A casting process optimization method and system based on interpretable artificial intelligence

Abstract

The application relates to a casting process optimization method and system based on an interpretable artificial intelligence, and belongs to the technical field of intelligent control. Through data acquisition and preprocessing, field knowledge driven process feature modeling, interpretable artificial intelligence modeling and optimization decision, decision explanation and visualization, and man-machine collaborative closed-loop feedback, field knowledge and an artificial intelligence model are deeply fused, an interpretable decision basis is provided while an optimized process scheme is output, and the model is continuously iterated and updated in combination with man-machine collaborative closed-loop feedback. The application effectively solves the problems that traditional casting process optimization relies on artificial experience, an intelligent model is a black box and cannot be interpreted, generalization is poor, and continuous optimization is impossible, and improves the engineering applicability, quality stability and production efficiency of casting process optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a casting process optimization method and system based on interpretable artificial intelligence. Background Technology

[0002] The casting process is characterized by numerous steps, complex parameter coupling, and strong nonlinearity. Product quality is affected by many factors, including raw material composition, melting state, pouring conditions, cooling methods, and equipment status. With the development of casting equipment towards larger scale, automation, and digitalization, the scale of production data continues to grow. Traditional optimization methods that rely on manual experience and single parameter adjustments are no longer sufficient to meet the demands for quality stability and production efficiency.

[0003] In recent years, data-driven artificial intelligence methods have been gradually applied to casting quality prediction, defect identification, and process parameter optimization, improving decision-making efficiency and accuracy. However, existing methods have significant drawbacks: First, the internal mechanisms of the models are opaque, and the decision-making basis is uninterpretable. Process engineers cannot understand the physical meaning and process logic of parameter adjustments, making engineering verification and risk control difficult and limiting practical applications. Second, pure data-driven modeling fails to effectively integrate knowledge from metallurgical mechanisms, solidification behavior, heat and mass transfer, resulting in insufficient model stability and generalization ability when operating conditions change or data distribution shifts. Third, the model outputs are numerical or abstract indicators that are not mapped to engineering language and process experience, making it difficult to directly guide on-site adjustments. Fourth, the human-machine interaction is one-way, preventing process engineers from intervening and providing feedback, hindering continuous model correction, and limiting long-term application value.

[0004] Therefore, there is an urgent need for a casting process optimization solution that takes into account optimization effect, interpretability, domain knowledge integration and human-machine collaboration, so as to make decisions understandable, verifiable and traceable, and improve the engineering applicability and promotion value of artificial intelligence in the casting field. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a casting process optimization method and system based on interpretable artificial intelligence. It is applicable to various casting process scenarios and production lines for various casting equipment, including cast iron, cast steel, and non-ferrous metals. It solves the technical problems of traditional casting process optimization relying on human experience, the uninterpretable nature of intelligent models (which are often black boxes), poor generalization, insufficient human-computer interaction, and the inability to continuously optimize.

[0006] In a first aspect, the present invention provides a casting process optimization method based on interpretable artificial intelligence, the method comprising the following steps:

[0007] Data is collected from multiple sources and heterogeneous data throughout the entire casting production process, including process parameters, equipment operating status, and casting quality inspection results. The data is then preprocessed to form standardized data for process feature modeling.

[0008] Based on engineering knowledge in the casting field, the process sequentially performs the following steps: process stage division and parameter classification, key process parameter identification, process feature reconstruction and mapping, multi-parameter coupled feature construction, process constraint and rule embedding, and feature validity verification and updating. This achieves deep integration of engineering knowledge and standardized data from multiple process sources, forming a set of process features.

[0009] Based on interpretable artificial intelligence models, the correlation between process characteristics and quality is explored. By combining process constraints and engineering logic, an intelligent decision-making model for casting process optimization is constructed. Through a traceable interpretation mechanism, intelligent decision results and structured interpretation data are generated, providing explanatory information corresponding to process characteristics to support engineering understanding and verification.

[0010] The abstract intelligent decision-making results are transformed into engineering expressions that fit the actual production. The content is visualized by relying on diverse charts and graphs, and a multi-scheme comparison and judgment mode is used to reflect the relationship between parameter control and casting quality. The output is standardized decision explanation information after engineering transformation, multi-type visualization analysis charts, and multi-dimensional optimization scheme comparison analysis results.

[0011] By deeply integrating intelligent decision-making results, process personnel experience, and on-site production verification data, process personnel review the process optimization plan, and then the approved plan is put into actual casting production and relevant data is collected to form a structured feedback dataset. After effectiveness evaluation, the feedback data is used to update and adaptively optimize the interpretable artificial intelligence model, thus building a human-machine collaborative closed loop.

[0012] Optionally, the standardization preprocessing is characterized by including five preprocessing steps performed sequentially: data cleaning, outlier handling, missing value completion, time alignment, and data standardization.

[0013] Optionally, the data is characterized by: data cleaning to remove invalid, redundant, erroneous and meaningless information from the original data; outlier handling to identify and correct or remove extreme data that deviates from the normal production range; missing value completion to restore missing data caused by data acquisition interruption, communication failure, or manual omission; time alignment to unify the time base of multi-source data and achieve accurate matching of process parameters, equipment status and quality data; and data standardization to eliminate differences in dimensions and numerical magnitudes between different parameters.

[0014] Optionally, the process stage division and parameter classification are characterized by the following: based on the actual process flow and physical change law of casting production, the complete casting production process is divided into five consecutive processes: melting stage, heat preservation stage, pouring stage, solidification and cooling stage and post-processing stage. The process is then accurately classified according to the actual production links, so that each standardized data corresponds to a specific process stage, and the data is bound to the process behavior and physical process.

[0015] Optionally, the identification of key process parameters specifically includes: combining casting metallurgy theory, solidification phase transformation mechanism, defect formation law and field engineering experience, analyzing the degree of influence and significance of all parameters contained in the standardized data in each stage, and screening out key process parameters from them.

[0016] Optionally, the key process parameters include variables of temperature, time, component ratio, and cooling conditions; non-key parameters are uniformly labeled as auxiliary variables or background variables.

[0017] Optionally, the process feature reconstruction and mapping specifically includes: carrying out feature reconstruction through parameter combination, proportional relationship construction, stage statistics, interval mapping, and trend extraction, so as to upgrade the standardized data from basic numerical form to process feature variables that can characterize the physical behavior of the casting process.

[0018] Optionally, the construction of the multi-parameter coupling feature specifically includes: based on standardized data of multiple key process parameters, mining the joint distribution patterns, synergistic change trends and interaction relationships contained in the data, and constructing coupling feature variables that reflect the combined effect of multiple parameters.

[0019] Optionally, the process constraints and rule embedding specifically includes: embedding process limitations, equipment operating limits, production safety specifications, quality control standards, and mature engineering experience rules in the casting field into the feature construction process in the form of constraints, and directly applying them to the standardized data on which the generated features are based.

[0020] Secondly, the present invention provides a casting process optimization system based on interpretable artificial intelligence, which is used to implement the casting process optimization method based on interpretable artificial intelligence. The system includes a data acquisition and preprocessing module, a domain knowledge-driven process feature modeling module, an interpretable artificial intelligence modeling and optimization decision module, a decision interpretation and visualization module, and a human-machine collaborative closed-loop feedback module.

[0021] The data acquisition and preprocessing module is used to acquire multi-source heterogeneous data, including process parameters, equipment operating status, and casting quality inspection results, throughout the entire casting production process, and to perform standardized preprocessing to form standardized data for process feature modeling.

[0022] The domain knowledge-driven process feature modeling module is used to perform process stage division and parameter classification, key process parameter identification, process feature reconstruction and mapping, multi-parameter coupled feature construction, process constraint and rule embedding, and feature validity verification and updating in sequence, based on engineering knowledge in the casting field, in order to achieve deep integration of engineering knowledge and standardized data from multiple sources of processes to form a set of process features.

[0023] An interpretable AI modeling and optimization decision-making module is used to explore the correlation between process characteristics and quality. It combines process constraints and engineering logic to build an intelligent decision-making model for casting process optimization. Through a traceable interpretation mechanism, it forms intelligent decision results and structured interpretation data, providing interpretation information corresponding to process characteristics to support engineering understanding and verification.

[0024] The decision interpretation and visualization module is used to transform abstract intelligent decision results into engineering expressions that fit the actual production. It uses a variety of charts to complete the content visualization display, and is equipped with a multi-scheme comparison and judgment mode to reflect the relationship between parameter control and casting quality. It outputs standardized decision interpretation information after engineering transformation, multiple types of visualization analysis charts, and multi-dimensional optimization scheme comparison analysis results.

[0025] The human-machine collaborative closed-loop feedback module is used to deeply integrate intelligent decision-making results, process personnel experience, and on-site production verification data. Process personnel review the process optimization plan, and then the approved plan is put into actual casting production and relevant data is collected to form a structured feedback dataset. After effectiveness evaluation, the feedback data is used to update and adaptively optimize the interpretable artificial intelligence model, thus constructing a human-machine collaborative closed loop.

[0026] This invention provides a casting process optimization method and system based on interpretable artificial intelligence. The technical solutions provided by the embodiments of this invention bring at least the following beneficial effects:

[0027] This invention employs domain knowledge-driven process feature modeling, integrating metallurgical mechanisms, process constraints, and empirical rules into the modeling process. This avoids the model deviating from actual processes due to purely data-driven approaches, thus improving modeling reliability. Simultaneously, it combines interpretable learning models to clarify the impact of process parameters on quality, enhancing decision interpretability and reducing engineering risks. Through decision interpretation and visualization, process engineers can easily verify optimization schemes; and a human-machine collaborative closed-loop feedback mechanism is established to achieve dynamic adaptive optimization of the model. The overall solution reduces experimental costs and risks, improves optimization efficiency and quality stability, and provides a reliable technical path for intelligent casting processes.

[0028] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart illustrating a casting process optimization method based on interpretable artificial intelligence.

[0031] Figure 2 This is a schematic diagram of the composition structure of a casting process optimization system based on explainable artificial intelligence. Detailed Implementation

[0032] 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 embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0033] Before describing the technical solution of the present invention in detail, the technical background and technical terms involved in the technical solution will be explained first:

[0034] Explainable Artificial Intelligence (XAI) is a collection of technologies, methods, and theories that enable humans to understand the decision-making process, internal logic, and prediction results of artificial intelligence models. In this invention, it does not refer to a single model, but rather encompasses two technical paths: first, intrinsically explainable models, where the model structure itself is transparent and the logic is traceable, such as linear regression and decision trees; second, model interpretability enhancement techniques, which use locally interpretable methods such as LIME, SHAP value analysis, and feature importance ranking to deconstruct traditional complex black-box models (such as deep neural networks and ensemble learning), quantifying the contribution and direction of influence of each input feature on the output result. The core purpose of introducing XAI in this invention is to break the "black box" effect of traditional intelligent models in the casting field, enabling process engineers to clearly understand why a certain process parameter is chosen, clarify key influencing factors and their physical meaning, thereby establishing decision-making trust and reducing the risks of technology application.

[0035] Domain Knowledge-Driven: Throughout the entire process of casting process optimization, this approach uses the professional theories, production principles, engineering experience, and process constraints of the casting discipline as core guidance to replace or assist purely data-driven blind modeling. Specifically, in this invention, it is manifested in the deep embedding of metallurgical mechanisms, solidification behavior, heat and mass transfer laws, casting microstructure evolution knowledge, equipment performance boundaries, and industry experience rules into each stage of data acquisition, feature engineering, model building, and optimization decision-making. This driving approach ensures that the process features input to the model have clear physical meaning and engineering orientation, rather than being simple numerical combinations; it ensures that the optimized solutions output by the model conform to the logic of the casting process and can be directly applied to guide production, effectively solving the technical defects of purely data-driven models, such as poor generalization and detachment from actual physical laws.

[0036] Knowledge Graph: A technical system for the structured representation and storage of domain knowledge using a graph structure, consisting of nodes (entities, concepts) and edges (relationships, attributes). In the casting process optimization scenario of this invention, the knowledge graph serves as an important carrier and supporting tool for domain knowledge-driven processes, systematically integrating knowledge throughout the entire casting lifecycle. Its nodes can include entities such as casting process stages, raw material composition, key process parameters (e.g., pouring temperature, cooling rate), casting quality indicators (e.g., tensile strength, shrinkage defects), and equipment models; edges are used to define the logical relationships between entities, such as "pouring temperature" affecting "solidification process," and "carbon equivalent" determining "microstructure properties." This invention utilizes the knowledge graph to achieve the standardization, structuring, and visualization of domain knowledge. It provides a logical framework and constraints for process feature reconstruction and serves as a knowledge source for interpretable AI output decision explanations, assisting in mapping model results into engineering language and improving the readability and applicability of the solution.

[0037] Figure 1 This is a flowchart illustrating a casting process optimization method based on interpretable artificial intelligence. The method includes the following steps:

[0038] Step S101: Collect multi-source heterogeneous data, including process parameters, equipment operating status, and casting quality inspection results, throughout the entire casting production process, and perform standardized preprocessing to form standardized data for process feature modeling.

[0039] The above steps are for data acquisition and preprocessing, and are the initial foundational steps of the casting process optimization method of this invention. They are used to comprehensively acquire multi-source heterogeneous data highly relevant to process optimization and quality control throughout the entire casting production process, constructing a stable, reliable, and standardized data input foundation. This provides data support for subsequent domain knowledge-driven process feature modeling and interpretable artificial intelligence model training. Data acquisition is the primary input step of the casting process optimization method and system of this invention. It is used to comprehensively, in real-time, and continuously acquire multi-source, multi-dimensional, and multi-modal raw information closely related to process optimization, quality formation, and equipment operation throughout the entire casting production process, providing a complete, authentic, and traceable data source for subsequent data preprocessing, process feature modeling, and artificial intelligence modeling. The acquisition objects in this step cover the entire casting production process, mainly including three categories of data:

[0040] The first category is process parameter data, which covers all key and auxiliary parameters in the melting, holding, casting, solidification and cooling, and post-processing stages. Specifically, it includes parameters directly related to metallurgical mechanisms and solidification behavior, such as melting temperature, holding temperature, casting temperature, casting speed, casting time, cooling method, cooling rate, cooling duration, alloy composition ratio, molding parameters, and mold performance.

[0041] The second category is equipment operation status data, including real-time data reflecting the health and stability of various casting equipment such as smelting furnaces, casting machines, cooling devices, molding equipment, and control systems, including start-up and shutdown status, operating conditions, workload, operating frequency, energy consumption information, fault signals, and action sequence.

[0042] The third category is casting quality inspection results data, which includes the casting's mechanical properties, metallographic structure, internal defect detection results, dimensional accuracy results, and surface quality evaluation information, used to establish the correspondence between process parameters and final quality.

[0043] The data acquisition method supports multiple approaches, including automatic sensor acquisition, system log acquisition, manual inspection and entry, PLC data reading from equipment, and direct connection to online inspection equipment. It can achieve synchronous acquisition and unified identification of information throughout the entire process within the same furnace, the same casting, and the same process cycle, ensuring that the data source is authentic, comprehensive, and chronologically complete, providing a solid and reliable data foundation for the entire optimization system.

[0044] After comprehensive data collection, systematic and standardized preprocessing operations were carried out to address issues such as noise, redundancy, distortion, missing values, temporal misalignment, inconsistent units, and large numerical ranges in the raw data, which were caused by factors such as the complex industrial environment, sensor fluctuations, manual recording bias, and equipment communication delays. Data preprocessing is a crucial data governance step after raw data collection and before process feature modeling. It systematically regulates and processes the multi-source, heterogeneous, noisy, missing, and temporally biased raw data collected from the industrial site to improve data quality, eliminate abnormal interference, and standardize format, providing a stable and reliable data foundation for subsequent modeling. Data preprocessing includes five key operations performed sequentially: data cleaning, outlier handling, missing value completion, time alignment, and data standardization. Each operation has a clearly defined processing object, implementation method, and engineering purpose, as detailed below:

[0045] 1. Data Cleaning

[0046] Data cleaning is a fundamental preprocessing operation used to remove invalid, redundant, erroneous, and meaningless information from the raw data. Addressing issues such as communication interference in the casting field, duplicate sensor reports, and manual errors, each data entry is verified to remove duplicate data, null fields, garbled messages, incorrectly formatted records, and invalid entries that clearly violate the basic logic of the process. Through cleaning, the uniqueness, accuracy, and usability of the data are ensured, providing a clean and standardized foundation for subsequent processing.

[0047] 2. Outlier Handling

[0048] Outlier handling is used to identify, correct, or remove extreme data that deviates from the normal production range. Based on casting process ranges, equipment operating limits, industry standard thresholds, and statistical methods, it locates outliers caused by sensor drift, transient interference, and equipment failure. Correctable outliers are corrected according to process rules, while severely outliers that cannot be corrected are filtered out to prevent outliers from misleading model learning and to improve the stability and reliability of subsequent modeling and decision-making.

[0049] 3. Missing value completion

[0050] Missing value completion is used to recover missing data caused by data acquisition interruptions, communication failures, or human error. For missing items in process parameters, equipment status, and quality results, appropriate methods such as nearest neighbor interpolation, batch mean / median filling, process rule derivation filling, and historical similar data fitting filling are used, depending on the data type and degree of missingness. After completion, data continuity and integrity are maintained, ensuring that process flow information is not interrupted or lost.

[0051] 4. Time alignment

[0052] Time alignment is used to unify the time reference of multi-source data, enabling precise matching of process parameters, equipment status, and quality data. Data from different acquisition nodes, sampling frequencies, and timestamps are calibrated along a unified timeline according to casting furnace number, casting number, and process stage, ensuring a one-to-one correspondence between data from smelting, pouring, cooling, and post-processing within the same production cycle and corresponding quality inspection results. Time alignment guarantees accurate data correlation, providing a time-consistent data foundation for process mechanism analysis and model training.

[0053] 5. Data Standardization

[0054] Data standardization is used to eliminate differences in units and magnitudes among different parameters. Data with different physical meanings, units, and numerical ranges, such as temperature, time, component ratios, energy consumption, and mechanical properties, are mapped to a unified range through methods like normalization and standardization. After standardization, the weights of each parameter's influence on the model are more reasonable, avoiding model training imbalances caused by excessively large or small values, and improving the convergence speed and prediction accuracy of interpretable artificial intelligence models.

[0055] Step S102: Based on engineering knowledge in the casting field, the following steps are executed in sequence: process stage division and parameter classification, key process parameter identification, process feature reconstruction and mapping, multi-parameter coupling feature construction, process constraint and rule embedding, and feature validity verification and updating, so as to achieve deep integration of engineering knowledge and standardized data from multiple sources of processes to form a set of process features.

[0056] The above steps achieve domain knowledge-driven process feature modeling. This step uses the standardized data output from the data acquisition and preprocessing steps as the sole input data source. This data is characterized by consistent timing, standardized format, absence of anomalies and missing data, and uniform dimensions, providing real, reliable, and traceable data support for the entire process modeling. Based on the standardized data, this step takes the casting metallurgical mechanism, solidification behavior, heat and mass transfer laws, process constraints, and engineering experience rules as the core guidance. Throughout the feature construction process, it consistently uses the fundamental theories of casting science, the physical changes in casting formation, the constraints that must be followed on the production site, and long-accumulated practical experience as judgment criteria, design principles, and constraints. It avoids pure data calculations that are divorced from engineering logic, ensuring that all feature extraction, parameter selection, and relationship construction have clear physical meaning and engineering rationality, thus guaranteeing the consistency of the model input mechanism from the root. Through six sequentially connected and progressively advancing sub-steps, the raw basic data is transformed into structured process features with clear physical meaning, engineering orientation, and consistency with technological mechanisms. This provides high-quality, mechanistically reliable model input for subsequent interpretable artificial intelligence modeling and optimization decision-making, achieving a smooth transition and effective connection from the data layer to the decision layer. Each sub-step is described below.

[0057] 1. Process Stage Division and Parameter Classification

[0058] This sub-step directly uses the standardized data output from data acquisition and preprocessing as its processing object. Based on the actual process flow and physical change laws of casting production, the complete casting production process is divided into five consecutive processes: melting, holding, pouring, solidification and cooling, and post-processing. Using the timestamps, process identifiers, equipment numbers, furnace information, and process stage affiliation markers inherent in the standardized data, all process parameter data, equipment operating status data, and quality inspection data are precisely categorized according to their actual production stages. This ensures that each piece of standardized data corresponds to a specific process stage, achieving a one-to-one binding of data with process behavior and physical processes. This forms a phased, structured, and logically clear data system, laying a standardized and orderly data framework for subsequent parameter analysis and feature construction.

[0059] 2. Identification of key process parameters

[0060] This sub-step uses standardized data, after process stage division and classification, as the analysis carrier. Combining casting metallurgy theory, solidification phase transformation mechanisms, defect formation laws, and field engineering experience, it analyzes the impact and significance of all parameters included in the standardized data within each stage. From this, parameters that have a direct, significant, and decisive effect on the casting forming process, internal microstructure, and final quality indicators are selected and defined as key process parameters, mainly including core variables such as temperature, time, composition ratio, and cooling conditions. Non-key parameters with weak influence, low correlation, or indirect effects are uniformly marked as auxiliary variables or background variables and retained only as supplementary information. This approach fully utilizes the complete information of the standardized data while achieving focused extraction of core data, effectively reducing redundancy in subsequent modeling and highlighting process factors that play a dominant role in quality.

[0061] 3. Reconstruction and Mapping of Process Features

[0062] This sub-step uses standardized data corresponding to key process parameters as its processing foundation. It deeply processes and transforms single, discrete raw data into engineering-oriented data. Through parameter combination, proportional relationship construction, staged statistics, interval mapping, and trend extraction, it reconstructs features, upgrading the standardized data from basic numerical form into process characteristic variables that can characterize the physical behavior of the casting process. The reconstructed features directly reflect core mechanisms such as energy input intensity, heat release rate, phase transition development trend, and microstructure evolution law. This achieves the transformation from standardized data to features with clear physical meaning and engineering orientation, ensuring that each feature highly corresponds to on-site process operations and metallurgical physical mechanisms, thus enhancing the engineering expressiveness and interpretability of the data.

[0063] 4. Construction of multi-parameter coupled features

[0064] This sub-step, based on standardized data of multiple key process parameters, fully explores the joint distribution patterns, synergistic change trends, and interactive relationships contained in the data. Addressing the characteristics of multiple parameters in the casting process—mutual influence, mutual constraint, and strong coupling—it constructs coupled feature variables that reflect the combined effects of multiple parameters. These features can comprehensively express the direction, magnitude, enhancement, inhibition, and synergistic influence relationships between parameters. By fully utilizing the multi-dimensional correlation information in the standardized data, it solves the problem that single parameters cannot describe complex process behaviors and are difficult to accurately reflect actual production mechanisms. This significantly improves the accuracy and coverage of the features in representing the entire casting process, making the model input closer to the real process logic.

[0065] 5. Process Constraints and Rule Embedding

[0066] This sub-step embeds casting process limitations, equipment operating limits, production safety regulations, quality control standards, and mature engineering experience rules into the feature construction process as constraints, directly impacting the standardized data upon which the generated features are based. The process features built on standardized data undergo value verification, rationality assessment, and compliance screening. Invalid features generated from standardized data that exceeds process limits, violates metallurgical principles, does not meet equipment capabilities, or poses production risks are eliminated. This ensures that all ultimately retained process features are built upon compliant, reliable, and executable standardized data, guaranteeing the engineering practicality and process safety of the features from the outset.

[0067] 6. Feature validity verification and updating

[0068] This sub-step again fully utilizes all standardized data output from data acquisition and preprocessing, along with the corresponding casting quality inspection results, to perform multi-dimensional quantitative verification of all process features constructed in the preceding steps. Evaluation indicators include the feature's contribution to quality indicators, correlation, stability, discriminative power, and robustness. Based on the feedback from standardized and quality data, features that fail verification, have low contribution, exhibit large fluctuations, or are easily affected by operating conditions are adjusted, replaced, or eliminated. Features with clear mechanisms, excellent performance, and strong applicability are retained and continuously optimized. Through iterative updates, a stable, efficient, reliable set of process features highly matched with standardized data is formed, providing high-quality, highly reliable, and directly usable standard input for subsequent interpretable artificial intelligence modeling and optimization decisions.

[0069] The process feature set formed by the above steps possesses physical meaning, engineering orientation, and consistency with the process mechanism. It is directly used as the input vector for an interpretable artificial intelligence model to establish the mapping relationship between process features and casting quality indicators, conduct process influence relationship analysis, generate process optimization decisions, and support the model in extracting explanatory information such as feature contribution and parameter sensitivity. This ensures that subsequent decisions have mechanistic consistency and engineering interpretability from the outset. Simultaneously, it provides a unified feature dimension and engineering semantic foundation for subsequent decision explanation and visualization, as well as human-machine collaborative closed-loop feedback. This allows explanatory information to be accurately mapped to process parameters, and human-machine feedback to effectively contribute to model updates, ensuring logical coherence, data consistency, and engineering feasibility throughout the entire process.

[0070] Step S103: Based on the interpretable artificial intelligence model, the correlation between process features and quality is mined. Combined with process constraints and engineering logic, an intelligent decision-making model for casting process optimization is constructed. Through a traceable interpretation mechanism, intelligent decision-making results and structured interpretation data are formed, providing interpretation information corresponding to process features to support engineering understanding and verification.

[0071] The above steps are used to achieve interpretable AI modeling and optimization decision-making, serving as a core link between "data preprocessing" and "on-site production implementation." Its core objective is to construct an interpretable, implementable, and traceable intelligent decision-making system based on high-quality data output from data acquisition and preprocessing, combined with expertise in the casting field. This completely solves the "black box" problem of traditional AI models, achieving scientific and practical process optimization. This step uses domain knowledge-driven process feature sets as the core input, and through a complete process of "model building - relationship modeling - optimization decision-making - interpretation extraction - structured output," achieves intelligent optimization and traceable decision-making in the casting process. The following sections describe each sub-step in detail.

[0072] 1. Model Input Construction

[0073] Model input construction is the fundamental prerequisite for interpretable AI modeling and a core step in ensuring the accuracy and rationality of subsequent modeling. The entire process strictly adheres to the principles of "data compliance, semantic clarity, and mechanism adaptation." First, the sole data source for model input is the standardized data output from the preceding data acquisition and preprocessing stages. This data has undergone noise reduction, redundancy removal, and time alignment, possessing completeness, consistency, and reliability, and can be directly used for model training and analysis. Second, the process feature variables in the process feature set are further screened and standardized, eliminating redundant features with low relevance to casting process optimization and quality control, retaining core feature variables with clear physical meaning and direct relevance to casting quality. These process feature variables have all been validated using domain knowledge and conform to the physical laws of casting production. Finally, according to the training requirements of interpretable AI models, the screened process feature variables undergo format standardization and range normalization to eliminate dimensional differences between different physical quantities, ensuring all input data are within a unified analytical dimension. This provides standardized and unified data support for subsequent model building and feature learning, achieving seamless integration with the preceding process feature modeling stages and ensuring the coherence and logic of the entire modeling process.

[0074] 2. Explainable AI model selection and construction

[0075] Model selection and construction are central to achieving "interpretability." The entire process revolves around the production characteristics, process complexity, and engineering implementation needs of the casting industry. Traditional "black box" models are resolutely rejected, with priority given to learning models possessing inherent interpretability, or interpretability enhancement mechanisms introduced into conventional models. This ensures that every step of the model's computation and every decision has clear physical meaning and engineering logic. In model architecture design, professional knowledge from the casting field is fully integrated, embedding core requirements such as metallurgical mechanisms, process specifications, and equipment operating boundaries into the model structure, clearly defining the model's input-output logic, feature weight allocation, and constraints. During model training, considering the continuity and multi-parameter coupling characteristics of casting production, the hyperparameter settings of the model are optimized, focusing on strengthening the model's learning of process mechanisms rather than simply data analysis and fitting. Simultaneously, the model construction process strictly adheres to a dual logic of "domain knowledge guidance + data-driven," ensuring both the model's predictive accuracy and the deconstructibility and traceability of the model's reasoning process. This ensures that the model's computational logic highly aligns with the physical laws of the casting process and actual production, avoiding modeling results that are divorced from engineering reality.

[0076] 3. Modeling of process influence relationships

[0077] Process influence relationship modeling is a crucial bridge connecting model training and actual production decisions. Its core lies in using interpretable artificial intelligence models to deeply explore the intrinsic relationships between process characteristics and casting quality, production efficiency, and equipment operation, overcoming the shortcomings of traditional data-driven modeling that "only looks at the phenomena, not the essence." Specifically, based on standardized input process characteristic data, combined with professional knowledge such as casting metallurgy mechanisms and solidification laws, model training and data analysis establish mapping relationships between process characteristic variables in the process characteristic set and casting quality indicators (such as mechanical properties and surface quality) and production process parameters (such as energy consumption and production efficiency). Simultaneously, it quantitatively analyzes the direction (positive promotion or negative inhibition) and magnitude of the influence of different process characteristics, clarifies the synergistic and antagonistic effects between characteristics, accurately identifies the core characteristics that dominate casting quality, and the mutual constraints and complementary relationships between different process parameters. Furthermore, this step can also uncover the sensitive ranges of process parameters, providing a clear theoretical basis for subsequent process optimization and parameter adjustment, ensuring that the optimization direction aligns with the physical essence of casting production.

[0078] 4. Process optimization decision generation

[0079] Process optimization decision generation is the core output of the entire interpretable AI modeling and optimization decision-making process. Its core objective is to generate feasible and executable process optimization solutions based on the previously established process influence relationship model, combined with actual on-site production needs, equipment capabilities, and quality standards. During the decision generation process, the principles of "mechanism compliance, engineering feasibility, and quality priority" are strictly followed. First, the optimization objectives are clearly defined (e.g., improving casting yield, reducing energy consumption, and minimizing defects). Then, based on the process influence relationship model, process parameters are iteratively analyzed across multiple dimensions and scenarios to select the optimal parameter combination that balances quality, efficiency, and cost. Optimization solutions include both single-parameter adjustment suggestions (e.g., fine-tuning temperature and time) and multi-parameter collaborative adjustment schemes (e.g., coordinating adjustments to component ratios and cooling rates). The basis, scope, and expected effects of each adjustment are clearly defined to ensure that the solutions comply with on-site production conditions, equipment operating limits, and process specifications. This avoids optimization suggestions that are unimplementable or inconsistent with engineering realities, truly achieving a dual guarantee of "data-driven + mechanism-supported" solutions.

[0080] 5. Information extraction for decision interpretation

[0081] Extracting explanatory information for decisions is a crucial step in achieving "explainability" and a core advantage that distinguishes it from traditional black-box models. Its purpose is to ensure that every step of intelligent decision-making is understandable and verifiable. While generating process optimization decisions, it comprehensively extracts explanatory information directly related to the decision results from the explainable AI model. This includes: the contribution of each process feature to the decision outcome (clarifying which features are key to the decision), the sensitivity thresholds of process parameters (such as the reasonable range of temperature and the optimal ratio of components), the process rules and metallurgical mechanisms followed in the decision-making process, and the logic behind the impact of different parameter adjustments on the final quality. This explanatory information is not abstract data analysis results, but rather engineering language closely integrated with actual casting production. It directly corresponds to on-site operations, allowing process engineers to clearly understand "why this optimization is done" and "what effect the optimization will achieve." This completely solves the pain points of traditional intelligent models—"uncontrollable decisions and difficult-to-verify results"—and provides strong support for the implementation of process optimization solutions.

[0082] 6. Structured output of interpretation results

[0083] The structured output of the explanation results is the final step in interpretable AI modeling and optimization decision-making. The core is to standardize and systematically organize the extracted decision explanation information into a structured report that fits the engineering reality and is easy for on-site personnel to use. First, the extracted explanation information is categorized according to "process steps, feature types, and decision basis" to ensure the information is logically structured. Second, the explanation information is mapped one-to-one with specific process parameters, optimization schemes, and quality objectives, clearly defining the explanation basis, influencing factors, and expected effects for each optimization decision, making the explanation information highly readable in engineering. Finally, in accordance with the patent specification and engineering application standards, the structured explanation information is organized into a standardized format. This provides directly usable data support for the subsequent decision visualization module, offers clear references for process engineers to verify the rationality of decisions and adjust process parameters, and provides traceable evidence for subsequent model iteration optimization and continuous process improvement, forming a complete closed loop of "modeling-decision-explanation-implementation."

[0084] The above steps build upon the process feature set output from the domain knowledge-driven process feature modeling stage, completing process rule mining, intelligent model construction, process optimization deduction, and decision interpretation information generation. As a key central hub for intelligent analysis and solution output in the overall technical process, it naturally connects to subsequent related execution and feedback steps. This step outputs two core results: first, the casting process optimization parameter scheme obtained through solving multiple constraints; and second, structured decision interpretation data matching the process features and production mechanisms. The optimized process scheme directly provides clear operational basis and parameter standards for downstream steps such as process execution, on-site production control, and process effect detection. Simultaneously generated and structured interpretation information can be directly used by subsequent visualization, process mechanism analysis, and human-machine collaborative verification modules, achieving an intuitive presentation of decision logic, influencing mechanisms, and optimization basis. Meanwhile, the model operation data, feature influence patterns, and process optimization effect feedback information accumulated in this step can also provide data support and theoretical reference for subsequent model iteration and updates, process rule improvement, and long-term process closed-loop optimization, thus opening up a complete link between intelligent modeling, optimization decision-making, field application, and dynamic iteration, ensuring logical coherence, data interoperability, and close functional connection between each step.

[0085] Step S104: Transform the abstract intelligent decision-making results into an engineering expression that fits the actual production situation, complete the content visualization display by relying on diverse charts, and combine it with a multi-scheme comparison and judgment mode to reflect the relationship between parameter control and casting quality. Output standardized decision explanation information after engineering transformation, multi-type visualization analysis charts, and multi-dimensional optimization scheme comparison analysis results.

[0086] The above steps build upon the output of interpretable AI modeling and optimization decision-making, focusing on the practical application needs of the foundry industry. Using intelligent decision-making results and structured interpretable data as input, they conduct a full-process processing encompassing process interpretation, mechanism analysis, intuitive visualization, and auxiliary judgment. This breaks down the information barriers between intelligent models and engineering applications, transforming abstract model computational logic into engineering content easily understood by process engineers, providing comprehensive support for process scheme verification, on-site parameter adjustment, and production process improvement. The following sections describe each sub-step in detail.

[0087] 1. Decision Result Analysis

[0088] This sub-step, as the primary foundational step in the aforementioned process, involves in-depth decomposition and systematic analysis of the complete set of process optimization results, which can be interpreted through AI modeling and optimization decision-making. It comprehensively reviews the overall optimization scheme generated by the model, accurately extracting key decision information highly correlated with core objectives such as casting quality improvement, production stability control, and process energy consumption optimization. It precisely identifies the specific adjustment direction, reasonable adjustment range, and parameter optimization boundary range for each process parameter, and, combined with the casting production mechanism, predicts the improvement effects and expected benefits in aspects such as casting internal structure, forming quality, and defect control after parameter adjustments. This completes the core framework of the optimization decision-making process, providing complete, standardized, and effective basic data support for subsequent multi-dimensional interpretation, analysis, and visualization.

[0089] 2. Explanation of Key Process Characteristics

[0090] This sub-step relies on the feature attribution and decision explanation information synchronously output by the interpretable artificial intelligence model, combined with the process feature system constructed from prior domain knowledge, to accurately screen and identify core process feature variables that play a dominant role and have high influence weight in the process optimization decision generation process. Combining the casting metallurgical mechanism, solidification and cooling laws, and multi-parameter coupling relationships, it deeply analyzes the positive or negative promoting effects of various key process features on casting quality indicators and overall optimization goals, quantifying the strength and proportion of the influence of different process features. It clearly elucidates the actual mechanism of each process feature in parameter matching, operating condition control, and quality management, clarifies the inherent logic of different process elements participating in intelligent decision-making, and achieves explicit interpretation of the model's implicit decision-making logic.

[0091] 3. Parameter sensitivity and impact analysis

[0092] This sub-step focuses on conducting refined sensitivity quantitative analysis of the core process parameters throughout the casting process. Combining multi-condition simulation data and model training patterns, it systematically characterizes the dynamic response relationships and changing patterns of various process parameters within reasonable ranges to quality indicators, optimization objectives, and production conditions. Through quantitative analysis, it accurately distinguishes between highly sensitive core parameters, generally influential parameters, and weakly correlated auxiliary parameters, precisely identifying key control parameters that are highly likely to cause casting defects and quality fluctuations. Simultaneously, it delineates the safe operating range, optimal control range, and critical warning range for each core process parameter, helping process engineers clearly understand the parameter control boundaries and providing a quantitative basis for precise on-site process adjustments and early risk prevention.

[0093] 4. Explain information engineering mapping

[0094] This sub-step focuses on the problems of weak universality and high professional barriers in model interpretation information, and carries out standardized engineering transformation and precise mapping processing. The abstract data indicators, mathematical relationships, and feature operation logic output by the interpretable artificial intelligence model are mapped one by one to various specific process parameters, melting, holding, and pouring stages of the entire casting process. Simultaneously, content calibration is completed by incorporating engineering constraints such as equipment operating limitations, production operation specifications, and metallurgical reaction conditions. Purely data-driven and model-based professional expressions are transformed into standardized engineering language that fits the workshop production and process control scenarios, abandoning obscure algorithmic logic descriptions. This ensures that the interpretation content aligns with on-site operational logic and industry technical understanding, comprehensively improving the practicality and applicability of the interpretation information.

[0095] 5. Visual presentation of decision results

[0096] This sub-step focuses on visualization and concretization, unifying and visually reconstructing process optimization schemes, key feature contribution patterns, parameter influence relationships, and decision explanations. Utilizing diverse visualization formats, it constructs visual charts such as parameter change trend graphs, key process feature contribution distribution graphs, and multi-dimensional process influence relationship diagrams. These charts intuitively display the numerical changes before and after process parameter adjustments, the weight distribution of different features, and the intrinsic relationship between parameter linkage changes and casting quality. Through graphical and modular presentation, it simplifies complex process logic, making cumbersome data analysis results clear and easy to understand, achieving an integrated and intuitive display of optimization decisions and mechanism explanations.

[0097] 6. Decision Support and Comparative Analysis

[0098] This sub-step possesses the capability for comprehensive evaluation and differential analysis of multiple solutions. It can simultaneously load multiple process optimization solutions generated under different constraints and optimization objectives, supporting process engineers in conducting horizontal comparisons and comprehensive evaluations. The system systematically analyzes the differentiated characteristics of each optimization solution in dimensions such as core process parameter configuration, parameter adjustment range, quality improvement potential, production energy consumption level, and equipment load intensity. It comprehensively analyzes the process implementation difficulties, production operation risks, and adaptability of different solutions. Through multi-dimensional comparison of advantages and disadvantages and risk warnings, it assists process engineers in scientifically selecting the most suitable process optimization solution based on actual workshop capacity, equipment conditions, production costs, and other realistic conditions. This further enhances the engineering practical value of intelligent decision-making and ensures the smooth implementation of process optimization measures.

[0099] Step S105: Deeply integrate the intelligent decision-making results, process personnel experience, and on-site production verification data. The process personnel review the process optimization plan, and then put the approved plan into actual casting production and collect relevant data to form a structured feedback dataset. After effectiveness evaluation, the feedback data is used to update and adaptively optimize the interpretable artificial intelligence model to build a human-machine collaborative closed loop.

[0100] The above steps, as the closed-loop core of the entire intelligent optimization of the casting process, closely follow the standardized decision interpretation information, multi-type visualization analysis charts, and multi-dimensional optimization scheme comparison analysis results output from the decision interpretation and visualization steps, which have undergone engineering transformation. With "human-machine collaboration, data feedback, and model iteration" as the core logic, through six coherent and progressive sub-steps, it achieves deep integration of the interpretable artificial intelligence model with on-site production practice, and promotes continuous improvement in process optimization capabilities. A detailed description is as follows:

[0101] 1. Process personnel intervention and confirmation

[0102] This step is the first step in human-machine collaboration. Its core is to leverage the on-site experience and professional capabilities of process engineers to comprehensively review, confirm, and make necessary adjustments to the process optimization solutions output by the decision interpretation and visualization modules. Process engineers, combining visual analysis charts and standardized interpretation information, compare the optimization solutions with the operating limits of on-site production equipment, actual production conditions, safety operating procedures, and raw material characteristics. They verify the rationality and feasibility of each process parameter in the optimization solution, focusing on whether the parameter adjustment range and control logic conform to the casting metallurgy mechanism and actual workshop production. Solutions that meet production conditions are directly approved, and production execution instructions are issued. For unreasonable or impractical parameter configurations, process engineers make targeted adjustments based on their engineering experience, clarifying the basis for the adjustments and the adjusted parameters. This ensures that the optimization solution not only aligns with the intelligent model's decision-making logic but also meets on-site production constraints and safety requirements, laying the foundation for subsequent production execution.

[0103] 2. Production Execution and On-site Validation

[0104] This step is the core process for implementing and verifying the actual effects of the manually confirmed (or adjusted) process optimization plan. The approved process optimization plan is synchronized to the entire casting production process, clarifying the parameter execution standards and operational requirements for each process stage, and guiding on-site operators to strictly follow the optimization plan in production operations such as melting, pouring, solidification, and cooling. During production execution, full-process production data is collected simultaneously, including real-time process parameter data, equipment operating status data, raw material consumption data, and casting quality inspection data after production (such as mechanical properties, surface quality, and defect status). This comprehensively records the actual application effects of the optimization plan, forming a complete production verification data chain, providing real and reliable on-site data support for subsequent feedback data collection and model optimization.

[0105] 3. Feedback Data Collection and Processing

[0106] This step focuses on collecting, classifying, and standardizing all types of data generated during production verification to create a structured feedback dataset. The collection scope covers three core data areas: first, process parameter data, including actual parameter values, fluctuations, and adjustment records during the optimization process; second, quality inspection data, including test results of various quality indicators of castings, defect types, and severity; and third, data on process personnel intervention, including manually adjusted parameters, reasons for adjustments, adjustment time, and post-adjustment effects. The collected raw data undergoes noise reduction, redundancy removal, and missing value completion. It is then categorized and archived according to casting process stages and data types, transforming it into a standardized and unified structured data format. This ensures the completeness, consistency, and usability of the feedback data, providing a standardized data carrier for subsequent evaluation of the feedback's effectiveness.

[0107] 4. Evaluation of the effectiveness of feedback information

[0108] The core of this step is to assess the effectiveness of the structured feedback dataset, eliminating invalid data and identifying influencing factors to ensure the reliability and relevance of the data used for model updates. Combining the casting process mechanism and actual production conditions, the feedback data undergoes a multi-dimensional evaluation: First, the authenticity of the data is determined, identifying abnormal and false data to ensure it accurately reflects the actual application effect of the optimization scheme; second, influencing factors are identified, clarifying whether the quality fluctuations and parameter deviations reflected in the feedback data stem from explainable AI model decision failures, or from changes in on-site process conditions (such as batch differences in raw materials, equipment wear), human error, or other factors; third, valid data is selected, retaining core data that accurately reflects the model's optimization effect and is valuable for model iteration, while eliminating irrelevant and meaningless redundant data. This provides high-quality, targeted feedback support for subsequent model updates, preventing invalid data from causing model optimization deviations.

[0109] 5. Model Update and Adaptive Optimization

[0110] This step is the core of continuous model improvement. Based on reliable feedback data after effectiveness evaluation, the previously constructed interpretable AI model is dynamically updated and adaptively retrained. Combining the process optimization effects, parameter deviations, and changes in operating conditions reflected in the feedback data, the model's hyperparameter settings and feature weight allocation are adjusted to optimize the mapping relationship between process features and casting quality, correcting any mismatches between the model and actual production conditions. Simultaneously, new patterns and experiences from the feedback data are integrated into the model training process, enabling the model to gradually adapt to changes in on-site process conditions, fluctuations in the production environment, and differences in raw material characteristics. This continuously improves the model's prediction accuracy, optimization capabilities, and environmental adaptability, ensuring that the optimized solutions output by the model always align with actual on-site production, thus achieving dynamic improvement in process optimization capabilities.

[0111] 6. Human-machine collaborative closed-loop iteration

[0112] This step is the final stage in building a complete closed-loop optimization mechanism. It involves repeatedly executing the entire process of "process personnel intervention and confirmation—production execution and on-site verification—feedback data collection and processing—feedback information effectiveness evaluation—model update and adaptive optimization," forming a virtuous cycle of human-machine collaboration and continuous iteration. After each iteration, the new optimization scheme output by the updated, interpretable AI model is submitted again to process personnel for review and confirmation, repeating the subsequent production verification and feedback process to achieve a closed-loop operation of "model optimization—on-site verification—data feedback—model iteration." This closed-loop mechanism deeply integrates the engineering experience of process personnel with the data analysis capabilities of the intelligent model, compensating for the limitations of pure model-based decision-making and the subjectivity of purely human experience, and continuously improving the stability, applicability, and optimization efficiency of the casting process optimization system in complex and variable production environments.

[0113] In summary, this intelligent optimization method for casting processes, guided by domain knowledge, centered on interpretable intelligence, and supported by a human-machine collaborative closed loop, constructs a complete intelligent optimization system from data input to model iteration, from decision output to on-site implementation. The entire process revolves around the precision of process optimization, interpretability of decisions, and feasibility of applications, with each step seamlessly connected and forming a logical closed loop. First, through domain knowledge-driven process feature modeling, casting production-related data is transformed into process feature variables with clear physical meaning and engineering orientation, providing standardized input for subsequent intelligent modeling. Then, relying on interpretable AI modeling and optimization decision-making steps, an interpretable model adapted to casting conditions is constructed to uncover the correlation between process features and casting quality, generating scientifically feasible process optimization solutions, and simultaneously extracting structured decision interpretation information. Next, through decision interpretation and visualization steps, abstract intelligent decisions are transformed into engineering representations and visual results, assisting process engineers in understanding the decision logic and conducting solution verification. Finally, through a human-machine collaborative closed-loop feedback step, process engineer experience, on-site production verification data, and model decision results are integrated. Through data feedback and effectiveness evaluation, the model is dynamically updated, building a virtuous cycle of human-machine collaboration through iterative iteration. The entire method connects the entire chain of "feature modeling - intelligent decision-making - visual interpretation - on-site implementation - model iteration", taking into account both data-driven and process mechanism, and solving the pain points of traditional intelligent models being difficult to implement as "black boxes" and having low efficiency in manual optimization. It realizes intelligent and precise optimization of casting process, continuously improves the stability and applicability of the system in complex production environment, and provides comprehensive technical support for improving casting production quality and optimizing efficiency.

[0114] Figure 2 This is a schematic diagram of the composition of a casting process optimization system based on explainable artificial intelligence. The system is used to implement the aforementioned casting process optimization method and includes a data acquisition and preprocessing module, a domain knowledge-driven process feature modeling module, an explainable artificial intelligence modeling and optimization decision-making module, a decision interpretation and visualization module, and a human-machine collaborative closed-loop feedback module.

[0115] The data acquisition and preprocessing module is used to acquire multi-source heterogeneous data, including process parameters, equipment operating status, and casting quality inspection results, throughout the entire casting production process, and to perform standardized preprocessing to form standardized data for process feature modeling.

[0116] The domain knowledge-driven process feature modeling module is used to perform process stage division and parameter classification, key process parameter identification, process feature reconstruction and mapping, multi-parameter coupled feature construction, process constraint and rule embedding, and feature validity verification and updating in sequence, based on engineering knowledge in the casting field, in order to achieve deep integration of engineering knowledge and standardized data from multiple sources of processes, forming a set of process features.

[0117] An interpretable AI modeling and optimization decision-making module is used to explore the correlation between process characteristics and quality. It combines process constraints and engineering logic to build an intelligent decision-making model for casting process optimization. Through a traceable interpretation mechanism, it generates intelligent decision results and structured interpretation data, providing interpretation information corresponding to process characteristics to support engineering understanding and verification.

[0118] The decision interpretation and visualization module is used to transform abstract intelligent decision results into engineering expressions that fit production realities. It uses diverse charts to visualize the content and combines them with a multi-scheme comparison and analysis mode to reflect the relationship between parameter control and casting quality. It outputs standardized decision interpretation information that has been transformed into engineering, multiple types of visual analysis charts, and multi-dimensional optimization scheme comparison and analysis results.

[0119] The human-machine collaborative closed-loop feedback module is used to deeply integrate intelligent decision-making results, process personnel experience, and on-site production verification data. Process personnel review the process optimization plan, and then the approved plan is put into actual casting production and relevant data is collected to form a structured feedback dataset. After effectiveness evaluation, the feedback data is used to update and adaptively optimize the interpretable artificial intelligence model, thus constructing a human-machine collaborative closed loop.

[0120] The specific implementation functions of each of the above modules have been introduced in the aforementioned method section and will not be repeated here.

[0121] The technical solution of this invention will be described in its entirety below through a casting process example of a ductile iron crankshaft. This process example uses a typical automotive engine ductile iron crankshaft as a casting, focusing on engineering problems such as shrinkage cavities, insufficient pearlite content, fluctuations in mechanical properties, and low casting yield. It follows five steps sequentially: data acquisition and preprocessing, domain knowledge-driven process feature modeling, interpretable AI modeling and optimization decision-making, decision interpretation and visualization, and human-machine collaborative closed-loop feedback. This fully demonstrates the implementation process of the entire technical solution, with each module acting as an independent entity to complete data processing, feature construction, intelligent decision-making, engineering transformation, and closed-loop iteration.

[0122] 1. Data Acquisition and Preprocessing (Step S101)

[0123] Execution subject: Data acquisition and preprocessing module (hereinafter referred to as the module).

[0124] Core objective: To acquire multi-source data from the entire casting process, complete standardized governance, and generate high-quality, standardized data that can be used for feature modeling.

[0125] This module serves as the initial foundation of the method, encompassing the entire process of ductile iron crankshaft smelting, spheroidizing inoculation, casting, solidification and cooling, and post-processing, and involves comprehensive data acquisition and systematic preprocessing.

[0126] (1) Multi-source heterogeneous data acquisition

[0127] The module automatically collects three types of core data from the production site: first, process parameter data, including melting temperature, holding temperature, pouring temperature, pouring time, spheroidizing agent dosage, inoculant dosage, cooling rate, mold temperature, and mold permeability; second, equipment operating status data, including medium-frequency furnace output power, operating load, pouring machine travel speed, cooling fan frequency, equipment start-up and shutdown sequence, and operational stability data; and third, casting quality inspection result data, including pearlite content, tensile strength, elongation, hardness, internal flaw detection defect level, surface quality, and dimensional accuracy data. The module supports real-time sensor acquisition, PLC data reading, direct connection to online inspection equipment, and manual input of inspection results, ensuring the completeness of data sequence and the authenticity and traceability of data sources within the same furnace, crankshaft, and process cycle.

[0128] (2) Standardized data preprocessing

[0129] The module performs five standardization processes on the raw data in sequence: data cleaning removes duplicates, garbled characters, and invalid records that violate process logic; outlier handling identifies and corrects extreme points caused by sensor drift and transient interference; missing value completion uses the average value of the same batch and process rules to restore data continuity; time alignment unifies the time axis according to furnace number, casting number, and process stage to achieve accurate matching of parameters, equipment, and quality data; and data standardization maps data of different dimensions such as temperature, time, composition ratio, and mechanical properties to a unified range.

[0130] Module output: time-aligned, noise-free, missing data with a consistent format.

[0131] 2. Domain knowledge-driven process feature modeling (step S102)

[0132] Execution entity: Domain knowledge-driven process feature modeling module (hereinafter referred to as the module).

[0133] Core objective: Guided by the metallurgical mechanism of casting, transform standardized data into a set of process features with clear physical meaning, engineering orientation, and consistent mechanism.

[0134] This module takes the standardized data output from the data acquisition and preprocessing module as its sole input. Throughout the process, it is constrained by the solidification law of ductile iron, the graphite spheroidization mechanism, the heat and mass transfer characteristics, and the shrinkage cavity formation mechanism. It does not perform pure data calculations that are divorced from engineering logic, ensuring that all features are engineering interpretable.

[0135] (1) Process stage division and parameter classification

[0136] The module divides the entire process into five stages according to the actual crankshaft production process: melting, spheroidizing and inoculation, casting, solidification and cooling, and post-treatment. All parameters are automatically categorized according to their respective processes.

[0137] (2) Identification of key process parameters

[0138] The module combines domain knowledge and correlation analysis to automatically identify key parameters that play a decisive role in crankshaft quality: pouring temperature, cooling rate, amount of spheroidizing agent added, inoculation amount, and mold temperature.

[0139] (3) Reconstruction and mapping of process features

[0140] The module transforms discrete fundamental parameters into features that characterize physical behavior, such as superheat, cooling intensity, composition-temperature matching degree, and spheroidization sufficiency, so that the features directly correspond to metallurgical mechanisms and quality formation laws.

[0141] (4) Construction of multi-parameter coupling features

[0142] The module explores the interactions between parameters, constructing temperature-composition coupling characteristics, cooling rate-casting wall thickness coupling characteristics, and pouring speed-filling stability coupling characteristics to comprehensively reflect the synergistic effects of multiple parameters.

[0143] (5) Process constraints and rule embedding

[0144] The module embeds casting process specifications, equipment operating limits, and production safety conditions into the process. For example, the pouring temperature is limited to 1380–1420℃, and features outside this range are automatically deemed invalid.

[0145] (6) Feature validity verification and updating

[0146] The module uses quality inspection data to quantitatively verify the features, retaining features that contribute significantly to shrinkage cavities, pearlite content, and mechanical properties and have strong stability, while eliminating features that are weakly correlated or highly volatile.

[0147] Module output: A set of process features that can be directly input into an interpretable artificial intelligence model.

[0148] 3. Explainable AI modeling and optimization decision-making (step S103)

[0149] Execution Entity: Interpretable Artificial Intelligence Modeling and Optimization Decision Module (hereinafter referred to as the Module).

[0150] Core objective: To build an interpretable model, explore the process-quality mapping relationship, and generate feasible and interpretable process optimization solutions and structured interpretable data.

[0151] This module takes the set of process features as input and serves as the intelligent decision-making center of the entire method. It rejects black-box models throughout the process, ensuring that decisions are traceable, verifiable, and interpretable in an engineering manner.

[0152] (1) Model input construction

[0153] The module further filters and standardizes the set of process features, retains core features with high contribution, completes format unification and normalization, and eliminates the impact of magnitude differences on the model.

[0154] (2) Explainable selection and construction of artificial intelligence models

[0155] The module adopts an inherently interpretable model architecture, embedding the metallurgical mechanism, process constraints, and equipment limits of ductile iron into the model structure to ensure that the model's operational logic is highly consistent with physical laws.

[0156] (3) Modeling of process influence relationships

[0157] The module establishes a quantitative mapping between process characteristics and quality indicators through model learning: increased casting temperature significantly increases the risk of shrinkage cavities and porosity; increased cooling rate can improve pearlite content and mechanical properties; insufficient spheroidizing agent directly leads to poor spheroidization and decreased strength; and the module also quantifies the synergistic and inhibitory relationships between parameters.

[0158] (4) Process optimization decision generation

[0159] Under the condition of meeting all process constraints and equipment limits, the module generates the optimal process scheme for crankshaft through multi-dimensional iterative deduction: casting temperature 1400℃, cooling rate 0.8℃ / s, spheroidizing agent addition 0.9%, and inoculant addition 0.6%. The scheme includes single parameter fine-tuning and multi-parameter collaborative matching, and clarifies the adjustment direction, magnitude and expected quality improvement.

[0160] (5) Extraction of decision interpretation information

[0161] The module synchronously extracts the decision-making basis: casting temperature contributes 42%, cooling rate contributes 28%, and spheroidizing agent contributes 18%, while outputting the parameter sensitivity range, influence trend, and defect suppression mechanism.

[0162] (6) Structured output of interpretation results

[0163] The module standardizes and organizes the explanatory information into structured content that corresponds one-to-one with process parameters, process stages, and quality objectives, which can be directly used for subsequent visualization and engineering verification.

[0164] Module outputs: process optimization schemes and structured interpretation data.

[0165] 4. Decision Interpretation and Visualization (Step S104)

[0166] Implementing entity: Decision interpretation and visualization module (hereinafter referred to as the module).

[0167] Core objective: To transform abstract intelligent decision-making into engineering language, and to present the decision-making logic intuitively through visualization, so as to assist process personnel in reviewing, verifying and selecting solutions.

[0168] This module closely follows the output of the previous module, transforming model decisions from algorithmic language into engineering information that process engineers can directly understand, thus breaking down the barriers between intelligent models and field applications.

[0169] (1) Analysis of decision results

[0170] The module breaks down the optimization scheme in depth, specifying that the casting temperature is reduced by 15℃, the cooling rate is increased by 0.2℃ / s, and the spheroidizing agent is finely adjusted by 0.1%, and gives the expected reduction in defects and the range of performance improvement.

[0171] (2) Explanation of key process characteristics

[0172] The characteristics that play a leading role in module identification decision-making are identified, and it is found that excessively high casting temperature is the primary factor for crankshaft shrinkage and insufficient cooling is the main reason for substandard pearlite. The influencing mechanism is explained using metallurgical principles.

[0173] (3) Parameter sensitivity and impact analysis

[0174] The module performs quantitative analysis of key parameters, determining that a casting temperature of ±5℃ significantly changes the probability of shrinkage cavities, and that a cooling rate of ±0.1℃ / s significantly affects the pearlite content, thus defining safe, optimal, and warning ranges.

[0175] (4) Explaining information engineering mapping

[0176] The module converts the abstract indicators of the model into process language, such as "excessive temperature → increased liquid shrinkage → insufficient solidification and feeding → easy to produce shrinkage cavities", so that the explanation content fits the on-site operation logic.

[0177] (5) Visual presentation of decision results

[0178] The module generates a bar chart of feature contribution, a temperature-defect trend chart, a cooling rate-pearlite relationship curve, and a parameter optimization comparison diagram, which intuitively demonstrate the relationship between parameter adjustment and quality change.

[0179] (6) Decision support and comparative analysis

[0180] The module provides a comparison of multiple feasible solutions, showing the differences between the solutions in terms of parameter configuration, expected pass rate, implementation risk, and equipment compatibility, to help process engineers quickly select the optimal solution.

[0181] Module outputs include: engineering decision explanation information, visual analysis charts, and comparative analysis results of multi-dimensional optimization schemes.

[0182] 5. Human-machine collaborative closed-loop feedback (step S105)

[0183] Execution entity: Human-machine collaborative closed-loop feedback module (hereinafter referred to as the module).

[0184] Core objective: To integrate model-based decision-making, process personnel experience, and field validation data to achieve dynamic model updates and adaptive iterations, and to build a continuous optimization closed loop.

[0185] This module serves as the closed-loop core of the entire system, enabling real-time linkage between the intelligent model and the production site, allowing the system to adapt to fluctuations in operating conditions, changes in raw materials, and equipment aging.

[0186] (1) Intervention and confirmation by process personnel

[0187] Process engineers review optimization plans based on visual charts and engineering explanations, confirming that the parameters are within safe ranges and comply with equipment capabilities and production specifications, and can issue execution instructions without adjustment.

[0188] (2) Production execution and on-site verification

[0189] The module distributes the confirmed solution to the production execution unit, and the site continuously produces according to the optimized process, while simultaneously collecting actual operating parameters, equipment status, casting flaw detection results and mechanical property data to form a complete verification data chain.

[0190] (3) Feedback data collection and processing

[0191] The module automatically collects process execution data, quality inspection data, and process personnel adjustment records, and after cleaning, alignment, and normalization, forms a structured feedback dataset.

[0192] (4) Evaluation of the effectiveness of feedback information

[0193] The module assesses the authenticity and validity of the data, distinguishing whether the quality improvement stems from process optimization, changes in raw materials, or fluctuations in equipment status, thus ensuring the reliability of the data used for model updates.

[0194] (5) Model update and adaptive optimization

[0195] The module uses effective feedback data to retrain the interpretable artificial intelligence model, corrects the process-quality mapping relationship, adjusts feature weights, and improves the model's prediction accuracy and optimization capabilities under the current working conditions.

[0196] (6) Human-machine collaborative closed-loop iteration

[0197] The module iteratively executes the processes of manual review, on-site verification, data feedback, and model update, forming a continuous optimization loop that enables the system to adapt to changes in the production environment and maintain optimal optimization results.

[0198] Final module results: shrinkage defects reduced by 72%, pearlite compliance rate reached 100%, casting qualification rate increased from 85% to 97.5%, and system stability and adaptability to working conditions were significantly improved.

[0199] In summary, this complete example of ductile iron crankshaft casting process uses five modules as independent execution entities to sequentially complete data governance → feature construction → intelligent decision-making → engineering transformation → closed-loop iteration. The entire process embodies the core workflow of "domain knowledge-driven, explainable AI decision-making, engineering implementation, and human-machine collaborative iteration". It not only achieves defect suppression and quality improvement, but also ensures that each optimization step has a clear mechanism, intuitive presentation, and verifiable basis, making it fully applicable to the industrial implementation of complex casting scenarios.

[0200] The present invention provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described casting process optimization method based on interpretable artificial intelligence.

[0201] The present invention may also provide a storage medium, which may be a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the above-described casting process optimization method based on interpretable artificial intelligence.

[0202] The computer-readable storage medium provided by this invention may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0203] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A casting process optimization method based on interpretable artificial intelligence, characterized in that, The method includes the following steps: Data is collected from multiple sources and heterogeneous data throughout the entire casting production process, including process parameters, equipment operating status, and casting quality inspection results. The data is then standardized and preprocessed to form standardized data for process feature modeling. Based on engineering knowledge in the casting field, the process is divided into stages and classified into parameters, key process parameters are identified, process features are reconstructed and mapped, multi-parameter coupled features are constructed, process constraints and rules are embedded, and feature validity is verified and updated in sequence. This process achieves deep integration of engineering knowledge and standardized data from multiple sources of processes to form a set of process features. Based on interpretable artificial intelligence models, the correlation between process characteristics and quality is explored. By combining process constraints and engineering logic, an intelligent decision-making model for casting process optimization is constructed. Through a traceable explanation mechanism, intelligent decision results and structured explanation data are generated, providing explanation information corresponding to process characteristics to support engineering understanding and verification. The abstract intelligent decision-making results are transformed into engineering expressions that fit the actual production. The content is visualized by relying on diverse charts and graphs. Combined with the multi-scheme comparison and judgment mode, it reflects the relationship between parameter control and casting quality. The output is standardized decision explanation information after engineering transformation, multi-type visualization analysis charts and multi-dimensional optimization scheme comparison analysis results. By deeply integrating intelligent decision-making results, process personnel experience, and on-site production verification data, process personnel review the process optimization plan, and then the approved plan is put into actual casting production and relevant data is collected to form a structured feedback dataset. After effectiveness evaluation, the feedback data is used to update and adaptively optimize the interpretable artificial intelligence model, thus building a human-machine collaborative closed loop.

2. The casting process optimization method based on interpretable artificial intelligence according to claim 1, characterized in that, Standardized preprocessing includes five preprocessing steps performed sequentially: data cleaning, outlier handling, missing value completion, time alignment, and data standardization.

3. The casting process optimization method based on interpretable artificial intelligence according to claim 2, characterized in that, Data cleaning is used to remove invalid, redundant, erroneous, and meaningless information from raw data; outlier handling is used to identify, correct, or remove extreme data that deviates from the normal production range; missing value completion is used to restore missing data caused by data acquisition interruptions, communication failures, or human error; time alignment is used to unify the time base of multi-source data to achieve accurate matching of process parameters, equipment status, and quality data; and data standardization is used to eliminate differences in dimensions and numerical magnitudes between different parameters.

4. The casting process optimization method based on interpretable artificial intelligence according to claim 1, characterized in that, The process stage division and parameter classification specifically include: based on the actual process flow and physical change law of casting production, the complete casting production process is divided into five continuous processes: melting stage, heat preservation stage, pouring stage, solidification and cooling stage, and post-processing stage. The process is then accurately classified according to the actual production links, so that each standardized data corresponds to a specific process stage, realizing one-to-one binding of data with process behavior and physical process.

5. The casting process optimization method based on interpretable artificial intelligence according to claim 1, characterized in that, The identification of key process parameters specifically includes: combining casting metallurgy theory, solidification phase transformation mechanism, defect formation law and field engineering experience, analyzing the degree of influence and significance of all parameters contained in the standardized data in each stage, and screening out key process parameters from them.

6. The casting process optimization method based on interpretable artificial intelligence according to claim 5, characterized in that, The key process parameters include variables of temperature, time, component ratio, and cooling conditions; non-key parameters are uniformly marked as auxiliary variables or background variables.

7. The casting process optimization method based on interpretable artificial intelligence according to claim 1, characterized in that, The process feature reconstruction and mapping specifically includes: carrying out feature reconstruction through parameter combination, proportional relationship construction, stage statistics, interval mapping, and trend extraction, upgrading standardized data from basic numerical form to process feature variables that can characterize the physical behavior of the casting process.

8. The casting process optimization method based on interpretable artificial intelligence according to claim 1, characterized in that, The construction of the multi-parameter coupling feature specifically includes: based on standardized data of multiple key process parameters, mining the joint distribution patterns, synergistic change trends and interaction relationships contained in the data, and constructing coupling feature variables that reflect the combined effect of multiple parameters.

9. The casting process optimization method based on interpretable artificial intelligence according to claim 1, characterized in that, The process constraints and rule embedding specifically include: embedding process limitations, equipment operating limits, production safety regulations, quality control standards, and mature engineering experience rules in the casting field into the feature construction process in the form of constraints, and directly applying them to the standardized data on which the generated features are based.

10. A casting process optimization system based on interpretable artificial intelligence, used to implement the casting process optimization method based on interpretable artificial intelligence as described in any one of claims 1 to 9, characterized in that, The system includes a data acquisition and preprocessing module, a domain knowledge-driven process feature modeling module, an interpretable artificial intelligence modeling and optimization decision-making module, a decision interpretation and visualization module, and a human-machine collaborative closed-loop feedback module. The data acquisition and preprocessing module is used to acquire multi-source heterogeneous data, including process parameters, equipment operating status, and casting quality inspection results, throughout the entire casting production process, and to perform standardized preprocessing to form standardized data for process feature modeling. The domain knowledge-driven process feature modeling module is used to perform process stage division and parameter classification, key process parameter identification, process feature reconstruction and mapping, multi-parameter coupled feature construction, process constraint and rule embedding, and feature validity verification and updating in sequence, based on engineering knowledge in the casting field, in order to achieve deep integration of engineering knowledge and standardized data from multiple sources of processes to form a set of process features. An interpretable AI modeling and optimization decision-making module is used to explore the correlation between process characteristics and quality. It combines process constraints and engineering logic to build an intelligent decision-making model for casting process optimization. Through a traceable interpretation mechanism, it forms intelligent decision results and structured interpretation data, providing interpretation information corresponding to process characteristics to support engineering understanding and verification. The decision interpretation and visualization module is used to transform abstract intelligent decision results into engineering expressions that fit the actual production. It uses a variety of charts to complete the content visualization display, and is equipped with a multi-scheme comparison and judgment mode to reflect the relationship between parameter control and casting quality. It outputs standardized decision interpretation information after engineering transformation, multiple types of visualization analysis charts, and multi-dimensional optimization scheme comparison analysis results. The human-machine collaborative closed-loop feedback module is used to deeply integrate intelligent decision-making results, process personnel experience, and on-site production verification data. Process personnel review the process optimization plan, and then the approved plan is put into actual casting production and relevant data is collected to form a structured feedback dataset. After effectiveness evaluation, the feedback data is used to update and adaptively optimize the interpretable artificial intelligence model, thus constructing a human-machine collaborative closed loop.

Images