Sand performance data management and addition amount prediction method and system

Abstract

This invention relates to the field of foundry sand processing technology, and discloses a method and system for managing molding sand performance data and predicting addition amounts. Addressing the technical problems of excessive reliance on manual experience in molding sand batching and the difficulty in accurately modeling nonlinear systems, this invention receives and integrates molding sand performance parameters and material addition records through a data access layer; performs data cleaning and feature construction through a business logic and calculation layer to generate derived features such as invalid clay; trains a prediction model using a machine learning regression algorithm combined with the Huber loss function and process constraints; and finally receives the current parameters and calls the model to output recommended addition amounts of dust, clay, and mixed soil in real time. This invention achieves data-driven standardized batching operations, significantly improving prediction accuracy and molding sand quality stability, and is suitable for efficient and intelligent batching of molding sand in foundry workshops.

Description

Technical Field

[0001] This invention relates to the field of foundry sand treatment technology, and in particular to a method and system for managing molding sand performance data and predicting the amount of sand to be added. Background Technology

[0002] Sand treatment is one of the core processes in the entire molding sand casting process. To ensure casting quality and significantly reduce casting defects such as sand holes, porosity, sand adhesion, sintering, and deformation, the molding sand must be kept in a highly stable state, possessing suitable strength, permeability, plasticity, and refractoriness. Sand treatment in modern high-quality foundries is no longer a simple physical mixing process, but a complex systems engineering project deeply intertwined with materials science, mechanical engineering, and environmental science. In the molding sand preparation process, controlling the addition of materials such as dust, clay, and other additives is a crucial step.

[0003] Currently, in this technological field, determining the amount of materials added relies heavily on the subjective experience of workshop workers, who primarily judge the state of the molding sand by hand, visual inspection, and pressing. This non-standardized operating mode has extremely poor stability and low repeatability. With the loss of experienced technicians, companies face a serious risk of technological gaps. Furthermore, historical molding sand performance testing data and raw material addition records are mostly in paper or scattered spreadsheet formats, completely lacking unified and structured database management. This data silo phenomenon makes historical data retrieval extremely difficult, and in-depth correlation analysis between data is almost impossible, making it difficult to transform valuable production experience into reusable digital assets. Although some companies have introduced molding sand performance testing equipment to obtain production data such as permeability and strength, the final decision on the addition amount still requires process engineers to consult manuals and combine them with personal experience for estimation. This testing plus manual decision-making model is not only inefficient and severely delayed in response, but also highly susceptible to fluctuations in the batching results due to differences in the judgment standards of different engineers.

[0004] In summary, the core technical problems that urgently need to be solved in this field are: how to overcome the defects of excessive reliance on manual experience and extremely fragmented data management in the molding sand preparation process, break down data silos to achieve standardization and real-time closed-loop control of material batching decisions; and how to establish an accurate quantitative prediction model for the molding sand system, which is characterized by numerous influencing factors and highly nonlinear relationships, so as to achieve high-precision prediction and forward-looking dynamic adjustment of material addition. Summary of the Invention

[0005] To address the technical problems of excessive reliance on manual experience in molding sand batching leading to decision-making delays, and the difficulty in establishing accurate quantitative prediction models due to the strong nonlinearity of molding sand systems, the present invention aims to provide a method and system for managing molding sand performance data and predicting the amount of molding sand added.

[0006] This invention provides a method for managing molding sand performance data and predicting its addition amount, comprising the following steps: Receive and integrate molding sand production-related data from different sources to construct a structured dataset. The acquired data includes molding sand performance parameters, historical material addition data corresponding to the molding sand performance parameters, and related process and environmental parameters. The acquired raw data is cleaned, standardized, and features are constructed to generate a high-quality feature dataset for use by machine learning models. A machine learning regression algorithm is used to train and cross-validate the model using the feature dataset. The Huber loss function is introduced into the objective function of the model and combined with the specific process constraints of the production line to generate the best prediction model after training. Receive the performance parameters and process objectives of the molding sand for the current batch, call the best prediction model to perform forward inference calculations, and output the recommended addition amount prediction values ​​of dust, clay and mixed materials in real time.

[0007] As a preferred technical solution of the present invention, when constructing the features, derivative features are created based on the acquired original data to enhance the model's understanding of the process logic. The derivative features include the amount of ineffective clay added and the proportion of core entry. The amount of ineffective clay added is configured as the mud content value minus the blue absorption value. The proportion of core entry is configured as the amount of sand core entering today divided by the amount of mixed sand. When calculating the amount of sand core entering today, the amount of cold core entering and the amount of hot core entering are counted independently.

[0008] As a preferred technical solution of the present invention, during the data cleaning process, outliers exceeding physical limits are automatically identified and processed, and missing values ​​in the data records are repaired using time-series-based linear interpolation, pre- and post-value filling, or statistical characteristic-based filling strategies; the data standardization includes normalizing parameters of different dimensions and ranges to a unified interval through mathematical transformation to eliminate the influence of dimensions.

[0009] As a preferred technical solution of the present invention, the machine learning regression algorithm adopts the gradient boosting tree algorithm or the random forest algorithm. During the model training stage, the feature dataset is divided into a training set and a validation set. In the process of adjusting the model hyperparameters through the validation set, the depth and learning rate of the decision tree are independently optimized for different production line types and different target prediction variables to capture the complex nonlinear interaction features between molding sand parameters.

[0010] As a preferred technical solution of the present invention, after outputting the predicted value of the recommended addition amount, the model prediction results and historical related data are presented in a graphical manner. The visualization presentation includes generating a comparison chart of the predicted value and the historical actual value to show the model prediction fidelity, a time series chart of key performance indicators to show the fluctuation trend of core indicators, a box-and-whisker plot to analyze the stability of indicator distribution, and a correlation heatmap to show the degree of statistical correlation between variables. The prediction results and analysis charts are then integrated into a structured report file and exported according to instructions.

[0011] The present invention also provides a molding sand performance data management and addition amount prediction system, the logical architecture of which includes a data access layer, a data storage layer, a business logic and calculation layer, and a user interaction layer; The data access layer is equipped with a data entry and integration module, which provides a human-computer interaction interface that supports manual structured entry and batch file import. It is used to receive diverse molding sand production-related data and perform data format verification and basic deduplication. The data storage layer is equipped with a central database for persistently storing the input raw data, the generated preprocessed data, the trained model parameters, and the generated historical prediction records. The business logic and computing layer is equipped with a data processing and feature engineering module and an intelligent prediction core module. The data processing and feature engineering module is configured to perform data cleaning, standardization, and feature construction operations to transform the raw data into a high-quality feature dataset. The intelligent prediction core module includes a model training unit and a model inference unit. The model training unit is configured to call a machine learning algorithm library and use the feature dataset to perform model training and optimization processes to generate a trained prediction model. The model inference unit integrates the prediction model to calculate and return the predicted value of material addition in real time based on new input parameters. The user interaction layer is equipped with a visualization and human-computer interaction module and a data output and report generation module. The visualization and human-computer interaction module is configured to generate and render multi-dimensional data analysis charts to display in-depth data trends, model performance and prediction results. The data output and report generation module is configured to encapsulate the prediction results and analysis charts into a report file of a specified format and provide an export function.

[0012] As a preferred technical solution of the present invention, the intelligent prediction core module configures the Huber loss function in the objective function to reduce overfitting during model training, and performs cross-validation by combining the specific process range of each production line extracted from the central database as a constraint condition. The model inference unit resides in memory and is instantly called through the application programming interface to achieve a second-level prediction response.

[0013] As a preferred embodiment of the present invention, when performing the feature construction operation, the data processing and feature engineering module calculates the amount of ineffective clay added and the proportion of core-injected material as derived features for use by the machine learning model.

[0014] As a preferred technical solution of the present invention, the data entry and integration module receives molding sand performance parameters including moisture content, air permeability, wet compressive strength, sand temperature, mud content, loss on ignition, blue absorption, and particle size distribution. The received historical material addition data includes the addition mass of dust, bentonite clay, and mixed soil.

[0015] As a preferred technical solution of the present invention, the multidimensional data analysis charts generated by the visualization and human-computer interaction module include a comparison chart of predicted values ​​and historical actual values ​​in the form of a double vertical axis line chart, a time series chart of key performance indicators in the form of a continuous curve, and a box-and-whisker plot based on quartiles.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention fundamentally realizes data-driven scientific decision-making and standardized operation of molding sand batching process by constructing an intelligent prediction model based on historical data. This solution successfully transforms the highly experience-dependent molding sand batching process into a strictly quantifiable system operation, completely eliminating the time lag caused by cumbersome manual data lookup and paper-and-pen calculations in traditional production models. This automated real-time decision-making system not only closely matches the fast-paced production demands of the foundry site but also forms an intelligent closed-loop control system integrating data acquisition, analysis and modeling, decision support, and result feedback. It effectively accumulates the tacit knowledge of experts and mitigates the risk of technological gaps in enterprises.

[0017] This invention addresses the core challenges of highly coupled massive parameters and extremely nonlinear relationships among influencing factors in molding sand systems. It creatively applies and optimizes machine learning algorithms such as gradient boosting trees, introduces the Huber loss function to effectively combat data noise, and incorporates process constraints from industrial settings for model training. This mechanism enables the system to automatically extract microscopic feature mappings from chaotic historical data, achieving a prediction accuracy and specificity far exceeding traditional thermodynamic empirical formulas for dust, clay, and mixed material additions. This significantly suppresses molding sand quality fluctuations caused by inconsistent human judgment standards, improving the quality stability and batch consistency of final casting products from the source, and establishing a high-value data asset system with self-iterative and evolving capabilities for enterprises. Attached Figure Description

[0018] Figure 1 This is a logical architecture diagram of the molding sand data management and prediction system provided in an embodiment of the present invention; Figure 2This is a schematic diagram of a data entry interface provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the training and application of the prediction model provided in an embodiment of the present invention.

[0019] Attached reference numerals: 201, Data access layer; 202, Data storage layer; 203, Business logic and computing layer; 204, User interaction layer. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the casting production process, sand treatment is one of the core steps in the entire molding sand casting process. It ensures the stability of the molding sand's properties, giving it suitable strength, permeability, plasticity, and refractoriness, thereby ensuring the quality of the castings and reducing casting defects such as sand holes, porosity, sand adhesion, sintering, and deformation. Traditional sand treatment often relies on simple material mixing and manual experience judgment. However, modern high-quality foundries require a stable, efficient, and intelligent sand treatment system, which is a complex systems engineering project involving materials science, mechanical engineering, automatic control, and environmental science. The preparation of molding sand, especially the amount of dust, clay, and other mixed materials added, is a crucial step. In the existing technology, most small foundries mainly rely on experienced craftsmen to subjectively judge the state of the molding sand by hand, visual inspection, and pressing, and determine the amount to add based on experience. This method, which relies entirely on manual experience, has poor stability and low repeatability, and faces the risk of technological discontinuity. Meanwhile, historical molding sand performance testing data and raw material addition records are mostly stored in paper or scattered spreadsheet formats, lacking unified structured database management, making historical data retrieval difficult and correlation analysis virtually impossible. Even existing office software with basic data entry and storage functions is merely equivalent to electronic notepads, lacking in-depth data mining and analysis. Furthermore, although some companies use molding sand performance testing equipment to obtain production data, decisions regarding the amount of materials added still require engineers to consult manuals and combine calculations with experience. This testing-plus-manual-decision-making model is inefficient and slow to respond. To address the shortcomings of existing technologies, such as over-reliance on personal experience leading to non-standardized and delayed decision-making, fragmented data management resulting in insufficient data value mining, and the difficulty in establishing accurate quantitative models due to the numerous influencing factors in the molding sand system, this invention provides a method and system for predicting molding sand quality and performance based on historical data.

[0021] This embodiment provides a molding sand performance data management and addition prediction system. Through manual entry or batch import, it collects historical molding sand performance data and process parameters from laboratory test records, production work orders, and process cards to build a unified data warehouse. Based on this data warehouse, the system uses data mining and machine learning techniques to establish a molding sand performance prediction model, thereby assisting process engineers in analyzing quality fluctuations, predicting future batch molding sand performance trends, and optimizing production. (See also...) Figure 1 The diagram shown illustrates the logical architecture of the molding sand data management and prediction system. It clearly demonstrates the layered architecture, illustrating the data and control flow relationships between layers, and showcasing the system's integration and completeness. Specifically, the architecture includes a data access layer 201, a data storage layer 202, a business logic and computation layer 203, and a user interaction layer 204. At the physical device implementation level, the system can be deployed on servers or industrial computer clusters with high computing power. The servers are equipped with multi-core CPUs to support multi-threaded machine learning model training and large-capacity high-speed random access memory to meet the memory loading and feature matrix operation requirements of massive historical molding sand data. The system's central database is persistently deployed on a high-performance solid-state drive array to ensure millisecond-level data read and storage response within the production cycle. On-site operation terminals communicate with the server in real time via the factory LAN or industrial Ethernet, thus establishing a seamless connection between cloud or edge computing and on-site operations.

[0022] Based on the aforementioned system architecture, the system comprises multiple interconnected modules to implement the method of this invention, specifically including a data entry and integration module, a central database, a data processing and feature engineering module, an intelligent prediction core module, a visualization and human-computer interaction module, and a data output and report generation module. The data entry and integration module provides a human-computer interaction interface, supports manual structured data entry and batch file import, and is used to receive diverse data. This module also includes data format verification and basic deduplication functions. The central database is used to persistently store the entered raw data, the generated preprocessed data, the trained model parameters, and the generated historical prediction records, thereby providing a unified, unique, and authoritative data source for the entire intelligent system. The data processing and feature engineering module is configured to perform low-level computational operations such as data cleaning, standardization, and feature construction, transforming raw data containing noise and missing values ​​into a high-quality feature dataset that can be directly used by machine learning models. The intelligent prediction core module, as a key device of this system, is divided into a model training unit and a model inference unit in its software architecture. The model training unit is configured to call the machine learning algorithm library, utilize historical data from the central database to perform model training and optimization processes, and generate a trained prediction model. The model inference unit resides permanently in memory and integrates the trained prediction model. When a new query request is received from the client, this unit instantly invokes the model in memory via the application programming interface (API), performs forward propagation and real-time calculations based on the latest input parameters, and immediately returns the predicted material addition amount. The visualization and human-computer interaction module is configured to generate and render various multi-dimensional data analysis charts, providing a graphical interface to display in-depth data trends, model performance evaluation indicators, and the final prediction results, forming the main window for user interaction with the system. The data output and report generation module is configured to package the prediction results and analysis charts into report files of a specified format according to a preset business template and provide users with flexible export functions.

[0023] The molding sand performance data management and addition prediction method provided by this invention mainly includes a series of sequential steps: multivariate data acquisition and structured input, data preprocessing and feature engineering, prediction model construction, training and deployment, visualization analysis and presentation of prediction results, and data output and decision support. In the multivariate data acquisition and structured input step, the system comprehensively receives and seamlessly integrates molding sand production-related data from different sources, constructing a highly structured relational or time-series dataset. (See also...) Figure 2The data entry interface shown is a schematic diagram. This interface uses a flat and ergonomic design language, intuitively demonstrating a graphical example of manual data entry and fully showcasing the system's ease of use in complex industrial environments. The system acquires data specifically covering molding sand performance parameters. Operators can manually enter single data points for detailed input, or efficiently import data from different batches of laboratory or online testing data using batch import methods such as spreadsheets or comma-separated files. These performance parameters cover core indicators characterizing the physical and chemical properties of molding sand, such as moisture content, permeability, wet compressive strength, sand temperature, clay content, loss on ignition, bluing absorption, and particle size distribution. Moisture content directly determines the clay's bonding ability; permeability affects the smoothness of gas expulsion from the mold cavity during casting; wet compressive strength ensures the sand mold does not collapse under handling and molten metal impact; and sand temperature directly affects the moisture evaporation rate, thus requiring specific water replenishment strategies. Along with the input of molding sand performance parameters, the system simultaneously acquires historical material addition data corresponding to these parameters. This primarily includes the specific mass or proportion of materials such as dust, clay, and mixed soil. Furthermore, to enable the machine learning model to learn patterns under different production environments, the system comprehensively acquires relevant process and environmental parameters. This includes production line identifiers to distinguish equipment differences between different production lines, production dates and times to extract time-series features and periodic patterns, and contextual information such as ambient temperature and humidity that may have a microscopic impact on molding sand moisture evaporation. Simultaneously with data reception, the data input and integration module's verification engine performs real-time basic verification during the input phase. For example, it uses primary key constraints to avoid duplicate data entry contamination and employs a numerical range screening mechanism based on industrial common sense to intercept input data that clearly violates physical laws.

[0024] After completing the collection and structuring of multi-source data, the system initiates data preprocessing and feature engineering steps. This involves deep cleaning, multi-dimensional transformation, and complex feature construction of the acquired raw data to prepare a high-quality dataset for subsequent model training. Data collected from industrial sites often includes significant noise and anomalies caused by occasional malfunctions, making data cleaning crucial. The system's built-in anomaly detection algorithm automatically identifies and handles outliers. For example, values ​​exceeding physical limits are removed or smoothed. For missing values ​​caused by equipment malfunctions or human error, the system employs time-series-based linear interpolation, before-and-after value imputation strategies, or mean-median imputation strategies based on statistical distribution characteristics for proper repair. Because different molding sand parameters have vastly different dimensions—for example, permeability values ​​may reach hundreds while moisture content is only in the single digits—directly inputting them into neural networks or distance metric models can lead to slow gradient descent or unbalanced weight distribution. Therefore, the system performs data standardization or normalization, normalizing parameters of different dimensions and ranges to a unified interval, such as zero to one, through mathematical transformations. This eliminates the influence of dimensions and significantly accelerates the model's convergence process. In the method of this invention, feature construction is an optional but extremely important step, translating engineers' domain knowledge into machine-understandable mathematical language. Based on casting process knowledge, the system creates derived features that are more predictive and explanatory for the target variable through mathematical operations between parameters. For example, the system constructs two key derived features: the amount of ineffective clay added and the core-injection ratio. The ineffective clay added is defined by the formula: mud content minus bluing absorption. Physically, mud content reflects the content of all fine powders in the sand, while bluing absorption specifically refers to the content of active bentonite. The difference between the two represents the content of dead clay that has lost its binding capacity but still absorbs a large amount of water. This derived feature is crucial for the model to accurately predict the amount of water and new clay replenishment. Furthermore, the core-injection ratio is calculated by dividing the amount of sand cores added today by the amount of mixed sand. During the casting process, the disintegration of sand cores mixed into the old sand system will significantly change the composition of the molding sand. Therefore, when calculating, the system not only needs to consider the total amount of sand entering, but also needs to strictly distinguish between the amount of cold core entering and the amount of hot core entering. This is because the resin binder used for cold cores and hot cores is different, and their impact mechanism on the loss of molding sand and strength after disintegration is also completely different. This fine-grained feature construction greatly enhances the machine learning model's deep understanding of complex process logic.

[0025] See Figure 3The diagram illustrating the training and application of the prediction model details the core process from historical data preparation, model training and selection, model deployment, to new data input and prediction output, highlighting the crucial role of the machine learning model in the system of this invention. In the prediction model construction, training, and deployment steps, the system establishes a nonlinear mapping model from multi-dimensional input variables to key output variables based on a rigorously preprocessed and feature-engineered historical dataset. The input vector includes, but is not limited to, parameters such as moisture content, sand temperature, permeability, strength, old sand moisture content, and environmental stability, while the output variable is set as the optimal daily addition amount of dust, clay, and blended materials. Due to the high coupling and nonlinear relationships among the various parameters in the molding sand system, traditional multiple linear regression cannot meet the accuracy requirements. Therefore, the prediction model is constructed using at least one advanced machine learning regression algorithm, such as gradient boosting trees, random forests, decision trees, or their multi-model ensemble architectures. These tree-based algorithms do not require strict assumptions about the data distribution, can naturally handle multi-feature inputs of different dimensions, and effectively capture the complex nonlinear relationships between parameters through feature splitting. During the model training phase, the system scientifically divides the historical dataset into training and validation sets according to time series or random sampling strategies. The model training unit uses a large training set to iteratively train the model, prompting the algorithm to continuously adjust its internal weights and tree structure, thereby deeply learning the complex mapping patterns hidden between molding sand performance parameters, dynamic process conditions, and optimal material addition amounts. To prevent overfitting—where the model performs well on the training set but has poor generalization ability on unknown data—the system continuously uses independent validation sets to dynamically adjust the model's hyperparameters during training and compares the performance of multiple candidate models, ultimately selecting the optimal model with the smallest prediction error on the validation set. Notably, considering the typically high noise and strong nonlinearity of molding sand production environment data, this invention introduces the Huber loss function into the model's objective function. The Huber loss function combines the advantages of mean squared error and mean absolute error, using squared loss for normal samples to accelerate convergence, while using linear loss for anomalous samples with large deviations to reduce their negative impact on the overall model gradient, thus effectively reducing overfitting and improving the model's robustness to production fluctuations. Meanwhile, the system also fully incorporates the actual constraints of the industrial site during the model training and verification process, and includes the specific parameter ranges of each production line as process constraints into the cross-validation system.In large-scale hyperparameter tuning experiments using grid search or random search, the system found that tree depth has the most significant impact on molding sand prediction accuracy, directly determining the model's ability to capture complex feature interactions. The learning rate is the most sensitive parameter in gradient boosting tree models; an excessively high learning rate leads to model oscillations and failure to converge, while an excessively low learning rate results in excessively long training times or even getting stuck in local optima. Therefore, the system performed independent parameter optimization for different production lines and different target prediction variables to ensure that each model perfectly fits the specific process scenario. After training and rigorous validation, the system enters the model deployment and application phase, solidifying the trained optimal model, saving its network structure and parameter weights, and seamlessly integrating it into the inference unit of the online system. In actual production, when the user inputs the performance testing parameters for a new batch of molding sand and the predetermined process target, the system's inference unit instantly calls the solidified model for forward inference calculations. Taking into account preset process constraints, it outputs the predicted values ​​for recommended dust addition, clay addition, and mixed soil addition in real time and with high accuracy.

[0026] The system not only provides accurate data predictions but also boasts powerful visualization and analysis capabilities for the prediction results. It visually presents the complex mathematical model predictions and multi-dimensional underlying data in a graphical format, assisting process managers in making informed decisions. This module offers a rich variety of data visualization components, including a comparison chart of predicted and historical actual values. This chart typically uses a dual-axis line graph to intuitively demonstrate the model's prediction fidelity within a historical time window and its ability to capture future trends. Furthermore, the system provides time series charts of key performance indicators, displaying historical fluctuations and future predicted trends of core indicators such as moisture content and wet compressive strength as continuous curves, helping managers identify potential quality drift that could lead to defective products. To analyze the stability of individual indicators, the system generates data distribution analysis charts, such as box-and-whisker plots based on quartiles, to show the central tendency, dispersion, and outlier distribution of data within a certain production cycle. The system also provides correlation heatmaps, which, by calculating Pearson or Spearman correlation coefficients between various variables, visually display the statistical correlation between moisture content, permeability, strength, and the amount of various materials added using different shades of color blocks, providing a theoretical basis for adjusting process parameters. In the data output and decision support phase, the system presents the predicted material addition amounts obtained through model inference to the end user in a clear and conspicuous interface format, serving as the core guidance for on-site batching operations. Simultaneously, responding to user commands, the data output module integrates and encapsulates the input data, model predictions, and various automatically generated in-depth analysis charts into a structured business report, such as a spreadsheet or portable document, for rapid export. This report is widely used for daily production guidance, shift handover records, or long-term quality traceability archiving.

[0027] To fully verify the feasibility and outstanding industrial application value of the technical solution provided by this invention, this embodiment also discloses in detail the software implementation scheme of a molding sand data management and prediction system and its deployment verification process in a real industrial scenario. During the system development and configuration phase, the R&D team used Python programming language as the core backend development language, fully utilizing its mature ecosystem in the field of data science. They called upon core libraries for implementing machine learning algorithms, analysis libraries for large-scale structured data processing and cleaning, and mathematical libraries for high-speed calculations of low-level multidimensional arrays, constructing an extremely robust core algorithm engine and data processing logic. Regarding the front-end human-computer interaction, the system uses a cross-platform graphical user interface framework to build a desktop client application. Initially, it focused on deep adaptation and performance optimization for the Windows operating system, which has the highest penetration rate in industrial settings. In terms of data interaction interface design, the system fully supports high-concurrency batch import operations of historical data from various versions of spreadsheet formats and comma-separated value formats. It also supports the reverse export of model prediction results and visualization charts to the same compatible format. This two-way channel design greatly facilitates seamless data interaction between this system and existing enterprise resource planning or manufacturing execution management systems in foundry enterprises. Through the integration of the above technology stack, the prototype system has successfully achieved a complete functional chain covering the entire lifecycle, including manual and automatic data entry, automated preprocessing, parallel training and cross-evaluation of multi-algorithm models, accurate prediction of real-time addition volume, multi-dimensional result visualization rendering, and customized report generation.

[0028] During the actual testing and application verification phase, the system was officially deployed in the core industrial control network of the sand processing workshop of a large automotive parts foundry, undergoing long-term and rigorous on-site operational testing. During this period, the system continuously collected daily multi-frequency test data of real molding sand from the production line for two hundred consecutive days, along with corresponding actual material addition records and complex equipment process parameters, thus constructing a massive historical dataset with highly representative characteristics and industrial noise features. In the daily standard operating procedure, on-site operators only need to input the molding sand performance test results obtained during their shift, such as key indicators like moisture content, wet compressive strength, and permeability index, into the system interface. The system will then extremely quickly call upon the adaptively trained prediction model in the cloud or locally, outputting precise recommended addition amounts of dust, bentonite clay, and mixed materials to the batching terminal within milliseconds to seconds. Furthermore, thanks to the system's built-in time series analysis algorithm engine, the system also possesses a forward-looking function for predicting macro-trends in molding sand material addition amounts for the next two consecutive days. This provides invaluable digital reference data for the workshop's material procurement, inventory preparation, and production scheduling. After hundreds of days of comparison with actual operating data and quantitative evaluation of model performance, the prediction models independently constructed for different materials all demonstrated remarkable high-precision fitting levels. In the generally accepted model evaluation system in the field of machine learning, this embodiment uses the coefficient of determination (COD) as the main quantitative evaluation index to measure the model's explanatory power and goodness of fit. Quantitative test results show that the dust addition prediction model constructed using the gradient boosting tree algorithm achieved a COD of 0.815; the clay (bentonite) addition prediction model constructed using the same algorithm achieved a COD of 0.820; and the mixed material addition prediction model constructed using the random forest algorithm achieved an astonishing COD of 0.871. In addition to the COD, the system continuously monitors indicators such as mean absolute error and root mean square error in the background to comprehensively understand the model's performance in penalizing extreme values ​​and its robustness drift over long periods. These solid quantitative results unequivocally demonstrate that the machine learning prediction model constructed in this invention can deeply and effectively learn the extremely complex physical and chemical correlations hidden in massive amounts of messy historical data. Compared with traditional methods that rely heavily on the subjective experience of senior craftsmen or on rudimentary linear empirical formulas, the prediction results provided by this system have unparalleled objectivity, consistency, and accuracy.

[0029] This invention, compared to existing molding sand processing control technologies, exhibits a series of significant and disruptive technological advantages and beneficial effects. First, it fundamentally realizes data-driven scientific decision-making and standardized process operation. By constructing an intelligent prediction model deeply integrated with historical data, it completely ends the fuzzy decision-making mode of traditional foundries that has long relied on manual experience or simple estimations. The system can stably and continuously output objective and highly consistent addition amount prediction suggestions, successfully transforming the key black-box process of molding sand batching, often considered an art of experience, into a standardized industrial operation that can be precisely quantified and infinitely replicated. This significantly reduces the severe fluctuations in molding sand quality caused by differences in human judgment standards among different operators, ultimately significantly improving the quality stability and batch consistency of the final casting product. Second, the method and system of this invention greatly improve the production response speed and batching decision efficiency in industrial settings, perfectly meeting the fast-paced, high-load production demands of modern manufacturing. The fully automated and rapid processing flow, from real-time input of raw data to instant generation of final prediction results, allows operators to input only a few key performance parameters, and the system can instantly output the optimal addition amount operation suggestion. This process completely eliminates the severe time delays caused by cumbersome manual data retrieval, inefficient paper-and-pen calculations, or lengthy technical discussion meetings in traditional production models. It ensures that the precise mixing of molding sand keeps pace with the high-speed production cycle, thereby comprehensively improving the operational efficiency of the sand processing department and the entire casting production line. More significantly, this invention successfully constructs a highly integrated intelligent closed-loop control system, greatly reducing the company's unhealthy reliance on the personal experience of specific senior technical personnel and achieving the digital transfer of valuable industrial knowledge at the system level. This system abandons the drawbacks of past information silos, seamlessly integrating previously fragmented production data management, in-depth multi-dimensional data analysis, and intelligent forward-looking predictive decision-making functions onto a unified and efficient industrial internet platform, forming a complete intelligent data flow closed loop encompassing data acquisition, analysis and modeling, decision support, and result feedback. This not only enables inexperienced young operators to prepare molding sand that meets stringent technical standards with the help of the system, lowering the operational threshold, but more importantly, it transforms the valuable production experience accumulated by the company over the years and the tacit knowledge of experts into clear neuron weights and decision tree branches through advanced mathematical algorithms. In the form of digital assets, it realizes the permanent solidification and sustainable intergenerational inheritance of the company's core process technology, fundamentally and effectively mitigating the fatal risk of technology gap caused by the retirement or departure of older employees.Furthermore, addressing the industry-recognized core technical challenge of the highly coupled physicochemical parameters and extremely complex nonlinear relationships among influencing factors in molding sand systems, this invention innovatively employs and finely trains advanced machine learning nonlinear regression models such as gradient boosting trees and random forests. This results in extremely high prediction accuracy, strong process targeting, and powerful core capabilities for handling complex multidimensional spatial mapping relationships. The model architecture possesses exceptional self-learning and automatic feature extraction capabilities, automatically mining deep implicit mapping relationships between various performance parameters, external process constraints, and optimal material proportions from vast and often disruptive historical datasets. The accuracy and fit of its actual prediction output far exceed the limits of relying on traditional thermodynamic empirical formulas or rough manual estimations. This provides the most solid, reliable, and irreplaceable underlying technical support for the refined batching process of high-precision castings, resulting in significant economic and practical benefits such as reduced scrap rates and material savings. Finally, through a systematic operating mechanism, this invention has endogenously formed a complete, traceable, and sustainably self-iteratory data asset system for foundry enterprises, laying an irreplaceable foundation for their future move towards fully intelligent manufacturing. The system's underlying database stores in extremely detailed detail all raw, coarse data entered from sensors or manual forms, preprocessing action records after cleaning and transformation, predicted numerical results from each model inference, and quality feedback information ultimately guiding actual production. These data are interwoven and interconnected, forming a highly structured, fully lifecycle-traceable core data asset goldmine for molding sand production. This not only greatly facilitates process engineers in daily management by enabling them to quickly and accurately retrospectively analyze and diagnose sudden abnormalities in molding sand quality, but also, more importantly, provides the highest quality data fuel for the continuous iteration and reconstruction optimization of the predictive model. This gives the system a life-like growth attribute, allowing it to continuously fine-tune model weights and evolve as daily production data accumulates in the workshop. This ensures that the system's intelligence level and predictive accuracy continue to rise over time, maintaining a long-term technological advantage.

[0030] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for managing molding sand performance data and predicting its addition amount, comprising the following steps: Receive and integrate molding sand production-related data from different sources to construct a structured dataset. The acquired data includes molding sand performance parameters, historical material addition data corresponding to the molding sand performance parameters, and related process and environmental parameters. The acquired raw data is cleaned, standardized, and features are constructed to generate a high-quality feature dataset for use by machine learning models. A machine learning regression algorithm is used to train and cross-validate the model using the feature dataset. The Huber loss function is introduced into the objective function of the model and combined with the specific process constraints of the production line to generate the best prediction model after training. Receive the performance parameters and process objectives of the molding sand for the current batch, call the best prediction model to perform forward inference calculations, and output the recommended addition amount prediction values ​​of dust, clay and mixed materials in real time.

2. The method for managing molding sand performance data and predicting addition amount according to claim 1, characterized in that, When constructing the features, derived features are created based on the acquired raw data to enhance the model’s understanding of the process logic. The derived features include the amount of ineffective clay added and the proportion of core entry. The amount of ineffective clay added is configured as the mud content value minus the blue absorption value. The proportion of core entry is configured as the amount of sand cores entering today divided by the amount of mixed sand. When calculating the amount of sand cores entering today, the amount of cold cores entering and the amount of hot cores entering are counted independently.

3. The method for managing molding sand performance data and predicting addition amount according to claim 1, characterized in that, During the data cleaning process, outliers exceeding physical limits are automatically identified and processed. Missing values ​​in the data records are repaired using time-series-based linear interpolation, fill-in of preceding and following values, or fill-in strategies based on statistical characteristics. The data standardization includes normalizing parameters of different dimensions and ranges to a unified interval through mathematical transformations to eliminate the influence of dimensions.

4. The method for managing molding sand performance data and predicting addition amount according to claim 1, characterized in that, The machine learning regression algorithm employs either gradient boosting tree or random forest algorithms. During the model training phase, the feature dataset is divided into a training set and a validation set. While adjusting the model hyperparameters using the validation set, the depth and learning rate of the decision tree are independently optimized for different production line types and different target predictor variables to capture the complex nonlinear interaction features between molding sand parameters.

5. The method for managing molding sand performance data and predicting addition amount according to claim 1, characterized in that, After outputting the predicted value of the recommended addition amount, the model prediction results and historical related data are presented in a graphical way. The visualization presentation includes generating a comparison chart of the predicted value and the historical actual value to show the fidelity of the model prediction, a time series chart of key performance indicators to show the fluctuation trend of core indicators, a box-and-whisker plot to analyze the stability of indicator distribution, and a correlation heatmap to show the degree of statistical correlation between variables. The prediction results and analysis charts are then integrated into a structured report file and exported according to the instructions.

6. A system for managing molding sand performance data and predicting its addition amount, characterized in that, The logical architecture of the system includes a data access layer (201), a data storage layer (202), a business logic and computing layer (203), and a user interaction layer (204). The data access layer (201) is equipped with a data entry and integration module, which provides a human-computer interaction interface that supports manual structured entry and batch file import, and is used to receive diverse molding sand production-related data and perform data format verification and basic deduplication. The data storage layer (202) is equipped with a central database for persistently storing the input raw data, the generated preprocessed data, the trained model parameters, and the generated historical prediction records. The business logic and computing layer (203) is equipped with a data processing and feature engineering module and an intelligent prediction core module. The data processing and feature engineering module is configured to perform data cleaning, standardization and feature construction operations to transform the raw data into a high-quality feature dataset. The intelligent prediction core module includes a model training unit and a model inference unit. The model training unit is configured to call a machine learning algorithm library and use the feature dataset to perform model training and optimization processes to generate a trained prediction model. The model inference unit integrates the prediction model to calculate and return the predicted value of material addition in real time based on new input parameters. The user interaction layer (204) is equipped with a visualization and human-computer interaction module and a data output and report generation module. The visualization and human-computer interaction module is configured to generate and render multi-dimensional data analysis charts to display deep data trends, model performance and prediction results. The data output and report generation module is configured to encapsulate the prediction results and analysis charts into a report file of a specified format and provide an export function.

7. The molding sand performance data management and addition prediction system according to claim 6, characterized in that, When training the model, the intelligent prediction core module configures the Huber loss function in the objective function to reduce overfitting, and performs cross-validation by combining the specific process range of each production line extracted from the central database as a constraint condition. The model inference unit resides in memory and is instantly called through the application programming interface to achieve a second-level prediction response.

8. The molding sand performance data management and addition prediction system according to claim 6, characterized in that, When performing the feature construction operation, the data processing and feature engineering module calculates the amount of ineffective clay added and the proportion of core-injected material as derived features for use by the machine learning model.

9. The molding sand performance data management and addition prediction system according to claim 6, characterized in that, The data entry and integration module receives molding sand performance parameters including moisture content, air permeability, wet compressive strength, sand temperature, mud content, loss on ignition, blue absorption, and particle size distribution. It also receives historical material addition data including the addition mass of dust, bentonite clay, and mixed soil.

10. The molding sand performance data management and addition prediction system according to claim 6, characterized in that, The visualization and human-computer interaction module generates multidimensional data analysis charts including a comparison chart of predicted values ​​and historical actual values ​​in the form of a double vertical axis line chart, a time series chart of key performance indicators in the form of a continuous curve, and a box-and-whisker plot based on quartiles.

Images