Intelligent identification method and system for unblocking state of blast furnace tap hole

Abstract

The invention discloses an intelligent identification method and system for the unblocking state of a blast furnace tap hole. The method comprises the following steps: (1) initializing equipment; (2) image acquisition and preprocessing; (3) model reasoning and target classification: inputting the standardized tensor into a lightweight deep learning model deployed on edge computing equipment, and identifying a dynamic slag iron flow area and an iron notch channel area through a multi-scale feature extraction and bidirectional feature fusion structure; (4) judging an iron opening or iron blocking state: based on a continuous multi-frame detection result, jointly judging an iron opening or iron blocking state through a region overlapping rate, an edge change rate and gray level distribution; and (5) outputting and storing data. According to the invention, accurate, real-time and automatic determination of iron notch opening and blocking time can be realized.

Description

Technical Field

[0001] This invention relates to the field of automated monitoring technology for blast furnace ironmaking processes, specifically to an intelligent identification method and system for the opening and closing status of the blast furnace taphole. It is particularly suitable for real-time and automatic identification of whether molten iron flows out during the opening and closing of the blast furnace taphole, and for accurately recording key operation time nodes. Background Technology

[0002] In modern blast furnace ironmaking, opening and closing the taphole are critical processes that determine tapping efficiency, ensure furnace stability, and maintain on-site safety. Current mainstream monitoring methods include: manual visual judgment (reliant on experience, with large errors), infrared thermal imaging (high cost, susceptible to ambient temperature interference, high false alarm rate), and vibration sensor methods (unable to distinguish between tapping and equipment impact, high missed detection rate). Therefore, this stage still heavily relies on the operator's personal experience and visual observation. Workers must operate in harsh environments with high temperatures (over 1500℃), dust, and strong radiation, making subjective judgments based on characteristics such as molten iron color and flame shape. This traditional model has three core problems: First, due to the lack of unified standards, it is easily affected by fatigue, emotions, and environmental factors, leading to inaccurate timing of taphole opening and closing, thus affecting furnace pressure balance and safe production; second, operators face extremely high safety risks, with prolonged exposure to high temperatures increasing the risk of personal injury; and finally, the lack of accurate production data recording makes effective process optimization and management difficult.

[0003] As a crucial indicator of blast furnace operation, tapping time not only directly reflects the state of the hearth, slag and iron characteristics, equipment matching, and the rationality of the smelting rhythm, but also plays a vital role in optimizing the entire ironmaking process. Specifically, tapping time comprehensively characterizes the amount of slag and iron accumulated in the hearth, the fluidity and thermal state of the slag and iron, the working condition of the taphole and its equipment, and the degree of matching between the smelting rhythm and operating parameters.

[0004] To address the aforementioned issues and accelerate the digital and intelligent transformation of the steel industry, particularly aiming to achieve unmanned operation in high-risk processes, the industry is actively seeking an intelligent identification method that can deeply understand operational scenarios and possess strong anti-interference capabilities. This method integrates multi-dimensional information based on visual technology to achieve high-precision, low-latency, and fully automatic identification of the opening and closing states of the taphole in complex industrial environments. This supports unmanned operation and provides a solid data foundation for building a safe and efficient intelligent scheduling system. Although existing research has used auxiliary means such as infrared thermal imaging or vibration sensors, the high cost and weak anti-interference capabilities of these technologies have not yet been effectively resolved. Therefore, developing a more comprehensive and adaptable intelligent solution is urgently needed. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies, such as low safety of manual monitoring and poor adaptability of sensor monitoring, and provides an intelligent identification method and system for the opening and closing status of blast furnace tapholes. Based on visual perception, edge computing, and intelligent algorithms, this method achieves accurate, real-time, and automated determination of the timing of taphole opening and closing. Its core technical means include: (1) The three-in-one hardware architecture of "high temperature resistant camera + water cooling jacket + directional blow" is adopted to ensure the stability of image acquisition in harsh environments; (2) Deploy lightweight deep learning models on edge computing devices to achieve real-time target detection; (3) Innovatively propose a multi-frame joint judgment logic based on process mechanism: determine whether the source is the iron tap by the "spatial overlap rate between the dynamic slag and iron flow area and the iron tap channel area", distinguish between real flow and static reflection by the "edge change rate", eliminate residual heat source interference by the "gray distribution threshold", and combine the temporal constraints of consecutive N frames (iron tap opening) and M frames (iron tap blocking) to achieve high-precision state judgment.

[0006] The technical solution for achieving the objective of this invention is as follows: A method for intelligently identifying the open / closed status of a blast furnace taphole includes the following steps: (1) Equipment initialization: Synchronize the industrial camera with the blast furnace distributed control system (DCS), start the integrated purging device to periodically clean the lens surface, and check the cooling system and field of view status; (2) Image acquisition and preprocessing: The video stream of the iron tap area is acquired by the industrial camera, and the edge computing device performs image decoding, region of interest (ROI) extraction, size standardization and numerical normalization in sequence. Median filtering combined with adaptive histogram equalization (CLAHE) algorithm is used to eliminate smoke and light interference and output standardized image tensors. (3) Model reasoning and target classification: The standardized tensor input is deployed on the lightweight deep learning model of the edge computing device. Through multi-scale feature extraction and bidirectional feature fusion structure, the "dynamic slag and iron flow area" and "iron tap channel area" are identified, and the bounding box information with confidence is output. (4) Determination of opening and blocking status: Based on the detection results of multiple consecutive frames, the opening or blocking status of iron is determined by the combined use of regional overlap rate, edge change rate and gray value distribution: When N consecutive frames meet the condition that "the overlap rate between the dynamic slag and iron flow area and the iron tap channel area is greater than the first threshold, the edge change rate is higher than the second threshold and the average gray value exceeds the third threshold", it is determined to be the opening status; when M consecutive frames meet the condition that "there is no dynamic slag and iron flow area, the integrity of the iron tap channel outline is higher than the fourth threshold and the maximum gray value is lower than the fifth threshold", it is determined to be the blocking status. (5) Data output and storage: Calculate the iron tapping time of a single batch, synchronize the opening and closing time, original images and recognition results to the production management system and store them in the traceable database. The data retention time shall not be less than three months.

[0007] The lightweight deep learning target classification model is based on the YOLOv8-nano architecture with a C2f module, FPN and PAN bidirectional feature fusion structure. Its parameter count is less than 5M and its inference speed is no less than 20fps.

[0008] The edge computing device is an industrial-grade embedded platform with GPU acceleration capabilities, preferably the NVIDIA Jetson Xavier NX platform, with a computing power of ≥21 TOPS, supporting industrial-grade protection (IP65), and a processing latency of ≤80ms.

[0009] The purging device uses dry compressed air for periodic cleaning. The dry compressed air is purified by three-stage filtration (dust content ≤0.1μm), with an air pressure of 0.3–0.7MPa, a purging cycle of 20–40 seconds / cycle, and a single cycle duration of 0.3–1 second. The preferred parameters are air pressure 0.4–0.6MPa, cycle 30 seconds / cycle, and single cycle duration 0.5 seconds.

[0010] The N is 3 and the M is 5; the preferred threshold combination is a first threshold of 75%–85%, a second threshold of 25%–35%, a third threshold of 180–220%, a fourth threshold of 85%–95%, and a fifth threshold of 60–90%.

[0011] A smart identification system for the opening and closing status of a blast furnace taphole includes: The visual perception module includes a camera equipped with a water-cooled jacket (cooling water temperature difference controlled at 5-10℃), which is installed on a protective frame 1.5-2m to the side of the iron outlet, with the lens facing the center of the iron outlet and at an angle of 20°-40° (preferably 30°) to the horizontal plane. The edge computing module includes an industrial-grade embedded processor and storage unit for performing image preprocessing and deep learning model inference; The data output module is used to synchronize the opening and closing time, images and recognition results to the blast furnace production management system via industrial Ethernet or 5G network; A traceable database for long-term storage of original images, annotation results, and timestamp data.

[0012] The high-temperature resistant industrial camera can withstand ambient temperatures of -20℃ to 120℃, can withstand slag and iron radiation temperatures of ≤1500℃, has a resolution of ≥1080P, a frame rate of ≥30fps, and the preferred model is Hikvision MV-CA016-10GM.

[0013] The synchronization delay of the data output module is ≤100ms.

[0014] The method described above is used to achieve unmanned monitoring of blast furnace tapping outlets, reduce the risk of human intervention, and improve the accuracy of production data.

[0015] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method.

[0016] The beneficial effects of this invention are: 1. Significantly improved recording accuracy: The dual logic of "customized optimization of image recognition + continuous flow verification" improves the recording accuracy of the opening and closing of the iron from ±3-10 minutes by manual methods to ±10 seconds, and the statistical error of the iron tapping time is ≤15 seconds, providing accurate data support for furnace condition analysis; 2. Significantly enhanced anti-interference capability: High-temperature resistant camera + directional purging device ensures image acquisition stability, CLAHE preprocessing reduces the accuracy fluctuation caused by light fluctuations from ±15% to ±8%, FPN+PAN feature fusion enables the non-rigid slag and iron target classification recall rate to reach 90%, and the linkage verification mechanism controls the false recognition rate to ≤2%; 3. Excellent automation and real-time performance: No manual intervention is required throughout the process. The latency from image acquisition to data output is ≤100ms. The YOLOv8-nano model has an inference speed of ≥20fps, which is adapted to the field requirements and reduces the labor intensity of furnace workers by more than 80%. 4. Strong compatibility and feasibility: The system can be seamlessly integrated with the existing blast furnace DCS system. The edge computing gateway (NVIDIA Jetson Xavier NX) and model framework (PyTorch) are mature commercial or open source solutions. All parameters have been determined through on-site testing and can be directly implemented by those skilled in the art. 5. Reduced deployment costs: No infrared or vibration sensors are required, reducing the overall system cost. Attached Figure Description

[0017] Figure 1 This is a flowchart of a blast furnace tapping spout image recognition and tapping time recording system.

[0018] Figure 2 A flowchart for identifying and tracing the timing of tapping at the blast furnace taphole. Detailed Implementation

[0019] The following is in conjunction with the appendix Figure 1-2 The technical solution of the present invention is described in detail using multi-scenario verification data to ensure that it can be repeatedly implemented by those skilled in the art.

[0020] (I) Basic Information of the Implementation Example This embodiment uses a 3200m³ blast furnace (No. 1 tapping spout) in a steel plant as the application scenario, and cross-validates it on a 2500m³ blast furnace (No. 2 tapping spout).

[0021] Visual perception module: High-temperature resistant industrial camera (Hikvision MV-CA016-10GM, 1080P resolution, temperature resistance -20℃~120℃, slag and iron radiation resistance ≤1500℃) + integrated blowing device (air pressure 0.4-0.6MPa, three-stage filtered compressed air, dust content ≤0.1μm). Edge computing module: NVIDIA Jetson Xavier NX (21 TOPS computing power, IP65 protection, processing latency ≤80ms); Data output module: Industrial Ethernet + 5G dual-mode transmission (latency ≤85ms); Traceable database: Industrial-grade solid-state drives (read / write speed 550MB / s, data retention ≥3 months).

[0022] (II) Implementation Process 1. Device Initialization Provides the hardware and timing foundation for system operation, matching purging parameters: Time synchronization: The high-temperature resistant industrial camera is synchronized with the blast furnace DCS system (Siemens PCS7) through NTP protocol calibration, with a synchronization error of ≤8ms, ensuring the consistency of timestamps at the opening and closing times. Purge Start-up: Activate the integrated purging device, using dry compressed air (dust content ≤0.1μm) filtered through three stages to clean the lens at a cycle of "0.5MPa air pressure, 0.5 seconds of purging every 30 seconds". The purging nozzle is at a 15° angle to the lens axis (to avoid airflow obstruction or water mist formation). The lens cleanliness retention rate is tested to be ≥95%. Equipment inspection: Manually confirm the water-cooled jacket's sealing (cooling water temperature difference 7℃), ensure the lens's field of view is not obstructed by the furnace body's protective plate (focus on the center of the taphole channel), and ensure the hardware's physical condition is normal.

[0023] 2. Image Acquisition and Preprocessing Image acquisition: The camera acquires JPEG format images at a frame rate of 30fps, with a single frame image size of 1920×1080 pixels (RGB three channels, pixel value 0-255).

[0024] Gateway processing (executed by edge computing devices): ① Decoding: Decode the JPEG image into a 1920×1080 pixel RGB pixel matrix; ②ROI Extraction: Based on the preset calibration position of the taphole channel, an 800×600 pixel rectangular area within the range of "taper channel + surrounding 50cm" is accurately extracted (excluding irrelevant backgrounds such as furnace frame and protective plate), reducing the amount of subsequent data processing by 32%; ③Size standardization: Scale to 640×640 pixels proportionally according to the original aspect ratio. During the scaling process, fill the image edges with gray pixels with RGB values ​​(128,128,128) to avoid distortion of the slag flow texture caused by stretching. ④ Numerical normalization: Convert the pixel integer value to a 32-bit floating-point number, and normalize it to the range of 0.0-1.0 according to the formula "normalized pixel value = original pixel value / 255.0", thereby reducing the computational complexity of the model; ⑤ Tensor reshaping: Convert the image format from "height-width-channel (HWC)" to "channel-height-width (CHW)" required by the YOLOv8-nano model, and add batch processing dimensions to finally form a four-dimensional tensor with shape [1,3,640,640].

[0025] Interference correction: ①Median filtering: Using a 3×3 convolution kernel to filter noise caused by dust adhering to the lens, the image signal-to-noise ratio was improved by 20% after testing, avoiding dust points being misjudged as "slag and iron flow points"; ②CLAHE + Gamma Enhancement: The dynamic range of brightness in the high-temperature and strong light area of ​​slag and iron is compressed by adaptive histogram equalization (CLAHE, clip Limit=2.0), while the edge details of the iron tapping channel are enhanced with Gamma=1.2. Actual tests show that the model recognition accuracy fluctuation is reduced from ±15% to ±3% under the difference between morning and evening light (illuminance difference of 10000 lux).

[0026] 3. Model Inference and Target Classification Target classification based on edge computing devices and lightweight models: Hardware and model deployment: The YOLOv8-nano model (3.2M < 5M parameters) was deployed on an NVIDIA Jetson Xavier NX edge computing device (21 TOPS computing power). The model was trained and optimized under five working conditions (≥10,000 labeled images) of “normal iron tapping, after plugging, splashing, slag formation, and slight iron leakage”.

[0027] Reasoning process: ①Feature extraction: After the tensor is input into the model, it is downsampled layer by layer through the backbone network stacked by C2f modules to generate feature maps of three scales: 80×80 (shallow layer: iron tapping edge, slag and iron texture), 40×40 (middle layer: details + semantics), and 20×20 (deep layer: slag and iron flow semantics). ② Bidirectional feature fusion: High-level semantic information and shallow localization information are fused through the FPN+PAN structure. According to actual tests, the classification recall rate for "initial small flow splash (flow area 50×50 pixels)" and "mid-term large flow (flow area 400×400 pixels)" both reached 98.5%. ③ Classification Head Prediction: Three parallel convolutional classification heads output three types of information: bounding box parameters (center point offset, normalized width and height values), target confidence (probability of target existence), and category probability ("Dynamic Slag and Iron Flow Area" category 0, "Iron Mine Channel Area" category 1).

[0028] Post-processing: ① Confidence threshold filtering: Set a confidence threshold of 0.5 (determined after 100 sets of on-site debugging) to remove false prediction boxes with low confidence (<0.5); ②NMS algorithm: Calculate the comprehensive confidence score according to "target confidence score × category probability", sort the candidate boxes in descending order, and remove duplicate boxes with IoU=0.45 (adapting to the irregular shape of slag and iron flow) as the threshold; ③ Coordinate restoration: The normalized coordinates of the predicted box are converted into the actual coordinates of the original image of 1920×1080 pixels through inverse operation. For example, the coordinates of the bounding box of the "dynamic slag and iron flow area" in the original image are (x1=320, y1=450, x2=680, y2=720), which clarifies the target location.

[0029] 4. Determining the timing of opening the iron dam Based on joint determination of multi-frame features: Iron-starting timing determination: The following three conditions must be met for three consecutive frames: ① The overlap rate between the dynamic slag and iron flow area and the taphole channel area is ≥80% (preferred value of the first threshold), ensuring that the slag and iron flow originates from the taphole channel (not from surrounding iron leakage). ② The edge change rate of the dynamic slag and iron flow area is >30% (preferred value of the second threshold), which distinguishes between "real flow" and "static reflection"; ③The average gray value of the dynamic slag and iron flow area is >200 (the preferred value of the third threshold), eliminating interference from low-temperature dust reflection; Once the above conditions are met, the acquisition time of the third frame image is recorded as the "valid iron-starting time". For example, if the acquisition time of the third frame is 14:25:30.8, then the iron-starting time is 14:25:30.8.

[0030] Determining when the railway is blocked: Five consecutive frames (M=5) must meet the following dual conditions (fourth and fifth thresholds): ① No “dynamic slag and iron flow area” was detected in 5 consecutive frames (the area of ​​the flow area is less than 5%). ② The outline integrity of the iron tapping channel area is ≥90% (preferred value of the fourth threshold), and the maximum gray value in the area is <80 (preferred value of the fifth threshold), so as to avoid the high temperature reflection of residual slag and iron after the tapping is blocked and misjudged as "not blocked"; Once the above conditions are met, the acquisition time of the 5th frame image is recorded as the "effective iron blockage time". For example, if the acquisition time of the 5th frame is 14:58:15.2, then the iron blockage time is 14:58:15.2.

[0031] 5. Data Output and Traceability Complete the synchronization and storage of iron production data.

[0032] Iron tapping time calculation: The iron tapping time of a single batch is obtained by "stopping iron time - stopping iron time", for example: 14:58:15.2-14:25:30.8=32 minutes 44.4 seconds; Dual-mode transmission: The data output module uses industrial Ethernet + 5G dual-mode transmission (latency ≤85ms) to synchronize the opening and closing time and iron tapping time to the blast furnace production management system (such as Rockwell ControlLogix). Traceable storage: The original image (time frame of blast furnace opening / closing) + identification and annotation map (including bounding box) + timestamp data are stored in a traceable database (industrial-grade solid-state drive), classified by "date-blast furnace number-tap number-batch", and the data is retained for 6 months (no less than 3 months).

[0033] (iii) Cross-validation data To verify the stability and superiority of the technical solution, tests were conducted in both a "3200m³ blast furnace (main scenario)" and a "2500m³ blast furnace (cross-scenario)" for comparison.

[0034] (iv) Closed-loop process logic The technical process of this embodiment forms a complete closed loop: visual perception module (acquiring video stream) → edge computing module (preprocessing + inference) → opening / closing determination (output time) → data output module (synchronization + storage). The technical parameters of all links have been verified through cross-validation in multiple scenarios to prove that this solution can operate stably under different blast furnace models and different harsh operating conditions, meeting the needs of unmanned monitoring of blast furnace tapping.

[0035] The embodiments described above can be further combined or replaced, and these embodiments are merely descriptions of preferred embodiments of the present invention, not limitations on the concept and scope of the present invention. Various changes and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the inventive concept are all within the protection scope of the present invention. The protection scope of the present invention is given by the appended claims and any equivalents.

Claims

1. A method for intelligently identifying the open state of a blast furnace tap hole, characterized in that, The method comprises the following steps: (1) Device initialization: synchronize the industrial camera with the blast furnace distributed control system (DCS) in time, start the integrated purging device to periodically clean the lens surface, and check the cooling system and the field of view state; (2) Image acquisition and preprocessing: collect the video stream of the tapping hole area through the industrial camera, sequentially perform image decoding, region of interest (ROI) extraction, size standardization and numerical normalization through the edge computing device, and eliminate smoke and light interference by using median filtering combined with adaptive histogram equalization (CLAHE) algorithm, and output the standardized image tensor; (3) Model inference and target classification: input the standardized tensor into the lightweight deep learning model deployed on the edge computing device, identify the "dynamic slag iron flow area" and "tapping hole channel area" through multi-scale feature extraction and bidirectional feature fusion structure, and output the bounding box information with confidence; (4) Open and block state judgment: based on continuous multiple frame detection results, judge the open iron or block iron state through the area overlap rate, edge change rate and gray distribution: when the continuous N frames satisfy "the overlap rate of dynamic slag iron flow area and tapping hole channel area is greater than the first threshold, the edge change rate is higher than the second threshold, and the average gray value exceeds the third threshold", it is judged as open iron state; when the continuous M frames satisfy "no dynamic slag iron flow area, the integrity of the tapping hole channel contour is higher than the fourth threshold, and the maximum gray value is lower than the fifth threshold", it is judged as block iron state; (5) Data output and storage: calculate the single batch tapping time, synchronize the opening and blocking time, original image and recognition result to the production management system, and store it in the traceable database, the data retention time is not less than three months.

2. The method of claim 1, wherein the lightweight deep learning target classification model is a YOLOv8-nano architecture based on C2f module, FPN and PAN bidirectional feature fusion structure, with a parameter amount less than 5M and an inference speed not less than 20fps.

3. The method of claim 1, wherein the edge computing device is an industrial-grade embedded platform with GPU acceleration capability, with a computing power ≥21TOPS, supporting industrial-grade protection IP65, and a processing delay ≤80ms.

4. The method of claim 1, wherein the purging device uses dry compressed air for periodic cleaning, the dry compressed air is purified through three-stage filtration, with a dust content ≤0.1μm, a gas pressure of 0.3-0.7MPa, a purging period of 20-40 seconds / time, and a single time length of 0.3-1 second; the preferred parameters are gas pressure 0.4-0.6MPa, period 30 seconds / time, and single time length 0.5 second.

5. The method of claim 1, wherein the N is 3 and the M is 5; the preferred threshold combination is first threshold 75%-85%, second threshold 25%-35%, third threshold 180-220, fourth threshold 85%-95%, and fifth threshold 60-90. The method comprises: ​ ​ ​ 6. A high furnace taphole opening state intelligent identification system, characterized in that, ​ A visual perception module, the camera is configured with a water cooling jacket, the cooling water temperature difference is controlled within 5-10℃, installed on the protective frame 1.5-2m away from the taphole side, the lens is directed to the center of the taphole with an angle of 20°-40° (preferably 30°) to the horizontal plane; An edge computing module, including an industrial-grade embedded processor and a storage unit, used for image preprocessing and deep learning model inference; A data output module, used for synchronizing the opening and blocking time, images and recognition results to the blast furnace production management system through industrial Ethernet or 5G network; A traceable database, used for long-term storage of original images, annotation results and timestamp data.

7. The system of claim 6, wherein, The high-temperature-resistant industrial camera can withstand an environmental temperature of -20℃-120℃, a slag-iron radiation temperature of ≤1500℃, a resolution of ≥1080P, and a frame rate of ≥30fps.

8. The system of claim 6, wherein, The synchronization delay of the data output module is ≤100ms.

9. The method or system of any one of claims 1-8, wherein, The method is used for realizing the unmanned monitoring of the blast furnace taphole, reducing the risk of manual intervention and improving the production data accuracy.

10. A computer readable storage medium having a computer program stored thereon, when the program is executed by a processor, the method of any one of claims 1-5 is realized.

Images