Material refining process efficiency optimization and real-time temperature control method and system based on artificial intelligence

Abstract

The invention discloses an artificial intelligence-based refining process efficiency optimization and real-time temperature control method and system, and the method achieves the efficient optimization and precise temperature control of a refining process through the innovation of four steps: 1, breaking through the limitation of traditional single-point monitoring, and deploying a multi-modal sensing array and a feed port near infrared spectrum sensor according to three-dimensional coverage; process-raw material two-dimensional data are collected in real time and associated, and the problem of data islands is solved; 2, innovatively constructing a'process-raw material 'two-dimensional feature vector, predicting a finished product rate by adopting an XGBoost model, and judging a dynamic power adjustment demand in combination with a threshold value to replace a traditional experience decision; thirdly, on the basis of the value output by the DQN model, an optimal dynamic temperature curve is calculated in a staged mode, and then the problem of fixed curve stiffness is solved through a generated precise power instruction; and fourthly, a three-level aging standard reaching mechanism is created for the first time, and an effect closed loop is formed corresponding to a differentiated scheme of model strengthening, local compensation and cross-furnace repairing.

Description

Technical Field

[0001] This invention relates to the field of refining process technology, specifically to a method and system for refining process efficiency optimization and real-time temperature control based on artificial intelligence. Background Technology

[0002] With the increasing demand for high-purity metals in the new energy and high-end manufacturing sectors, the limitations of traditional technical systems are becoming increasingly apparent. On the one hand, human experience cannot cope with complex operating conditions involving multiple coupled parameters; on the other hand, simple models and fixed curves are insufficient to meet the comprehensive needs of "improving quality, increasing efficiency, and reducing costs." The industry urgently needs a new technological approach that can achieve "multi-source data collaboration, AI-based precise decision-making, and hierarchical compensation closed-loop" to break through the bottlenecks of traditional technologies. Therefore, an artificial intelligence-based method and system for optimizing refining process efficiency and real-time temperature control has emerged.

[0003] Existing technologies, such as the invention patent application CN102222128B, disclose a method for optimizing the combustion of waste plastics in oil refining. Existing methods primarily rely on the experience of operators. This invention, however, utilizes data collection from the main waste plastics refining process. It establishes a data mining-based model to analyze the interrelationships between the axial temperature distribution in the reactor and the operating parameters of each burner, as well as the feed rate and product output. By applying parallel optimization algorithms, a method for optimizing the combustion of waste plastics in oil refining is established. This invention effectively controls the temperature distribution within the reactor during waste plastics refining, improving reaction efficiency and product quality. It can be implemented for both offline and online real-time combustion optimization.

[0004] Regarding the above-mentioned solutions, the applicant of this invention has discovered at least the following technical problems: 1. Traditional technologies adopt a "single-point monitoring + offline analysis" mode. Sensor deployment only covers a few key nodes such as furnace wall temperature and flue gas, lacking core data such as local temperature distribution of the molten pool and real-time changes in raw material composition. Moreover, process dimension and raw material dimension data are stored in independent systems, with inconsistent timestamps and incompatible formats, forming "data silos." At the same time, the system only has basic storage and alarm functions, lacking data cleaning and standardization processing capabilities. The feedback of offline raw material detection results is delayed, making it impossible to adjust the process in a timely manner when raw material composition fluctuates. Historical data is also difficult to reuse as a basis for optimization due to the lack of structured archiving.

[0005] 2. Traditional yield prediction relies solely on linear regression models based on two process parameters: heating power and melting time. It fails to incorporate key characteristics such as raw material composition and furnace gas concentration, resulting in high prediction errors and providing no reliable reference for power adjustment. Heating power adjustment, on the other hand, depends primarily on operator experience based on dashboard observations, lacking quantitative calculation logic. This easily leads to "over-adjustment" or "under-adjustment," compromising finished product quality stability and wasting energy, making it unsuitable for the precision requirements of large-scale production.

[0006] 3. Traditional smelting furnaces operate with fixed process curves and do not take into account the differences in raw material composition. This results in incomplete separation of impurities during the smelting of Grade B raw materials. When the temperature does not meet the standard, a "one-size-fits-all" compensation strategy is adopted without distinguishing the fault type. The adjustment data is not linked to the operating conditions and archived. Similar problems occur repeatedly, increasing the rate of unplanned downtime and failing to meet the requirements of high-quality production for temperature control flexibility and targeted compensation. Summary of the Invention

[0007] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a method and system for optimizing refining process efficiency and real-time temperature control based on artificial intelligence.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In the first aspect, the present invention provides a method for optimizing the efficiency of refining process and real-time temperature control based on artificial intelligence, including: Step 1, real-time acquisition of multi-source data and composition of refining materials: deploying multi-modal sensor arrays in each smelting furnace of the target plant, and simultaneously adding near-infrared spectral sensors at the feed inlet of each smelting furnace, thereby acquiring the process dimension data corresponding to each smelting furnace at each monitoring time point, and associating it with the raw material dimension data corresponding to each smelting furnace.

[0009] Step 2: Synergistic evaluation of refining process and composition: Based on the process dimension data and raw material dimension data corresponding to each smelting furnace at each monitoring time point, predict the yield value corresponding to each smelting furnace at each monitoring time point, and evaluate whether the heating power of each smelting furnace at each monitoring time point needs to be dynamically adjusted.

[0010] Step 3: Composition-oriented process parameter optimization: If a smelting furnace needs to dynamically adjust its heating power at a certain monitoring time point, the optimal dynamic temperature curve corresponding to the smelting furnace at that monitoring time point is analyzed, and the heating power adjustment command corresponding to the smelting furnace at that monitoring time point is also analyzed.

[0011] Step 4: Implementation and Effect Closure of Compensation Scheme: Send the heating power adjustment command corresponding to the smelting furnace at the monitoring time point to the smelting furnace actuator, monitor the temperature return time corresponding to the smelting furnace in real time, and analyze the compensation scheme corresponding to the smelting furnace.

[0012] In a second aspect, the present invention provides an artificial intelligence-based method for optimizing refining process efficiency and real-time temperature control, comprising: a real-time acquisition module for multi-source data and composition of refining materials: used to deploy multi-modal sensor arrays in each smelting furnace of the target plant, and simultaneously add near-infrared spectral sensors at the feed inlet of each smelting furnace, thereby acquiring process dimension data corresponding to each smelting furnace at each monitoring time point, and associating it with raw material dimension data corresponding to each smelting furnace.

[0013] The refining process and composition co-evaluation module is used to predict the yield of each smelting furnace at each monitoring time point based on the process dimension data and raw material dimension data corresponding to each smelting furnace at each monitoring time point, and to evaluate whether the heating power of each smelting furnace at each monitoring time point needs to be dynamically adjusted.

[0014] The composition-oriented process parameter optimization module is used to analyze the optimal dynamic temperature curve of a smelting furnace at a certain monitoring time point and the corresponding heating power adjustment command when the heating power of a smelting furnace needs to be dynamically adjusted at a certain monitoring time point.

[0015] The compensation scheme implementation and effect closed-loop module is used to send the heating power adjustment command corresponding to the smelting furnace at the monitoring time point to the smelting furnace actuator, monitor the temperature return time corresponding to the smelting furnace in real time, and analyze the compensation scheme corresponding to the smelting furnace.

[0016] The beneficial effects of this invention are as follows: 1. In the embodiments of this invention, through the collaborative deployment of a three-dimensional coverage multimodal sensing array and a near-infrared spectral sensor at the feed inlet, real-time correlation acquisition of process and raw material data is achieved for the first time. The sensor sampling frequency is increased to 1Hz, and the raw material composition analysis time is shortened from offline 3-4 hours to real-time, completely solving the traditional "data fragmentation" problem. At the same time, the edge computing node standardizes the four types of core data to form a 4-dimensional state vector input DQN model. Historical data is archived in a structured manner according to "furnace-parameter-effect", which significantly reduces the fluctuation of yield and provides high-value data support for subsequent process optimization, improving data reuse rate.

[0017] 2. In this embodiment of the invention, a multi-layer decision-making system of "XGBoost yield prediction + DQN temperature curve optimization + PID-AI power adjustment" is constructed. The yield prediction is input with 9-dimensional process-raw material characteristics to reduce prediction errors and provide a scientific basis for power adjustment. The optimal dynamic temperature curve is calculated according to a staged formula to adapt to different raw material compositions and equipment conditions, avoiding purity problems caused by traditional fixed curves. The heating power adjustment is carried out through "PID basic calculation + W value correction + safety constraint verification". The single adjustment range is accurately controlled within 10% of the current power, the power fluctuation of a single furnace is reduced from ±12% to ±3%, and energy consumption is reduced by 8%-10%, solving the energy waste problem of "over-adjustment" or "under-adjustment".

[0018] 3. This invention breaks through the traditional "one-size-fits-all" compensation model. Based on temperature regression time, it categorizes compliance into three levels: Level 1 (within 5 minutes), Level 2 (5-10 minutes), and Level 3 (not meeting the standard within 10 minutes). Differentiated compensation schemes are implemented accordingly. Level 1 compliance is achieved by accumulating optimal parameters through positive model reinforcement; Level 2 compliance is achieved by eliminating temperature unevenness through precise local compensation; and Level 3 compliance is achieved by resolving complex faults through cross-furnace parameter migration and root cause repair, significantly reducing unplanned downtime. Simultaneously, the compensation effect data and data from preceding steps form a closed-loop archive, reducing the recurrence rate of similar faults. This ensures high-quality production and further reduces long-term production costs through continuous optimization. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the implementation steps of the method of the present invention.

[0021] Figure 2 This is a schematic diagram of the system module connections of the present invention. Detailed Implementation

[0022] 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.

[0023] Examples of embodiments of the present invention Figure 1 As shown, an artificial intelligence-based method for optimizing refining process efficiency and real-time temperature control includes: Step 1, real-time acquisition of multi-source data and composition of refining materials: deploying multi-modal sensor arrays in each smelting furnace of the target plant, and simultaneously adding near-infrared spectral sensors at the feed inlet of each smelting furnace, thereby acquiring process dimension data corresponding to each smelting furnace at each monitoring time point, and associating it with raw material dimension data corresponding to each smelting furnace.

[0024] In one specific embodiment, the deployment of multimodal sensor arrays in each smelting furnace of the target factory, and the simultaneous addition of near-infrared spectral sensors at the feed inlet of each smelting furnace, are carried out as follows: the multimodal sensor arrays in each smelting furnace of the target factory are deployed according to the principle of three-dimensional coverage: one K-type thermocouple is set every 1m along the vertical direction of the furnace wall, two sets of dual-band infrared thermal imagers are symmetrically deployed 2m directly above the molten pool, and a laser gas analyzer is installed on the side wall of the flue outlet pipe; a power sensor is embedded in the furnace body power distribution box; at the same time, the near-infrared spectral sensor is installed directly above the feed conveyor belt of the smelting furnace, 30cm away from the raw material conveying surface, and all sensors are connected to the edge computing node through shielded cables, and the node is connected to the cloud platform through a 5G industrial module.

[0025] It should be noted that the K-type thermocouple uses a ceramic protective tube model with a temperature resistance of over 1800℃ to ensure stable operation in the high-temperature environment of the furnace wall; the dual-band infrared thermal imager uses a model with a resolution of 640×512 and a temperature measurement range of 0-2000℃, with a dust cover added to the lens and equipped with an automatic purging device to prevent dust from the molten pool from affecting the temperature measurement accuracy; the laser gas analyzer uses a model with strong anti-interference capabilities based on TDLAS technology, and the sampling probe is installed at a 45° angle to the flue gas flow direction to reduce probe damage caused by airflow impact; the power sensor uses a Hall current sensor with an accuracy of less than 0.5, and when embedded in the distribution box, it maintains a safe distance of ≥10cm from high-voltage lines to prevent electromagnetic interference; the near-infrared spectral sensor uses a model with a wavelength range of 900-1700nm and a resolution of ≤2nm, and the sensor housing is equipped with a high-temperature resistant heat insulation layer to prevent heat dissipation from the conveyor belt from affecting the detection results. All shielded cables are flame-retardant twisted-pair shielded cables, laid through galvanized steel conduits and separately from power cables to reduce electromagnetic interference. Edge computing nodes utilize industrial-grade fanless mainframes with built-in data encryption modules. 5G industrial modules employ VPN (Virtual Private Network) data transmission to prevent data leakage or tampering. Simultaneously, a regular sensor maintenance plan is implemented: K-type thermocouples are calibrated monthly; infrared thermal imagers have their lenses and purging devices cleaned quarterly; laser gas analyzers have their zero point and range calibrated monthly; and near-infrared spectral sensors are calibrated weekly with standard samples. This ensures the accuracy and reliability of data collected by each sensor, providing a high-quality data foundation for subsequent process analysis and optimization.

[0026] Step 2: Synergistic evaluation of refining process and composition: Based on the process dimension data and raw material dimension data corresponding to each smelting furnace at each monitoring time point, predict the yield value corresponding to each smelting furnace at each monitoring time point, and evaluate whether the heating power of each smelting furnace at each monitoring time point needs to be dynamically adjusted.

[0027] In a specific embodiment, the predicted yield value for each smelting furnace at each monitoring time point is predicted as follows: A gradient boosting decision tree regression model is used to construct a yield prediction model. The training samples for the yield prediction model are derived from the target plant's historical refining dataset over the past three years, with a training sample size of ≥5000 sets, covering different raw material types (A, B, and C grades) and various operating conditions including normal production, parameter fluctuations, and equipment maintenance. Input features are used to construct a two-dimensional feature vector of process and raw materials: the process dimension data includes four types of features: furnace wall temperature standard deviation, average molten pool temperature, average heating power, and CO concentration; the raw material dimension data includes five types of features: copper content, nickel content, and impurity percentage as resolved by near-infrared spectroscopy sensors, as well as raw material weight and raw material grade code, for a total of nine input features.

[0028] It should be noted that the standard deviation of the furnace wall temperature is calculated as follows: Based on the K-type thermocouples deployed on the furnace wall according to the three-dimensional coverage principle, the real-time temperature of each thermocouple is collected once every minute. For example, if the temperatures at the six measuring points are 1450℃, 1452℃, 1448℃, 1451℃, 1449℃, and 1453℃, the average temperature of the six measuring points within that minute is first calculated ((1450+1452+1448+1451+1449+1453) / 6=1450.5℃). Then, the deviation of each measuring point temperature from the average value is calculated using the standard deviation formula, such as standard deviation = [((1450-1450.5)] 2 +(1452-1450.5) 2 +...+(1453-1450.5) 2 [(6)≈1.87℃], and the average standard deviation every 5 minutes is taken as the characteristic value of the standard deviation of the furnace wall temperature at this monitoring time point.

[0029] Average temperature of the molten pool: Collected by two sets of dual-band infrared thermal imagers symmetrically deployed 2m directly above the molten pool. A thermal map of the molten pool temperature field is generated every 30 seconds. Temperature data of 100 uniformly distributed sampling points in the thermal map are extracted by edge computing nodes. For example, if the temperature range of the sampling points is 1445-1455℃, the arithmetic mean of the 100 sampling points is calculated, such as 1450.2℃. The average value every 2 minutes is taken as the characteristic value of the average temperature of the molten pool at the monitoring time point.

[0030] Average heating power: The current and voltage data of the heating circuit are collected in real time by a 0.5-level precision Hall current sensor embedded in the furnace body distribution box. The instantaneous power is calculated according to "power = voltage × current × power factor (taken as 0.92)". The average instantaneous power is calculated once every minute. If the instantaneous power fluctuates between 270-280kW in a certain minute, the average value is 275.3kW. The average value every 5 minutes is taken as the characteristic value of the average heating power at that monitoring time point.

[0031] CO concentration: Collected by a TDLAS laser gas analyzer installed on the side wall of the flue outlet pipe. The CO concentration in the flue is output once every minute. The real-time detection value at the monitoring time point is directly taken as the CO concentration characteristic value. If data fluctuation occurs, such as the difference between two adjacent detection values ​​exceeding 20 ppm, the median of the three detection values ​​before and after is taken as the final value.

[0032] Five characteristics of raw materials are acquired: copper content, nickel content, and impurity percentage. These are collected by a near-infrared spectral sensor installed 30cm above the feed conveyor belt of the smelting furnace. As the raw material moves with the conveyor belt, the sensor performs a spectral scan on the surface of the raw material every 2 seconds. The data is calibrated based on the offline test data of the target factory's raw materials over the past 3 years through a preset near-infrared spectral quantitative analysis model, such as the linear relationship between copper content and absorbance at a specific wavelength of 1200nm. For example, the analysis shows that the copper content is 99.2%, the nickel content is 0.5%, and the impurity percentage is 0.3% in a single scan. The average value of the scan data of each batch of raw materials is taken as the corresponding characteristic value at that monitoring time point.

[0033] Raw material weight: A weighing sensor is installed in the middle of the feeding conveyor belt to collect the raw material weight per unit length of the conveyor belt in real time. Combined with the running speed of the conveyor belt, the raw material feeding weight per minute is calculated according to "instantaneous weight = weight per unit length × speed × time". The total feeding weight of the 30 minutes before the monitoring time point is accumulated as the raw material weight characteristic value.

[0034] Raw material grade coding: Based on the batch information of the target factory's raw material management system, raw materials of grades A, B, and C are numerically coded, such as grade A = 1, grade B = 2, and grade C = 3. When raw materials are fed, the batch tag information of the raw materials is read by the RFID reader next to the conveyor belt, and the corresponding grade code is automatically matched. For example, if the current feed is grade A, the code is 1, which is directly used as the raw material grade code feature value to ensure that it is compatible with the format of the process-raw material two-dimensional feature vector.

[0035] During prediction, the process-raw material two-dimensional feature vectors corresponding to each smelting furnace at each monitoring time point are input into the well-trained XGBoost model, and the model outputs the predicted yield value corresponding to each smelting furnace at each monitoring time point.

[0036] It should be noted that the process-raw material two-dimensional feature vectors collected at the current monitoring time point are first preprocessed to ensure consistency with the feature format during model training: Process dimension feature processing: Furnace wall temperature standard deviation is retained to two decimal places, such as "2.35℃"; the average molten pool temperature is taken as the average of 100 sampling points within 5 minutes from the infrared thermal imager and rounded to the nearest integer, such as "1352℃"; the average heating power is calculated as the average of 12 5-minute power data points within one hour prior to the monitoring time point, in kW, and retained to one decimal place, such as "285.3kW"; CO concentration is retained to three decimal places, such as "0.215%". Raw material dimension feature processing: copper content, nickel content, and impurity percentage are all retained to two decimal places, such as "98.52%", "0.85%", and "0.63%". The raw material weight is accurate to 0.1 kg, such as "800.5 kg". The raw material grade code is converted to an integer according to "Grade A = 3, Grade B = 2, Grade C = 1". The preprocessed 9-dimensional feature vector is arranged in the order of "process dimension (4 items) → raw material dimension (5 items)" to form a standardized input array, such as "[2.35,1352,285.3,0.215,98.52,0.85,0.63,800.5,3]".

[0037] The mature XGBoost model is called through the model service interface of the cloud AI platform. The specific calling process is as follows: Interface parameter settings: The input parameters include "furnace ID", "monitoring timestamp", and "standardized feature array". The model version is also specified. The latest iteration version is called by default. The version number format is "V + date + batch number", such as "V2025092501"; Model calculation process: The model first loads the pre-trained decision tree parameters, including the split threshold of each node and the weight of the leaf nodes. It then traverses the input feature array dimension by dimension: 1. It judges whether the feature values ​​meet the split conditions according to the decision tree level, such as "furnace ID", "monitoring timestamp", and "standardized feature array". If the average pool temperature is >1350℃, proceed to the left subtree; otherwise, proceed to the right subtree. 2. After traversing all 100 decision trees, summarize the output prediction values ​​of each tree, and sum them according to the subsample ratio (0.8) to obtain the preliminary prediction results. 3. Apply activation function processing to the preliminary prediction results, using the Sigmoid function to map to the 0-100% range, and finally output the yield prediction value, retaining 2 decimal places, such as "96.85%". Calculation time control: Accelerate calculation through cloud GPU to ensure that the time for a single prediction is ≤500ms, meeting the real-time requirements. The monitoring interval is 5 minutes to reserve sufficient processing time.

[0038] After completing the prediction result verification, perform the following operations: Data storage: Encapsulate "furnace ID, monitoring timestamp, standardized feature array, predicted value, deviation flag, and verification status" into JSON format data and store it in a time series database, such as InfluxDB. At the same time, associate it with the original process data and raw material data at the monitoring time point to form a complete traceability chain.

[0039] In a specific embodiment, the evaluation process for determining whether each smelting furnace needs dynamic adjustment of heating power at each monitoring time point is as follows: The predicted yield and process dimension data corresponding to each smelting furnace at each monitoring time point are compared with the set threshold values ​​for predicted yield and process dimension data, respectively. If the predicted yield is less than 92% or any process dimension data is greater than the set threshold, then the smelting furnace at that monitoring time point is evaluated as needing dynamic adjustment of heating power. If the predicted yield is greater than or equal to 92%, and all process dimension data are less than or equal to the set threshold, then the smelting furnace at that monitoring time point is evaluated as not needing dynamic adjustment of heating power.

[0040] Step 3: Composition-oriented process parameter optimization: If a smelting furnace needs to dynamically adjust its heating power at a certain monitoring time point, the optimal dynamic temperature curve corresponding to the smelting furnace at that monitoring time point is analyzed, and the heating power adjustment command corresponding to the smelting furnace at that monitoring time point is also analyzed.

[0041] In a specific embodiment, the analysis of the optimal dynamic temperature curve corresponding to the smelting furnace at the monitoring time point is specifically performed as follows: A1. Obtain the molten pool temperature Q, raw material composition deviation W, heating power E, and smelting time R corresponding to the state vector of the smelting furnace in each dimension at the monitoring time point.

[0042] A2. If the smelting furnace is in the heating stage at the monitoring time point, then the heating stage curve corresponding to the smelting furnace at the monitoring time point is as follows: Target heating temperature: calculated based on the molten pool temperature Q value, according to the formula: Target temperature = Basic melting temperature 1450 + (Q value - 7.0) × 10; Heating rate: calculated based on the heating power E value, according to the formula: Rate = Basic rate 4 + (E value - 6.5) × 0.8; Remaining heating time: based on the smelting time R value, the current actual temperature is 1380℃, the remaining heating temperature difference = Target heating temperature - Current actual temperature, and the remaining heating time = Remaining heating temperature difference ÷ 4.5. That is, the heating stage will continue until: The smelting time R value + the remaining heating time. Taking the actual time corresponding to the smelting time R value as the current starting point and the current actual temperature as the starting temperature, the temperature increases with time minute by minute according to the heating rate, forming a continuous optimal dynamic temperature curve for the heating stage.

[0043] A2. If the smelting furnace is in the heat preservation stage at the monitoring time point, then the corresponding heating stage curve for the smelting furnace at the monitoring time point is as follows: Heat preservation temperature reference: calculated based on the molten pool temperature Q value, according to heat preservation reference temperature = heating stage target temperature + (Q value - 7.5) × 5; Temperature fluctuation range: calculated based on the raw material composition deviation W value, according to fluctuation range = ±3 - (W value - 6.5) × 0.5; Heat preservation duration: calculated based on the smelting duration R value, according to the total heat preservation stage duration = 80 - (R value - 6.0) × 10. Taking the actual time corresponding to the smelting duration R value as the starting point and the heat preservation reference temperature as the center, within a fluctuation range of ±2.8℃, a small-amplitude temperature curve is plotted in combination with power compensation requirements to form the optimal dynamic temperature curve for the heat preservation stage.

[0044] A3. If the smelting furnace is in the cooling stage at the monitoring time point, then the corresponding heating stage curve for the smelting furnace at the monitoring time point is as follows: Cooling start temperature: Based on the reference temperature of the holding stage, it is obtained by calculating the cooling start temperature = holding reference temperature - (Q value - 7.5) × 2; Cooling rate: Based on the heating power E value, it is calculated by calculating the rate = 3 + (E value - 6.5) × 2; Remaining cooling time: Based on the smelting time R value, the cooling end temperature, the remaining cooling temperature difference = cooling start temperature - cooling end temperature, the remaining cooling time = remaining cooling temperature difference ÷ 3.8; Cooling stage end time: The smelting time R value + the remaining cooling time, taking the actual time corresponding to the smelting time R value as the starting point, the cooling start temperature as the starting temperature, and plotting the temperature decreasing with time minute by minute according to the cooling rate, forming the optimal dynamic temperature curve of the cooling stage.

[0045] In a specific embodiment, the acquisition process for obtaining the molten pool temperature Q, raw material composition deviation W, heating power E, and smelting time R corresponding to the state vector of the smelting furnace at the monitoring time point is as follows: Four types of core data corresponding to the smelting furnace at the monitoring time point are collected and standardized to form a state vector for the input model: standardized molten pool temperature, standardized raw material composition deviation, standardized heating power, and standardized smelting time. The standardized molten pool temperature, standardized raw material composition deviation, standardized heating power, and standardized smelting time data are integrated in the order of molten pool temperature, composition deviation, power, and time to form a 4-dimensional state vector corresponding to the smelting furnace at the monitoring time point. This 4-dimensional state vector is input into the DQN model to calculate the molten pool temperature Q, raw material composition deviation W, heating power E, and smelting time R corresponding to the state vector of the smelting furnace at the monitoring time point.

[0046] It should be noted that the calculation process for the molten pool temperature Q, raw material composition deviation W, heating power E, and melting time R is as follows: The standardized molten pool temperature is denoted as S1: The average temperature of the molten pool is collected by the infrared thermal imager within 5 minutes and standardized according to "S1=(actual temperature-minimum melting temperature 1200℃) / (maximum tolerable temperature 1800℃-1200℃)", with the value range mapped to 0-1.

[0047] The standardized raw material composition deviation is denoted as S2: Based on the copper, nickel, and impurity content analyzed by the near-infrared spectral sensor, the comprehensive composition deviation is calculated. For example, if the copper content deviation is -0.2%, the nickel content deviation is +0.1%, and the impurity content deviation is +0.3%, the comprehensive deviation is (-0.2+0.1+0.3) / 3=0.067%). The deviation is standardized according to "S2=(comprehensive deviation-minimum deviation-2%) / (maximum deviation2%-(-2%))", with a value range of 0-1.

[0048] The standardized heating power is denoted as S3: the real-time output power of the power sensor is collected and standardized according to "S3=(actual power-minimum sustaining power 150kW) / (rated maximum power 500kW-150kW)", with a value range of 0-1.

[0049] The standardized melting time is denoted as S4: It is standardized according to "S4 = melting time / total melting time", and the value range is 0-1.

[0050] The final result is a 4-dimensional normalized state vector: [S1,S2,S3,S4], formatted as a TensorFlow-compatible float32 array, which is used as input to the DQN model.

[0051] The DQN model adopts a network structure of "convolutional layer + fully connected layer". It maps the 4-dimensional state vector into Q value, W value, E value and R value through pre-trained weight parameters. The values ​​range from 0 to 10, which meets the requirements of process evaluation. The specific steps are as follows: Input layer: Receives the 4-dimensional state vector and converts it into a feature matrix of "[1,4]" through the Reshape layer to adapt to the input format of the fully connected layer.

[0052] Fully connected layer 1 (64 neurons): The ReLU activation function is used, and the calculation formula is "F1=ReLU(W1×X+B1)", where W1 is a 64×4 weight matrix (e.g., the weight of a certain row is [0.2,0.5,0.1,0.2]), X is the input feature matrix, B1 is a 64-dimensional bias vector (e.g., [0.1,0.05,...,0.08]), and the output is a 64-dimensional feature vector F1.

[0053] Fully connected layer 2 (32 neurons): It also uses the ReLU activation function, and the calculation formula is "F2=ReLU(W2×F1+B2)", where W2 is a 32×64 weight matrix, B2 is a 32-dimensional bias vector, and the output is a 32-dimensional feature vector F2, which completes feature compression and key information extraction.

[0054] The output layer contains four independent linear neurons, corresponding to the Q, W, E, and R values, respectively. The hidden layer output is converted into a process evaluation value of 0-10 through "linear mapping + range scaling". All weight matrices (W1, W2, W_Q, W_W, W_E, W_R) and bias vectors (B1, B2, B_Q, B_W, B_E, B_R) are trained using three years of historical data from the target factory. The training process uses "maximizing yield + minimizing energy consumption" as the dual objective function and adopts the Adam optimizer for iterative updates. The model loss rate (MSE) is stable below 0.02.

[0055] Calculation of the Q value of the molten pool temperature: The formula for calculating the output layer neuron 1 is "Q_raw=W_Q×F2+B_Q", where W_Q is a 1×32 weight vector (e.g., [0.3,0.15,...,0.2]) and B_Q is a bias term (e.g., 0.2), resulting in Q_raw (example: Q_raw=0.78); scaled to the range of 0-10 according to "Q value=Q_raw×10" (example: Q=0.78×10=7.8), corresponding to the molten pool temperature adaptability evaluation (7.8 points, indicating that the temperature is close to the optimal range).

[0056] Raw material composition deviation W value calculation: The calculation formula for output layer neuron 2 is "W_raw=W_W×F2+B_W", where W_W is a 1×32 weight vector and B_W is the bias term (example: W_raw=0.68); scaled by "W value=W_raw×10" (example: W=0.68×10=6.8).

[0057] Heating power E value calculation: The calculation formula for output layer neuron 3 is "E_raw=W_E×F2+B_E", where W_E is a 1×32 weight vector and B_E is the bias term (example: E_raw=0.69); scaled by "E value=E_raw×10" (example: E=0.69×10=6.9).

[0058] Calculation of the melting time R value: The formula for the output layer neuron 4 is "R_raw=W_R×F2+B_R", where W_R is a 1×32 weight vector and B_R is the bias term (example: R_raw=0.4); scaled by "R value=R_raw×10" (example: R=0.4×10=4.0).

[0059] In a specific embodiment, the analysis of the heating power adjustment command corresponding to the smelting furnace at the monitoring time point is as follows: First, obtain the current actual temperature of the smelting furnace at the monitoring time point and the target temperature at the corresponding time point of the optimal dynamic temperature curve, and calculate the temperature deviation = target temperature - current actual temperature; use a traditional PID controller to calculate the basic power adjustment amount, and set the core PID parameters as: proportional coefficient Kp = 2.5, integral coefficient Ki = 0.1, and derivative coefficient Kd = 0.5. The calculation formula is: basic power adjustment amount = proportional coefficient Kp × temperature deviation + integral coefficient Ki × total temperature deviation value of the past 5 minutes + derivative coefficient Kd × (current temperature deviation value - temperature deviation value of the previous minute).

[0060] Then, the raw material composition deviation W value of the smelting furnace at the monitoring time point is obtained, and the raw material composition deviation W value of the smelting furnace at the monitoring time point is compared with the set standard raw material composition deviation W value. If the raw material composition deviation W value of the smelting furnace at the monitoring time point is 0.1% lower than the set standard raw material composition deviation W value, the basic adjustment amount is increased by 1.5%; if it is 0.1% higher, the basic adjustment amount is decreased by 1%. The adjusted basic adjustment amount is recorded as the heating power adjustment value of the smelting furnace at the monitoring time point.

[0061] Finally, the heating power E value of the smelting furnace at the monitoring time point is obtained, and power adjustment safety constraints are set: single adjustment range: not exceeding 10% of the heating power E value of the smelting furnace at the monitoring time point; upper limit of adjusted power: not exceeding 90% of the rated maximum power of the smelting furnace at the monitoring time point; lower limit of adjusted power: not lower than the minimum power of the smelting furnace maintained at the monitoring time point; finally, the heating power adjustment command corresponding to the smelting furnace at the monitoring time point is generated.

[0062] Step 4: Implementation and Effect Closure of Compensation Scheme: Send the heating power adjustment command corresponding to the smelting furnace at the monitoring time point to the smelting furnace actuator, monitor the temperature return time corresponding to the smelting furnace in real time, and analyze the compensation scheme corresponding to the smelting furnace.

[0063] In a specific embodiment, the analysis of the compensation scheme corresponding to the smelting furnace is carried out as follows: B1. Obtain the timeliness achievement type corresponding to the smelting furnace. The timeliness achievement type includes first-level optimization achievement, second-level compensation achievement and third-level collaborative intervention. If the timeliness achievement level corresponding to the smelting furnace is first-level optimization achievement, then execute the no-additional-compensation + model positive reinforcement scheme.

[0064] It should be noted that when the temperature of the smelting furnace returns to the optimal curve range within 5 minutes, the scheme of no additional compensation + positive model enhancement is implemented. The core is to optimize the model through data accumulation. During implementation, the edge computing nodes first automatically extract the full-link data for the furnace operation: including a 4D standardized state vector, evaluation parameters such as Q and W values ​​output by the DQN model, as well as temperature regression process data and energy consumption change data, and these data are marked as "high-confidence positive samples". Next, the samples are encrypted and synchronized to the factory's federated learning nodes through a 5G industrial module. In the monthly model iteration, the samples are assigned a weight of 1.3 times, focusing on strengthening the state-action mapping relationship between the "Q and W values" intervals in the DQN model, while adjusting the feature weights of the XGBoost yield prediction model. Finally, a structured "uncompensated confirmation report" is generated, recording key related information such as "Grade A copper raw material, heating rate of 4.2℃ / min, and yield of 94.5%", and stored in the process knowledge graph "optimal case library". When calling parameters for smelting furnaces under the same operating conditions in the future, the system can directly match the data in the case library, reducing the decision time from 5 seconds to within 2 seconds, and eliminating the need for additional power adjustments, thus avoiding energy waste.

[0065] B2. If the aging compliance level corresponding to the smelting furnace is Level 2 compensation compliance, then implement the local precise compensation + parameter fine-tuning solidification scheme.

[0066] It should be noted that a localized precise compensation + parameter fine-tuning solidification scheme is implemented when the temperature of the smelting furnace returns to the optimal curve range within 5-10 minutes. The first step involves using a temperature field thermogram generated by a dual-band infrared thermal imager directly above the molten pool to locate localized low-temperature areas. The second step, based on a preset "localized temperature deviation - power adjustment" mapping rule, increases the heating power of the corresponding area by 0.5% for every 0.1℃ decrease in temperature. The local compensation amount is calculated: the current power of the heating unit on the right side of the furnace wall is 275kW, requiring an increase of 0.5% × (0.8 / 0.1) = 4%, or 11kW. The adjusted power is 286kW. A localized compensation command is then generated and sent to the smelting furnace zone power controller. The three-step process involves collecting local temperature data every 10 seconds via furnace wall thermocouples for 3 minutes to ensure the temperature deviation in that area remains stable within ±0.5℃. After compensation, the relationships between parameters such as "right side of furnace wall, W value, and local power increase of 4%" and raw material composition and furnace area are solidified into the process knowledge graph, establishing a mapping table of "raw material composition deviation - local area - power fine-tuning amount". Subsequently, when a low temperature of 0.6℃ and a W value of 6.5 appear on the left side of the furnace wall, the system can directly call the parameters in the table without recalculation.

[0067] B3. If the aging compliance level corresponding to the smelting furnace is Level 3 collaborative intervention, then the root cause repair compensation + cross-furnace parameter migration scheme shall be implemented.

[0068] It should be noted that when the temperature in the smelting furnace fails to return to the optimal curve range within 10 minutes, the root cause repair compensation + cross-furnace parameter migration solution is implemented. The core is to solve the fundamental problem and reuse similar experience. During implementation, the root cause of the fault is first located by combining multimodal sensor data and a digital twin model: if the power sensor detects that the actual output power of the heating element is 260kW, which is 15kW lower than the rated 280kW, it is determined to be an equipment problem. The power limit of the heating element is then adjusted to 300kW, and a maintenance work order is sent, noting "Heating element usage time 8000 hours, replacement recommended within 72 hours"; if the near-infrared spectral sensor detects a sudden decrease of 0.8% in the content of a certain raw material, it is determined to be a raw material problem. The raw material ratio at the feed inlet is increased from 10% to 15%, and the holding time is extended by 5 minutes; the second step involves calling up data from within the plant area... Real-time desensitized data from furnace F002 of the same type and raw material batch was used to extract effective compensation parameters under the scenario of "raw material content low by 0.7%". Combined with the current aging status of the furnace body, the increase in the heat preservation temperature was fine-tuned to 8℃ to generate the final compensation scheme. In the third step, after the compensation was executed, the temperature and yield changes were continuously monitored for 15 minutes to ensure that the temperature deviation was ≤±1℃. Finally, information such as "equipment aging - power upper limit adjustment - cross-furnace reference furnace F002" was entered into the knowledge graph "fault response library" to update the fault handling branch of the DQN model, thereby increasing the success rate of subsequent autonomous compensation for similar faults from the original 75% to over 90%.

[0069] In a specific embodiment, the process of obtaining the timeliness compliance type corresponding to the smelting furnace is as follows: obtain the temperature return time corresponding to the smelting furnace. If the temperature of the smelting furnace returns to the optimal curve range within 5 minutes, it is recorded as Level 1 optimization compliance. If the temperature of the smelting furnace returns to the optimal curve range within 5 to 10 minutes, it is recorded as Level 2 compensation compliance. If the temperature of the smelting furnace does not return to the optimal curve range within 10 minutes, it is recorded as Level 3 collaborative intervention.

[0070] Examples of embodiments of the present invention Figure 2 As shown, an artificial intelligence-based refining process efficiency optimization and real-time temperature control system includes: a refining multi-source data and composition real-time acquisition module: used to deploy multi-modal sensor arrays in each smelting furnace of the target plant, and simultaneously add near-infrared spectral sensors at the feed inlet of each smelting furnace, so as to collect the process dimension data corresponding to each smelting furnace at each monitoring time point, and associate it with the raw material dimension data corresponding to each smelting furnace.

[0071] The refining process and composition co-evaluation module is used to predict the yield of each smelting furnace at each monitoring time point based on the process dimension data and raw material dimension data corresponding to each smelting furnace at each monitoring time point, and to evaluate whether the heating power of each smelting furnace at each monitoring time point needs to be dynamically adjusted.

[0072] The composition-oriented process parameter optimization module is used to analyze the optimal dynamic temperature curve of a smelting furnace at a certain monitoring time point and the corresponding heating power adjustment command when the heating power of a smelting furnace needs to be dynamically adjusted at a certain monitoring time point.

[0073] The compensation scheme implementation and effect closed-loop module is used to send the heating power adjustment command corresponding to the smelting furnace at the monitoring time point to the smelting furnace actuator, monitor the temperature return time corresponding to the smelting furnace in real time, and analyze the compensation scheme corresponding to the smelting furnace.

[0074] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in this specification, they should all fall within the protection scope of the present invention.

Claims

1. A method for optimizing refining process efficiency and real-time temperature control based on artificial intelligence, characterized in that, include: Step 1: Real-time acquisition of multi-source data and composition of raw materials: Deploy multi-modal sensor arrays in each smelting furnace of the target plant, and simultaneously add near-infrared spectral sensors at the feed inlet of each smelting furnace to collect process dimension data corresponding to each smelting furnace at each monitoring time point, and associate it with raw material dimension data corresponding to each smelting furnace. Step 2: Synergistic evaluation of refining process and composition: Based on the process dimension data and raw material dimension data corresponding to each smelting furnace at each monitoring time point, predict the yield value corresponding to each smelting furnace at each monitoring time point, and evaluate whether the heating power of each smelting furnace at each monitoring time point needs to be dynamically adjusted. Step 3: Composition-oriented process parameter optimization: If a smelting furnace needs to dynamically adjust its heating power at a certain monitoring time point, analyze the optimal dynamic temperature curve corresponding to the smelting furnace at that monitoring time point, and analyze the heating power adjustment command corresponding to the smelting furnace at that monitoring time point. Step 4: Implementation and Effect Closure of Compensation Scheme: Send the heating power adjustment command corresponding to the smelting furnace at the monitoring time point to the smelting furnace actuator, monitor the temperature return time corresponding to the smelting furnace in real time, and analyze the compensation scheme corresponding to the smelting furnace.

2. The method for optimizing refining process efficiency and real-time temperature control based on artificial intelligence as described in claim 1, characterized in that, The deployment of multimodal sensor arrays in each smelting furnace of the target plant, along with the simultaneous addition of near-infrared spectral sensors at the feed inlet of each furnace, is described in the following steps: The target factory deploys multimodal sensor arrays in each smelting furnace according to the principle of three-dimensional coverage: one K-type thermocouple is installed every 1m along the vertical direction of the furnace wall, two sets of dual-band infrared thermal imagers are symmetrically deployed 2m directly above the molten pool, and a laser gas analyzer is installed on the side wall of the flue outlet pipe; power sensors are embedded in the furnace body power distribution box; at the same time, near-infrared spectral sensors are installed directly above the smelting furnace feed conveyor belt, 30cm away from the raw material conveying surface. All sensors are connected to edge computing nodes through shielded cables, and the nodes are connected to the cloud platform through 5G industrial modules.

3. The method for optimizing refining process efficiency and real-time temperature control based on artificial intelligence as described in claim 2, characterized in that, The predicted yield values ​​for each smelting furnace at each monitoring time point are determined as follows: A gradient boosting decision tree regression model was used to construct a yield prediction model. The training samples of the yield prediction model came from the target plant’s refining history association dataset over the past 3 years, with a training sample size of ≥5000 sets, covering different raw material types of Grade A, Grade B, and Grade C, as well as various operating conditions such as normal production, parameter fluctuations, and equipment maintenance. The input features construct a two-dimensional feature vector of process and raw materials: the process dimension data includes four types of features: standard deviation of furnace wall temperature, average temperature of molten pool, average heating power and CO concentration; the raw material dimension data includes five types of features: copper content, nickel content and impurity ratio as resolved by near-infrared spectroscopy sensor, as well as raw material weight and raw material grade code, for a total of nine input features; During prediction, the process-raw material two-dimensional feature vectors corresponding to each smelting furnace at each monitoring time point are input into the well-trained XGBoost model, and the model outputs the predicted yield value corresponding to each smelting furnace at each monitoring time point.

4. The method for optimizing refining process efficiency and real-time temperature control based on artificial intelligence as described in claim 3, characterized in that, The assessment process for determining whether each smelting furnace needs dynamic adjustment of its heating power at each monitoring time point is as follows: The predicted yield and process dimension data for each smelting furnace at each monitoring time point are compared with the set thresholds for predicted yield and process dimension data. If the predicted yield is less than 92% or any process dimension data is greater than the set threshold, the smelting furnace is assessed to need dynamic adjustment of heating power at that monitoring time point. If the predicted yield is greater than or equal to 92% and all process dimension data are less than or equal to the set threshold, the smelting furnace is assessed to not need dynamic adjustment of heating power at that monitoring time point.

5. The method for optimizing refining process efficiency and real-time temperature control based on artificial intelligence as described in claim 4, characterized in that, The analysis of the optimal dynamic temperature curve of the smelting furnace at the monitoring time point is as follows: A1. Obtain the molten pool temperature Q, raw material composition deviation W, heating power E, and smelting time R corresponding to the state vector of the smelting furnace in each dimension at the monitoring time point; A2. If the smelting furnace is in the heating stage at the monitoring time point, then the heating stage curve corresponding to the smelting furnace at the monitoring time point is as follows: Target heating temperature: calculated based on the molten pool temperature Q value, according to the formula: Target temperature = Basic melting temperature 1450 + (Q value - 7.0) × 10; Heating rate: calculated based on the heating power E value, according to the formula: Rate = Basic rate 4 + (E value - 6.5) × 0.8; Remaining heating time: based on the smelting time R value, the current actual temperature is 1380℃, the remaining heating temperature difference = Target heating temperature - Current actual temperature, and the remaining heating time = Remaining heating temperature difference ÷ 4.

5. That is, the heating stage will continue until: The smelting time R value + the remaining heating time. Taking the actual time corresponding to the smelting time R value as the current starting point and the current actual temperature as the starting temperature, the temperature increases with time minute by minute according to the heating rate, forming a continuous optimal dynamic temperature curve for the heating stage. A2. If the smelting furnace is in the heat preservation stage at the monitoring time point, then the corresponding heating stage curve for the smelting furnace at the monitoring time point is as follows: Heat preservation temperature reference: calculated based on the molten pool temperature Q value, according to heat preservation reference temperature = heating stage target temperature + (Q value - 7.5) × 5; Temperature fluctuation range: calculated based on the raw material composition deviation W value, according to fluctuation range = ±3 - (W value - 6.5) × 0.5; Heat preservation time: Based on the melting time R value, it is calculated as follows: total heat preservation time = 80 - (R value - 6.0) × 10. Taking the actual time corresponding to the melting time R value as the starting point and the heat preservation reference temperature as the center, within the fluctuation range of ±2.8℃, a small fluctuation temperature curve is plotted in combination with the power compensation requirements to form the optimal dynamic temperature curve for the heat preservation stage. A3. If the smelting furnace is in the cooling stage at the monitoring time point, then the corresponding heating stage curve for the smelting furnace at the monitoring time point is as follows: Cooling start temperature: Based on the reference temperature of the holding stage, it is obtained by calculating the cooling start temperature = holding reference temperature - (Q value - 7.5) × 2; Cooling rate: Based on the heating power E value, it is calculated by calculating the rate = 3 + (E value - 6.5) × 2; Remaining cooling time: Based on the smelting time R value, the cooling end temperature, the remaining cooling temperature difference = cooling start temperature - cooling end temperature, the remaining cooling time = remaining cooling temperature difference ÷ 3.8; Cooling stage end time: The smelting time R value + the remaining cooling time, taking the actual time corresponding to the smelting time R value as the starting point, the cooling start temperature as the starting temperature, and plotting the temperature decreasing with time minute by minute according to the cooling rate, forming the optimal dynamic temperature curve of the cooling stage.

6. The method for optimizing refining process efficiency and real-time temperature control based on artificial intelligence as described in claim 5, characterized in that, The specific process for obtaining the molten pool temperature Q, raw material composition deviation W, heating power E, and smelting time R corresponding to the state vectors of the smelting furnace at each monitoring time point is as follows: Four types of core data corresponding to the smelting furnace at the monitoring time point are collected and standardized to form a state vector for the input model: standardized molten pool temperature, standardized raw material composition deviation, standardized heating power, and standardized smelting time. The standardized molten pool temperature, standardized raw material composition deviation, standardized heating power, and standardized smelting time data are integrated in the order of molten pool temperature, composition deviation, power, and time to form a 4-dimensional state vector corresponding to the smelting furnace at the monitoring time point. The 4-dimensional state vector corresponding to the smelting furnace at the monitoring time point is input into the DQN model to calculate the molten pool temperature Q value, raw material composition deviation W value, heating power E value, and smelting time R value corresponding to each dimension of the smelting furnace at the monitoring time point.

7. The method for optimizing refining process efficiency and real-time temperature control based on artificial intelligence as described in claim 6, characterized in that, The analysis of the heating power adjustment command corresponding to the smelting furnace at the monitoring time point is as follows: First, obtain the current actual temperature of the smelting furnace at the monitoring time point and the target temperature at the corresponding time point of the optimal dynamic temperature curve, and calculate the temperature deviation = target temperature - current actual temperature; use a traditional PID controller to calculate the basic power adjustment, and set the core PID parameters as: proportional coefficient Kp = 2.5, integral coefficient Ki = 0.1, and derivative coefficient Kd = 0.

5. The calculation formula is: basic power adjustment = proportional coefficient Kp × temperature deviation + integral coefficient Ki × total temperature deviation value of the past 5 minutes + derivative coefficient Kd × (current temperature deviation value - temperature deviation value of the previous minute). Then, the raw material composition deviation W value of the smelting furnace at the monitoring time point is obtained, and the raw material composition deviation W value of the smelting furnace at the monitoring time point is compared with the set standard raw material composition deviation W value. If the raw material composition deviation W value of the smelting furnace at the monitoring time point is 0.1% lower than the set standard raw material composition deviation W value, the basic adjustment amount is increased by 1.5%; if it is 0.1% higher, the basic adjustment amount is decreased by 1%. The adjusted basic adjustment amount is recorded as the heating power adjustment value of the smelting furnace at the monitoring time point. Finally, the heating power E value of the smelting furnace at the monitoring time point is obtained, and power adjustment safety constraints are set: single adjustment range: not exceeding 10% of the heating power E value of the smelting furnace at the monitoring time point; upper limit of adjusted power: not exceeding 90% of the rated maximum power of the smelting furnace at the monitoring time point; lower limit of adjusted power: not lower than the minimum power of the smelting furnace maintained at the monitoring time point; finally, the heating power adjustment command corresponding to the smelting furnace at the monitoring time point is generated.

8. The method for optimizing refining process efficiency and real-time temperature control based on artificial intelligence as described in claim 7, characterized in that, The analysis of the compensation scheme corresponding to the smelting furnace is as follows: B1. Obtain the timeliness achievement type corresponding to the smelting furnace. The timeliness achievement type includes Level 1 optimization achievement, Level 2 compensation achievement and Level 3 collaborative intervention. If the timeliness achievement level corresponding to the smelting furnace is Level 1 optimization achievement, then execute the no additional compensation + model positive reinforcement scheme. B2. If the aging compliance level corresponding to the smelting furnace is Level II compensation compliance, then implement the local precise compensation + parameter fine-tuning solidification scheme. B3. If the aging compliance level corresponding to the smelting furnace is Level 3 collaborative intervention, then the root cause repair compensation + cross-furnace parameter migration scheme shall be implemented.

9. The method for optimizing refining process efficiency and real-time temperature control based on artificial intelligence as described in claim 8, characterized in that, The specific process for obtaining the aging compliance type corresponding to the smelting furnace is as follows: Obtain the temperature regression time corresponding to the smelting furnace. If the temperature of the smelting furnace returns to the optimal curve range within 5 minutes, it is recorded as Level 1 optimization achievement. If the temperature of the smelting furnace returns to the optimal curve range within 5 to 10 minutes, it is recorded as Level 2 compensation achievement. If the temperature of the smelting furnace does not return to the optimal curve range within 10 minutes, it is recorded as Level 3 collaborative intervention.

10. An AI-based refining process efficiency optimization and real-time temperature control system for implementing the AI-based refining process efficiency optimization and real-time temperature control method according to any one of claims 1-9, characterized in that, include: Real-time acquisition module for multi-source data and composition of raw materials: It is used to deploy multi-modal sensor arrays in each smelting furnace of the target plant and simultaneously add near-infrared spectral sensors at the feed inlet of each smelting furnace, so as to collect the corresponding process dimension data of each smelting furnace at each monitoring time point and associate it with the corresponding raw material dimension data of each smelting furnace. The refining process and composition co-evaluation module is used to predict the yield of each smelting furnace at each monitoring time point based on the process dimension data and raw material dimension data corresponding to each smelting furnace at each monitoring time point, and to evaluate whether the heating power of each smelting furnace at each monitoring time point needs to be dynamically adjusted. The composition-oriented process parameter optimization module is used to analyze the optimal dynamic temperature curve of the smelting furnace at a certain monitoring time point and the heating power adjustment command of the smelting furnace at that monitoring time point when the heating power of the smelting furnace needs to be dynamically adjusted at a certain monitoring time point. The compensation scheme implementation and effect closed-loop module is used to send the heating power adjustment command corresponding to the smelting furnace at the monitoring time point to the smelting furnace actuator, monitor the temperature return time corresponding to the smelting furnace in real time, and analyze the compensation scheme corresponding to the smelting furnace.

Images