A remote unmanned control method for slagging of a ladle

Abstract

This invention provides a remote, unmanned method for slag removal from a molten iron ladle, relating to the field of slag removal and detection technology. This method uses dynamic visual information collected during the pouring process to infer the possible distribution area of ​​the slag from its source, generating an initial distribution. After the molten iron is poured in, the slag is actually distributed on the liquid surface. At this point, a first image is collected to correct the deviation of the initial distribution, ensuring the inferred result matches the actual distribution, thus obtaining the distribution to be checked and planning the slag removal route. During the slag removal process, as the slag is gradually removed, the liquid surface state continuously changes. A second image is collected in real time to monitor the slag layer coverage, and force feedback information is used to sense the force between the slag removal tool and the slag layer. Both factors are combined to judge the slag removal status and adjust the subsequent route accordingly until slag removal is complete. This invention achieves accurate acquisition of slag distribution through multi-stage progressive sensing and real-time correction, ensuring the effectiveness of unmanned intelligent slag removal.

Description

Technical Field

[0001] This invention relates to the field of slag removal and detection technology for molten iron ladles, and in particular to a remote, unmanned method for slag removal from molten iron ladles. Background Technology

[0002] Slag removal from molten iron ladles is a crucial process affecting steel quality in steelmaking. Traditional manual slag removal relies on workers' visual inspection and experience, operating in high-temperature, dusty environments using tools like shovels. This method is labor-intensive, carries high safety risks, and produces inconsistent slag removal quality. To address these issues, the metallurgical industry has gradually introduced mechanized equipment such as tilting trolleys and slag removers, further integrating industrial cameras and liquid level radar sensors to develop automated slag removal systems, improving operational consistency. With the development of intelligent control technology, slag removal from molten iron ladles is rapidly evolving towards unmanned and intelligent processes.

[0003] However, existing unmanned slag removal systems are essentially open-loop or semi-open-loop controls. Their core flaw lies in their inability to accurately obtain the true distribution of the slag layer within the molten iron ladle. This is because the flow and impact of the molten iron during the pouring process constantly alters the migration path and aggregation location of the slag. Existing solutions only generate the path once based on static images or preset rules before slag removal. However, since the slag is suspended on the surface of the molten iron, its shape and volume below the surface cannot be determined. As a result, the slag distribution cannot be accurately predicted, leading to low slag removal efficiency and poor performance. Summary of the Invention

[0004] This invention provides a remote, unmanned method for cleaning slag from a molten iron ladle, which solves the problem that the distribution of slag cannot be accurately predicted and the cleaning efficiency is poor.

[0005] This invention provides a remote, unmanned method for cleaning slag from a molten iron ladle, the method comprising: During the process of pouring molten iron from the upstream container into the ladle, dynamic visual information of the molten iron during the pouring process is collected, and the initial distribution of slag in the ladle is predicted based on the dynamic visual information. After the molten iron is poured in, a first image of the liquid surface of the molten iron ladle is acquired; based on the first image information, the initial distribution is corrected for deviation to obtain the distribution to be checked; An initial slag removal route is generated based on the distribution to be checked; The slag cleaning machine is controlled to perform slag cleaning operations according to the initial slag cleaning route. During the slag cleaning process, the second image information of the liquid surface of the molten iron ladle and the force feedback information of the slag cleaning tools in the slag cleaning machine are collected in real time. The initial slag cleaning route is corrected in real time according to the second image information and the force feedback information until the slag cleaning is completed.

[0006] In one possible implementation, the acquisition of dynamic visual information during the pouring process of molten iron, and the prediction of the initial distribution of slag within the ladle based on the dynamic visual information, includes: The system continuously acquires time-series image sequences and flow rate information during the pouring process of molten iron, and generates the dynamic visual information based on the time-series image sequences and the flow rate information. Based on the time sequence image, the timing changes of the landing point when molten iron falls into the ladle are identified, and the flow and diffusion path of molten iron in the ladle is calculated based on the timing changes of the landing point and the flow rate information. The slag's buoyancy and aggregation pattern is obtained. Based on the buoyancy and aggregation pattern and the flow diffusion path, the aggregation probability distribution of slag in each region of the molten iron ladle is predicted, and the aggregation probability distribution is used as the initial distribution.

[0007] In one possible implementation, the prediction of the slag aggregation probability distribution in different regions within the molten iron ladle further includes: The fluctuation characteristics of the flow rate information during the pouring of molten iron were obtained; The distribution of turbulence intensity on the surface of molten iron is identified based on the aforementioned wave characteristics; In regions where the turbulence intensity distribution is higher than a first threshold, the aggregation probability distribution in the corresponding region is reduced; in regions where the turbulence intensity distribution is lower than a second threshold, the aggregation probability distribution in the corresponding region is increased.

[0008] In one possible implementation, the step of correcting the initial distribution based on the first image information to obtain the distribution to be checked includes: The area covered by the initial distribution is divided into multiple grid cells, and each grid cell is assigned an initial confidence value for the amount of slag. Identify the slag-iron boundary based on the first image information; The initial value of the slag quantity confidence is adjusted according to the position of each grid cell relative to the slag-iron boundary; Based on the adjusted initial confidence value of the slag quantity, the distribution to be verified is obtained.

[0009] In one possible implementation, adjusting the initial value of the slag quantity confidence based on the position of each grid cell relative to the slag-iron boundary includes: Determine whether each of the grid cells is located on one side of the slag region or the molten iron region of the slag-iron boundary; For the grid cell located on one side of the slag region, increase the initial value of the corresponding slag quantity confidence. For the grid cell located on one side of the molten iron region, the initial confidence value of the corresponding slag quantity is lowered.

[0010] In one possible implementation, the force feedback information includes the thrust and lateral force of the slag-cleaning tool in the slag-cleaning propulsion direction; the real-time correction of the initial slag-cleaning path based on the second image information and the force feedback information includes: The slag layer coverage status at the current slag removal location is identified based on the second image information; The adhesion strength of the slag layer in the current slag removal tool area is identified based on the thrust and the lateral force. By integrating and analyzing the slag layer coverage state and the slag layer bonding strength, an evaluation result of the slag layer state in the current slag cleaning tool area is obtained; The initial slag removal route is adjusted in real time based on the evaluation results.

[0011] In one possible implementation, identifying the slag layer adhesion strength in the current slag removal tool area based on the thrust and the lateral force includes: The time-domain variation of the thrust during the process of the slag removal tool cutting into the slag layer was obtained; Based on the aforementioned time-domain variation, the feature value of bond strength is extracted; The fluctuation characteristics of the lateral force are obtained, and the bonding strength characteristic value is corrected according to the fluctuation characteristics to obtain the bonding strength and bonding grade of the slag layer.

[0012] In one possible implementation, the integrated analysis of the slag layer coverage state and the slag layer adhesion strength to obtain an evaluation result of the slag layer state in the current slag cleaning tool area includes: The slag layer coverage area is determined based on the second image information; The slag layer covered area is discretized into multiple evaluation sub-regions; The bonding level of each evaluation sub-region is determined based on the force feedback information; The adhesion levels of each of the evaluation sub-regions are superimposed onto the corresponding positions of the covered area to form the evaluation results with spatial distribution.

[0013] In one possible implementation, the real-time adjustment of the initial slag removal route based on the evaluation results includes: The areas in the evaluation results where the adhesion level is higher than a preset threshold are marked as priority processing areas; The initial slag removal route is biased toward the priority treatment area; The areas where the slag layer has been removed in the assessment results will be excluded from the subsequent slag removal route.

[0014] In one possible implementation, the step of "until the slag removal is complete" includes: When the second image information shows that the slag coverage of the molten iron ladle is lower than a preset threshold, and the force feedback information shows that the slag removal resistance of the entire molten iron ladle surface is lower than a preset threshold, the slag removal is determined to be complete.

[0015] This invention provides a remote, unmanned method for slag removal from a molten iron ladle. During the detection process, dynamic visual information is collected during the pouring process to infer the possible distribution area of ​​the slag from its source, generating an initial distribution. After the molten iron is poured in, the slag is actually distributed on the liquid surface. At this point, a first image is collected to correct the deviation of the initial distribution, ensuring the inferred result matches the actual distribution, thus obtaining the distribution to be checked and planning the slag removal route. During the slag removal process, as the slag is gradually removed, the liquid surface state continuously changes. A second image is collected in real time to monitor the slag layer coverage, and force feedback information is used to sense the force between the slag removal tool and the slag layer. Both factors are combined to judge the slag removal status and adjust the subsequent route accordingly until slag removal is complete. This invention achieves accurate acquisition of slag distribution through multi-stage progressive sensing and real-time correction, ensuring the effectiveness of unmanned intelligent slag removal. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic flowchart of a remote, unmanned slag removal method for molten iron ladles provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, this embodiment of the invention provides a remote, unmanned method for cleaning slag from a molten iron ladle. The method includes steps S101-S104.

[0020] S101. During the process of pouring molten iron from the upstream container into the ladle, collect dynamic visual information of the molten iron during the pouring process, and predict the initial distribution of slag in the ladle based on the dynamic visual information.

[0021] Specifically, during the process of pouring molten iron from an upstream container (such as the blast furnace taphole or blast furnace) into the ladle, due to the impact of pouring and surface fluctuations, the slag (mainly silicates, oxides, and other slag with a lower density than molten iron) will migrate, float, and gradually accumulate on the surface along with the flow of molten iron. The core function of step S101 is to predict the final distribution trend of the slag in advance, before the molten iron has been completely poured and the surface is still in a dynamic change stage, thereby gaining a time window for subsequent slag cleaning route planning. This prediction does not rely on static surface images, but utilizes dynamic visual information available within the "pouring process" time window (including the shape of the molten iron stream, the landing point, and the direction of surface ripple diffusion, etc.), combined with the physical laws of jet impact and surface expansion in fluid mechanics, to deduce the filling sequence and flow path of the molten iron in the ladle, and then, based on the inherent characteristics of slag floating and accumulating, to calculate the area where the slag is most likely to accumulate after pouring. This initial distribution, expressed in the form of a probability density map or a confidence heatmap, provides a priori benchmark for subsequent deviation correction. Its accuracy directly affects the economy and safety of the subsequent slag removal route.

[0022] As one possible implementation, step S101 can be specifically implemented as steps S1011-S1013.

[0023] S1011. Continuously acquire time-series image sequences and flow information of molten iron during the pouring process, and generate dynamic visual information based on the time-series image sequences and flow information.

[0024] Specifically, the time-series image sequence is continuously acquired by a radiation-proof industrial camera installed above the molten iron ladle at a fixed frame rate (e.g., 25fps or 30fps), recording the changes in the molten iron flow and the liquid level within the ladle from the start to the end of the pouring process. Flow rate information is calculated in real time using a tilt angle sensor of the upstream container and the cross-sectional area of ​​the taphole, or obtained by measuring the velocity at the taphole with a laser Doppler velocimeter and then converting it to the cross-sectional area. The timestamp of each frame is synchronized with the instantaneous flow rate value at the corresponding moment, forming a "time-flow-image" triplet sequence. The dynamic visual information is composed of this triplet sequence, which not only contains spatial image texture information but also the flow rate calibration value corresponding to each moment, providing a quantitative basis for subsequent calculations of the accumulated volume and diffusion range of molten iron within the ladle. This information provides a time-consistent data foundation for the flow diffusion path calculation in the subsequent step S1012.

[0025] S1012. Identify the timing changes of the landing point when molten iron falls into the ladle based on the timing image sequence, and calculate the flow and diffusion path of molten iron in the ladle based on the timing changes of the landing point and the flow rate information.

[0026] Specifically, in the time-series image sequence, the location where the molten iron stream impacts the bottom of the ladle or the existing molten iron surface is called the landing point. Because the upstream container's tilt angle continuously changes during the pouring process, the landing point is not fixed but moves along a certain direction from the ladle bottom. By tracking the pixel coordinates of the intersection point between the stream's end and the liquid surface in the image sequence, the trajectory of the landing point can be extracted. Furthermore, using this trajectory as the spatial boundary condition for the molten iron entering the ladle, combined with the input velocity provided by the flow rate information, numerical simulations of flow diffusion are performed on the two-dimensional plane at the ladle bottom using the shallow water wave equation or the finite volume method. This simulates the entire process of molten iron spreading from the landing point outwards, reflecting off the ladle walls, and superimposing onto the surface, ultimately obtaining the temporal sequence of molten iron filling in different areas of the ladle and the curves showing the change in the height of the rising liquid surface. This flow diffusion path is actually a dynamic flow field distribution sequence with time as the axis; it characterizes the "filling footprint" of the molten iron within the ladle and is a key intermediate variable for subsequent prediction of slag aggregation and distribution.

[0027] S1013. Obtain the buoyancy and aggregation pattern of slag. Based on the buoyancy and aggregation pattern and the flow diffusion path, predict the aggregation probability distribution of slag in each region of the molten iron ladle, and use the aggregation probability distribution as the initial distribution.

[0028] Specifically, the slag buoyancy and aggregation law refers to the following: Under static or low-turbulence conditions of molten iron, slag droplets with a density lower than that of molten iron rise to the surface at a certain speed due to buoyancy. Under the combined influence of surface tension and slag viscosity, they tend to merge and aggregate towards areas where the flow stagnates at the surface, areas with a large velocity gradient near the ladle wall, or previously accumulated slag layers. Based on this law, the flow diffusion path calculated in step S1012 is coupled with the slag droplet buoyancy model: For each moment of newly entered molten iron, the time it takes for the slag droplets to rise from their current position to the surface and the horizontal drift distance after rising are calculated. By summing the distribution of the buoyancy endpoints at all moments, the cumulative probability density of slag in each region of the surface can be obtained. This aggregation probability distribution is output in the form of a two-dimensional matrix. Each element in the matrix represents the probability that slag exists in the corresponding region when the pouring is completed. A higher probability value indicates that the region is more likely to become a major slag enrichment area. This distribution is the initial distribution, serving as a rough but a priori reasonable starting point for subsequent fine-tuning.

[0029] As one possible implementation, step S1013 can be specifically implemented as steps S10131-S10133.

[0030] S10131. Obtain the fluctuation characteristics of the flow rate information of molten iron during the pouring process.

[0031] Specifically, the fluctuation characteristics of flow information refer to the statistical properties of the instantaneous flow rate changing over time, including parameters such as mean, variance, peak frequency, and fluctuation amplitude. In actual pouring, due to potential blockages at the upstream container's taphole, limited tilt angle control precision, or sloshing of the molten iron surface, the instantaneous flow rate is not constant but exhibits varying degrees of pulsation. By performing time-series differential processing on the flow information obtained in step S1011, calculating the absolute value of the flow rate change rate between adjacent moments, and conducting power spectral density analysis on the flow sequence over the entire pouring duration, the amplitude spectrum and dominant frequency of the fluctuation can be extracted. This fluctuation characteristic macroscopically reflects the impact intensity of the molten iron input on the ladle's surface, providing excitation input for identifying the turbulence intensity distribution in step S10132.

[0032] S10132. Identify the turbulence intensity distribution on the surface of molten iron based on wave characteristics.

[0033] Specifically, the flow fluctuation characteristics extracted in step S10131 are used as the excitation source and input into the flow diffusion model established in step S1012. In the model, a disturbance component with the same frequency and amplitude as the fluctuation characteristics is superimposed on the injection velocity at the impact point. Then, the velocity field and vorticity field of the liquid surface within the bag are solved again. By calculating the ratio of the root mean square value of the velocity fluctuations to the average flow velocity within each grid cell, the dimensionless turbulence intensity of each cell is obtained. The turbulence intensity distribution reflects the degree of disorder in different regions of the liquid surface within the bag due to impact disturbances: the turbulence intensity is higher near the impact point and along its main diffusion path, while the turbulence intensity is lower at the corners of the bag wall and in flow stagnation areas. This distribution is an important environmental parameter for subsequent adjustment of the slag aggregation probability.

[0034] S10133. In regions where the turbulence intensity distribution is higher than the first threshold, reduce the aggregation probability distribution of the corresponding region; in regions where the turbulence intensity distribution is lower than the second threshold, increase the aggregation probability distribution of the corresponding region.

[0035] Specifically, turbulence intensity has a dual effect on slag aggregation: In high turbulence intensity regions (above the first threshold, e.g., turbulence intensity > 0.3), strong velocity fluctuations and vortex motions hinder slag droplet uplift and cause slag layers that have already risen to the liquid surface to be re-dispersed or pushed to other regions. Therefore, the probability of slag aggregation in these regions should be reduced. In low turbulence intensity regions (below the second threshold, e.g., turbulence intensity < 0.05), the flow tends to be laminar or quasi-static, the slag droplet uplift path is undisturbed, and the slag layer on the liquid surface is easily kept stable. Therefore, the aggregation probability should be enhanced. For the transition region where the turbulence intensity is between the two thresholds, the probability values ​​output by the basic model in step S1013 are kept unchanged. The adjustment operation is as follows: multiply the elements in the initial aggregation probability distribution matrix corresponding to the high turbulence region by a reduction factor (e.g., 0.3~0.6), and multiply the elements in the low turbulence region by an enhancement factor (e.g., 1.5~2.0). The coefficient values ​​are inversely or directly proportional to the turbulence intensity value. After adjustment, normalization is performed to ensure that the sum of probabilities across the entire region remains consistent, resulting in a corrected initial distribution that better reflects the actual physical process.

[0036] S102. After the molten iron is poured in, the first image information of the liquid surface of the molten iron ladle is acquired; based on the first image information, the initial distribution is corrected for deviation to obtain the distribution to be checked.

[0037] Specifically, after the molten iron is poured in, the liquid surface is no longer subjected to the impact of the pouring, and the distribution of slag on the liquid surface enters a relatively stable state. At this time, a static liquid surface image is acquired using an industrial camera, which is the first image information. This image directly presents the actual shape and coverage area of ​​the slag layer on the liquid surface. However, due to liquid surface reflection, uneven slag layer thickness, and viewing angle projection effects, it is difficult to accurately determine the slag layer thickness or amount of slag by relying solely on the image. Therefore, it is necessary to combine the initial distribution predicted in step S101 for fusion correction. The core logic of the correction is as follows: the initial distribution provides a spatial probability prior of the slag amount, and the first image information provides the positional observation of the slag layer boundary. The two are fused through a Bayesian update or confidence adjustment strategy—strengthening the high-probability areas in the initial distribution and correcting the low-probability areas within the slag area identified by the image, and reducing the probability of the initial distribution at the corresponding position in the non-slag area (i.e., the exposed molten iron area) identified by the image. The corrected distribution to be checked retains the physical prior based on the pouring dynamics process and integrates the real boundary information provided by the static image, and its reliability is significantly improved compared to the initial distribution.

[0038] As one possible implementation, step S102 can be specifically implemented as steps S1021-S1023.

[0039] S1021. Divide the area covered by the initial distribution into multiple grid cells, and assign an initial value of the slag quantity confidence level to each grid cell.

[0040] Specifically, the two-dimensional planar region corresponding to the molten iron ladle surface is discretized into multiple grid cells with a fixed resolution (e.g., one grid per centimeter, or N×N uniform grids based on the ladle opening diameter). For each grid cell, its corresponding probability value is read from the initial aggregation probability distribution output in step S101, and this probability value is used as the initial confidence value for "the presence of slag" within that cell. The initial confidence value ranges from [0,1], where a value closer to 1 indicates that the cell is more likely to be covered by slag, and a value closer to 0 indicates that it is more likely to be an exposed molten iron area. This discretization operation converts the continuous probability distribution field into a confidence array of a finite number of discrete points, facilitating grid-by-grid comparison and adjustment with image pixel data in subsequent steps.

[0041] S1022. Identify the slag and iron boundary based on the first image information.

[0042] Specifically, the acquired first image information undergoes preprocessing, including denoising, grayscale transformation, and contrast enhancement, to highlight the visual difference between the slag area (typically dark or brown with a rough texture) and the exposed molten iron area (bright white or bright yellow with strong specular reflection). Subsequently, a threshold segmentation algorithm (such as Otsu adaptive thresholding) or a deep learning-based semantic segmentation model (such as U-Net) is used to binarize the image, classifying pixels as either "slag" or "molten iron." Based on the binarized image, the classification boundary pixel chain is extracted, and a smooth, continuous slag-iron boundary curve is obtained through polygon fitting or spline interpolation. This boundary curve divides the liquid surface plane into two non-overlapping subspaces: the slag area and the molten iron area, serving as the spatial reference for confidence adjustment in step S1023.

[0043] S1023. Adjust the initial value of the slag quantity confidence based on the position of each grid cell relative to the slag-iron boundary.

[0044] Specifically, the discretized grid cells are spatially overlaid with the slag-iron boundary curve extracted in step S1022: For each grid cell, the positional relationship between its center point or coverage area and the boundary curve is determined—if it is located within the slag area, it indicates that the cell contains slag, supported by image evidence, and its initial confidence value should be increased; if it is located within the molten iron area, it indicates that the cell is initially judged to be a slag-free area, and its initial confidence value should be decreased; if it intersects with the boundary curve, interpolation adjustment is performed according to the proportion of the intersection area. The adjustment range is weighted according to the distance from the center of the grid cell to the boundary curve, with smaller adjustment ranges for cells closer to the boundary (preserving prior uncertainty) and larger adjustment ranges for cells farther from the boundary (more sufficient image evidence). Through this positional relationship-driven adjustment, the confidence distribution of the discrete grid tends to be consistent with the spatial distribution of slag and iron observed in the image.

[0045] As a possible implementation, step S1023 can be specifically implemented as steps S10231 to S10233.

[0046] S10231: Determine whether each grid cell is located on the slag region side or the molten iron region side of the slag-iron boundary.

[0047] Specifically, the ray casting method or the winding number method is used to determine the position of the center point of each grid cell relative to the polygonal region: if the center point is inside the slag region polygon enclosed by the slag-iron boundary curve, it is determined that the cell is located on the slag region side; if it is outside the polygon, it is determined that the cell is located on the molten iron region side. For grid cells intersecting the boundary curve, further calculate the ratio r (0<r<1) of the area of the cell falling within the slag region, and use this ratio as a continuous quantitative indicator for position determination instead of a simple binary classification. This determination result is used as the basic category label for the subsequent adjustment amplitude calculation.

[0048] S10232: For grid cells located on the slag region side, increase the corresponding initial slag amount confidence value.

[0049] Specifically, for grid cells determined to be on the slag region side (or intersecting cells with an area ratio r>0.5), a positive adjustment is applied to their initial slag amount confidence value. The adjustment formula is: new confidence value = initial confidence value + α×(1-initial confidence value), where α is the adjustment step size, and its value is negatively correlated with the distance from the center of the grid cell to the boundary curve: the closer the distance, the smaller α is (e.g., 0.1~0.2); the farther the distance, the larger α is (e.g., 0.5~0.7), which reflects the increasing reliability of image evidence as it moves away from the boundary. Meanwhile, an upper limit of 0.95 is set to prevent overfitting. After this adjustment, the confidence of cells in the slag region is significantly improved, which is consistent with image observation.

[0050] S10233: For grid cells located on the molten iron region side, decrease the corresponding initial slag amount confidence value.

[0051] Specifically, for grid cells determined to be on the molten iron region side (or intersecting cells with an area ratio r<0.5), a negative adjustment is applied to their initial slag amount confidence value. The adjustment formula is: new confidence value = initial confidence value - β×initial confidence value, where β is the attenuation coefficient, which is also negatively correlated with the distance from the cell center to the boundary curve: the closer the distance, the smaller β is (e.g., 0.1~0.2); the farther the distance, the larger β is (e.g., 0.6~0.8). A lower limit of 0.05 is set to avoid complete zeroing to retain uncertainty. This adjustment greatly reduces the confidence in the molten iron region, and effectively eliminates false alarm regions that may exist in the initial distribution.

[0052] S1024: Based on the adjusted initial slag amount confidence values, obtain the distribution to be checked.

[0053] Specifically, the new confidence values ​​of slag quantity for all grid cells adjusted in step S1023 are recombined into a two-dimensional confidence matrix according to their original grid spatial locations. Gaussian filtering or median filtering is then used to smooth the matrix, eliminating local outliers caused by grid discretization and boundary jumps, ensuring spatial continuity and a reasonable transition in the distribution field. The smoothed confidence matrix is ​​the distribution to be verified. This distribution integrates dynamic priors (initial distribution) and static image observations (slag-iron boundary), exhibiting higher accuracy and robustness than a single information source, providing a precise spatial slag quantity estimate for subsequent slag removal route generation.

[0054] S103. Generate an initial slag removal route based on the distribution to be checked.

[0055] One embodiment of this application is that the force feedback information includes the thrust and lateral force of the slag removal tool in the slag removal propulsion direction.

[0056] Specifically, the force feedback information is collected in real time by a six-dimensional force / torque sensor installed at the connection between the slag cleaning machine's robotic arm and the slag cleaning tool. The thrust refers to the force component along the direction of the slag cleaning tool's advance (i.e., the direction in which the tool cuts into the slag layer and pushes the slag outwards). Its magnitude directly reflects the resistance of the slag layer to the tool's advancement and is positively correlated with the slag layer thickness and adhesion strength. The lateral force refers to the force component perpendicular to the advance direction and within the liquid surface plane. Its magnitude and direction changes reflect whether the tool encounters asymmetric resistance such as uneven slag layer thickness, local hard lumps, or interference from the wall during advancement, and are important indicators for judging the spatial heterogeneity of the slag layer. Both the thrust and lateral forces are time-varying signals, with a sampling frequency typically not lower than 100Hz to ensure the sensitivity of the real-time response.

[0057] As one possible implementation, step S104 can be specifically implemented as steps S1041-S1044.

[0058] S1041. Identify the slag layer coverage status at the current slag removal location based on the second image information.

[0059] Specifically, the second image information consists of real-time images of the liquid surface continuously acquired during the slag removal operation. The content changes over time, reflecting the dynamic process of the slag layer being gradually removed. For the current slag removal location (i.e., the real-time projection area of ​​the slag removal tool on the liquid surface and its surrounding neighborhood), a sub-image of the corresponding region is extracted from the second image. Using an image segmentation method similar to step S1022, the proportion of slag pixels within the sub-image is identified. Simultaneously, the slag layer boundary changes between the current frame and the previous frame are analyzed using the inter-frame difference method to extract the newly added area of ​​the exposed molten iron region after the slag layer is removed. Combining the above information, a description of the slag layer coverage status at the current slag removal location is formed, including: whether the slag layer fully covers the location, whether it partially covers it, the shape of the slag layer edge, and the removal progress relative to the previous moment. This status description provides a spatial matching benchmark for subsequent adhesion strength estimation.

[0060] S1042. Identify the slag layer adhesion strength in the current slag removal tool area based on thrust and lateral force.

[0061] Specifically, when the slag removal tool advances to a certain position, the resistance that the tool needs to overcome to cut into the slag layer is mainly determined by the adhesion strength of the slag layer. The higher the adhesion strength, the stronger the adhesion between the slag and the molten iron surface or ladle wall, the greater the peak value of the required thrust, and the greater the irregular pulsation of the lateral force. From the force signal time window corresponding to the current slag removal position determined in step S1041, characteristic parameters such as the slope and peak value of the rising segment of the thrust curve, as well as the root mean square and peak-to-peak value of the lateral force, are extracted. By inputting these characteristic parameters into a pre-calibrated adhesion strength mapping model (which is obtained through laboratory testing or regression of historical slag removal data), the force signal characteristics are converted into a quantitative estimate of the slag layer adhesion strength. This adhesion strength value reflects the ease or difficulty of slag removal and is a key basis for subsequent adjustments to the slag removal strategy.

[0062] As one possible implementation, step S1042 can be specifically implemented as steps S10421-S10423.

[0063] S10421. Obtain the time-domain variation of thrust during the process of the slag removal tool cutting into the slag layer; Specifically, as the cleaning tool advances towards the slag layer from a position away from it, the thrust signal undergoes a complete process of "no-load—contact—cutting in—penetration—detachment." The thrust time-domain signal from the "contact" to "penetration" stages typically exhibits a three-stage characteristic: a sharp initial rise (the tool's leading edge squeezes the slag layer), a peak value (the slag layer is disrupted and begins to slide), and a slight decrease followed by stabilization (the slag layer is pushed aside and slides along the tool surface). This time-domain signal is saved as its original waveform, and the corresponding time axis is recorded simultaneously, providing complete data input for feature value extraction in subsequent steps.

[0064] S10422. Extracting bond strength feature values ​​based on time-domain variation.

[0065] Specifically, the thrust time-domain signal obtained in step S10421 undergoes feature engineering processing: First, the maximum slope of the thrust rise segment is calculated, reflecting the initial shear stiffness of the slag layer; second, the thrust peak value F_max is extracted, reflecting the ultimate failure load of the slag layer; third, the time T_peak of the thrust peak occurrence is calculated, reflecting the speed of the failure response; finally, the mean thrust F_plateau during the penetration stage is calculated, reflecting the dynamic sliding resistance of the slag layer. These four features are then fused using a multiple linear regression or neural network model to obtain the bond strength feature value σ_bond, with units of N / mm² equivalent. This feature value comprehensively reflects the mechanical response characteristics of the slag layer under both static and dynamic conditions, and is more representative than a single peak value.

[0066] S10423. Obtain the fluctuation characteristics of the lateral force, and correct the characteristic value of the bonding strength according to the fluctuation characteristics to obtain the bonding strength and bonding grade of the slag layer.

[0067] Specifically, the fluctuation characteristics of lateral force include: the standard deviation σ_F_lat of the lateral force signal during propulsion, the maximum transient offset amplitude A_lat, and the fluctuation frequency flt. If the lateral force fluctuation is severe (σ_F_lat and A_lat are large), it indicates the presence of hard inclusions or severe thickness unevenness within the slag layer. In this case, the bond strength characteristic value extracted solely from the thrust may be too high or too low, requiring correction. The correction rule is as follows: the lateral force fluctuation characteristics are used as a penalty factor or gain factor to perform weighted correction on σ_bond obtained in step S10422—in the case of high fluctuation, the weighted average of σ_bond and the root mean square of the lateral force is taken as the final bond strength to balance the different mechanical information reflected by the thrust and lateral force. The corrected bond strength value is further mapped to four bond levels (light, medium, heavy, and super heavy) according to preset grading thresholds (e.g., <0.5, 0.5~1.0, 1.0~2.0, >2.0) for subsequent path planning.

[0068] S1043. Integrate the analysis of slag layer coverage and slag layer bonding strength to obtain the evaluation results of the slag layer status in the current slag cleaning tool area.

[0069] Specifically, slag cover status reflects the spatial distribution of slag's "presence" and "amount," while adhesion strength reflects the mechanical property of slag's "difficulty in removal." Both describe slag characteristics from visual and mechanical perspectives, respectively, but neither alone is sufficient to comprehensively evaluate the complexities of slag removal operations. The integrated analysis employs a decision-level fusion strategy: cross-combining slag thickness / area information with adhesion level in the cover status—for example, a light adhesion level in a fully covered area is assessed as "easy to remove large areas of floating slag"; a heavy adhesion level in a partially covered area is assessed as "difficult to remove residual hard slag"; if the cover status shows newly exposed molten iron areas but the adhesion strength is abnormally increased, it may indicate that the tool has touched the ladle wall or infiltrated the molten iron. The evaluation results are output in a structured data format, including location coordinates, cover status labels, adhesion level, and a comprehensive recommended operation type (normal advance, decelerated advance, multiple reciprocating movements, detour avoidance, etc.), providing direct decision-making basis for route correction.

[0070] As one possible implementation, step S1043 can be specifically implemented as steps S10431-S10434.

[0071] S10431. Determine the slag layer coverage area based on the second image information.

[0072] Specifically, the second image information undergoes full-image semantic segmentation to extract all pixel sets labeled "slag body," and the pixel sets are divided into several unconnected slag layer coverage regions through connected component analysis. Each coverage region is represented by its contour point sequence and minimum bounding rectangle, and the area and coverage density (slag pixel ratio) of the region are calculated simultaneously. The geometric boundary of this coverage region differs from the slag-iron boundary in step S1022 in that it reflects the distribution of residual slag layers that have not yet been removed at the "current moment." As the slag removal operation progresses, this region will gradually shrink and split into multiple scattered areas.

[0073] S10432. Discretize the slag layer covered area into multiple evaluation sub-regions.

[0074] Specifically, each coverage area determined in step S10431 is further divided into smaller evaluation sub-areas according to a fixed step size (e.g., 5cm × 5cm), with each sub-area serving as an independent evaluation unit. For large, continuous coverage areas, a uniform grid is used; for scattered or elongated areas, adaptive segmentation is performed along their length to ensure that the area of ​​each sub-area is roughly balanced and facilitates single-action coverage by the cleaning tool. After division, each sub-area is assigned a unique identifier (ID), and its geometric center coordinates and occupied area are recorded to form a spatial index table of evaluation sub-areas, providing a refined spatial carrier for subsequent bonding grade superposition.

[0075] S10433. Determine the bonding grade of each evaluation sub-region based on the stress feedback information; Specifically, the cleaning tool passes through multiple evaluation sub-zones during its advancement, and the signals collected by the force sensors correspond to the spatial position of the tool at different times. Using the kinematic model of the cleaning machine (provided by a joint encoder or GPS / IMU positioning system to indicate the real-time position of the tool's end effector), each force signal sampling point is mapped to its corresponding evaluation sub-zone ID. For all force signal sampling points falling within the same sub-zone, the effective adhesion strength of that sub-zone is calculated and mapped to an adhesion grade according to steps S10421-S10423. If a sub-zone has not yet been directly contacted by the tool, its adhesion grade is temporarily marked as "unknown," awaiting subsequent cleaning actions to cover it before being filled. This step achieves precise spatial registration of the mechanical sensing information.

[0076] S10434. The adhesion levels of each evaluation sub-region are superimposed onto the corresponding positions of the covered area to form an evaluation result with spatial distribution.

[0077] Specifically, the adhesion levels of each evaluation sub-region obtained in step S10433 are used as attribute values ​​and assigned back to the sub-region spatial index established in step S10432, forming a triplet data of "location-coverage status-adhesion level" for each sub-region. All sub-region triplet data are organized into a two-dimensional attribute map according to spatial adjacency relationships; this map represents the spatially distributed evaluation results. These evaluation results are displayed on the user interface or as internal data within the algorithm in the form of a heatmap overlay, where different colors or grayscale levels represent different adhesion levels, and the transparency is adjusted according to the coverage status (cleared areas are completely transparent). This spatially distributed evaluation result intuitively reflects the spatial distribution of the current residual slag layer removal difficulty, providing a refined, gridded decision map for real-time route adjustments in step S1044.

[0078] S1044. Adjust the initial slag removal route in real time based on the evaluation results.

[0079] Specifically, the spatial distribution evaluation results obtained in step S1043 are overlaid and compared with the initial slag removal route generated in step S103. For sub-regions with a adhesion level of "heavy" or "overheavy" in the evaluation results, if the initial route only allows a single pass, it is adjusted to multiple passes or the advancing speed is reduced to increase the removal force; for sub-regions marked as "slag layer cleared" in the evaluation results, they are removed from the remaining routes to avoid repeated operations; for areas where new slag layer accumulation occurs in the evaluation results (due to slag re-aggregation caused by liquid surface sloshing during slag removal), new access points are inserted into the route. The adjustment strategy adopts a greedy algorithm or dynamic programming, prioritizing high adhesion level areas while ensuring slag removal efficiency. The adjusted route is output as a new path point sequence and immediately sent to the motion controller of the slag remover for execution, realizing a real-time optimization closed loop of "monitoring while cleaning and adjusting immediately after testing".

[0080] As one possible implementation, step S1044 can be specifically implemented as steps S10441-S10443.

[0081] S10441. Mark the areas with an adhesion level higher than the preset threshold in the evaluation results as priority treatment areas.

[0082] Specifically, the preset threshold is set according to the power of the slag remover and the tool type. For example, a adhesion level ≥ 2.0 (heavy) is used as the priority treatment threshold. All evaluation sub-regions with adhesion levels higher than this threshold are selected from the evaluation results of step S10434, and the union of these sub-regions is taken as the priority treatment area. The priority treatment area is marked in the internal data with an outline or highlighted block and assigned the highest operation priority (Priority=1). The purpose of this marking operation is to schedule the most difficult-to-remove areas for treatment in advance during route planning, avoiding difficulties in removal due to subsequent tool wear or changes in liquid level.

[0083] S10442, Shift the initial slag removal route toward the priority treatment area.

[0084] Specifically, when replanning the remaining unexecuted routes, a path bias strategy is adopted: using the initial slag removal route as a baseline, an attraction field function is introduced. This function takes a high value in the priority treatment area, a low value in the already cleared area, and a zero value in the normal area. The corrected path affected by this attraction field is solved using the gradient descent method. This causes the corrected route to deflect inwards towards the center of the priority treatment area when passing through it, ensuring an increase in the number of times the tool covers and the dwell time in that area. The bias amount is limited by the kinematic constraints of the slag remover (such as the minimum turning radius), ensuring the feasibility of the adjusted route. This bias treatment concentrates route resources on difficult-to-clean areas, achieving differentiated slag removal intensity configuration.

[0085] S10443. Areas where the slag layer has been removed in the assessment results will be excluded from the subsequent slag removal route.

[0086] Specifically, based on the real-time monitoring of the slag layer coverage status in step S10431, when the slag layer coverage ratio of a certain evaluation sub-area drops below a set threshold (e.g., 5%) and remains in this state for a certain period of time (e.g., 2 seconds), the sub-area is marked as "cleared". In subsequent route planning, all marked cleared sub-areas are set as obstacles or restricted areas, and the path planning algorithm (e.g., A) is used. The algorithm (or DWA algorithm) automatically avoids these areas when searching the path. This rejection operation gradually narrows the path of the slag remover, limiting the working area to the residual area of ​​the slag layer that has not been removed, effectively avoiding repeated operations and energy waste, and accelerating the convergence of the slag removal process.

[0087] S104. Control the slag cleaning machine to perform slag cleaning operation according to the initial slag cleaning route. During the slag cleaning process, collect the second image information of the liquid surface of the molten iron ladle and the force feedback information of the slag cleaning tool in the slag cleaning machine in real time. Based on the second image information and the force feedback information, correct the initial slag cleaning route in real time until the slag cleaning is completed.

[0088] Specifically, during the slag removal operation, an industrial camera continuously acquires liquid surface images at a frame rate of no less than 15fps, while a six-dimensional force sensor simultaneously acquires force data at a sampling rate of no less than 200Hz. These two sets of real-time data streams serve as dynamic feedback signals, inputting to the online path correction module. This module employs a model predictive control (MPC) framework: within each control cycle, starting from the current slag removal position, it combines the latest second image information (providing updates on the spatial distribution of the residual slag layer) and force feedback information (providing real-time evaluation of the slag removal effect in the current operating area) to perform rolling optimization of the remaining path's future finite step length, generating corrected local path instructions and executing them immediately. This correction includes not only adjustments to path point positions but also multi-parameter coordinated optimization of propulsion speed, tool angle, and reciprocating frequency. This real-time correction process iterates continuously until the second image information shows that the liquid surface slag coverage is below a preset completion threshold (e.g., <3%) and the force feedback information shows that the slag removal resistance across the entire area is stable at the no-load level. At this point, the system automatically determines that slag removal is complete and terminates the operation, achieving a fully autonomous slag removal control closed loop without manual intervention.

[0089] As one possible implementation, step S104 can be specifically implemented as step S1041.

[0090] S1041. When the second image information shows that the slag coverage rate of the molten iron ladle is lower than the preset threshold, and the force feedback information shows that the slag removal resistance of the entire area of ​​the molten iron ladle surface is lower than the preset threshold, it is determined that the slag removal is completed.

[0091] Specifically, the determination of slag removal completion adopts a dual-channel confirmation mechanism of "visual + force" to avoid misjudgment from a single information source. Visual channel: The latest frame of the second image information is used to segment the slag body across the entire image. The ratio of the total area of ​​slag body pixels to the total area of ​​the liquid surface is calculated as the slag coverage rate. When this ratio is lower than a preset threshold (e.g., 3%, corresponding to a small amount of scattered slag that can be ignored), the visual condition is met. Force channel: The slag remover is controlled to traverse the entire liquid surface area at low speed (or multiple representative locations are randomly sampled), and the slag removal resistance at each location is monitored in real time (characterized by the root mean square value of thrust). When the slag removal resistance at all locations is lower than a preset threshold (e.g., no-load baseline value + 5% margin), it indicates that there is no significant slag layer adhesion, and the force condition is met. Both conditions must be met simultaneously and continuously verified for more than a set time window (e.g., 3-5 seconds) to eliminate false triggering caused by liquid surface fluctuations or transient interference from sensors. Once the judgment is completed, the system generates a slag removal completion report, recording key indicators such as the total slag removal time, the removal effect of each area, and the maximum adhesion strength, and automatically resets the slag removal machine to a safe position, completing the entire process.

[0092] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A remote, unmanned method for cleaning slag from a molten iron ladle, characterized in that, include: During the process of pouring molten iron from the upstream container into the ladle, dynamic visual information of the molten iron during the pouring process is collected, and the initial distribution of slag in the ladle is predicted based on the dynamic visual information. After the molten iron is poured in, a first image of the liquid surface of the molten iron ladle is acquired; based on the first image information, the initial distribution is corrected for deviation to obtain the distribution to be checked; An initial slag removal route is generated based on the distribution to be checked; The slag cleaning machine is controlled to perform slag cleaning operations according to the initial slag cleaning route. During the slag cleaning process, the second image information of the liquid surface of the molten iron ladle and the force feedback information of the slag cleaning tools in the slag cleaning machine are collected in real time. The initial slag cleaning route is corrected in real time according to the second image information and the force feedback information until the slag cleaning is completed.

2. The method for remotely and unmanned slag removal from a molten iron ladle according to claim 1, characterized in that, The process of collecting dynamic visual information during the pouring of molten iron and predicting the initial distribution of slag within the ladle based on this dynamic visual information includes: The system continuously acquires time-series image sequences and flow rate information during the pouring process of molten iron, and generates the dynamic visual information based on the time-series image sequences and the flow rate information. Based on the time sequence image, the timing changes of the landing point when molten iron falls into the ladle are identified, and the flow and diffusion path of molten iron in the ladle is calculated based on the timing changes of the landing point and the flow rate information. The slag's buoyancy and aggregation pattern is obtained. Based on the buoyancy and aggregation pattern and the flow diffusion path, the aggregation probability distribution of slag in each region of the molten iron ladle is predicted, and the aggregation probability distribution is used as the initial distribution.

3. The method for remotely and unmanned slag removal from a molten iron ladle according to claim 2, characterized in that, The predicted probability distribution of slag aggregation in different regions within the ladle also includes: The fluctuation characteristics of the flow rate information during the pouring of molten iron were obtained; The distribution of turbulence intensity on the surface of molten iron is identified based on the aforementioned wave characteristics; In regions where the turbulence intensity distribution is higher than a first threshold, the aggregation probability distribution in the corresponding region is reduced; in regions where the turbulence intensity distribution is lower than a second threshold, the aggregation probability distribution in the corresponding region is increased.

4. The method for remotely and unmanned slag removal from a molten iron ladle according to claim 1, characterized in that, The step of correcting the initial distribution based on the first image information to obtain the distribution to be checked includes: The area covered by the initial distribution is divided into multiple grid cells, and each grid cell is assigned an initial confidence value for the amount of slag. Identify the slag-iron boundary based on the first image information; The initial value of the slag quantity confidence is adjusted based on the position of each grid cell relative to the slag-iron boundary; Based on the adjusted initial confidence value of the slag quantity, the distribution to be verified is obtained.

5. A remotely controlled, unmanned method for cleaning slag from a molten iron ladle according to claim 4, characterized in that, The step of adjusting the initial value of the slag quantity confidence based on the position of each grid cell relative to the slag-iron boundary includes: Determine whether each of the grid cells is located on one side of the slag region or the molten iron region of the slag-iron boundary; For the grid cell located on one side of the slag region, increase the initial value of the corresponding slag quantity confidence. For the grid cell located on one side of the molten iron region, the initial confidence value of the corresponding slag quantity is lowered.

6. The method for remotely and unmanned slag removal from a molten iron ladle according to claim 1, characterized in that, The force feedback information includes the thrust and lateral force of the slag removal tool in the slag removal propulsion direction; the real-time correction of the initial slag removal route based on the second image information and the force feedback information includes: The slag layer coverage status at the current slag removal location is identified based on the second image information; The adhesion strength of the slag layer in the current slag removal tool area is identified based on the thrust and the lateral force. By integrating and analyzing the slag layer coverage state and the slag layer bonding strength, an evaluation result of the slag layer state in the current slag cleaning tool area is obtained; The initial slag removal route is adjusted in real time based on the evaluation results.

7. A remotely controlled, unmanned method for cleaning slag from a molten iron ladle according to claim 6, characterized in that, The step of identifying the slag layer adhesion strength in the current slag removal tool area based on the thrust and the lateral force includes: The time-domain variation of the thrust during the process of the slag removal tool cutting into the slag layer was obtained; Based on the aforementioned time-domain variation, the feature value of bond strength is extracted; The fluctuation characteristics of the lateral force are obtained, and the bonding strength characteristic value is corrected according to the fluctuation characteristics to obtain the bonding strength and bonding grade of the slag layer.

8. A remotely controlled, unmanned method for cleaning slag from a molten iron ladle according to claim 7, characterized in that, The process of integrating and analyzing the slag layer coverage state and the slag layer adhesion strength to obtain the evaluation result of the slag layer state in the current slag cleaning tool area includes: The slag layer coverage area is determined based on the second image information; The slag layer covered area is discretized into multiple evaluation sub-regions; The bonding level of each evaluation sub-region is determined based on the force feedback information; The adhesion levels of each of the evaluation sub-regions are superimposed onto the corresponding positions of the covered area to form the evaluation results with spatial distribution.

9. A remotely controlled, unmanned method for cleaning slag from a molten iron ladle according to claim 8, characterized in that, The real-time adjustment of the initial slag removal route based on the evaluation results includes: The areas in the evaluation results where the adhesion level is higher than a preset threshold are marked as priority processing areas; The initial slag removal route is biased toward the priority treatment area; The areas where the slag layer has been removed in the assessment results will be excluded from the subsequent slag removal route.

10. A remotely controlled, unmanned method for cleaning slag from a molten iron ladle according to claim 1, characterized in that, The process until the slag removal is complete includes: When the second image information shows that the slag coverage of the molten iron ladle is lower than a preset threshold, and the force feedback information shows that the slag removal resistance of the entire molten iron ladle surface is lower than a preset threshold, the slag removal is determined to be complete.

Images