Intelligent monitoring method for blast furnace tuyere working condition based on visual monitoring and deep learning

Abstract

This invention discloses an intelligent monitoring method for blast furnace tuyere operating conditions based on visual monitoring and deep learning. The method includes: acquiring monitoring video containing tuyere flames, injection nozzles, and surrounding furnace walls from the existing blast furnace monitoring system; decoding the monitoring video and extracting image frames at a preset frame extraction interval of N seconds (preferably one frame every 60 seconds); in the sample construction stage, cropping the tuyere area based on tuyere layout and camera calibration, and performing quality screening and data augmentation to train a blast furnace tuyere operating condition monitoring model; in the online monitoring stage, inputting the real-time extracted full-frame images into the model to output the tuyere location and operating condition category; when the full-frame detection does not obtain results that meet the output conditions, enabling candidate ROI region detection; mapping the detection results within the ROI back to the full-frame coordinate system; and selecting the optimal output from the candidate results to improve the robustness and stability of the monitoring; finally, jointly analyzing the results with on-site operating parameters such as air volume, air pressure, and cooling water flow rate for real-time evaluation of blast furnace tuyere operating conditions.

Description

Technical Field

[0001] This invention relates to the field of condition monitoring and intelligent diagnosis technology for blast furnace ironmaking processes in the metallurgical industry, and in particular to an intelligent monitoring method for blast furnace tuyere conditions that utilizes blast furnace monitoring video data and combines it with a deep learning target detection network to monitor tuyere conditions. Background Technology

[0002] The tuyere is a crucial component in the blast furnace ironmaking process, responsible for both the input of hot blast and the entry of injected fuel into the furnace. The flame morphology, injection status, and surrounding furnace wall characteristics of the tuyere directly reflect the combustion and permeability within the furnace. Abnormal conditions at the tuyere, such as slag buildup, coal shortage, excessive coal injection, or abnormal flame morphology, can easily lead to furnace condition fluctuations, decreased smelting efficiency, and even tuyere burn-out, furnace condition deterioration, and safety accidents. Therefore, real-time monitoring and assessment of the tuyere's operating conditions are essential for the stable operation of the blast furnace.

[0003] Current production sites typically rely on manual inspections by staff or manual observation and judgment of air vent conditions through monitoring screens. However, this approach has the following shortcomings in practical applications:

[0004] (1) Manual judgment is highly dependent on human experience, highly subjective, difficult to form a unified standard, and difficult to monitor continuously for a long time and trace back afterward.

[0005] (2) The blast furnace tuyeres area is exposed to high temperature, strong light, smoke and dust, heat wave refraction and other environments for a long time. The monitoring video is prone to blurring, occlusion, overexposure / underexposure and other phenomena, resulting in high false alarm and false alarm rates of traditional algorithms, which are difficult to adapt to long-term stable operation.

[0006] (3) The amount of monitoring video data is large. If reasonable frame extraction and image quality control are not performed, the computational load of online monitoring will increase significantly and it will be easily affected by invalid frames. At the same time, the vent images show scale and position changes under different camera installation positions, field of view differences and vent position offsets. If there is a lack of unified regional constraints and adaptive ROI mechanism, the online monitoring results are prone to missed detections, false detections or positioning drift, making it difficult to achieve stable and reliable working condition judgment.

[0007] (4) Blast furnaces typically record operating parameters such as air volume, air pressure, and cooling water flow rate simultaneously. Existing solutions often fail to correlate these parameters with visual monitoring results, making it difficult to support more reliable anomaly diagnosis and operational decisions.

[0008] Therefore, it is necessary to propose an intelligent monitoring method for blast furnace tuyere scenarios, forming an integrated process from video frame extraction by N seconds, tuyere area cropping and quality control during the training phase, data annotation and enhancement, target detection model training, to full-frame monitoring and ROI adaptive detection in the online phase, and comprehensive analysis combined with on-site operating parameters, in order to improve the accuracy and robustness of tuyere abnormal operating condition monitoring, while meeting the real-time requirements. Summary of the Invention

[0009] To address the problems of existing blast furnace tuyere condition monitoring methods, such as reliance on manual experience, instability under conditions of smoke and dust obstruction and fluctuating flame brightness, and high computational load and susceptibility to invalid frames, this invention proposes an intelligent monitoring method for blast furnace tuyere conditions based on visual monitoring and deep learning. This method reduces redundancy by sampling frames at time intervals while ensuring monitoring coverage, achieves joint output of tuyere location and condition category through a target detection network, and introduces an ROI detection mechanism in the online phase to enhance robustness to complex scenarios such as obstruction and target offset, thereby achieving continuous, stable, and traceable monitoring of tuyere conditions.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent monitoring of blast furnace tuyeres operating conditions based on visual monitoring and deep learning includes the following steps: (1) On-site data acquisition: Acquire monitoring video data including tuyer flame, injection port and surrounding furnace wall from the blast furnace monitoring system; and combine on-site operating parameters, including at least one of air volume, air pressure and cooling water flow rate. (2) Video processing: Decode the surveillance video and extract image frames at a preset frame extraction interval of N seconds; (3) Sample construction: Based on the blast furnace tuyere layout and camera calibration information, crop the area image containing the target tuyere from the image frame; adjust the contrast of the cropped image to enhance the details; and remove invalid samples that are severely obscured, abnormally exposed, too dark or too bright. (4) Data labeling: Manually label the valid samples and mark each vent area image as "normal" or "abnormal". Abnormalities include slag hanging, coal shortage, excessive coal injection and abnormal flame shape. (5) Image enhancement: The labeled samples are enhanced, including one or more of blur transformation, brightness adjustment, contrast adjustment, color perturbation, noise superposition and scale change; (6) Model training: The enhanced samples are input into the deep learning target detection model for training to obtain the blast furnace tuyere status detection model. The model is used to output the tuyere location and its operating condition category. (7) Online detection: The image obtained by real-time frame extraction is input into the blast furnace tuyere status detection model. If the output confidence is not lower than the preset threshold, the tuyere detection result is directly output; if the condition is not met, the candidate ROI detection mechanism is triggered. (i) Generate a set of candidate ROI regions, which includes a preset region covering the possible location of the wind vent, or dynamically adjusted based on the detection results of the previous frame; (ii) Crop the image for each candidate ROI region and input it into the model for detection, then map the detection results back to the coordinate system of the entire frame image; (iii) Select the detection box with the highest confidence from the mapped candidate results and use it as the final output result of this frame; (8) Results output and analysis: The test results are overlaid and displayed, and are jointly analyzed with the on-site working parameters for air outlet condition assessment.

[0011] The preset frame-skipping interval N in step (2) is 60 to 120 seconds, preferably 60 seconds.

[0012] The sample construction in step (3) includes: Before generating the sample, the details of the flames at the vent are highlighted by reducing the contrast of the background area. The sampling area is cropped based on the spatial layout of the wind vent and camera calibration information; An effective sample screening mechanism is introduced to remove samples that are severely occluded, have abnormal brightness, or lack details.

[0013] The data annotation in step (4) is completed using a manual annotation tool. The annotation results are then stored in a one-to-one correspondence with the sample images.

[0014] The image enhancement in step (5) includes: (1) Blur enhancement: Random blur, simulating heat wave / out of focus / shaking; (2) Brightness / contrast adjustment: Randomly adjust brightness and contrast to simulate exposure fluctuations; (3) Color perturbation: Randomly adjust hue / saturation / brightness to simulate color shift; (4) Scale transformation: random scaling to adapt to changes in target scale; (5) Geometric perturbation: random rotation / translation / shearing to adapt to changes in viewpoint.

[0015] The blast furnace tuyere status detection model in step (6) is constructed using the YOLOv8 framework to simultaneously output the tuyere detection frame position and its operating condition category.

[0016] The candidate ROI region set in step (7) is a configurable parameter, and its position and size can be adaptively updated according to historical detection results; Specifically, when the previous frame detects the output detection box... At that time, calculate its center point. , in A local candidate ROI set is generated centered on this point, with its width and height as follows: ,in >1 is the expansion coefficient.

[0017] The coordinate mapping in step (7) is based on the following formula: Let the coordinates of the top-left corner of the ROI in the entire image frame be ( The coordinates of the detection box in the ROI subplot coordinate system are: Then the coordinates of the detection box mapped to the coordinate system of the entire image frame are: .

[0018] A smart monitoring system for blast furnace tuyeres operating conditions, based on the method described above, includes: The video acquisition module is used to acquire the video stream from the blast furnace tuyeres monitoring system. The image processing module is used to decode the video and extract image frames at a preset frame extraction interval of N seconds. The sample processing module is used for trimming air vent areas, quality control, data annotation, and enhancement. The model training module is used to train the blast furnace tuyeres condition detection model. The online monitoring module is used to perform whole-frame detection and ROI detection mechanisms; The results output module is used to overlay and display the test results and to perform joint analysis with the operating parameters.

[0019] Beneficial effects of the present invention

[0020] Compared with the prior art, the present invention has at least the following beneficial effects:

[0021] (1) The lightweight YOLOv8 model is used to realize the joint output of wind vent location and working condition category, which makes it easy to overlay the detection box and working condition label for display and record the results and backtrack the analysis. The model can run stably for a long time on resource-constrained devices such as edge gateways, reducing the dependence on computing power.

[0022] (2) To realize automatic monitoring and prompting of the working conditions of the air outlet, reduce the reliance on human experience judgment, and eliminate the need for on-site personnel to continuously monitor the monitoring screen or conduct frequent manual inspections. They can make judgments on the working conditions based on the model output results, thereby improving monitoring efficiency and consistency.

[0023] (3) Introduce an ROI detection mechanism. When the detection of the whole frame does not obtain the result that meets the output conditions, the detection of candidate ROIs is supplemented by enabling coordinate mapping and filtering output, thereby improving the robustness and continuous stability of online monitoring.

[0024] (4) Under on-site deployment conditions, the model inference speed is greater than 25fps, which meets the real-time requirements. At the same time, the recognition accuracy is higher than 90% and the detection recall rate is not less than 92%, which can effectively improve the timeliness and reliability of wind vent monitoring.

[0025] (5) Visual monitoring results can be correlated with on-site operating parameters such as air volume, air pressure, and cooling water flow to improve the reliability and interpretability of anomaly diagnosis and early warning. Attached Figure Description

[0026] Figure 1 This is a flowchart of intelligent monitoring of blast furnace tuyeres, which includes modules such as video acquisition, frame extraction, sample construction, model training, online detection (whole frame → ROI), result output, and operating parameter analysis.

[0027] Figure 2 , Figure 3 Example of online blast furnace tuyere monitoring in a steel plant, showing the detection results under normal and abnormal slag-laden conditions, including the detection frame, category labels, and flame pattern comparison. Detailed Implementation

[0028] To make the objectives and technical solutions of this invention clearer, the embodiments of this invention will be described below with reference to the accompanying drawings. It should be understood that the following embodiments are only for explaining this invention and are not intended to limit the scope of protection of this invention; without departing from the concept of this invention, those skilled in the art can make equivalent substitutions or modifications to the order of steps, parameter forms, or module divisions, all of which should fall within the scope of protection of this invention.

[0029] System input and online frame extraction

[0030] like Figure 1As shown, this invention first acquires a video stream of tuyere monitoring through the existing blast furnace monitoring system. The video stream includes the tuyere flame, injection nozzle, and surrounding furnace wall area. The monitoring system is preferably an industrial camera with a resolution of at least 1080p and a frame rate of at least 25fps. The acquired video stream is decoded to obtain continuous image frames, and frames are extracted from the video stream at a preset frame extraction interval of N seconds to obtain full-frame images for online monitoring. N is a configurable parameter, preferably 60 seconds, to reduce data redundancy and online computation while covering changes in tuyere operating conditions. During online operation, this invention also simultaneously incorporates on-site ironmaking process parameters, including but not limited to air volume, air pressure, and cooling water flow rate. These parameters can be accessed via industrial Ethernet to the on-site PLC control system and stored after being aligned with the extracted frame images using a unified timestamp for subsequent joint analysis, operating condition assessment, and anomaly early warning.

[0031] Model building and acquisition

[0032] like Figure 1 As shown, during the data preparation stage, the video stream from the existing monitoring system of the blast furnace is decoded. The video encoding format supports H.264 / HEVC. After decoding, a continuous RGB image frame sequence is obtained. Every N seconds (default 60s, value range 60~120s), a frame is extracted to form a frame-skipping sequence, so as to ensure the capture of key operating conditions while reducing the amount of online computation. Based on the tuyere layout and camera calibration information, the tuyere in the whole frame image is mapped to the cropped area A. Contrast enhancement and brightness correction are performed on this area. The brightness scaling factor is set to [0.8, 1.2], and the contrast scaling factor is set to [0.9, 1.1] to enhance the details of the flame edge and tuyere boundary. When the proportion of effective visible pixels of the air vent in the clipping region A is less than 5%, it is judged as occlusion and removed. When the mean brightness of the air vent in the clipping region A is lower than μ−3σ or higher than μ+3σ, it is judged as exposure abnormal (too dark / too bright) and removed. When the noise intensity of the air vent in the clipping region A is abnormal (e.g., the estimated variance of Gaussian noise is greater than 20), it is judged as noise abnormal and removed, thereby ensuring the validity of the samples and the quality of the annotation.

[0033] During the sample construction phase, the selected wind vent ROI images were manually labeled. A detection box was drawn for each cropped region A using LabelImg, and a "normal / abnormal" label was assigned. Abnormal conditions were further subdivided into slag adhesion, coal shortage, excessive coal injection, and abnormal flame morphology. The labeled files were in TXT format and a one-to-one mapping relationship was established with the corresponding images, stored uniformly in the / dataset / train / and / dataset / label / directories. To improve the model's adaptability to on-site imaging disturbances, further data augmentation was applied to the samples, including blur enhancement, brightness / contrast adjustment, color perturbation, scale variation, and geometric perturbation. The augmented samples and the original samples together constituted the training set for subsequent model training and evaluation. Specific augmentation parameters are shown in the table below.

[0034] Enhanced type Parameter range illustrate Blur Enhancement Gaussian blur, blur kernel size: blur_limit ∈ [3, 5]; trigger probability p=0.3 Simulated heat wave refraction, slight defocus / lens shake causing blur. Brightness / contrast adjustment Brightness adjustment range: brightness_limit=0.15 (approximately ±15%); Contrast adjustment range: contrast_limit=0.15 (approximately ±15%); Trigger probability p=0.5 Simulates flame intensity fluctuations and exposure changes inside the furnace. Color disturbance Hue shift: hue_shift_limit = ±3; Saturation shift: sat_shift_limit = ±8; Brightness shift: val_shift_limit = ±8; Trigger probability: p = 0.3 Simulates changes in flame color, and color shifts and brightness variations caused by smoke. Scale transformation Scaling range: scale=0.1 Simulates the difference in target scale caused by changes in focal length / shooting distance Geometric perturbation Rotation: degrees = ±2; Translation: translate = 0.1 (approximately ±10%); Shear: shear = 0.3 Simulates slight changes in installation angle, minor vibrations, and changes in viewing angle.

[0035] During the model training phase, YOLOv8 was selected as the object detection network. The detection head simultaneously outputs class prediction and bounding box regression. The backbone network uses the C2f module as a lightweight feature extraction network to balance accuracy and speed, making it suitable for deployment on-site GPUs or edge computing devices. The training configuration example is: Batch Size=16, Epochs=150, LearningRate=1e-4, with cosine annealing for learning rate decay. During training, the loss function consists of bounding box regression loss and classification loss, jointly optimizing localization accuracy and operational condition classification ability. The training hardware used is an NVIDIA RTX 3060. After training, 10% of the training set was used as the validation set for performance evaluation. The model achieved an accuracy of 95%, mAP@0.5=92.3%, and a recall ≥92%. The obtained blast furnace tuyeres operational condition monitoring model was used for on-site online monitoring and result output.

[0036] Online monitoring

[0037] like Figure 1As shown, during the online monitoring phase, the blast furnace tuyeres status detection model has been deployed on a field GPU server (equipped with an RTX 4080 SUPER graphics card) with a real-time inference speed of 100fps. It can also be deployed to an edge gateway to achieve a real-time inference speed of at least 25fps, depending on the field network and power consumption conditions. The system periodically extracts frames from the real-time video stream at N=60s (the actual time interval), inputs the extracted full-frame images into the detection model for full-frame detection, filters candidate boxes using a first confidence threshold CONF_THRES_MAIN=0.20, and outputs the target location and operating condition category. If the full-frame detection fails to obtain a valid result that meets the above threshold conditions (i.e., no valid detection output), the ROI detection mechanism is automatically activated, using a second confidence threshold CONF_THRES_ROI=0.12 to obtain ROI candidate results, and taking the detection with the highest confidence among all ROI results as the final output. Finally, the detection box and the working condition category are overlaid on the image and saved. At the same time, the frame extraction timestamp, frame number, saved file name and working condition category are written into a table or database for working condition statistics, backtracking analysis and anomaly warning.

[0038] ROI detection mechanism

[0039] like Figure 1 As shown, to improve the robustness of monitoring under conditions of smoke and dust interference, brightness fluctuations, and wind vent position shifts, a ROI detection mechanism is activated when no detection result meeting the output conditions is obtained for the entire frame. This mechanism includes candidate ROI generation, ROI-in-region detection, coordinate mapping, and output filtering.

[0040] (1) Generation of candidate ROIs

[0041] A set of candidate ROI regions is preset within the entire image frame. This set is a configurable parameter and can be distributed using a multi-grid or sliding window layout to cover the spatial range where wind vents may appear. Assuming the entire image frame size is W×H, the relative parameter can be used for the k-th candidate ROI. It means that among them The relative position ratio of the ROI center point Given the relative proportions of the ROI's width and height, its pixel scale and position can be calculated using formula (1-3) rounded down: (1); (2); (3); in( () represents the coordinates of the top-left corner of the ROI in the entire frame's image coordinate system, and then... Boundary constraints are applied to ensure that the Region of Interest (ROI) does not exceed the entire frame's image range. Simultaneously, the position and size of the candidate ROI can be adaptively updated based on historical detection results. Let the wind vent detection bounding box output from the previous full-frame detection be... Its center point is shown in formula (4): (4); by A set of local candidate ROIs is generated around the center, and the scale of the ROIs is expanded to improve the offset tolerance. Assuming the expansion coefficient is α>1, the width and height of the local ROIs can be calculated by formula (5): (5); Therefore, if the location of the air vent shifts or the detection fails briefly, detection can be quickly restored through local candidate ROIs.

[0042] (2) ROI coordinate mapping

[0043] For each ROI in the candidate ROI region set, a ROI sub-image is cropped from the entire image frame, and then input into the blast furnace tuyere status detection model for detection to obtain the detection box in the ROI coordinate system. ,in( )and( ) are the coordinates of the top left and bottom right corners of the detection box in the ROI sub-graph coordinate system, respectively. Since the above detection box coordinates are located in the ROI sub-graph coordinate system, they need to be mapped back to the full frame image coordinate system. The mapping relationship is shown in formula (6): (6); in( The coordinates of the detection box are shown in the coordinate system of the entire image frame. Through the above back mapping, the detection results within all ROIs can be unified in the coordinate system of the entire frame, which facilitates subsequent filtering and output.

[0044] Output Results

[0045] All the detection boxes mapped back to the coordinate system constitute a candidate detection result set. This set is then filtered. First, detection results with a confidence level below a preset threshold are removed. Then, the detection box with the highest confidence level is selected from the candidate results as the final output detection result for this frame. The system periodically extracts frames from the real-time video stream and saves the frame extraction monitoring results as image files. The monitoring results for each frame are then recorded in a MySQL table in the database. The program automatically creates the result table `tuyere_results`, which contains a timestamp field, a frame sequence number field `frame_idx`, a frame extraction image file name field `filename`, and a condition category field `class`. The timestamp records the acquisition / processing time of the frame, `frame_idx` identifies the frame's sequence number in the video stream, `filename` associates the saved frame extraction image file, and `class` records the condition category detected for this frame. To ensure record traceability and deduplication, the database sets a unique constraint on the combination of (timestamp, frame_idx, filename). During writes, it uses an "update if exists, insert if not" approach to save results, thus avoiding duplicate writes and supporting updates to the same frame. Furthermore, the database automatically generates a unique ID as the primary key for each record and automatically records the entry time (created_at) for subsequent time-based retrieval, statistics, and backtracking analysis.

[0046] Brief description of on-site implementation effects and abnormal characteristics

[0047] In field trials at two different blast furnaces, the model achieved an average recall rate of 94.6% and an accuracy of 93.7% during three months of continuous operation, with an inference speed of approximately 100 fps, meeting the requirements for real-time monitoring. Through the ROI detection mechanism, it was still able to stably identify the tuyere position even when there were sudden changes or positional shifts. Furthermore, the accuracy of the anomaly warning, which was integrated with process parameters, was improved by 1.2%, reducing the false alarm rate. Figure 2 and Figure 3 Images from the tuyeres of different blast furnaces, among which Figure 2 For air vents operating under normal conditions, the flame edges are regular. Figure 3 For tuyeres exhibiting abnormal operating conditions, the presence of slag buildup is observed through on-site investigation and verification. Visually, abnormal slag buildup typically manifests as irregular flame shapes at the tuyere edge, with jagged or notched outlines. This is due to slag adhering to the inner wall of the tuyere, causing partial obstruction of the flame jet and resulting in flame shape distortion. By effectively utilizing the aforementioned visual characteristics, this invention can provide intuitive evidence for on-site operating condition assessment and abnormal early warning.

[0048] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. The scope of protection of the present invention is defined by the appended claims and any equivalent technical solutions.

Claims

1. A method for intelligent monitoring of blast furnace tuyere operating conditions based on visual monitoring and deep learning, characterized in that, Includes the following steps: (1) On-site data acquisition: Acquire monitoring video data including tuyer flame, injection port and surrounding furnace wall from the blast furnace monitoring system; and combine on-site operating parameters, including at least one of air volume, air pressure and cooling water flow rate. (2) Video processing: Decode the surveillance video and extract image frames at a preset frame extraction interval of N seconds; (3) Sample construction: Based on the blast furnace tuyere layout and camera calibration information, crop the area image containing the target tuyere from the image frame; adjust the contrast of the cropped image to enhance the details; and remove invalid samples that are severely obscured, abnormally exposed, too dark or too bright. (4) Data labeling: Manually label the valid samples and mark each vent area image as "normal" or "abnormal". Abnormalities include slag hanging, coal shortage, excessive coal injection and abnormal flame shape. (5) Image enhancement: The labeled samples are enhanced, including one or more of blur transformation, brightness adjustment, contrast adjustment, color perturbation, noise superposition and scale change; (6) Model training: The enhanced samples are input into the deep learning target detection model for training to obtain the blast furnace tuyere status detection model. The model is used to output the tuyere location and its operating condition category. (7) Online detection: Input the image obtained by real-time frame extraction into the blast furnace tuyere status detection model. If the output confidence is not lower than the preset threshold, the tuyere detection result is directly output. If the conditions are not met, the candidate ROI detection mechanism will be triggered: (i) Generate a set of candidate ROI regions, which includes a preset region covering the possible location of the wind vent, or dynamically adjusted based on the detection results of the previous frame; (ii) Crop the image for each candidate ROI region and input it into the model for detection, then map the detection results back to the coordinate system of the entire frame image; (iii) Select the detection box with the highest confidence from the mapped candidate results and use it as the final output result of this frame; (8) Results output and analysis: The test results are overlaid and displayed, and are jointly analyzed with the on-site working parameters for air outlet condition assessment.

2. The method according to claim 1, characterized in that, The preset frame-sampling interval N in step (2) is 60 to 120 seconds.

3. The method according to claim 1, characterized in that, The sample construction in step (3) includes: Before generating the sample, the details of the flames at the vent are highlighted by reducing the contrast of the background area. The sampling area is cropped based on the spatial layout of the wind vent and camera calibration information; An effective sample screening mechanism is introduced to remove samples that are severely occluded, have abnormal brightness, or lack details.

4. The method according to claim 1, characterized in that, The data annotation in step (4) is completed using a manual annotation tool. The annotation results are then stored in a one-to-one correspondence with the sample images.

5. The method according to claim 1, characterized in that, The image enhancement in step (5) includes: (1) Blur enhancement: Random blur, simulating heat wave / out of focus / shaking; (2) Brightness / contrast adjustment: Randomly adjust brightness and contrast to simulate exposure fluctuations; (3) Color perturbation: Randomly adjust hue / saturation / brightness to simulate color shift; (4) Scale transformation: random scaling to adapt to changes in target scale; (5) Geometric perturbation: random rotation / translation / shearing to adapt to changes in viewpoint.

6. The method according to claim 1, characterized in that, The blast furnace tuyere status detection model in step (6) is constructed using the YOLOv8 framework to simultaneously output the tuyere detection frame position and its operating condition category.

7. The method according to claim 1, characterized in that, The candidate ROI region set in step (7) is a configurable parameter, and its position and size can be adaptively updated according to historical detection results; Specifically, when the previous frame detects the output detection box... At that time, calculate its center point. , in A local candidate ROI set is generated centered on this point, with its width and height as follows: ,in >1 is the expansion coefficient.

8. The method according to claim 1, characterized in that, The coordinate mapping in step (7) is based on the following formula: Let the coordinates of the top-left corner of the ROI in the entire image frame be ( The coordinates of the detection box in the ROI subplot coordinate system are: Then the coordinates of the detection box mapped to the coordinate system of the entire image frame are: 。 9. A smart monitoring system for blast furnace tuyeres, characterized in that, The method based on claim 1 includes: The video acquisition module is used to acquire the video stream from the blast furnace tuyeres monitoring system. The image processing module is used to decode the video and extract image frames at a preset frame extraction interval of N seconds. The sample processing module is used for trimming air vent areas, quality control, data annotation, and enhancement. The model training module is used to train the blast furnace tuyeres condition detection model. The online monitoring module is used to perform whole-frame detection and ROI detection mechanisms; The results output module is used to overlay and display the test results and to perform joint analysis with the operating parameters.

Images