A cemented carbide production control method and system based on digital twin

Abstract

The present invention discloses a cemented carbide production control method and system based on digital twins, which relates to the field of digital twin technology. A cemented carbide production control system based on digital twins includes: a twin processing model construction module and a cemented carbide production control module. The present invention constructs a digital twin model for the full-process alloy production equipment, and introduces historical production data and equipment loss data for calibration and regulation. The virtual model can truly map the equipment operation status and process behavior, significantly improving the model prediction accuracy and adaptability; under the premise of meeting quality goals and cost constraints, candidate process paths are generated in multiple equipment clusters, and combined with raw material feature prediction portraits and historical databases for screening, thereby improving the scientificity and operability of path planning, and effectively reducing the blindness and resource waste of process path selection.

Description

Technical Field

[0001] The present invention relates to the field of digital twin technology, and in particular to a cemented carbide production control method and system based on digital twin. Background Art

[0002] Due to its high hardness and wear resistance, cemented carbide is a key base material for cutting tools, mining tools, and wear-resistant parts. Its production process involves multiple tightly coupled physical and chemical steps, and the final product performance is highly sensitive to raw material batches, process parameters, and equipment status. However, existing cemented carbide production control technologies are generally limited. They often rely on fixed standard operating procedures and independent process controls, lacking the ability to adjust to dynamic disturbances such as raw material batch fluctuations and real-time equipment wear and tear. Minor deviations in previous processes are not only difficult to capture, but can also be amplified in subsequent processes.

[0003] Therefore, an intelligent production control method based on digital twins is needed to achieve global optimization and dynamic control from raw materials to finished products through virtual modeling and simulation of the entire cemented carbide process. Summary of the Invention

[0004] The present invention aims to provide a cemented carbide production control method and system based on digital twins, to perform virtual modeling and simulation of the entire cemented carbide process, and to achieve global optimization and dynamic control from raw materials to finished products.

[0005] A cemented carbide production control method based on digital twins includes the following steps:

[0006] Establish digital twin models for all alloy production equipment to obtain several alloy production models; cluster the several alloy production models according to cemented carbide production processes to establish a virtual cemented carbide production model; include N combined alloy production equipment clusters in the virtual cemented carbide production model, where N is the total number of cemented carbide production processes; obtain historical cemented carbide production data; perform initial control on the virtual cemented carbide production model based on the historical cemented carbide production data to obtain a controlled cemented carbide processing model;

[0007] Obtain the current data of raw materials to be produced and the current data of orders to be produced; perform process prediction based on the current data of raw materials to be produced, the current data of orders to be produced and the control cemented carbide processing model to obtain the predicted quality results of the finished product;

[0008] Before the start of cemented carbide production process n, obtain the real-time processing quality results of cemented carbide production process n-1; based on the predicted finished product quality results and the real-time processing quality results, reverse optimization is performed on the control cemented carbide processing model to obtain a new control cemented carbide processing model; the new control cemented carbide processing model is used to control the cemented carbide processing of cemented carbide production process n; and the above steps are repeated until the current cemented carbide production order is completed.

[0009] As a preferred technical solution of the present invention, the specific steps of initially regulating the virtual cemented carbide production model based on cemented carbide production history data include:

[0010] Cemented carbide production history data includes equipment operation data, environmental data, quality control data, and product quality data for each cemented carbide production process; obtains loss data for each current alloy production model equipment;

[0011] Based on the historical data of cemented carbide production, characteristic analysis is performed to obtain the key production-influencing characteristics of each cemented carbide production process; based on the loss data and equipment operation data, characteristic comparative analysis is performed to obtain the equipment loss characteristics corresponding to each cemented carbide production process;

[0012] An equipment loss compensation relationship is established based on equipment loss characteristics; the virtual cemented carbide production model is regulated by combining the equipment loss compensation relationship and key production influencing characteristics to obtain a regulated cemented carbide processing model.

[0013] As a preferred technical solution of the present invention, the specific steps of performing process prediction based on the current raw material data to be produced, the current order data to be produced, and the control of the cemented carbide processing model include:

[0014] Identify processing quality targets and processing cost constraints based on current pending production order data; generate M candidate alloy processing paths based on the N combined alloy production equipment clusters in the control carbide processing model based on the processing quality targets and processing cost constraints;

[0015] Extract raw material features based on the data of raw materials to be produced to obtain the characteristics of alloy processing raw materials; build a historical raw material feature association database; use the alloy processing raw material features to match the historical raw material feature association database to obtain a predicted profile of the raw materials to be produced;

[0016] The predicted image of the raw material to be produced is used as the initial output, and simulation processing is performed in a controlled cemented carbide processing model marked with M candidate alloy processing paths to obtain the predicted finished product quality results.

[0017] As a preferred technical solution of the present invention, the specific steps of performing simulation processing in a controlled cemented carbide processing model marked with M candidate alloy processing paths include:

[0018] For each candidate alloy processing path, simulation processing is performed based on the predicted image of the raw materials to be produced to obtain a predicted processing result matrix; based on the predicted processing result matrix, identification is performed to obtain the key bottleneck process; parameter sensitivity analysis is performed on the key bottleneck process to obtain the processing quality risk index;

[0019] I candidate alloy processing paths with better processing quality risk indexes are extracted, and Monte Carlo simulation is performed using the predicted profile of the raw materials to be produced to obtain I virtual finished product quality result; judgment is made based on all virtual finished product quality results to obtain a predicted finished product quality result.

[0020] As a preferred technical solution of the present invention, the specific steps of reversely optimizing the cemented carbide processing model based on the predicted finished product quality results and the real-time processing quality results include:

[0021] Compare and analyze the real-time processing quality results and the predicted finished product quality results to obtain the production result deviation value;

[0022] Based on the production result deviation value and the preset threshold, a first deviation result or a second deviation result is obtained;

[0023] If the result is the first deviation, the swarm optimization algorithm is used to calculate and output the compensation parameter scheme from cemented carbide production process n to cemented carbide production process N. Among them, only the compensation parameter scheme of cemented carbide production process n is used to build a new control cemented carbide processing model, and the compensation parameter scheme of cemented carbide production process n+1 to cemented carbide production process N is retained as the optimized initial control parameters;

[0024] If it is the second deviation result, the process prediction is re-performed based on the real-time processing quality result to obtain a new predicted finished product quality result; based on the new predicted finished product quality result and the real-time processing quality result, the control carbide processing model is reversely optimized until a new control carbide processing model is output.

[0025] As a preferred technical solution of the present invention, the swarm optimization algorithm is a particle swarm optimization algorithm.

[0026] A cemented carbide production control system based on digital twin, including:

[0027] The twin processing model construction module includes a model construction unit and an initial control unit; the model construction unit is used to establish a digital twin model for all alloy production equipment to obtain several alloy production models; based on the several alloy production models, they are clustered according to the cemented carbide production process to establish a virtual cemented carbide production model; the virtual cemented carbide production model includes N combined alloy production equipment clusters, where N is the total number of cemented carbide production processes; the initial control unit is used to obtain historical cemented carbide production data; based on the historical cemented carbide production data, the virtual cemented carbide production model is initially controlled to obtain a controlled cemented carbide processing model;

[0028] The cemented carbide production control module includes a production prediction unit and a production control unit; the production prediction unit is used to obtain the current raw material data to be produced and the current order data to be produced; the process is predicted for the current raw material data to be produced, the current order data to be produced and the control cemented carbide processing model to obtain the predicted finished product quality result; the production control unit is used to obtain the real-time processing quality result of the cemented carbide production process n-1 before the start of the cemented carbide production process n; based on the predicted finished product quality result and the real-time processing quality result, the control cemented carbide processing model is reversely optimized to obtain a new control cemented carbide processing model; the new control cemented carbide processing model is used to control the cemented carbide processing of the cemented carbide production process n; and the above steps are repeated until the current cemented carbide production order is completed.

[0029] The present invention has the following advantages:

[0030] 1. The present invention constructs a digital twin model for the entire alloy production equipment and introduces historical production data and equipment loss data for calibration and control. The virtual model can truly map the equipment's operating status and process behavior, and has the ability to evolve synchronously with the actual production line, significantly improving the model's prediction accuracy and adaptability. On the premise of meeting quality goals and cost constraints, candidate process paths are automatically generated in multiple equipment clusters, and combined with raw material feature prediction portraits and historical databases for screening, the scientific nature and operability of path planning are improved, and the blindness and resource waste of process path selection are effectively reduced. By constructing a raw material feature database and extracting raw material maps in real time, raw material differences can be perceived in advance in process prediction and simulation, avoiding large-scale defects caused by raw material variations and realizing feedforward control.

[0031] 2. The present invention can identify high-risk processes and key control parameters through sensitivity analysis and quality risk index assessment during the simulation process in a virtual model, providing a clear basis for quality warning and strategic intervention; the predicted finished product quality results are compared and analyzed with the real-time processing quality results to form a deviation judgment mechanism, and whether to perform local compensation optimization or full-process re-forecasting is selected according to the deviation level, with the ability of autonomous learning and continuous correction; the particle swarm algorithm is used to perform group optimization of key process parameters, avoiding full-process readjustment while controlling deviations, maximizing the protection of existing process stability, and achieving quality closed-loop control with minimal disturbance. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a structural diagram of a cemented carbide production control system based on digital twins adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0034] Example 1, a cemented carbide production control method based on digital twin, comprising the following steps:

[0035] Establish digital twin models for all alloy production equipment to obtain several alloy production models; cluster the several alloy production models according to cemented carbide production processes to establish a virtual cemented carbide production model; include N combined alloy production equipment clusters in the virtual cemented carbide production model, where N is the total number of cemented carbide production processes; obtain historical cemented carbide production data; perform initial control on the virtual cemented carbide production model based on the historical cemented carbide production data to obtain a controlled cemented carbide processing model;

[0036] Taking the design drawings, control logic, real-time sensor streams and historical maintenance records of the entire line of alloy production equipment as input, a digital twin model is established for each device through multi-physics field simulation, data fusion and object-oriented model packaging methods; the output is a set of independent and interactive alloy production equipment models, which are expressed as hierarchical objects in the data structure. The top layer encapsulates geometry and process parameters, and the lower layer encapsulates control variables and health status. They can synchronize field data in real time and perform offline operation predictions; the beneficial effect of this is to mirror the physical equipment in the virtual space, laying a high-reliability unit foundation for subsequent process-level optimization.

[0037] Taking the above-mentioned digital twin models as input, the functional characteristics of each cemented carbide equipment in the specific process (such as powder batching, pressing, vacuum sintering, etc.) are first extracted, and then the similarity feature vector is sent to the improved density-based process clustering algorithm, and the output is a virtual cemented carbide production model that is automatically merged according to the process; the data structure of the model is a directed acyclic graph, in which the nodes are the digital twins of the equipment, the edges represent the process sequence constraints, and the key process parameter ranges are recorded in the node attributes; its method can map scattered equipment into a complete process, and the beneficial effect is the intuitive presentation of process dependencies, which is convenient for subsequent overall production scheduling and bottleneck identification.

[0038] Within the virtual cemented carbide production model, topological division is further performed according to the process sequence to form N clusters of combined cemented carbide production equipment. The input is the directed graph generated in the second step, and the output is N cluster subgraphs marked in the graph. Their data structure is a labeled subgraph object, and each subgraph contains a list of belonging equipment, process number and constraint matrix. This link adopts an automatic segmentation method based on topological sorting, so that each cluster corresponds to only a single process and retains cross-process interface information. This helps to modularize complex processes and supports targeted analysis or horizontal expansion of any process.

[0039] Taking the cemented carbide production history records accumulated in the enterprise's manufacturing execution system, field data acquisition system and enterprise resource planning system as input, after completing the extraction, conversion and loading steps, a historical data set is obtained; the data structure of this data set is mainly a table of time series and batch dimensions, and the fields include equipment number, recipe parameters, energy consumption, yield, etc.; the methods used cover outlier elimination, missing value filling and feature normalization; the beneficial effect of data processing is to provide high-quality samples for subsequent model calibration, so that the virtual model can show accurate spatiotemporal response to real production.

[0040] By combining parameter identification with supervised machine learning, the virtual cemented carbide production model is initially regulated, and the output is a regulated cemented carbide processing model. In terms of data structure, it is mainly manifested in the updated equipment parameter dictionary, process window range and control algorithm weights. In terms of method, historical data is first used to train the gradient boosting decision tree to predict key quality indicators, and then particle swarm optimization is used to search for feasible control parameters in the virtual model. The beneficial effect obtained in this way is to make the digital twin as close as possible to the actual production and processing logic during simulation processing, which can not only predict production line performance, but also provide feasible control strategies to reduce defects, save energy and reduce emissions.

[0041] In this embodiment, taking tungsten carbide-cobalt cemented carbide processing as an example, the production process can generally be divided into the following main steps: powder preparation, wet ball milling mixing, spray granulation, forming and pressing, pre-sintering dewaxing, vacuum sintering and hot isostatic pressing, precision machining and final surface coating and inspection: In the powder preparation link, tungsten powder, cobalt powder and trace amounts of modified carbides such as tantalum carbide, titanium carbide, and niobium carbide are weighed in a predetermined ratio and sent to a wet ball mill to be ground together with ethanol or acetone and organic binders such as paraffin and polyvinyl alcohol; the slurry is then granulated through a spray drying tower to obtain spherical particles with good fluidity; the granulated material is pressed into a blank by a mold press or cold isostatic pressing equipment; the blank enters a vacuum or hydrogen furnace to implement staged heating to remove paraffin, and then completes the carbon thermal reaction in the high temperature zone. And densification, if necessary, high-pressure argon is applied in the hot isostatic pressing furnace to further eliminate the remaining porosity; after leaving the furnace, the sintered block is corrected in size and shape on a CNC cylindrical grinder, internal hole grinder, wire cutting or laser processing center, and then a titanium aluminum nitride or titanium carbide film is deposited in a chemical vapor deposition or physical vapor deposition device to improve wear resistance; the main equipment involved in the whole process includes wet grinding ball mill, spray drying tower, hydraulic or isostatic pressing press, vertical and horizontal vacuum sintering furnace, hot isostatic pressing furnace, precision grinding center and coating reaction furnace, and the key raw material data covers the average particle size of tungsten powder, cobalt powder purity, added carbide ratio, organic binder content, granulation moisture content and sintering hydrogen or vacuum atmosphere composition, these data together determine the structure, hardness and toughness of the cemented carbide grade.

[0042] The specific steps for initial adjustment of the virtual cemented carbide production model based on historical cemented carbide production data include:

[0043] Cemented carbide production history data includes equipment operation data, environmental data, quality control data, and product quality data for each cemented carbide production process; obtains loss data for each current alloy production model equipment;

[0044] Taking the original records derived from the manufacturing execution system, equipment data collector, environmental monitoring node and laboratory quality inspection database as input, equipment operation data refers to the real-time or periodic records collected directly from production equipment such as powder preparation machines, spray drying towers, presses, vacuum furnaces, hot isostatic pressing furnaces, grinding centers, etc., including motor speed, temperature curve, vacuum degree, pressure, vibration amplitude, bearing temperature rise, energy consumption, current, voltage and the switch status and alarm codes of various actuators, which are used to reflect the current process execution status and health status of the equipment; environmental data describes the surrounding conditions of the production workshop, such as air temperature, relative humidity, oxygen content, hydrogen purity, dust particle concentration, illumination, power grid fluctuations and noise levels, etc. They are collected through distributed environmental sensors, atmosphere analyzers and power quality monitors, and are used to Assess the impact of external factors on process stability. Quality control data refers to information obtained through random inspections or online testing of intermediate products or process status during the production process. This includes information such as the solid content of the mixed slurry, granulation moisture, briquette density, green body size deviation, dewaxing weight loss rate, sintering shrinkage rate, and online micropore assessment results. This is used to determine in real time whether each process remains within the predetermined process window. Product quality data refers specifically to the comprehensive performance test results of the final product or sample after it leaves the furnace, covering hardness, flexural strength, impact toughness, microstructure distribution, cobalt pool ratio, surface roughness, wear resistance test life, coating thickness uniformity, and appearance defect ratings. Its sources are laboratory physical and chemical tests, automated appearance inspections, and user feedback. It is the final basis for evaluating the production line process effect and product qualification rate.

[0045] Wear data for each alloy production model device is a numerical record used to quantify the performance degradation and component wear experienced by the device from new to its current state. This data covers a wide range of mechanical, thermal, electrical, and airtight aspects. Examples include vibration acceleration increments, noise levels, and clearance growth for spindles and rolling bearings; pressure compensation amplitude and response delay for hydraulic or servo systems under constant load; thermal resistance increase for vacuum furnace insulation, resistance drift rate for heating elements, and cooling water flow rate attenuation; diameter wear, blade blunting angle, and self-excited vibration mark increase for grinding wheels or cutting tools; pressure drop increase and clogging coefficient for filters, atmosphere lines, or powder conveying screws; current harmonic content and efficiency degradation curves for motors, power supplies, and inverters; and runtime, start-stop counts, and maintenance replacement timestamps corresponding to these physical quantities. All metrics are typically indexed by device and timestamp, with fields including wear type, measurement value, threshold, estimated error, and health score. These fields are used to update device health in real time within the digital twin, predict remaining life, and trigger appropriate process compensation logic.

[0046] Based on the historical data of cemented carbide production, characteristic analysis is performed to obtain the key production-influencing characteristics of each cemented carbide production process; based on the loss data and equipment operation data, characteristic comparative analysis is performed to obtain the equipment loss characteristics corresponding to each cemented carbide production process;

[0047] In the step of feature analysis based on historical cemented carbide production data, the input is a historical wide table that has been cleaned and aligned. The table records the equipment parameters, environmental indicators and corresponding final quality results of each batch in the powder preparation, granulation, pressing, sintering and other processes. The Spearman correlation coefficient is used to screen out variables that are almost irrelevant to the target quality indicators. Then, the dimension is reduced by recursive feature elimination. The random forest and gradient boosting models are trained separately to calculate the importance of the remaining variables, and the stability of the ranking is cross-validated using the Shapley value. The output is a process-feature mapping table with fields including process number, feature name, importance score, confidence interval and influence direction (positive or negative). This approach based on multi-model consistency testing can clearly locate the control variables that are most sensitive to hardness, bending strength and micro-defect rate while avoiding the bias of a single algorithm, providing a reliable basis for subsequent parameter compensation and process optimization.

[0048] In the feature comparison analysis step based on wear data and equipment operation data, the input is the equipment wear data and the operation curve for the same time window. The wear data records indicators such as vibration amplitude, thermal resistance increase, current harmonics, and grinding wheel diameter reduction by equipment ID and date. The operation curve stores temperature, pressure, power, and online quality inspection results within the same step length. The two types of data are aligned by equipment and time window, and the residuals of key process parameters and quality indicators are calculated at different wear levels. Multiple linear regression and dynamic time warping are then used to extract significant wear-driven drift. Variables with similar performance are grouped into wear feature families using hierarchical clustering. The output is an equipment wear feature matrix with fields including process number, wear feature name, normalized drift amplitude, and significance p-value. A dictionary is also included to map each wear feature to a specific equipment and physical component. This analysis quantifies the impact of equipment aging on the process window, providing computable input for establishing wear compensation relationships and dynamic control, enabling the virtual model to maintain predictive accuracy as equipment health changes.

[0049] Establishing an equipment loss compensation relationship based on equipment loss characteristics; regulating the virtual cemented carbide production model by combining the equipment loss compensation relationship and key production-influencing characteristics to obtain a regulated cemented carbide processing model;

[0050] In the step of establishing an equipment wear compensation relationship based on equipment wear characteristics, the input is the equipment wear characteristic matrix obtained in the previous stage and the dictionary mapping wear characteristics to physical components. The quality indicators of historical batches are also read as a reference. The wear indicator of each device is paired with the offset of the key control variable in its process. Bayesian multi-task regression is then used to fit the functional relationship between wear and control parameter drift. To prevent overfitting, the model uses hierarchical priors to restrict the slope distribution of components of the same type and selects the optimal hyperparameters through cross-validation using the holdout method. The output is an equipment wear compensation table. The data structure uses equipment identification, wear indicator name, compensation slope, compensation intercept, applicable wear range, and residual variance as fields. At the same time, a callable compensation function object is generated and stored in the model library. The control parameters that originally drifted gradually with wear are quantitatively mapped into compensation amounts that can be corrected in real time, thereby providing a mathematical basis for parameter adaptation in subsequent production and a reference for the equipment maintenance department to determine when recalibration or component replacement is needed.

[0051] In the process of combining the compensation relationship with production-influencing features and regulating the virtual production model, the input includes the complete virtual cemented carbide production model, the compensation function object generated in the previous step, and the process-feature importance table. The compensation function is first injected into the model's control layer. At each simulation step, parameters such as the speed curve, temperature set point, and holding time are dynamically updated based on the real-time wear value. The updated process parameters are then weightedly coupled with the importance table. The built-in particle swarm optimization is used to search for the optimal control combination that can simultaneously meet the hardness, flexural strength, and porosity targets under the current equipment health. The output is the regulated cemented carbide processing model, whose data structure consists of a new process node object, a control parameter dictionary, a real-time compensation coefficient table, and a quality prediction distribution. The model is then run and replayed in a sandbox to verify that the finished product quality distribution is within the target range. The beneficial effect of this regulation step is that the digital twin can automatically adjust the process window to closely follow the equipment aging trajectory, maintaining high-precision predictions of the real production line and continuously providing operators and scheduling systems with optimization solutions for energy saving, consumption reduction, and defect reduction.

[0052] Obtain the current data of raw materials to be produced and the current data of orders to be produced; perform process prediction based on the current data of raw materials to be produced, the current data of orders to be produced and the control cemented carbide processing model to obtain the predicted quality results of the finished product;

[0053] The specific steps for predicting the process based on the current raw material data to be produced, the current order data to be produced, and the control of the cemented carbide processing model include:

[0054] Identify processing quality targets and processing cost constraints based on current pending production order data; generate M candidate alloy processing paths based on the N combined alloy production equipment clusters in the control carbide processing model based on the processing quality targets and processing cost constraints;

[0055] The processing quality target is the quantitative requirement of the order for the function and appearance of the final cemented carbide product. Typical contents include specific indicators such as hardness, bending strength, impact toughness, microporosity, coating thickness, dimensional tolerance, surface roughness and defect rate. Each item gives a numerical range or maximum allowable deviation that must be achieved to ensure that the parts meet customer or industry standards in terms of service life, reliability and regulatory compliance; the processing cost constraint is the economic and resource boundary allowed by the order to achieve this goal, usually covering the upper limit of raw material procurement amount, upper limit of energy consumption per unit part, calculated machine depreciation and labor time cost, tool wear cost, upper limit of scrap loss rate and even emission or carbon cost control target; these constraints limit the production system to meet quality requirements while keeping total costs within the budget and ensuring delivery cycle and profitability.

[0056] Each equipment cluster is treated as a node in the graph. The node's production capacity, energy consumption curve, and calibrated parameter window are simultaneously read. A multi-objective heuristic search algorithm is then used to enumerate feasible paths that satisfy the constraints on the directed acyclic graph. Fast non-dominated sorting is then used to retain M paths on the cost-quality Pareto front. The output is a candidate path library consisting of a path number, process sequence, set of key process parameters, estimated energy consumption, and estimated processing time. This step provides a set of cost-quality-balanced process routes for subsequent simulations, narrowing the search space and reducing the computational load.

[0057] Extract raw material features based on the data of raw materials to be produced to obtain the characteristics of alloy processing raw materials; build a historical raw material feature association database; use the alloy processing raw material features to match the historical raw material feature association database to obtain a predicted profile of the raw materials to be produced;

[0058] The data of raw materials to be produced are stored in a structured manner using fields such as batch number, average particle size of tungsten powder, purity of cobalt powder, proportion of modified carbide, oxygen content, moisture content, and specific surface area. The original fields are subjected to distribution statistics, missing data are filled in, and unit correction is performed. Secondary indicators such as skewness and kurtosis that describe the particle morphology are calculated, and a standardized raw material feature vector is finally output. The data structure is a vector table with batch number as the key and containing more than 30 numerical dimensions. This link condenses chemical and physical indicators into a unified numerical expression, providing a consistent format for database matching and simulation model input.

[0059] When constructing a historical raw material feature association database, the input is the cleaned raw material feature vectors and the corresponding finished product quality results from historical production records. A one-to-one association is first established between raw material features and finished product quality based on batch number. A hash index is then used to accelerate vector retrieval. The database stores foreign key relationships between four logical tables: raw material features, equipment parameters, process paths, and final quality. The output is a columnar database that can be queried in parallel. The data structure includes a vector index tree, feature table, path table, and quality table. This allows for rapid subsequent retrieval of historical records similar to the current raw material, providing empirical data support for profiling and simulation.

[0060] Taking alloy processing raw material characteristics and a database of historical raw material characteristics as input, a metric learning algorithm based on neighborhood similarity is used to retrieve the closest historical vectors. The average quality indicators, common process parameters, and corresponding fluctuation ranges of these historical batches are weighted and synthesized into a prediction profile. The output is a raw material profile object, whose data structure consists of the batch number, estimated sinterable density, predicted shrinkage range, possible defect probability distribution, and confidence interval. This profile provides the simulation model with initial conditions that closely resemble real-world raw material behavior, thereby improving prediction accuracy.

[0061] Taking the predicted image of the raw material to be produced as the initial output, simulate the processing in the controlled cemented carbide processing model marked with M candidate alloy processing paths to obtain the predicted finished product quality result;

[0062] The specific steps of simulating machining in the controlled cemented carbide machining model marked with M candidate alloy machining paths include:

[0063] For each candidate alloy processing path, simulation processing is performed based on the predicted image of the raw materials to be produced to obtain a predicted processing result matrix; based on the predicted processing result matrix, identification is performed to obtain the key bottleneck process; parameter sensitivity analysis is performed on the key bottleneck process to obtain the processing quality risk index;

[0064] Extracting I candidate alloy processing paths with relatively good processing quality risk indices, performing Monte Carlo simulation using the predicted profile of the raw material to be produced, and obtaining I virtual finished product quality result; performing judgment based on all virtual finished product quality results to obtain a predicted finished product quality result;

[0065] During the candidate path-by-path simulation phase, the input includes a process sequence of M candidate alloy processing paths and their calibrated parameter windows, as well as more than 30 dimensional feature vectors such as particle size distribution, oxygen content, and binder residue contained in the predicted portrait of the raw material to be produced. The model injects the raw material portrait into each path and simultaneously solves the heat transfer, sintering densification, and stress field equations within the virtual equipment cluster according to a physical-data hybrid algorithm. It also calls a machine learning agent model in real time to compensate for microstructure evolution, and the output is a predicted processing result matrix, which records variables such as temperature, pressure, porosity, shrinkage, and immediate strength according to the batch-process-time triple index. This can provide a quality and energy consumption baseline covering all processes for each path without consuming physical production capacity.

[0066] In the bottleneck identification phase, the input is the predicted processing result matrix generated in the previous stage; first, the sensitivity of each process to the target quality indicator is calculated, and then the process time and energy consumption coefficient are combined, and a sorting method based on multi-objective tensor decomposition is used to mark steps with high defect risk, high energy consumption or high waiting time as bottlenecks; the output is a bottleneck process table that records the process number, risk weight and contribution confidence interval. The data structure uses column storage for fast query; this step helps production engineers focus on the few links that truly limit quality or cycle time.

[0067] In the bottleneck process parameter sensitivity analysis phase, the input is the bottleneck process table and the adjustable control parameter range of the corresponding process; Latin hypercube sampling is used to generate large-scale parameter combinations and parallel simulations are performed in a virtual model. Subsequently, a global sensitivity analysis based on the Sobol index is used to separate the contribution rate of each parameter to the variance of hardness, toughness, and porosity. The output is a processing quality risk index vector, in which each element is a composite of the parameter variance contribution and the process risk weight; this quantitative method reveals in which parameter ranges the quality fluctuations are most severe, providing an objective score for path screening.

[0068] In the candidate path screening stage, the input is the processing quality risk index corresponding to each path; the paths are sorted from low to high according to the risk index, and I paths with the best performance are selected on the Pareto curve based on a pre-set threshold or budget. The output is a path selection list with fields including path number, cumulative risk index, estimated energy consumption and expected cycle time; by eliminating high-risk paths, the computational complexity of subsequent random simulations can be greatly reduced, while ensuring that the remaining solutions are more likely to meet the standards in real production in one go.

[0069] In the Monte Carlo simulation stage, the input is a predicted profile of the raw materials to be produced and I screened paths; a joint probability distribution is generated for the raw material characteristics, equipment compensation coefficients and environmental fluctuations, and a large number of random sampling simulations are run on each path to record the full distribution of the finished product hardness, flexural strength, porosity and energy consumption. The output is I virtual finished product quality result objects, each of which contains the quality indicator percentile value, compliance probability and cost distribution. The data structure is encapsulated in an object-oriented manner to facilitate subsequent query and visualization. This process fully injects the uncertainty in reality into the model to obtain more reliable quality predictions for future batches.

[0070] In the finished product quality result summary and judgment stage, the input is I virtual finished product quality result objects; the probability of each path meeting the order quality target is first calculated, and then the score is normalized according to the cost weight. Finally, the path with the highest comprehensive score is selected through the Bayesian decision rule, and a predicted finished product quality result is output, giving the expected values, confidence intervals and compliance probabilities of hardness, flexural strength, porosity, energy consumption and production cycle.

[0071] Before cemented carbide production process n begins, obtain the real-time processing quality results of cemented carbide production process n-1; perform reverse optimization on the control cemented carbide processing model based on the predicted finished product quality results and the real-time processing quality results to obtain a new control cemented carbide processing model; use the new control cemented carbide processing model to control the cemented carbide processing of cemented carbide production process n; repeat the above steps until the current cemented carbide production order is completed;

[0072] The specific steps for reverse optimization of the cemented carbide machining model based on the predicted finished product quality results and the real-time machining quality results include:

[0073] Compare and analyze the real-time processing quality results and the predicted finished product quality results to obtain the production result deviation value;

[0074] In the calculation of production result deviation values, the input is the real-time processing quality results and the predicted finished product quality results of the corresponding batch. Both data tables use batch number and process number as joint indexes. The fields include indicators such as hardness, flexural strength, porosity, dimensional deviation and energy consumption. First, an inner join is performed on the prediction table and the real-time table according to the index, and the unit correction and dimension unification are performed on the numerical fields. Then, the relative error and absolute error of each quality indicator are calculated, and the differences in multiple indicators are integrated through the Mahalanobis distance to generate a comprehensive deviation score; the output is a deviation value table with fields such as batch number, process number, deviation of each individual indicator, comprehensive deviation score and calculation timestamp. The data structure uses column storage to facilitate rapid aggregation. The purpose of this step is to use rigorous quantitative methods to make the gap between model prediction and on-site performance visible in real time, providing basic data for subsequent judgment of model credibility and compensation strategy.

[0075] Based on the deviation value of the production result and the preset threshold, a first deviation result or a second deviation result is obtained; the preset threshold is manually set by professional technicians;

[0076] The first deviation result indicates that the combined deviation between the real-time processing quality and the predicted result is still within the set tolerance range, indicating that the virtual model's description of the current raw material characteristics and equipment health status is still basically reliable. Only limited compensation of the control parameters within the scope of subsequent processes that have not yet been executed is required to reduce the error back to the target range; the second deviation result indicates that at least one key quality indicator or comprehensive deviation exceeds the threshold, indicating that the deviation between the original model, process path or parameter assumptions and the on-site conditions is so significant that it cannot be eliminated through local compensation. It is necessary to re-evaluate the raw material profile, re-screen the process path and re-simulate to obtain a new quality prediction baseline. The role of this two-tier judgment mechanism is to divide real-time feedback into two situations: those that can be quickly corrected and those that require remodeling: when encountering the first deviation result, the system only calls particle swarm optimization for rapid fine-tuning to ensure uninterrupted production; when encountering the second deviation result, the full process re-prediction and iterative optimization are initiated to avoid blind compensation in the case of model inaccuracy, which may lead to greater quality risks, thereby achieving a balance between efficiency and robustness.

[0077] In the deviation judgment link, the input is the deviation value table and the threshold vector preset by the system. The threshold vector gives the upper limit, mean fluctuation allowable degree and comprehensive deviation score threshold according to the quality index. The deviation value table is traversed in batches. If all single errors and comprehensive errors are lower than the corresponding threshold, it is marked as the first deviation result; as long as any indicator or comprehensive difference exceeds the threshold, it is marked as the second deviation result, and the exceeded field is written into the abnormal field set; the output is a deviation classification report, the fields of which include batch number, classification result, list of exceeded indicators and deviation overview summary. In this way, the complex multi-indicator differences are summarized into two clear results, laying a decision-making basis for the subsequent adoption of different strategies of gradual correction or re-forecasting.

[0078] If the result is the first deviation, a swarm optimization algorithm is used to calculate and output a compensation parameter solution from cemented carbide production process n to cemented carbide production process N. Only the compensation parameter solution of cemented carbide production process n is used to construct a new control cemented carbide processing model, and the compensation parameter solution from cemented carbide production process n+1 to cemented carbide production process N is retained as the initial control parameter for optimization; the swarm optimization algorithm is a particle swarm algorithm;

[0079] When the deviation is determined to be the first deviation result, the particle swarm algorithm is used to optimize the compensation parameters of the remaining processes; the input includes the boundary of the parameters to be optimized from the current process n to the final process N, the error vector after n processes in the deviation value table, and the calibrated virtual production model; the position vector of each particle represents the temperature gradient, holding time, heating rate and other control parameters of the n process, and the fitness function is defined as the weighted sum of the residual minimization of the real-time quality result after the model is updated and the energy consumption constraint penalty; in the initialization stage, points are evenly spread in the parameter space according to the historical optimal solution and the current equipment wear, and then the speed and position are updated in each iteration And calculate the fitness, and write the local optimum and global optimum into the shared buffer in real time; the iteration termination condition is that the fitness improvement rate is lower than the set threshold or the maximum number of iterations is reached, and the output is a compensation parameter scheme table with fields such as process number, control parameter name, compensation amount, expected residual and number of iterations; among them, only the parameters of process n are extracted and immediately written into the new control carbide processing model, and the compensation parameters corresponding to processes n+1 to N are recorded as the initial values ​​for subsequent optimization. The effect of this is to compress the error to an acceptable range with a small number of calculation steps without interrupting production, while retaining the search memory of subsequent processes to accelerate the next round of optimization.

[0080] If the result is the second deviation, the process prediction is re-performed based on the real-time processing quality result to obtain a new predicted finished product quality result; based on the new predicted finished product quality result and the real-time processing quality result, the control carbide processing model is reversely optimized until a new control carbide processing model is output;

[0081] When the deviation is judged to be the second deviation result, it is considered that the original model's description of the current raw material and equipment status is no longer sufficient to support the prediction, and the process prediction needs to be re-performed; the input is the real-time processing quality results, the predicted portrait of the raw materials to be produced and the candidate path library; the system re-executes the path selection, Monte Carlo simulation and quality prediction process to obtain a new predicted finished product quality result, and then compares it with the latest real-time data to generate a new deviation value table; if the deviation still exceeds the threshold, the reverse optimization cycle will continue until convergence or the maximum number of iterations is reached; in each cycle, the latest compensation parameters and model structure will be persisted to ensure that the model gradually approaches the actual production line behavior over time, so that it can still maintain prediction and control capabilities when facing raw material fluctuations or sudden equipment failures.

[0082] In this embodiment, a specific implementation plan is provided. Taking tungsten carbide-cobalt cemented carbide processing as an example, after a batch of tungsten carbide-cobalt cemented carbide orders for the manufacture of wear-resistant cutting blades enters the system, the scheduling system first obtains the order data. The order requires a product hardness of 92.5HRA or above, a bending strength of no less than 2800MPa, a batch size of 2000 pieces, a cost budget of less than 15 yuan per piece, and a delivery cycle of no more than 72 hours. After parsing the order, the system extracts clear processing quality targets (hardness, strength, size) and processing cost constraints (raw materials, energy consumption, time, equipment utilization rate) as optimization inputs. Next, the system retrieves N equipment clusters in the current state from the regulated virtual cemented carbide processing model and, based on the quality targets and cost constraints, generates M candidate processing paths in the constrained directed process graph. These include traditional routes (wet ball milling, spray granulation, compression molding, vacuum sintering, precision grinding) and energy-saving variant routes (dry mixing, hot pressing, integrated sintering).

[0083] Then, the raw material batch data currently to be used is imported, in which the tungsten powder particle size is 1.2 microns, the cobalt content is 6.0%, the oxygen content is 0.65%, and 0.5% of tantalum carbide and titanium carbide are mixed in. The raw material is tested to have a slightly high moisture content; the system extracts high-dimensional vectors such as morphology, oxygen content, and proportion from the raw material characteristics, and performs similarity matching in the historical raw material database to obtain the historical sintering shrinkage range (17.2%-18.1%), dewaxing temperature window (200°C-350°C), and typical hardness fluctuation range (91.8-93.0HRA) of the raw material under similar proportions and particle sizes. A predicted raw material profile is constructed. The system then inputs this predicted raw material profile into all M candidate paths, performs a full-process simulation on each path, generates process response records including the powder densification process, the dewaxing process temperature control curve, and the density evolution curve after molding, and outputs a predicted processing result matrix containing the predicted values ​​of the finished product hardness, strength, porosity, and energy consumption for each path. The system further performs a sensitivity analysis on the output distribution of each process in the simulation results, identifying a path where the pressure fluctuation of the pressing process and the heating rate of the sintering process contribute more than 60% to the final hardness, making it a key bottleneck process. Based on the processing quality risk index, the system selects five lower-risk paths and performs Monte Carlo simulation using the raw material profile and the uncertainty of the simulation parameters. It simulates the quality distribution of the finished product under different combinations of particle size fluctuations, oxygen content deviations, and wear conditions, and outputs a virtual finished product quality report. The report shows that a certain path has a 97% probability of meeting the target hardness and strength range and has the lowest unit energy consumption, making it the optimal processing path for this batch. After executing actual production, the system collects online inspection data during the processing in real time, including pressing density, sintering reduction ratio, dewaxing residual carbon rate and the initial hardness measurement value. After the processing is completed, the real-time quality results are compared with the predicted values, and it is found that the average hardness is 0.4HRA lower and the porosity is slightly higher than the predicted value by 0.2%. The system determines this as the first deviation result. Because the error is within the threshold, it decides to start parameter compensation from the current process (pressing). In the compensation link, the system uses the particle swarm optimization algorithm to optimize the parameters such as the specific pressure, holding time and sintering heating rate of the pressing process. With error minimization as the objective function, combined with the current equipment wear parameters (such as reduced press oil pressure stability and increased heating element resistance), the system completes 25 iterations and outputs a new set of process control parameters: the pressing specific pressure is increased by 2.5MPa, the holding time is extended by 3 seconds, and the sintering heating curve is adjusted to a double-step mode. The new parameters are immediately loaded into the virtual model for subsequent batch execution.If the system detects that subsequent batches still have hardness deviations exceeding 0.8HRA and porosity fluctuations exceeding 0.5% after adjustment, the system will activate the second deviation processing mechanism, re-match the current raw material profile, re-screen the path, and perform new predictions and simulation verifications to ensure that the model still has dynamic adaptability in the context of further development of equipment wear. Through the above process, the prediction, simulation, feedback and optimization closed-loop control of the entire tungsten carbide-cobalt cemented carbide processing process is realized, which effectively improves the consistency of finished products, reduces the defect rate and energy consumption, and provides highly automated and intelligent support for the manufacturing process.

[0084] Example 2, a cemented carbide production control system based on digital twin, see Figure 1 As shown, including:

[0085] The twin processing model construction module includes a model construction unit and an initial control unit; the model construction unit is used to establish a digital twin model for all alloy production equipment to obtain several alloy production models; based on the several alloy production models, they are clustered according to the cemented carbide production process to establish a virtual cemented carbide production model; the virtual cemented carbide production model includes N combined alloy production equipment clusters, where N is the total number of cemented carbide production processes; the initial control unit is used to obtain historical cemented carbide production data; based on the historical cemented carbide production data, the virtual cemented carbide production model is initially controlled to obtain a controlled cemented carbide processing model;

[0086] The cemented carbide production control module includes a production prediction unit and a production control unit; the production prediction unit is used to obtain the current raw material data to be produced and the current order data to be produced; the process is predicted for the current raw material data to be produced, the current order data to be produced and the control cemented carbide processing model to obtain the predicted finished product quality result; the production control unit is used to obtain the real-time processing quality result of the cemented carbide production process n-1 before the start of the cemented carbide production process n; based on the predicted finished product quality result and the real-time processing quality result, the control cemented carbide processing model is reversely optimized to obtain a new control cemented carbide processing model; the new control cemented carbide processing model is used to control the cemented carbide processing of the cemented carbide production process n; and the above steps are repeated until the current cemented carbide production order is completed.

[0087] It should be understood that those skilled in the art may make improvements or modifications based on the above description, and all such improvements and modifications shall fall within the scope of protection of the appended claims. Any portion of this specification not described in detail is prior art known to those skilled in the art.

Claims

1. A cemented carbide production control method based on digital twin, characterized in that: The following steps are involved: Build digital twin models for all alloy production equipment to obtain several alloy production models. Cluster the cemented carbide production processes based on these alloy production models to establish a virtual cemented carbide production model. The virtual cemented carbide production model includes N clusters of combined alloy production equipment, where N is the total number of cemented carbide production processes. Obtain historical cemented carbide production data. Perform initial control on the virtual cemented carbide production model based on the cemented carbide production history data to obtain a controlled cemented carbide processing model; Obtain the current data of raw materials to be produced and the current data of orders to be produced; The process is predicted based on the current raw material data to be produced, the current order data to be produced and the control carbide processing model to obtain the predicted finished product quality results; The specific steps for predicting the process based on the current raw material data to be produced, the current order data to be produced, and the control of the cemented carbide processing model include: Identify processing quality targets and processing cost constraints based on current pending production order data; generate M candidate alloy processing paths based on the N combined alloy production equipment clusters in the control carbide processing model based on the processing quality targets and processing cost constraints; Extract raw material features based on the data of raw materials to be produced to obtain the characteristics of alloy processing raw materials; build a historical raw material feature association database; use the alloy processing raw material features to match the historical raw material feature association database to obtain a predicted profile of the raw materials to be produced; Taking the predicted image of the raw material to be produced as the initial output, simulate the processing in the controlled cemented carbide processing model marked with M candidate alloy processing paths to obtain the predicted finished product quality result; Before cemented carbide production process n begins, obtain the real-time processing quality results of cemented carbide production process n-1; perform reverse optimization on the control cemented carbide processing model based on the predicted finished product quality results and the real-time processing quality results to obtain a new control cemented carbide processing model; use the new control cemented carbide processing model to control the cemented carbide processing of cemented carbide production process n; repeat the above steps until the current cemented carbide production order is completed; The specific steps for reverse optimization of the cemented carbide machining model based on the predicted finished product quality results and the real-time machining quality results include: Compare and analyze the real-time processing quality results and the predicted finished product quality results to obtain the production result deviation value; Based on the production result deviation value and the preset threshold, a first deviation result or a second deviation result is obtained; If the result is the first deviation, the swarm optimization algorithm is used to calculate and output the compensation parameter scheme from cemented carbide production process n to cemented carbide production process N. Among them, only the compensation parameter scheme of cemented carbide production process n is used to build a new control cemented carbide processing model, and the compensation parameter scheme of cemented carbide production process n+1 to cemented carbide production process N is retained as the optimized initial control parameters; If the result is the second deviation, the process prediction is re-performed based on the real-time processing quality result to obtain a new predicted finished product quality result; based on the new predicted finished product quality result and the real-time processing quality result, the control carbide processing model is reversely optimized until a new control carbide processing model is output; The first deviation result indicates that the comprehensive deviation between the real-time processing quality and the predicted result is still within the set tolerance range; the second deviation result indicates that at least one key quality indicator or comprehensive deviation exceeds the preset threshold.

2. A cemented carbide production control method based on digital twin according to claim 1, characterized in that: The specific steps for initial adjustment of the virtual cemented carbide production model based on historical cemented carbide production data include: Cemented carbide production history data includes equipment operation data, environmental data, quality control data, and product quality data for each cemented carbide production process; obtains loss data for each current alloy production model equipment; Based on the historical data of cemented carbide production, characteristic analysis is performed to obtain the key production-influencing characteristics of each cemented carbide production process; based on the loss data and equipment operation data, characteristic comparative analysis is performed to obtain the equipment loss characteristics corresponding to each cemented carbide production process; An equipment loss compensation relationship is established based on equipment loss characteristics; the virtual cemented carbide production model is regulated by combining the equipment loss compensation relationship and key production influencing characteristics to obtain a regulated cemented carbide processing model.

3. A cemented carbide production control method based on digital twin according to claim 2, characterized in that: The specific steps of simulating machining in the controlled cemented carbide machining model marked with M candidate alloy machining paths include: For each candidate alloy processing path, simulation processing is performed based on the predicted image of the raw materials to be produced to obtain a predicted processing result matrix; based on the predicted processing result matrix, identification is performed to obtain the key bottleneck process; parameter sensitivity analysis is performed on the key bottleneck process to obtain the processing quality risk index; I candidate alloy processing paths with better processing quality risk indexes are extracted, and Monte Carlo simulation is performed using the predicted profile of the raw materials to be produced to obtain I virtual finished product quality result; judgment is made based on all virtual finished product quality results to obtain a predicted finished product quality result.

4. A cemented carbide production control method based on digital twin according to claim 3, characterized in that: The swarm optimization algorithm is the particle swarm optimization algorithm.

5. A cemented carbide production control system based on digital twin, characterized in that: The system is a cemented carbide production control method based on digital twin according to any one of claims 1 to 4, comprising: The twin processing model construction module includes a model construction unit and an initial control unit; the model construction unit is used to establish a digital twin model for all alloy production equipment to obtain several alloy production models; based on the several alloy production models, they are clustered according to the cemented carbide production process to establish a virtual cemented carbide production model; the virtual cemented carbide production model includes N combined alloy production equipment clusters, where N is the total number of cemented carbide production processes; the initial control unit is used to obtain historical cemented carbide production data; based on the historical cemented carbide production data, the virtual cemented carbide production model is initially controlled to obtain a controlled cemented carbide processing model; The cemented carbide production control module includes a production prediction unit and a production control unit; the production prediction unit is used to obtain the current raw material data to be produced and the current order data to be produced; the process is predicted for the current raw material data to be produced, the current order data to be produced and the control cemented carbide processing model to obtain the predicted finished product quality result; the production control unit is used to obtain the real-time processing quality result of the cemented carbide production process n-1 before the start of the cemented carbide production process n; based on the predicted finished product quality result and the real-time processing quality result, the control cemented carbide processing model is reversely optimized to obtain a new control cemented carbide processing model; the new control cemented carbide processing model is used to control the cemented carbide processing of the cemented carbide production process n; and the above steps are repeated until the current cemented carbide production order is completed.

Images