An artificial intelligence-based liquid metal smelting detection analysis method and system

Abstract

This application relates to an artificial intelligence-based detection and analysis method and system for liquid metal smelting. The method includes: acquiring real-time detection signals from a constant oxygen probe and a temperature probe in a smelting furnace, as well as the material formula identification information of the constant oxygen probe; outputting a basic dataset containing initial oxygen activity, current temperature, and material formula identification. The basic dataset is input into a polarization error compensation model corresponding to the material formula, outputting an initial polarization error compensation value. The cumulative usage time of the probe is input into a probe performance decay model to generate a dynamic error amplification coefficient, which is multiplied by the initial compensation value to obtain the actual polarization error correction value. The error correction value is used to correct the initial oxygen activity, outputting target oxygen activity data. A feeding control command sequence is generated by matching the smelting process specification database. The command sequence is desensitized to generate a desensitized log, encrypted and uploaded to a cloud server, while simultaneously sending commands to the execution terminal. This application meets the dual practical needs of modern metallurgical industry for refined production control and secure data assets.

Description

Technical Field

[0001] This application relates to the field of industrial data analysis technology, and in particular to an artificial intelligence-based detection and analysis method and system for liquid metal smelting. Background Technology

[0002] As global manufacturing transforms towards intelligent and green practices, liquid metal smelting (such as steel and aluminum alloy smelting), as a core link in the production of basic raw materials, faces increasingly stringent requirements for process precision, efficiency, and safety. Online monitoring technology, through sensors such as oxygen level probes and temperature probes, collects key parameters like oxygen activity and temperature in molten steel in real time, providing a basis for process decisions such as deoxidation alloy addition and furnace temperature control. This forms the foundation for achieving intelligent goals such as "one-click steelmaking." In recent years, many companies have made breakthroughs in sensor materials (such as nano-doped zirconia solid electrolyte tubes and highly active reference electrodes). However, converting detection signals into precise control commands still faces multiple challenges, including error compensation, data security, and process adaptation.

[0003] Currently, when facing extreme smelting conditions, oxygen-controlled probes exhibit complex polarization effects in high-temperature, low-oxygen environments. Furthermore, due to variations in the types and proportions of doped elements in the zirconium oxide solid electrolytes of different probes, their polarization characteristics are highly nonlinear and individualized. Existing liquid metal detection and analysis technologies cannot adapt to the individual differences in probe formulations with varying materials, nor can they account for the accelerated aging effects of dynamic changes in operating conditions, resulting in persistently low detection accuracy. In addition, existing smelting systems lack standardized safety mechanisms and closed-loop optimization capabilities in various data processing stages, hindering the intelligent collaboration of the entire smelting process and restricting the achievement of data-driven precision control objectives. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides an artificial intelligence-based method and system for detecting and analyzing liquid metal smelting.

[0005] Firstly, this application provides an artificial intelligence-based method for detecting and analyzing liquid metal smelting, employing the following technical solution: The system acquires real-time detection signals from oxygen constant probe and temperature probe in the smelting furnace, as well as the material formula identification information of the current oxygen constant probe. Based on the real-time detection signals, it calculates the initial oxygen activity data and outputs a basic dataset containing the initial oxygen activity, current temperature, and material formula identification. The current temperature and initial oxygen activity in the basic dataset are input into the polarization error compensation model corresponding to the material formulation identifier, and the polarization error compensation model outputs an initial polarization error compensation value that matches the current operating condition. The cumulative usage time data of the current oxygen determination probe is obtained and input into the pre-established probe performance degradation model to generate a dynamic error amplification coefficient. The dynamic error amplification coefficient is multiplied with the initial polarization error compensation value to output the actual polarization error correction value. The initial oxygen activity in the basic dataset is corrected by difference using the actual polarization error correction value, and the target oxygen activity data is output. The target oxygen activity data is matched with a preset smelting process specification database to generate a feeding control instruction sequence; Data desensitization processing is performed on the feeding control command sequence to remove the underlying potential signal features while retaining the process feeding features to generate a desensitized feeding log; The desensitized feeding log is encrypted and uploaded to the cloud server, while the feeding control command sequence is sent to the workshop execution terminal, and the cloud storage index and terminal execution status identifier are output.

[0006] By adopting the above technical solutions, a polarization error mapping mechanism driven by material formulation identification and a dynamic attenuation compensation mechanism based on cumulative usage time are introduced. This precisely solves the bottleneck of measurement distortion caused by probe polarization effect and material aging in the extreme conditions of high temperature and low oxygen in liquid metal smelting. It achieves a technical leap from traditional general signal noise reduction to material-level exclusive dynamic error correction, thereby continuously outputting highly reliable target oxygen activity data throughout the entire life cycle of the probe. On this basis, industrial data security is ensured through desensitized and encrypted transmission. By directly anchoring high-precision detection data to the smelting process specification database to generate precise feeding control instructions, alloy waste and steel defects caused by human experience intervention are effectively avoided, providing key technical support for the intelligent goal of "one-click steelmaking".

[0007] Secondly, this application provides an artificial intelligence-based detection and analysis system for liquid metal smelting, employing the following technical solution: The dataset construction module is used to acquire the real-time detection signals of the oxygen constant probe and temperature probe in the smelting furnace, as well as the material formula identification information of the current oxygen constant probe. Based on the real-time detection signals, the module calculates the initial oxygen activity data and outputs a basic dataset containing the initial oxygen activity, current temperature, and material formula identification. The polarization error mapping compensation module is used to input the current temperature and initial oxygen activity in the basic dataset into the polarization error compensation model corresponding to the material formulation identifier, and output the initial polarization error compensation value that matches the current working condition through the polarization error compensation model. The error comprehensive correction module is used to obtain the cumulative usage time data of the current oxygen probe and input it into the pre-established probe performance decay model to generate a dynamic error amplification coefficient. The dynamic error amplification coefficient is multiplied with the initial polarization error compensation value to output the actual polarization error correction value. The oxygen activity correction module is used to perform difference correction on the initial oxygen activity in the basic dataset using the actual polarization error correction value, and output the target oxygen activity data. The feeding control module is used to match the target oxygen activity data with a preset smelting process specification database to generate a feeding control instruction sequence. The data desensitization module is used to perform data desensitization processing on the feeding control command sequence, remove the underlying potential signal features, retain the process feeding features, and generate a desensitized feeding log. The data transmission module is used to encrypt and upload the desensitized feeding log to the cloud server, and at the same time send the feeding control command sequence to the workshop execution terminal, and output the cloud storage index and the terminal execution status identifier.

[0008] In summary, this application includes at least one of the following beneficial technical effects: This application uses material formulation identification and cumulative usage time as key variables, and performs dual compensation correction on the initial oxygen activity through a polarization error compensation model and a probe performance decay model, respectively, thereby improving the accuracy of oxygen activity detection data under complex smelting conditions; based on this high-precision data, the automatically generated feeding control instructions can effectively reduce the excessive or insufficient addition of deoxidizing alloys, reducing the cost of production auxiliary materials while stabilizing the final composition and quality of molten steel; furthermore, by performing underlying potential characteristic elimination and encrypted isolated transmission on the instruction sequence, while ensuring the real-time and accurate execution of feeding actions by the workshop execution terminal, the outflow of core sensing mechanisms and microscopic material data is avoided, meeting the dual practical needs of modern metallurgical industry for refined production control and secure data assets. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the first process of an artificial intelligence-based liquid metal smelting detection and analysis method according to one embodiment of this application.

[0010] Figure 2 This is a schematic diagram of the second process of an artificial intelligence-based liquid metal smelting detection and analysis method according to one embodiment of this application.

[0011] Figure 3 This is a schematic diagram of the third process of an artificial intelligence-based liquid metal smelting detection and analysis method according to one embodiment of this application.

[0012] Figure 4This is a schematic diagram of the fourth process of an artificial intelligence-based liquid metal smelting detection and analysis method according to one embodiment of this application.

[0013] Figure 5 This is a schematic diagram of the fifth process of an artificial intelligence-based liquid metal smelting detection and analysis method according to one embodiment of this application.

[0014] Figure 6 This is a schematic diagram of the sixth process of an artificial intelligence-based liquid metal smelting detection and analysis method according to one embodiment of this application.

[0015] Figure 7 This is a schematic diagram of the seventh process of an artificial intelligence-based liquid metal smelting detection and analysis method according to one embodiment of this application.

[0016] Figure 8 This is a schematic diagram of the eighth process of an artificial intelligence-based liquid metal smelting detection and analysis method according to one embodiment of this application.

[0017] Figure 9 This is a schematic diagram of the ninth process of an artificial intelligence-based liquid metal smelting detection and analysis method according to one embodiment of this application. Detailed Implementation

[0018] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-9 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0019] This application discloses an artificial intelligence-based detection and analysis method for liquid metal smelting.

[0020] Reference Figure 1 An artificial intelligence-based detection and analysis method for liquid metal smelting, specifically including: Step S101: Obtain the real-time detection signals of the oxygen constant probe and temperature probe in the smelting furnace, as well as the material formula identification information of the current oxygen constant probe. Calculate the initial oxygen activity data based on the real-time detection signals, and output the basic dataset containing the initial oxygen activity, current temperature, and material formula identification. Among them, the oxygen constant probe and the temperature probe are the core sensing units in the liquid metal smelting site, respectively responsible for monitoring oxygen content and temperature. The oxygen constant probe is based on the principle of oxygen concentration cell, reflecting the difference in oxygen partial pressure between molten steel and reference gas by measuring the oxygen concentration potential signal; the temperature probe uses the thermoelectric effect (such as a platinum-rhodium thermocouple) to convert temperature into a thermoelectric potential signal. The real-time detection signals output by both constitute the raw data layer.

[0021] Meanwhile, the material formulation identification information serves as the probe's identity label, recording its key material characteristics, including the types and proportions of doping elements (such as magnesium, yttrium, praseodymium, and erbium) in the zirconia solid electrolyte tube (which affect ionic conductivity and thermal shock resistance), and the reactivity level of the reference electrode (which determines potential stability). This identification is a prerequisite for the accurate adaptation of the subsequent error compensation model.

[0022] Furthermore, calculating initial oxygen activity data based on real-time detection signals converts physical signals into chemical quantities that can be directly interpreted by the metallurgical process. This can be achieved by relying on fundamental electrochemical laws (such as the Nernst equation): the oxygen concentration potential signal has a logarithmic relationship with oxygen activity, while the thermoelectric potential signal is used to correct for the interference of temperature in oxygen activity calculation (because oxygen activity is directly related to temperature in the Nernst equation). By substituting the two types of potential values ​​into the equation and solving it simultaneously, the influence of temperature can be eliminated, yielding the initial oxygen activity without error correction. The final output dataset (including initial oxygen activity, current temperature, and material formulation identifiers) provides structured input for subsequent error compensation and process decisions, ensuring that the entire process data is traceable and no element is omitted.

[0023] Step S102: Input the current temperature and initial oxygen activity in the basic dataset into the polarization error compensation model corresponding to the material formulation identifier, and output the initial polarization error compensation value that matches the current working condition through the polarization error compensation model. In the liquid metal smelting environment, the oxygen probe experiences a polarization effect when current passes through the solid electrolyte. This polarization effect, where a redox reaction occurs on the electrode surface, creating a polarization potential that hinders ion migration and causing the measured oxygen activity to deviate from the true value. Therefore, this step quantifies and compensates for this error using a polarization error compensation model.

[0024] Specifically, the key to the polarization error compensation model lies in the "corresponding material formulation identifier": probes with different formulations (such as differences in the proportion of doped elements) have vastly different polarization characteristics. Based on historical calibration test data (static experiments of the same formulation probe under multiple temperature and oxygen activity conditions), a multidimensional response surface is fitted, with temperature as the x-axis, initial oxygen activity as the y-axis, and polarization potential deviation as the y-axis. When the current operating condition (temperature and initial oxygen activity in the basic dataset) and the corresponding material formulation identifier are input, the model automatically locates that point on the surface and outputs the initial polarization error compensation value. This quantifies the oxygen activity measurement deviation caused by the polarization effect under the current operating condition, providing a basis for subsequent corrections.

[0025] Step S103: Obtain the cumulative usage time data of the current oxygen probe and input it into the pre-established probe performance decay model to generate a dynamic error amplification coefficient. Multiply the dynamic error amplification coefficient with the initial polarization error compensation value and output the actual polarization error correction value. As a consumable device, the oxygen determination probe's cumulative usage time leads to material performance degradation (such as microcrack propagation in the zirconia electrolyte tube and decreased activity of the reference electrode), thereby amplifying measurement errors. Therefore, this step introduces a probe performance degradation model and a dynamic error amplification coefficient to construct a time-dimensional error superposition compensation logic.

[0026] In this embodiment, the probe performance degradation model is designed as a nonlinear exponential function (e.g., y=a·e). bx +c, where x is the cumulative usage time), this exponential function is derived from empirical observations in materials science: slow decay in the early stage (during the material's stable performance period), accelerated decay in the middle stage (accumulation of micro-defects), and slowing down in the later stage (before reaching the failure threshold). The dynamic error amplification factor is calculated by the probe performance decay model based on the cumulative usage time, and its value range is limited to [1, maximum decay threshold] (e.g., 1~1.5) to avoid overcompensation.

[0027] Next, the dynamic error amplification factor is multiplied with the initial polarization error compensation value. Essentially, this is to superimpose the "error amplification effect caused by material aging" with the "instantaneous polarization error" to output the actual polarization error correction value. This value includes both the polarization deviation under the current working condition and the additional error generated by the probe due to long-term use, thus achieving a comprehensive quantification of the two-dimensional error.

[0028] Step S104: Use the actual polarization error correction value to perform difference correction on the initial oxygen activity in the basic dataset, and output the target oxygen activity data. The initial oxygen activity exhibits a systematic deviation due to polarization effects and material attenuation, which needs to be offset by a correction value. If the correction value indicates that the measured value is too high, the value is reduced through difference calculations (e.g., target oxygen activity = initial oxygen activity - correction value); conversely, it is increased. Subsequently, the corrected target oxygen activity data has been freed from all known error sources, accurately reflecting the oxygen content of the molten steel in the smelting furnace. This provides a reliable basis for subsequent process decisions, ensuring that the accuracy of the measurement results directly determines the precision of the feeding control.

[0029] Step S105: Match the target oxygen activity data with the preset smelting process specification database to generate a feeding control instruction sequence. Among them, the target oxygen activity data needs to be converted into executable process actions, which can be achieved through mapping and matching through the smelting process specification database. This database stores the correspondence table between the target oxygen activity and the amount and type of deoxidized alloys (such as aluminum wire and ferrosilicon) added under different smelting stages (such as converter blowing and LF refining) (based on metallurgical thermodynamics and process experience).

[0030] In this embodiment, based on the current smelting stage (implied in the process context of the basic dataset) and the target oxygen activity, the "gap" that needs to be deoxidized is determined by looking up a table. The difference is calculated in combination with the target composition requirements of the molten steel (such as residual aluminum content), and finally a feeding control instruction sequence is generated. Specifically, it includes the alloy type (such as "aluminum wire"), the added weight (such as "50kg"), the feeding equipment start signal (such as "wire feeder speed setting") and its associated underlying data traceability information.

[0031] Step S106: Perform data desensitization processing on the feeding control command sequence, remove the underlying potential signal features, retain the process feeding features, and generate a desensitized feeding log. It should be noted that while generating feeding action parameters (such as the type and weight of the target alloy to be added) based on the target oxygen activity, the system also packages and attaches the underlying data traceability information that triggers the feeding decision (including the time-series waveform data of the original oxygen concentration potential and thermoelectric potential, electrode impedance parameters, etc., on which the instruction is based) as extended fields into the data object. This composite design allows the instruction sequence to issue action instructions to the execution terminal (such as the wire feeder) and provide complete decision traceability basis to the local control system when circulating within the workshop's local area network.

[0032] However, when it is necessary to upload the operation record corresponding to the instruction sequence to the cloud server for cross-domain data sharing or model training, the underlying potential timing waveform and other characteristics can directly deduce the company's core technical secrets such as the micro-nano doped structure and electrolyte interface reaction mechanism of the field oxygen determination probe. Therefore, it is necessary to perform data desensitization processing on this composite data object.

[0033] In this embodiment, the data anonymization logic is "removing the underlying layer and retaining the process": that is, deleting the time-series waveform data of the original oxygen concentration potential and thermoelectric potential that the generation instruction depends on (such data belongs to the company's core technical secrets), and only retaining the process feeding characteristics (alloy type, weight, equipment signals). Ultimately, the generated anonymized feeding log meets the production audit requirements while avoiding the leakage of sensitive information.

[0034] Step S107: The desensitized feeding log is encrypted and uploaded to the cloud server. At the same time, the feeding control instruction sequence is sent to the workshop execution terminal, and the cloud storage index and terminal execution status identifier are output.

[0035] In some specific embodiments, the desensitized feeding log can be encrypted using the national cryptographic SM4 algorithm (symmetric encryption, high efficiency and security), transmitted through a 5G network slicing channel (ensuring low latency and high reliability), and uploaded to a cloud server for centralized storage. The cloud storage index (such as a URL or hash value) is output for quick location. At the same time, the original feeding control command sequence is sent to the workshop execution terminal (such as a PLC controller) to drive the actions of equipment such as wire feeders and feeding hoppers, and outputs the terminal execution status identifier (such as "in execution", "completed", "fault"), ultimately realizing cloud-edge collaborative control.

[0036] In the above embodiments, a polarization error mapping mechanism driven by material formulation identification and a dynamic attenuation compensation mechanism based on cumulative usage time are introduced. This precisely solves the bottleneck of measurement distortion caused by probe polarization effect and material aging in the extreme conditions of high temperature and low oxygen in liquid metal smelting. It realizes a technological leap from traditional general signal noise reduction to material-level exclusive dynamic error correction, thereby continuously outputting highly reliable target oxygen activity data throughout the entire life cycle of the probe. On this basis, industrial data security is ensured through desensitized and encrypted transmission. By directly anchoring high-precision detection data to the smelting process specification database to generate precise feeding control instructions, alloy waste and steel defects caused by human experience intervention are effectively avoided, providing key technical support for the intelligent goal of "one-click steelmaking".

[0037] Reference Figure 2 As one implementation of step S101, the step of calculating initial oxygen activity data based on real-time detection signals and outputting a basic dataset containing initial oxygen activity, current temperature, and material formulation identifiers includes: Step S201: Obtain the oxygen concentration difference potential signal of the oxygen constant probe and the thermoelectric potential signal of the temperature measuring probe, and output the original signal dataset; Based on the principle of solid electrolyte concentration difference oxygen determination technology, the core component inside the oxygen determination probe is a zirconia solid electrolyte tube with a specific formula. When the probe is inserted into molten steel, due to the oxygen concentration difference between the inside and outside of the electrolyte tube, an electromotive force, i.e., an oxygen concentration difference potential signal, will be generated between the reference electrode and the measuring electrode. At the same time, the thermocouple in the temperature measuring probe generates a thermoelectric potential signal in the high-temperature molten steel due to the temperature difference between its two ends.

[0038] It should be noted that these two signals are extremely weak and susceptible to strong electromagnetic interference, and are low-level physical electrical quantities. They must be acquired in real time through high-precision sensor interface modules and analog-to-digital conversion circuits to form an unprocessed raw signal dataset containing timestamps.

[0039] Step S202: Perform time synchronization processing on the original signal dataset, align the signal timestamps through the edge computing gateway, and output the aligned synchronization signal data; In rapid dynamic processes such as steel refining, temperature fluctuations and changes in oxygen content are strongly physically coupled. If there is a millisecond-level misalignment between the oxygen signal and the temperature signal on the time axis, the oxygen content calculated subsequently will not be the corresponding value at the current actual temperature.

[0040] Therefore, in this embodiment, by introducing a unified high-frequency clock source near the edge computing gateway close to the data source, and performing timestamp embedding and resampling alignment on heterogeneous signals, strict consistency of the output synchronization signal data in the absolute time dimension can be ensured. This provides a time reference guarantee for subsequent joint solutions based on rigorous physicochemical equations. For example, if the oxygen concentration signal sampling rate is 10Hz and the temperature measurement signal is 20Hz, after synchronization processing, both are output at equal intervals of 15Hz, thereby ensuring that each time point corresponds to a unique combination of oxygen concentration potential and thermoelectric potential. Step S203: Extract the thermoelectric potential signal from the synchronization signal data, perform temperature conversion calculation on the thermoelectric potential signal, and output the current temperature data; Thermoelectric potential signals are essentially millivolt-level voltage differences, which cannot be directly used in the logical judgments of metallurgical processes or the solution of complex electrochemical equations. Therefore, they must be converted into current temperature data with clear physical meaning. Specifically, the current temperature data can be extracted through thermocouple signal conversion (such as mapping the thermoelectric potential mV value to degrees Celsius by looking up a scale table), and the output value is a numerical value in degrees Celsius.

[0041] In this embodiment, the conversion process removes useless hardware noise representations at the electrical signal level and extracts core parameters that can truly reflect the real-time thermodynamic state of molten steel. This is the core data processing logic that meets the requirements of steelmaking enterprises for rapid and accurate temperature measurement on site.

[0042] Step S204: Based on the oxygen concentration difference potential signal and current temperature data in the synchronization signal data, calculate the initial oxygen activity data using the Nernst electrochemical equation and output the initial oxygen activity data. Specifically, calculating and outputting initial oxygen activity data using the Nernst electrochemical equation is a physicochemical solution process that converts electrical signals into chemical concentration indices using concentration-based oxygen determination technology. According to the Nernst principle, the oxygen concentration potential depends not only on the oxygen activity in the molten steel but also on the current absolute temperature. Therefore, the oxygen concentration potential aligned in the previous steps and the converted current temperature data must be substituted into the equation for joint solution.

[0043] It should be noted that, since no correction has yet been introduced for the polarization effect of the probe material under high temperature and low oxygen conditions, the initial oxygen activity data calculated at this time includes the inherent measurement bias caused by the physical properties of the material. However, this step establishes the core algorithm framework for fusing the two signals into a single displacement parameter based on the fundamental theory of electrochemistry.

[0044] Step S205: The initial oxygen activity data, current temperature data and pre-acquired material formulation identification information of the current oxygen determination probe are mapped and integrated to output the basic dataset.

[0045] Among these, probes with different formulations exhibit significant differences in polarization response and degradation curves under the same molten steel conditions. By employing data field mapping technology, the initial oxygen activity and current temperature, characterizing real-time operating conditions, can be strongly bound and integrated with material formulation information (including dopant element ratios and activity level identifiers), characterizing the intrinsic properties of the sensor. This results in a multi-dimensional, structured basic dataset, enabling subsequent error calibration models to accurately identify and utilize proprietary compensation strategies matched to the current probe material. This fundamentally addresses the technical pain point of lacking customized technology development tailored to the specific smelting conditions of various enterprises.

[0046] In the above embodiments, through rigorous signal timing alignment and physicochemical equation solving, the chaotic underlying electrical signals are accurately transformed into smelting state parameters with clear physical meaning. Furthermore, the intrinsic material formulation properties of the sensor are deeply bound and fused with real-time detection data, providing high-dimensional and strongly correlated structured data support for subsequent customized polarization error mapping and dynamic attenuation compensation for different doped zirconium oxide materials.

[0047] Reference Figure 3 As one implementation of step S102, the step of inputting the current temperature and initial oxygen activity from the basic dataset into the polarization error compensation model corresponding to the material formulation identifier, and outputting an initial polarization error compensation value matching the current operating condition through the polarization error compensation model includes: Step S301: Extract material formulation identification information from the basic dataset, perform retrieval and matching based on the material formulation identification information in the preset model database, and output the target polarization error compensation model. In the extreme high-temperature, low-oxygen conditions of liquid metal smelting, the polarization effect of a constant-oxygen probe is not a constant value, but rather highly dependent on the specific material properties of the solid electrolyte and reference electrode inside the probe. Different batches or formulations of solid electrolytes (e.g., different stabilizer doping ratios) and reference electrodes prepared using different processes exhibit objective physical differences in their internal oxygen ion mobility and interfacial electron exchange rate. This directly determines that the nonlinear curves of their polarization effects under high-temperature, low-oxygen conditions are drastically different.

[0048] Therefore, by extracting the material formulation identification information that characterizes the intrinsic properties of the probe, the target polarization error compensation model that is completely corresponding to the probe currently inserted with molten steel can be accurately retrieved from the pre-stored model database. This achieves a technological leap from the traditional "general error estimation" to "material-level customized error mapping", ensuring that the subsequent error compensation strategy can truly reflect the physicochemical characteristics of the current specific probe.

[0049] Step S302: Extract current temperature and initial oxygen activity data from the basic dataset, perform input standardization processing on the current temperature and initial oxygen activity data, and output standardized input data; Specifically, the current temperature obtained in real time during the refining process is usually a macroscopic thermodynamic value of thousands of degrees Celsius, while the initial oxygen activity calculated based on the Nernst equation is a concentration characterization parameter of a very small order of magnitude. If these two raw data of vastly different orders of magnitude are directly input into the calculation model, the model will be completely dominated by the temperature parameter with a larger value when solving gradients or measuring spatial distance, thereby shielding the sensitive influence of oxygen activity fluctuations on polarization under low oxygen conditions.

[0050] Therefore, by performing input standardization processing, including data normalization and range scaling, physical parameters of different dimensions are uniformly mapped to a dimensionless standardized numerical space, so that the weight distribution of temperature features and oxygen activity features on the multidimensional response surface tends to be balanced, ensuring the numerical stability of the subsequent model search and calculation process and the ability to accurately capture the joint change features of the two variables.

[0051] Step S303: Input the standardized input data into the target polarization error compensation model, find the initial polarization potential deviation corresponding to the standardized input data through the multidimensional response surface in the target polarization error compensation model, convert the initial polarization potential deviation into a value with the same dimension as oxygen activity, and output the initial polarization error compensation value.

[0052] Among them, the polarization error compensation model is a multidimensional response surface with temperature and initial oxygen activity as independent variables and polarization potential deviation as dependent variable. This response surface is generated by fitting static experimental data accumulated by the probe with the corresponding material formulation in historical calibration tests.

[0053] It is understandable that the direct physical manifestation of the polarization effect at the sensor level is that an additional initial polarization potential deviation is superimposed on the oxygen concentration potential signal. This deviation exhibits a highly complex nonlinear coupling relationship with temperature and oxygen activity. This relationship can be pre-fitted and constructed using a large amount of smelting data collected from the front line of steel enterprises to obtain a multidimensional response surface.

[0054] In this embodiment of the application, after finding the initial polarization potential deviation corresponding to the current operating condition through the multidimensional response surface, since the initial oxygen activity data calculated in the previous steps is based on the total potential including the polarization deviation, in order to achieve the subsequent direct difference correction of the initial oxygen activity, it is necessary to use the inverse function relationship of the Nernst electrochemical equation to reversely convert the found potential deviation into the corresponding oxygen activity deviation value, thereby outputting an initial polarization error compensation value with consistent dimensions that can be directly used for algebraic superposition to eliminate measurement errors.

[0055] Specifically, the core of this step lies in using the local mapping slope of the Nernst equation at a specific operating point to perform an equivalent conversion of micro-deviations. Assuming a specific operating condition in steel refining, the current temperature calculated through previous steps is 1600 degrees Celsius, the initial oxygen activity is 10 ppm, and the initial polarization potential deviation corresponding to the current operating condition is found to be 5 mV using a multidimensional response surface. However, due to the complex nonlinear relationship between potential and oxygen activity, 5 mV cannot be directly treated as a constant for subtraction. The system must combine the current specific operating point of 1600 degrees Celsius and 10 ppm to obtain the partial derivative of the Nernst equation with respect to oxygen activity, i.e., the formula for calculating the partial derivative. E / This allows us to derive the dynamic equivalent conversion ratio between potential change and oxygen activity change within this small neighborhood. Through this partial derivative calculation, we can determine that, under the current specific high-temperature, low-oxygen condition, this dynamic proportionality coefficient is assumed to be 10 mV / ppm. Based on this ratio, the system divides the found 5 mV potential deviation by the proportionality coefficient of 10 mV / ppm, thus accurately calculating the current hidden oxygen activity deviation as 0.5 ppm. This outputs an initial polarization error compensation value with completely consistent dimensions, which can be directly used for subsequent algebraic superposition to eliminate measurement errors.

[0056] In the above embodiments, a multi-dimensional error mapping mechanism driven by both material properties and real-time operating conditions is constructed. Through rigorous dimensional standardization and inverse physical equation calculation logic, the polarization potential deviation at the microscopic material level is accurately converted into an oxygen activity compensation parameter that can be directly executed by the computer. This reduces the poor measurement accuracy and low stability caused by the significant polarization effect of the oxygen electrode material under high temperature and low oxygen conditions of molten steel.

[0057] Reference Figure 4 As one implementation of step S103, the step of obtaining the cumulative usage time data of the current oxygen determination probe and inputting it into a pre-established probe performance degradation model to generate a dynamic error amplification coefficient includes: Step S401: Obtain the unique device identification code of the current oxygen determination probe and the probe life cycle database. Query the probe life cycle database based on the unique device identification code and output the cumulative usage time data. In the high-temperature and highly corrosive environment of liquid metal smelting, the zirconium oxide solid electrolyte tube and reference electrode inside the oxygen determination probe will undergo irreversible microstructural degradation due to repeated immersion in molten steel. This degradation directly leads to a decrease in its ionic conductivity and interfacial reactivity, thereby affecting the measurement accuracy.

[0058] In this embodiment, by configuring a unique device identifier for each probe, the system can accurately locate and extract the total time the probe has spent in the smelting furnace from its initial use to the current moment in the probe lifecycle database. This cumulative usage time data, expressed in hours or minutes, eliminates errors and confusion from manual recording and constitutes the most direct and tamper-proof initial physical indicator for assessing the basic aging state of the probe itself.

[0059] Step S402: Obtain the real-time operating parameters of the current smelting furnace, perform operating condition correction processing on the cumulative usage time data, and output the corrected usage time data. The real-time operating parameters include smelting temperature and steel composition type, and the operating condition correction process includes temperature compensation coefficient calculation and composition weight adjustment.

[0060] Specifically, under harsh refining conditions such as high temperature and low oxygen, the thermal shock stress caused by extremely high temperatures and the chemical corrosion of specific steel components will exponentially accelerate the crystalline phase aging of zirconia tubes. Assuming the physical aging degree of immersion in steel at normal temperature for 1 hour is a baseline of 1, under certain ultra-high temperature and high alkalinity refining conditions, the corrosion acceleration factor calculated based on the temperature and composition characteristics in the acquired real-time operating parameters may be 1.5. In this case, the system does not simply accumulate the time, but performs an equivalent conversion of "corrected usage time = absolute time × operating condition weight coefficient", converting the actual physical time under extreme conditions into an equivalent time under mild conditions, thereby outputting corrected usage time data that can truly reflect the degree of microscopic damage to the material.

[0061] Step S403: Input the corrected usage time data into the pre-established probe performance attenuation model, calculate the probe performance attenuation ratio corresponding to the corrected usage time data through the probe performance attenuation model, configure the probe performance attenuation ratio as a dynamic error amplification factor and output it.

[0062] The probe performance attenuation model is an exponential function model that increases non-linearly with the cumulative usage time. The value range of the dynamic error amplification coefficient is limited to a range greater than or equal to 1 and less than or equal to the preset maximum attenuation threshold.

[0063] Understandably, probe performance degradation is not a simple, uniform linear process, but rather typically exhibits an exponential decay curve with a gradual initial decline followed by a rapid deterioration in the later stages. This step utilizes a mathematical model to transform the microscopic degradation degree of the material into a macroscopic amplification and adjustment parameter for polarization error. Since probe aging exacerbates concentration polarization, the degradation model, based on the corrected usage time data, first assesses the degradation ratio of the probe performance relative to the factory standard, and then calculates the polarization amplification factor caused by this degradation ratio according to a preset mapping relationship. The system directly configures this error amplification factor, which is greater than the baseline value of 1, as the dynamic error amplification coefficient output. This ensures that when multiplied by the initial polarization error compensation value in the main process, the compensation force is scientifically and reasonably amplified, accurately reflecting the actual error deviation state of the aging probe.

[0064] In some embodiments, the attenuation coefficient is close to 1.0 in the initial stage (0-500 hours) when the material is stable, increases exponentially with the accumulation of micro-defects in the middle stage (500-2000 hours) (e.g., y=1+e^(0.001x)), and slows down in the later stage (>2000 hours) as the material failure threshold approaches. By inputting the corrected usage time (e.g., 1584 hours) into the model, a dynamic error amplification factor (e.g., 1.3) is calculated and output through a function. This factor represents the error amplification factor caused by material aging. For example, if the initial polarization error compensation value is 0.5mV, multiplying it by the dynamic error amplification factor of 1.3 results in an actual polarization error correction value of 0.65mV, thereby achieving superimposed compensation for aging errors.

[0065] In the above embodiments, the simple physical timing of the sensor is upgraded to an equivalent lifetime assessment that can accurately characterize the microscopic degradation state of the material. Based on the nonlinear decay model, the equivalent lifetime is transformed into an error amplification factor that characterizes the degree of polarization intensification. This provides an adaptively adjustable mathematical benchmark for the dynamic and accurate compensation of polarization error, eliminating the systematic bias that static compensation cannot cope with the individual aging differences of the probe.

[0066] Reference Figure 5 As a further implementation of the liquid metal smelting detection and analysis method, before the step S103, which involves acquiring the cumulative usage time data of the current oxygen determination probe and inputting it into the pre-established probe performance decay model to generate the dynamic error amplification coefficient, the method further includes: Step S501: Obtain the material degradation feature vector of the current oxygen-determining probe, including the microcrack density value of the zirconia tube and the activity decay rate of the reference electrode; Specifically, in the high-temperature, high-scouring, and multiphase reaction system of liquid metal smelting, the zirconium oxide solid electrolyte of the oxygen determination probe is not only subject to simple thermodynamic aging, but is also prone to micro-crack networks due to frequent thermal shocks or mechanical stress. These micro-cracks block the direct conduction path of oxygen ions and increase the ohmic resistance inside the concentration polarization. At the same time, the reference electrode inside the probe is usually composed of a mixture of metal and its oxide powder. During long-term use, it will undergo irreversible phase transitions due to the disruption of local chemical equilibrium or the penetration of impurities, causing its ability to generate a standard potential to decline as its activity decreases.

[0067] Therefore, by extracting the microscopic physical features of these two dimensions to construct vectors, the system can break away from the traditional crude logic of relying solely on macroscopic usage time to measure probe lifespan, and truly capture the underlying sensing performance degradation caused by physical damage and chemical deactivation of the probe.

[0068] Step S502: Based on the material formulation identification information, the corresponding set of benchmark performance parameters is obtained from the preset material property database, including the standard zirconia lattice constant and the theoretical activity threshold of the reference electrode. Among them, the material formulation identifier records the characteristics of the core materials of the probe (such as the proportion of zirconium oxide doping elements and the composition of the reference electrode), and is the probe's "digital ID card".

[0069] The pre-built material property database stores parameters for different formulations under ideal conditions: the standard zirconia lattice constant represents the optimal ion conduction channel size under a specific doping stabilizer, determining the initial oxygen ion vacancy concentration; the theoretical activity threshold of the reference electrode represents the most perfect electrochemical response capability when uncontaminated, determined by the thermodynamic equilibrium constant. For example, there is a slight but crucial difference in the lattice constant between yttrium-doped zirconia and magnesium-doped zirconia. By retrieving the reference parameter set through identifier matching, subsequent deviation calculations are ensured to use the current probe's specific "factory ideal state" as a reference system, avoiding misjudgments caused by universal benchmarks, and locking in the correct material science anchor for measuring the true degree of degradation.

[0070] Step S503: Calculate the difference based on the material degradation feature vector and the benchmark performance parameter set to generate a relative deviation matrix; Specifically, the measured microcrack density values ​​are correlated with the rate of change of lattice parameters under ideal crack-free conditions (e.g., microcrack propagation leads to lattice distortion, manifested as a negative weakening of the ion conduction cross-sectional area). The deviation rate of the reference electrode activity decay rate from the theoretical activity threshold is quantified (e.g., decreased activity manifests as a reduced ability to build up electrochemical potential). The generated relative deviation matrix not only contains the absolute deviation magnitude but also implicitly includes the deviation direction and physical coupling relationship (e.g., the interactive acceleration effect of microcracks and electrode poisoning). By dimensionality reduction and translation of the complex microscopic damage mechanism of materials science into a high-dimensional feature space, it becomes a data bridge that can be directly analyzed by subsequent artificial intelligence models.

[0071] Step S504: Input the relative deviation matrix into the pre-established material degradation compensation model to calculate and generate the material performance compensation factor; In the embodiments of this application, the material degradation compensation model is deeply trained with a large amount of historical probe dissection data and offline electrochemical impedance spectroscopy data, which enables it to understand the interaction of each element in the relative deviation matrix, such as how the microcrack network accelerates the poisoning process of the reference electrode, or how the lattice micro-distortion superimposes to affect the local oxygen partial pressure.

[0072] After the model receives the matrix, it performs complex internal analysis and maps it into a scalar or low-dimensional vector material property compensation factor. This factor essentially quantifies the proportion of "pathological" errors in the probe measurement signal under the current microscopic physical damage state, in addition to normal polarization error and simple time aging, thus realizing an intelligent leap from pathological analysis to prescription generation.

[0073] Step S505: The material performance compensation factor is fused with the initial model parameters of the pre-established probe performance attenuation model to generate an enhanced probe performance attenuation model. Among them, the initial model parameters of the pre-established probe performance degradation model are the time-performance degradation curve in the statistical average sense (such as the slope / intercept of a large population), which belongs to the static universal model.

[0074] In this embodiment, by fusing the material performance compensation factor with the initial parameters of the model, it is equivalent to applying a bias or scaling transformation in the mathematical space that is determined by the current real physical state of the probe (microcrack density, electrode activity): if the microcrack density is high or the electrode activity is severely degraded, the fusion calculation will distort the distribution of the original model parameters, so that the smooth theoretical curve is dynamically adapted to the steep actual failure curve, and the static model is transformed into a dynamic digital twin, so that the attenuation model can sense the individual physical state of the probe.

[0075] Step S506: Replace the probe performance degradation model with the enhanced probe performance degradation model to generate an enhanced dynamic error amplification factor based on the cumulative usage time data.

[0076] In particular, through seamless model replacement, the cumulative usage time data, which was originally just a single independent variable input, has its mathematical mapping channel completely changed by the preceding physical damage factor after entering the enhanced probe performance degradation model.

[0077] Ultimately, the output enhanced dynamic error amplification factor is no longer a mechanical value that monotonically depends on calendar time or number of uses, but a vivid indicator that deeply integrates the microstructural integrity of the current probe's zirconia tube and the electrochemical activity of the reference electrode. This factor can extremely sensitively identify abnormal probes that have experienced premature collapse of their internal physical structure due to adverse furnace conditions despite the same usage time, thus providing the most practically guiding correction weight for subsequent product calculations with the initial polarization error compensation value.

[0078] In the above implementation, a dynamic intervention mechanism from microscopic material physical degradation to macroscopic algorithm model enhancement is constructed at the front end of the attenuation model. By extracting microscopic features such as zirconia microcracks and reference electrode activity and performing deviation matrix processing with material reference parameters, and using an independent model to refine them into performance compensation factors, the mathematical expression of the attenuation model is reconstructed from the bottom layer. This enables the accurate quantification of the probe's actual physical lesions and their injection into the attenuation extrapolation in the time dimension. Without changing the main process calculation framework, the measurement drift caused by individual physical damage differences of the probe is eliminated, and the robustness and absolute accuracy of liquid metal oxygen activity detection under complex smelting conditions are improved.

[0079] Reference Figure 6 As one implementation of step S104, the step of using the actual polarization error correction value to perform difference correction on the initial oxygen activity data in the basic dataset and outputting the target oxygen activity data includes: Step S601: Receive the initial oxygen activity data and the actual polarization error correction value from the basic dataset, subtract the actual polarization error correction value from the initial oxygen activity data, and output the intermediate oxygen activity data. In this embodiment, since the preceding polarization error compensation model has invoked a dedicated multidimensional response surface based on the material formulation identifier, the millivolt-level polarization potential deviation is equivalently converted into a value with the same dimensions as the initial oxygen activity through the partial derivative of the Nernst equation with respect to oxygen activity. Moreover, this value has been multiplied by the dynamic error amplification factor that reflects the degree of probe aging. At this time, the actual polarization error correction value is mathematically equivalent to the interference increment directly superimposed on the true oxygen activity.

[0080] Therefore, this step directly performs an algebraic operation by subtracting the actual polarization error correction value from the initial oxygen activity data. In a physical sense, this is equivalent to accurately stripping away the artificially inflated oxygen ion concentration caused by the electrode polarization effect. For example, when the initial oxygen activity is measured to be 10 ppm, and the dimensionless correction value after rigorous conversion and dynamic amplification is determined to be 0.6 ppm, the system directly subtracts to obtain an intermediate oxygen activity data of 9.4 ppm, thereby achieving the highest fidelity approximation of the measurement signal to the true metallurgical physical parameters.

[0081] Step S602: Perform boundary constraint verification on the intermediate oxygen activity data based on the preset effective range threshold. When the intermediate oxygen activity data is within the effective range threshold, output the target oxygen activity data; otherwise, acquire and trigger the pre-configured exception handling protocol.

[0082] Specifically, in the high-temperature, high-pressure, and electromagnetically interference environment of steel refining, even after prior material matching correction and error stripping, the calculated intermediate oxygen activity data may still violate the common sense of metallurgical thermodynamics due to occasional physical breakage of the probe, external strong current intrusion, or extreme working conditions exceeding the fitting boundary of the algorithm model. For example, it may show negative values ​​or absurd values ​​that far exceed the oxygen solubility limit at the current molten steel temperature.

[0083] In this embodiment, the preset effective range threshold is dynamically set according to the type of steel being smelted. This threshold is essentially a digital representation of the reasonable oxygen level range for the current steel grade and process stage. By comparing the intermediate oxygen activity data with this threshold, the system can sensitively identify pathological data caused by model failure or hardware damage. Only when the data is within the safe threshold will it be released and officially output as the target oxygen activity data. Once it exceeds the limit, the pre-configured anomaly handling protocol will be immediately obtained. By executing actions such as data blocking, marking invalid, or rolling back to the previous steady-state value, the path of erroneous data to the subsequent feeding instruction generation module will be cut off, thus ensuring industrial data protection and production safety.

[0084] For example, low-carbon steel at 1600°C may have an effective range threshold of 0.5 to 20 ppm. If the intermediate oxygen activity is -2 ppm or 50 ppm, a protocol (discard, rollback to the last valid reading, etc.) is triggered to prevent catastrophic feeding errors downstream.

[0085] In the above implementation, the true oxygen level of molten steel is restored by difference calculation, and a robust barrier against extreme disturbances in the industrial field is constructed by relying on boundary constraint verification and anomaly handling protocol. The two work together to ensure that the oxygen activity data input to the subsequent smelting process specification database for matching has reliability and security, laying a data quality foundation for opening up a smart manufacturing closed loop of high-precision oxygen and aluminum determination and precise alloy addition.

[0086] Reference Figure 7As one implementation of step S105, the step of matching the target oxygen activity data with a preset smelting process specification database to generate a feeding control command sequence includes: Step S701: Obtain the current smelting context information, retrieve the preset smelting process specification database based on the target oxygen activity data and the current smelting context information, and output the matching process specification entries. It should be noted that in the physicochemical process of steel smelting, the absolute value of oxygen activity cannot determine the deoxidation or alloying operation path in isolation. The same oxygen activity value has completely different thermodynamic meanings in different steel matrices or different smelting stages.

[0087] The current smelting context information typically includes macroscopic operating parameters such as the steel type being smelted, the current refining stage, and the cumulative status of each heat. The target oxygen activity data represents the actual microscopic oxygen level in the molten steel. The pre-defined smelting process specification database is essentially a digital refinement and structured storage of a large amount of past successful smelting experience. Internally, it establishes a topological mapping network with macroscopic operating conditions and microscopic oxygen levels as joint index keys, storing a set of rules based on metallurgical thermodynamics and historical experience.

[0088] In this embodiment, the system performs a joint retrieval of target oxygen activity data and current smelting context information. In fact, it accurately locates the standardized operating criteria that meet the current specific steel grade composition requirements and the current physicochemical reaction progress in the multi-dimensional state space, thereby outputting matching process procedure entries. This provides a theoretical basis for subsequent alloying operations that is highly consistent with the current physicochemical environment of the molten pool.

[0089] Step S702: Determine the target alloy addition type based on the matching process specification items, calculate the corresponding target addition weight in combination with the target oxygen activity data, and generate the initial feeding control instruction sequence. Among them, the matched process specification items define the specific categories of deoxidizers or fine-tuning alloys that must be used to achieve the target steel composition at the macro level, which solves the problem of what to add. The accurate solution of the target added weight depends on the high-precision target oxygen activity data output from the previous steps, which solves the problem of how much to add.

[0090] Specifically, the system retrieves the corresponding stoichiometric reaction equation and activity interaction coefficient model based on the determined alloy type. Using the target oxygen activity data as the fundamental input for reactant concentration, it derives the theoretical alloy mass required to completely neutralize or balance the current free oxygen through strict laws of conservation of mass. Since the target oxygen activity data input here has been completely freed from sensor polarization error interference, the alloy weight calculated based on this data possesses extremely high physical accuracy. The system encapsulates this qualitative type information with quantitative weight information, thereby generating an initial feeding control command sequence that conforms to thermodynamic expectations under ideal thermodynamic conditions.

[0091] In this embodiment, the instruction sequence generation is achieved through "alloy type mapping" and "weight calculation": alloy type mapping determines the additive type based on the deoxidation requirements of the steel grade (e.g., aluminum wire is preferred for stainless steel, and ferrosilicon is used for carbon steel); weight calculation combines the target oxygen activity with the correspondence between "oxygen activity - alloy quantity" in the specification entry, and considers the residual element content of the current molten steel (implied in the smelting context information) to make differential adjustments. For example, the specification entry benchmark is "30kg of aluminum wire corresponds to an oxygen activity of 5×10⁻⁶". -5 If the current target oxygen activity is 4 × 10⁻⁶ -5 (Further deoxidation is required), then the weight calculation module outputs "35kg aluminum wire". Finally, the output initial feeding control instruction sequence is a structured data object, which includes the target alloy addition type (e.g., "aluminum wire"), the addition weight (e.g., "35kg"), and the feeding equipment identifier (e.g., "wire feeder #3"), thus providing the original instruction framework for subsequent resource verification.

[0092] Step S703: Obtain the pre-configured resource configuration constraints, perform resource optimization verification on the initial feeding control command sequence based on the resource configuration constraints, associate with the underlying data traceability information, and output the final feeding control command sequence.

[0093] Understandably, in a real factory workshop environment, the initial material feeding control command sequence is often constrained by rigid physical boundaries such as real-time changes in material warehouse inventory, the maximum single conveying capacity of the alloy chute, and the scheduling sequence of the feeding crane. The pre-configured resource allocation constraints are a structured mathematical expression of these underlying physical realities and supply chain states of the workshop. The resource optimization verification process is essentially a multi-objective optimization algorithm process with strong constraints. The system compares the idealized initial material feeding control command sequence with the current resource allocation constraints.

[0094] When the system detects that the instruction demand exceeds the existing constraints, it does not simply report an error and stop. Instead, it performs a secondary reconstruction based on the preset equivalent substitution rules or batch feeding strategy. For example, when the inventory of a certain deoxidizer is insufficient, it automatically triggers the calculation of the equivalent substitute, or when the single feeding exceeds the weight limit, it automatically splits it into a multi-step execution sub-sequence that meets the equipment's load-bearing limit. After this physical feasibility verification and dynamic adaptive adjustment, the output feeding control instruction sequence not only maintains the original metallurgical deoxidation purpose, but also ensures that it can be directly and physically implemented by the execution equipment at the bottom of the workshop without any obstacles.

[0095] In this embodiment, resource optimization verification includes inventory availability checking (verifying whether the alloy inventory meets the required weight) and cost constraint assessment (comparing the alloy unit price with the budget and prioritizing the low-cost combination). For example, if the initial instruction requires "35kg of aluminum wire," but the inventory only has 20kg remaining, the verification module automatically adjusts it to "20kg of aluminum wire + 10kg of ferrosilicon" (based on the deoxidation efficiency conversion of the alternative alloys). If the cost assessment shows that the unit price of aluminum wire is too high, the incremental substitution logic for ferrosilicon is triggered. After successful verification, a feeding control instruction sequence (verified instruction sequence) is output to ensure that the instructions meet both process requirements and adapt to the company's real-time resource status.

[0096] It should be noted that during the generation of the feeding control instruction sequence, the system also captures all related input data residing in the smelting context within the current instruction generation cycle (including the original oxygen concentration differential potential timing waveform, thermoelectric potential signal, probe impedance parameters, and other underlying hardware features obtained from previous steps), and automatically mounts and merges this underlying data as underlying data traceability information onto the core feeding parameters output by the business logic layer. In other words, the final feeding control instruction sequence is a composite structured object that combines business execution instructions and underlying traceability data.

[0097] In the above implementation, the precise matching of metallurgical operation criteria was achieved through multi-dimensional contextual joint retrieval, and a rigorous chemometric quantitative conversion was completed based on high-fidelity oxygen activity data. By introducing resource allocation constraints for verification, a feasibility bridge was built between ideal theory and physical reality. The three are progressive and tightly coupled, eliminating the theoretical deviation and physical lag problems that may occur in the issuance and execution of traditional feeding instructions. This provides a core instruction engine with industrial executability for the fully automatic, high-precision, and unmanned closed-loop control of the liquid metal smelting process.

[0098] Reference Figure 8 As one implementation of step S106, the step of performing data desensitization processing on the feeding control command sequence, removing the underlying potential signal features, and retaining the process feeding features to generate a desensitized feeding log includes: Step S801: Receive the feeding control command sequence and parse the data structure therein to identify the initial data field set containing the underlying potential signal characteristics and process feeding characteristics; To meet the dual requirements of high-precision motion control of the execution terminals within the workshop's local area network and full lifecycle traceability of the local central control system, the material feeding control command sequence is not a single action level signal, but a composite structured data object automatically intercepted and encapsulated by the system's underlying data bus at the moment the command is generated. This object not only contains the material feeding action parameters calculated by the business logic layer, but also snapshots the original hardware input characteristics that triggered the decision in the form of extended fields.

[0099] Therefore, the process of parsing the data structure is essentially to use deserialization technology to read the schema definition of the data object. By identifying field type identifiers and field value range attributes, the originally coupled heterogeneous data streams are broken up, and it is possible to accurately define which key-value pairs belong to the underlying potential signal characteristics (such as oxygen concentration difference potential timing waveform code, electrode impedance dynamic parameters, and probe impedance parameters) that characterize the underlying electrochemical reaction physical process, and which key-value pairs belong to the process feeding characteristics (such as target alloy addition type and target addition weight) that characterize the macroscopic material mixing action. In this way, an initial data field set is constructed to ensure that the subsequent desensitization operation can be targeted and avoid accidental damage to core business data.

[0100] Step S802: Based on the pre-configured sensitive field filtering rules, perform feature separation processing on the initial data field set, remove the underlying potential signal features, retain the process feeding features, and output the intermediate feature set; It should be noted that in the field of intelligent manufacturing in metallurgy, the underlying oxygen concentration difference potential and thermoelectric potential time-series waveforms, as well as the probe impedance parameters, contain the polarization response law and attenuation dynamics of the oxygen determination probe under specific high temperature and low oxygen environments. If these data are illegally extracted, competitors can easily reverse engineer them to deduce the company's core technical secrets, such as the micro-nano doped structure of the oxygen determination probe, the electrode catalytic activity, and the specific material formulation ratio.

[0101] The pre-configured sensitive field filtering rules in this application embodiment are dynamically generated based on this data security hierarchical defense logic. When the system performs feature separation processing, it uses this rule as a judgment benchmark to perform a security scan on each node in the initial data field set, erasing highly sensitive fields such as potential waveform feature codes and impedance parameters from the memory space, while strictly retaining process feeding features such as alloy type and weight used to describe the feeding facts. This precise removal based on field-level granularity not only blocks the path of core technical parameters leaking to the outside through the network, but also completely retains the objective business facts reflecting production cycle and material consumption, outputting a pure and secure intermediate feature set.

[0102] Step S803: Perform log structure encapsulation processing on the intermediate feature set to generate desensitized material feeding logs.

[0103] Specifically, the system calls a predefined log serialization template to reorganize the process feeding features in the intermediate feature set according to a standardized data exchange format. It can be encapsulated in JSON-LD format and automatically injects necessary contextual metadata such as timestamps and steel type identifiers during the encapsulation process, so that the generated desensitized feeding log is physically transformed into a self-contained and self-explanatory standardized data entity.

[0104] In the above implementation, without interfering with the front-end business logic calculation, the full amount of data required for control execution in the industrial control network and the desensitized data required for cloud sharing are precisely decoupled. This not only safeguards the security and compliance bottom line of the underlying hardware secrets such as the core material formula of the reverse probe, but also maximizes the release of the data element value of process feeding characteristics in the cloud-based intelligent manufacturing closed loop.

[0105] Reference Figure 9 As a further implementation of the liquid metal smelting detection and analysis method, after step S107, which outputs the cloud storage index and terminal execution status identifier, the method further includes: Step S901: Obtain the composition analysis data of the finished steel after the current smelting furnace is completed, and obtain the historical target oxygen activity data corresponding to the generation of the feeding control command sequence. The composition analysis data of the finished molten steel typically requires offline processes such as sampling, sample preparation, and spectral analysis. This data represents the absolute true physicochemical state that can only be obtained after the entire smelting furnace cycle is completed. By retrieving historical target oxygen activity data with a timestamp strictly bound to that furnace cycle, the system precisely aligns the offline, delayed final test results with the online, real-time decision-making basis in time and space. This provides an unalterable, two-way comparison source for measuring the accuracy of the preceding measurement links, ensuring that subsequent error tracing has a clear physical orientation.

[0106] Step S902: Based on the actual residual amount of deoxidized alloy in the composition analysis data of the finished steel, calculate the corresponding theoretical target oxygen activity benchmark value, and calculate the difference between the theoretical target oxygen activity benchmark value and the historical target oxygen activity data to generate oxygen activity measurement deviation characteristics. Specifically, by utilizing inverse metallurgical thermodynamics to eliminate interference in the feeding process, the systematic residuals of the measurement model itself can be accurately quantified. Because the feeding equipment may experience physical interference such as mechanical wear or uneven steel mixing during the feeding action, directly comparing the final oxygen content with historical target oxygen activity can easily introduce noise caused by non-measurement factors.

[0107] Therefore, by extracting the actual residual amount of deoxidized alloys in the finished steel (such as the actual remaining aluminum and silicon content), the system uses the principle of thermodynamic equilibrium to deduce the absolute equilibrium oxygen concentration that the steel should possess at the current temperature and the concentration of the alloy residue, i.e., the theoretical target oxygen activity benchmark value. The difference between this benchmark value derived from thermodynamic laws and the previously output historical target oxygen activity data generates an oxygen activity measurement deviation characteristic that eliminates physical fluctuations from subsequent feeding operations, reflecting the net measurement deviation caused by the incomplete elimination of polarization effects or insufficient estimation of material aging during this measurement cycle.

[0108] Step S903: Determine whether the oxygen activity measurement deviation characteristic is greater than the pre-configured error correction threshold; if yes, proceed to step S904; if no, do not perform any operation and end the process.

[0109] The purpose of setting an error correction threshold is to filter out meaningless disturbances caused by normal metallurgical process fluctuations or minor laboratory errors, and to trigger deep model intervention only when the deviation exceeds the allowable boundary.

[0110] Step S904: Obtain the historical basic dataset recorded when generating the feeding control command sequence, and extract the historical current temperature, historical initial oxygen activity and historical cumulative usage time data from it; Since the oxygen activity measurement bias characteristic is only a scalar result, in order to use it to correct complex nonlinear decay models, the complete context in which the bias occurred must be reconstructed.

[0111] In this embodiment, the extracted historical current temperature and historical initial oxygen activity constitute the thermodynamic polarization load boundary that the probe experiences at that instant, reflecting the intensity of ion migration inside the probe; while the historical cumulative usage time data objectively records the degree of fatigue accumulation of the probe's solid electrolyte under the scouring and chemical corrosion of high-temperature molten steel.

[0112] Step S905: Calculate the historical smelting condition characteristics based on the historical current temperature and historical initial oxygen activity, and integrate the historical smelting condition characteristics with the historical cumulative usage time data to generate the condition-related aging compensation value. Specifically, under the triggering conditions, the system does not simply attribute aging to time, but rather deeply recognizes that under extreme conditions such as high temperature and low oxygen, the phase transition acceleration of the probe electrolyte and the electrode poisoning effect are amplified exponentially. Therefore, by mapping the historical current temperature and historical initial oxygen activity to historical smelting condition characteristics that characterize the severity of the environment, and using these as acceleration weights to fuse with historical cumulative usage time data, the generated condition-related aging compensation value is no longer a monotonic time function, but a coupled physical quantity that deeply integrates the severity of the environment and the service time, truly restoring the actual equivalent aging damage experienced by the probe under specific extreme conditions.

[0113] Step S906: Using the working condition-related aging compensation value as the correction target value and the historical cumulative usage time data as the input variable, the model parameters of the probe performance degradation model are iteratively updated, and the updated probe performance degradation model is output.

[0114] Specifically, the working condition-related aging compensation value serves as the standard answer label, while the historical cumulative usage time data serves as the independent variable feature. The system continuously adjusts the weights or slope parameters of the nonlinear function inside the probe performance degradation model through optimization algorithms (such as gradient descent), so that when the model is input for a specific duration, it can accurately approximate and output a compensation value that includes the working condition acceleration effect.

[0115] Understandably, this iterative update mechanism breaks through the limitations of traditional static lookup table methods or fixed curves that cannot adapt to slight changes in material formulations and drastic changes in operating conditions. This enables the probe performance degradation model to have the ability of "experience memory" and "continuous learning," ensuring that when the model faces similar high temperature and low oxygen harsh conditions again, it can predict in advance and output a larger dynamic error amplification factor, thereby fundamentally eliminating the measurement drift caused by aging underestimation.

[0116] In the above implementation, a closed-loop self-evolving data link is constructed to optimize and iterate the underlying sensing error compensation model from the cloud-based execution terminal results. By using thermodynamic reverse deduction to remove non-measurement interference, pure measurement deviation is accurately extracted. The severity of extreme working conditions is used as an acceleration factor and fused into the cumulative usage time of the probe to generate working condition-related aging compensation values. These values ​​are then used as monitoring signals to drive the parameter iteration of the probe performance degradation model. This effectively overcomes the technical defects of existing technologies that cannot take into account both the acceleration effect of extreme working conditions and personalized probe aging. Without increasing any hardware costs, adaptive growth of the model at the software level is achieved, providing a continuously reliable data-driven engine for high-precision, maintenance-free detection throughout the entire life cycle of liquid metal smelting.

[0117] It should be noted that in practical applications, due to potential sampling costs or accuracy limitations in microscopic physical detection (steps S501-S506), it is impossible to fully capture all complex aging mechanisms. Therefore, steps S901-S906 are needed to continuously optimize the basic parameters of the degradation model through closed-loop feedback from macroscopic smelting results. Conversely, the model parameter updates in steps S901-S906 often have a lag (updated only after a heat of steel is completed), making them ineffective against sudden microscopic physical damage in the current heat (such as drastic polarization changes caused by accidental fragmentation). In this case, steps S501-S506 enhance the model by introducing real-time material degradation feature vectors, effectively compensating for this time lag. These two technical solutions, one preceding and one following, one microscopic and one macroscopic, together construct a highly robust polarization error elimination system.

[0118] This application also discloses an artificial intelligence-based detection and analysis system for liquid metal smelting.

[0119] An artificial intelligence-based detection and analysis system for liquid metal smelting specifically includes: The dataset construction module is used to acquire the real-time detection signals of the oxygen constant probe and temperature probe in the smelting furnace, as well as the material formula identification information of the current oxygen constant probe. Based on the real-time detection signals, the initial oxygen activity data is calculated, and the basic dataset containing the initial oxygen activity, current temperature and material formula identification is output. The polarization error mapping compensation module is used to input the current temperature and initial oxygen activity in the basic dataset into the polarization error compensation model corresponding to the material formulation identifier, and output the initial polarization error compensation value that matches the current working condition through the polarization error compensation model. The error comprehensive correction module is used to obtain the cumulative usage time data of the current oxygen probe and input it into the pre-established probe performance degradation model to generate a dynamic error amplification coefficient. The dynamic error amplification coefficient is multiplied with the initial polarization error compensation value to output the actual polarization error correction value. The oxygen activity correction module is used to perform difference correction on the initial oxygen activity in the basic dataset using the actual polarization error correction value, and output the target oxygen activity data. The feeding control module is used to match the target oxygen activity data with the preset smelting process specification database to generate a feeding control instruction sequence. The data desensitization module is used to perform data desensitization processing on the feeding control command sequence, remove the underlying potential signal features, retain the process feeding features, and generate desensitized feeding logs. The data transmission module is used to encrypt and upload the desensitized feeding logs to the cloud server, while simultaneously sending the feeding control command sequence to the workshop execution terminal and outputting the cloud storage index and terminal execution status identifier.

[0120] An artificial intelligence-based liquid metal smelting detection and analysis system according to an embodiment of this application can implement any of the above methods, and the specific working process of each module in the system can refer to the corresponding process in the above method embodiments.

[0121] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0122] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for detecting and analyzing liquid metal smelting based on artificial intelligence, characterized in that, The method includes: The system acquires real-time detection signals from oxygen constant probe and temperature probe in the smelting furnace, as well as the material formula identification information of the current oxygen constant probe. Based on the real-time detection signals, it calculates the initial oxygen activity data and outputs a basic dataset containing the initial oxygen activity, current temperature, and material formula identification. The current temperature and initial oxygen activity in the basic dataset are input into the polarization error compensation model corresponding to the material formulation identifier, and the polarization error compensation model outputs an initial polarization error compensation value that matches the current operating condition. The cumulative usage time data of the current oxygen determination probe is obtained and input into the pre-established probe performance degradation model to generate a dynamic error amplification coefficient. The dynamic error amplification coefficient is multiplied with the initial polarization error compensation value to output the actual polarization error correction value. The initial oxygen activity in the basic dataset is corrected by difference using the actual polarization error correction value, and the target oxygen activity data is output. The target oxygen activity data is matched with a preset smelting process specification database to generate a feeding control instruction sequence; Data desensitization processing is performed on the feeding control command sequence to remove the underlying potential signal features while retaining the process feeding features to generate a desensitized feeding log; The desensitized feeding log is encrypted and uploaded to the cloud server, while the feeding control command sequence is sent to the workshop execution terminal, and the cloud storage index and terminal execution status identifier are output.

2. The method for detecting and analyzing liquid metal smelting based on artificial intelligence according to claim 1, characterized in that, The steps for calculating initial oxygen activity data based on the real-time detection signal and outputting a basic dataset containing initial oxygen activity, current temperature, and material formulation identifiers include: Acquire the oxygen concentration difference potential signal from the oxygen constant probe and the thermoelectric potential signal from the temperature measuring probe, and output the raw signal dataset; The original signal dataset is subjected to time synchronization processing. The signal timestamps are aligned through the edge computing gateway, and the aligned synchronization signal data is output. Extract the thermoelectric potential signal from the synchronization signal data, perform temperature conversion calculation on the thermoelectric potential signal, and output the current temperature data; Based on the oxygen concentration potential signal in the synchronization signal data and the current temperature data, the initial oxygen activity data is calculated using the Nernst electrochemical equation, and the initial oxygen activity data is output. The initial oxygen activity data, the current temperature data, and the pre-acquired material formulation identification information of the current oxygen determination probe are mapped and integrated to output a basic dataset.

3. The method for detecting and analyzing liquid metal smelting based on artificial intelligence according to claim 2, characterized in that, The steps of inputting the current temperature and initial oxygen activity from the basic dataset into the polarization error compensation model corresponding to the material formulation identifier, and outputting an initial polarization error compensation value matching the current operating condition through the polarization error compensation model include: Material formulation identification information is extracted from the basic dataset, and a search and matching is performed in the preset model database based on the material formulation identification information to output the target polarization error compensation model. Extract current temperature and initial oxygen activity data from the base dataset, perform input standardization processing on the current temperature and initial oxygen activity data, and output standardized input data; The standardized input data is input into the target polarization error compensation model. The initial polarization potential deviation corresponding to the standardized input data is found through the multidimensional response surface in the target polarization error compensation model. The initial polarization potential deviation is converted into a value with the same dimension as oxygen activity, and the initial polarization error compensation value is output.

4. The method for detecting and analyzing liquid metal smelting based on artificial intelligence according to claim 1, characterized in that, The steps for obtaining the cumulative usage time data of the current oxygen determination probe and inputting it into a pre-established probe performance degradation model to generate a dynamic error amplification factor include: Obtain the unique device identifier of the current oxygen determination probe and the probe lifecycle database, query the probe lifecycle database based on the unique device identifier, and output the cumulative usage time data; Obtain the real-time operating parameters of the current smelting furnace, perform operating condition correction processing on the cumulative usage time data, and output the corrected usage time data; The corrected usage duration data is input into a pre-established probe performance attenuation model. The probe performance attenuation ratio corresponding to the corrected usage duration data is calculated using the probe performance attenuation model. The probe performance attenuation ratio is configured as a dynamic error amplification factor and output.

5. The method for detecting and analyzing liquid metal smelting based on artificial intelligence according to claim 4, characterized in that, Before the step of obtaining the cumulative usage time data of the current oxygen determination probe and inputting it into the pre-established probe performance degradation model to generate the dynamic error amplification factor, the following steps are also included: Obtain the material degradation feature vector of the current oxygen-determining probe, including the microcrack density value of the zirconia tube and the activity decay rate of the reference electrode; Based on the material formulation identification information, the corresponding set of benchmark performance parameters, including the standard zirconia lattice constant and the theoretical activity threshold of the reference electrode, are obtained from the pre-set material property database. A relative deviation matrix is ​​generated by calculating the difference between the material degradation feature vector and the benchmark performance parameter set. The relative deviation matrix is ​​input into a pre-established material degradation compensation model to calculate and generate a material performance compensation factor; The material performance compensation factor is fused with the initial model parameters of the pre-established probe performance attenuation model to generate an enhanced probe performance attenuation model. The enhanced probe performance degradation model is replaced by the enhanced probe performance degradation model to generate an enhanced dynamic error amplification factor based on the cumulative usage time data.

6. The method for detecting and analyzing liquid metal smelting based on artificial intelligence according to claim 1, characterized in that, The steps of using the actual polarization error correction value to perform difference correction on the initial oxygen activity data in the basic dataset and outputting the target oxygen activity data include: Receive initial oxygen activity data and actual polarization error correction value from the basic dataset, subtract the actual polarization error correction value from the initial oxygen activity data, and output intermediate oxygen activity data; The intermediate oxygen activity data is subjected to boundary constraint verification based on a preset effective range threshold. When the intermediate oxygen activity data is within the effective range threshold, the target oxygen activity data is output; otherwise, a pre-configured exception handling protocol is acquired and triggered.

7. The method for detecting and analyzing liquid metal smelting based on artificial intelligence according to claim 6, characterized in that, The steps of matching the target oxygen activity data with a preset smelting process specification database to generate a feeding control command sequence include: Obtain the current smelting context information, retrieve the preset smelting process specification database based on the target oxygen activity data and the current smelting context information, and output the matching process specification entries. The target alloy addition type is determined based on the matched process procedure entries, and the corresponding target addition weight is calculated in combination with the target oxygen activity data to generate an initial feeding control instruction sequence. Obtain the pre-configured resource configuration constraints, perform resource optimization verification processing on the initial feeding control command sequence based on the resource configuration constraints, associate with the underlying data traceability information, and output the final feeding control command sequence.

8. The method for detecting and analyzing liquid metal smelting based on artificial intelligence according to claim 7, characterized in that, The steps of performing data desensitization processing on the feeding control command sequence, removing the underlying potential signal features, and retaining the process feeding features to generate a desensitized feeding log include: Receive the feeding control command sequence and parse the data structure therein to identify the initial data field set containing the underlying potential signal characteristics and process feeding characteristics; Based on the pre-configured sensitive field filtering rules, feature separation processing is performed on the initial data field set to remove the underlying potential signal features, retain the process feeding features, and output an intermediate feature set. The intermediate feature set is subjected to log structure encapsulation processing to generate the desensitized material feeding log.

9. A method for detecting and analyzing liquid metal smelting based on artificial intelligence according to any one of claims 1 to 8, characterized in that, After the steps of outputting the cloud storage index and the terminal execution status identifier, the following are also included: Obtain the composition analysis data of the finished molten steel after the current smelting furnace cycle ends, and obtain the historical target oxygen activity data corresponding to the generation of the feeding control command sequence; Based on the actual residual amount of deoxidized alloy in the composition analysis data of the finished steel, the corresponding theoretical target oxygen activity benchmark value is calculated, and the difference between the theoretical target oxygen activity benchmark value and the historical target oxygen activity data is calculated to generate oxygen activity measurement deviation characteristics. Determine whether the oxygen activity measurement deviation characteristic is greater than the pre-configured error correction threshold; if so, obtain the historical basic dataset recorded when generating the feeding control command sequence, and extract the historical current temperature, historical initial oxygen activity, and historical cumulative usage time data from it; Based on the historical current temperature and the historical initial oxygen activity, the characteristics of historical smelting conditions are calculated, and the characteristics of historical smelting conditions are fused with the historical cumulative usage time data to generate condition-related aging compensation values. Using the working condition-related aging compensation value as the correction target value and the historical cumulative usage time data as the input variable, the model parameters of the probe performance degradation model are iteratively updated, and the updated probe performance degradation model is output.

10. An artificial intelligence-based liquid metal smelting detection and analysis system, used to execute the artificial intelligence-based liquid metal smelting detection and analysis method according to any one of claims 1 to 9, characterized in that, The system includes: The dataset construction module is used to acquire the real-time detection signals of the oxygen constant probe and temperature probe in the smelting furnace, as well as the material formula identification information of the current oxygen constant probe. Based on the real-time detection signals, the module calculates the initial oxygen activity data and outputs a basic dataset containing the initial oxygen activity, current temperature, and material formula identification. The polarization error mapping compensation module is used to input the current temperature and initial oxygen activity in the basic dataset into the polarization error compensation model corresponding to the material formulation identifier, and output the initial polarization error compensation value that matches the current working condition through the polarization error compensation model. The error comprehensive correction module is used to obtain the cumulative usage time data of the current oxygen probe and input it into the pre-established probe performance decay model to generate a dynamic error amplification coefficient. The dynamic error amplification coefficient is multiplied with the initial polarization error compensation value to output the actual polarization error correction value. The oxygen activity correction module is used to perform difference correction on the initial oxygen activity in the basic dataset using the actual polarization error correction value, and output the target oxygen activity data. The feeding control module is used to match the target oxygen activity data with a preset smelting process specification database to generate a feeding control instruction sequence. The data desensitization module is used to perform data desensitization processing on the feeding control command sequence, remove the underlying potential signal features, retain the process feeding features, and generate a desensitized feeding log. The data transmission module is used to encrypt and upload the desensitized feeding log to the cloud server, and at the same time send the feeding control command sequence to the workshop execution terminal, and output the cloud storage index and the terminal execution status identifier.

Images