Visual detection method and system for liquid metal online detection sensor

Abstract

The application provides a visual detection method and system of an online liquid metal detection sensor, relates to the technical field of liquid metal detection, and the method comprises the following steps: performing feature fusion processing on first monitoring data to generate a first fusion feature vector; inputting the first fusion feature vector into a preset stage judgment model to obtain a blowing stage; performing weighted fusion on each data in second monitoring data based on the blowing stage to generate a second fusion feature vector; inputting the second fusion feature vector into a preset quality prediction model to obtain a quality state prediction result of the liquid metal; calculating a theoretical value of a liquid metal composition based on a thermodynamic and kinetic equilibrium model of converter smelting; comparing the quality state prediction result with the theoretical value, if the deviation is greater than an allowable range, correcting the quality state prediction result according to the theoretical value to obtain a detection result of the liquid metal; and directly and intuitively displaying the detection result, the first monitoring data and the second monitoring data through a visual interface.

Description

Technical Field

[0001] This application relates to the field of liquid metal detection technology, and in particular to a visual detection method and system for an online liquid metal detection sensor. Background Technology

[0002] In the field of metal smelting, the quality control of liquid metal directly determines the performance of the final product. Especially in the smelting of steel and alloys, it is necessary to monitor the chemical composition and temperature of the molten liquid metal in real time to determine the smelting endpoint. Traditional offline testing methods involving manual sampling and sample preparation are cumbersome and time-consuming, which can easily lead to substandard performance of metallurgical products and energy waste.

[0003] Although online detection devices based on laser-induced breakdown spectroscopy and other technologies have emerged, their complex optical structures make them difficult to effectively detect the plasma characteristic spectra of important non-metallic elements such as carbon, sulfur, and phosphorus in harsh environments with high temperatures and dust, where the air strongly absorbs ultraviolet and deep ultraviolet spectra. Furthermore, the probes cannot penetrate deep into the liquid metal and are easily affected by slag interference. Probes primarily designed for long-distance in-situ detection suffer from problems such as liquid metal backflow, time-consuming laser ranging and focusing adjustments, liquid level fluctuations, and slag interference due to their open bottom structure, resulting in large detection deviations and time limitations.

[0004] Therefore, there is an urgent need for a visual detection method and system for online detection sensors of liquid metals. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a visual detection method and system for an online liquid metal detection sensor, thereby improving detection efficiency and accuracy.

[0006] A first aspect of this application provides a visual detection method for an online liquid metal detection sensor, comprising:

[0007] The first monitoring data is subjected to feature fusion processing to generate a first fused feature vector; the first monitoring data includes temperature, pressure and oxygen flow rate inside the converter;

[0008] The first fused feature vector is input into a preset stage judgment model to obtain the refining stage;

[0009] The second monitoring data is obtained based on the second type of sensor group corresponding to the blowing stage control. The second monitoring data includes temperature distribution data, carbon content data, oxygen content data, and furnace gas composition data.

[0010] Based on the blowing stage, the data in the second monitoring data are weighted and fused to generate a second fused feature vector;

[0011] The second fused feature vector is input into a preset quality prediction model to obtain the quality state prediction result of the liquid metal;

[0012] The theoretical values ​​of the composition of the liquid metal are calculated based on the thermodynamic and kinetic equilibrium model of converter smelting.

[0013] The predicted quality state is compared with the theoretical value. If the deviation is greater than the allowable range, the predicted quality state is corrected according to the theoretical value to obtain the detection result of the liquid metal.

[0014] The test results, first monitoring data, and second monitoring data are displayed intuitively through a visual interface.

[0015] A second aspect of this application provides a visual detection system for an online liquid metal detection sensor, comprising:

[0016] The first fusion module is used to perform feature fusion processing on the first monitoring data to generate a first fusion feature vector; the first monitoring data includes temperature, pressure and oxygen flow rate inside the converter;

[0017] The stage judgment module is used to input the first fused feature vector into a preset stage judgment model to obtain the refining stage;

[0018] The data acquisition module is used to acquire second monitoring data based on the second type of sensor group corresponding to the blowing stage control. The second monitoring data includes temperature distribution data, carbon content data, oxygen content data, and furnace gas composition data.

[0019] The second fusion module is used to perform weighted fusion of each data in the second monitoring data based on the blowing stage to generate a second fusion feature vector.

[0020] The quality prediction module is used to input the second fused feature vector into a preset quality prediction model to obtain the quality state prediction result of the liquid metal;

[0021] The theoretical calculation module is used to calculate the theoretical values ​​of the composition of the liquid metal based on the thermodynamic and kinetic equilibrium model of converter smelting;

[0022] The detection result module is used to compare the predicted quality state with the theoretical value. If the deviation is greater than the allowable range, the predicted quality state is corrected according to the theoretical value to obtain the detection result of the liquid metal.

[0023] The results display module is used to intuitively display the detection results, first monitoring data, and second monitoring data through a visual interface.

[0024] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described visualization detection method for an online liquid metal detection sensor.

[0025] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the visualization detection method of the above-described online liquid metal detection sensor.

[0026] The beneficial effects of the visualization detection method and system for online liquid metal detection sensors provided in this application are as follows: Firstly, by collecting multi-dimensional data such as temperature distribution and carbon-oxygen content in stages, and correcting the quality prediction results based on a thermodynamic and kinetic equilibrium model, this application effectively avoids deviations caused by slag interference and spectral absorption in traditional detection methods, improving detection accuracy and making quality predictions more realistic, thus providing a reliable basis for controlling the performance of finished metal products. Secondly, it provides real-time online monitoring throughout the entire process, eliminating the need for manual sampling and sample preparation, saving the waiting time of offline detection, and enabling rapid feedback on the liquid metal state, thereby improving detection efficiency. This also helps to adjust smelting parameters in a timely manner, reducing energy waste caused by detection delays. Furthermore, by activating corresponding sensor groups based on different blowing stages, it can work stably in complex smelting environments such as high-temperature fumes, adapting to complex smelting conditions. Finally, the visualization interface intuitively presents various types of data and detection results, helping staff quickly grasp the smelting situation and reducing operational difficulty. Attached Figure Description

[0027] Figure 1 A flowchart illustrating a visual detection method for an online liquid metal detection sensor provided in an embodiment of this application;

[0028] Figure 2 This is a structural block diagram of a visual detection system for an online liquid metal detection sensor provided in an embodiment of this application;

[0029] Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0030] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0031] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1-3 The following is an explanation using specific examples.

[0032] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a visual detection method for an online liquid metal detection sensor provided in an embodiment of this application. The method includes:

[0033] S101: Perform feature fusion processing on the first monitoring data to generate a first fused feature vector; the first monitoring data includes the temperature, pressure and oxygen flow rate inside the converter.

[0034] In this embodiment, a first-class sensor group deployed at key locations in the converter collects temperature, pressure, and oxygen flow data in real time. The temperature sensor is a K-type thermocouple, the pressure sensor is a high-temperature resistant piezoelectric sensor, and the oxygen flow sensor is a vortex flow meter. The acquisition frequency is set to 1 time / second to ensure coverage of the entire smelting cycle.

[0035] The feature fusion processing in this embodiment aims to effectively fuse preprocessed temperature, pressure, and oxygen flow data to generate a first fused feature vector that comprehensively reflects the initial state of converter smelting. This embodiment uses principal component analysis and an attention mechanism to process the collected temperature, pressure, and oxygen flow data to obtain the first fused feature vector. Specifically, outliers are removed based on the 3σ principle, and the data after outlier removal is standardized to map to the [0,1] interval, resulting in preprocessed data. Principal component analysis is used to reduce the dimensionality of the preprocessed data, extracting principal component features from three sets of data with a cumulative contribution rate ≥95%. An attention mechanism is used to assign weights to different principal components, with the weights determined based on correlation analysis of historical data on the influence of each parameter on the blowing stage. Finally, the first fused feature vector is generated, achieving efficient integration and redundancy removal of data information.

[0036] The process utilizes an attention mechanism to assign weights to different principal components, with these weights determined based on correlation analysis of historical data on the impact of each parameter on the blowing stage. Specifically, an attention model is constructed, and its weight coefficients are determined by training to learn the correlation between temperature, pressure, oxygen flow rate data and the converter smelting state. For example, in the early stages of blowing, oxygen flow rate has a greater impact on the smelting process, so the attention model assigns it a higher weight; while in the middle stages of blowing, changes in temperature and pressure are more critical to reflecting the smelting state, and the weights are shifted towards these two data points. After determining the weight coefficients, the standardized temperature, pressure, and oxygen flow rate data are multiplied by their respective weight coefficients, and then summed to obtain the fused feature values. Furthermore, to further enhance the expressive power of the feature vector, temporal features of each data point are extracted and combined with the fused feature values ​​to generate a first fused feature vector. This first fused feature vector reflects the basic smelting conditions within the converter, providing comprehensive and accurate input information for subsequent blowing stage judgments.

[0037] S102: Input the first fused feature vector into the preset stage judgment model to obtain the refining stage.

[0038] In this embodiment, the preset stage judgment model is a classification model built based on machine learning algorithms. Its core function is to accurately determine the current blowing stage of the converter based on the input first fusion feature vector. The converter blowing process is usually divided into three stages: the initial blowing stage, the middle blowing stage, and the final blowing stage. The smelting reaction characteristics, temperature and pressure change patterns, and oxygen flow requirements of different stages are significantly different. Therefore, it is necessary to construct a classification model that can distinguish between these three stages.

[0039] Specifically, the model building process includes dataset preparation, model training, and model optimization.

[0040] For example, firstly, a large amount of historical data from the converter smelting process is collected, including first monitoring data of different blowing stages and corresponding actual blowing stage labels, to construct training datasets and test datasets, where the training dataset accounts for 70% of the total data and the test dataset accounts for 30%.

[0041] Then, the random forest algorithm was chosen as the base model. The random forest algorithm has advantages such as strong resistance to overfitting, insensitivity to noisy data, and the ability to handle high-dimensional features, making it suitable for classification tasks in the refining stage. During the training process of the random forest algorithm, the classification accuracy of the model was improved by adjusting hyperparameters such as the number of decision trees, the maximum tree depth, and the number of features at node splits, resulting in a well-trained random forest algorithm model.

[0042] Finally, the performance of the trained random forest algorithm model was evaluated using a test dataset. The evaluation metrics included accuracy, precision, recall, and F1 score. When the accuracy of the trained random forest algorithm model reached 95% or higher and the F1 score reached 0.94 or higher, the trained random forest algorithm model was considered to meet the practical application requirements and was used as the preset stage judgment model.

[0043] In this embodiment, after the generated first fused feature vector is input into the preset stage judgment model, the stage judgment model analyzes and classifies the feature vector according to its internal decision rules, and outputs the corresponding blowing stage result. The result output by this stage judgment model has important application value. It is not only the basis for subsequent control of the second type of sensor group to obtain the second monitoring data, but also provides guidance for real-time control of the converter smelting process. For example, if the judgment result is the initial blowing stage, the main oxidation reaction of elements such as silicon and manganese in the converter is at this time, and the oxygen flow rate needs to be controlled at a low level to avoid the reaction being too violent; if the judgment result is the middle blowing stage, carbon elements begin to oxidize in large quantities, and the oxygen flow rate needs to be appropriately increased to promote the reaction; if the judgment result is the final blowing stage, the oxygen flow rate needs to be reduced to prevent the molten steel from over-oxidizing and to ensure the quality of the molten steel.

[0044] S103: Based on the second type of sensor group corresponding to the blowing stage control, the second monitoring data is acquired. The second monitoring data includes temperature distribution data, carbon content data, oxygen content data, and furnace gas composition data.

[0045] In this embodiment, different blowing stages have different requirements for the type of sensor and operating parameters. Therefore, it is necessary to select the appropriate sensor and control the parameters according to the actual blowing stage.

[0046] The temperature distribution data was acquired using an infrared thermal imaging sensor, as this sensor can non-contactly acquire images of the temperature distribution of molten steel within the converter, thereby extracting the temperature distribution data. Specifically, in the initial stage of blowing, the temperature distribution within the converter is uneven, requiring the infrared thermal imaging sensor to have a resolution of 640×512 pixels and a sampling frequency of 0.5 times / second to clearly capture detailed changes in the temperature distribution. In the middle and final stages of blowing, the temperature distribution within the furnace is relatively uniform, allowing the resolution to be appropriately reduced to 320×240 pixels and the sampling frequency increased to 1 time / second to balance the accuracy and real-time performance of the data acquisition.

[0047] Carbon content data is collected using a fixed-carbon probe, which needs to be inserted into the molten steel for measurement. Therefore, at different stages of blowing, the insertion depth and residence time of the sensor need to be adjusted according to the depth of the molten steel and the stirring conditions. In the early stage of blowing, when the stirring of the molten steel is weak, the sensor insertion depth should be deeper and the residence time longer. In the middle and late stages of blowing, when the stirring of the molten steel is stronger, the insertion depth can be appropriately reduced and the residence time shortened to minimize the impact on the flow of the molten steel.

[0048] Oxygen content data was acquired using a zirconia oxygen sensor. This sensor utilizes the oxygen ion conductivity of zirconia solid electrolyte at high temperatures to measure the oxygen content in molten steel. The zirconia oxygen sensor needs to be inserted into the molten steel for measurement; therefore, the insertion depth and residence time need to be adjusted according to the depth of the molten steel and the agitation conditions at different blowing stages. In the early stages of blowing, when the agitation of the molten steel is weak, the sensor insertion depth should be deeper and the residence time longer. In the middle and late stages of blowing, when the agitation of the molten steel is stronger, the insertion depth can be appropriately reduced and the residence time shortened to minimize the impact on the flow of the molten steel.

[0049] Data on furnace gas composition was collected using a gas chromatograph. The gas was collected from the converter outlet, and the composition of CO, etc., was analyzed. , , The components were analyzed to obtain furnace gas composition data. The gas chromatograph sampling frequency was set to 1 time / minute. At different stages of the blowing process, the separation conditions and detection parameters of the chromatographic column were adjusted according to the changing characteristics of the furnace gas composition to ensure accurate detection of each component's content. For example, during the middle stage of blowing, the CO content in the furnace gas was high, requiring optimization of the column temperature program to improve CO separation efficiency and detection accuracy.

[0050] S104: Based on the blowing stage, the data in the second monitoring data are weighted and fused to generate a second fused feature vector.

[0051] In this embodiment, the various data points in the second monitoring data are weighted and fused based on the blowing stage to generate a second fused feature vector. The importance of temperature distribution data, carbon content data, oxygen content data, and furnace gas composition data for predicting liquid metal quality varies at different blowing stages; therefore, it is necessary to determine the weight of each data point according to the blowing stage.

[0052] Specifically, in the initial stage of blowing, attention is paid to whether the temperature distribution of the molten steel is uniform and the changes in the composition of the furnace gas. Therefore, the weights of temperature distribution data and furnace gas composition data are relatively high, set to 0.35 and 0.3 respectively; while the weights of carbon content data and oxygen content data are relatively low, set to 0.2 and 0.15 respectively.

[0053] The oxidation of carbon during the mid-stage of blowing is the main reaction, and the carbon content data has the greatest impact on the quality of molten steel, with a weight of 0.4. At the same time, the oxygen content data and furnace gas composition data are also relatively important, with weights of 0.25 and 0.2 respectively; the weight of temperature distribution data is set to 0.15.

[0054] In the final stage of blowing, it is necessary to control the composition and temperature of the molten steel. The oxygen content data and carbon content data have the highest weights, both set to 0.35; the weight of temperature distribution data is set to 0.2; and the weight of furnace gas composition data is set to 0.1.

[0055] In this embodiment, after determining the weights of each data point, the second monitoring data is preprocessed (including data cleaning and standardization, similar to the processing method for the first monitoring data). Then, each preprocessed data point is multiplied by its corresponding weight, and the results are summed to obtain the fused feature value. Furthermore, feature parameters of each data point are extracted, and these feature parameters are combined with the fused feature value to generate a second fused feature vector. These feature parameters include, for example, the highest, lowest, and average temperatures of the temperature distribution data; the rate of change of carbon content data; the fluctuation range of oxygen content data; and the proportions of each component in the furnace gas composition data. The second fused feature vector of this embodiment can comprehensively and accurately characterize the quality status information of liquid metal in the current blowing stage. This embodiment can also call the corresponding weight allocation strategy from the preset fusion strategy library based on the blowing stage. The weight allocation strategy includes the initial weights of temperature distribution data, carbon content data, oxygen content data and furnace gas composition data in the second monitoring data; the initial weights are adjusted based on the quality confidence of each type of data in the second monitoring data to obtain the target weights; the target weights are multiplied by the normalized data and then superimposed to generate the second fusion feature vector; the quality confidence is calculated by analyzing the signal strength, noise level and deviation from historical data of each type of data.

[0056] S105: Input the second fusion feature vector into the preset quality prediction model to obtain the quality state prediction result of liquid metal.

[0057] In this embodiment, the preset quality prediction model is a regression model built based on a deep learning algorithm. This model is used to predict the quality state of the liquid metal based on the input second fusion feature vector, thus obtaining a quality state prediction result. The predicted quality state of the liquid metal includes predicted values ​​for indicators such as carbon content, silicon content, manganese content, phosphorus content, sulfur content, oxygen content, and temperature.

[0058] A large amount of historical data from the converter smelting process was collected, including second monitoring data from different blowing stages, corresponding second fusion feature vectors, and actual detection values ​​of liquid metal quality obtained through laboratory chemical analysis, to construct training and testing datasets; among them, the training dataset accounted for 75% and the testing dataset accounted for 25%.

[0059] This embodiment constructs a quality prediction model based on convolutional neural networks and long short-term memory networks, including: convolutional neural network layers, long short-term memory networks, and fully connected layers; wherein, the convolutional neural network can effectively extract local features from the second fused feature vector, and the long short-term memory network can capture the temporal features of the data. The quality prediction model of this embodiment can fully mine the effective information in the feature vector and improve the accuracy of quality prediction.

[0060] During model training, mean squared error (MSE) was used as the loss function, and the model parameters were continuously adjusted through backpropagation to minimize the loss function. Dropout was employed to prevent overfitting, with a dropout rate of 0.2. After training, the model's performance was evaluated using a test dataset. Models whose mean absolute error (MAE) for each quality indicator met certain conditions were adopted as the preset quality prediction model. These conditions included, for example, MSE of carbon content ≤ 0.02% and MSE of temperature ≤ 5℃.

[0061] In this embodiment, after the second fusion feature vector is input into the quality prediction model, the model will output the predicted values ​​of various quality indicators of liquid metal. These predicted values ​​together constitute the quality status prediction result of liquid metal, which can provide workers with a preliminary basis for judging the quality of liquid metal.

[0062] S106: Calculate the theoretical values ​​of the composition of liquid metal based on the thermodynamic and kinetic equilibrium model of converter smelting.

[0063] In this embodiment, the theoretical value of the liquid metal composition is calculated based on the thermodynamic and kinetic equilibrium model of converter smelting. This theoretical value is an important reference for measuring the accuracy of the quality state prediction results.

[0064] A thermodynamic equilibrium model is constructed based on the fundamental laws of thermodynamics and the chemical reaction equations in the converter smelting process. The model considers the Gibbs free energy changes of the reactions, and by calculating the Gibbs free energy changes of each reaction, the direction and extent of the reaction are determined, thereby determining the theoretical content of each element in the liquid metal under thermodynamic equilibrium. Furthermore, the converter smelting process is a dynamic reaction process and is not entirely in thermodynamic equilibrium; therefore, a kinetic equilibrium model is also required. This kinetic equilibrium model is based on reaction kinetics theory, analyzing the reaction rate constants, reaction orders, and factors affecting the reaction rate of each smelting reaction, and establishing reaction rate equations. The reaction rate constants and reaction orders are determined by fitting experimental data, and then the theoretical carbon content in the liquid metal at the current reaction stage is calculated based on the reaction time and initial concentration. This embodiment, by combining the thermodynamic equilibrium model and the kinetic equilibrium model, can more accurately calculate the theoretical values ​​of the liquid metal composition based on the reaction equilibrium state and reaction rate, providing a reliable standard for correcting the mass state prediction results.

[0065] S107: Compare the predicted quality status with the theoretical value. If the deviation is greater than the allowable range, correct the predicted quality status based on the theoretical value to obtain the detection result of liquid metal.

[0066] In this embodiment, the predicted quality state is compared with the theoretical value calculated by the thermodynamic and kinetic equilibrium model, and the deviation between the two is calculated. For each quality indicator, an allowable deviation range is set, which is determined according to production requirements and process standards. For example, the allowable deviation range for carbon content is ±0.03%, and the allowable deviation range for temperature is ±8℃.

[0067] If the deviation of a quality indicator is less than or equal to the allowable range, the predicted quality status of that indicator is considered accurate and requires no correction. If the deviation exceeds the allowable range, the predicted quality status of that indicator needs to be corrected based on the theoretical value. The correction method can be a linear correction method. By collecting a large amount of sample data with deviations exceeding the allowable range, the least squares method is used to fit and determine the correction coefficients and correction constants to ensure that the corrected predicted value is closer to the theoretical value, thus improving the accuracy of the quality prediction. After correcting all quality indicators with deviations exceeding the limit, the final detection result of the liquid metal is obtained, which can truly and accurately represent the quality status of the liquid metal. Alternatively, the correction method in this embodiment can use the theoretical value as the prior distribution, the predicted quality status result and its uncertainty as the likelihood function, calculate the posterior distribution using Bayes' theorem, and use the expected value of the posterior distribution as the detection result.

[0068] S108: The test results, first monitoring data and second monitoring data are displayed intuitively through a visual interface.

[0069] In this embodiment, the visual interface design adopts an industrial-grade human-machine interface, and the interface layout is divided into three parts: a data monitoring area, a quality result area, and a stage display area; among which,

[0070] The data monitoring area displays the first and second monitoring data in real-time curve format. This curve can trace back to historical data of the past hour and supports data export. Among them, the temperature distribution is displayed using a heat map, while the carbon content, oxygen content, and furnace gas composition are displayed using bar charts and real-time numerical values.

[0071] The quality results area displays the test results in tabular form, and marks the theoretical value and the quality prediction result before correction. Indicators with deviations exceeding the allowable range are highlighted in red.

[0072] The stage display area shows the current refining stage in the form of icons and text, and is accompanied by audio and visual prompts when switching stages.

[0073] Display function optimization: The interface supports multi-window switching, allowing users to view single-furnace smelting data or compare historical multi-furnace data separately. It also features an anomaly alarm function. When the detection results exceed the process requirements, an alarm pop-up window will appear on the interface, triggering an on-site audible and visual alarm. The alarm threshold can be customized according to the process requirements of different steel grades.

[0074] As can be seen from the above, this application effectively avoids the deviations caused by slag interference and spectral absorption in traditional detection by collecting multi-dimensional data such as temperature distribution and carbon and oxygen content in stages, and corrects the quality prediction results based on thermodynamic and kinetic equilibrium models. This improves the accuracy of detection and makes the quality prediction more realistic, providing a reliable basis for controlling the performance of finished metal products. Secondly, the entire process is monitored online in real time, eliminating the need for manual sampling and sample preparation, saving the waiting time of offline detection, and enabling rapid feedback on the state of liquid metal, thus improving detection efficiency. It also helps to adjust smelting parameters in a timely manner and reduce energy waste caused by detection delays. In addition, by using corresponding sensor groups for different blowing stages, it can work stably in complex smelting environments such as high-temperature fumes and dust, and can adapt to complex smelting environments. Finally, the visualization interface intuitively presents multiple types of data and detection results, helping staff to quickly grasp the smelting situation and reducing the difficulty of operation.

[0075] In one embodiment of this application, a visual detection method for an online liquid metal detection sensor further includes:

[0076] The test results are compared with the target quality indicators corresponding to the blowing stage to obtain the comparison results;

[0077] If the comparison result is greater than or equal to the preset deviation threshold, then based on the comparison result and the blowing stage, the parameter adjustment strategy library is queried to obtain the corresponding process parameter adjustment strategy; the parameter adjustment strategy library is built based on the expert rule library and the historical process database.

[0078] The blowing process parameters are adjusted based on the process parameter adjustment strategy to obtain the first blowing process parameters;

[0079] If the first blowing process parameters exceed the physical boundaries of the equipment, a level one alarm will be triggered, prompting manual emergency intervention through the visual interface;

[0080] If the first refining process parameters do not exceed the physical boundaries of the equipment, then the first refining process parameters will be used as the target refining process parameters.

[0081] In this embodiment, the target quality indicators corresponding to the blowing stage are formulated based on the product standards, process requirements, and smelting objectives of different stages of converter smelting. They need to be matched with the reaction characteristics and quality control priorities of the initial, middle, and final stages of blowing to form a phased indicator system. The target quality indicators are adjustable and can be updated in real time according to changes in steel grade, fluctuations in raw material composition, or equipment status to ensure the rationality and guidance of the indicators.

[0082] The corrected liquid metal detection results are compared one by one with the target quality indicators of the corresponding stage. A dual comparison method of single indicator deviation and comprehensive deviation is used to assess the gap between the current quality status and the target.

[0083] The single-index deviation calculation involves calculating the absolute or relative deviation between the measured value and the target value for each quality index. For example, if the measured carbon content during the mid-blowing process is 0.72% and the target value is 0.6%, the absolute deviation is 0.12%; if the measured furnace temperature rise rate is 6℃ / min and the target value is 5℃ / min, the relative deviation is 20%.

[0084] To avoid a single indicator's deviation masking overall quality issues, a comprehensive deviation index needs to be constructed for comprehensive deviation assessment. Weights are assigned to each indicator based on its importance, consistent with the weighting logic of the second monitoring data weighting and fusion. The comprehensive deviation index, based on established grading standards, provides a priority basis for subsequent parameter adjustments. For example, Level 1: Comprehensive deviation index ≤ 5%, quality meets standards; Level 2: 5% < Comprehensive deviation index ≤ 15%, slight deviation; Level 3: Comprehensive deviation index > 15%, serious deviation.

[0085] In this embodiment, a parameter adjustment strategy library is constructed based on the bidirectional support of an expert rule base and a historical process database. This parameter adjustment strategy library includes a mapping relationship between deviation type, refining stage, adjustment parameter, and adjustment range.

[0086] The expert rule base, developed by metallurgical experts based on converter smelting mechanisms, reaction kinetics, and field experience, forms qualitative adjustment rules for different deviation scenarios. These adjustment rules include: composition deviation adjustment rules, temperature deviation adjustment rules, and furnace gas composition deviation adjustment rules.

[0087] For example, the composition deviation adjustment rules are as follows: When the carbon content is too high, if it is in the middle of the blowing process, the oxygen flow rate needs to be increased by 5%-10% to accelerate carbon oxidation; if it is in the middle or late stage of the blowing process, the blowing time needs to be extended by 2-3 minutes to avoid over-oxidation caused by a sudden increase in flow rate. When the oxygen content is too high, the oxygen flow rate should be reduced by 3%-5%, or ferrosilicon alloy should be added.

[0088] Temperature deviation adjustment rules: When the furnace temperature is too low, if it is in the initial stage, increase the oxygen flow rate by 4%-6% to enhance the exothermic oxidation reaction and raise the temperature; if it is in the final stage, add coke. When the furnace temperature is too high, reduce the oxygen flow rate by 2%-4%, or add scrap steel.

[0089] Furnace gas composition deviation adjustment rules: When the carbon monoxide content in the furnace gas is too low and the carbon content is too high, it indicates that the carbon oxidation reaction is insufficient. The oxygen lance position needs to be adjusted by lowering the lance position by 50-100mm to enhance the contact between oxygen and molten steel. When the carbon dioxide content in the furnace gas is too low, it indicates that the silicon-manganese oxidation is incomplete. The initial blowing time should be extended by 1-2 minutes.

[0090] By utilizing historical converter smelting data, including quality inspection results at each stage, process parameter adjustment records, and final product quality, a data mining algorithm (decision tree classification) is employed to quantitatively verify and optimize expert rules, forming a closed-loop feedback loop of deviation-adjustment-effect. For example, analysis of historical data revealed that when the carbon content deviation was 0.1% during the mid-blowing stage, increasing the oxygen flow rate by 8% resulted in a 92% success rate of reducing the carbon content to the target value within 5 minutes, significantly higher than adjustments of 5% or 10%. Therefore, the adjustment range for this scenario was optimized to 8% ± 1%.

[0091] In this embodiment, based on the three-dimensional query conditions of refining stage, comprehensive deviation level, and single index deviation type, a unique or optimal process parameter adjustment strategy is matched from the parameter adjustment strategy library. The specific process is as follows:

[0092] First, identify the current refining stage. Then, determine the deviation level based on the comprehensive deviation index. Finally, identify the main deviation indicators. Filter all strategies corresponding to the deviation level and main deviation indicators for this stage from the strategy library. If multiple matching strategies exist, sort them according to their historical execution success rate and select the strategy with the highest success rate as the process parameter adjustment strategy.

[0093] In this embodiment, the physical boundary of the equipment is the limit value of the process parameters set based on the design parameters, safety standards and service life requirements of the converter equipment. It is used to determine whether the first blowing process parameters are within the safe and operable range, so as to avoid equipment damage or safety accidents caused by parameters exceeding the limits.

[0094] The physical boundaries of the equipment include oxygen flow rate boundaries, oxygen lance position boundaries, furnace temperature boundaries, and bottom blowing flow rate boundaries. Specifically, the upper limit of the oxygen flow rate boundary is the maximum design flow rate of the oxygen lance, and the lower limit is the minimum stable flow rate. The upper limit of the oxygen lance position boundary is the maximum lifting height of the oxygen lance, and the lower limit is the minimum safe distance. The upper limit of the furnace temperature boundary is the furnace lining's withstand temperature, and the lower limit is the minimum pouring temperature of the molten steel. The upper limit of the bottom blowing flow rate boundary is the maximum withstand flow rate of the bottom blowing element, and the lower limit is the minimum stirring flow rate.

[0095] The first blowing process parameters are compared one by one with the corresponding physical boundaries of the equipment to determine whether they exceed the limits. For example, if the first oxygen flow rate parameter is 3024 m³ / h and the maximum design flow rate of the oxygen lance is 3500 m³ / h, it does not exceed the limits; however, if the first furnace temperature parameter is 1720℃ and the furnace lining withstand temperature is 1700℃, it is determined to exceed the limits.

[0096] Multi-parameter collaborative verification: Some parameters may not exceed limits individually, but their combination may exceed the overall load of the equipment, requiring collaborative verification. For example, if the oxygen flow rate is 3200 m³ / h (not exceeding the upper limit of 3500 m³ / h) and the bottom blowing flow rate is 1800 m³ / h (not exceeding the upper limit of 200 m³ / h), but both are running simultaneously, the converter gas generation will exceed the processing capacity of the gas recovery system (the maximum processing capacity corresponds to an oxygen flow rate of 3000 m³ / h and a bottom blowing flow rate of 150 m³ / h), and it will still be judged as exceeding the limits.

[0097] In this embodiment, if the first blowing process parameter exceeds the physical boundary of the equipment, a first-level alarm is immediately triggered. The alarm signal is simultaneously transmitted to the visualization interface and displayed in the form of a red flashing pop-up window, accompanied by an audible and visual alarm. The prompt content includes that the oxygen flow rate exceeds the limit and manual emergency intervention is required. At the same time, the process adjustment command is automatically suspended to prevent damage to the equipment after the parameter is executed.

[0098] If all the first blowing process parameters are within the physical boundaries of the equipment, they are determined as the target blowing process parameters and automatically sent to the executing equipment through the control system for parameter adjustment.

[0099] In addition, a new quality deviation comparison panel has been added to the visualization interface, displaying the difference between the detected value and the target value in the form of a bar chart. Red bars represent the detected value, and green bars represent the target value, with the deviation visually displayed through the difference in bar height. It also displays the single-index deviation and the comprehensive deviation index, and labels the deviation level. Below the deviation comparison module, the matching process adjustment strategy is clearly displayed in text boxes, including adjustment parameters, adjustment range, execution steps, and monitoring requirements. The changing trend of the target process parameters and the actual execution parameters is displayed as a real-time curve, with the blue line representing the target value and the red line representing the actual value. The physical boundaries of the equipment are marked; if the actual value approaches the boundary, the curve turns yellow as a warning. In this embodiment, when a level one alarm is triggered, a red pop-up window appears at the top of the interface, displaying the alarm type, out-of-limit parameters, current value, and boundary value.

[0100] As can be seen from the above, after obtaining the test results, quality deviations can be detected in a timely manner by comparing them with the target quality indicators; process adjustment schemes can be obtained based on the deviation query parameter adjustment strategy library, realizing the dynamic optimization of the blowing process; judging whether the adjustment parameters exceed the equipment boundary and triggering corresponding alarms not only ensures the feasibility of process adjustment, but also prompts manual intervention in a timely manner when abnormalities occur, thereby improving the stability and safety of the smelting process.

[0101] In one embodiment of this application, a visual detection method for an online liquid metal detection sensor further includes:

[0102] Based on the target blowing process parameters, the second type of sensor group is controlled to reacquire the second monitoring data and perform the next iteration of detection and adjustment.

[0103] If the number of consecutive iterations reaches the preset number within the preset time window, and the comparison result is greater than or equal to the preset deviation threshold, a level two alarm will be triggered, prompting engineers to conduct process verification and intervention through a visual interface.

[0104] In this embodiment, after the target process parameters are stably executed, the second type of sensor group is controlled to re-collect the second monitoring data. The acquisition process is consistent with the acquisition logic of the second monitoring data in the above embodiment, but the acquisition parameters need to be optimized according to the current blowing stage and process adjustment direction; the re-collected second monitoring data must undergo the same preprocessing process as in the above embodiment to remove outliers caused by equipment fluctuations or environmental interference to ensure data accuracy.

[0105] The preprocessed new second monitoring data is weighted and fused according to the method in the above embodiment (the weights are determined according to the current blowing stage, consistent with the logic of the first fusion) to generate a new second fusion feature vector; then it is input into the quality prediction model to obtain a new quality state prediction result, and corrected by combining the theoretical value calculated by the thermodynamic and kinetic equilibrium model to obtain the iterative liquid metal detection result.

[0106] The iterative detection results are compared with the target quality indicators of the corresponding stage to obtain new comparison results. If the new comparison results are less than the preset deviation threshold, the current process parameter adjustment is deemed effective and the liquid metal quality meets the standard. The iteration is stopped, and the current target blowing process parameters are maintained until the end of the blowing stage. If the new comparison results are still greater than or equal to the preset deviation threshold, the parameter adjustment strategy library needs to be queried again, and the adjustment strategy is optimized based on the deviation change trend of this iteration.

[0107] In this embodiment, the preset time window refers to the longest allowable time from parameter adjustment to data re-collection and quality assessment in a single iteration, which is set according to the reaction characteristics of the blowing stage and the effective time of parameter adjustment. If the quality index is detected to have met the standard within the window, the iteration can be ended early without waiting for the window to end, thus improving control efficiency.

[0108] The preset number of iterations in this embodiment refers to the maximum number of times the detection and adjustment process can be repeated within a preset time window. It needs to be set based on the severity of the quality deviation and the feasibility of process adjustment to avoid invalid iterations.

[0109] The secondary alarm in this embodiment is an important intervention mechanism after iterative control fails. An alarm is triggered when both the number of consecutive iterations within a preset time window reaches a preset number and the comparison result is still greater than or equal to a preset deviation threshold. The secondary alarm has a higher priority than the primary alarm. When both are triggered simultaneously, the visual interface prioritizes displaying the secondary alarm information, using a flashing orange pop-up window and high-frequency audio-visual warning.

[0110] As can be seen from the above, this embodiment, through an iterative detection and adjustment mechanism, can continuously track the effect of process parameter adjustments, ensuring that quality deviations are effectively controlled. When the number of iterations reaches the upper limit within the preset time window and there is still a large deviation, a secondary alarm is triggered, prompting engineers to conduct in-depth process verification, further improving the quality control capability of complex smelting processes and reducing the generation of unqualified products.

[0111] In one embodiment of this application, the second type of sensor group includes a spectral sensor, and the method further includes:

[0112] Calculate the difference between the spectral analysis data and the detection results of liquid metal;

[0113] If the difference is greater than the preset difference threshold, anomaly detection is performed on the data stream of each sensor in the second type of sensor group to identify abnormal sensors that have drifted or malfunctioned.

[0114] In this embodiment, the spectral sensor illuminates the liquid metal surface with a specific wavelength of light, receives the characteristic spectra emitted by the excited metal atoms, and determines the content of each element in the liquid metal based on the wavelength and intensity of the characteristic spectra, i.e., spectral analysis data. The sampling frequency of the spectral sensor must match the overall data acquisition rhythm of the second type of sensor group, and is set to 1 time / minute. During the acquisition process, the raw spectral data and the calculated content data of each element are recorded simultaneously.

[0115] Due to high-temperature radiation from the liquid metal surface, interference from furnace gas and dust, and sensor noise, the raw spectral data may suffer from baseline drift and overlapping characteristic peaks, necessitating preprocessing to improve data accuracy. Specifically, a polynomial fitting algorithm is used to eliminate spectral baseline drift. By selecting background regions without characteristic peaks in the spectrum, a baseline curve is fitted. The baseline intensity is then subtracted from the original spectral intensity to obtain the corrected spectral data. Wavelet transform is used to denoise the corrected spectral data, enhancing the signal intensity of characteristic peaks. Peak shape fitting is then used to determine the center wavelength and peak area of ​​each element's characteristic peak, ensuring the accuracy of element content calculations. Finally, the preprocessed element content data is processed according to the standardization method for the second monitoring data, ensuring dimensional consistency between the spectral analysis data and the detection results.

[0116] This embodiment requires difference calculation based on the elemental content indicators in the spectral analysis data and detection results, performing the calculation from two dimensions: single-element difference and comprehensive difference, to represent the consistency between the spectral data and the detection results. Single-element difference calculation: For the same elemental content indicators in the spectral analysis data and detection results, both absolute and relative differences are calculated. For example, if the spectral analysis yields a carbon content of 0.65%, and the detection result shows a carbon content of 0.60%, then the absolute difference for carbon is 0.05%, and the relative difference is 8.33%. Comprehensive difference calculation: Weights are assigned to each element based on its importance in the liquid metal quality assessment. These weights are consistent with the elemental weights in the weighted fusion of the second monitoring data. This comprehensive difference index represents the overall degree of deviation between the spectral analysis data and the detection results. In this embodiment, the calculation process requires pairing each collected spectral analysis data with the corresponding detection result to ensure the timeliness and correspondence of the difference. Simultaneously, the single-element difference and comprehensive difference index obtained from each calculation are stored in a database.

[0117] The preset difference threshold in this embodiment is a standard for judging whether the deviation between the spectral analysis data and the detection result is greater than the normal range. It needs to be determined based on the accuracy index of the spectral sensor, the error range of the detection result, and actual production needs. Considering the above factors, the initial preset difference threshold is divided into a single-element threshold and a comprehensive difference threshold.

[0118] In this embodiment, when the difference between the spectral analysis data and the detection result (single element difference or comprehensive difference) is greater than the preset difference threshold, it indicates that there is an anomaly in the spectral analysis data. Further anomaly detection is required for the data streams of each sensor in the second type of sensor group to locate the sensor that has drifted or malfunctioned.

[0119] This embodiment constructs a training dataset based on feature parameter data from historical sensors under normal and abnormal (drift, malfunction) conditions, and trains an anomaly detection model using the Isolation Forest algorithm. The normalized feature parameters of the current sensor are input into the trained Isolation Forest model, which outputs an anomaly score for each sensor. This score ranges from [0,1], with a score closer to 1 indicating a higher probability of sensor anomaly. An anomaly score threshold of 0.8 is set; if a sensor's anomaly score is ≥0.8, it is considered a confirmed anomaly; if the anomaly score is between 0.6 and 0.8, it is considered a suspicious anomaly.

[0120] This embodiment categorizes fault types into two types based on the characteristic parameters of abnormal sensors: drift and malfunction. Sensor drift refers to a slow deviation of the sensor's measured value from the true value, but without damage to the device hardware, caused by environmental factors and calibration deviations. Sensor malfunction refers to damage to the sensor hardware or functional failure, resulting in severe distortion of the measurement data. By classifying fault types, targeted handling measures can be developed, improving the efficiency of anomaly resolution.

[0121] As can be seen from the above, this embodiment uses the difference between spectral analysis data and detection results to make anomaly judgments, which can promptly detect possible drift or faulty sensors in the second type of sensor group, ensuring the reliability of sensor data; it avoids detection result deviations caused by sensor anomalies, and provides an accurate data basis for subsequent quality assessment and process adjustment.

[0122] In one embodiment of this application, a visual detection method for an online liquid metal detection sensor further includes:

[0123] If an abnormal sensor is detected, the drift compensation coefficient is calculated based on the statistical characteristics of its historical normal data and current data, and the calibration parameters of the sensor are updated.

[0124] If no abnormal sensor is identified, the prediction confidence of the quality prediction model for the current second fused feature vector is calculated.

[0125] If the prediction confidence level is lower than the preset confidence threshold, the need to optimize the quality prediction model will be triggered.

[0126] In this embodiment, when sensor drift is detected, a scientifically reasonable drift compensation coefficient is calculated based on the statistical characteristics of its historical normal data and current data. The data deviation is corrected by updating the calibration parameters, the sensor detection accuracy is restored, and the detection results are avoided due to drift.

[0127] Specifically, based on the statistical distribution differences between historical normal data and current data, a method of dimensional statistics and weighted fusion is used to calculate the drift compensation coefficient. For the sensor's detection indicators, statistical parameters, including mean and standard deviation, are calculated for both historical normal data and current data. For example, the historical mean of carbon content in the spectral sensor is 0.60%, while the current mean is 0.65%, indicating a positive drift. A linear compensation algorithm is then used to calculate the compensation coefficient. Wherein, if it is a positive drift, k is a negative value; if it is a negative drift, k is a positive value.

[0128] In this embodiment, when no anomalies are detected in the second type of sensor group, the quality prediction model is further evaluated. The prediction confidence of the current second fused feature vector is used to determine whether the model's prediction accuracy has decreased due to factors such as changes in data distribution or process updates. If the confidence is insufficient, optimization is triggered. Specifically, the confidence of the quality prediction model represents the reliability of the prediction result for the current input data (second fused feature vector), calculated from three dimensions: data matching degree, model output stability, and error distribution characteristics.

[0129] Specifically, the matching degree between the current second fusion feature vector and the model training dataset is calculated to evaluate whether the model makes predictions within the scope of known knowledge: the feature space distance calculation method is used to calculate the Euclidean distance between the current feature vector and the feature vectors of all samples in the training dataset; the average distance between the current feature vector and the 100 most recent samples in the training dataset is calculated; if the average distance is less than 1.5 times the average distance of samples within the training dataset, it indicates that the current data and the training data distribution are similar, and the data matching degree score is 0.8-1.0, with a higher score for a smaller average distance; if the average distance is more than twice the average distance, the data matching degree score is 0.2-0.5, with a lower score for a larger average distance.

[0130] The model output stability assessment determines the confidence level by judging the volatility of the prediction results of the quality prediction model. The specific method is as follows: a small perturbation is added to the current second fusion feature vector to generate 10 sets of perturbed feature vectors; the 10 sets of perturbed vectors are input into the quality prediction model to obtain 10 sets of prediction results, and the coefficient of variation of the prediction results is calculated; if the coefficient of variation is <5%, the model output stability score is 0.9-1.0; if 5% ≤ coefficient of variation < 10%, the model output stability score is 0.6-0.8; if the coefficient of variation is ≥ 10%, the model output stability score is 0.2-0.5.

[0131] Error distribution characteristic assessment is based on the statistical distribution of recent prediction errors of the model. It evaluates the error probability of the current prediction result, extracts the prediction error data of the quality prediction model within the most recent period (30 days), constructs an error distribution histogram, and calculates the mean and standard deviation of the error. Assuming that the error follows a normal distribution, it calculates the probability that the error of the current prediction result is less than the allowable error. If the probability is ≥95%, the current prediction error is likely within the allowable range, and the error distribution characteristic score is 0.9-1.0; if 80% ≤ probability < 95%, the error distribution characteristic score is 0.6-0.8; if the probability < 80%, the error distribution characteristic score is 0.2-0.5.

[0132] Based on the scores from the three dimensions, a weighted summation method is used to calculate the overall confidence level as the prediction confidence level. The weights are set according to the degree of influence of each dimension on the model's reliability. For example, the data matching degree has a weight of 0.4, the model output stability has a weight of 0.3, and the error distribution characteristics have a weight of 0.3. The prediction confidence level ranges from 0 to 1, with values ​​closer to 1 indicating more reliable model predictions. The preset confidence threshold in this embodiment is determined based on the liquid metal quality requirements and model performance indicators. When the prediction confidence level is less than the preset confidence threshold, model optimization is automatically triggered.

[0133] As can be seen from the above, for the identified abnormal sensors, dynamic calibration of the sensors is achieved by calculating the drift compensation coefficient and updating the calibration parameters, thus extending the effective service life of the sensors. When no abnormal sensors are found, the prediction confidence can be evaluated in a timely manner to identify potential problems in the quality prediction model, providing a trigger mechanism for model optimization and ensuring the accuracy of quality assessment.

[0134] In one embodiment of this application, a visual detection method for an online liquid metal detection sensor is characterized by further comprising:

[0135] In response to the need for model optimization, based on the current second fusion feature vector and spectral analysis data, an evolutionary algorithm is used to optimize the parameters of the quality prediction model to obtain the optimized model parameters.

[0136] The quality prediction model is updated based on the optimized model parameters, and subsequent quality status predictions are performed using the updated quality prediction model.

[0137] In this embodiment, the newly generated second fusion feature vector when the model optimization requirement is triggered is extracted; spectral analysis data with the same timestamp as the second fusion feature vector is selected; the actual quality index obtained from the current liquid metal through laboratory chemical analysis is used as the optimization label data, ensuring that the timestamp deviation between the label data and the input feature matrix is ​​≤1 minute. Based on historical similar detection data, it is integrated with the current preprocessed input feature matrix and label data, and divided into an optimization training set and an optimization validation set in a 7:3 ratio. The optimization training set is used for model parameter evaluation in the covariance matrix adaptive evolution strategy iteration; the optimization validation set is used for optimal parameter verification. In this embodiment, the fitness function is constructed with the goal of minimizing the prediction error of the quality prediction model. The solution space dimension is determined according to the number of parameters to be optimized in the model.

[0138] As can be seen from the above, responding to the model optimization requirements and using evolutionary algorithms to optimize the parameters of the quality prediction model can automatically optimize the model parameters, improve the model's adaptability to the current smelting conditions and the prediction accuracy; updating the model based on the optimized parameters and using it for subsequent evaluation realizes the continuous evolution of the model and improves the reliability of long-term detection.

[0139] In one embodiment of this application, the evolutionary algorithm is an adaptive evolutionary strategy based on the covariance matrix;

[0140] Before using evolutionary algorithms to optimize the parameters of the quality prediction model, the following steps are also included:

[0141] Determine the solution space of the adaptive evolution strategy for the covariance matrix based on the structure and parameter range of the quality prediction model;

[0142] The hyperparameters of the adaptive evolution strategy for the covariance matrix are determined based on the structural complexity, number of parameters, and empirical values ​​of the quality prediction model.

[0143] Hyperparameters include population size, initial step size, and covariance matrix;

[0144] Among them, the population size is positively correlated with the number of parameters, while the initial step size is negatively correlated with the structural complexity.

[0145] In this embodiment, the structure of the quality prediction model includes: Convolutional cloud network layer parameters: number of convolutional kernels, 2D; convolutional kernel size, 2D; pooling window size, 2D; convolutional layer activation function weights, 4D; convolutional dropout rate, 2D; Long Short-Term Memory network layer parameters: number of hidden layer nodes, 2D; time step, 1D; LSTM layer dropout rate, 2D; forget gate / input gate weight initialization range, 3D; Fully connected layer parameters: number of neurons, 2D; L2 regularization coefficient, 2D; fully connected layer activation function bias, 2D; Training parameters: learning rate, 1D; batch size, 1D; number of iterations, 1D.

[0146] For discrete parameters, they are mapped to discrete integer indices, and the range of values ​​for that dimension in the search vector is set to [1,2]. Subsequently, the index is used to reverse map to the actual parameter. For continuous parameters, the range of values ​​for the corresponding dimension in the search vector is directly set as the actual range of values ​​for the parameter, and a real number encoding method is used. For parameters with constraints, constraints are added to the solution space, and search vectors that do not meet the constraints are filtered out by a feasibility judgment function.

[0147] Constraints based on the physical meaning of the parameters: the boundaries of probabilistic parameters are set to [0,1] to avoid negative or invalid values ​​greater than 1; the boundaries of counting parameters (number of convolutional kernels, number of hidden layer nodes) are set to positive integers and must meet the minimum functional requirements of the model structure, such as ≥8 convolutional kernels to ensure feature extraction capability; ≥32 hidden layer nodes to avoid underfitting; the boundaries of training parameters need to be adapted to hardware computing power, such as when the GPU memory is 16GB, the maximum batch size is set to 64 and the maximum number of iterations is set to 200.

[0148] Constraints based on model performance: The lower bound of the learning rate is set to... To avoid excessively low learning rates leading to slow training convergence, the upper limit of the learning rate is set to... To avoid excessively high learning rates leading to parameter oscillations and non-convergence, the lower bound of the L2 regularization coefficient is set to... To avoid overfitting due to weak regularization; the upper limit of the L2 regularization coefficient is set to... To avoid overly strong regularization leading to underfitting, the convolution kernel size is limited to 3×3 or 5×5. Through these constraints, the search range of the covariance matrix adaptive evolution strategy algorithm is strictly confined to the effective parameter space, significantly improving optimization efficiency and the effectiveness of parameter combinations.

[0149] In this embodiment, the population size is a key hyperparameter affecting search diversity and convergence speed in the covariance matrix adaptive evolution strategy. It needs to be positively correlated with the dimension of the solution space to ensure that the population can cover the key regions of the solution space.

[0150] The initial step size controls the covariance matrix in the adaptive evolution strategy. The step size of each search iteration is negatively correlated with the structural complexity of the quality prediction model. The more complex the structure, the smaller the initial step size, avoiding instability caused by an excessively large step size; conversely, the simpler the structure, the larger the initial step size, accelerating convergence. The structural complexity is quantified using the product of the number of layers × the average number of parameters per layer.

[0151] The covariance matrix describes the correlation and search direction of parameters in each dimension of the solution space. It needs to be initialized based on the correlation characteristics and value range of the model parameters in the quality prediction model to ensure efficient searching along key directions in the initial stages of the algorithm. If there is no significant correlation between parameters, the covariance of the corresponding dimension of the covariance matrix is ​​set to 0; if there is a strong correlation between parameters, the covariance is set to a positive value; if there is a negative correlation between parameters, the covariance is set to a negative value. The diagonal elements of the covariance matrix are positively correlated with the value range of the parameter; the larger the value range, the larger the variance.

[0152] As can be seen from the above, this embodiment adopts an adaptive evolutionary strategy based on the covariance matrix for parameter optimization, which improves the efficiency and accuracy of optimization. By determining the solution space based on the model structure and parameter range, and determining the hyperparameters based on structural complexity, number of parameters and empirical values, the use of evolutionary algorithms is more in line with the optimization needs of the quality prediction model, ensuring the pertinence and effectiveness of the optimization process.

[0153] In one embodiment of this application, a visual detection method for an online liquid metal detection sensor further includes:

[0154] Obtain population diversity and convergence rate indices for the covariance matrix adaptive evolution strategy;

[0155] If the population diversity index is less than the first threshold and the convergence speed index is less than the second threshold, then the step size of the adaptive evolution strategy based on the first step size is increased and the preset proportion of individuals with the worst fitness in the population is reinitialized.

[0156] If the population diversity index is greater than the third threshold and the convergence speed index is greater than the fourth threshold, then the step size of the adaptive evolution strategy based on the second step size is reduced and the elite retention ratio is increased.

[0157] In this embodiment, population diversity reflects the degree of difference among individuals within the population, directly affecting the algorithm's global search capability; convergence speed reflects the efficiency of the algorithm in approaching the optimal solution, determining the optimization cycle. Population diversity and convergence speed indicators need to be quantitatively evaluated through scientific indicator definitions and calculation methods.

[0158] This embodiment uses a two-dimensional index—spatial distribution diversity and fitness distribution diversity—to assess population diversity, avoiding misjudgments caused by a single dimension. Among these,

[0159] The spatial distribution diversity index is calculated based on the distribution density of individuals in the solution space, characterizing the degree of dispersion of individuals in the parameter space. Its value ranges from [0, +∞). A larger value indicates a more dispersed distribution and higher diversity in the solution space; a smaller value indicates a more concentrated distribution and lower diversity. For example, when individuals in the population are concentrated in a local area of ​​the solution space, the spatial distribution diversity index will be less than 0.2; when individuals are evenly distributed in the solution space, the spatial distribution diversity index will be greater than 0.8.

[0160] The fitness distribution diversity index is calculated based on the distribution of fitness values ​​of individuals in a population, characterizing the degree of difference between individuals in the performance space. The spatial distribution diversity index ranges from [0, +∞). A larger value indicates greater differences in fitness among individuals within the population, suggesting multiple combinations of performance levels and higher diversity. Conversely, a smaller value indicates smaller fitness differences, with individuals showing similar performance and lower diversity. For example, when the population fitness values ​​are concentrated between 0.8 and 0.85 (mean 0.82, standard deviation 0.015), the spatial distribution diversity index is approximately 0.018, indicating low diversity. When the fitness values ​​are distributed between 0.6 and 0.9 (mean 0.75, standard deviation 0.08), the spatial distribution diversity index is approximately 0.107, indicating high diversity.

[0161] In this embodiment, the spatial distribution diversity index and the fitness distribution diversity index are weighted and summed to obtain the population diversity index. The weights of spatial distribution diversity and fitness distribution diversity are set according to the degree of influence of the index on optimization. For example, the weight of spatial distribution diversity is 0.6, and the weight of fitness distribution diversity is 0.4.

[0162] In this embodiment, the convergence speed index is calculated based on the rate of change of the optimal fitness value at different iteration stages, reflecting the efficiency of the algorithm in approaching the optimal solution. The short-term convergence speed and the long-term convergence speed are used for evaluation.

[0163] The short-term convergence rate is calculated as the average rate of change of the optimal fitness value over the last five generations, characterizing the recent convergence trend of the algorithm; the range of this short-term convergence rate is... In this context, a positive value indicates an increase in fitness (convergence), while a negative value indicates a decrease in fitness (divergence). The larger the absolute value, the faster the short-term convergence or divergence.

[0164] The long-term convergence rate is calculated as the total rate of change of the optimal fitness value from the start of the iteration to the current generation, characterizing the overall convergence efficiency of the algorithm; the range of this long-term convergence rate is [value missing]. Positive values ​​indicate overall convergence, while negative values ​​indicate overall divergence. The larger the absolute value, the higher the long-term convergence or divergence efficiency.

[0165] This embodiment calculates a convergence speed index by weighted summation of short-term and long-term convergence speeds. Short-term convergence speed, being a better indicator of the current algorithm state, is given a higher weight; for example, the short-term convergence speed weight is 0.7, and the long-term convergence speed weight is 0.3. The convergence speed index ranges from [0, +∞), with larger values ​​indicating faster overall convergence and smaller values ​​indicating slower convergence.

[0166] In this embodiment, the first threshold, the second threshold, the third threshold, and the fourth threshold are set based on the characteristics of the CMA-ES algorithm, the optimization requirements of the quality prediction model parameters, and historical optimization data.

[0167] Specifically, the first threshold is used to determine if the population diversity is too low. When the population diversity index is less than the first threshold, it indicates that the population is overly concentrated and prone to getting trapped in local optima. The first threshold is based on historical optimization data and is the minimum diversity value when the algorithm is normally optimizing. Based on the influence of the solution space dimension, the higher the dimension, the higher the diversity is needed to maintain the global search capability. Therefore, the first threshold can be set to 0.2.

[0168] The third threshold is used to determine if the population diversity is too high. When the population diversity index exceeds the third threshold, it indicates that the population is overly dispersed and has low convergence efficiency. The third threshold is the highest diversity value during normal optimization by the statistical algorithm, and it is based on the population size. The larger the population size, the higher the diversity can be allowed. The third threshold can be set to 0.7.

[0169] The second threshold is used to determine if the convergence speed is too slow. When the convergence speed index is less than the second threshold, it indicates that the algorithm has low convergence efficiency and needs to be accelerated to approach the optimal solution. The second threshold is based on the optimization objective of the quality prediction model, which is used to deduce the minimum convergence speed. Combined with the iteration stage, in the early stage, rapid exploration is required, and the second threshold can be set to 0.03; in the middle and later stages, fine search is required, and the second threshold can be set to 0.015.

[0170] The fourth threshold is used to determine if the convergence speed is too fast. When the convergence speed index exceeds the fourth threshold, it indicates that the algorithm converges too quickly and is prone to missing the optimal solution. The fourth threshold is based on the highest convergence speed of the statistical algorithm during normal convergence. When the fitness is close to the theoretical maximum value, the convergence speed should be slowed down, and the fourth threshold can be set to 0.06; when the fitness is low, it can be set to 0.09.

[0171] In this embodiment, based on the combination of population diversity index and convergence speed index, two scenarios are implemented for regulation. By adjusting the step size and population state, the algorithm's optimization efficiency is improved.

[0172] Scenario 1: Low population diversity and slow convergence speed: The population diversity index is less than the first threshold and the convergence speed index is less than the second threshold. Scenario 1 indicates that the population is overly concentrated in a local area and is difficult to converge to a better solution. It is necessary to increase the step size and reconstruct part of the population to enhance the global search capability.

[0173] The adaptive evolution strategy increases the step size of the covariance matrix based on the first step size and reinitializes the population with the worst fitness of individuals by a preset proportion. The first step size is 20%-50% of the current step size, and the specific value is positively correlated with the deviation of the current diversity and convergence speed. At the same time, the updated step size is less than twice the initial step size to avoid the search getting out of control due to an excessively large step size.

[0174] The preset ratio is set to 20%-40% of the population size. The ratio is positively correlated with the degree of diversity deviation. The closer the degree of diversity deviation is to 0, the higher the ratio. Individuals are sorted from smallest to largest according to their fitness values, and the bottom 30% of individuals are selected as individuals to be reconstructed. New individuals are generated based on the solution space boundary and the current covariance matrix. The new individuals follow a multivariate normal distribution.

[0175] Scenario 2: High population diversity and fast convergence speed: If the population diversity index is greater than the third threshold and the convergence speed index is greater than the fourth threshold, then the step size of the adaptive evolution strategy based on the second step size is reduced and the elite retention ratio is increased. Scenario 2 indicates that the population is too dispersed, the convergence efficiency is too high, and it is easy to miss local optima. It is necessary to reduce the step size and increase the elite retention ratio to enhance the fine search capability.

[0176] The adaptive evolution strategy is based on reducing the covariance matrix by a second step size and increasing the elite retention ratio. The second step size is set to 10%-30% of the current step size, and the specific value is positively correlated with the degree of excess of the current diversity and convergence speed. The elite retention ratio is adjusted. If the original elite retention ratio is 25%, it is adjusted to 35%, that is, the top 35% of fitness individuals are selected as parents to improve the probability of high-quality parameters being inherited.

[0177] As can be seen from the above, by monitoring population diversity and convergence speed indicators, and adjusting the step size and population of the covariance matrix adaptive evolution strategy, the search ability can be enhanced by increasing the step size and reinitializing some individuals when the algorithm gets stuck in a local optimum. When the convergence is too fast, the optimization accuracy can be improved by decreasing the step size and increasing the elite retention ratio. This balances the algorithm's exploration and utilization capabilities and improves the effect of parameter optimization.

[0178] In one embodiment of this application, a visual detection method for an online liquid metal detection sensor further includes:

[0179] Based on the detection results and corresponding target quality indicators in the historical optimization process, the accuracy score of the quality prediction model is calculated.

[0180] If the accuracy score is greater than the first score threshold, then the initial value of the step size for the adaptive evolution strategy based on the third step size to reduce the covariance matrix is ​​determined.

[0181] If the accuracy score is less than the second score threshold, then the initial step size of the adaptive evolution strategy based on the fourth step size increases the covariance matrix.

[0182] In this embodiment, the accuracy score is based on the statistical deviation between the model's detection results and the corresponding target quality index during the historical optimization process, representing the accuracy of the quality prediction model in predicting the quality of liquid metal.

[0183] In this embodiment, the weighted average method is used to calculate the accuracy score of the quality prediction model. Based on the prediction deviation rate of each quality indicator, the accuracy score is calculated by weighting the indicators according to their importance. The accuracy score ranges from [0,1]. The closer it is to 1, the higher the historical prediction accuracy of the quality prediction model.

[0184] The calculation formula is:

[0185] in, For the accuracy score, i represents the i-th quality indicator, n represents the number of quality indicators; wi represents the weight of the i-th quality indicator, which is set based on industrial process requirements; j represents the j-th furnace, and m represents the number of furnaces in the historical data. Let i be the target quality value of the i-th quality indicator for the j-th furnace. This is the predicted value of the i-th quality index for the j-th furnace after optimization.

[0186] In this embodiment, if the accuracy score is greater than the first score threshold, it indicates that the historical prediction accuracy of the quality prediction model is high, and the current parameter region is close to the optimum. No large step size is needed for exploration. The initial value of the basic step size is reduced based on the third step size to reduce parameter oscillations during the optimization process. If the accuracy score is less than the second score threshold, it indicates that the historical prediction accuracy of the quality prediction model is insufficient, and the model will be far from the optimal solution in the current parameter region. Therefore, the initial exploration range needs to be expanded. The initial step size of the covariance matrix adaptive evolution strategy is increased based on the fourth step size to improve the exploration efficiency of high-quality parameter regions. If the accuracy score is less than or equal to the first score threshold and greater than or equal to the second threshold, it indicates that the historical accuracy of the quality prediction model is moderate, and the initial value of the basic step size is maintained. Here, the first score threshold represents excellent historical model accuracy, requiring no large-scale parameter exploration; the second score threshold represents insufficient historical model accuracy, requiring an expanded initial exploration range; the third step size, based on experimental results, is used to reduce the initial step size when accuracy is excellent; and the fourth step size, based on experimental results, is used to increase the initial step size when accuracy is insufficient.

[0187] As can be seen from the above, this embodiment dynamically adjusts the initial step size of the evolutionary strategy based on the historical accuracy score of the quality prediction model. When the accuracy of the quality prediction model is high, the step size is reduced to perform fine optimization, and when the accuracy is low, the step size is increased to expand the search range. This allows the evolutionary algorithm to adaptively adjust according to the actual performance of the quality prediction model, further improving the efficiency and pertinence of model parameter optimization and ensuring the continuous stability of model performance.

[0188] Corresponding to the visualization detection method of the online liquid metal detection sensor in the above embodiment, Figure 2 This is a structural block diagram of a visual detection system for an online liquid metal detection sensor provided in one embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 2 The visualization detection system 20 of the online liquid metal detection sensor includes: a first fusion module 21, a data acquisition module 23, a stage judgment module 22, a second fusion module 24, a quality prediction module 25, a theoretical calculation module 26, a detection result module 27, and a result display module 28.

[0189] The first fusion module 21 is used to perform feature fusion processing on the first monitoring data to generate a first fusion feature vector; the first monitoring data includes the temperature, pressure and oxygen flow rate inside the converter;

[0190] The stage judgment module 22 is used to input the first fused feature vector into the preset stage judgment model to obtain the refining stage;

[0191] Data acquisition module 23 is used to acquire second monitoring data based on the second type of sensor group corresponding to the blowing stage control. The second monitoring data includes temperature distribution data, carbon content data, oxygen content data and furnace gas composition data.

[0192] The second fusion module 24 is used to perform weighted fusion of each data in the second monitoring data based on the blowing stage to generate a second fusion feature vector.

[0193] The quality prediction module 25 is used to input the second fused feature vector into the preset quality prediction model to obtain the quality state prediction result of the liquid metal.

[0194] Theoretical calculation module 26 is used to calculate the theoretical values ​​of the composition of liquid metal based on the thermodynamic and kinetic equilibrium model of converter smelting.

[0195] The detection result module 27 is used to compare the predicted quality status with the theoretical value. If the deviation is greater than the allowable range, the predicted quality status is corrected according to the theoretical value to obtain the detection result of liquid metal.

[0196] The results display module 28 is used to intuitively display the test results, first monitoring data and second monitoring data through a visual interface.

[0197] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 2 The functions of the first fusion module 21, data acquisition module 23, stage judgment module 22, second fusion module 24, quality prediction module 25, theoretical calculation module 26, detection result module 27, and result display module 28 are shown.

[0198] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0199] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0200] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0201] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the visualization detection method of the online liquid metal detection sensor provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0202] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0203] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0204] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0205] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0206] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0207] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0208] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0209] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A visual detection method for an online liquid metal detection sensor, characterized in that, include: The first monitoring data is subjected to feature fusion processing to generate a first fused feature vector; The first monitoring data includes temperature, pressure, and oxygen flow rate inside the converter; The first fused feature vector is input into a preset stage judgment model to obtain the refining stage; The second monitoring data is acquired based on the second type of sensor group corresponding to the blowing stage control. The second monitoring data includes temperature distribution data, carbon content data, oxygen content data, and furnace gas composition data. Among them, the temperature distribution data is acquired using an infrared thermal imaging sensor. In the early stage of blowing, the temperature distribution in the converter is uneven. The resolution of the infrared thermal imaging sensor is set to 640×512 pixels, and the sampling frequency is 0.5 times / second to clearly capture the detailed changes in temperature distribution. In the middle and late stages of blowing, the temperature distribution in the furnace is relatively uniform. The resolution is reduced to 320×240 pixels, and the sampling frequency is increased to 1 time / second to balance the accuracy and real-time performance of data acquisition. Carbon content data is collected using a fixed carbon probe, which is inserted into the molten steel for measurement. Oxygen content data is collected using a zirconia oxygen sensor, which is inserted into the molten steel for measurement. At different blowing stages, the insertion depth and residence time of the sensor are adjusted according to the depth of the molten steel and the stirring conditions. The data on furnace gas composition were collected using a gas chromatograph. At different blowing stages, the separation conditions and detection parameters of the chromatographic column were adjusted according to the changing characteristics of the furnace gas composition to ensure accurate detection of the content of each component. Based on the blowing stage, the data in the second monitoring data are weighted and fused to generate a second fused feature vector; The second fused feature vector is input into a preset quality prediction model to obtain the quality state prediction result of the liquid metal; The theoretical values ​​of the composition of the liquid metal are calculated based on the thermodynamic and kinetic equilibrium model of converter smelting. The predicted quality state is compared with the theoretical value. If the deviation is greater than the allowable range, the predicted quality state is corrected according to the theoretical value to obtain the detection result of the liquid metal. The test results, first monitoring data, and second monitoring data are displayed intuitively through a visual interface.

2. The visual detection method for an online liquid metal detection sensor according to claim 1, characterized in that, Also includes: The test results are compared with the target quality indicators corresponding to the blowing stage to obtain the comparison results; If the comparison result is greater than or equal to the preset deviation threshold, then based on the comparison result and the blowing stage, the parameter adjustment strategy library is queried to obtain the corresponding process parameter adjustment strategy. The parameter adjustment strategy library is built based on an expert rule base and a historical process database. The blowing process parameters are adjusted based on the aforementioned process parameter adjustment strategy to obtain the first blowing process parameters; If the first blowing process parameters exceed the physical boundaries of the equipment, a level one alarm will be triggered, prompting manual emergency intervention through a visual interface. If the first blowing process parameters do not exceed the physical boundaries of the equipment, then the parameters based on the first blowing process parameters will be used as the target blowing process parameters.

3. The visual detection method for an online liquid metal detection sensor according to claim 2, characterized in that, Also includes: Based on the target blowing process parameters, the second type of sensor group is controlled to reacquire the second monitoring data and perform the next iteration of detection and adjustment. If the comparison result is greater than or equal to the preset deviation threshold after a preset number of consecutive iterations within a preset time window, a level two alarm is triggered, prompting engineers to conduct process verification and intervention through a visual interface.

4. The visualization detection method for an online liquid metal detection sensor according to claim 1, characterized in that, The second type of sensor group includes a spectral sensor, and the method further includes: Calculate the difference between the spectral analysis data of the liquid metal and the detection result; If the difference is greater than a preset difference threshold, then anomaly detection is performed on the data stream of each sensor in the second type of sensor group to identify abnormal sensors that have drifted or malfunctioned.

5. The visualization detection method for an online liquid metal detection sensor according to claim 4, characterized in that, Also includes: If the abnormal sensor is detected, the drift compensation coefficient is calculated based on the statistical characteristics of its historical normal data and current data, and the calibration parameters of the sensor are updated. If no abnormal sensor is identified, the prediction confidence of the quality prediction model for the current second fused feature vector is calculated. If the prediction confidence level is less than the preset confidence threshold, the quality prediction model optimization requirement is triggered.

6. The visualization detection method for an online liquid metal detection sensor according to claim 5, characterized in that, Also includes: In response to the model optimization requirement, based on the current second fusion feature vector and spectral analysis data, an evolutionary algorithm is used to optimize the parameters of the quality prediction model to obtain the optimized model parameters. The quality prediction model is updated based on the optimized model parameters, and subsequent quality status predictions are performed using the updated quality prediction model.

7. The visual detection method for an online liquid metal detection sensor according to claim 6, characterized in that, The evolutionary algorithm is a covariance matrix adaptive evolutionary strategy; Before optimizing the parameters of the quality prediction model using an evolutionary algorithm, the method further includes: The solution space of the adaptive evolution strategy for the covariance matrix is ​​determined based on the structure and parameter range of the quality prediction model. The hyperparameters of the adaptive evolution strategy for the covariance matrix are determined based on the structural complexity, number of parameters, and empirical values ​​of the quality prediction model. The hyperparameters include population size, initial step size, and covariance matrix; The population size is positively correlated with the number of parameters, and the initial step size is negatively correlated with the structural complexity.

8. The visual detection method for an online liquid metal detection sensor according to claim 7, characterized in that, Also includes: Obtain the population diversity index and convergence speed index of the adaptive evolution strategy of the covariance matrix; If the population diversity index is less than the first threshold and the convergence speed index is less than the second threshold, then the step size of the covariance matrix adaptive evolution strategy is increased based on the first step size, and the preset proportion of individuals with the worst fitness in the population is reinitialized. If the population diversity index is greater than the third threshold and the convergence speed index is greater than the fourth threshold, then the step size of the adaptive evolution strategy based on the covariance matrix is ​​reduced and the elite retention ratio is increased.

9. The visual detection method for an online liquid metal detection sensor according to claim 7, characterized in that, Also includes: Based on the detection results and corresponding target quality indicators in the historical optimization process, the accuracy score of the quality prediction model is calculated. If the accuracy score is greater than the first score threshold, then the initial value of the step size of the covariance matrix adaptive evolution strategy is reduced based on the third step size. If the accuracy score is less than the second score threshold, the initial value of the step size of the covariance matrix adaptive evolution strategy is increased based on the fourth step size.

10. A visual detection system for an online liquid metal detection sensor, characterized in that, include: The first fusion module is used to perform feature fusion processing on the first monitoring data to generate a first fusion feature vector; The first monitoring data includes temperature, pressure, and oxygen flow rate inside the converter; The stage judgment module is used to input the first fused feature vector into a preset stage judgment model to obtain the refining stage; The data acquisition module is used to acquire second monitoring data based on the second type of sensor group corresponding to the blowing stage control. The second monitoring data includes temperature distribution data, carbon content data, oxygen content data, and furnace gas composition data. Among them, the temperature distribution data is acquired using an infrared thermal imaging sensor. In the early stage of blowing, the temperature distribution in the converter is uneven. The resolution of the infrared thermal imaging sensor is set to 640×512 pixels, and the sampling frequency is 0.5 times / second to clearly capture the detailed changes in temperature distribution. In the middle and late stages of blowing, the temperature distribution in the furnace is relatively uniform. The resolution is reduced to 320×240 pixels, and the sampling frequency is increased to 1 time / second to balance the accuracy and real-time performance of data acquisition. Carbon content data is collected using a fixed carbon probe, which is inserted into the molten steel for measurement. Oxygen content data is collected using a zirconia oxygen sensor, which is inserted into the molten steel for measurement. At different blowing stages, the insertion depth and residence time of the sensor are adjusted according to the depth of the molten steel and the stirring conditions. The data on furnace gas composition were collected using a gas chromatograph. At different blowing stages, the separation conditions and detection parameters of the chromatographic column were adjusted according to the changing characteristics of the furnace gas composition to ensure accurate detection of the content of each component. The second fusion module is used to perform weighted fusion of each data in the second monitoring data based on the blowing stage to generate a second fusion feature vector. The quality prediction module is used to input the second fused feature vector into a preset quality prediction model to obtain the quality state prediction result of the liquid metal; The theoretical calculation module is used to calculate the theoretical values ​​of the composition of the liquid metal based on the thermodynamic and kinetic equilibrium model of converter smelting; The detection result module is used to compare the predicted quality state with the theoretical value. If the deviation is greater than the allowable range, the predicted quality state is corrected according to the theoretical value to obtain the detection result of the liquid metal. The results display module is used to intuitively display the detection results, first monitoring data, and second monitoring data through a visual interface.

Images