; a second type of parameter used to correct the thermodynamic simulation boundary conditions in the synergistic enhancement processing; and a third type of parameter used to fine-tune the fusion weight allocation strategy in the hybrid prediction modeling, includes: If the cause classification result includes a data quality category, then based on the dynamic data quality risk value sequence corresponding to the systematic prediction deviation, the benchmark value of the evaluation weight is updated through regression analysis to obtain the first type of parameter; If the causal classification result includes a physical mechanism category, then based on the systematic prediction deviation sequence and the corresponding material ratio parameter sequence, the correction amount of the thermodynamic simulation boundary condition is fitted to obtain the second type of parameter. If the cause classification result includes a model fusion category, then based on the distribution difference of the systematic prediction bias in different segments of the fusion weight historical record, the compensation coefficient of the fusion weight allocation strategy is determined, and the third type of parameter is obtained.

9. A cement material proportioning intelligent analysis system, characterized in that, The system includes: The processing module performs multi-dimensional reliability quantification on multi-source heterogeneous data from cement production to obtain dynamic data quality risk values ​​and associated risk levels. The multi-dimensional reliability quantification includes coupled calculations of the basic reliability of the data source, instantaneous anomalies, and time-series logical consistency. The multi-source heterogeneous data includes material chemical composition, proportion settings, process parameters, and microstructure image data. Based on the risk levels, a target operation mode is determined from multiple predefined data augmentation operation modes, each associated with a different risk level. Data augmentation operations corresponding to the target operation mode are executed on the multi-source heterogeneous data. These operations include generating virtual performance data associated with the material proportions in the multi-source heterogeneous data and extracting quantitative features from the microstructure images in the multi-source heterogeneous data. The virtual performance data, the quantitative features, and the quality-compensated multi-source heterogeneous data are then fused to generate a physical information augmented dataset. The strategy selection module is used to select a target modeling strategy from multiple preset hybrid prediction modeling strategies based on the risk level associated with the dynamic data quality risk value, and to use the target modeling strategy to perform hybrid prediction modeling of mechanism and data fusion on the physical information augmentation dataset to obtain the matching performance prediction result and the corresponding uncertainty interval. The generation module is used to generate a recommended ratio scheme by performing multi-objective optimization based on the ratio performance prediction results and the uncertainty interval, and according to the optimization objective set corresponding to the risk level. The adjustment module is used to coordinately adjust the parameters of dynamic quality risk assessment, the thermodynamic simulation boundary conditions of the synergistic enhancement processing, the fusion weight of the hybrid prediction model, and the criteria for the risk level based on the deviation between the actual production feedback and the predicted performance of the formula. The actual production feedback is the actual performance result obtained by producing using the recommended formula.

Images