Silicon carbide smelting whole process data linkage ERP system and management method

Abstract

The invention relates to the technical field of intelligent manufacturing, and discloses a silicon carbide smelting whole process data linkage ERP management method, which comprises the following steps: collecting multi-source data of a silicon carbide smelting whole process in real time; constructing a dynamic tensor model, and updating node features based on a dynamic graph convolutional network; embedding a silicon carbide smelting process equation as a constraint condition into the tensor decomposition process, and solving a tensor factor matrix through a constraint optimization algorithm; performing incremental tensor decomposition on the newly added data slices, and updating real-time data change; dynamically adjusting weight coefficients of the process target and the business target according to real-time business requirements; and issuing the optimization instruction to production equipment. By constructing a four-dimensional dynamic tensor model, multi-dimensional data such as equipment operation states, process parameters and business requirements are uniformly mapped into structured feature representation, the problem of local optimization caused by data islands in a traditional method is solved, and high-precision data support is provided for global decision-making in the silicon carbide smelting process.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and in particular to an ERP system and management method for data linkage throughout the entire silicon carbide smelting process. Background Art

[0002] In the field of silicon carbide smelting, traditional data management methods generally suffer from insufficient ability to fuse multi-source heterogeneous data. Existing technologies typically utilize independent data acquisition and processing modules, performing local optimization for equipment operating parameters, process control indicators, and business demand data. This results in a lack of dynamic correlation between data at the equipment, process, and business levels, making it difficult to achieve collaborative decision-making throughout the entire process. Due to the failure to establish a unified, cross-dimensional data model, real-time state perception and optimization instruction generation for the smelting process often rely on empirical rules or offline simulation, which cannot effectively respond to high-frequency, dynamically changing production environments.

[0003] Existing optimization methods often focus on a single objective (such as minimizing energy consumption or maximizing output) and often simplify process constraints to fixed threshold limits, ignoring the nonlinear coupling relationships between key parameters such as temperature and carbon-silicon ratio during the silicon carbide smelting process. This simplification can easily lead to optimization results that deviate from actual physical laws and even pose safety risks to the equipment. Furthermore, traditional tensor decomposition methods rely on recalculating the entire data set, making it difficult to efficiently process newly added data in real time. This can cause optimization instructions to lag, impacting the timeliness of production scheduling.

[0004] When it comes to adapting to business needs, existing technologies typically employ static weight allocation strategies, failing to adjust optimization target priorities based on dynamic factors like real-time electricity price fluctuations and changes in order urgency. This leads to conflicts between process optimization and business needs. Furthermore, the lack of a closed-loop feedback mechanism makes it difficult to promptly correct equipment execution deviations, and the lack of self-healing capabilities under abnormal operating conditions hinders the long-term stability and reliability of the silicon carbide smelting system. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides an ERP system and management method for the entire process of silicon carbide smelting data linkage, which solves the problems of global collaborative decision-making difficulties and dynamic control inaccuracy caused by multi-source data islands, rigid process constraints, real-time optimization lags and multi-objective conflicts in the silicon carbide smelting process.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a data linkage ERP management method for the entire silicon carbide smelting process, comprising the following steps: Real-time collection of multi-source data from the entire silicon carbide smelting process, including equipment-level data, process-level data, and business-level data; Constructing a dynamic tensor model to map the multi-source data into a four-dimensional tensor containing time, equipment, process, and business dimensions, and updating node features based on a dynamic graph convolutional network; Embedding the silicon carbide smelting process equation as a constraint condition into the tensor decomposition process, and solving the tensor factor matrix that satisfies the process constraint condition by a constrained optimization algorithm; Perform incremental tensor decomposition on newly added data slices and update the local factor matrix to adapt to real-time data changes; Dynamically adjust the weight coefficients of process objectives and business objectives according to real-time business needs to generate multi-objective optimization instructions; The optimization instructions are sent to the production equipment to complete the closed-loop control of the entire process from data collection to production execution.

[0007] Preferably, the equipment layer data includes the temperature value collected by the furnace temperature sensor, the current intensity value collected by the current sensor, and the raw material quality recorded by the raw material weighing equipment; the process layer data includes the carbon-silicon ratio parameter and the crystallization time parameter; the business layer data includes the order urgency level, inventory level and real-time electricity price.

[0008] Preferably, the data preprocessing includes: Outlier filtering: Eliminate invalid data based on the preset effective range of process parameters, including furnace temperature range, carbon-silicon ratio range, and crystallization time range; Unit standardization: standardize the unit of current intensity to ampere and the unit of raw material mass to kilogram.

[0009] Preferably, the dimension of the four-dimensional tensor is defined as: Time dimension: continuous division by millisecond time slices; Equipment dimension: coded as unique identifiers of smelting furnaces, sensors, and sorting equipment; Process dimensions: including temperature, current, carbon-silicon ratio and crystallization time; Business dimensions: including order urgency, inventory levels, and real-time electricity prices.

[0010] Preferably, the adjacency matrix updating method of the dynamic graph convolutional network is: ; in: Representation node In time slice The eigenvector of is the bandwidth parameter of the Gaussian kernel function, which is used to control the decay rate of node similarity.

[0011] Preferably, the process equation is a quadratic function relationship equation of furnace efficiency, temperature, and carbon-silicon ratio, and its expression is: ; in is the process coefficient calibrated by historical smelting data, and the temperature Need to meet the preset process temperature range constraints .

[0012] Preferably, the constrained optimization algorithm is an alternating direction multiplier method, comprising the following steps: Convert the process equation constraints into auxiliary variables and decompose the original tensor decomposition problem into unconstrained subproblems; Alternately update the original variables and auxiliary variables to ensure that the tensor reconstruction error is minimized and the process equation constraints are satisfied; The deviation between the original variables and the auxiliary variables is dynamically balanced by Lagrange multipliers until convergence.

[0013] Preferably, the incremental tensor decomposition includes: Only the time dimension factor matrix of the four-dimensional tensor is updated for the newly added time slice data, and the equipment, process and business dimension factor matrices remain unchanged; The incremental update of the time dimension factor matrix is calculated by stochastic gradient descent method, so that the tensor reconstruction error of the newly added data slice is minimized.

[0014] Preferably, the weight coefficient adjustment method is: Process target weight Real-time electricity prices Positive correlation; Business goal weight and order urgency level Positive correlation; in , and the weight value is updated in real time.

[0015] A data-linked ERP system for the entire silicon carbide smelting process, including: Multi-source data acquisition module, used to collect data from the equipment layer, process layer and business layer in real time, and transmit the collected raw data to the dynamic tensor modeling module; A dynamic tensor modeling module is connected to the multi-source data acquisition module, receives the raw data and constructs a four-dimensional dynamic tensor model, and transmits the model output to the process constraint optimization module; A process constraint optimization module is connected to the dynamic tensor modeling module, embeds the silicon carbide smelting process equation as a constraint into the tensor decomposition process, and outputs the optimized tensor factor matrix to the incremental calculation module; An incremental calculation module, connected to the process constraint optimization module, updates the local factor matrix for the newly added data slices, and feeds the updated factor matrix back to the dynamic tensor modeling module in real time to update the model; A multi-objective decision-making module is connected to the incremental calculation module and the multi-source data acquisition module, receives real-time business data, including order urgency, real-time electricity prices, and updated factor matrices, dynamically adjusts the weight coefficients of process and business objectives, generates optimization instructions, and sends them to the instruction execution module; The instruction execution module is connected to the multi-objective decision-making module, and sends the optimization instructions to the production equipment through the industrial communication protocol, while returning the equipment execution status data to the multi-source data acquisition module to form a closed-loop control link.

[0016] The present invention provides an ERP system and management method for data linkage throughout the entire silicon carbide smelting process. This system has the following beneficial effects: 1. This invention constructs a four-dimensional dynamic tensor model to uniformly map multi-dimensional data such as equipment operating status, process parameters and business requirements into a structured feature representation, solving the local optimization problem caused by data silos in traditional methods and providing high-precision data support for global decision-making in the silicon carbide smelting process.

[0017] 2. The present invention embeds the physical laws of silicon carbide smelting as hard constraints into the tensor decomposition process to ensure that the optimization results conform to both the data distribution laws and the actual production process limitations, avoiding the non-physical feasible solution problem that may be caused by a purely data-driven model and significantly improving process compliance.

[0018] 3. The present invention adopts incremental tensor decomposition technology, and only updates the local factor matrix for the newly added data slices, avoiding the consumption of computing resources caused by recalculating the entire data, supporting millisecond-level response to dynamic changes, and meeting the stringent requirements of the silicon carbide smelting process for real-time optimization.

[0019] 4. The present invention dynamically adjusts the weights of multiple objectives based on real-time electricity prices and order urgency, achieving flexible adaptation of energy consumption optimization and delivery timeliness, solving the rigidity problem of traditional fixed weight strategies in complex production scenarios, and improving the economy and agility of production scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the method of the present invention; Figure 2 is a schematic diagram of the system level of the present invention; Figure 3 Schematic diagram of data flow and closed-loop link of the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] Example 1: Please see the attached Figure 1 -Attached Figure 3 The embodiment of the present invention provides an ERP management method for data linkage of the entire silicon carbide smelting process, including: In this embodiment, the data linkage management method for the entire silicon carbide smelting process collects and preprocesses multi-source heterogeneous data, providing standardized input for the dynamic tensor modeling module and ensuring data compatibility with the subsequent process constraint optimization module. The data collection and preprocessing process is described in detail below, combined with specific implementations.

[0023] In this embodiment, equipment-level data is acquired in real time via data acquisition units deployed on the smelting furnace, sensors, and sorting equipment. For example, the furnace temperature sensor utilizes a thermocouple array, positioned at varying heights on the furnace wall, collecting temperature distribution data at a millisecond sampling rate. Current sensors, integrated into the power busbar, measure current intensity waveforms using the Hall effect principle. Raw material weighing equipment, installed at the end of the batching conveyor, uses high-precision strain gauges to record the mass of carbon and silicon powder feeds.

[0024] The process layer data is generated by the real-time database analysis of the process control system. For example, the carbon-silicon ratio parameter is dynamically calculated through the raw material weighing data. The specific formula is: ; in Indicates the quality of silicon powder, Indicates the mass of carbon powder in kilograms (kg). The crystallization time parameter is obtained through the furnace temperature threshold trigger mechanism. When the temperature in the furnace reaches the preset crystallization temperature The timer starts when the sorting machine starts and the timing ends when the start signal is triggered.

[0025] Business-level data is synchronized with external power grid data interfaces through the enterprise resource planning system interface. For example, order urgency is dynamically categorized into three levels (urgent, high, and normal) based on the number of days remaining on the delivery deadline. Inventory data is obtained through the real-time inventory table of the warehouse management system, and real-time electricity price data is updated on a minute-by-minute basis through the application programming interface (API) of the power grid open platform.

[0026] In some embodiments, outlier filtering is performed based on a dynamic threshold interval defined in a process knowledge base. For example, if the furnace temperature data exceeds the process allowable range, When the current intensity data exceeds the rated working current continuously, the data elimination rule is triggered; 120%, it is marked as abnormal and replaced by the sliding average of the previous time series window. Exceeding the preset reasonable range When the raw material feeding record is combined, a logic check is performed. If the check fails, the interpolation compensation algorithm is activated. In some embodiments, the unit unification processing is achieved by pre-set conversion rules. For example, if the current intensity limit data is in kiloamperes (kA), the formula is: ; in, : current intensity (ampere); : original current data (kA); Convert to ampere (A); if the original data of raw material mass is in tons (t), use the formula: ; in, : Raw material mass (kg); : original mass data (tons); The converted data is stored in a time series database and mapped with the equipment unique identifier, process parameter label, and business dimension metadata.

[0027] The preprocessed data is transmitted to the central data storage module via an industrial communication protocol. This module utilizes a distributed time-series database architecture, sharding data by millisecond timestamp and establishing a time-aligned index with the dynamic tensor modeling module. For example, each time-slice data packet contains a device identifier, a process parameter vector, and a business dimension label. Its data structure strictly corresponds to the time, device, process, and business dimensions of the subsequent four-dimensional tensor.

[0028] This step addresses the heterogeneity of data at the equipment, process, and business layers through multi-source data collection and preprocessing, providing standardized input for subsequent feature updates in the dynamic graph convolutional network. For example, millisecond-level timestamp indexing directly supports the time dimension slicing requirements of the dynamic tensor modeling module; device unique identifier mapping ensures consistency between node feature vectors and the device dimension encoding of the graph convolutional network; and the pre-association of process parameters with business labels provides real-time decision-making support for the multi-objective optimization instruction generation module.

[0029] In this embodiment, a data linkage management method for the entire silicon carbide smelting process uses dynamic tensor modeling and feature updating to map preprocessed multi-source data into a four-dimensional tensor structure. This method also uses a dynamic graph convolutional network to capture the dynamic relationships between equipment, process, and business dimensions, providing structured input for the subsequent process constraint optimization module. The dynamic tensor modeling process is described in detail below, combined with specific implementations.

[0030] In this embodiment, the construction of the four-dimensional tensor is based on the multi-source data pre-processed in step S1. Specifically, the dimensions of the four-dimensional tensor are defined as time, equipment, process, and business dimensions. For example, the time dimension is divided into millisecond time slices, each of which contains synchronized data on the equipment's operating status, process parameters, and business tags; the equipment dimension is encoded as a unique identifier for the smelting furnace, sensor, and sorting equipment, which is used to distinguish the data sources of different physical entities; the process dimension includes temperature, current intensity, carbon-silicon ratio, and crystallization time parameters; and the business dimension includes order urgency level, inventory level, and real-time electricity price tags.

[0031] The four-dimensional tensor is initialized by pre-populating historical data. For example, during the first run, the past 24 hours of smelting data are loaded, the initial tensor structure is constructed in time slice order, and a mapping between device identifiers and tensor device dimensions is established. When adding data dynamically, the pre-processed real-time data is aligned by timestamp and appended to the end of the tensor's time dimension.

[0032] The adjacency matrix update method of dynamic graph convolutional network uses Gaussian kernel function to measure the similarity between nodes. Specifically, the adjacency matrix elements The calculation formula is: ; in: Representation node In time slice The characteristic vector of is formed by normalizing the process parameters corresponding to the equipment dimensions; is the bandwidth parameter of the Gaussian kernel function, which is used to control the decay rate of node similarity. Its value is positively correlated with the standard deviation of the eigenvector. represents the Euclidean distance.

[0033] The adjacency matrix is updated at the same frequency as the time slice division. For example, after receiving new time slice data every 100 milliseconds, the adjacency matrix is recalculated based on the node feature vectors and the topology of the dynamic graph convolutional network is updated.

[0034] Node features aggregate information of adjacent nodes through graph convolution operations. In general, the node feature matrix The update formula is: ; in: Time slice The node feature matrix of is a trainable weight matrix used for linear transformation of feature space; is the bias vector; ReLU is an activation function used to introduce nonlinear expression capabilities. The weight matrix With the bias vector Initialized through pre-training with historical data. During dynamic updates, a sliding window mechanism is used to retain the recent feature matrix, for example, retaining the feature states of the last 10 time slices to support incremental learning.

[0035] This step converts the normalized data provided in step S1 into a structured feature representation through four-dimensional tensor modeling and dynamic graph convolutional network updates. For example, the millisecond-level division of the time dimension is strictly aligned with the timestamp index of the preprocessing module to ensure data temporal consistency; the unique identifier mapping of the equipment dimension supports the correspondence between nodes and physical devices in the dynamic graph convolutional network; and the parameter labels of the process and business dimensions provide a decision basis for the subsequent multi-objective optimization instruction generation module. The real-time update mechanism of the dynamic adjacency matrix can capture the dynamic correlation of equipment operating status. For example, when the temperature of a smelting furnace rises abnormally, the feature aggregation weights of its adjacent nodes (such as associated sensors) are automatically adjusted, thereby enhancing the propagation sensitivity of the abnormal state.

[0036] In this embodiment, a data-linked management method for the entire silicon carbide smelting process utilizes process constraint embedding and tensor decomposition to transform the physical laws of the smelting process into mathematical constraints. These constraints are then integrated with a dynamic tensor model to ensure that the optimization results meet actual process constraints. The following describes the process equation definition and constraint optimization algorithm in detail, combined with specific implementations.

[0037] In this embodiment, the process equation is a nonlinear relationship equation between furnace efficiency, temperature, and vibration-to-silicon ratio, which is constructed based on the process dimension parameters in the dynamic tensor model provided in step S2. Specifically, the mathematical expression of the equation is: ; in: represents furnace efficiency (dimensionless), which is defined as the silicon carbide output per unit energy consumption; is the furnace temperature (unit: °C), taken from the process dimension of the dynamic tensor model; is the carbon-silicon ratio (dimensionless), which is calculated in real time using the preprocessed data from step S1; is the process coefficient, which is calibrated through historical smelting data, for example, by using the least squares fitting method; and The lower and upper limits of the temperature allowed by the process, such as 800°C and 2500°C.

[0038] The calibration process of the process equation includes the following steps: extracting temperature, carbon-silicon ratio and corresponding furnace efficiency data from historical data, building a sample set , solve the coefficients through nonlinear regression , and verify the goodness of fit of the equation in the preset temperature range.

[0039] The constrained optimization algorithm uses the alternating direction method of multipliers (ADMM) to embed the process equation constraints into the tensor decomposition process. Specifically, it includes the following steps: Variable decomposition: Convert the original tensor decomposition problem into a constrained optimization problem. Define the original variables (Process dimension factor matrix) and auxiliary variables , so that the optimization objective is: ; in: represents the process dimension factor matrix; represents the auxiliary variable matrix; Represents the process equation constraint, that is .

[0040] Alternate update: Original variable update: fixed auxiliary variables , solve the unconstrained tensor decomposition subproblem, update To minimize the reconstruction error; Auxiliary variable update: fixed , solve the constrained subproblem, and update Make it satisfy the process equation; Multiplier adjustment: through Lagrange multipliers Dynamically balance the deviation between the original variable and the auxiliary variable, and the update formula is: ; in is the penalty factor that controls the weight of constraint violation.

[0041] Convergence judgment: When the original residual and the dual residual The iteration is terminated when both are less than a preset threshold. In some embodiments, the auxiliary variable The update is achieved through numerical optimization algorithms. For example, for nonlinear constraints , using the Newton-Raphson method to solve A feasible solution to ensure the temperature exist Within the range.

[0042] This step deeply integrates the dynamic tensor model generated in step S2 with the physical laws of smelting by combining the process equation constraints with the ADMM algorithm. For example, the process dimension factor matrix The update process considers both data reconstruction error and process compliance, ensuring that the decomposition results are consistent with both data distribution and actual production constraints. Compared with carbon silicon As the input parameter of the process equation, it directly relates to the pre-processed data of step S1 and the incremental decomposition requirements of step S4. In addition, the Lagrange multiplier The adjustment mechanism provides quantitative feedback of constraint deviations for subsequent multi-objective optimization modules and supports dynamic adaptation of weight coefficients.

[0043] In this embodiment, the data linkage management method for the entire silicon carbide smelting process uses incremental tensor decomposition to efficiently update the local factor matrix for newly added data slices, ensuring real-time model performance while reducing the consumption of full computing resources. The following describes the incremental calculation process in detail with reference to specific implementation methods.

[0044] In this embodiment, the trigger conditions for incremental tensor decomposition are the optimized tensor factor matrix output from step S3 and the arrival of new time slice data. Specifically, after the preprocessing module (step S1) transmits the new time slice data to the dynamic tensor modeling module (step S2), and the process constraint optimization module (step S3) completes the factor matrix optimization for the current time slice, the incremental calculation module automatically initiates the local update process.

[0045] The newly added data slice is the latest extension of the dynamic tensor model in the time dimension. For example, a time slice data packet is generated every 100 milliseconds, containing a synchronized record of the equipment status, process parameters, and business tags during that period.

[0046] Incremental tensor decomposition only updates the time dimension factor matrix , and the factor matrix of equipment, process and business dimensions , , Specifically, the objective function is defined as minimizing the tensor reconstruction error of the newly added data slice: ; in: Represents the four-dimensional tensor slice corresponding to the newly added time slice data; is the factor matrix of the previous time slice; is the incremental update amount of the time dimension factor matrix; Represents a tensor shrinking operation.

[0047] Incremental update amount Calculated by stochastic gradient descent (SGD). The specific process includes: initialization is a zero matrix, and the gradient is calculated iteratively , and according to the learning rate renew: ; Represents the learning rate (scalar), which controls the update step size; Represents the gradient of the objective function with respect to the increment matrix; Until the gradient change is less than the preset threshold or the maximum number of iterations is reached.

[0048] This step dynamically merges the optimized factor matrix output from step S3 with the newly added data by updating the local factor matrix, thus avoiding the computational overhead of full tensor decomposition. For example, the device dimension factor matrix The fixedness is based on the assumption that the equipment topology remains unchanged in the short term; the process dimension factor matrix The stability of depends on the process constraint embedding mechanism in step S3; Business dimension factor matrix The persistence of is determined by the low frequency of business tag updates. The updated time dimension factor matrix Real-time feedback is provided to the dynamic tensor modeling module (step S2) to ensure that the model continues to adapt to the latest data distribution and provide the latest feature representation for the multi-objective decision module (step S5).

[0049] In this embodiment, a data linkage management method for the entire silicon carbide smelting process achieves closed-loop data control by dynamically adjusting the weights of multiple objectives, balancing process optimization goals with business needs in real time. This method generates executable instructions to drive production equipment. The following describes the weight coefficient adjustment and optimization instruction generation process in detail, combined with specific implementation methods.

[0050] In this embodiment, the weight coefficient adjustment is based on the factor matrix updated in step S4 and the real-time business data collected in step S1. Specifically, the process target weight Real-time electricity prices Positive correlation, business goal weight and order urgency level Positive correlation and satisfying normalization constraints The calculation formula is: ; in: Indicates time slice The real-time electricity price (unit: yuan / kWh) is obtained from the business layer data interface in step S1; Indicates the order urgency level (dimensionless), which is dynamically mapped to a numerical label from 1 to 3 based on the delivery deadline.

[0051] Real-time electricity prices and order urgency Normalization is required. For example, scaling electricity price data to the [0,1] interval: ; in and The lower and upper limits of the preset electricity price are, for example, 0.5 yuan / kWh and 1.2 yuan / kWh. Similarly, the order urgency level Normalization is achieved by directly dividing by the maximum level value 3.

[0052] Optimization instruction generation is achieved through a weighted multi-objective programming model. Specifically, the objective function is defined as the weighted sum of process objectives and business objectives: ; in: is the control variable vector, including the furnace temperature setting value, carbon-silicon ratio adjustment value and separator priority parameters; is the process objective function, such as maximizing furnace efficiency or minimizing energy consumption; is the business objective function, such as minimizing order delivery time or minimizing inventory costs.

[0053] Control variables For example, the furnace temperature setting value must meet , the carbon-silicon ratio adjustment amount needs to be Within the range. Description of the relevance of technical solutions: This step uses a dynamic weight adjustment mechanism to deeply integrate the latest factor matrix output by the incremental calculation module in step S4 with the real-time business data in step S1 to generate optimization instructions that adapt to the current production environment. For example, the process objective function The construction depends on the furnace efficiency model provided by the process constraint optimization module in step S3; the business objective function The order delivery time is calculated based on the inventory and sorting equipment status data in step S1. Weight coefficient and Dynamic changes directly respond to electricity price fluctuations and order priority adjustments, ensuring that optimization instructions meet both process compliance and business economic requirements.

[0054] In this embodiment, the data linkage management method for the entire silicon carbide smelting process uses a closed-loop control system for command execution and status feedback. The multi-objective optimization command generated in step S5 is issued to production equipment, and equipment execution status data is collected in real time, forming a complete closed loop of "decision-execution-feedback." The following describes the command execution and feedback mechanism in detail, in conjunction with specific implementations.

[0055] In this embodiment, optimization instructions are transmitted to the production equipment control unit via an industrial communication protocol. Specifically, these instructions include adjusting the furnace temperature setpoint, correcting the carbon-to-silicon ratio, and setting sorter priorities. For example, the furnace temperature setpoint instruction is converted into a 4-20mA analog signal via the OPC UA protocol to drive the heater power regulation module; the sorter priority parameter is written to the equipment controller register via the Modbus TCP protocol, triggering an update to the sorting logic.

[0056] The instruction issuance process involves protocol adaptation and signal conversion. For example, when the optimization instruction includes discrete control parameters (such as sorting machine priority), digital signal transmission is used; when it involves continuously adjustable parameters (such as furnace temperature setpoint), analog signals or pulse width modulation (PWM) signal output are used.

[0057] Equipment execution status data is transmitted in real time via the sensor network to the multi-source data acquisition module in step S1. For example, actual furnace temperature data is collected by thermocouples, converted to analog-to-digital (A / D) and uploaded to the central data storage module. The sorting machine's operating status (e.g., sorting speed, fault codes) is periodically read via the equipment controller's status register. In some embodiments, the deviation between feedback data and the original optimization instructions is calculated using the following formula: ; in: The furnace temperature setting value issued in step S5 (unit: °C); The actual furnace temperature measurement value fed back by the equipment (unit: ℃); It is the absolute value of temperature deviation, which is used to trigger abnormal alarm or re-optimization.

[0058] Exception handling and closed-loop verification: In one possible implementation, when When the preset threshold is exceeded (for example, 5% of the set value), the following actions are triggered: Sending a re-optimization request to the multi-objective decision module (step S5) with the current device status data; Suspend the instruction issuing channel until new optimization instructions are generated; Record abnormal events to the process knowledge base for subsequent model training.

[0059] Closed-loop verification is achieved through a timestamp alignment mechanism. For example, the timestamp of the feedback data strictly matches the timestamp of the corresponding instruction, ensuring that the causal relationship between the state data and the instruction is traceable.

[0060] This step implements the optimization decision results of step S5 as equipment control actions through a closed-loop link between instruction execution and state feedback. The execution results are then fed back to the data acquisition module (step S1), forming a data linkage for the entire process. For example, the issuance of the furnace temperature setpoint depends on the temperature safety range provided by the process constraint optimization module in step S3; the sorting machine priority adjustment is based on the business objective weight calculation results in step S5. The real-time transmission of feedback data provides the latest input to the dynamic tensor modeling module (step S2), supporting continuous iterative optimization of the model. The anomaly detection mechanism, through deviation threshold determination, is directly linked to the incremental calculation module in step S4, triggering local model updates to adapt to changes in equipment status.

[0061] Example 2: An embodiment of the present invention provides an ERP system for data linkage of the entire silicon carbide smelting process, including: Multi-source data acquisition module: used to collect data from the equipment layer, process layer and business layer in real time.

[0062] Equipment layer data: Parameters such as furnace temperature, current intensity, and raw material quality are obtained through temperature sensors, current sensors, and weighing equipment.

[0063] Process layer data: Extract process parameters such as carbon-silicon ratio and crystallization time from the process control system.

[0064] Business layer data: Obtain order urgency level, inventory level and real-time electricity price through ERP interface.

[0065] Data preprocessing: This includes outlier filtering (based on process parameter thresholds) and unit normalization (converting current units to amperes and mass units to kilograms).

[0066] Dynamic tensor modeling module: Connected to the multi-source data acquisition module, it receives preprocessed data and constructs a four-dimensional dynamic tensor.

[0067] Four-dimensional tensor structure: time dimension (millisecond-level slicing), equipment dimension (equipment unique identifier), process dimension (temperature, current, carbon-silicon ratio, crystallization time), and business dimension (order urgency, inventory level, electricity price).

[0068] Dynamic graph convolutional network update: The adjacency matrix calculates node similarity through the Gaussian kernel function, the formula is: ; in is the node feature vector, is the bandwidth parameter.

[0069] Process constraint optimization module: Connected with the dynamic tensor modeling module, it embeds the silicon carbide smelting process equation as a constraint into the tensor decomposition.

[0070] Process equation definition: furnace efficiency and temperature , carbon-silicon ratio The quadratic relationship equation is: ; Constrained optimization algorithm: The alternating direction method of multipliers (ADMM) is used to decompose the problem into unconstrained subproblems and constraint satisfaction subproblems, and the original variables and auxiliary variables are updated alternately.

[0071] Incremental calculation module: Connected with the process constraint optimization module to update the time dimension factor matrix for the newly added data slices.

[0072] Local update strategy: only update the time dimension factor matrix , other dimension factors are fixed.

[0073] Stochastic Gradient Descent (SGD): The objective function is to minimize the reconstruction error of the newly added data slices. The formula is: ; Connect with the incremental calculation module and multi-source data acquisition module to dynamically adjust the weights of process and business objectives.

[0074] Weight calculation rules: process weight Positively correlated with real-time electricity prices, business weight Positively correlated with order urgency, the formula is: ; , : weight coefficient; : Real-time electricity prices; :Order urgency level Multi-objective optimization instruction generation: objective function is in the form of weighted sum The quantities include furnace temperature setting value and carbon-silicon ratio adjustment.

[0075] Instruction execution module: Connected with the multi-objective decision-making module, it sends optimization instructions to production equipment and provides feedback on the execution status.

[0076] Command protocol adaptation: Convert commands into control signals through industrial protocols such as OPC UA and Modbus.

[0077] State feedback mechanism: collects actual device temperature, current and other data and calculates deviations ,Re-optimization is triggered when the threshold is exceeded.

[0078] System operation process: Data acquisition and modeling: After preprocessing, multi-source data is input into the dynamic tensor modeling module to construct a four-dimensional tensor and update the node features of the dynamic graph convolutional network.

[0079] Process constraint fusion: The process equation is embedded in the tensor decomposition process as a hard constraint, and the compliance factor matrix is solved using the ADMM algorithm.

[0080] Incremental update and decision-making: New data triggers the update of the local factor matrix, and the multi-objective decision-making module generates optimization instructions based on real-time business data.

[0081] Closed-loop control and feedback: Instructions are sent to the device for execution, and actual status data is transmitted back to the acquisition module. When the deviation exceeds the limit, it is re-optimized to form a closed loop.

[0082] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A data-linked ERP management method for the entire silicon carbide smelting process, characterized in that: The following steps are involved: Real-time collection of multi-source data from the entire silicon carbide smelting process, including equipment-level data, process-level data, and business-level data; Constructing a dynamic tensor model to map the multi-source data into a four-dimensional tensor containing time, equipment, process, and business dimensions, and updating node features based on a dynamic graph convolutional network; Embedding the silicon carbide smelting process equation as a constraint condition into the tensor decomposition process, and solving the tensor factor matrix that satisfies the process constraint condition by a constrained optimization algorithm; Perform incremental tensor decomposition on newly added data slices and update the local factor matrix to adapt to real-time data changes; Dynamically adjust the weight coefficients of process objectives and business objectives according to real-time business needs to generate multi-objective optimization instructions; The optimization instructions are sent to the production equipment to complete the closed-loop control of the entire process from data collection to production execution.

2. The method for data linkage ERP management of the entire silicon carbide smelting process according to claim 1 is characterized in that: The equipment layer data includes the temperature value collected by the furnace temperature sensor, the current intensity value collected by the current sensor, and the raw material quality recorded by the raw material weighing equipment; the process layer data includes the carbon-silicon ratio parameter and the crystallization time parameter; the business layer data includes the order urgency level, inventory level and real-time electricity price.

3. The method for data linkage ERP management of the entire silicon carbide smelting process according to claim 1 is characterized in that: The data preprocessing includes: Outlier filtering: Eliminate invalid data based on the preset effective range of process parameters, including furnace temperature range, carbon-silicon ratio range, and crystallization time range; Unit standardization: standardize the unit of current intensity to ampere and the unit of raw material mass to kilogram.

4. The method for data linkage ERP management of the entire silicon carbide smelting process according to claim 1 is characterized in that: The dimensions of the four-dimensional tensor are defined as: Time dimension: continuous division by millisecond time slices; Equipment dimension: coded as unique identifiers of smelting furnaces, sensors, and sorting equipment; Process dimensions: including temperature, current, carbon-silicon ratio and crystallization time; Business dimensions: including order urgency, inventory levels, and real-time electricity prices.

5. The method for data linkage ERP management of the entire silicon carbide smelting process according to claim 1 is characterized in that: The adjacency matrix update method of the dynamic graph convolutional network is: ； in: Representation node In time slice The eigenvector of is the bandwidth parameter of the Gaussian kernel function, which is used to control the decay rate of node similarity.

6. The method for data linkage ERP management of the entire silicon carbide smelting process according to claim 1 is characterized in that: The process equation is a quadratic function relationship equation of furnace efficiency, temperature, and carbon-silicon ratio, and its expression is: ； in is the process coefficient calibrated by historical smelting data, and the temperature Need to meet the preset process temperature range constraints .

7. The method for data linkage ERP management of the entire silicon carbide smelting process according to claim 1 is characterized in that: The constrained optimization algorithm is an alternating direction multiplier method, which includes the following steps: Convert the process equation constraints into auxiliary variables and decompose the original tensor decomposition problem into unconstrained subproblems; Alternately update the original variables and auxiliary variables to ensure that the tensor reconstruction error is minimized and the process equation constraints are satisfied; The deviation between the original variables and the auxiliary variables is dynamically balanced by Lagrange multipliers until convergence.

8. The method for data linkage ERP management of the entire silicon carbide smelting process according to claim 1 is characterized in that: The incremental tensor decomposition includes: Only the time dimension factor matrix of the four-dimensional tensor is updated for the newly added time slice data, and the equipment, process and business dimension factor matrices remain unchanged; The incremental update of the time dimension factor matrix is calculated by stochastic gradient descent method, so that the tensor reconstruction error of the newly added data slice is minimized.

9. The method for data linkage ERP management of the entire silicon carbide smelting process according to claim 1, characterized in that: The weight coefficient adjustment method is: Process target weight Real-time electricity prices Positive correlation; Business goal weight and order urgency level Positive correlation; in , and the weight value is updated in real time.

10. A data-linked ERP system for the entire silicon carbide smelting process, according to a data-linked ERP management method for the entire silicon carbide smelting process according to any one of claims 1 to 9, characterized in that: include: Multi-source data acquisition module, used to collect data from the equipment layer, process layer and business layer in real time, and transmit the collected raw data to the dynamic tensor modeling module; A dynamic tensor modeling module is connected to the multi-source data acquisition module, receives the raw data and constructs a four-dimensional dynamic tensor model, and transmits the model output to the process constraint optimization module; A process constraint optimization module is connected to the dynamic tensor modeling module, embeds the silicon carbide smelting process equation as a constraint into the tensor decomposition process, and outputs the optimized tensor factor matrix to the incremental calculation module; An incremental calculation module, connected to the process constraint optimization module, updates the local factor matrix for the newly added data slices, and feeds the updated factor matrix back to the dynamic tensor modeling module in real time to update the model; A multi-objective decision-making module is connected to the incremental calculation module and the multi-source data acquisition module, receives real-time business data, including order urgency, real-time electricity prices, and updated factor matrices, dynamically adjusts the weight coefficients of process and business objectives, generates optimization instructions, and sends them to the instruction execution module; The instruction execution module is connected to the multi-objective decision-making module, and sends the optimization instructions to the production equipment through the industrial communication protocol, while returning the equipment execution status data to the multi-source data acquisition module to form a closed-loop control link.

Images