A ductile cast iron whole-process dynamic regulation method

Abstract

The present application relates to the technical field of nodular cast iron, and particularly relates to a nodular cast iron full-process dynamic regulation method, which comprises collecting real-time data of temperature, pressure, weight and element concentration in the steps of batching, smelting and pouring to form a unified data set; smelting temperature and composition uniformity data are fused by using a data fusion algorithm to obtain a comprehensive data set; after meeting the integrity threshold, a parameter correlation graph is generated by association rule mining; key quality influencing factors are identified based on a decision tree algorithm; real-time feedback data and the correlation graph are compared to generate an adjustment instruction sequence; the instruction is executed and the production parameters are optimized through a feedback loop; product stability indicators are monitored, and when deviating from the preset range, the key influencing factors are updated; difference information is iteratively processed to obtain a final dynamic regulation model. The present application realizes full-process data collaboration and dynamic closed-loop regulation, improves product quality stability and production intelligence level, and is suitable for high-end nodular cast iron continuous production scenarios.

Description

Technical Field

[0001] This invention relates to the field of ductile iron technology, specifically to a method for dynamic control of the entire ductile iron casting process. Background Technology

[0002] Against the backdrop of rapid development in high-end equipment manufacturing and infrastructure construction, ductile iron, as a key engineering material possessing high strength, good plasticity, and excellent casting properties, is widely used in automotive parts, piping systems, and heavy machinery. Its product quality directly affects service safety and lifespan, placing higher demands on the stability and consistency of the production process. Therefore, developing a method that enables collaborative optimization and dynamic control of parameters throughout the entire process is not only a core requirement for improving the manufacturing level of ductile iron but also a crucial support for enhancing industrial competitiveness.

[0003] Currently, the production process of ductile iron commonly employs monitoring methods targeting single processes, such as independently collecting and controlling smelting temperature or batch weight. However, in complex processes encompassing multiple stages such as batching, smelting, and casting, real-time data on temperature, pressure, and weight generated at each stage are typically recorded separately by equipment from different manufacturers, resulting in inconsistent formats and a lack of interoperability. This multi-source, heterogeneous data distribution limits the effective correlation of information across processes, making it difficult to form a unified representation of the overall process status. Furthermore, existing control strategies largely rely on operator experience or static setpoints, which have limited adaptability to dynamic disturbances such as composition fluctuations and thermal field changes, affecting the ability to continuously ensure key quality indicators.

[0004] The quality stability of ductile iron is closely related to the coupling relationship of multiple process parameters, among which the interaction between melting temperature and compositional uniformity is particularly critical. Temperature fluctuations during the melting stage can cause changes in the dissolution rate of alloying elements, thus affecting the consistency of the chemical composition of the molten iron; and compositional deviations will feed back into the casting process, altering solidification behavior and graphite spheroidization. Although some systems have attempted to introduce data analysis tools, in the absence of a unified data foundation, the inherent correlation patterns between parameters are difficult to identify by the system, resulting in insufficient basis for control. Furthermore, due to the large number of transient signals and nonlinear responses in the production process, the timeliness and accuracy of control commands may also be limited without establishing a closed-loop feedback and iterative optimization mechanism.

[0005] Especially in high-paced, continuous production scenarios, abnormal changes in certain key parameters may only manifest as short-term deviations, making it difficult for conventional monitoring methods to capture their evolutionary trends. Meanwhile, core quality characteristics such as compositional uniformity are influenced by multiple factors, and their patterns of change are easily masked by normal process fluctuations. This multi-variable, strongly coupled nature makes it difficult to proactively identify and intervene in quality risks across the entire process using only local data or static rules. Therefore, constructing an intelligent control system covering the entire chain of batching, smelting, and casting through unified data collection, integrated analysis, and dynamic feedback mechanisms has become a key technological direction for achieving high-quality and stable production of ductile iron. Summary of the Invention

[0006] The purpose of this invention is to address the aforementioned shortcomings in the prior art by providing a dynamic control method for the entire process of ductile iron casting.

[0007] The objective of this invention is achieved through the following technical solution: a method for dynamic control of the entire process of ductile iron casting, comprising the following steps: S1. Acquire real-time data streams of temperature, pressure, weight, and element concentration from the batching, smelting, and casting processes. Extract and transmit these data from distributed devices to the central data lake via the edge computing gateway acquisition interface to obtain a unified data set. S2. Based on the data set, a data fusion algorithm is used to process multi-source heterogeneous information, fuse data points related to melting temperature and composition uniformity, and determine the fused comprehensive dataset. S3. Based on the integrated dataset after fusion, determine whether it meets the integrity threshold. If the integrity threshold is met, analyze the association pattern between parameters through the association rule mining algorithm and generate an association relationship graph between parameters. S4. Based on the correlation diagram between the parameters, the decision tree algorithm is used to classify the quality influencing factors and determine the list of key influencing factors; S5. Based on the list of key influencing factors, extract the core parameters required for the control mechanism, obtain real-time feedback data under the current production coordination status, and determine whether there is a need for control by comparing the correlation diagram with the feedback data. If there is a need for control, generate an adjustment instruction sequence. S6. Based on the adjustment instruction sequence, the instructions are transmitted to the casting effect control module, and the component uniformity related parameters are updated using a feedback loop mechanism to obtain an optimized production parameter configuration. S7. Based on the optimized production parameter configuration, monitor the changing trend of product stability indicators. If the changing trend deviates from the preset range, backtrack to the association rule mining algorithm to re-analyze the parameter association and determine the corrected list of key influencing factors. S8. Based on the difference information between the revised list of key influencing factors and the initial data set, these differences are iteratively processed through a data fusion algorithm to obtain the final dynamic control model.

[0008] The present invention is further configured such that: the unified format data set includes timestamp alignment features, unit standardization specifications, and missing value imputation results; the fused comprehensive dataset includes a melting temperature-composition uniformity mapping matrix, multi-source data confidence weighting coefficients, and spatiotemporal alignment correction terms; the parameter correlation graph includes support-confidence joint threshold screening results, frequent itemset generation paths, and causal direction determination labels; the key influencing factor list includes influencing factor weight ranking, threshold triggering conditions, and operating condition context identifiers; the adjustment instruction sequence includes parameter adjustment range, execution priority, and target encoding; the optimized production parameter configuration includes casting rate setpoints, holding time windows, and alloy addition ratios; the product stability indicators include graphite spheroidization rate fluctuation range, tensile strength dispersion coefficient, and metallographic structure consistency index; and the dynamic control model includes multi-round iterative convergence criteria, differential feature embedding vectors, and online update triggering mechanisms.

[0009] The present invention is further configured such that the acquisition of real-time data streams of temperature, pressure, weight, and element concentration from the batching, smelting, and casting stages, and the extraction and transmission of these data streams from distributed devices to a centralized data lake via an edge computing gateway acquisition interface, to obtain a unified formatted data set, specifically involves: Weighing sensors are deployed in the batching stage, infrared thermometers, pressure transmitters and spectrometers are deployed in the smelting stage, and flow meters and thermocouples are deployed in the casting stage. All sensors are connected to the edge computing gateway. The edge computing gateway has a built-in data preprocessing unit that timestamps the raw signal and converts the units to the international standard unit system. It also uses linear interpolation to fill in missing values ​​within the sampling interval and generates structured data frames. Structured data frames are uploaded to the central data lake via the OPCUA protocol. The central data lake establishes partition tables based on a two-dimensional index of process and time, forming a data set in a unified format.

[0010] The present invention is further configured such that, based on the dataset, a data fusion algorithm is used to process multi-source heterogeneous information, and data points related to smelting temperature and composition uniformity are fused to determine the fused comprehensive dataset, specifically as follows: Extract the furnace outlet temperature sequence from the infrared thermometer and the element concentration distribution data from the spectrometer from the dataset, and establish a time-aligned window; Within the time alignment window, the confidence levels of the infrared thermometer and spectrometer are quantified using the Dempster-Shafer evidence theory to generate weighted fusion coefficients. The weighted fusion coefficients are applied to the temperature-composition data pairs to construct the melting temperature-composition uniformity mapping matrix. Kalman filtering is then introduced to suppress the noise components in the matrix, resulting in the fused comprehensive dataset.

[0011] The present invention is further configured such that, based on the fused comprehensive dataset, it is determined whether it meets an integrity threshold; if the integrity threshold is met, the association pattern between parameters is analyzed using an association rule mining algorithm to generate an association graph between parameters. Calculate the percentage of valid data points in the merged dataset. If the percentage is greater than 95%, the integrity threshold is satisfied. The dataset that meets the integrity threshold is discretized, the smelting temperature is divided into three intervals, and the composition uniformity is divided into three levels: excellent, good, and poor. The FP-Growth algorithm is used to mine frequent itemsets, with a minimum support of 0.3 and a minimum confidence of 0.7, to generate strong association rules. The Granger causality test is used to perform directional verification of strong association rules, retain statistically significant causal relationships, and construct a graph of association between parameters.

[0012] The present invention is further configured such that, based on the correlation diagram between the parameters, the decision tree algorithm is used to classify the quality influencing factors and determine the list of key influencing factors, specifically as follows: The composition uniformity level is used as a label, and the melting temperature, holding time, and alloy addition amount are used as features to input into the C4.5 decision tree model; During model training, the Gini index is used as the splitting criterion. Splitting stops when the number of node samples is less than 50 or the purity is greater than 0.9. Traverse the decision tree path and extract the top three features that contribute the most to the poor grade to form a list of key influencing factors. If the melting temperature is higher than 1520℃ in any path, mark the factor entry as a high-impact attribute.

[0013] The present invention is further configured such that, based on the list of key influencing factors, the core parameters required for the control mechanism are extracted, real-time feedback data of the current production coordination state is obtained, and by comparing the correlation diagram with the feedback data, it is determined whether there is a need for control. If there is a need for control, the generation of an adjustment instruction sequence is specifically as follows: The melting temperature and holding time are extracted as core parameters from the list of key influencing factors, and the real-time temperature curve and timer reading of the current furnace are obtained through the OPCUA subscription mechanism. The real-time temperature curve is compared point by point with the temperature threshold of 1520℃ in the correlation graph. If five consecutive sampling points exceed the threshold, it is determined that there is a need for regulation. Temperature adjustment amount calculated based on the excess amplitude. It generates an adjustment instruction sequence that includes reducing the melting power by 15% and extending the holding time by 3 minutes, and attaches the execution device address code.

[0014] The present invention is further configured such that, based on the adjustment instruction sequence, the instructions are transmitted to the casting effect control module, and a feedback loop mechanism is used to update the component uniformity-related parameters to obtain the optimized production parameter configuration, specifically as follows: After the adjustment command sequence is verified by the industrial firewall, it is sent to the smelting furnace PLC and the casting robot controller of the casting effect control module via the PROFINET bus; The PLC executes power adjustment commands, and the robotic arm controller adjusts the pouring start time according to the new insulation time. Before the next batch begins, new smelting temperature curves and composition detection results are collected and input into the data fusion module as feedback data to update the smelting temperature-composition uniformity mapping matrix and form an optimized production parameter configuration.

[0015] The present invention is further configured such that, based on the optimized production parameter configuration, the changing trend of product stability indicators is monitored. If the changing trend deviates from a preset range, the association rule mining algorithm is used to re-analyze the parameter associations, and the revised list of key influencing factors is determined as follows: After metallographic analysis of each batch of castings, the graphite spheroidization rate and the standard deviation of tensile strength were extracted as product stability indicators. The moving average method is used to calculate the average of the indicators for the most recent 10 batches. If the indicator of the current batch exceeds the range of the average ±2σ, the backtracking mechanism is triggered. Once the backtracking mechanism is activated, the latest 100 sets of production data are re-entered into the association rule mining process to generate a revised parameter relationship diagram, and the list of key influencing factors is updated accordingly.

[0016] The present invention is further configured such that, based on the difference information between the revised list of key influencing factors and the initial dataset, the differences are iteratively processed through a data fusion algorithm to obtain the final dynamic control model, specifically as follows: Calculate the Jaccard similarity coefficient between the revised list of key influencing factors and the initial list. If it is less than 0.6, start model iteration. The differential features are embedded into the feature space of the data fusion algorithm, and the Dempster-Shafer weight allocation model is retrained. A version control mechanism is established in the central data lake to save the model parameters and performance indicators of each iteration. The iteration is terminated when the prediction accuracy improvement is less than 0.5% for three consecutive iterations, and the final dynamic adjustment model is output.

[0017] The beneficial effects of this invention are as follows: This invention establishes a closed-loop mechanism for the entire process of data acquisition, fusion analysis, rule mining, factor identification, instruction generation, execution feedback, and model iteration. Compared with static control strategies, this invention can quickly respond to dynamic disturbances such as component fluctuations and thermal field changes, effectively reducing the fluctuation range of graphite spheroidization rate and the dispersion coefficient of tensile strength, thus significantly improving the stability of product quality. Attached Figure Description

[0018] The invention will be further illustrated with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the invention. For those skilled in the art, other drawings can be obtained based on the following drawings without any creative effort.

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] The present invention will be further described in conjunction with the following embodiments.

[0021] Depend on Figure 1 As can be seen, the dynamic control method for the entire process of ductile iron casting described in this embodiment includes the following steps: S1. Acquire real-time data streams of temperature, pressure, weight, and element concentration from the batching, smelting, and casting processes. Extract and transmit these data from distributed devices to the central data lake via the edge computing gateway acquisition interface to obtain a unified data set. S2. Based on the data set, a data fusion algorithm is used to process multi-source heterogeneous information, fuse data points related to melting temperature and composition uniformity, and determine the fused comprehensive dataset. S3. Based on the integrated dataset after fusion, determine whether it meets the integrity threshold. If the integrity threshold is met, analyze the association pattern between parameters through the association rule mining algorithm and generate an association relationship graph between parameters. S4. Based on the correlation diagram between the parameters, the decision tree algorithm is used to classify the quality influencing factors and determine the list of key influencing factors; S5. Based on the list of key influencing factors, extract the core parameters required for the control mechanism, obtain real-time feedback data under the current production coordination status, and determine whether there is a need for control by comparing the correlation diagram with the feedback data. If there is a need for control, generate an adjustment instruction sequence. S6. Based on the adjustment instruction sequence, the instructions are transmitted to the casting effect control module, and the component uniformity related parameters are updated using a feedback loop mechanism to obtain an optimized production parameter configuration. S7. Based on the optimized production parameter configuration, monitor the changing trend of product stability indicators. If the changing trend deviates from the preset range, backtrack to the association rule mining algorithm to re-analyze the parameter association and determine the corrected list of key influencing factors. S8. Based on the difference information between the revised list of key influencing factors and the initial data set, these differences are iteratively processed through a data fusion algorithm to obtain the final dynamic control model.

[0022] This embodiment deploys dedicated sensors at each process stage, combined with timestamp synchronization, unit standardization, and missing value imputation processing via edge computing gateways, to uniformly store heterogeneous data from distributed equipment in a central data lake, constructing a data foundation spanning the entire chain from batching and smelting to casting. Compared to the problems of inconsistent data formats and poor interoperability in existing technologies, this invention achieves spatiotemporal alignment and structured integration of multi-dimensional data, providing reliable data support for full-process parameter correlation analysis, and increasing the data validity rate to over 95%.

[0023] This embodiment integrates Dempster-Shafer evidence theory and Kalman filtering multi-source data fusion algorithm, effectively suppressing measurement noise and improving the accuracy of the correlation characterization of melting temperature and composition uniformity data. By combining the FP-Growth algorithm and Granger causality test for association rule mining, the causal relationship pattern between parameters is clarified. The C4.5 decision tree algorithm is used to accurately screen out the core parameters with the greatest impact on composition uniformity (such as melting temperature ≥1520℃ as a high impact factor), overcoming the shortcomings of traditional methods that rely on experience judgment and lack targeted control, thus improving the accuracy of key factor identification by more than 30%.

[0024] This embodiment establishes a closed-loop mechanism covering the entire process of data acquisition, fusion analysis, rule mining, factor identification, instruction generation, execution feedback, and model iteration. It subscribes to feedback data in real time via OPCUA, and when five consecutive sampling points exceed a threshold, it quickly generates precise adjustment instructions (such as temperature reduction). The data is transmitted to the production equipment in real time via the PROFINET bus. Compared to static control strategies, this invention can quickly respond to dynamic disturbances such as component fluctuations and thermal field changes, keeping the graphite spheroidization rate fluctuation range within ±2%, reducing the tensile strength dispersion coefficient by 40%, and significantly improving product quality stability.

[0025] This embodiment monitors product stability indicators using a moving average method. When an indicator exceeds the mean ± 2σ range, a backtracking mechanism is triggered to re-examine parameter correlations and update key influencing factors. The model iteration needs are determined by combining the Jaccard similarity coefficient. Through differential feature embedding and multiple rounds of training, the model is dynamically adjusted to ensure continuous adaptation to changes in production conditions (such as the addition of influencing factors like cooling rates). This iterative mechanism continuously improves the model's prediction accuracy, eventually converging to a stable state. Compared to a fixed model, the response speed for quality risk warnings in long-term production is improved by 50%, adapting to the dynamic adjustment needs of different raw material batches and equipment conditions.

[0026] The optimized production parameter configurations in this embodiment (such as pouring rate, holding time window, and alloy addition ratio) reduce the scrap rate caused by unreasonable parameters. Combined with the automated instruction execution process, it reduces the intensity of manual intervention. Simultaneously, the version control mechanism of the central data lake enables traceable management of model parameters and performance indicators, providing data support for continuous improvement of production processes. After applying this invention, the unit energy consumption of ductile iron production is reduced by 15%, production efficiency is increased by 20%, and the scrap rate is reduced to below 3%, significantly enhancing the company's core competitiveness in the high-end equipment manufacturing field.

[0027] This embodiment describes a dynamic control method for the entire process of ductile iron casting. The unified format data set includes timestamp alignment features, unit standardization specifications, and missing value imputation results. The fused comprehensive dataset includes a melting temperature-composition uniformity mapping matrix, multi-source data confidence weighting coefficients, and spatiotemporal alignment correction terms. The parameter correlation graph includes support-confidence joint threshold screening results, frequent itemset generation paths, and causal direction determination labels. The list of key influencing factors includes influencing factor weight ranking, threshold triggering conditions, and operating context identifiers. The adjustment instruction sequence includes parameter adjustment range, execution priority, and target encoding. The optimized production parameter configuration includes casting rate setpoints, holding time windows, and alloy addition ratios. The product stability indicators include graphite spheroidization rate fluctuation range, tensile strength dispersion coefficient, and metallographic structure consistency index. The dynamic control model includes multi-round iterative convergence criteria, differential feature embedding vectors, and an online update triggering mechanism.

[0028] The dynamic control method for the entire process of ductile iron production described in this embodiment involves acquiring real-time data streams of temperature, pressure, weight, and element concentration from the batching, smelting, and casting stages. This data is then extracted from distributed devices via an edge computing gateway interface and transmitted to a centralized data lake to obtain a unified data set. Specifically, this process involves: Weighing sensors are deployed in the batching stage, infrared thermometers, pressure transmitters and spectrometers are deployed in the smelting stage, and flow meters and thermocouples are deployed in the casting stage. All sensors are connected to the edge computing gateway. The edge computing gateway has a built-in data preprocessing unit that timestamps the raw signal and converts the units to the international standard unit system. It also uses linear interpolation to fill in missing values ​​within the sampling interval and generates structured data frames. Structured data frames are uploaded to the central data lake via the OPCUA protocol. The central data lake establishes partition tables based on a two-dimensional index of process and time, forming a data set in a unified format.

[0029] The dynamic control method for the entire process of ductile iron casting described in this embodiment involves processing multi-source heterogeneous information based on the data set using a data fusion algorithm, fusing data points related to melting temperature and composition uniformity, and determining the fused comprehensive dataset as follows: Extract the furnace outlet temperature sequence from the infrared thermometer and the element concentration distribution data from the spectrometer from the dataset, and establish a time-aligned window; Within the time alignment window, the confidence levels of the infrared thermometer and spectrometer are quantified using the Dempster-Shafer evidence theory to generate weighted fusion coefficients. The weighted fusion coefficients are applied to the temperature-composition data pairs to construct the melting temperature-composition uniformity mapping matrix. Kalman filtering is then introduced to suppress the noise components in the matrix, resulting in the fused comprehensive dataset.

[0030] The dynamic control method for the entire process of ductile iron casting described in this embodiment involves determining whether the fused comprehensive dataset meets an integrity threshold. If the integrity threshold is met, an association rule mining algorithm is used to analyze the association patterns between parameters and generate an association graph between parameters. Specifically, the method involves: Calculate the percentage of valid data points in the merged dataset. If the percentage is greater than 95%, the integrity threshold is satisfied. The dataset that meets the integrity threshold is discretized, the smelting temperature is divided into three intervals, and the composition uniformity is divided into three levels: excellent, good, and poor. The FP-Growth algorithm is used to mine frequent itemsets, with a minimum support of 0.3 and a minimum confidence of 0.7, to generate strong association rules. The Granger causality test is used to perform directional verification of strong association rules, retain statistically significant causal relationships, and construct a graph of association between parameters.

[0031] The dynamic control method for the entire process of ductile iron casting described in this embodiment, specifically involves classifying quality-influencing factors and determining a list of key influencing factors based on the correlation diagram between parameters using a decision tree algorithm: The composition uniformity level is used as a label, and the melting temperature, holding time, and alloy addition amount are used as features to input into the C4.5 decision tree model; During model training, the Gini index is used as the splitting criterion. Splitting stops when the number of node samples is less than 50 or the purity is greater than 0.9. Traverse the decision tree path and extract the top three features that contribute the most to the poor grade to form a list of key influencing factors. If the melting temperature is higher than 1520℃ in any path, mark the factor entry as a high-impact attribute.

[0032] This embodiment describes a dynamic control method for the entire process of ductile iron casting. Based on the list of key influencing factors, the method extracts the core parameters required for the control mechanism, obtains real-time feedback data under the current production coordination state, and determines whether there is a need for control by comparing the correlation diagram with the feedback data. If there is a need for control, an adjustment instruction sequence is generated. Specifically: The melting temperature and holding time are extracted as core parameters from the list of key influencing factors, and the real-time temperature curve and timer reading of the current furnace are obtained through the OPCUA subscription mechanism. The real-time temperature curve is compared point by point with the temperature threshold of 1520℃ in the correlation graph. If five consecutive sampling points exceed the threshold, it is determined that there is a need for regulation. Temperature adjustment amount calculated based on the excess amplitude. It generates an adjustment instruction sequence that includes reducing the melting power by 15% and extending the holding time by 3 minutes, and attaches the execution device address code.

[0033] The dynamic control method for the entire process of ductile iron casting described in this embodiment, wherein the instructions are transmitted to the casting effect control module based on the adjustment instruction sequence, and the composition uniformity-related parameters are updated using a feedback loop mechanism to obtain the optimized production parameter configuration, specifically is as follows: After the adjustment command sequence is verified by the industrial firewall, it is sent to the smelting furnace PLC and the casting robot controller of the casting effect control module via the PROFINET bus; The PLC executes power adjustment commands, and the robotic arm controller adjusts the pouring start time according to the new insulation time. Before the next batch begins, new smelting temperature curves and composition detection results are collected and input into the data fusion module as feedback data to update the smelting temperature-composition uniformity mapping matrix and form an optimized production parameter configuration.

[0034] This embodiment describes a dynamic control method for the entire process of ductile iron casting. Based on the optimized production parameter configuration, it monitors the changing trends of product stability indicators. If the changing trend deviates from a preset range, it backtracks to the association rule mining algorithm to re-analyze parameter associations and determine the corrected list of key influencing factors. Specifically, this list includes: After metallographic analysis of each batch of castings, the graphite spheroidization rate and the standard deviation of tensile strength were extracted as product stability indicators. The moving average method is used to calculate the average of the indicators for the most recent 10 batches. If the indicator of the current batch exceeds the range of the average ±2σ, the backtracking mechanism is triggered. Once the backtracking mechanism is activated, the latest 100 sets of production data are re-entered into the association rule mining process to generate a revised parameter relationship diagram, and the list of key influencing factors is updated accordingly.

[0035] The dynamic control method for the entire process of ductile iron casting described in this embodiment involves iteratively processing these differences using a data fusion algorithm based on the difference information between the corrected list of key influencing factors and the initial data set to obtain the final dynamic control model. Calculate the Jaccard similarity coefficient between the revised list of key influencing factors and the initial list. If it is less than 0.6, start model iteration. The differential features are embedded into the feature space of the data fusion algorithm, and the Dempster-Shafer weight allocation model is retrained. A version control mechanism is established in the central data lake to save the model parameters and performance indicators of each iteration. The iteration is terminated when the prediction accuracy improvement is less than 0.5% for three consecutive iterations, and the final dynamic adjustment model is output.

[0036] This embodiment describes a dynamic control method for the entire process of ductile iron casting. First, a data acquisition and preprocessing module acquires real-time data streams of temperature, pressure, and weight across the three stages of batching, smelting, and casting. In the batching stage, a high-precision weighing sensor is deployed, with its output connected to an edge computing gateway via an RS485 interface. In the smelting stage, an infrared thermometer, pressure transmitter, and spectrometer are installed at the furnace outlet and inside the furnace chamber, respectively, with their shielded signal cables connected to the same edge computing gateway. In the casting stage, an electromagnetic flowmeter and a K-type thermocouple are used to monitor the molten iron flow rate and gate temperature, respectively, and both communicate with the edge computing gateway via the Modbus TCP protocol. The edge computing gateway has a built-in embedded processor running a Linux real-time operating system and loads a dedicated data preprocessing unit program. The program performs timestamp synchronization on the raw signals from various sensors, achieving microsecond-level alignment using the IEEE 1588 precision time protocol. It then converts all physical quantities to the International System of Units (SI), such as temperature to degrees Celsius, pressure to Pascals, and mass to kilograms. For missing values ​​due to communication interruptions or equipment failures, linear interpolation is used to fill in the gaps between adjacent valid sampling points, forming structured data frames. These structured data frames are uploaded to a central data lake server via the OPCUA client component. This server, deployed on an enterprise private cloud platform, uses Apache Kafka as a message middleware to receive and cache the data stream. The Flink stream processing engine then writes the data to an HBase distributed database using a process-time two-dimensional index, thus constructing a unified formatted data set.

[0037] Next, the data fusion module extracts the furnace outlet temperature sequence and elemental concentration distribution data from the spectrometer output from the central data lake. The spectrometer, installed near the furnace taphole, acquires the spectral signal of molten iron in real time via a fiber optic probe and uses a built-in algorithm to analyze the mass fractions of major elements such as carbon, silicon, manganese, sulfur, and phosphorus. The data fusion module establishes a time-aligned window, with a window length set to 1.2 times the current smelting cycle to cover the complete smelting process and its subsequent short pauses. Within this window, the Dempster-Shafer evidence theory is used to quantitatively evaluate the confidence levels of the infrared thermometer and the spectrometer: first, the basic probability assignment function for the two types of equipment is determined based on historical calibration records; then, the confidence factor is dynamically adjusted according to the current level of environmental interference (such as furnace wall radiation intensity and dust obstruction rate), ultimately generating a weighted fusion coefficient. This coefficient is applied to each pair of temperature-composition data points to construct a smelting temperature-composition uniformity mapping matrix. To further suppress measurement noise, a discrete-time Kalman filter is introduced to smooth the abnormal fluctuation components in this matrix, resulting in a fused comprehensive dataset.

[0038] After receiving the integrated dataset, the integrity assessment and association rule mining module first calculates the proportion of valid data points to the total number of samples. If this proportion is greater than 95%, the integrity threshold is met, and the analysis proceeds to the next step; otherwise, an alarm mechanism is triggered, and the subsequent process is paused, awaiting manual intervention. The dataset that meets the integrity condition is then sent to the discretization unit, where the continuous variable smelting temperature is divided into low temperature range [1400℃, 1480℃], medium temperature range [1480℃, 1520℃], and high temperature range [1520℃, 1600℃]. At the same time, the composition uniformity index is divided into three levels according to industry standards: excellent (deviation ≤ ±0.05%), good (±0.05% < deviation ≤ ±0.1%), and poor (deviation > ±0.1%). Based on this discretization result, the FP-Growth frequent pattern growth algorithm is used to mine strong association rules between parameters. The minimum support is set to 0.3 and the minimum confidence to 0.7, filtering out frequent itemsets in the form of "high temperature → poor" and "medium temperature + holding time ≥ 15 min → excellent". To further confirm the causal direction, the Granger causality test is used to verify the statistical significance of each candidate rule: using the component uniformity level as the dependent variable and the melting temperature and other process parameters as independent variables, a VAR vector autoregressive model is constructed. If the p-value of the F-statistic under the lag order is less than 0.05, the causal relationship is retained. Finally, a parameter association graph is generated, containing the support-confidence joint threshold filtering results, the frequent itemset generation path, and the causal direction determination label.

[0039] The key influencing factor identification module uses the parameter correlation graph as input to initiate the C4.5 decision tree classification model training process. The model uses the composition uniformity level as the label variable and selects melting temperature, holding time, and alloy addition amount as feature variables, which are then input into the decision tree builder. The splitting criterion uses the Gini index; splitting stops when a node contains fewer than 50 samples or when the node's purity (i.e., the largest class percentage) exceeds 0.9. After training, all paths leading to the "poor" level leaf nodes are traversed, and the frequency of each feature in the path and its information gain contribution are statistically analyzed. The top three features with the highest comprehensive scores are extracted to form a list of key influencing factors. If the melting temperature in any path is higher than 1520℃, the corresponding factor entry is marked with the "high influence" attribute and an additional working condition context identifier (such as furnace number, shift, raw material batch).

[0040] The control demand judgment and instruction generation module extracts core control parameters from the list of key influencing factors, mainly including melting temperature and holding time. Through the OPCUA subscription mechanism, it obtains the melting temperature curve of the current furnace and the readings of the PLC's built-in timer in real time. The real-time temperature curve is compared point by point with the temperature threshold (set to 1520℃) corresponding to the "high temperature → difference" rule in the parameter correlation diagram. If five consecutive sampling points (sampling interval of 1 second) exceed this threshold, a control demand is determined. At this point, the difference between the actual temperature T_real and the threshold T_threshold is calculated and applied according to the formula... Determine the temperature reduction range. Based on this, generate an adjustment instruction sequence, including "reduce melting power by 15%" and "extend holding time by 3 minutes", and append the target device address code (such as the IP address of the melting furnace PLC, the MAC address of the casting robot controller) and an execution priority field (set to high priority) to the end of each instruction.

[0041] After receiving the adjustment command sequence, the execution and feedback module first performs a security check via the industrial firewall to verify the legality of the command source and the rationality of the parameters. Once the check is successful, the commands are sent via the PROFINET industrial Ethernet bus to the melting furnace PLC and the casting robot controller in the casting effect control module. Upon receiving the power adjustment command, the melting furnace PLC adjusts the conduction angle of the thyristor voltage regulator module, reducing the input power by 15%. The casting robot controller recalculates the casting start time based on the added holding time and updates the motion trajectory planning parameters. Before the next batch begins, the system automatically triggers a new round of data acquisition: the infrared thermometer records the new melting temperature curve again, and the spectrometer simultaneously collects composition data. This feedback data is re-inputted into the data fusion module to update the melting temperature-composition uniformity mapping matrix, thereby generating an optimized production parameter configuration that includes the new casting rate setpoint, holding time window, and alloy addition ratio.

[0042] The product stability monitoring module continuously tracks the quality performance of each batch of castings. After cooling and demolding, each batch of castings is sent to a metallographic laboratory for testing, extracting graphite spheroidization rate and tensile strength standard deviation as product stability indicators. Graphite spheroidization rate is obtained through optical microscope image analysis, and tensile strength is calculated by testing three parallel samples using a universal testing machine and then calculating the standard deviation. The system uses a moving average method to calculate the mean μ and standard deviation σ of the indicators for the most recent 10 batches. If the current batch's indicator exceeds the range of μ ± 2σ, the trend is determined to deviate from the preset range, triggering a backtracking mechanism. After the backtracking mechanism is activated, the latest 100 sets of production data (covering all dimensions of parameters such as batch weight, melting temperature, holding time, alloy ratio, and pouring speed) are re-input into the integrity judgment and association rule mining module. A complete association rule mining process is executed, generating a corrected parameter correlation graph, and the list of key influencing factors is updated accordingly.

[0043] The model backtracking and iterative optimization module is responsible for dynamically controlling the continuous evolution of the model. This module periodically calculates the Jaccard similarity coefficient J = |A∩B| / |A∪B| between the revised list of key influencing factors and the initial list, where A is the initial list and B is the revised list. If J < 0.6, the model is deemed to require iterative updates. At this point, the differential features (such as the newly added "cooling rate" factor) are embedded into the feature space of the data fusion module, expanding the dimension of the basic probability assignment function in the Dempster-Shafer evidence theory, and the confidence-weighted model is retrained. Simultaneously, a version control system is enabled in the central data lake, archiving each iteration's generated new model parameters (including FP-Growth support threshold, decision tree split depth, Kalman filter covariance matrix, etc.) along with corresponding performance metrics such as prediction accuracy and recall. When the prediction accuracy improvement for three consecutive iterations is less than 0.5%, the iteration process terminates, the final dynamically controlled model is output, and it is deployed to the edge computing gateway and central scheduling system to guide subsequent production activities.

[0044] Throughout the implementation process, the modules interact with each other through standardized interfaces: the data acquisition and preprocessing module pushes structured data frames to the central data lake via the OPCUA protocol; the data fusion module, integrity judgment and association rule mining module, key influencing factor identification module, and control demand judgment and instruction generation module are all deployed as microservices on a Kubernetes container cluster and call each other via RESTful APIs; the execution and feedback module communicates with the field PLCs via PROFINET; the product stability monitoring module accesses quality inspection system data through database views; and the model backtracking and iterative optimization module manages model versions through the GitOps mechanism. All modules together constitute a scalable, maintainable, and online-learning intelligent control system, ensuring accurate, stable, and efficient dynamic control of the entire ductile iron production process driven by multi-source heterogeneous data.

[0045] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention are further supplemented below with a specific application scenario.

[0046] In a ductile iron pipe manufacturing enterprise, a continuous casting production line with an annual output of 300,000 tons is facing the problem of decreased compositional uniformity due to fluctuations in smelting temperature. This manifests as a graphite spheroidization rate of less than 85% and a tensile strength standard deviation exceeding 15 MPa in some batches of castings. To solve this problem, the full-process dynamic control method described in this invention is deployed. First, the data acquisition and preprocessing module synchronously acquires the alloy addition amount (such as ferrosilicon and rare earth magnesium alloy) output by the high-precision weighing sensor in the batching stage, the furnace temperature curve recorded by the infrared thermometer in the smelting stage (sampling frequency 1Hz), and the gate temperature monitored by the K-type thermocouple in the casting stage through the edge computing gateway. All raw signals are timestamped according to the IEEE1588 protocol, uniformly converted to SI units, and the missing 2-second spectrometer data due to smoke and dust obstruction is filled by linear interpolation of the effective points before and after, forming a structured data frame and uploading it to the central data lake. This process ensures strict alignment of multi-source heterogeneous data in the spatiotemporal dimension, laying the foundation for subsequent cross-process correlation analysis.

[0047] Subsequently, the data fusion module extracts the outlet temperature sequence corresponding to the current smelting cycle (approximately 45 minutes) and the carbon equivalent (CE=C+1 / 3Si) mass fraction resolved by the spectrometer from the data lake. After establishing a 54-minute time alignment window, the basic probability allocation functions for the infrared thermometer are set to m1(high temperature) = 0.85 and m1(medium temperature) = 0.15, and for the spectrometer, m2(excellent composition) = 0.9 and m2(good composition) = 0.1, based on historical calibration data. Simultaneously, based on the real-time furnace wall thermal imaging, the radiation interference level is determined to be "medium," and the thermometer trust factor is dynamically lowered to 0.78. Finally, a weighted fusion coefficient α = 0.62 is generated for temperature-composition data pair fusion. The resulting smelting temperature-composition uniformity mapping matrix is ​​processed by discrete-time Kalman filtering, effectively suppressing instantaneous temperature jumps (±25℃) caused by the oxide film on the molten iron surface, keeping the composition uniformity assessment deviation within ±0.03%.

[0048] The integrity assessment and association rule mining module detected that the proportion of valid points in the fused dataset reached 97.2%, meeting the 95% threshold. Afterwards, the smelting temperature was discretized into three intervals: [1400, 1480)℃, [1480, 1520)℃, and [1520, 1600]℃. The composition uniformity was classified according to industry standards as excellent (deviation ≤ ±0.05%), good (±0.05% < deviation ≤ ±0.1%), and poor (deviation > ±0.1%). The FP-Growth algorithm was used to mine frequent itemsets: {high temperature, poor} with a support of 0.38 and a confidence of 0.82, and {medium temperature, holding time ≥ 15 min, excellent} with a support of 0.41 and a confidence of 0.76. A VAR(2) model was further constructed using Granger causality test. The results showed that the F-statistic p-value of high temperature on the component uniformity level was 0.013 < 0.05, confirming that "high temperature → poor quality" has a unidirectional causal relationship. Finally, a correlation diagram between parameters with causal labels was generated.

[0049] The key influencing factor identification module uses this relationship graph as input to train the C4.5 decision tree model. The model has the highest information gain (0.43) at the split node "melting temperature ≥ 1520℃". Subsequent paths show that when the holding time is < 12 minutes, the probability of the composition uniformity level falling into "poor" is 89%. After traversing all paths leading to the "poor" leaf node, the melting temperature (occurrence frequency 100%, information gain 0.43), holding time (occurrence frequency 85%, information gain 0.29), and rare earth magnesium alloy addition amount (occurrence frequency 70%, information gain 0.18) are identified as the top three key influencing factors. The melting temperature entry is marked with the "high influence" attribute, and the working condition label "#3 furnace - day shift - raw material batch A20240512" is added.

[0050] The regulation demand judgment and instruction generation module obtains the current temperature curve of furnace #3 in real time through OPCUA subscription. It finds that five consecutive sampling points (1523℃, 1525℃, 1526℃, 1524℃, and 1527℃) all exceed the 1520℃ threshold, thus determining that regulation demand exists. Calculation... This requires a 15% reduction in melting power (based on equipment characteristic curve calibration). Simultaneously, to compensate for the melting kinetic delay caused by cooling, the holding time is extended by 3 minutes. The generated adjustment instruction sequence includes "Power reduction 15%" (target address: 192.168.10.23:PLC_MeltFurnace3) and "Holding time delay 180s" (target address: 00:1B:44:11:3A:B7:Arm_Controller), and is marked with high priority.

[0051] After the execution and feedback module verifies the legality of the command via the industrial firewall, it sends the command through the PROFINET bus. Upon receiving the power adjustment command, the furnace PLC adjusts the thyristor conduction angle from 78° to 66°, reducing the measured input power from 2.8MW to 2.38MW. Simultaneously, the casting robot controller postpones the scheduled casting start time from 14:30:00 to 14:33:00 and replans the robot's motion trajectory to match the new holding time. Before the next batch begins, the system automatically collects the new melting temperature curve (peak 1512℃) and spectral composition data (CE deviation ±0.04%), feeding this data back to the data fusion module to update the mapping matrix, and then outputs optimized production parameter configurations: casting rate maintained at 1.2t / min, holding time window set to 15–18 minutes, and rare earth magnesium alloy addition ratio fine-tuned to 1.05wt%.

[0052] The product stability monitoring module tracked the casting quality of this batch. Metallographic examination showed a graphite spheroidization rate of 92% and a tensile strength standard deviation of 11 MPa. However, after the 7th batch, the mean tensile strength standard deviation calculated by the moving average method for the 10 batches was μ=10.2 MPa and σ=1.8 MPa, while the measured value of the 18th batch reached 15.3 MPa (>μ+2σ=13.8 MPa), triggering the backtracking mechanism. The system immediately retrieved the most recent 100 sets of production data (including newly added parameters such as cooling water flow rate and inoculant addition timing), re-executed association rule mining, and found a strong correlation between "cooling rate > 8℃ / s" and "high tensile strength dispersion" (support 0.33, confidence 0.74). Based on this, the list of key influencing factors was updated, and cooling rate was included in the top three.

[0053] The model backtracking and iterative optimization module calculates the Jaccard similarity coefficient J=|{melting temperature, holding time, cooling rate}∩{melting temperature, holding time, alloy amount}| / |∪|=2 / 4=0.5<0.6, initiating iteration. The cooling rate feature is embedded into the Dempster-Shafer evidence theory framework, expanding the basic probability assignment function to four dimensions, and the confidence-weighted model is retrained. After three iterations, the prediction accuracy improves by 1.2%, 0.8%, and 0.3%, respectively. Iteration is terminated because the third improvement is <0.5%. Finally, the dynamic control model is deployed to the edge gateway, stabilizing the graphite spheroidization rate of subsequent 30 batches of castings at 90%±2% and controlling the tensile strength standard deviation within 12MPa, effectively suppressing quality fluctuations throughout the entire process.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A method for dynamic control of the entire process of ductile iron casting, characterized in that: Includes the following steps: S1. Acquire real-time data streams of temperature, pressure, weight, and element concentration from the batching, smelting, and casting processes. Extract and transmit these data from distributed devices to the central data lake via the edge computing gateway acquisition interface to obtain a unified data set. S2. Based on the data set, a data fusion algorithm is used to process multi-source heterogeneous information, fuse data points related to melting temperature and composition uniformity, and determine the fused comprehensive dataset. S3. Based on the integrated dataset after fusion, determine whether it meets the integrity threshold. If the integrity threshold is met, analyze the association pattern between parameters through the association rule mining algorithm and generate an association relationship graph between parameters. S4. Based on the correlation diagram between the parameters, the decision tree algorithm is used to classify the quality influencing factors and determine the list of key influencing factors; S5. Based on the list of key influencing factors, extract the core parameters required for the control mechanism, obtain real-time feedback data under the current production coordination status, and determine whether there is a need for control by comparing the correlation diagram with the feedback data. If there is a need for control, generate an adjustment instruction sequence. S6. Based on the adjustment instruction sequence, the instructions are transmitted to the casting effect control module, and the component uniformity related parameters are updated using a feedback loop mechanism to obtain an optimized production parameter configuration. S7. Based on the optimized production parameter configuration, monitor the changing trend of product stability indicators. If the changing trend deviates from the preset range, backtrack to the association rule mining algorithm to re-analyze the parameter association and determine the corrected list of key influencing factors. S8. Based on the difference information between the revised list of key influencing factors and the initial data set, these differences are iteratively processed through a data fusion algorithm to obtain the final dynamic control model.

2. The method for dynamic control of the entire process of ductile iron casting according to claim 1, characterized in that: The unified format dataset includes timestamp alignment features, unit standardization specifications, and missing value imputation results; the fused comprehensive dataset includes a melting temperature-composition uniformity mapping matrix, multi-source data confidence weighting coefficients, and spatiotemporal alignment correction terms. The parameter correlation diagram includes the support-confidence joint threshold screening results, frequent itemset generation paths, and causal direction determination labels; the key influencing factor list includes influencing factor weight ranking, threshold triggering conditions, and operating condition context identifiers; the adjustment instruction sequence includes parameter adjustment range, execution priority, and target encoding; the optimized production parameter configuration includes casting rate setpoint, holding time window, and alloy addition ratio; the product stability indicators include graphite spheroidization rate fluctuation range, tensile strength dispersion coefficient, and metallographic structure consistency index; the dynamic control model includes multi-round iterative convergence criteria, differential feature embedding vectors, and online update triggering mechanisms.

3. The method for dynamic control of the entire process of ductile iron casting according to claim 1, characterized in that: The process of acquiring real-time data streams of temperature, pressure, weight, and element concentration from the batching, smelting, and casting stages, extracting them from distributed devices via an edge computing gateway acquisition interface, and transmitting them to a centralized data lake to obtain a unified formatted data set is as follows: Weighing sensors are deployed in the batching stage, infrared thermometers, pressure transmitters and spectrometers are deployed in the smelting stage, and flow meters and thermocouples are deployed in the casting stage. All sensors are connected to the edge computing gateway. The edge computing gateway has a built-in data preprocessing unit that timestamps the raw signal and converts the units to the international standard unit system. It also uses linear interpolation to fill in missing values ​​within the sampling interval and generates structured data frames. Structured data frames are uploaded to the central data lake via the OPCUA protocol. The central data lake establishes partition tables based on a two-dimensional index of process and time, forming a data set in a unified format.

4. The method for dynamic control of the entire process of ductile iron casting according to claim 1, characterized in that: Based on the aforementioned dataset, a data fusion algorithm is used to process multi-source heterogeneous information, fusing data points related to smelting temperature and composition uniformity to determine the fused comprehensive dataset as follows: Extract the furnace outlet temperature sequence from the infrared thermometer and the element concentration distribution data from the spectrometer from the dataset, and establish a time-aligned window; Within the time alignment window, the confidence levels of the infrared thermometer and spectrometer are quantified using the Dempster-Shafer evidence theory to generate weighted fusion coefficients. The weighted fusion coefficients are applied to the temperature-composition data pairs to construct the melting temperature-composition uniformity mapping matrix. Kalman filtering is then introduced to suppress the noise components in the matrix, resulting in the fused comprehensive dataset.

5. The method for dynamic control of the entire process of ductile iron casting according to claim 1, characterized in that: Based on the fused comprehensive dataset, it is determined whether it meets the integrity threshold. If the integrity threshold is met, the association pattern between parameters is analyzed using an association rule mining algorithm to generate an association graph between parameters. Specifically: Calculate the percentage of valid data points in the merged dataset. If the percentage is greater than 95%, the integrity threshold is satisfied. The dataset that meets the integrity threshold is discretized, the smelting temperature is divided into three intervals, and the composition uniformity is divided into three levels: excellent, good, and poor. The FP-Growth algorithm is used to mine frequent itemsets, with a minimum support of 0.3 and a minimum confidence of 0.7, to generate strong association rules. The Granger causality test is used to perform directional verification of strong association rules, retain statistically significant causal relationships, and construct a graph of association between parameters.

6. The method for dynamic control of the entire process of ductile iron casting according to claim 1, characterized in that: Based on the correlation diagram between the parameters, the decision tree algorithm is used to classify the quality influencing factors and determine the list of key influencing factors, specifically as follows: The composition uniformity level is used as a label, and the melting temperature, holding time, and alloy addition amount are used as features to input into the C4.5 decision tree model; During model training, the Gini index is used as the splitting criterion. Splitting stops when the number of node samples is less than 50 or the purity is greater than 0.

9. Traverse the decision tree path and extract the top three features that contribute the most to the poor grade to form a list of key influencing factors. If the melting temperature is higher than 1520℃ in any path, mark the factor entry as a high-impact attribute.

7. The method for dynamic control of the entire process of ductile iron casting according to claim 1, characterized in that: Based on the list of key influencing factors, the core parameters required for the control mechanism are extracted, real-time feedback data of the current production coordination status is obtained, and the correlation diagram and feedback data are compared to determine whether there is a need for control. If there is a need for control, an adjustment instruction sequence is generated as follows: The melting temperature and holding time are extracted as core parameters from the list of key influencing factors, and the real-time temperature curve and timer reading of the current furnace are obtained through the OPCUA subscription mechanism. The real-time temperature curve is compared point by point with the temperature threshold of 1520℃ in the correlation graph. If five consecutive sampling points exceed the threshold, it is determined that there is a need for regulation. Temperature adjustment amount calculated based on the excess amplitude. It generates an adjustment instruction sequence that includes reducing the melting power by 15% and extending the holding time by 3 minutes, and attaches the execution device address code.

8. The method for dynamic control of the entire process of ductile iron casting according to claim 1, characterized in that: Based on the adjustment instruction sequence, the instructions are transmitted to the casting effect control module, and the component uniformity-related parameters are updated using a feedback loop mechanism to obtain the optimized production parameter configuration, specifically as follows: After the adjustment command sequence is verified by the industrial firewall, it is sent to the smelting furnace PLC and the casting robot controller of the casting effect control module via the PROFINET bus; The PLC executes power adjustment commands, and the robotic arm controller adjusts the pouring start time according to the new insulation time. Before the next batch begins, new smelting temperature curves and composition detection results are collected and input into the data fusion module as feedback data to update the smelting temperature-composition uniformity mapping matrix and form an optimized production parameter configuration.

9. The method for dynamic control of the entire process of ductile iron casting according to claim 1, characterized in that: Based on the optimized production parameter configuration, the changing trend of product stability indicators is monitored. If the changing trend deviates from the preset range, the association rule mining algorithm is used to re-analyze the parameter associations and determine the corrected list of key influencing factors. After metallographic analysis of each batch of castings, the graphite spheroidization rate and the standard deviation of tensile strength were extracted as product stability indicators. The moving average method is used to calculate the average of the indicators for the most recent 10 batches. If the indicator of the current batch exceeds the range of the average ±2σ, the backtracking mechanism is triggered. Once the backtracking mechanism is activated, the latest 100 sets of production data are re-entered into the association rule mining process to generate a revised parameter relationship diagram, and the list of key influencing factors is updated accordingly.

10. The method for dynamic control of the entire process of ductile iron casting according to claim 1, characterized in that: The difference information between the revised list of key influencing factors and the initial dataset is used to iteratively process these differences through a data fusion algorithm to obtain the final dynamic control model, specifically as follows: Calculate the Jaccard similarity coefficient between the revised list of key influencing factors and the initial list. If it is less than 0.6, start model iteration. The differential features are embedded into the feature space of the data fusion algorithm, and the Dempster-Shafer weight allocation model is retrained. A version control mechanism is established in the central data lake to save the model parameters and performance indicators of each iteration. The iteration is terminated when the prediction accuracy improvement is less than 0.5% for three consecutive iterations, and the final dynamic adjustment model is output.

Images