Intelligent self-adaptive real-time calibration method and system for liquid metal detection error

Abstract

The application relates to an intelligent self-adaptive real-time calibration method and system for liquid metal detection errors, which comprises the following steps: collecting three types of signals of liquid metal through temperature, oxygen potential and internal resistance sensors, generating a synchronous multi-source signal sequence through time stamp alignment; extracting the internal resistance change rate, combining temperature and oxygen potential to construct a polarization feature vector; according to the smelting working condition stage identifier, the initial reference curve is called, the dynamic reference correction model is input for offset compensation, and a dynamic compensation reference sequence is generated; the deviation of the oxygen potential and the dynamic compensation reference is calculated, and the segmented mapping model is input to output the target calibration parameter set; the sensor gain and reference voltage are adjusted to re-quantize the oxygen content data after calibration; the working condition feature set is uploaded to the cloud, the segmented mapping model is refitted, and the model update configuration file is issued. The application realizes full-link adaptive calibration from signal collection to cloud optimization, and improves the oxygen content detection precision and long-term stability.

Description

Technical Field

[0001] This application relates to the field of metallurgical testing and automation control technology, and in particular to an intelligent adaptive real-time calibration method and system for liquid metal detection errors. Background Technology

[0002] With the profound transformation of the modern steel metallurgical industry towards refinement and intelligence, accurate component detection in the liquid metal smelting process has become a core link in ensuring steel quality, reducing deoxidizer consumption, and optimizing process control. Among numerous metallurgical parameters, real-time monitoring of oxygen content is particularly critical. Currently, the industry commonly uses oxygen-determining probes based on the principle of solid electrolyte concentration cells to achieve this goal. These probes typically integrate a temperature sensor, an oxygen potential sensor, and an internal resistance sensor, which are used to sense the molten pool temperature, oxygen concentration potential, and the probe's own internal resistance, respectively. The collected potential signals are then used to calculate the oxygen activity or oxygen content in the liquid metal using the Nernst equation.

[0003] However, since the steel smelting process involves multiple complex stages such as melting, oxidation, and reduction, the physical and chemical properties and temperature ranges of molten steel vary greatly at different stages. In addition, the smelting site is usually accompanied by high-frequency and strong electromagnetic interference generated by electric arc furnaces, frequency conversion equipment, etc. This places extremely stringent requirements on the signal acquisition accuracy, anti-interference ability, and subsequent signal fusion processing of multi-source sensors.

[0004] Currently, existing technologies for detecting oxygen content in liquid metals still face measurement inaccuracies in harsh smelting environments. Oxygen probes inevitably experience concentration and electrochemical polarization in unsteady, highly fluctuating smelting environments. Furthermore, as the probe ages with continued use or experiences severe thermal shock, the ion conductivity of its electrolyte interface undergoes transient fluctuations. This causes traditional static reference curves to fail to accurately track and reflect the drift patterns of the probe's true state. Moreover, existing sensor signal conditioning hardware circuits typically use fixed gain coefficients and reference voltages for analog signal quantization. The inherent contradiction between this fixed configuration and complex nonlinear error characteristics leads to the continuous accumulation of systematic deviations in the detection signal chain. Consequently, the final output oxygen content data fails to accurately reflect the actual chemical composition of the liquid metal, thus limiting the efficiency and reliability of closed-loop control in the smelting process. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides an intelligent adaptive real-time calibration method and system for liquid metal detection errors.

[0006] Firstly, this application provides an intelligent adaptive real-time calibration method for liquid metal detection errors, employing the following technical solution: Temperature signals, oxygen potential signals, and probe internal resistance signals of liquid metal are collected by a temperature sensor, an oxygen potential sensor, and an internal resistance sensor, respectively. The signals are then time-stamped to generate a synchronous multi-source signal sequence. Extract the rate of change of the probe internal resistance signal in the synchronous multi-source signal sequence, and construct a polarization feature vector by combining the temperature signal and the oxygen potential signal; Obtain the current smelting process condition stage identifier, and retrieve the corresponding initial reference curve based on the condition stage identifier. The polarization feature vector and the working condition stage identifier are input into a preset dynamic benchmark correction model to perform offset compensation on the initial benchmark curve, generating a dynamic compensation benchmark sequence. Calculate the deviation between the oxygen potential signal in the synchronous multi-source signal sequence and the dynamic compensation reference sequence, input the deviation and the polarization feature vector into a pre-constructed piecewise mapping model, and output the target calibration parameter set; Based on the target calibration parameter set, hardware control instructions are generated to adjust the gain coefficient and reference voltage of the oxygen potential sensor, and the oxygen potential signal is requantized to generate calibrated oxygen content detection data. The deviation value and the working condition stage identifier are aggregated into a working condition feature set and uploaded to the cloud server. The cloud server refits the segmented mapping model based on the working condition feature set, generates an updated model configuration file, and distributes it to replace the local configuration file.

[0007] By adopting the above technical solution, the measurement inaccuracies caused by probe polarization effect, drastic fluctuations in operating conditions, and device aging in liquid metal detection are effectively solved, achieving closed-loop calibration across the entire chain. This technical solution not only improves the real-time accuracy and stability of oxygen content detection and overcomes the shortcomings of traditional static calibration in adapting to environmental changes, but also endows the sensor system with self-learning and optimization capabilities through an edge-cloud collaborative mechanism, reducing the frequency of manual intervention and providing a highly reliable data foundation for the intelligent and precise control of the steel smelting process.

[0008] Secondly, this application provides an intelligent adaptive real-time calibration system for liquid metal detection errors, employing the following technical solution: The multi-source signal synchronous acquisition module is used to acquire the temperature signal, oxygen potential signal and probe internal resistance signal of liquid metal through temperature sensor, oxygen potential sensor and internal resistance sensor respectively, and perform time stamp alignment processing on the signals to generate a synchronous multi-source signal sequence. The polarization feature vector construction module is used to extract the rate of change of the probe internal resistance signal in the synchronous multi-source signal sequence, and to construct a polarization feature vector by combining the temperature signal and the oxygen potential signal. The working condition identification and reference curve retrieval module is used to obtain the working condition stage identifier of the current smelting process and retrieve the corresponding initial reference curve according to the working condition stage identifier. The dynamic benchmark compensation module is used to input the polarization feature vector and the working condition stage identifier into a preset dynamic benchmark correction model to perform offset compensation on the initial benchmark curve and generate a dynamic compensation benchmark sequence. The target calibration parameter generation module is used to calculate the deviation between the oxygen potential signal in the synchronous multi-source signal sequence and the dynamic compensation reference sequence, input the deviation and the polarization feature vector into a pre-constructed piecewise mapping model, and output the target calibration parameter set. The hardware control and signal quantization module is used to generate hardware control instructions based on the target calibration parameter set, adjust the gain coefficient and reference voltage of the oxygen potential sensor, requantize the oxygen potential signal, and generate calibrated oxygen content detection data. The cloud-based collaborative model update module is used to aggregate the deviation value and the working condition stage identifier into a working condition feature set and upload it to the cloud server. The cloud server refits the segmented mapping model according to the working condition feature set, generates an updated model configuration file, and distributes it to replace the local configuration file.

[0009] Thirdly, this application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the methods in the first aspect.

[0010] In summary, this application includes at least one of the following beneficial technical effects: This application breaks through the limitations of static calibration and fixed hardware configuration in traditional liquid metal detection, which cannot adapt to complex smelting conditions. By strictly synchronizing multi-source signals and accurately extracting polarization features, combined with staged dynamic reference compensation, it effectively eliminates the probe polarization effect and measurement drift caused by drastic changes in operating conditions. At the same time, relying on the segmented mapping model and software-hardware collaborative mechanism, the complex error characteristics are transformed into hardware-level real-time adjustment of the gain of the front-end conditioning circuit and the reference voltage, cutting off the accumulation path of nonlinear errors in the signal link. Finally, with the help of the edge-cloud collaborative mechanism, the detection system is endowed with self-learning and evolution capabilities throughout its entire life cycle. Thus, high-precision and high-robust real-time monitoring of oxygen content is achieved in extremely harsh smelting environments, providing a solid and reliable data foundation for intelligent closed-loop control of the steel smelting process. Attached Figure Description

[0011] Figure 1 This is a first flowchart illustrating an intelligent adaptive real-time calibration method for liquid metal detection error according to one embodiment of this application.

[0012] Figure 2 This is a second flowchart illustrating the intelligent adaptive real-time calibration method for liquid metal detection error according to one embodiment of this application.

[0013] Figure 3 This is a schematic diagram of the third process of an intelligent adaptive real-time calibration method for liquid metal detection error according to one embodiment of this application.

[0014] Figure 4 This is a schematic diagram of the fourth process of an intelligent adaptive real-time calibration method for liquid metal detection error according to one embodiment of this application.

[0015] Figure 5 This is a schematic diagram of the fifth step of the intelligent adaptive real-time calibration method for liquid metal detection error according to one embodiment of this application.

[0016] Figure 6 This is a schematic diagram of the sixth process of an intelligent adaptive real-time calibration method for liquid metal detection error according to one embodiment of this application.

[0017] Figure 7 This is a schematic diagram of the seventh process of an intelligent adaptive real-time calibration method for liquid metal detection error according to one embodiment of this application.

[0018] Figure 8 This is a schematic diagram of the eighth step of the intelligent adaptive real-time calibration method for liquid metal detection error according to one embodiment of this application. Detailed Implementation

[0019] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-8 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.

[0020] This application discloses an intelligent adaptive real-time calibration method for liquid metal detection errors.

[0021] Reference Figure 1 A smart adaptive real-time calibration method for liquid metal detection error, the specific method includes: Step S101: The temperature signal, oxygen potential signal and probe internal resistance signal of liquid metal are collected by temperature sensor, oxygen potential sensor and internal resistance sensor respectively, and the signals are time-stamp aligned to generate a synchronous multi-source signal sequence. Since temperature sensors, oxygen potential sensors, and internal resistance sensors are usually sampled independently and have different signal transmission paths, direct fusion can lead to feature misalignment.

[0022] Therefore, firstly, a finite impulse response filter is used to perform low-pass filtering on each signal to filter out high-frequency electromagnetic noise generated by electric arc furnaces, frequency converters, etc., to ensure signal purity. Then, according to a preset constant sampling frequency, a uniform timestamp is added to each set of filtered data. The temperature, oxygen potential, and internal resistance values ​​captured at the same time point are encapsulated into a three-dimensional vector, and a synchronous multi-source signal sequence is constructed in chronological order. This step provides a strict time reference for multivariate coupling analysis, ensuring the accuracy of subsequent polarization feature extraction.

[0023] Step S102: Extract the rate of change of probe internal resistance signal in the synchronous multi-source signal sequence, and construct a polarization feature vector by combining temperature signal and oxygen potential signal; Specifically, the internal resistance change rate is obtained by calculating the difference between the probe's internal resistance signal at the current moment and the previous moment and dividing it by the base value. This index directly reflects the transient fluctuations in the ion conduction characteristics of the electrolyte interface. At the same time, the temperature gradient is calculated to characterize the stress effect of thermal shock on the probe's physical structure. Subsequently, the internal resistance change rate, temperature gradient, and current oxygen potential signal are normalized to eliminate the magnitude differences between different physical dimensions, and then they are concatenated into a polarization feature vector.

[0024] In this embodiment, the polarization feature vector is a multi-dimensional feature descriptor that can accurately characterize the polarization degree of the probe under unsteady smelting conditions, providing a mathematical basis for distinguishing normal measurement signals from polarization interference signals.

[0025] Step S103: Obtain the current working condition stage identifier of the smelting process, and retrieve the corresponding initial reference curve according to the working condition stage identifier. Since steel smelting includes different stages such as melting, oxidation, and reduction, and the physicochemical properties and temperature ranges of molten steel vary greatly in each stage, a single reference curve cannot cover the entire cycle. Therefore, the system first identifies the current operating stage and then retrieves the standard oxygen potential decay curve matching that stage from the local database as the initial reference curve.

[0026] Understandably, this phased call mechanism ensures that the initial reference system for calibration is highly consistent with the current process environment, avoiding systematic deviations caused by large spans in operating conditions, and laying a correct benchmark framework for subsequent dynamic compensation.

[0027] Step S104: Input the polarization feature vector and the working condition stage identifier into the preset dynamic benchmark correction model to perform offset compensation on the initial benchmark curve and generate a dynamic compensation benchmark sequence. The dynamic reference correction model incorporates a mapping relationship between polarization effects and potential shifts. After receiving a polarization feature vector representing the current polarization level, the model calculates the potential shift caused by factors such as concentration polarization and electrochemical polarization. By subtracting this real-time calculated shift from the values ​​at each time point on the initial reference curve, a dynamic compensation reference sequence is generated. This process effectively eliminates the contamination of measured values ​​by the probe's own polarization, allowing the reference curve to be updated in real-time according to the probe's condition. This solves the reference failure problem caused by probe aging or sudden environmental changes in traditional methods.

[0028] Step S105: Calculate the deviation between the oxygen potential signal in the synchronous multi-source signal sequence and the dynamic compensation reference sequence, input the deviation and polarization feature vector into the pre-constructed piecewise mapping model, and output the target calibration parameter set; In particular, the piecewise mapping model typically constructs a two-dimensional piecewise mapping table with the deviation value as the horizontal axis and key indicators in the polarization feature vector (such as the rate of change of internal resistance) as the vertical axis.

[0029] In this embodiment, the system first determines the grid interval where the current data point is located, reads the preset gain adjustment and offset adjustment values ​​at the vertices of that interval, and then uses a bilinear interpolation algorithm to accurately calculate the actual gain adjustment and actual offset adjustment values ​​required under the current operating conditions. This table lookup and interpolation-based calculation method ensures the real-time performance of parameter output and can flexibly handle complex nonlinear error characteristics, quickly generating a calibration parameter set adapted to the current environment without complex online iterative calculations.

[0030] Step S106: Generate hardware control instructions based on the target calibration parameter set, adjust the gain coefficient and reference voltage of the oxygen potential sensor, requantize the oxygen potential signal, and generate calibrated oxygen content detection data. The system converts the actual gain and offset adjustments in the target calibration parameter set into corresponding digital control words. These control words are then written to the digital-to-analog converter control register in the signal conditioning module via a serial bus. Changes in the register value are directly converted into analog adjustment signals, driving the programmable gain amplifier to change its amplification factor and simultaneously adjusting the reference voltage of the analog-to-digital converter circuit. This hardware-level real-time adjustment capability allows the sensor front-end processing circuit to dynamically adjust the signal amplification factor and zero-point reference according to changes in operating conditions, thereby performing pre-calibration of the signal at the physical level and greatly expanding the dynamic response range of the oxygen potential sensor.

[0031] Next, after the hardware circuitry completes gain and reference adjustments, the analog-to-digital converter resamples and quantizes the conditioned analog oxygen potential signal to obtain a high-precision quantized digital value. This quantized digital value is then input into a pre-defined Nernst equation mapping table for inverse calculation, converting the potential value into the corresponding oxygen activity or oxygen content value. This process, through hardware and software collaboration, eliminates nonlinear errors and drift in the signal chain, ensuring that the final output oxygen content data accurately reflects the actual chemical composition of the liquid metal.

[0032] Step S107: The deviation value and the working condition stage identifier are aggregated into a working condition feature set and uploaded to the cloud server. The cloud server refits the segmented mapping model according to the working condition feature set, generates an updated model configuration file, and distributes it to replace the local configuration file.

[0033] The cloud server aggregates a set of operational features from a massive number of edge devices and evaluates the prediction accuracy of the current model by calculating the root mean square error of the deviation values ​​within each time window. When the accumulated error under a specific operational condition exceeds a threshold, the cloud uses a node drift evaluation model to calculate the optimal adjustment step size of the vertex coordinates of each grid interval in the two-dimensional segmented mapping table and updates the vertex coordinates by translation.

[0034] Once the updated model's prediction residuals on the validation set converge to a acceptable range, the cloud packages and sends the new vertex coordinates and parameters to the edge to replace the local configuration file, thereby enabling continuous optimization of the model throughout its entire lifecycle and constantly approaching the optimal solution.

[0035] In the above implementation, by constructing a polarization feature vector and a dynamic benchmark correction model, precise decoupling and real-time compensation for complex nonlinear errors are achieved; by using a piecewise mapping model to dynamically adjust the hardware gain and reference voltage, the accuracy and long-term stability of oxygen content detection are improved; at the same time, by combining the edge-cloud collaboration mechanism, cloud big data analysis is used to continuously evolve and self-optimize the model throughout its entire life cycle, which not only completely eliminates the risk of industrial data leakage, but also provides a high-confidence data foundation for the intelligent closed-loop control of steel smelting, and significantly reduces operation and maintenance costs.

[0036] Reference Figure 2 As one implementation of step S102, the step of extracting the rate of change of the probe internal resistance signal in the synchronous multi-source signal sequence and constructing a polarization feature vector by combining the temperature signal and the oxygen potential signal includes: Step S201: Extract the probe internal resistance signal from the synchronous multi-source signal sequence, calculate the difference between the probe internal resistance signal at the current moment and the probe internal resistance signal at the previous moment, divide the difference by the probe internal resistance signal at the previous moment to generate the rate of change of the probe internal resistance signal, and arrange the internal resistance rate components in chronological order. Among them, because the liquid metal detection probe is in a high-temperature and corrosive environment for a long time, the microstructure of its internal zirconium oxide solid electrolyte tube will change due to thermal shock and chemical corrosion. The most direct external manifestation of this change is the fluctuation of the probe's internal resistance.

[0037] Therefore, the system first locks the probe internal resistance signal channel in the synchronous multi-source signal sequence, calculates the absolute change by continuously monitoring the difference between the internal resistance value at the current moment and the previous moment, and divides the change by the reference internal resistance value at the previous moment to obtain the standardized internal resistance change rate.

[0038] This step eliminates the dimensional influence of initial impedance differences between different batches of probes, making the rate of change a dimensionless relative indicator that can objectively reflect the severity of probe polarization. By arranging this calculation result along the time axis, the system generates a sequence of internal resistance change rates, which not only records the polarization state at a single point but also forms a polarization trajectory that evolves over time. This provides crucial time-domain characteristics for subsequent determination of whether the probe has entered the nonlinear distortion region.

[0039] Step S202: Obtain the temperature signal and extract the temperature gradient component from the synchronous multi-source signal sequence, and at the same time obtain the oxygen potential signal from the synchronous multi-source signal sequence. Temperature is a key variable affecting the electromotive force of the oxygen concentration cell. Simply relying on the rate of change of internal resistance cannot completely eliminate the interference of thermodynamic factors on the measurement. Therefore, it is necessary to introduce the temperature gradient as an auxiliary feature.

[0040] In this embodiment, the system reads the temperature signal in real time from the synchronous multi-source signal sequence, calculates the temperature difference between the current moment and the previous moment, and obtains the temperature gradient value. This value characterizes the thermal shock intensity experienced by the probe. High thermal gradients are often accompanied by abrupt changes in grain boundary stress and ion mobility, which are important factors leading to measurement drift.

[0041] Meanwhile, the system directly extracts the raw oxygen potential signal, which contains information about the oxidation potential of the liquid metal under test, but is also mixed with noise generated by polarization. Incorporating the temperature gradient and the raw oxygen potential signal into the processing flow is essentially constructing a multi-dimensional observation window that includes thermodynamic parameters, electrochemical response, and structural stability, ensuring that subsequent feature vectors can comprehensively cover various physical factors affecting measurement accuracy.

[0042] Step S203: Normalize the internal resistance change rate component, temperature gradient component, and oxygen potential signal and then splice them together to generate a polarization feature vector.

[0043] Since the internal resistance change rate, temperature gradient and oxygen potential have different physical dimensions and numerical magnitudes, direct numerical calculations will cause the model to be dominated by a large number of features, thereby masking the small but crucial polarization information.

[0044] Therefore, the system uses a normalization algorithm to map these three sets of data into a unified numerical range, such as [0,1], so that parameters with different physical meanings have equal contribution weights in the feature space. Subsequently, the system concatenates the normalized internal resistance change rate sequence, temperature gradient value, and oxygen potential signal along the feature dimension to form a high-dimensional polarization feature vector.

[0045] Understandably, this polarization feature vector is no longer an isolated data point, but a comprehensive digital fingerprint that integrates probe aging status, environmental thermodynamic disturbances, and instantaneous electrochemical response. It can be directly analyzed by subsequent dynamic benchmark correction models and piecewise mapping models, thereby accurately identifying the polarization components that deviate from the true value of the current measurement, providing a solid data foundation for intelligent adaptive calibration.

[0046] In the above embodiments, by constructing a multidimensional polarization feature vector that includes the dynamic change rate of internal resistance, temperature gradient and oxygen potential, the problem that a single parameter cannot characterize complex nonlinear interference is effectively solved. This achieves deep decoupling and precise quantification of polarization effects in the liquid metal detection process, improves the sensor's anti-interference capability and measurement stability under extreme conditions, and provides high-confidence data support for precise control of oxygen content in the steel smelting process.

[0047] Reference Figure 3 As one implementation of step S104, the step of inputting the polarization feature vector and the operating condition stage identifier into a preset dynamic benchmark correction model to perform offset compensation on the initial benchmark curve and generate a dynamic compensation benchmark sequence includes: Step S301: Input the polarization feature vector and the working condition stage identifier into the preset dynamic benchmark correction model, and obtain the initial benchmark curve; Since the steelmaking process includes several stages with distinct physicochemical environments, such as the melting and refining stages, the thermodynamic behavior of molten steel differs significantly in each stage. Therefore, the dynamic benchmark correction model first needs to identify the input operating condition stage identifier. This identifier serves as an index key, triggering the model to retrieve the standard oxygen potential decay curve, i.e., the initial benchmark curve, that matches the specific stage from the locally stored mapping relationship.

[0048] Understandably, this curve represents the theoretical trajectory of the probe's output potential over time under ideal conditions under specific operating conditions. It serves as the benchmark for all subsequent compensation operations, ensuring that the calibration process is always anchored in the correct process context and avoiding systematic deviations caused by cross-stage references.

[0049] In some embodiments, the process of constructing a dynamic benchmark correction model includes: collecting standard oxygen potential decay curves under different operating conditions during historical smelting processes; fitting the mapping relationship between the benchmark curves and polarization feature vectors for each operating condition based on experimental data; and encoding the mapping relationship into a piecewise linear function and storing it in a local database.

[0050] Step S302: Obtain the corresponding basic offset coefficient from the preset offset mapping table according to the working condition stage identifier, and extract the internal resistance change rate component in the polarization feature vector. The preset offset mapping table is a lookup table built based on a large amount of historical experimental data. It records the reference drift tendency value that is common under different operating conditions, namely the basic offset coefficient, which reflects the inherent process noise level of that stage.

[0051] At the same time, the system accurately extracts the internal resistance change rate component from the input polarization feature vector. This value is a direct quantitative indicator of the influence of polarization effect on the probe at the micro level, and characterizes the signal distortion rate caused by the change in electrolyte interface impedance.

[0052] Understandably, by incorporating these two parameters into subsequent calculations, we are essentially combining macroscopic process characteristics with microscopic probe conditions, providing a dual basis for accurately calculating real-time compensation amounts.

[0053] Step S303: Input the internal resistance change rate component into the preset polarization sensitivity mapping table and output the polarization sensitivity coefficient. The polarization sensitivity mapping table is a dataset obtained through calibration experiments on the response characteristics of the probe material under different aging degrees and temperature changes. Using the internal resistance change rate as an index, the polarization sensitivity coefficient obtained by looking up the table essentially describes the potential drift amplitude corresponding to a unit change in internal resistance.

[0054] In the embodiments of this application, since the microstructure of the zirconium oxide electrolyte of probes of different batches or different service durations is different, the same internal resistance change rate may correspond to different potential offsets. Therefore, the introduction of this coefficient can realize differentiated compensation for different individual probes, and solve the problem that the general compensation algorithm cannot adapt to the discreteness of probes.

[0055] Step S304: Multiply the polarization sensitivity coefficient by the basic offset coefficient to generate a temperature-independent offset. The baseline offset coefficient represents the reference drift baseline determined by the operating environment, while the polarization sensitivity coefficient quantifies the probe's sensitivity to polarization effects. The product of these two coefficients essentially convolves environmental interference with the probe's response characteristics, resulting in a temperature-independent offset that eliminates the instantaneous fluctuations in temperature, purely reflecting the potential deviation caused by factors such as electrochemical polarization and concentration polarization. This step decouples complex multidimensional interference into a single core compensation value, simplifying subsequent calculations while ensuring the targeted nature of the compensation.

[0056] Step S305: Extract the temperature gradient component from the polarization feature vector and input it into the preset temperature compensation function to calculate the temperature compensation bias. The temperature gradient component characterizes the intensity of the thermal shock experienced by the probe, while the temperature compensation function is a mathematical model fitted based on the Nernst equation and the thermodynamic properties of the probe material. This function can calculate the impact of thermodynamic effects such as changes in electrolyte conductivity and electrode potential drift caused by sudden temperature changes on the output potential, i.e., the temperature compensation bias.

[0057] This step effectively separates thermal noise from electrochemical noise, ensuring that the calculation of the total offset not only covers polarization effects but also accurately covers interference in the thermodynamic dimension, greatly improving the robustness of measurements under high-temperature and harsh environments.

[0058] Step S306: Add the temperature-independent offset to the temperature-compensated offset to calculate the total offset; This addition operation is not a simple numerical superposition, but rather a vector synthesis of the offset caused by electrochemical polarization and the offset caused by thermodynamic temperature effects on the same dimension. The total offset, as a comprehensive error correction value, fully characterizes all deviations that the measured signal should have relative to the ideal reference signal at the current moment, under the current operating conditions, and with the current probe state. Accurate acquisition of this value is a crucial bridge connecting feature analysis and hardware calibration, providing precise quantitative basis for subsequent correction of the reference curve.

[0059] Step S307: Subtract the total offset from the values ​​at each time point on the initial reference curve to perform offset compensation operation, and generate and output the dynamic compensation reference sequence.

[0060] Traditional calibration methods often use a fixed initial reference curve, but in actual smelting, probe aging and operating condition fluctuations can cause the actual reference to drift continuously.

[0061] In this embodiment, the initial reference curve is vertically shifted in real time to ensure it always conforms to the current actual physical state. The generated dynamic compensation reference sequence is no longer a rigid theoretical value, but an adaptive reference that evolves in real time with changes in probe polarization and temperature. This sequence serves as the latest reference system for subsequent calculations of measurement deviations, fundamentally solving the problems of low accuracy and poor stability caused by probe polarization, and breaking through the accuracy bottleneck of foreign sensors due to their inability to dynamically adapt to the environment.

[0062] In the above embodiments, a dynamic reference correction mechanism based on multi-physics coupling is constructed, which effectively solves the reference drift problem caused by probe polarization, sudden temperature changes and operating condition switching in liquid metal detection. By decoupling electrochemical polarization and thermodynamic interference and correcting the reference curve in real time, the accuracy stability and operating condition adaptability of oxygen content detection throughout the entire life cycle are improved, providing a high-fidelity core data source for intelligent closed-loop control of steel smelting.

[0063] Reference Figure 4 As one implementation of step S105, the steps of calculating the deviation between the oxygen potential signal in the synchronous multi-source signal sequence and the dynamic compensation reference sequence, inputting the deviation value and polarization feature vector into a pre-constructed piecewise mapping model, and outputting the target calibration parameter set include: Step S401: Calculate the initial deviation value corresponding to each timestamp between the oxygen potential signal in the synchronous multi-source signal sequence and the dynamic compensation reference sequence; Since the dynamic compensation reference sequence has been corrected in real time according to polarization characteristics, it should theoretically be highly consistent with the true value. Therefore, the difference between the two directly reflects the residual nonlinear error and system noise in the current signal link. In this embodiment, by subtracting the actual acquired oxygen potential signal, which has been strictly timestamped, from the dynamic compensation reference sequence at each discrete timestamp, the system essentially extracts the high-frequency residual components that are not yet covered by the reference curve at the microscopic scale. These initial deviation values ​​not only include the true quantization error that needs to be eliminated by hardware circuitry, but also abnormal jumps caused by transient interferences such as unsteady turbulence of liquid metal, concentration polarization caused by sudden changes in local concentration gradients, and electrochemical polarization.

[0064] Step S402: Preprocess the initial deviation values ​​corresponding to each timestamp within the continuous time window to obtain the preprocessed deviation values ​​corresponding to each timestamp. In particular, due to the extremely harsh conditions at the smelting site, the initial deviation value is highly susceptible to contamination by sudden strong polarization interference. Directly inputting this into the model would cause severe calibration oscillations. By defining a continuous time window, the system establishes a sliding observation interval, and within this interval, independently performs variable-coefficient weighted fusion based on the degree of fluctuation in the rate of change of internal resistance for the initial deviation value at each timestamp.

[0065] Specifically, when it is determined that the polarization fluctuation is severe at the current moment, the weight of the current initial deviation value can be significantly reduced, so that its output moves closer to the historical stable trajectory to smooth out the spikes; when it is determined that the polarization is stable, the original characteristics of the current initial deviation value are retained.

[0066] Step S403: Input the preprocessed deviation value and polarization feature vector within the continuous time window into the pre-constructed piecewise mapping model, and determine the current mapping segment based on the internal resistance change rate component in the polarization feature vector; When the oxygen detector is in long-term service or encounters severe thermal shock, the ion conduction mechanism of its electrolyte interface will change fundamentally, resulting in a significant nonlinear piecewise characteristic between the detection error and the actual physical quantity. The traditional global single mapping function cannot take into account the error law under different polarization depths.

[0067] Therefore, this step inputs the pre-processed deviation value after cleaning along with the polarization feature vector containing multi-dimensional states, enabling the model to not only grasp the magnitude of the error but also perceive the physical background that generates the error. The model specifically extracts the internal resistance change rate component, which is most sensitive to the polarization response, as the core criterion, and searches and locks it in multiple nonlinear intervals pre-divided by massive calibration data, thereby accurately determining the current mapping segment where the probe is located.

[0068] For example, in the steady-state linear region where the rate of change of internal resistance is extremely low, the model enters the basic correction segment; while in the deep polarization nonlinear region where the rate of change of internal resistance jumps, the model switches to the higher-order compensation segment. This mechanism completely resolves the inherent contradiction that fixed configuration cannot adapt to complex nonlinear errors.

[0069] Understandably, this segmented processing mechanism allows the system to invoke different calibration strategies under different operating conditions and probe states, avoiding the problem of insufficient generalization ability of a single model across the entire range, and ensuring that the most suitable calibration parameters can be obtained regardless of whether the probe is in the early or late stages of aging.

[0070] In some embodiments, the process of constructing the segmented mapping model includes: collecting oxygen potential deviation samples and corresponding calibration parameters under different polarization states; dividing the samples into multiple data segments based on a clustering algorithm; training a linear regression function for each data segment to establish a mapping relationship between deviation features and calibration parameters.

[0071] Step S404: Calculate the mean and variance of the preprocessed deviation values ​​within the continuous time window, and concatenate the mean and variance to generate statistical features. The mean represents the average level of deviation within a set window period, reflecting the direction of the system error offset (positive or negative); the variance quantifies the dispersion of the deviation, revealing the fluctuation stability and noise intensity of the measurement signal.

[0072] In this embodiment, by concatenating these two seemingly simple statistics into a two-dimensional feature vector, the distribution characteristics of the error are summarized in a very concise way: the mean indicates how much adjustment is needed, and the variance indicates the urgency and risk level of the adjustment. This method of extracting statistical features has low computational overhead, making it suitable for real-time computation in embedded edge devices, while effectively filtering out instantaneous spike interference and ensuring the stability of calibration decisions.

[0073] Step S405: Obtain the preset linear regression function corresponding to the current mapping segment, input the statistical characteristic quantity into the linear regression function, calculate the gain coefficient adjustment amount and the reference voltage offset, and output them as the target calibration parameter set.

[0074] Each mapping segment has a pre-set linear regression function that has been trained and optimized offline. This function defines the optimal linear transformation relationship between statistical features (mean, variance) and hardware control parameters (gain coefficient, reference voltage) within a specific interval.

[0075] In this embodiment, the system uses the calculated mean and variance as inputs and quickly calculates the specific gain adjustment and reference voltage offset using a regression function. These two values ​​directly correspond to the adjustment amounts of the programmable gain amplifier and the analog-to-digital converter reference voltage in the signal conditioning circuit, and together they constitute the target calibration parameter set. Through this mechanism, the system can automatically generate precise hardware control commands, achieving real-time dynamic calibration of the sensor signal chain without manual intervention.

[0076] In the above implementation, by constructing a piecewise mapping model based on the rate of change of internal resistance and dynamically generating hardware adjustment parameters using the statistical characteristics of time-series deviation, the nonlinear drift problem caused by aging, polarization and environmental changes during the entire life cycle of the sensor is effectively solved, improving the accuracy and long-term stability of oxygen content detection and reducing maintenance costs and the frequency of manual calibration.

[0077] Reference Figure 5 As one implementation of step S402, the step of preprocessing the initial deviation values ​​corresponding to each timestamp within a continuous time window to obtain the preprocessed deviation values ​​corresponding to each timestamp includes: Step S501: Obtain the historical polarization feature vector corresponding to the historical time adjacent to the current time, and extract the historical internal resistance change rate component in the historical polarization feature vector. Specifically, in the environment of strong electromagnetic interference and unsteady thermal shock in liquid metal smelting, the ion conduction characteristics of the probe electrolyte interface are not static, but exhibit a continuous dynamic evolution process. The cross-sectional data at a single moment cannot accurately distinguish whether the current measurement signal is in a real physical drift or a pseudo jump caused by sudden polarization interference.

[0078] Therefore, by extracting the internal resistance change rate component of adjacent historical moments, the system can establish the historical evolution trajectory of the probe polarization state on the time axis. This historical internal resistance change rate component essentially reflects the basic polarization response level of the probe under specific operating conditions in the previous sampling period, providing an indispensable reference system for subsequent quantification of the degree of anomaly in the polarization state at the current moment.

[0079] Step S502: Extract the internal resistance change rate component from the polarization feature vector at the current moment, calculate the absolute value of the change between the internal resistance change rate component and the historical internal resistance change rate component, and generate a transient fluctuation evaluation index. The principle behind this step lies in utilizing the first-order differential characteristics of the signal to sensitively capture unsteady abrupt changes at the probe's physical interface. When concentration polarization and electrochemical polarization encounter severe thermal shock or drastic changes in operating conditions, the ion migration rate at the electrolyte interface undergoes an instantaneous step change, which is directly reflected in a dramatic change in the rate of change of internal resistance.

[0080] In this embodiment, by calculating the absolute value of the difference between the current rate of change of internal resistance and the historical rate of change of internal resistance, the system is actually calculating the acceleration of the polarization state change. The magnitude of this transient fluctuation evaluation index is directly proportional to the intensity of the polarization interference currently experienced by the probe. For example, when encountering a sudden high-frequency electromagnetic pulse or an instantaneous thermal shock caused by the violent tumbling of molten steel, the value of this index will jump significantly, thereby accurately depicting the transient drift law that cannot be tracked and reflected by traditional static reference curves at the mathematical level.

[0081] Step S503: Input the transient fluctuation assessment index into the pre-configured weight mapping relationship table for matching, and calculate the target smoothing weight coefficient; The weighted mapping table is essentially a nonlinear mapping function fitted based on massive prior smelting data. It establishes a quantitative correspondence between the intensity of polarization fluctuations and the depth of data smoothing.

[0082] In this embodiment of the application, when the transient fluctuation evaluation index is at a high level, it indicates that the current deviation value is mixed with a large amount of polarization high-frequency glitch noise. If it is directly input into the model, it will cause violent oscillation of subsequent hardware control commands. At this time, the mapping table will output a target smoothing weight coefficient that gives a very high weight to historical data and a very low weight to current data.

[0083] Conversely, when the transient fluctuation evaluation index is low, it indicates that the probe is in a stable linear drift period. The mapping table will output a weighting coefficient that gives the current data a higher weight, so as to ensure that the calibration system has sufficient real-time response sensitivity to the real physical drift. This mechanism breaks the inherent limitation of traditional sensor signal conditioning hardware circuits using fixed configurations that cannot adapt to complex nonlinear error characteristics.

[0084] Step S504: Obtain the historical initial deviation values ​​corresponding to multiple historical moments within a preset continuous time window before the current moment, and perform weighted fusion calculation on the historical initial deviation values ​​and the initial deviation value at the current moment according to the target smoothing weight coefficient to obtain the preprocessed deviation value.

[0085] Specifically, a data buffer pool is constructed by introducing historical initial deviation values ​​within a time window, and combined with the aforementioned dynamically calculated target smoothing weight coefficient, the current initial deviation value is adjusted or followed. For example, during periods of severe polarization fluctuations, the current initial deviation value is significantly weakened, and its output is more of a continuation of the previous stable deviation trajectory, thus effectively filtering out abnormal deviation spikes caused by transient fluctuations; while during stable periods, the current initial deviation value is retained with high weight.

[0086] Understandably, this processing method completely eliminates the systematic interference noise introduced by polarization transient fluctuations from the source of the data without changing the physical dimensions and basic trend of the original deviation value, so that the deviation value of the final input piecewise mapping model can purely and truly reflect the actual physical error of liquid metal oxygen content detection.

[0087] In the above implementation, by dynamically evaluating the transient fluctuation state of the internal resistance change rate and adaptively adjusting the smoothing weight, high-frequency deviation spikes caused by non-steady-state polarization interference are accurately filtered out in the data preprocessing stage. This avoids the transmission of erroneous deviation values ​​to subsequent mapping models, eliminates the accumulation of systematic deviations caused by the contradiction between nonlinear error characteristics and fixed hardware configuration, effectively solves the problem of input model data distortion caused by probe polarization effect and drastic fluctuations in operating conditions in liquid metal detection, improves the stability and accuracy of subsequent target calibration parameter set output, and provides a high signal-to-noise ratio and high reliability data foundation for closed-loop control of the smelting process.

[0088] In practical applications, this technical solution adds a dynamically adaptive data cleaning barrier between deviation calculation and model mapping, precisely blocking the transmission of abnormal spike noise generated by unsteady polarization transient fluctuations to the subsequent segmented mapping model. By effectively identifying and filtering polarization interference glitches and retaining the true physical drift trend, it avoids abnormal deviation values ​​causing extreme calibration parameters in the model output, which could lead to over-adjustment and malfunction of the back-end hardware conditioning circuit. This eliminates the hidden danger of systematic deviation accumulation caused by complex nonlinear errors in the signal link.

[0089] It should be noted that when the transient fluctuation assessment index is greater than the preset fluctuation threshold, the historical deviation value and the deviation value at the current moment can be weighted and fused according to the target smoothing weight coefficient in step S504 above; when the transient fluctuation assessment index is not greater than the preset fluctuation threshold, it is determined that the index is within the normal range, and the deviation value at the current moment can be directly used as the preprocessed deviation value.

[0090] Reference Figure 6 As one implementation of the piecewise mapping model, the steps for pre-constructing the piecewise mapping model include: Step S601: Obtain a pre-collected historical oxygen potential deviation sample set; wherein, the historical oxygen potential deviation sample set contains oxygen potential time-series deviation data under multiple different polarization states, the corresponding polarization feature vectors, and target calibration parameters. Specifically, the historical oxygen potential deviation sample set covers a large amount of data collected under various real smelting conditions. Each data point contains three core dimensions: oxygen potential time-series deviation data consisting of the difference between the actual measured value and the theoretical reference value; polarization characteristic vector reflecting the physical state of the probe at that time (containing key indicators such as internal resistance change rate and temperature gradient); and target calibration parameters (gain coefficient adjustment and reference voltage offset) obtained by senior engineers to enable accurate measurement recovery under that state.

[0091] Understandably, the purpose of this step is to accumulate sufficiently rich empirical data to ensure that the model can learn the intrinsic mapping between error characteristics and optimal hardware adjustment under various complex polarization states, thus providing data support for subsequent intelligent partition calibration.

[0092] Step S602: Extract the internal resistance change rate component from each polarization feature vector in the historical oxygen potential deviation sample set, divide the numerical distribution of the internal resistance change rate component into intervals to generate multiple preset threshold intervals, configure each threshold interval as an independent mapping segment, and establish a matching relationship between the internal resistance change rate component and the mapping segment identifier. As the probe undergoes continuous changes from a new probe to an aged probe throughout its entire life cycle, its nonlinear characteristics are extremely complex and difficult to describe with a single model.

[0093] Therefore, by discretizing the most critical health indicator—the rate of change of continuous internal resistance—into segments (e.g., dividing it into "low polarization region" for 0-5%, "medium polarization region" for 5-15%, and "high polarization region" for greater than 15%), the system decomposes the complex nonlinear global problem into several relatively linear local subproblems. This physically-based segmentation strategy makes the data characteristics within each mapping segment tend to be consistent, greatly reducing the difficulty of subsequent model fitting.

[0094] It should be noted that the internal resistance change rate components extracted in the above steps are the same characteristic component, namely the internal resistance change rate component generated in step S201. Subsequent steps directly call this component and do not need to calculate it repeatedly.

[0095] Step S603: Based on the matching relationship, the historical oxygen potential deviation sample set is allocated to the corresponding mapping segment. For the oxygen potential time series deviation data in each mapping segment, the time series statistical characteristics within the set time window are calculated, and a deviation feature sample set corresponding to each mapping segment is generated. Specifically, for each sample within a mapping segment, the system extracts deviation data within a continuous time window and calculates statistics that characterize its distribution, such as the mean (reflecting the direction of system deviation) and variance (reflecting the level of fluctuation noise). This processing method effectively filters out the interference of instantaneous random noise and retains the steady-state characteristics that reflect the true working state of the probe, thereby constructing a high-quality deviation feature sample set. This sample set transforms the original, high-dimensional time-series data into low-dimensional, high-information-density feature vectors, improving the efficiency and accuracy of model training.

[0096] Step S604: Linearly fit the deviation feature sample set corresponding to each mapping segment with the corresponding target calibration parameter to generate a linear regression function corresponding to each mapping segment, and establish a mapping relationship from time series statistical features to target calibration parameters. Specifically, within each mapping segment, a linear regression algorithm (such as least squares) can be used to fit the functional relationship between statistical characteristics (mean, variance) and calibration parameters (gain, offset). For example, the generated linear regression function is: Gain = A mean + B Variance + C is a function that defines a calibration strategy that allows any input feature to be calculated to produce accurate output parameters using a defined formula.

[0097] Among them, coefficient A is the mean influence factor, which quantifies the linear contribution of the average deviation of the system error to the gain adjustment. If the mean is large (indicating that the measured values ​​are generally higher than the reference value), A is usually negative to achieve reverse compensation. Its absolute value represents the sensitivity of the system to steady-state deviation. Coefficient B is the variance influence factor, which characterizes the correction weight of the severity of signal fluctuation (i.e., noise level) on the gain adjustment. When strong electromagnetic interference in the metallurgical field leads to large variance, B will guide the system to appropriately reduce the gain to smooth the noise and prevent control instability caused by signal jitter. The constant term C is the basic gain bias, which represents the basic gain level required for the system to maintain normal operation even without deviation within a specific polarization range, ensuring a smooth transition between different mapping segments. Through the synergistic effect of these three parameters, the model transforms abstract statistical characteristics into specific hardware control quantities.

[0098] Step S605: Associate and combine the linear regression functions and threshold intervals corresponding to each generated mapping segment to configure and generate a segmented mapping model.

[0099] The system encapsulates each independent linear function with its corresponding threshold interval (such as function F1 for the 0-5% interval, function F2 for the 5%-15% interval, etc.) to establish a hierarchical model structure.

[0100] In practical applications, the model first determines the interval based on the rate of change of the input internal resistance, and then calls the corresponding linear function for calculation. This structure is not only highly efficient and suitable for real-time operation in embedded devices, but also fully reveals the entire logical path of the model from input to output through clear interval division and function definition.

[0101] In the above implementation, the piecewise mapping model that is finally constructed can decouple the complex nonlinear calibration problem into a linear sub-interval problem based on the rate of change of internal resistance. It uses an explicit linear regression function to achieve a transparent mapping from error statistical characteristics to hardware control parameters, effectively solving the problem of nonlinear drift caused by polarization, aging and drastic changes in operating conditions throughout the sensor's life cycle, and improving the accuracy and long-term stability of oxygen content detection.

[0102] Reference Figure 7 As one implementation of step S106, the steps of generating hardware control instructions based on the target calibration parameter set, adjusting the gain coefficient and reference voltage of the oxygen potential sensor, requantizing the oxygen potential signal, and generating calibrated oxygen content detection data include: Step S701: Analyze the target calibration parameter set to obtain the corresponding gain coefficient adjustment and reference voltage offset; The target calibration parameter set is a digital result calculated by the segmented mapping model based on the statistical characteristics of timing deviation. The gain coefficient adjustment is used to correct nonlinear attenuation or amplification distortion in the signal link, while the reference voltage offset is used to compensate for zero-point drift.

[0103] Specifically, the system first parses these parameters, converting them from floating-point or integer logical values ​​into binary data format that conforms to the hardware register communication protocol. This ensures that every tiny parameter change can accurately correspond to the adjustment of the circuit's physical quantity, laying the data foundation for the subsequent direct drive of the signal conditioning circuit.

[0104] Step S702: Encode the gain coefficient adjustment amount into a first digital control word, encode the reference voltage offset amount into a second digital control word, and merge the first digital control word and the second digital control word to generate a hardware control instruction. The digital control word is not a simple numerical value, but a specific byte sequence composed of device address, register address, data bits and parity bits.

[0105] In the embodiments of this application, the first digital control word is specifically for the configuration register of the programmable gain amplifier (PGA), and its bit width and encoding method determine whether the amplification factor is increased or decreased; the second digital control word is for the control register of the digital-to-analog converter (DAC) or the reference voltage source, and is used to set the reference level of the analog-to-digital converter (ADC).

[0106] By packaging and merging these two control words according to the frame format of the Serial Peripheral Interface (SPI) or Integrated Circuit Bus (I2C) to form a complete hardware control instruction package, the integrity and anti-interference capability of the instruction during transmission are ensured, enabling the sensor hardware to accurately identify and execute the corresponding adjustment actions.

[0107] Step S703: Send hardware control commands to the hardware circuit of the oxygen potential sensor to adjust the gain coefficient and reference voltage of the hardware circuit. The hardware control instructions are transmitted to the microcontroller of the signal conditioning module via a serial bus. After parsing the instructions, the microcontroller writes configuration data to the programmable gain amplifier (PGA) and the reference voltage chip, respectively.

[0108] Specifically, adjusting the gain coefficient essentially changes the resistance ratio of the PGA's internal resistor feedback network, thereby altering the operational amplifier's closed-loop gain. This allows weak or saturated oxygen potential signals to be amplified to a full-scale range suitable for analog-to-digital converter (ADC) sampling. Adjusting the reference voltage involves outputting a new reference level through the DAC, changing the ADC's conversion reference and correcting zero-point offset during quantization. This hardware-level real-time dynamic adjustment enables the same sensor circuit to adaptively handle significant differences in output signal amplitude across different steel grades and temperatures, greatly expanding the measurement's dynamic range.

[0109] Step S704: The oxygen potential signal is amplified using the adjusted gain coefficient, and the amplified oxygen potential signal is sampled and quantized using the adjusted reference voltage to generate a quantized digital signal. The oxygen potential signal, which has been calibrated at the front end, is usually a weak voltage in the millivolt range. It first enters the PGA with the new gain parameters for in-phase amplification to improve the signal-to-noise ratio. Then, the amplified analog signal is sent to the analog-to-digital converter (ADC). At this time, the ADC no longer uses the factory default fixed reference voltage, but a reference voltage that is dynamically adjusted according to the calibration parameters.

[0110] Understandably, this floating reference sampling method effectively avoids quantization overflow or resolution degradation caused by overall signal drift. Analog-to-digital converters (ADCs) discretize continuous analog voltages into digital quantized values, which directly reflect the signal strength under the current calibration reference, providing accurate physical quantity input for subsequent digital inverse calculations.

[0111] Step S705: Perform linear inverse calculation on the quantized digital signal based on the adjusted gain coefficient and reference voltage to generate an oxygen potential calibration value; Since the gain and reference of the circuit were changed in the previous step, the raw quantization value of the analog-to-digital converter (ADC) read directly is not equal to the actual oxygen potential, so reverse mathematical operations are required.

[0112] Specifically, the system uses a linear formula that is completely inverse of the hardware adjustment to divide the analog-to-digital converter (ADC) output value by the current gain coefficient and add a numerical compensation corresponding to the reference voltage offset, thereby restoring the "scaled" and "shifted" signal to a standardized oxygen potential calibration value. This step ensures that regardless of how the hardware circuitry adaptively adjusts, the final output potential value remains consistent in both numerical value and physical meaning, achieving closed-loop calibration through a combination of hardware and software.

[0113] Step S706: Input the oxygen potential calibration value into the preset oxygen content conversion function for calculation, calculate the oxygen partial pressure value based on the concentration difference oxygen determination principle of solid electrolyte, and generate and output the calibrated oxygen content detection data according to the mapping relationship between the oxygen partial pressure value and the oxygen content.

[0114] The concentration-based oxygen determination principle of solid electrolytes is mainly based on the Nernst equation, which describes the relationship between the oxygen concentration difference and the cell electromotive force. The oxygen content conversion function incorporates the mathematical model of the Nernst equation and the characteristic constant of the probe at a specific temperature. By receiving the oxygen potential calibration value and the current temperature signal as input, it calculates the chemical potential of oxygen molecules in molten steel, i.e., the oxygen partial pressure.

[0115] Subsequently, based on thermodynamic equilibrium, the system maps the oxygen partial pressure value to commonly used oxygen content values ​​(such as ppm). Finally, after multiple rounds of filtering, compensation, calibration, and back-calculation, the interference from polarization effects, temperature drift, and circuit noise is completely eliminated, ultimately outputting high-precision oxygen content detection data.

[0116] In the above embodiments, by dynamically adjusting the programmable gain amplifier and the analog-to-digital conversion reference voltage, and combining the precise inverse calculation of the Nernst equation, the measurement inaccuracy problem caused by signal amplitude drift and nonlinear distortion in liquid metal detection is effectively solved, the accuracy, resolution and long-term stability of oxygen content detection are improved, and the hardware and software collaborative adaptive calibration of the sensor signal link is realized, providing a reliable data source for the precise control of the steel smelting process.

[0117] Reference Figure 8 As one implementation of step S107, the steps of aggregating deviation values ​​and operating condition stage identifiers into an operating condition feature set and uploading it to a cloud server, and then refitting the segmented mapping model based on the operating condition feature set to generate an updated model configuration file and distributing it to replace the local configuration file include: Step S801: Align the deviation value with the working condition stage identifier in terms of dimensions and concatenate them to generate a multidimensional array as the working condition feature set; When edge devices collect data, the deviation values ​​are usually a continuous data stream in the form of a time series, while the operating condition stage identifier is a discrete classification label representing the current smelting process.

[0118] In this embodiment of the application, in order for the cloud server to understand the specific process background of the data generation, the system must perform a dimension alignment operation to accurately match each deviation value data point with its corresponding working condition stage on the time axis, ensuring that the two have a one-to-one correspondence in the time dimension.

[0119] After alignment, the system concatenates the deviation value sequence with the operating condition stage identifier sequence to construct a multidimensional array. This multidimensional array is not merely a stack of data; it is actually a structured dataset containing both error distribution information reflecting measurement accuracy and operating condition labels defining the process environment. This provides semantically rich feature inputs for unsupervised learning and model retraining in the cloud, ensuring the targeted and effective nature of model updates.

[0120] Step S802: Encapsulate the working condition feature set to generate a data packet, upload the data packet to the cloud server, and store it in the time series database of the cloud server; In this embodiment, considering the complexity and security requirements of the network environment at the metallurgical plant, the system encapsulates the operating condition feature set using protocols before uploading. This typically employs lightweight message queue telemetry transmission protocols or hypertext transfer protocols for data packet encapsulation, and includes encryption layers and message authentication codes to prevent data tampering or leakage during transmission. After being uploaded to the cloud, the data is not directly used for computation but is first written to a time-series database.

[0121] Among them, the time series database can efficiently store and retrieve sequence data that evolves over time. By storing the set of operating conditions in the time series database, it not only achieves the persistence of massive historical data, but also provides an efficient data access interface for subsequent data mining, trend analysis and model backtracking, supporting the data foundation of the entire system.

[0122] Step S803: The cloud server obtains the root mean square error of all deviation values ​​within a preset time window from the set of operating condition features, inputs the root mean square error and the operating condition stage identifier into the preset node drift evaluation model, and outputs the adjustment step size of the mapping node parameters of the segmented mapping model. Root mean square error (RMSE) is a crucial metric for measuring model prediction accuracy, reflecting the degree of deviation of the piecewise mapping model currently deployed at the edge in practical applications. The cloud server quantifies the current health status of the model by calculating the RMSE within a specific time window.

[0123] Subsequently, this accuracy metric, along with the corresponding operating condition stage identifier, is input into the node drift evaluation model. This model is a regression or classification model trained based on historical optimization experience, capable of diagnosing that the performance degradation is due to parameter drift under specific operating conditions. The adjustment step size output by the model indicates which grid intervals in the two-dimensional piecewise mapping table of the piecewise mapping model require vertex coordinates to be moved, as well as the magnitude and direction of the movement.

[0124] In some embodiments, the steps for constructing the node drift evaluation model described above specifically include: Obtain the historical operating condition feature set, and use a clustering algorithm to divide the distribution of deviation values ​​in the historical operating condition feature set into segments. Classify the mapping nodes of the segmented mapping model into several typical operating condition clusters according to their physical meaning (e.g., "high polarization low drift region", "high polarization high drift region" and "low polarization stable region"). For each operating condition cluster, the statistical characteristics of the deviation value within a preset time window (such as root mean square error, mean and variance) are calculated, and the statistical characteristics are associated with the corresponding operating condition stage identifier to construct a training sample set; The main structure of the model, consisting of explicit mapping rules, is established based on the training sample set. When the current root mean square error and the working condition stage identifier are input, the current working condition cluster is determined according to the preset threshold range. Based on the historical data corresponding to the identified working condition clusters, a linear regression function is fitted. The direction and magnitude of the adjustment required for the mapping node parameters of the piecewise mapping model are calculated through the linear regression function to generate the adjustment step size of the mapping node parameters. For example, if it is determined that the current region is in the "high polarization and high drift region" and the root mean square error exceeds the threshold of 20%, the model will output an explicit instruction to shift the vertex coordinates of the region by 0.5 units in the negative direction based on the historical best adjustment record.

[0125] Furthermore, this node drift assessment incorporates a residual feedback mechanism. After each output adjustment step and corresponding update of the piecewise mapping model, the updated predicted residual is compared with a preset convergence threshold. If the threshold is not met, a new round of fine-tuning based on explicit mapping rules is triggered. Through this combination of cluster analysis, rule-based decision-making, and linear regression, the internal structure of the node drift assessment model is completely transparent, and its decision-making logic can be traced back to specific physical conditions and statistical laws.

[0126] Step S804: The mapping node parameters of the segmented mapping model are shifted and updated according to the adjustment step size. When the prediction residual of the updated segmented mapping model on the validation set is less than the preset convergence threshold, the updated mapping node parameters and the corresponding gain adjustment amount and offset adjustment amount are packaged to generate the updated model configuration file. In this process, the cloud uses the calculated adjustment step size to fine-tune the vertex coordinates of poorly performing grid intervals. This gradient-based fine-tuning can gradually push the model toward the optimal solution.

[0127] Furthermore, to ensure the updated model doesn't introduce new errors, the system tests the updated mapping table using a separate validation set. The update is considered valid only if the prediction residuals (i.e., the difference between the model output and the actual value) shrink to an acceptable range. Once validation is successful, the system extracts the updated mapping node parameters and their associated gain and offset parameters, packaging them into a structured model configuration file. This configuration file, containing all calibration parameters, serves as the carrier of model knowledge and is ready to be distributed to edge devices.

[0128] Step S805: Distribute the updated model configuration file to the local machine and perform digital signature verification on the updated model configuration file locally. In this process, the cloud pushes the configuration file to the edge device through a secure transmission protocol. To prevent malicious code injection or tampering with the configuration file during transmission, the edge device immediately performs digital signature verification upon receiving the file.

[0129] Specifically, a digital signature is a unique identifier generated using asymmetric encryption technology. The device uses a pre-stored public key to decrypt and verify the signature, and only configuration files with matching signatures and trusted sources are accepted. This step effectively prevents cyberattacks against smart manufacturing equipment, ensuring the security and stability of the control system.

[0130] Step S806: Load the verified updated model configuration file, overwrite the existing local configuration file, and activate the updated segmented mapping model.

[0131] Specifically, once verification is successful, the system loads the new configuration parameters into memory, replacing the old, outdated model file. Subsequently, the system resets the segmented mapping model's runtime state, directing it to the new parameter set. At this point, the calibration algorithm within the edge device has already self-corrected and optimized based on the latest operating data.

[0132] Understandably, this edge-cloud collaborative continuous learning mechanism enables the sensor system to continuously accumulate experience during use, automatically adapt to new operating condition changes and probe aging trends, without the need for frequent manual recalibration.

[0133] In the above implementation, an intelligent calibration ecosystem based on edge-cloud collaboration was constructed. Through in-depth analysis and model refitting of massive operating condition data in the cloud, combined with security verification and dynamic loading at the edge, the shortcomings of traditional sensor models being fixed and unable to adapt to long-term operating condition evolution were effectively solved. The calibration model was able to self-evolve and continuously optimize throughout its entire life cycle, thereby improving the long-term accuracy maintenance and intelligence level of the liquid metal detection system.

[0134] This application also discloses an intelligent adaptive real-time calibration system for liquid metal detection errors.

[0135] An intelligent adaptive real-time calibration system for liquid metal detection errors, specifically comprising: The multi-source signal synchronous acquisition module is used to acquire the temperature signal, oxygen potential signal and probe internal resistance signal of liquid metal through temperature sensor, oxygen potential sensor and internal resistance sensor respectively, and perform time stamp alignment processing on the signals to generate synchronous multi-source signal sequence. The polarization feature vector construction module is used to extract the rate of change of the probe internal resistance signal in the synchronous multi-source signal sequence, and to construct the polarization feature vector by combining the temperature signal and the oxygen potential signal. The working condition identification and reference curve retrieval module is used to obtain the working condition stage identifier of the current smelting process and retrieve the corresponding initial reference curve according to the working condition stage identifier. The dynamic benchmark compensation module is used to input the polarization feature vector and the working condition stage identifier into the preset dynamic benchmark correction model to perform offset compensation on the initial benchmark curve and generate a dynamic compensation benchmark sequence. The target calibration parameter generation module is used to calculate the deviation between the oxygen potential signal in the synchronous multi-source signal sequence and the dynamic compensation reference sequence. The deviation value and polarization feature vector are input into the pre-built piecewise mapping model, and the target calibration parameter set is output. The hardware control and signal quantization module is used to generate hardware control instructions based on the target calibration parameter set, adjust the gain coefficient and reference voltage of the oxygen potential sensor, requantize the oxygen potential signal, and generate calibrated oxygen content detection data. The cloud-based collaborative model update module is used to aggregate deviation values ​​and operating condition stage identifiers into an operating condition feature set and upload it to the cloud server. The cloud server refits the segmented mapping model based on the operating condition feature set, generates an updated model configuration file, and distributes it to replace the local configuration file.

[0136] The intelligent adaptive real-time calibration system for liquid metal detection error according to the embodiments 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.

[0137] 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.

[0138] This application also discloses a computer-readable storage medium.

[0139] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above in any of the intelligent adaptive real-time calibration methods for liquid metal detection errors.

[0140] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0141] 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 smart adaptive real-time calibration method for liquid metal detection error, characterized in that, The method includes: Temperature signals, oxygen potential signals, and probe internal resistance signals of liquid metal are collected by a temperature sensor, an oxygen potential sensor, and an internal resistance sensor, respectively. The signals are then time-stamped to generate a synchronous multi-source signal sequence. Extract the rate of change of the probe internal resistance signal in the synchronous multi-source signal sequence, and construct a polarization feature vector by combining the temperature signal and the oxygen potential signal; Obtain the current smelting process condition stage identifier, and retrieve the corresponding initial reference curve based on the condition stage identifier. The polarization feature vector and the working condition stage identifier are input into a preset dynamic benchmark correction model to perform offset compensation on the initial benchmark curve, generating a dynamic compensation benchmark sequence. Calculate the deviation between the oxygen potential signal in the synchronous multi-source signal sequence and the dynamic compensation reference sequence, input the deviation and the polarization feature vector into a pre-constructed piecewise mapping model, and output the target calibration parameter set; Based on the target calibration parameter set, hardware control instructions are generated to adjust the gain coefficient and reference voltage of the oxygen potential sensor, and the oxygen potential signal is requantized to generate calibrated oxygen content detection data. The deviation value and the working condition stage identifier are aggregated into a working condition feature set and uploaded to the cloud server. The cloud server refits the segmented mapping model based on the working condition feature set, generates an updated model configuration file, and distributes it to replace the local configuration file.

2. The intelligent adaptive real-time calibration method for liquid metal detection error according to claim 1, characterized in that, The steps of extracting the rate of change of the probe internal resistance signal in the synchronous multi-source signal sequence and constructing a polarization feature vector by combining the temperature signal and the oxygen potential signal include: Extract the probe internal resistance signal from the synchronous multi-source signal sequence, calculate the difference between the probe internal resistance signal at the current moment and the probe internal resistance signal at the previous moment, divide the difference by the probe internal resistance signal at the previous moment to generate the rate of change of the probe internal resistance signal, and arrange the internal resistance rate components in chronological order. The temperature signal is obtained from the synchronous multi-source signal sequence and the temperature gradient component is extracted. At the same time, the oxygen potential signal is obtained from the synchronous multi-source signal sequence. The internal resistance change rate component, the temperature gradient component, and the oxygen potential signal are normalized and then concatenated to generate the polarization feature vector.

3. The intelligent adaptive real-time calibration method for liquid metal detection error according to claim 1, characterized in that, The steps of inputting the polarization feature vector and the operating condition stage identifier into a preset dynamic benchmark correction model to perform offset compensation on the initial benchmark curve and generate a dynamic compensation benchmark sequence include: The polarization feature vector and the operating condition stage identifier are input into a preset dynamic benchmark correction model to obtain the initial benchmark curve; According to the operating condition stage identifier, the corresponding basic offset coefficient is obtained from the preset offset mapping table, and the internal resistance change rate component in the polarization feature vector is extracted. The internal resistance change rate component is input into a preset polarization sensitivity mapping table, and the polarization sensitivity coefficient is output. Multiply the polarization sensitivity coefficient by the basic offset coefficient to generate a temperature-independent offset; Extract the temperature gradient component from the polarization feature vector and input it into a preset temperature compensation function to calculate the temperature compensation bias. The total offset is calculated by adding the temperature-independent offset to the temperature-compensated offset. The total offset is subtracted from the values ​​at each time point on the initial reference curve to perform an offset compensation operation, thereby generating and outputting the dynamic compensation reference sequence.

4. The intelligent adaptive real-time calibration method for liquid metal detection error according to claim 2, characterized in that, The steps of calculating the deviation between the oxygen potential signal in the synchronous multi-source signal sequence and the dynamic compensation reference sequence, inputting the deviation and the polarization feature vector into a pre-constructed piecewise mapping model, and outputting the target calibration parameter set include: Calculate the initial deviation value between the oxygen potential signal in the synchronous multi-source signal sequence and the dynamic compensation reference sequence at each timestamp; The initial deviation values ​​corresponding to each timestamp within a continuous time window are preprocessed to obtain the preprocessed deviation values ​​corresponding to each timestamp. The preprocessed deviation value and the polarization feature vector within the continuous time window are input into the pre-constructed piecewise mapping model, and the current mapping segment is determined according to the internal resistance change rate component in the polarization feature vector. Calculate the mean and variance of the preprocessed deviation values ​​within a continuous time window, and concatenate the mean and variance to generate a statistical feature. Obtain the preset linear regression function corresponding to the current mapping segment, input the statistical feature quantity into the linear regression function, calculate the gain coefficient adjustment amount and the reference voltage offset, and output them as the target calibration parameter set.

5. The intelligent adaptive real-time calibration method for liquid metal detection error according to claim 4, characterized in that, The steps of preprocessing the initial deviation values ​​corresponding to each timestamp within a continuous time window to obtain the preprocessed deviation values ​​corresponding to each timestamp include: Obtain the historical polarization feature vector corresponding to the historical moments preceding the current moment, and extract the historical internal resistance change rate component from the historical polarization feature vector; Extract the internal resistance change rate component from the polarization feature vector at the current moment, calculate the absolute value of the change between the internal resistance change rate component and the historical internal resistance change rate component, and generate a transient fluctuation evaluation index. The transient fluctuation assessment index is input into a pre-configured weight mapping table for matching, and the target smoothing weight coefficient is calculated. Obtain the historical initial deviation values ​​corresponding to multiple historical moments within a preset continuous time window before the current moment, and perform weighted fusion calculation on the historical initial deviation values ​​and the initial deviation value at the current moment according to the target smoothing weight coefficient to obtain the preprocessed deviation value.

6. The intelligent adaptive real-time calibration method for liquid metal detection error according to claim 4, characterized in that, The steps for pre-constructing the segmented mapping model include: Obtain a pre-collected historical oxygen potential deviation sample set; wherein, the historical oxygen potential deviation sample set includes oxygen potential time-series deviation data under multiple different polarization states, corresponding polarization feature vectors, and target calibration parameters; Extract the internal resistance change rate component from each polarization feature vector in the historical oxygen potential deviation sample set, divide the numerical distribution of the internal resistance change rate component into intervals to generate multiple preset threshold intervals, configure each threshold interval as an independent mapping segment, and establish a matching relationship between the internal resistance change rate component and the mapping segment identifier. Based on the matching relationship, the historical oxygen potential deviation sample set is assigned to the corresponding mapping segment. For the oxygen potential time-series deviation data in each mapping segment, the time-series statistical features within a set time window are calculated, and a deviation feature sample set corresponding to each mapping segment is generated. The deviation feature sample set corresponding to each mapping segment is linearly fitted with the corresponding target calibration parameter to generate a linear regression function corresponding to each mapping segment, thus establishing a mapping relationship from the time-series statistical features to the target calibration parameter; The linear regression function and the corresponding threshold interval corresponding to each generated mapping segment are associated and combined to configure and generate the segmented mapping model.

7. The intelligent adaptive real-time calibration method for liquid metal detection error according to claim 4, characterized in that, The steps of generating hardware control instructions based on the target calibration parameter set, adjusting the gain coefficient and reference voltage of the oxygen potential sensor, requantizing the oxygen potential signal, and generating calibrated oxygen content detection data include: The target calibration parameter set is analyzed to obtain the corresponding gain coefficient adjustment and reference voltage offset; The gain coefficient adjustment is encoded as a first digital control word, the reference voltage offset is encoded as a second digital control word, and the first digital control word and the second digital control word are combined to generate hardware control instructions. The hardware control command is sent to the hardware circuit of the oxygen potential sensor to adjust the gain coefficient and reference voltage of the hardware circuit. The oxygen potential signal is amplified using the adjusted gain coefficient, and the amplified oxygen potential signal is sampled and quantized using the adjusted reference voltage to generate a quantized digital signal. The quantized digital signal is linearly inversely calculated based on the adjusted gain coefficient and the reference voltage to generate an oxygen potential calibration value. The oxygen potential calibration value is input into a preset oxygen content conversion function for calculation. The oxygen partial pressure value is calculated based on the concentration difference oxygen determination principle of solid electrolyte. The calibrated oxygen content detection data is generated and output according to the mapping relationship between the oxygen partial pressure value and the oxygen content.

8. The intelligent adaptive real-time calibration method for liquid metal detection error according to any one of claims 1 to 7, characterized in that, The steps of aggregating the deviation value and the working condition stage identifier into a working condition feature set and uploading it to the cloud server, and then refitting the segmented mapping model based on the working condition feature set to generate an updated model configuration file and then distributing it to replace the local configuration file include: The deviation value and the working condition stage identifier are dimensionally aligned and concatenated to generate a multi-dimensional array as the working condition feature set. The operating condition feature set is encapsulated to generate a data packet, the data packet is uploaded to the cloud server, and stored in the time series database of the cloud server; The cloud server obtains the root mean square error of all deviation values ​​within a preset time window from the set of operating conditions features, inputs the root mean square error and the operating condition stage identifier into a preset node drift evaluation model, and outputs the adjustment step size of the mapping node parameters of the segmented mapping model. The mapping node parameters of the segmented mapping model are shifted and updated according to the adjustment step size. When the prediction residual of the updated segmented mapping model on the validation set is less than the preset convergence threshold, the updated mapping node parameters and the corresponding gain adjustment and offset adjustment are packaged to generate the updated model configuration file. The updated model configuration file is distributed to the local machine, and the updated model configuration file is digitally signed and verified locally. Load the updated model configuration file that has passed verification, overwrite the existing local configuration file, and activate the updated segmented mapping model.

9. An intelligent adaptive real-time calibration system for liquid metal detection error, characterized in that, The system includes: The multi-source signal synchronous acquisition module is used to acquire the temperature signal, oxygen potential signal and probe internal resistance signal of liquid metal through temperature sensor, oxygen potential sensor and internal resistance sensor respectively, and perform time stamp alignment processing on the signals to generate a synchronous multi-source signal sequence. The polarization feature vector construction module is used to extract the rate of change of the probe internal resistance signal in the synchronous multi-source signal sequence, and to construct a polarization feature vector by combining the temperature signal and the oxygen potential signal. The working condition identification and reference curve retrieval module is used to obtain the working condition stage identifier of the current smelting process and retrieve the corresponding initial reference curve according to the working condition stage identifier. The dynamic benchmark compensation module is used to input the polarization feature vector and the working condition stage identifier into a preset dynamic benchmark correction model to perform offset compensation on the initial benchmark curve and generate a dynamic compensation benchmark sequence. The target calibration parameter generation module is used to calculate the deviation between the oxygen potential signal in the synchronous multi-source signal sequence and the dynamic compensation reference sequence, input the deviation and the polarization feature vector into a pre-constructed piecewise mapping model, and output the target calibration parameter set. The hardware control and signal quantization module is used to generate hardware control instructions based on the target calibration parameter set, adjust the gain coefficient and reference voltage of the oxygen potential sensor, requantize the oxygen potential signal, and generate calibrated oxygen content detection data. The cloud-based collaborative model update module is used to aggregate the deviation value and the working condition stage identifier into a working condition feature set and upload it to the cloud server. The cloud server refits the segmented mapping model according to the working condition feature set, generates an updated model configuration file, and distributes it to replace the local configuration file.

10. A computer-readable storage medium, characterized in that: The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 8.

Images