Blast furnace molten iron silicon content prediction method and system based on iron-like section multi-modal morphology

Abstract

The application discloses a blast furnace molten iron silicon content prediction method and system based on a multi-modal morphology of an iron sample section, samples the molten iron, and cools the molten iron, and then prepares a standardized section at a unified position under the action of pressure; multi-angle and multi-magnification optical images and three-dimensional morphologies of the section are collected to obtain a multi-scale and multi-modal morphology data set; a deep learning model with a dynamic updating mechanism is used to generate a plurality of silicon content prediction values and confidence, and an optimal prediction value is obtained through confidence weighted fusion, so that rapid and accurate prediction of the silicon content of the blast furnace molten iron is realized, and the real-time regulation and control demand in the smelting process is met.

Description

Technical Field

[0001] This invention relates to a method and system for predicting the silicon content of molten iron in blast furnaces based on the multimodal morphology of iron sample cross sections, belonging to the field of iron and steel smelting technology. Background Technology

[0002] In blast furnace smelting, the silicon content of molten iron is a key indicator reflecting furnace temperature stability. Excessive silicon content increases the viscosity of the molten iron and makes slag-iron separation difficult, while insufficient silicon content indicates insufficient furnace temperature and increased fuel consumption. Currently, steel companies generally use manual sampling followed by laboratory analysis to detect silicon content. This involves manually operating refractory molds to collect samples from the molten iron trough, cooling them into circular iron cakes, and then transporting them to a testing center via pneumatic pipelines for analysis using chemical titration or X-ray fluorescence spectrometry. This method has the following problems: sampling and sample preparation rely on manual labor, requiring operators to work in high-temperature, dusty environments, resulting in high labor intensity and safety risks; manual operation leads to variations in sample forming time and cooling rate, resulting in poor sample consistency and poor test repeatability; the testing cycle is long, and the data lags significantly behind changes in furnace conditions, making it unsuitable for real-time adjustment of process parameters such as blast temperature and pulverized coal injection.

[0003] While existing automated sampling devices reduce human intervention, they lack standardized sample preparation procedures, making it difficult to maintain consistency between the cross-sectional location of the iron sample and the detection area. A few image-based detection methods require the prior preparation of metallographic specimens, involving cutting, mounting, grinding, and polishing the iron sample—a time-consuming process that cannot meet the demands of rapid detection. Furthermore, variations in sampling time intervals and cooling rates can also affect the accuracy of subsequent imaging data.

[0004] In terms of data modeling, existing models are either single, lightweight models or large models. Lightweight models lack accuracy with few training samples, while large models require a large number of samples for effective training and perform poorly with low sample sizes. Neither type of model can simultaneously balance detection speed and accuracy during sample accumulation. These issues make it difficult for existing detection technologies to meet the real-time feedback requirements of silicon content in blast furnace smelting, limiting the fine-grained control and efficiency improvement of the smelting process. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for predicting the silicon content of molten iron in blast furnaces based on the multimodal morphology of iron sample cross sections, thereby solving the technical problems of long cycle and poor consistency in existing methods for detecting silicon content in molten iron in blast furnaces.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0007] In a first aspect, the present invention provides a method for predicting the silicon content of blast furnace hot metal based on the multimodal morphology of iron sample cross-sections, the method comprising:

[0008] Molten iron is sampled, and after the sample is cooled, a cross-section is prepared at a uniform location using pressure to obtain an iron sample with a standardized cross-section. Optical images of the cross-section of the iron sample are acquired at different angles and magnifications, and three-dimensional morphology is scanned to obtain a multi-scale, multi-modal morphology dataset, which includes two-dimensional feature information and three-dimensional feature information. Multiple sets of feature data from the morphology dataset are input into a deep learning model to generate multiple silicon content prediction values ​​and their corresponding confidence levels. The deep learning model adopts a dynamic update mechanism. When the accumulated sample amount reaches a preset threshold or the prediction error of the deep learning model exceeds a preset threshold, the model update training is started, and after verification, it automatically switches to the new model that has completed training for prediction. Multiple silicon content predictions are weighted and fused according to their respective confidence levels to obtain the optimal silicon content prediction.

[0009] In conjunction with the first aspect, the cooling conditions are further as follows: the iron sample is first naturally cooled to 500°C, and then cooled to below 100°C by water cooling.

[0010] Furthermore, the cross-section preparation at a uniform location using pressure is achieved through hydraulic pressing, and the deviation of the pressing position, after visual calibration, does not exceed ±0.5mm.

[0011] Furthermore, optical images of the iron sample cross-section were acquired at different angles and magnifications, and three-dimensional morphology scanning was performed, including:

[0012] Low-magnification optical imaging is performed at magnifications of 0.1 to 10 times to acquire the overall morphology and color gradient features of the iron sample cross-section from multiple angles. The multi-angle acquisition is achieved through automatic rotation. High-magnification microscopic imaging was performed at magnifications of 50 to 200 times to collect microscopic pore and crack features of the iron sample cross-section; three-dimensional morphological data of the iron sample cross-section were obtained by laser scanning to acquire information on pore depth and crack width.

[0013] Furthermore, the dynamic update mechanism includes:

[0014] When the number of samples is less than 1500, a lightweight model is used for prediction. The lightweight model is used to extract boundary contours and pore features by reducing the number of network layers. When the number of samples reaches 3000, or the continuous prediction error of the lightweight model exceeds 0.05%, the large model training is started. After the training is completed and the verification is passed, the large model containing residual connections and attention mechanisms is switched to make predictions. The large model is used to extract features such as porosity, crack distribution and color gradient change rate.

[0015] Furthermore, weighted fusion is performed based on their respective confidence levels, including:

[0016] Assign a weight that is positively correlated with its confidence level to each predicted value;

[0017] Predictions with a confidence level below 0.6 are removed, and the remaining valid predictions are weighted and fused to obtain the optimal silicon content prediction.

[0018] Furthermore, the method also includes:

[0019] The newly added imaging data and actual silicon content data during the production process are added to the morphology dataset, and incremental training is performed on the deep learning model according to the preset cycle. The system monitors the operational status of each stage of sampling, imaging, and prediction in real time. When sampling is interrupted, image quality does not meet preset indicators, or data transmission is interrupted, it switches to the backup prediction model.

[0020] Secondly, the present invention provides a blast furnace hot metal silicon content prediction system based on multimodal morphology of iron sample cross-sections, comprising: The sampling unit is used to sample molten iron. After cooling the obtained iron sample, a cross-section is prepared at a uniform location by applying pressure to obtain an iron sample with a standardized cross-section. The imaging unit is used to acquire optical images of the cross-section of the iron sample at different angles and magnifications, and to perform three-dimensional morphology scanning to obtain a multi-scale, multi-modal morphology dataset. The prediction unit has a built-in deep learning model with a dynamic update mechanism, which is used to predict the silicon content of the morphology dataset and to start model update training and automatic switching based on the amount of accumulated samples or prediction error. The output unit is used to weight and fuse multiple silicon content prediction values ​​generated based on multiple sets of feature data according to their respective confidence levels, and output the optimal silicon content prediction value.

[0021] In conjunction with the second aspect, furthermore, the prediction unit initiates model update training and automatic switching based on the accumulated sample size or the continuous prediction error exceeding 0.05%;

[0022] The system also includes a self-diagnosis unit and an iteration unit. The self-diagnosis unit is used to monitor the operating status of each unit in real time, and automatically alarm and switch to the backup prediction model when an abnormality occurs. The iteration unit is used to supplement the morphology dataset with the imaging data and silicon content annotation data added in the production process, and perform incremental training on the model according to a preset period.

[0023] Thirdly, the present invention provides a computer device, comprising:

[0024] Memory, used to store instructions;

[0025] A processor for executing the instructions, causing the device to perform operations implementing any of the methods of the first aspect.

[0026] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0027] This invention provides a method and system for predicting the silicon content of molten iron in blast furnaces based on the multimodal morphology of iron sample cross sections. By preparing cross sections at a unified location under pressure and then visually calibrating them, the detection areas of all iron samples are made consistent, thereby improving detection efficiency.

[0028] Multimodal imaging, which combines multi-angle, multi-magnification optical images with three-dimensional topography scanning, is used to obtain spatial features such as cross-sectional two-dimensional texture, color gradient, pore depth, and crack width. Through multi-scale multimodal feature fusion and dynamic evolution model, the prediction accuracy is improved.

[0029] The deep learning model automatically switches between a lightweight model and a large model based on the amount of accumulated samples or the prediction error, and updates after verification, balancing detection speed and accuracy. When there are insufficient samples, the lightweight model ensures prediction speed, while the large model ensures prediction accuracy when there are sufficient samples, so as to adapt to different blast furnace operating conditions.

[0030] By weighting and fusing multiple silicon content predictions according to their confidence levels and eliminating predictions with a confidence level below 0.6, the random error of a single prediction is reduced. The entire process requires no manual intervention, avoiding harm to operators from high-temperature and dusty environments, and standardized sample preparation ensures test repeatability.

[0031] By adding new data to the dataset and performing incremental training at preset intervals, the model can continuously adapt to different blast furnace operating conditions. By monitoring the operating status of each link in real time and switching to the backup prediction model in case of anomalies, the continuity of the detection process is ensured, and smelting efficiency and product quality are improved. Attached Figure Description

[0032] Figure 1 A flowchart illustrating the method for predicting silicon content in blast furnace molten iron based on multimodal morphology of iron sample cross-sections, provided in an embodiment of the present invention. Detailed Implementation

[0033] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0034] This invention provides a method for predicting the silicon content of blast furnace hot metal based on multimodal morphology of iron sample cross-sections, such as... Figure 1 As shown, the specific steps include:

[0035] Step 1: Automatic Sampling and Sample Preparation

[0036] A mobile platform adapted to the working conditions of a blast furnace molten iron trough, equipped with an actuator, automatically grasps the sampling mold to complete molten iron sampling. The mobile platform has autonomous navigation capabilities and can accurately reach the preset sampling point. The actuator can automatically complete mold grasping, molten iron sampling, iron sample transfer, and docking with the sample preparation device. The sample preparation process uses precise positioning control to ensure the consistency of the detection area of ​​all iron sample cross-sections, avoiding secondary damage to the cross-section.

[0037] After the iron sample is formed, it is first naturally cooled to 500℃, and then cooled to below 100℃ using water cooling. After cooling, the robotic arm moves to the crushing device to break off the refractory outer shell. The robotic arm uses its end effector to visually identify the iron sample, grasps it, and transfers it to the hydraulic sample preparation device. The hydraulic sample preparation device is servo-controlled, with a breaking pressure of not less than 20MPa, a breaking speed of 0.5~1mm / s, and the breaking position is calibrated by a visual positioning system with a deviation of no more than ±0.5mm. After sample preparation, the robotic arm transfers the iron sample to the stage of the multi-scale multimodal imaging unit. The entire process takes no more than 90 seconds.

[0038] Step 2: Multi-scale and multi-modal imaging

[0039] Images of the prepared iron sample cross-section were acquired at different magnifications. By combining optical imaging and three-dimensional morphology scanning in a multimodal manner, multi-angle data were obtained through angle adjustment to comprehensively capture the relevant features of the cross-section.

[0040] After the iron sample is fixed on the stage, low-magnification optical imaging is performed at magnifications of 0.1 to 10x. The overall morphology and color gradient characteristics of the iron sample cross-section are acquired from three angles: 0°, 120°, and 240°, through automatic stage rotation. Then, the system switches to a high-magnification microscopic imaging module, acquiring microscopic pore and crack features of the cross-section at magnifications of 50 to 200x. Finally, a laser scanning module is activated to obtain three-dimensional morphological data of the cross-section, including pore depth and crack width information. Uniform illumination is provided by an LED cold light source during the imaging process. The total imaging time does not exceed one minute, generating optical images and three-dimensional point cloud data.

[0041] Step 3: Deep Learning Prediction

[0042] A dataset covering the silicon content range of 0.15% to 1.00% in blast furnace hot metal was constructed, with 10 to 30 samples collected for each silicon content gradient, resulting in a multi-scale, multi-modal image-silicon content association dataset with over 2000 valid samples. XRF fluorescence analysis was used to obtain the true silicon content values ​​as labeled data. Data preprocessing included image denoising, 3D point cloud denoising, and feature normalization.

[0043] The deep learning model employs a dynamic evolution mechanism adapted to the sample size. When the number of samples is less than 1500, a lightweight model is used for prediction, which extracts boundary contours and pore features by reducing the number of network layers. When the accumulated number of samples reaches 3000, or when the continuous prediction error of the small model exceeds 0.05%, large model training is initiated. The large model includes residual connections and attention mechanisms to extract features such as pore proportion, crack distribution, and color gradient change rate. After training and validation are completed, the system automatically switches to the large model for prediction. The system automatically checks the accumulated sample size and the prediction error of the small model every 24 hours, and if the conditions are met, large model training and validation are initiated. New samples added during production are automatically added to the dataset, and incremental training of the large model is performed monthly.

[0044] Step 4: Result Optimization and Output

[0045] By weighted fusion of multiple imaging data and comparison of results from different prediction modes, the optimal silicon content prediction value is selected. The silicon content prediction value, confidence level, detection time and other data are pushed to the monitoring terminal in the blast furnace control room in real time via industrial Ethernet and stored in the database.

[0046] The specific steps are as follows:

[0047] (1) The collected multi-scale and multi-modal morphology data of the iron sample cross-section is divided into multiple independent feature subsets. Each subset is input into a pre-trained deep learning model that integrates two-dimensional visual features and three-dimensional structural features to generate corresponding independent prediction results. Specifically, low-magnification optical feature subset: Features are extracted from low-magnification images collected from 3 to 5 angles to form 3 to 5 independent low-magnification angle feature subsets. Each subset generates a silicon content prediction value and confidence level.

[0048] High-magnification microscopic feature subsets: 30 to 50 non-overlapping microscopic regions are randomly selected in the central area of ​​the iron sample cross-section, and high-magnification microscopic features of each region are extracted to form 30 to 50 independent high-magnification position feature subsets. Each subset corresponds to one silicon content prediction value and confidence level.

[0049] 3D topography feature subset: Extract global structural features from the full-domain 3D point cloud data obtained by laser scanning to form an independent 3D topography feature subset, which corresponds to generating a silicon content prediction value and confidence level.

[0050] The confidence level of each predicted value is calculated by combining the model uncertainty score, feature quality score, and historical bias score.

[0051] Model uncertainty score: In the prediction phase, the model's fully connected layer is forward-propagated 20 times to obtain 20 predicted values. The standard deviation of the predicted values ​​is calculated. The smaller the standard deviation, the lower the model uncertainty and the higher the score.

[0052] Feature quality score: For low-magnification images, the three indicators of image sharpness, gray-level uniformity, and surface integrity are weighted in a 1:1:1 ratio; for high-magnification images, the three indicators of image sharpness, noise level, and feature quantity are weighted in a 1:1:1 ratio; for 3D point clouds, the three indicators of point cloud integrity, point cloud density uniformity, and normal vector consistency are weighted in a 1:1:1 ratio to obtain the final feature quality score.

[0053] Historical bias score: Calculated based on the model’s average prediction error over the past 50 similar samples. The smaller the average error, the higher the score. (2) Remove all prediction results with a confidence level below 0.6. If the number of valid prediction results retained is not less than 2, proceed to the subsequent weighted fusion stage. If the number of valid results is less than 2, the quality of the detection data is deemed unqualified and the re-imaging process is triggered.

[0054] (3) For all the retained valid prediction results, adaptive weight allocation is performed according to their confidence level. The confidence levels of all valid prediction results are normalized so that the sum of the weights is 1. Each valid prediction value is multiplied by its corresponding fusion weight and then summed to obtain the final comprehensive optimal silicon content prediction value.

[0055] (4) Determine whether the predicted value is within the normal production fluctuation range of the blast furnace. If it exceeds the range, mark it as an abnormal result.

[0056] Step 5: Model Iteration and Self-Diagnosis

[0057] New imaging data and actual silicon content data added during the production process are automatically added to the dataset. The deep learning model is incrementally trained regularly to continuously optimize its adaptability to iron sample characteristics under different working conditions. At the same time, the operation status of each stage of sampling, imaging, and prediction is monitored in real time. When sampling is interrupted, image quality does not meet preset indicators, or data transmission is interrupted, the system switches to a backup prediction model to ensure the continuity of the detection process.

[0058] During the model iteration and optimization process, historical detection data and newly added data are used together for training to ensure that the model's adaptability is continuously improved; self-diagnostic alarm information is recorded synchronously to the system database to facilitate fault tracing and equipment maintenance.

[0059] The present invention also provides a blast furnace hot metal silicon content prediction system based on multimodal morphology of iron sample cross-section, which specifically includes the following units.

[0060] The sampling unit is used to sample molten iron. After cooling the obtained iron sample, a cross-section is prepared at a uniform location using pressure to obtain an iron sample with a standardized cross-section. This unit is fundamental to achieving rapid testing and ensuring the standardization and automated production of iron samples.

[0061] The sampling unit consists of an all-terrain mobile platform, a six-degree-of-freedom robotic arm, a hydraulic sampling device, a sampling mold storage box, various sensors, and communication modules. The all-terrain mobile platform features a tracked design, adaptable to complex terrain around the molten iron ditch, and is equipped with a visual tracking navigation module, enabling it to autonomously move along a route to a preset sampling point. The six-degree-of-freedom robotic arm is equipped with high-temperature protection, a load capacity greater than 10kg, and a repeatability accuracy better than 1mm. It can automatically grab the sampling mold from the material box; the mold body is made of refractory material, with a temperature resistance of ≥1800℃, and can be inserted into the molten iron ditch to collect samples in no more than 30 seconds.

[0062] The sample preparation process is as follows:

[0063] After the iron sample is formed, it is first naturally cooled to 500℃, and then water-cooled to below 100℃. After cooling, the robotic arm moves to the crushing device to break off the refractory outer shell. The iron sample is then visually identified by the end effector of the robotic arm, grasped, and transferred to the hydraulic sample preparation device. This device uses a servo control system with a crushing pressure of not less than 20MPa and a crushing speed of 0.5~1mm / s. The crushing position is calibrated by a visual positioning system with a deviation of no more than ±0.5mm to ensure the consistency of the detection area of ​​all iron sample cross-sections. After sample preparation, the robotic arm transfers the iron sample to the stage of the multi-scale multimodal imaging unit. The entire process is automated and takes no more than 90 seconds.

[0064] The imaging unit is used to acquire optical images of the cross-section of the iron sample from different angles and magnifications, and to perform three-dimensional morphological scanning to obtain a multi-scale, multi-modal morphological dataset. This unit is used to acquire comprehensive feature data of the iron sample cross-section, solving the problem of insufficient feature extraction by a single imaging method.

[0065] The multi-scale, multi-modal imaging unit comprises a low-magnification optical imaging module, a high-magnification microscopic imaging module, a laser scanning module, an angle-adjustable stage, and an LED cold light source module. The low-magnification optical imaging module supports magnifications from 0.1 to 10x, while the high-magnification microscopic imaging module supports magnifications from 50 to 200x. The low-magnification module captures the overall cross-sectional morphology, boundaries, large cracks, and color gradients, while the high-magnification module captures microscopic pores with a diameter of at least 10 micrometers. The laser scanning module uses a semiconductor laser with a wavelength of 405 nm, a scanning accuracy of 0.5 to 3 micrometers, and a scanning step size of 0.1 to 0.3 micrometers. It can acquire three-dimensional cross-sectional morphology data, such as pore depth and crack width, compensating for the limitations of two-dimensional information in optical imaging.

[0066] The imaging process is as follows:

[0067] After the iron sample is fixed on the stage, low-magnification optical imaging is first performed at three angles: 0°, 120°, and 240°. Then, high-magnification optical imaging is switched at the same angle. Finally, laser scanning is initiated to acquire full-domain 3D data. During the imaging process, a cold LED light source provides uniform illumination, avoiding the influence of light, shadows, and temperature. The total imaging time is ≤1 minute, generating no fewer than 23 optical images, including one low-magnification image at each of the three angles, 20 high-magnification images at one angle, and one set of 3D point cloud data.

[0068] The prediction unit, which incorporates a deep learning model with a dynamic update mechanism, is used to predict the silicon content of the morphology dataset. It initiates model update training and automatic switching based on the amount of accumulated samples or the prediction error. This unit is the core of achieving quantitative prediction of silicon content, and it balances speed and accuracy through a dynamic evolution model.

[0069] The construction of the morphology dataset included: collecting iron samples with silicon content ranging from 0.15% to 1.00% in blast furnace molten iron, covering more than 99% of the production conditions in steel enterprises. For each silicon content gradient, 10-30 samples were collected at intervals of 0.02% to 0.05%, resulting in a dataset containing over 2000 valid samples of multi-scale, multi-modal images and silicon content. XRF fluorescence analysis was used to obtain the true silicon content values, with a detection accuracy of ±0.01%. Data preprocessing included image denoising, 3D point cloud denoising, and feature normalization to ensure data quality.

[0070] Model architecture design: A dynamic architecture is adopted, which involves initializing small models and evolving large models.

[0071] Small model stage: Based on a lightweight architecture, the number of network layers is simplified, the number of depthwise separable convolutional layers is reduced by 20%, and the focus is on extracting easily identifiable features such as boundary contours and large-sized pores. It is suitable for scenarios with less than 1500 samples, with a model inference time of ≤ 3 seconds and a prediction accuracy of ≥ 90%.

[0072] Large model stage: Based on the large model architecture, residual connections and attention mechanisms are introduced to enhance the extraction capabilities of key features such as the proportion of micropores, crack distribution, and color gradient change rate. When the number of samples is ≥3000, the system automatically triggers large model training with an inference time of ≤5 seconds and a prediction accuracy of ≥98%.

[0073] Dynamic evolution mechanism: The system automatically detects the accumulated sample volume and the prediction error of the small model every 24 hours. If the sample volume meets the standard or the small model error exceeds 0.05% for three consecutive times, the large model training and validation will be started. After the validation is passed, the system will automatically switch to large model prediction. New samples added during the production process will be automatically added to the dataset. The large model will be incrementally trained once a month to continuously optimize the accuracy.

[0074] Feature extraction and training: First, based on the essential differences between image and 3D topography data, visual features of the image and spatial features of the 3D structure are extracted separately. The visual features include cross-sectional texture distribution, geometric shape and color gradient related representations, while the spatial features focus on the cross-sectional depth dimension and three-dimensional structural information. The two types of heterogeneous features are integrated through feature fusion technology to construct a multi-dimensional feature expression that can comprehensively characterize the cross-section of the iron sample and the strong related information of silicon content, avoiding the information limitations of a single feature dimension.

[0075] The output unit is used to weight and fuse multiple silicon content prediction values ​​generated based on multiple sets of feature data according to their respective confidence levels, and output the optimal silicon content prediction value. This unit is used to optimize the prediction results and provide real-time feedback to ensure the reliability of production applications.

[0076] The output unit's result optimization employs a multi-set feature data weight fusion strategy. Multi-angle, multi-scale, and multi-modal feature data of the iron sample are input into a model that integrates visual and 3D structural features, generating multiple sets of prediction results based on features of different dimensions. Weights are assigned based on the confidence level of each prediction result: high-confidence results are given higher weights, medium-confidence results are given appropriate adaptation weights, and low-confidence results are directly discarded to avoid interference from invalid information. A weighted fusion algorithm integrates the effective prediction results, outputting the optimal comprehensive silicon content prediction value. This approach fully leverages the advantages of a single model in fusing multi-source features while reducing random errors in single predictions through cross-validation of multiple sets of feature data, further enhancing the reliability of the results.

[0077] Data transmission: Through industrial Ethernet, the predicted silicon content, confidence level, detection time and other data are pushed to the monitoring terminal in the blast furnace control room in real time and stored in the database, supporting historical data traceability and model optimization.

[0078] Iteration unit: Used to supplement the multi-scale multimodal morphology dataset with newly added imaging data and silicon content annotation data in the production process, and to perform incremental training on the model according to a preset period.

[0079] Self-diagnostic unit: Used to monitor the operating status of each unit in real time, automatically alarm and switch to the backup predictive model when an abnormality occurs.

[0080] The present invention also provides a computer device, including a memory and a processor; wherein the memory is used to store instructions, and the processor is used to execute the instructions, causing the device to perform the operation of the above-described method for predicting the silicon content of molten iron based on the multimodal morphology of iron sample cross sections.

[0081] The computer device is deployed as a server, and its hardware configuration includes a CPU, GPU, memory, and hard disk storage. The operating system is Linux or Windows Server, and the deep learning framework is PyTorch.

[0082] Example 1:

[0083] This embodiment is used to verify the sampling consistency, imaging effectiveness, model prediction accuracy, and detection efficiency of a blast furnace molten iron silicon content prediction system based on multimodal morphology of iron sample cross sections, providing data support for industrial applications.

[0084] The specific experimental equipment and parameters are as follows:

[0085] Sampling unit: All-terrain tracked mobile platform, load capacity 50kg, navigation accuracy ±5cm; six-degree-of-freedom robotic arm, repeatability ±0.1mm, load capacity 12kg; hydraulic sample preparation device, maximum pressure 20MPa, crushing speed 0.8mm / s; sampling mold, diameter 35mm, thickness 15mm, temperature resistance 1800℃.

[0086] Imaging units: 2K industrial camera, 0.2x magnification; 4K microscopic imaging, 80x magnification; line laser scanner model: Keyence LJ-X8000, scanning accuracy 0.5 micrometers; motorized rotating stage positioning accuracy ±0.1°.

[0087] Prediction unit: The server CPU is an Intel Xeon 8375C, the GPU is an NVIDIA A100, and the deep learning framework is PyTorch 1.12.

[0088] Comparison testing equipment: XRF fluorescence analyzer, model: XRF-1800, accuracy ±0.01%.

[0089] The experimental procedure is as follows:

[0090] Sample preparation: Simulate the composition of molten iron in a blast furnace, with a silicon content of 0.25%~0.75% at intervals of 0.05%, for a total of 10 gradients. Prepare 30 iron samples for each gradient, for a total of 300 samples. Molten iron is melted in an induction furnace, and standardized cross-sectional iron samples are obtained by sampling, cooling, and pressing through a sampling unit.

[0091] Imaging data acquisition: The iron sample is placed on the imaging unit stage, and low-magnification images are acquired according to the preset procedure. The magnification is 0.2 times, and the images are acquired from three angles: 0°, 120°, and 240°. The high-magnification images are acquired at 80 times, and 15 images are acquired from different positions. These images are combined with the laser scanning 3D data to generate 18 images and 1 set of point cloud data. The average imaging time is 60 seconds per sample.

[0092] Dataset construction: The true silicon content of all samples was detected using an XRF fluorescence analyzer and used as labeled data; 260 samples were randomly selected as the training set, 30 as the validation set, and 10 as the test set for data preprocessing.

[0093] Model training and testing: In the small model stage, the model was iterated 100 times based on the training set, and the error on the validation set was stabilized at 0.1%. The small model was used to predict test samples, and the prediction error and detection time were recorded. The accuracy on the test set was 90.5%.

[0094] The total time for a single sample is as follows: sampling and preparation takes 2 minutes and 30 seconds, imaging takes 30 seconds, prediction takes 5 seconds, and result output takes 5 seconds, for a total of 3 minutes and 10 seconds, which meets the requirements for minute-level detection.

[0095] Example 2:

[0096] This embodiment demonstrates an industrial application in a blast furnace of a steel group. The original method used was traditional manual sampling and XRF testing, with a testing cycle exceeding 30 minutes. The system of this invention is deployed in this blast furnace to monitor the silicon content of molten iron in real time and guide process adjustments.

[0097] The system is deployed as follows: the sampling unit is an all-terrain mobile platform adapted to the high temperature and dust conditions around the iron ditch, and is equipped with a dust cover and a heat insulation cover; the robotic arm is equipped with a high temperature resistant protective sleeve, and the sampling position is calibrated by laser positioning; the hydraulic sampling device maintains a safe distance of 20m from the iron ditch, and the sampling mold storage material box can accommodate 50 molds.

[0098] The imaging unit is deployed inside a sealed inspection box, maintaining a safe distance of 20 meters from the molten iron ditch. The prediction unit's server is deployed in a computer room and connected to the imaging unit and production system via an industrial Ethernet network.

[0099] Application Process: The system starts and initializes, performing self-checks every time it's powered on or every 24 hours, including robotic arm positioning and imaging unit calibration, to ensure proper equipment operation. During real-time monitoring of blast furnace tapping, the system automatically completes sampling, imaging, and prediction every 10 minutes, outputting the predicted silicon content value and confidence level. The control room adjusts process parameters based on real-time silicon content data.

[0100] After one month of field operation, the number of valid samples accumulated exceeded 2,000. The system automatically triggered large-scale model training, and after training was completed, it switched to large-scale model prediction while retaining a small model as a backup. In field applications, the system's average total process time was 6 minutes, significantly reducing detection time. The large-scale model prediction achieved an average accuracy of 98%, with a monthly failure rate of less than 1%. The equipment demonstrated good operational stability, meeting production requirements.

[0101] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the silicon content of blast furnace hot metal based on multimodal morphology of iron sample cross-sections, characterized in that, include: Molten iron is sampled, and after the sample is cooled, a cross-section is prepared at a uniform location using pressure to obtain an iron sample with a standardized cross-section. Optical images of the cross-section of the iron sample are acquired at different angles and magnifications, and three-dimensional morphology is scanned to obtain a multi-scale, multi-modal morphology dataset, which includes two-dimensional feature information and three-dimensional feature information. Multiple sets of feature data from the morphology dataset are input into a deep learning model to generate multiple silicon content prediction values ​​and their corresponding confidence levels. The deep learning model adopts a dynamic update mechanism. When the accumulated sample amount reaches a preset threshold or the prediction error of the deep learning model exceeds a preset threshold, the model update training is started, and after verification, it automatically switches to the new model that has completed training for prediction. Multiple silicon content predictions are weighted and fused according to their respective confidence levels to obtain the optimal silicon content prediction.

2. The method according to claim 1, characterized in that, The cooling conditions are as follows: the iron sample is first naturally cooled to 500°C, and then cooled to below 100°C by water cooling.

3. The method according to claim 1, characterized in that, The cross-section is prepared by pressure application at a uniform location using hydraulic pressing, and the deviation of the pressing position after visual calibration does not exceed ±0.5mm.

4. The method according to claim 1, characterized in that, Optical images of the cross-section of the iron sample were acquired at different angles and magnifications, and three-dimensional morphological scanning was performed, including: Low-magnification optical imaging is performed at magnifications of 0.1 to 10 times to acquire the overall morphology and color gradient features of the iron sample cross-section from multiple angles. The multi-angle acquisition is achieved through automatic rotation. High-magnification microscopic imaging was performed at magnifications of 50 to 200 times to collect microscopic pore and crack features of the iron sample cross-section; three-dimensional morphological data of the iron sample cross-section were obtained by laser scanning to acquire information on pore depth and crack width.

5. The method according to claim 1, characterized in that, The dynamic update mechanism includes: When the number of samples is less than 1500, a lightweight model is used for prediction. The lightweight model is used to extract boundary contours and pore features by reducing the number of network layers. When the number of samples reaches 3000, or the continuous prediction error of the lightweight model exceeds 0.05%, the large model training is started. After the training is completed and the verification is passed, the large model containing residual connections and attention mechanisms is switched to make predictions. The large model is used to extract features such as porosity, crack distribution and color gradient change rate.

6. The method according to claim 1, characterized in that, Weighted fusion is performed based on their respective confidence levels, including: Assign a weight that is positively correlated with its confidence level to each predicted value; Predictions with a confidence level below 0.6 are removed, and the remaining valid predictions are weighted and fused to obtain the optimal silicon content prediction.

7. The method according to claim 1, characterized in that, The method further includes: The newly added imaging data and actual silicon content data during the production process are added to the morphology dataset, and incremental training is performed on the deep learning model according to the preset cycle. The system monitors the operational status of each stage of sampling, imaging, and prediction in real time. When sampling is interrupted, image quality does not meet preset indicators, or data transmission is interrupted, it switches to the backup prediction model.

8. A system for predicting silicon content in blast furnace hot metal based on multimodal morphology of iron sample cross-sections, characterized in that, include: The sampling unit is used to sample molten iron. After cooling the obtained iron sample, a cross-section is prepared at a uniform location by applying pressure to obtain an iron sample with a standardized cross-section. The imaging unit is used to acquire optical images of the cross-section of the iron sample at different angles and magnifications, and to perform three-dimensional morphology scanning to obtain a multi-scale, multi-modal morphology dataset. The prediction unit has a built-in deep learning model with a dynamic update mechanism, which is used to predict the silicon content of the morphology dataset and to start model update training and automatic switching based on the amount of accumulated samples or prediction error. The output unit is used to weight and fuse multiple silicon content prediction values ​​generated based on multiple sets of feature data according to their respective confidence levels, and output the optimal silicon content prediction value.

9. The system according to claim 8, characterized in that, The prediction unit initiates model update training and automatic switching based on the accumulated sample size or the continuous prediction error exceeding 0.05%. The system also includes a self-diagnosis unit and an iteration unit. The self-diagnosis unit is used to monitor the operating status of each unit in real time, and automatically alarm and switch to the backup prediction model when an abnormality occurs. The iteration unit is used to supplement the morphology dataset with the imaging data and silicon content annotation data added in the production process, and perform incremental training on the model according to a preset period.

10. A computer device, characterized in that, include: Memory, used to store instructions; A processor for executing the instructions, causing the device to perform operations implementing the method as described in any one of claims 1-7.

Images