Automatic smelting system based on AI deep learning

Abstract

The invention relates to the technical field of smelting process automatic control in the metallurgical industry, and discloses an automatic smelting system based on AI deep learning. According to the invention, a closed-loop control system composed of the automatic execution unit and the AI decision control unit replaces a traditional high-risk operation mode depending on human experience, so that the safety risk and the labor cost are reduced; according to the system, real-time processing and intelligent decision making are conducted on component data through an AI large model, the alloy adding process depending on artificial experience is replaced, component fluctuation caused by artificial judgment differences is reduced, and the consistency and stability of smelting components are improved; meanwhile, the period from sampling to adding is shortened through automatic process control and dynamic optimization, the production efficiency is improved, and a reliable basis is provided for process optimization through complete recording of data in the whole process.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology for smelting processes in the metallurgical industry, specifically to an automatic smelting system based on AI deep learning. Background Technology

[0002] In the smelting and production of metals such as cast iron and cast steel, precise control of the chemical composition of the molten metal is a crucial step in ensuring the quality of the final product. Currently, this step largely relies on manual operation and experience-based judgment. A typical process flow is as follows: Operators first manually sample the metal near the high-temperature furnace, sending the samples to a laboratory equipped with a spectrometer or carbon-sulfur analyzer for chemical composition analysis. After the laboratory personnel complete the analysis and issue a report, the operators interpret the data based on their experience and estimate the types and quantities of alloys to be added. Finally, they direct an overhead crane or other mechanical devices to add the alloy material into the furnace. This "sampling-analysis-decision-addition" process usually needs to be repeated multiple times until the chemical composition of the molten metal reaches the target range.

[0003] However, existing methods require operators to be exposed to high temperatures and radiation near furnaces for extended periods, facing numerous potential dangers such as burns, heat radiation hazards, and molten metal splashes. The working environment is extremely harsh, posing a serious threat to personnel safety. Furthermore, this position is not only physically demanding but also requires a high level of experience from operators, leading to high labor costs for enterprises. Simultaneously, the entire process is time-consuming, with a long cycle from sampling to completion of addition, limiting smelting production efficiency and increasing energy and time consumption. Moreover, because the entire control loop is highly dependent on human experience, subjective differences exist in the judgment standards of different operators, and even the decisions of the same operator may fluctuate under different circumstances, resulting in insufficient stability of the chemical composition of the smelting results and difficulty in ensuring product consistency. Furthermore, manually recorded data is often discontinuous and incomplete, making it difficult to trace and analyze the correlation between "composition data - addition operation - final result," hindering the data-driven, model-based, and continuous optimization of the production process. Although existing technologies have produced partially automated devices such as robotic arms and automated overhead cranes, these devices are usually "islands of automation" that operate independently and have failed to form a complete intelligent system that integrates real-time perception, intelligent decision-making, execution, and closed-loop feedback.

[0004] Therefore, there is an urgent need in this field for a fully automated melting control system that can overcome the above-mentioned defects, so as to realize unmanned and intelligent closed-loop control of the entire melting process from sampling and analysis to alloy addition decision-making and execution, thereby improving product quality, ensuring production safety, reducing operating costs and achieving continuous self-optimization of the process. To this end, an automated melting system based on AI deep learning is proposed. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an automated melting system based on AI deep learning to solve the problems mentioned in the background.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an automated melting system based on AI deep learning, comprising: The smelting unit includes at least one furnace for melting pig iron and scrap steel raw materials into molten metal. The automated execution unit includes a sampling robotic arm for extracting smelting sample blocks from the smelting unit, a sample preparation robotic arm for grinding and preparing the sample blocks, an overhead crane and alloy addition robotic arm for adding pig iron, scrap steel and various alloys to the smelting unit, and a slag removal robotic arm for performing slag removal operations. The component analysis unit, located in the spectroscopy and CS laboratory, is used to receive the sample processed by the sample preparation robotic arm, automatically analyze its chemical composition, and upload the analysis results to a designated cloud or local database. AI decision and control unit, which includes: The AI ​​big model, as the core analysis engine, accesses the database through cloud API to retrieve real-time and historical chemical composition data. It combines the preset target composition, smelting tonnage, raw material information and target temperature to perform multimodal data fusion and deep learning analysis, and generates decision instructions for the type and amount of alloy addition as well as temperature control instructions. The AI ​​agent, acting as a decision executor, receives decision instructions from the large AI model and transforms them into specific, executable control commands. The process control console communicates with the AI ​​agent, receives the control commands, and automatically drives the corresponding equipment in the automated execution unit to complete alloy addition, sampling, sample preparation, and slag removal actions, and controls the temperature of the smelting unit to achieve heating / heating, thereby forming a complete closed-loop control circuit.

[0007] Preferably, the AI ​​big model is configured to process continuous, multi-batch data streams and dynamically adjust its decision output each time new component analysis data is received; the AI ​​agent is configured to automatically trigger a new round of "sampling-analysis-decision-addition" cycle when the component analysis results do not meet the preset standard, until the components are qualified.

[0008] Preferably, the sampling robotic arm is specifically configured to automatically immerse itself in the molten metal in the furnace at a specific time point during the smelting process or after receiving an instruction from the AI ​​agent, obtain a smelting sample block, and transfer the sample block to a designated sample preparation station.

[0009] Preferably, the overhead crane and the alloy-adding robotic arm are controlled collaboratively by the process control console; wherein, the overhead crane is responsible for hoisting and feeding bulk raw materials, and the alloy-adding robotic arm is responsible for adding small batches of alloy elements in trace amounts according to AI decision instructions.

[0010] Preferably, the system further includes a centralized data recording module for automatically recording and storing all key data throughout the smelting process, including but not limited to: sampling time for each time, composition analysis results, AI decision logs, alloy addition records, and smelting temperature curves, forming a complete and traceable electronic production log.

[0011] Preferably, the AI ​​big model adopts a neural network model based on the Transformer architecture and is jointly trained with historical smelting data and process rules, so that its decision logic conforms to both data patterns and metallurgical principles.

[0012] Preferably, the system is configured to support parallel management of multiple smelting units; the AI ​​decision and control unit independently and concurrently sends control commands to multiple process control consoles according to the production progress and status of each furnace, and schedules multiple sets of automated execution units to work together to realize workshop-level multi-task automated smelting production.

[0013] Preferably, the component analysis unit specifically includes a spectrometer and a carbon-sulfur analyzer, used for quantitative analysis of the chemical composition of the sample.

[0014] Preferably, the process control console is further configured to: automatically trigger a safety alarm and pause the automatic process when the component analysis results fail to meet the preset standard multiple times in a row, or when the decision instructions of the AI ​​big model exceed the preset safety threshold, and switch to waiting for manual intervention.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention replaces the traditional high-risk operation mode that relies on human experience with a closed-loop control system consisting of an automated execution unit and an AI decision control unit, thereby reducing safety risks and labor costs. The system uses a large AI model to process composition data in real time and make intelligent decisions, replacing the alloy addition process that relies on human experience. This reduces composition fluctuations caused by differences in human judgment and improves the consistency and stability of the smelting composition. At the same time, automated process control and dynamic optimization shorten the cycle from sampling to addition, improve production efficiency, and provide a reliable basis for process optimization through complete recording of data throughout the entire process.

[0016] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0017] Figure 1 This is a block diagram of the automated melting system based on AI deep learning of the present invention; Figure 2 This is a flowchart of the automated melting process based on AI deep learning, as described in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 and Figure 2 The present invention discloses an automated melting system based on AI deep learning. The core of this system lies in constructing a fully intelligent closed loop of "perception-decision-execution-learning," and its specific workflow is as follows: Step 1: Task Initialization and Smelting Preparation The entire process begins at the process control console. The operator inputs the complete process target package for this smelting task on this interface. This package includes: the target steel grade (e.g., GCr15 bearing steel), for which the system has a pre-set standard chemical composition target vector, denoted as... The planned weight of molten steel to be melted is denoted as... ,For example =80 tons; and the target tapping temperature, denoted as ,For example These parameters are packaged into a task instruction by the process console and sent to the AI ​​agent.

[0020] As the command center of the entire system, the AI ​​agent immediately performs task analysis and resource scheduling upon receiving a task.

[0021] First, a command is sent to the overhead crane subsystem. Based on the command, the overhead crane uses a magnetic hoist or grab bucket to lift a total of [materials / materials] from the raw material area. Tons of pig iron blocks and scrap steel are fed into the electric arc furnace of the smelting unit according to a pre-set feeding pattern. The status of the entire feeding process, including the type, weight, and order of raw materials, is fed back and recorded in real time in a centralized data recording module. After feeding is completed, the AI ​​agent sends a "start smelting" command to the electric arc furnace control system. The process control console communicates with the furnace control system, and after the command is issued, the electrodes automatically descend to ignite the arc, heating and melting the raw materials in the furnace according to the pre-set initial power curve. At this point, the system officially enters automatic monitoring mode, and real-time temperature, current, and voltage data of the smelting unit are continuously collected and transmitted to the data recording module.

[0022] Step 2: Intelligent Triggering and Automated Sampling Traditional timed sampling has been replaced by the dynamic sampling strategy of this system. When the thermocouple inside the furnace detects that the molten metal temperature has reached a preliminary, empirical sampling point... ,For example At that time, this temperature data refers to the current real-time temperature of the molten pool. The data is transmitted in real time to the AI ​​model. The AI ​​model then retrieves the current energy consumption curve and historical melting rate data from similar smelting processes to make a rapid comprehensive judgment. The model uses an internal state evaluation function... Calculations are performed, and if the current molten pool is determined to be sufficiently uniform and representative, an AI agent proactively sends an "immediate sampling" command to the process control console. This method avoids the ineffectiveness of sampling when the molten pool is not uniform, thus improving the quality of the test data.

[0023] Upon receiving the command, the process control console immediately activates the sampling robotic arm. This robotic arm's end effector is equipped with a sampler. Based on a pre-stored 3D model of the furnace, the robotic arm moves to above the optimal sampling point, then immerses the sampler at a specific angle and speed to approximately 300 mm below the molten pool surface, holding it for 3-5 seconds to allow the molten metal to fill the sampler cavity before quickly lifting it. The retrieved sample can be a rapidly solidified rod-shaped specimen with white iron structure, accurately reflecting the microscopic composition of the molten pool. Subsequently, the robotic arm transfers the high-temperature sample to a specialized fixture on the sample preparation robotic arm via a handover operation.

[0024] Step 3: Sample Preparation and Precise Analysis After receiving the sample block, the robotic arm initiates a standardized sample preparation process. First, an integrated high-powered grinding wheel is used to coarsely grind one end face of the sample block to remove oxide scale and surface defects. Then, a finer-grit grinding wheel is used for fine grinding until a mirror-like metallic surface is exposed. Finally, compressed air may be used to purge and remove all grinding debris. The entire sample preparation process is completed within a sealed protective chamber, ensuring a contamination-free and safe environment. The prepared sample is then placed by the robotic arm onto the spectrometer sample stage of the component analysis unit.

[0025] The spectrometer in the composition analysis unit is automatically triggered. Its excitation torch is aimed at the smooth surface of the sample, generating a high-voltage spark that excites and causes local metal atoms to transition. By analyzing the generated characteristic spectra, the precise content of elements in the molten pool is obtained, forming the current composition vector, denoted as . Simultaneously, the carbon-sulfur analyzer performs a secondary, precise determination of carbon and sulfur in the sample using a combustion method, cross-validating the results with the spectral data. After analysis, the results... Information such as sample ID, analysis timestamp, and furnace number is automatically uploaded to a designated data table in the cloud database via the factory network. This process completely eliminates human error in sample delivery and manual data entry.

[0026] Step 4: AI Core Decision Making and Command Generation This is the intelligent hub of the system. The AI ​​big data model (a Transformer model trained on massive amounts of historical smelting data) continuously monitors database updates through its data interface. Once new composition data from the electric arc furnace is detected... Initiate a decision-making cycle immediately.

[0027] The model first constructs a high-dimensional feature vector. As input:

[0028] in: It is a vector of the total amount of various alloys added in this furnace so far, retrieved from the data recording module. This is the duration of this smelting process. It is the real-time unit price vector of various alloy materials, which is obtained in real time by the process control console through the interface of the enterprise manufacturing execution system or ERP system.

[0029] AI large model right A forward propagation calculation is performed. The output, after passing through a fully connected layer, first yields a set of "theoretical additions". However, a purely data-driven approach may violate the laws of metallurgical physics. Therefore, a metallurgical rule constraint layer is introduced into the system.

[0030] Based on the yield of key elements For example, it is a variable dynamically fine-tuned by an AI model. Its base value comes from a metallurgical handbook, but the model will adjust it according to the current furnace conditions (such as slag basicity and temperature). Ultimately, for a certain alloying element... The amount of decision-making additions Determined by the following formula:

[0031] in: It is the current weight of molten steel estimated by the system based on the input raw materials and the burn-off model; It is the mass fraction of element i in the alloy material to be used in the plan; This is a single safe addition limit set to prevent excessive addition due to model misjudgment; This is the target concentration of element i, which is determined by the target steel grade. It represents the desired mass percentage (%) of element i in the final molten steel, derived from the target composition vector set during initialization. ; This represents the current concentration of element i, which is the actual mass percentage (%) of element i in the current molten pool obtained after this sampling analysis, derived from the current composition vector. .

[0032] Simultaneously, the model will perform multi-objective optimization. It will optimize all components that meet the requirements. Among feasible alloy proportioning schemes, the system calculates the scheme with the lowest total cost. For example, to simultaneously adjust the content of silicon and manganese, the system will search for the optimal combination of various alloys such as ferrosilicon, ferromanganese, and ferromanganese, taking into account not only cost. It also needs to satisfy the final component vector Constraints on all elements are applied, and the final selection is the ratio scheme with the lowest total cost and technical feasibility.

[0033] Ultimately, the AI ​​model outputs two core instructions: an alloy addition decision instruction. And temperature control commands (such as "raise the furnace temperature to 1600°C within 10 minutes and hold it at that temperature").

[0034] After receiving these instructions, the AI ​​agent "translates" them into a low-level command sequence that the equipment can understand, such as: "Command the alloy adding robotic arm to take 12.5 kg of high-carbon ferrochrome from bin A and 8.0 kg of silicon-manganese alloy from bin 5 and put them into the electric arc furnace in sequence"; and "Command the electric arc furnace control system to adjust the temperature setpoint to 1600°C".

[0035] Step 5: Precise Execution and Closed-Loop Feedback After receiving commands from the AI ​​agent, the process control console performs parallel control: It instructs the alloy-adding robotic arm to move to the designated hopper. The hopper outlet is equipped with a high-precision weighing sensor; the robotic arm controls the hopper's vibration to feed micro-amounts until the weighing display shows a certain value. Then the robotic arm moves to the furnace charging port and feeds in the alloy.

[0036] It simultaneously sends a new temperature setpoint to the furnace control system. The furnace control system employs a classic PID control algorithm to adjust the electrode power, and its output... for:

[0037] in The output of the controller; The temperature deviation at time t is the set temperature. With the current measured temperature of the molten pool The difference between them, i.e. This is the core deviation signal that the entire control loop needs to minimize. Real-time adjustments are made to make the furnace temperature approach the set value quickly and without overshoot. , , These are the proportional gain coefficient, integral gain coefficient, and differential gain coefficient, respectively. The integral term of the deviation is used to eliminate the steady-state error of the system, that is, to correct the small, persistent temperature deviations that cannot be overcome by the proportional term by accumulating the sum of historical deviations. This is the differential of the deviation.

[0038] After execution, the system enters an intelligent waiting period. This waiting time... The AI ​​model determines the alloy type, melting point, and current molten pool temperature based on the newly added alloy. The system dynamically predicts stirring conditions to ensure that the alloy is fully melted and diffused evenly. After the waiting period, the AI ​​agent automatically triggers a new closed-loop cycle, that is, returns to the second step to sample, analyze, make decisions and add materials again.

[0039] Step 6: Iteration, endpoint determination, and steel tapping The intelligent closed loop formed by steps four and five above will continue to operate. Each cycle will cause... Towards Taking it one step closer. When a component analysis shows that all elements fall within the target range, and the temperature reaches... At that time, the AI ​​model determined that the smelting was qualified.

[0040] At this point, the AI ​​agent issues the final sequence of commands: first, it instructs the slag-removing robotic arm to move to the side of the furnace, extend the slag-removing plate, and thoroughly remove the slag from the furnace door area; then, it instructs the process control console to illuminate the "melting complete" indicator light and prepare the ladle car for tapping steel. All data throughout the entire process, including each... , Temperature curves and equipment status logs are all captured completely by a centralized data recording module, forming an electronic archive that can be used for quality traceability throughout the entire lifecycle.

[0041] Step 7: Safety Monitoring and Parallel Scheduling Throughout the process, the safety mechanism runs in the background. The system monitors two key risks in real time: first, the risk of process stagnation, if the composition does not improve or continues to deteriorate after N consecutive adjustments (e.g., 3 times); second, the risk of decision-making exceeding limits, if the AI ​​calculates... The preset absolute safety threshold has been exceeded. Once any condition is triggered, the process control console will immediately pause the automated process, activate the highest level of audible and visual safety alarm, and display detailed alarm information on the control interface. This will forcefully return system control to the operator and put the system into a state awaiting manual intervention, thereby avoiding the extreme risks that the automated system may bring.

[0042] In multi-furnace scenarios, the AI ​​decision and control unit acts as the central scheduling brain, simultaneously processing data streams and task queues from multiple furnaces (such as furnace 1 and furnace 2). Embedded within the AI ​​is a task scheduler based on priority rules and resource conflict detection. The scheduler dynamically calculates the priority for each furnace task, with priority factors including: task urgency (e.g., whether it's nearing a delivery date), smelting stage (e.g., refining stage takes precedence over melting stage), and waiting time. When multiple tasks request the same scarce resource (e.g., overhead crane, the same sample preparation robotic arm), the scheduler allocates the resource to the highest-priority task, while other tasks enter a queue to wait, thus achieving efficient parallel production.

Claims

1. An AI deep learning-based automatic smelting system, characterized in that, a smelting unit comprising at least one smelting furnace for melting pig iron, scrap steel raw materials into molten metal; an automatic execution unit comprising a sampling mechanical arm for extracting smelting samples from the smelting unit, a sample preparation mechanical arm for grinding and preparing samples, a crane and alloy adding mechanical arm for adding pig iron, scrap steel and various alloys to the smelting unit, and a slagging mechanical arm for performing slagging operation; a component analysis unit arranged in a spectrum and C-S laboratory for receiving samples processed by the sample preparation mechanical arm and automatically analyzing their chemical components, and uploading the analysis results to a designated cloud or local database; an AI decision and control unit, comprising: an AI large model as a core analysis engine, which accesses the database through a cloud API, calls real-time and historical chemical component data, and combines preset target components, smelting tonnage, raw material information and target temperature for multi-modal data fusion and deep learning analysis to generate alloy addition type and quantity decision instructions and temperature control instructions; an AI agent as a decision executor, which receives the decision instructions of the AI large model and converts them into specific and executable control commands; a process console in communication connection with the AI agent, receiving the control commands and automatically driving the corresponding devices in the automatic execution unit to complete alloy addition, sampling, sample preparation and slagging actions, and controlling the temperature of the smelting unit to realize temperature rising / holding, thereby forming a complete closed-loop control circuit.

2. The automatic smelting system based on AI deep learning according to claim 1, wherein, The AI large model is configured to process continuous and multi-batch data streams and dynamically adjust its decision output after receiving new component analysis data each time; the AI agent is configured to automatically trigger a new round of "sampling-analysis-decision-addition" cycle when the component analysis result does not meet the preset standard until the component is qualified.

3. The automatic smelting system based on AI deep learning according to claim 1, wherein The sampling mechanical arm is specifically configured to automatically immerse in the molten metal in the smelting furnace at a specific time point in the smelting process or after receiving instructions from the AI agent, take smelting samples, and transfer the samples to a designated sample preparation station.

4. The automatic smelting system based on AI deep learning according to claim 1, wherein The crane and alloy adding mechanical arm are controlled by the process console; wherein the crane is responsible for lifting and feeding bulk raw materials, and the alloy adding mechanical arm is responsible for small-batch and high-precision trace addition of alloy elements according to AI decision instructions.

5. The automatic smelting system based on AI deep learning according to claim 1, wherein The system further comprises a centralized data recording module for automatically recording and storing all key data in the entire smelting process, including but not limited to: sampling time, component analysis result, AI decision log, alloy addition record, smelting temperature curve, forming a complete and traceable electronic production log.

6. The automatic smelting system based on AI deep learning according to claim 1, wherein The AI large model adopts a neural network model based on the Transformer architecture and is jointly trained with historical smelting data and process rules, so that its decision logic conforms to both data rules and metallurgical principles.

7. The automatic smelting system based on AI deep learning according to claim 1, wherein The system is configured to support parallel management of multiple smelting units; the AI decision and control unit independently and concurrently sends control instructions to multiple process control consoles according to the production progress and state of each smelting furnace, schedules multiple sets of automatic execution units to work cooperatively, and realizes multi-task automatic smelting production at the workshop level.

8. The automatic smelting system based on AI deep learning according to claim 1, wherein, The component analysis unit specifically includes a spectrum analyzer and a carbon-sulfur analyzer, which are used for quantitative chemical component analysis of the sample.

9. The automatic smelting system based on AI deep learning according to claim 2, wherein, The process control console is further configured to automatically trigger a safety alarm and suspend the automatic process and switch to manual intervention when the component analysis result fails to meet the preset standard for multiple times in succession or the decision instruction of the AI large model exceeds the preset safety threshold.

Images