Vision and pressure fused intelligent detection method for depth of insertion of RH dip pipe into molten steel

Abstract

The invention relates to the technical field of ferrous metallurgy, and discloses a vision and pressure fused intelligent detection method for the depth of an RH dip pipe inserted into molten steel. The method comprises the following steps: when an RH vacuum tank is vacuumized, starting a built-in camera to collect and preprocess a tank bottom image signal to obtain an RH tank bottom preprocessed image; extracting an average pixel gray value and a local maximum brightness gradient of each frame of image as key features; constructing a two-dimensional feature image based on the key features, constructing a model by using an artificial intelligence algorithm, and determining the moment when the molten steel overlays the bottom of the RH vacuum tank; and reading the vacuum pressure at the moment, and calculating the depth of the dip pipe inserted into the molten steel by combining the preset length of the dip pipe, the molten steel density, the slag density and the slag thickness. Through visual feature extraction and vacuum pressure data fusion, the depth of the dip pipe inserted into the molten steel in the RH production process is accurately calculated in cooperation with the neural network model, and the problem that depth detection is inaccurate due to the fact that traditional detection is prone to being disturbed by the molten steel and slag is solved.

Description

Technical Field

[0001] This application relates to the field of iron and steel metallurgy technology, and in particular to an intelligent detection method for the insertion depth of an RH-immersed tube into molten steel by integrating vision and pressure. Background Technology

[0002] RH refining technology is a core refining method for producing high-quality steel in the iron and steel metallurgical industry, and its intelligent and precise process control is a key direction for the steel industry's development towards intelligent manufacturing. Precise control of the depth to which the immersion tube is inserted into the molten steel in the ladle directly affects the steel circulation, compositional uniformity, and refining efficiency during the RH refining process, and is crucial for ensuring steel quality. However, the harsh environment of high temperature and high dust in real-world production poses a significant challenge to the accurate detection of the immersion tube insertion depth.

[0003] Currently, the industry still relies primarily on manual observation to determine the insertion depth of RH immersion tubes into molten steel. Due to the high temperatures and hazards at production sites, operators cannot closely and accurately observe the contact state between the molten steel and the immersion tube, making it difficult to capture the precise moment when the molten steel enters the vacuum tank. This results in significant errors in judging the insertion depth. This traditional method not only significantly affects the degassing and composition adjustment effects of RH refining but also reduces the predictive accuracy of RH refining models for the production process, hindering the intelligent upgrading of the refining process. While existing related patent technologies attempt to improve detection methods, they have significant shortcomings: For example, patent CN202411286994.8 only identifies the ladle lifting status through image processing, without incorporating key physical parameters such as vacuum chamber pressure and immersion tube length, and only focuses on the lifting height, failing to directly obtain the actual depth of the immersion tube inserted into the molten steel; patent CN202310456840.8 controls the lifting by monitoring the bright surface area of ​​the molten steel, similarly failing to consider dynamic physical parameters such as vacuum pressure, resulting in insufficient detection accuracy under dynamic conditions during the vacuuming process; patent CN202323385978.8 relies on multiple sensors to detect the lifting height, but ordinary sensors are easily affected by the high temperature and dust environment of RH refining, the equipment structure is complex and lacks adaptive adjustment capabilities for different RH equipment, making it difficult to achieve universal application, and none of them can completely solve the problem of accurate detection of immersion tube insertion depth.

[0004] To address the aforementioned issues, this application combines visual feature extraction with vacuum pressure data fusion and a neural network model to accurately calculate the insertion depth of the immersion tube into the molten steel during RH production. This solves the problem of inaccurate depth detection caused by molten steel disturbance and slag interference in traditional detection methods, significantly improving the intelligence level of the RH refining process. Summary of the Invention

[0005] This application provides an intelligent detection method for the insertion depth of RH immersion tubes into molten steel by fusing visual and pressure data. By extracting visual features and fusing them with vacuum pressure data, and combining them with a neural network model, the method accurately calculates the insertion depth of the immersion tubes into molten steel during RH production. This solves the problem of inaccurate depth detection caused by molten steel disturbance and slag interference in traditional detection methods, and significantly improves the intelligence level of the RH refining process.

[0006] In a first aspect, this application provides a visual and pressure-integrated intelligent detection method for the immersion depth of an RH-impregnated tube in molten steel, the method comprising: Step S101: During the vacuuming process of the RH vacuum tank, the built-in camera is activated to collect image signals of the bottom area of ​​the RH vacuum tank in real time, and the image signals are preprocessed to obtain a preprocessed image of the bottom of the RH tank. Step S102: Extract the average pixel grayscale value of each frame in the preprocessed image of the RH tank bottom, and calculate the local maximum brightness gradient of the image to obtain key features characterizing the brightness change of the image. Step S103: Generate a two-dimensional feature image with spatial structure based on the key features, apply artificial intelligence algorithms to construct a smart model for molten steel positioning, and determine the moment when RH molten steel enters the RH vacuum tank and fills its bottom. Step S104: When the RH molten steel is determined to cover the bottom of the RH vacuum tank, read the vacuum pressure inside the RH vacuum tank at this time. Based on the vacuum pressure and the preset impregnation tube length, and in combination with the molten steel density, slag density and slag thickness, calculate the depth to which the RH impregnation tube is inserted into the molten steel in the ladle.

[0007] Optionally, step S101 includes: During the evacuation process of the RH vacuum tank, the built-in camera is activated and continuously captures image signals of the bottom area of ​​the RH vacuum tank in real time at a preset frame rate. The continuous image signal is converted to grayscale to obtain a first preprocessed image signal; The first preprocessed image signal is subjected to spatial filtering to obtain the second preprocessed image signal; The second preprocessed image signal is normalized to obtain the RH tank bottom preprocessed image.

[0008] Optionally, step S102 includes: For the preprocessed image of the RH tank bottom, the average pixel grayscale value representing the overall brightness change trend of each frame is extracted frame by frame; Meanwhile, for the preprocessed image of the RH tank bottom, the Sobel operator is used to calculate the gradient magnitude of each frame image, and the maximum gradient magnitude of each frame image is selected as the local maximum brightness gradient of that frame image. By integrating the average pixel grayscale value and the local maximum brightness gradient of each frame image, key features characterizing the brightness changes of the preprocessed image of the RH tank bottom are obtained.

[0009] Optionally, step S103 includes: The key features are arranged in a one-to-one correspondence with their spatial locations in the original image to generate a two-dimensional feature image with spatial structure. The two-dimensional feature image is input into the intelligent model for molten steel positioning, and the probability of RH molten steel reaching the bottom of the RH vacuum tank is output. Calculate the luminance slope and luminance amplitude increment of the average luminance between adjacent frames; When the probability of molten steel reaching the bottom of the RH vacuum tank is greater than or equal to a preset first threshold, the brightness slope is greater than a preset slope threshold, and the brightness amplitude increment is less than a preset amplitude threshold, it is preliminarily determined that the RH molten steel has entered the RH vacuum tank and filled its bottom. The moment of the preliminary determination is recorded as the moment when the RH molten steel enters the RH vacuum tank and fills its bottom.

[0010] Optionally, the intelligent model for molten steel positioning in step S103 includes a feature extraction module, a time series analysis module, and a decision output module. These modules work together to calculate the probability of RH molten steel reaching the bottom of the RH vacuum tank, as detailed below: The feature extraction module is constructed using a convolutional neural network. It takes the two-dimensional feature image as input and outputs a low-dimensional feature vector that can characterize the spatial features of the molten steel region. The convolutional neural network includes an input layer, at least four convolutional layers, at least four pooling layers, and one fully connected layer. The time series analysis module is constructed using a long short-term memory network. It takes the low-dimensional feature vector as input and outputs the time-optimized feature vector. The long short-term memory network contains at least two LSTM layers, one dropout layer, and one fully connected layer. The output module uses a sigmoid activation function to construct a binary classification output layer, which maps the time-optimized feature vector to a probability value between 0 and 1. This probability value is the probability that the RH molten steel reaches the bottom of the RH vacuum tank.

[0011] Optionally, in step S103, after determining that the RH molten steel has entered the RH vacuum tank and filled its bottom, a step to verify the stability of the molten steel surface needs to be added, specifically including: Calculate the standard deviation of brightness fluctuation and the mean of brightness change within a preset number of consecutive frames after the initial determination time; Determine whether both of the following conditions are met: the standard deviation of the brightness fluctuation is less than a preset fluctuation threshold and the mean value of the brightness change is less than a preset change threshold. If so, then the determination time is confirmed to be the final moment when the RH molten steel enters the RH vacuum tank and fills its bottom; If not, the brightness slope and brightness amplitude increment between adjacent frames are recalculated, and the arrival probability of molten steel is combined with the re-determination until a determination time that meets the stability requirements is obtained.

[0012] Optionally, step S104 includes: After determining the moment when the RH molten steel has completely filled the bottom of the RH vacuum tank, the vacuum pressure inside the RH vacuum tank is read using the pressure detection device configured in the RH vacuum tank. ; Get the preset impregnation tube length Density of molten steel and slag density And read the current slag thickness in the ladle. ; Based on the vacuum pressure The length of the impregnation tube The density of the molten steel The density of the slag and the thickness of the slag Using formula Calculate the depth H of the RH immersion tube inserted into the molten steel in the ladle, where the length is in meters, the density is in t / m³, and the pressure is in Pa.

[0013] This application provides an intelligent detection method for the insertion depth of an RH-treated tube into molten steel by fusing visual and pressure data. It accurately calculates the insertion depth of the tube into molten steel during RH production by extracting visual features and fusing them with vacuum pressure data, combined with a neural network model. This solves the problem of inaccurate depth detection caused by molten steel disturbance and slag interference in traditional methods. Compared with existing technologies, the beneficial effects of this application's technical solution are at least as follows: First, by utilizing the existing built-in camera and vacuum pressure sensor of the RH equipment, the insertion depth is determined by combining the camera's captured signal with the vacuum pressure of the vacuum tank. This eliminates the need for additional complex equipment or modifications to RH equipment with different structural parameters, making it adaptable to various RH refining devices and solving the problem of some existing technologies being only applicable to specific equipment and having poor compatibility.

[0014] Secondly, by applying artificial intelligence algorithms such as pattern recognition and neural networks, the images captured by the camera are preprocessed by grayscale conversion and spatial filtering, and the average brightness and local maximum brightness gradient features are extracted. Combined with brightness slope-amplitude detection, a robust protection layer is established to accurately identify the moment when molten steel fills the bottom of the vacuum tank. Then, the depth is calculated by integrating vacuum pressure, which solves the problems of poor accuracy caused by the on-site environment in traditional manual observation, and the fact that existing technologies do not consider physical parameters or are easily interfered with.

[0015] Third, by completing the identification of the moment when the molten steel is fully spread, the reading of the vacuum pressure, and the calculation of the depth online during the RH vacuuming process, an online and accurate method is provided for the evaluation and adjustment of the immersion tube depth. This solves the problem that traditional methods are difficult to obtain accurate timing, which affects the RH refining effect and the prediction accuracy of the refining model, and supports the industrial control system to adjust process parameters in real time. Attached Figure Description

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

[0017] Figure 1 This is a flowchart illustrating the intelligent detection method for the RH immersion tube insertion depth in molten steel, which integrates vision and pressure in this application. Figure 2 This is a schematic diagram of the convolutional neural network structure in the intelligent detection process of RH immersion tube insertion depth into molten steel using vision and pressure fusion in this application. Figure 3 This is a schematic diagram of the intelligent detection results of the RH immersion tube insertion depth into molten steel using the vision and pressure fusion method of this application. In the diagram, 1-camera, 2-vacuum chamber, 3-immersion tube, 4-steel slag, 5-steel ladle, 6-molten steel. Detailed Implementation

[0018] This application provides an intelligent detection method for the immersion depth of an RH-impregnated tube in molten steel, integrating vision and pressure. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data used can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent detection method for the insertion depth of RH-impregnated tubes into molten steel, which integrates vision and pressure, includes: Step S101: During the vacuuming process of the RH vacuum tank, the built-in camera is activated to collect image signals of the bottom area of ​​the RH vacuum tank in real time, and the image signals are preprocessed to obtain a preprocessed image of the bottom of the RH tank.

[0020] In one specific embodiment, step S101 may specifically include the following steps: During the evacuation process of the RH vacuum tank, the built-in camera is activated and continuously captures image signals of the bottom area of ​​the RH vacuum tank in real time at a preset frame rate. The continuous image signal is converted to grayscale to obtain a first preprocessed image signal; The first preprocessed image signal is subjected to spatial filtering to obtain the second preprocessed image signal; The second preprocessed image signal is normalized to obtain the RH tank bottom preprocessed image.

[0021] Specifically, when the RH vacuum tank starts its vacuuming process, the built-in camera is simultaneously activated. The camera must be synchronized with the RH refining equipment beforehand to ensure the deviation between the vacuuming procedure and image acquisition start time does not exceed 0.5 seconds. The preset frame rate is set to 25-30 frames per second. This frame rate range is determined based on the flow rate characteristics of molten steel entering the vacuum tank during RH refining. As the pressure inside the vacuum tank gradually decreases from atmospheric pressure to the required process value, the speed at which molten steel enters the vacuum tank from the riser and downcomer changes with the vacuum level. A frame rate of 25-30 frames per second can completely capture the dynamic process of molten steel from its initial entry to filling the bottom of the tank, avoiding missing key image frames due to a low frame rate or increasing the processing burden due to excessively high frame rates causing data redundancy. The acquired continuous image signal is an RGB three-channel color image, with each channel having a pixel value range of 0-255. The resolution of a single frame image is set to 640×480. This resolution can cover the bottom area of ​​the vacuum tank while ensuring that the image data volume is suitable for the real-time transmission and storage needs of the industrial control equipment, avoiding transmission delays caused by excessively high resolution.

[0022] When performing grayscale conversion on continuous image signals, a weighted average method is used to convert the RGB three-channel pixel values ​​into a single grayscale value. The conversion formula is as follows: ,in , and These represent the pixel values ​​of the red, green, and blue channels in a color image, respectively. This represents the converted grayscale value. Through this formula, each frame of color image generates a corresponding grayscale image, which is the first preprocessed image signal. The pixel value range of this signal remains 0~255. Grayscale processing eliminates information redundancy between channels in the color image, reducing subsequent data processing. It also highlights the image's brightness characteristics. Because there is a significant difference in brightness between the molten steel and the metal at the bottom of the vacuum tank, the grayscale image more clearly reflects the brightness changes as the molten steel overflows the bottom, providing a more focused data basis for subsequent identification of the molten steel's location.

[0023] The first preprocessed image signal needs to undergo spatial filtering, using a 5×5 Gaussian filter with the following kernel function: ,in σ is the standard deviation of the Gaussian function, with a value of 1.0. The coordinates are within the filter kernel. During the filtering process, a 5×5 pixel area is taken around each pixel in the first preprocessed image signal as the center. The pixel values ​​within this area are multiplied by the corresponding Gaussian filter kernel weights and summed. The result is used as the new value of the central pixel. This process is repeated for all pixels to obtain the second preprocessed image signal. The interior of the RH vacuum tank is in a high-temperature environment. When the camera captures images, it is susceptible to high-frequency noise caused by equipment vibration and high-temperature radiation. This noise can cause random bright spots or dark spots in the first preprocessed image signal, obscuring the true characteristics of the brightness changes in the molten steel. Gaussian filtering can effectively filter out high-frequency noise while retaining the brightness gradient information of the molten steel edges in the image, ensuring that the second preprocessed image signal can accurately reflect the actual distribution of the molten steel.

[0024] The second preprocessed image signal is normalized to obtain the preprocessed image of the RH tank bottom. This is achieved using the formula... Map pixel values ​​to the range [0,1], where, For the pixels in the second preprocessed image signal grayscale value, The minimum grayscale value in this frame of the image. The maximum grayscale value. This refers to the normalized grayscale value in this frame of the image. During the RH refining process, the temperature of the molten steel and the dust concentration inside the vacuum ladle may vary between different batches, causing fluctuations in the grayscale value range of the second preprocessed image signal acquired at different times. For example, the grayscale value range might be 50-200 at one time and 80-220 at another. If directly used for subsequent feature extraction, the different pixel scales would affect the stability of the algorithm's analysis. After normalization, the pixel values ​​of the RH ladle bottom preprocessed images in all frames are at the same scale, ensuring that the subsequent artificial intelligence algorithm can identify changes in molten steel brightness based on a unified standard when analyzing images at different times, thus solving the problem of inconsistent data scales under different acquisition conditions.

[0025] Step S102: Extract the average pixel grayscale value of each frame in the preprocessed image of the RH tank bottom, and calculate the local maximum brightness gradient of the image to obtain key features characterizing the brightness change of the image.

[0026] In one specific embodiment, step S102 may specifically include the following steps: For the preprocessed image of the RH tank bottom, the average pixel grayscale value representing the overall brightness change trend of each frame is extracted frame by frame; Meanwhile, for the preprocessed image of the RH tank bottom, the Sobel operator is used to calculate the gradient magnitude of each frame image, and the maximum gradient magnitude of each frame image is selected as the local maximum brightness gradient of that frame image. By integrating the average pixel grayscale value and the local maximum brightness gradient of each frame image, key features characterizing the brightness changes of the preprocessed image of the RH tank bottom are obtained.

[0027] Specifically, when extracting the average pixel grayscale value frame by frame for the preprocessed image of the RH tank bottom, it is necessary to first determine the effective pixel area of ​​each frame. This area consists of the metal plane at the bottom of the vacuum tank and the area covered by molten steel, which are retained after preprocessing. Invalid pixels caused by the equipment structure at the image edges need to be excluded. Let the total number of effective pixels in a single frame of the preprocessed RH tank bottom image be N, and the th... i Each valid pixel in the frame image The normalized gray value is Then the first i Average pixel grayscale value of the frame image Through formula Calculations show that during the RH refining process, when the molten steel does not cover the bottom of the vacuum tank, the normalized grayscale value of the bottom metal surface stabilizes in the range of 0.6 to 0.7, and the average pixel grayscale value... The grayscale value remains between 0.62 and 0.68; as the molten steel gradually overflows the bottom, the normalized grayscale value of the molten steel is lower than that of the metal plane, resulting in a decrease in the average pixel grayscale value. The value decreases linearly with increasing molten steel coverage area, stabilizing at 0.35–0.42 after the molten steel completely covers the bottom. This was calculated frame by frame. It can capture the overall brightness trend of the image in real time, providing data support for the overall brightness dimension to determine the moment when the molten steel is fully covered.

[0028] When calculating the local maximum brightness gradient, a 3×3 Sobel operator is used for each frame of the RH tank bottom preprocessed image. This operator includes a horizontal direction operator. with vertical direction operator ,in: During processing, a 3×3 neighborhood pixel matrix is ​​taken as the center of each valid pixel in the image, and this matrix is ​​then compared with... and Perform convolution operations separately to obtain the horizontal gradient values. gradient value in the vertical direction The calculation formulas are as follows: , , where m corresponds to the horizontal offset of the image and n corresponds to the vertical offset of the image, both ranging from -1, 0, to 1. Then, using the formula... Calculate the gradient magnitude of the pixel, and after traversing all valid pixels, select the largest gradient magnitude value as the local maximum brightness gradient of the frame image. In the RH refining scenario, the brightness difference between the molten steel and the metal plane creates a significant gradient change at their boundary: when the molten steel does not cover the bottom, there is no obvious brightness boundary in the image, and the maximum local brightness gradient is observed. Below 0.2; when the molten steel begins to diffuse and form a boundary, the gradient amplitude of the pixels at the boundary increases sharply to 0.5~0.7, and the local maximum brightness gradient... The brightness then increases; until the molten steel completely covers the bottom, the boundary disappears, and the local maximum brightness gradient is reached. It fell back below 0.2. (Through a local maximum brightness gradient) The fluctuation reflects the existence or absence of the molten steel boundary, providing data support for determining the moment when the molten steel is fully covered.

[0029] When integrating the average pixel grayscale value and local maximum brightness gradient of each frame as key features, it is necessary to establish the correlation between the two and the timestamp, with each frame corresponding to a collection timestamp. The frame , and t The data is bound to storage to form a three-dimensional data set of "timestamp-average grayscale value-maximum gradient". During the RH refining process, this data set needs to be synchronized with the pressure values ​​inside the vacuum tank. This correlation ensures that when determining the moment when molten steel is fully saturated, the pressure data for that specific moment can be retrieved simultaneously. For example, when a frame of an image... Stable between 0.35 and 0.42 for 3 consecutive frames When the pressure is below 0.2, it can be determined that the molten steel has covered the bottom. At this point, the pressure value P at that moment is retrieved by matching the timestamp, providing a key parameter for subsequent calculation of the insertion depth of the impregnation tube. Through multi-dimensional data correlation, it is ensured that the brightness change characteristics can directly serve the depth calculation, avoiding judgment delays or errors caused by data gaps.

[0030] Step S103: Generate a two-dimensional feature image with spatial structure based on the key features, apply artificial intelligence algorithms to construct a smart model for molten steel positioning, and determine the moment when RH molten steel enters the RH vacuum tank and fills its bottom.

[0031] In one specific embodiment, step S103 may specifically include the following steps: The key features are arranged in a one-to-one correspondence with their spatial locations in the original image to generate a two-dimensional feature image with spatial structure. The two-dimensional feature image is input into the intelligent model for molten steel positioning, and the probability of RH molten steel reaching the bottom of the RH vacuum tank is output. Calculate the luminance slope and luminance amplitude increment of the average luminance between adjacent frames; When the probability of molten steel reaching the bottom of the RH vacuum tank is greater than or equal to a preset first threshold, the brightness slope is greater than a preset slope threshold, and the brightness amplitude increment is less than a preset amplitude threshold, it is preliminarily determined that the RH molten steel has entered the RH vacuum tank and filled its bottom. The moment of the preliminary determination is recorded as the moment when the RH molten steel enters the RH vacuum tank and fills its bottom.

[0032] Specifically, the key features integrated in step S102 are arranged according to their spatial positions in the original image, forming a 2-channel 64×64 two-dimensional feature image. The first channel is the average gray value matrix of each sub-region, and the second channel is the maximum gradient value matrix of each sub-region. Collecting 10 consecutive preprocessed images generates a feature image sequence with dimensions (10, 64, 64, 2). When constructing the intelligent model for molten steel positioning using artificial intelligence algorithms, a CNN-LSTM hybrid neural network structure is adopted. The input to this model is the feature image sequence. During the model training phase, more than 100 sets of historical data collected in RH refining production are used. Each set of data contains a "feature sequence - molten steel state label," where the molten steel state label is divided into "not fully covered" (label 0) and "fully covered" (label 1). The labeling is based on manual observation records and vacuum pressure change curves. The CNN layer is responsible for extracting the spatial correlation (such as average pixel gray value) in the feature image sequence. With local maximum brightness gradient The numerical correspondence is calculated using four convolutional layers and four pooling layers, compressing the output dimension to half that of the original feature image sequence. The LSTM layer is responsible for capturing the temporal dependencies of the feature image sequence (such as the average pixel grayscale value between consecutive frames). With local maximum brightness gradient (Based on the fluctuation pattern), one hidden layer is set, and the output is the probability of molten steel reaching the bottom of the vacuum tank for each frame. (Value range [0,1]). The model construction process solves the problems of low accuracy caused by reliance on experience in manual observation. It quantifies the state of molten steel through algorithms, avoiding subjective judgment errors.

[0033] When calculating the luminance slope and luminance amplitude increment of the average luminance between adjacent frames, the average pixel grayscale value extracted in step S102 is used. Based on this, let the first i Frame and the i+ The time interval for acquiring one frame is (Determined by the camera frame rate; at a frame rate of 30 frames per second) (seconds), then the brightness slope Through formula Calculation. This value reflects the rate of change of average brightness, the average pixel grayscale value during the molten steel diffusion stage. Continuing to decrease, brightness slope (The larger the absolute value, the faster the decrease); after the molten steel is fully spread, the average pixel grayscale value Stable, brightness slope Approaching 0. Brightness amplitude increment Through formula The calculation reflects the change in average brightness between adjacent frames, specifically the increase in brightness when the molten steel is not fully spread. Larger (usually 0.02~0.05); after the molten steel is fully spread, the brightness increases. Less than 0.01. In the RH refinement scenario, the calculation results need to be smoothed using a 3-frame moving average (applied to three consecutive brightness slopes). With brightness amplitude increment The average value is used to eliminate abnormal fluctuations caused by single-frame noise and ensure the stability of parameter calculation. The slope and amplitude increments are used to supplement the determination of probability values, avoiding misjudgments caused by relying solely on model probabilities.

[0034] To initially determine whether the molten steel has completely filled the vacuum tank, three conditions must be met simultaneously: first, the probability of the molten steel reaching the bottom of the vacuum tank. Greater than or equal to the preset first threshold ( The value is set to 0.9, determined based on the model's test accuracy; a probability exceeding 0.9 indicates high model reliability. Secondly, the brightness slope... Greater than the preset slope threshold ( Set to -0.05 / second, based on the average pixel grayscale value before the molten steel covers the entire surface. The rate of descent slows down, and the absolute value of the slope must be less than 0.05 / second, i.e., the brightness slope. >-0.05 / second indicates that the brightness change has stabilized); thirdly, the increase in brightness amplitude. Less than the preset amplitude threshold ( Set to 0.01, the increase in brightness after the molten steel has fully covered the surface. The value must be less than 0.01 to ensure no significant fluctuations in brightness. When three consecutive frames simultaneously meet the above three conditions, it is determined that the molten steel has entered the vacuum tank and filled the bottom, and the result is recorded. i +1 frame capture timestamp This serves as the initial determination point. The initial determination point must be synchronized with the pressure values ​​collected by the pressure sensor inside the vacuum tank. This provides pressure parameters for subsequent calculations of the insertion depth of the impregnation tube. Through multi-condition collaborative judgment, taking into account probability quantification, rate change, and amplitude stability, the accuracy of real-time recording is ensured, avoiding judgment delays or misjudgments caused by a single condition trigger.

[0035] In one specific embodiment, the intelligent model for molten steel positioning in step S103 includes a feature extraction module, a time series analysis module, and a decision output module. These modules work together to calculate the probability of RH molten steel reaching the bottom of the RH vacuum tank, as detailed below: The feature extraction module is constructed using a convolutional neural network. It takes the two-dimensional feature image as input and outputs a low-dimensional feature vector that can characterize the spatial features of the molten steel region. The convolutional neural network includes an input layer, at least four convolutional layers, at least four pooling layers, and two fully connected layers. The time series analysis module is constructed using a long short-term memory network. It takes the low-dimensional feature vector as input and outputs the time-optimized feature vector. The long short-term memory network contains at least two LSTM layers, one dropout layer, and one fully connected layer. The output module uses a sigmoid activation function to construct a binary classification output layer, which maps the time-optimized feature vector to a probability value between 0 and 1. This probability value is the probability that the RH molten steel reaches the bottom of the RH vacuum tank.

[0036] Specifically, please refer to Figure 2 The feature extraction module is constructed using a convolutional neural network. The input is a two-dimensional feature image with spatial structure. R frames of preprocessed images are continuously acquired to generate a feature image sequence with dimensions (R, 64, 64, 2), where R is the number of consecutive input frames. Here, 10 frames are used to cover the temporal segment of molten steel diffusion. The input layer receives data through this matrix and passes it to the convolutional layer. Four convolutional layers are set up. The first layer has a kernel size of 3×3, 32 kernels, and a stride of 1. This is achieved using the formula... Calculate the feature map and extract the basic correlation features between the average pixel gray value and the local maximum brightness gradient of each frame image, where For the first layer Each feature map element These are the weights of the first layer convolutional kernel. For the input matrix elements, For biasing; the second and third convolutional layers have 2×2 kernels, 32 kernels, and a stride of 1 to further enhance local features; the fourth convolutional layer has 2×2 kernels, 64 kernels, and a stride of 1 to deepen the level of feature abstraction. Four pooling layers are set, one after each convolutional layer, using max pooling (3×3 kernels, stride 2), calculated using the formula... Compressing feature dimensions reduces computational cost while retaining key features, where For the pooled feature map at the location The value, This represents the value of the feature map at the corresponding position before pooling. The fully connected layer receives the features after pooling in the fourth layer and outputs a low-dimensional feature vector of 1×32 for each frame. After processing 10 frames consecutively, a 10×32 feature sequence is formed, which is then input into an LSTM for temporal analysis. This vector integrates the spatial correlation of key features, avoiding judgment bias caused by feature fragmentation.

[0037] The time series analysis module is constructed using a Long Short-Term Memory (LSTM) network. The input is a low-dimensional feature vector (10×32) output from the feature extraction module. It employs two LSTM layers, each with 32 hidden units. The first LSTM layer is constructed using the formula... , , , , , Processing time series data ( For the Gate of Oblivion For input gate, For output gate, Neuron state, In hidden state, For the sigmoid function, The first LSTM layer (elemental product) captures the temporal dependencies of features in consecutive frames. The second LSTM unit further optimizes the hidden states output by the first layer, enhancing the stability of temporal features. A dropout layer, located between the two LSTM layers with a dropout rate of 0.2, randomly masks some neurons to prevent overfitting and outputs a temporally optimized feature vector (10×32), ensuring the continuity of feature analysis during the dynamic diffusion of molten steel. Subsequently, a fully connected layer following the second LSTM layer maps the 32-dimensional features of each frame to 1-dimensional fractions using a weight matrix. This fully connected layer transforms the 10×32 feature sequence into a 10×1 fraction sequence, providing a foundation for subsequent probability mapping.

[0038] The output module uses a sigmoid activation function to construct a binary classification output layer. The input is a time-optimized feature vector, which is then processed using the formula... Map the feature vector of each frame to a probability value between 0 and 1. ( For output layer weights, This represents the final hidden state of the timing analysis module. (For bias). In the RH refining scenario, when the molten steel does not completely cover the bottom of the vacuum tank, the probability value for each frame is... Typically less than 0.3; probability value for each frame as molten steel gradually covers the surface. Gradually increase; after the molten steel is fully covered, the probability value for each frame... A probability value consistently above 0.9 quantifies the likelihood of molten steel reaching the bottom, providing a precise probabilistic basis for subsequent multi-condition judgments. During the collaborative process of each module, the feature extraction module is responsible for spatial feature abstraction, the temporal analysis module for temporal feature optimization, and the judgment output module for probability mapping, forming a complete data chain of "feature input-processing-output" to ensure stable output of probability values ​​under different molten steel temperatures and vacuum conditions.

[0039] In one specific embodiment, after determining that the RH molten steel has entered the RH vacuum tank and filled its bottom in step S103, a step to verify the stability of the molten steel surface is also required, specifically including: Calculate the standard deviation of brightness fluctuation and the mean of brightness change within a preset number of consecutive frames after the initial determination time; Determine whether both of the following conditions are met: the standard deviation of the brightness fluctuation is less than a preset fluctuation threshold and the mean value of the brightness change is less than a preset change threshold. If so, then the determination time is confirmed to be the final moment when the RH molten steel enters the RH vacuum tank and fills its bottom; If not, the brightness slope and brightness amplitude increment between adjacent frames are recalculated, and the arrival probability of molten steel is combined with the re-determination until a determination time that meets the stability requirements is obtained.

[0040] Specifically, after initially determining that the RH molten steel has entered the vacuum tank and filled the bottom, the subsequent 10 consecutive frames of RH tank bottom preprocessed images are selected, with the determination time as the starting point. The preset frame number is set to 10 to adapt to the camera's frame rate of 30 frames per second. This frame number can cover a duration of about 0.33 seconds and can capture short-term fluctuations in the liquid surface. Then, the average pixel grayscale value of each frame in these 10 images is extracted. The extraction object is the images from the 1st to the 10th frame after the determination time.

[0041] To calculate the standard deviation of brightness fluctuation, first, take the arithmetic mean of the average pixel grayscale values ​​of the extracted 10 frames to obtain the mean of the average brightness of these 10 frames. Then, subtract the mean from the average pixel grayscale value of each frame, square each difference, sum all the squared results, divide by 10, and finally take the square root of the result, which is the standard deviation of brightness fluctuation. This value reflects the degree of dispersion of brightness. When the molten steel surface is stable, the standard deviation of brightness fluctuation is less than 0.005; when the surface fluctuates (such as when molten steel continues to flow in or splash), the standard deviation of brightness fluctuation is greater than 0.01. When calculating the average brightness change, first calculate the absolute value of the difference between the average pixel grayscale values ​​of the 1st and 2nd frames, the 2nd and 3rd frames, ... the 9th and 10th frames after the judgment time, to obtain the brightness change amplitude of 9 adjacent frames. Then take the arithmetic mean of these 9 change amplitudes, which is the average brightness change. This value reflects the intensity of the average brightness change. When the liquid surface is stable, the average brightness change is less than 0.003; when the liquid surface fluctuates, the average brightness change is greater than 0.008.

[0042] The system checks whether the standard deviation of brightness fluctuation is less than a preset fluctuation threshold (set to 0.005) and whether the mean brightness change is less than a preset change threshold (set to 0.003). If both conditions are met, the original judgment time is confirmed as the final time. If either condition is not met, the image data of the period that failed the verification is retrieved again, and the brightness slope and brightness amplitude increment of adjacent frames are calculated according to the original method in step S103. At the same time, the relevant data is input into the intelligent model for molten steel positioning to obtain a new probability of molten steel arrival, and the judgment process is repeated. This verification step avoids misjudging the final time due to instantaneous coverage (such as brief coverage by splashing molten steel) and ensures that the time used for subsequent depth calculation corresponds to a stable liquid surface state.

[0043] Step S104: When the RH molten steel is determined to cover the bottom of the RH vacuum tank, read the vacuum pressure inside the RH vacuum tank at this time. Based on the vacuum pressure and the preset impregnation tube length, and in combination with the molten steel density, slag density and slag thickness, calculate the depth to which the RH impregnation tube is inserted into the molten steel in the ladle.

[0044] In one specific embodiment, step S104 may specifically include the following steps: After determining the moment when the RH molten steel has completely filled the bottom of the RH vacuum tank, the vacuum pressure inside the RH vacuum tank is read using the pressure detection device configured in the RH vacuum tank. ; Get the preset impregnation tube length Density of molten steel and slag density And read the current slag thickness in the ladle. ; Based on the vacuum pressure The length of the impregnation tube The density of the molten steel The density of the slag and the thickness of the slag Using formula Calculate the depth H of the RH immersion tube inserted into the molten steel in the ladle, where the length is in meters, the density is in t / m³, and the pressure is in Pa.

[0045] For details, please refer to Figure 3 After determining the moment when RH molten steel completely fills the bottom of the vacuum ladle, the vacuum pressure P inside the ladle is read using a pressure detection device configured in the ladle. The pressure detection device and the image judgment system are synchronized via a real-time clock to ensure that the time difference between the pressure data acquisition moment and the final moment when the molten steel stably fills the ladle does not exceed 0.1 seconds. The pressure detection device has a range of 50–500 Pa, and the measurement error is controlled within ±1 Pa to ensure the accuracy of the basic data. The preset impregnation tube length L is obtained; this parameter is pre-stored in the control system according to the ladle specifications and is automatically matched using the ladle identification code when retrieved. The molten steel density is then obtained. Based on the current steel grades being smelted, the database shows that the density of molten low-carbon steel is 7.0 t / m³, and the density of molten high-carbon steel is 7.2 t / m³. The slag density is then obtained. Based on the slag composition analysis, the density of the slag, mainly composed of CaO-SiO2, was determined to be 2.8 t / m³. The current slag thickness H was read by a laser sensor installed on the top of the ladle. S This value is transmitted to the control system in real time, and all parameters are automatically converted to a unified unit when called: length in meters, density in t / m³, and pressure in Pa.

[0046] Using formula Calculate the depth to which the RH immersion tube is inserted into the molten steel in the ladle. H First, we need to handle the operational relationships between the parameters in the formula. In the formula, (P−103000) is the difference between vacuum pressure and atmospheric pressure, and 103000Pa is the standard value of local atmospheric pressure; 9800 is the product of gravitational acceleration 9.8m / s² and 1000, used to achieve unit conversion; This item reflects the combined effect of the length of the immersion tube and the density of the molten steel; The study reflects the influence of slag thickness and density, and ultimately obtains the depth H through formula integration. By precisely synchronizing pressure data with time, combined with equipment parameters and material characteristic parameters, the depth can be quantitatively calculated, avoiding deviations caused by experience-based judgments and providing reliable data support for subsequent process adjustments.

[0047] The method will be described below with reference to specific embodiments: A steel plant uses a 200-ton RH refining unit to produce low-carbon steel. The immersion tube length L of the RH unit is 1.65m. The target vacuum level during the vacuuming process is 90Pa, and the density of the molten steel is... The density of the slag is slag thickness .

[0048] The method for calculating the insertion depth of the impregnated tube as described in this invention includes the following steps: During the RH vacuuming process, the built-in camera of the vacuum tank acquires 640×480 color images at a rate of 25 frames per second, with a sequence length of 10 frames. After grayscale conversion, 5×5 Gaussian filtering for noise reduction, and normalization, the average pixel grayscale value and local maximum brightness gradient features are extracted.

[0049] The intelligent model for molten steel positioning was trained and validated using data from 120 RH operation processes. During training, the Adam optimizer (learning rate 0.001) was used with a batch size of 16. After 60 iterations, the loss function on the validation set converged, achieving an accuracy of 97.5% on the validation set.

[0050] The extracted key features are arranged one-to-one with their spatial locations in the original image to generate a two-dimensional feature image with spatial structure. This image is then input into the intelligent model for molten steel localization. The processing results show that during the 0-3s period, the average pixel gray value is stable at 0.65, which matches the brightness characteristics of the bottom metal plane when the molten steel is not covered, and the local maximum brightness gradient remains below 0.15. At 3.2s, the average pixel gray value suddenly drops to 0.35, indicating that the molten steel coverage has reduced the brightness, and the local maximum brightness gradient increases to 0.55, forming a clear boundary between the molten steel and the metal plane, with the gradient value matching the boundary characteristics. The CNN module performs spatial correlation extraction on the two-dimensional feature images before and after 3.2s, outputting a low-dimensional feature vector with a dimension of 1×32. The LSTM module captures the temporal changes of features during this period and outputs a temporally optimized feature vector. The decision output module uses the sigmoid activation function to map the temporally optimized feature vector to a confidence level of 0.92 at 3.3s, which is the probability that the molten steel has reached the bottom of the vacuum tank. This value is greater than the preset first threshold of 0.9.

[0051] Simultaneously, the brightness slope between 3.2s and 3.3s is calculated (the difference in average pixel grayscale values ​​divided by the time interval 0.1s, resulting in -0.02 / second), which is greater than the preset slope threshold of -0.05 / second; the brightness amplitude increment is calculated (the absolute value of the difference in average pixel grayscale values ​​between the two frames, 0.005), which is less than the preset amplitude threshold of 0.01. Combining the constraints on the brightness slope and amplitude, all judgment conditions are met, and it is preliminarily determined that the molten steel has filled the bottom of the vacuum tank. 3.3s is recorded as the preliminary judgment time.

[0052] Steel molten surface stability verification: Ten consecutive pre-processed images of the bottom of the RH ladle were selected after 3.3 seconds. The average pixel grayscale values ​​of each frame were extracted as follows: 0.349, 0.352, 0.348, 0.351, 0.349, 0.350, 0.347, 0.352, 0.348, and 0.350. The calculated standard deviation of brightness fluctuation (0.0015) was less than the preset fluctuation threshold (0.005), and the mean value of brightness change (0.0028) was less than the preset change threshold (0.003), which met the judgment criteria of "solid surface stability". It can be finally confirmed that the moment when the molten steel filled the bottom of the vacuum ladle is valid.

[0053] The vacuum pressure inside the vacuum tank was read at the corresponding moment as P=30387Pa. Combined with the RH impregnation tube length L=1.65m, and based on the preset insertion depth calculation formula: = =0.57, and the calculated insertion depth of the RH-impregnated tube is H=0.57m.

[0054] As described in the above embodiments, those skilled in the art will understand that this application provides an intelligent detection method for the insertion depth of an RH immersion tube into molten steel, which integrates visual and pressure measurements. First, when the RH vacuum tank is evacuated, a built-in camera is activated to acquire and preprocess image signals from the bottom of the tank, obtaining a preprocessed image of the RH tank bottom. Second, the average pixel grayscale value and local maximum brightness gradient of each frame are extracted as key features. Furthermore, a two-dimensional feature image is constructed based on these key features, and an artificial intelligence algorithm is used to build a model to determine the moment when the molten steel completely covers the bottom of the RH vacuum tank. Finally, the vacuum pressure at this time is read, and combined with the preset immersion tube length, molten steel density, slag density, and slag thickness, the insertion depth of the immersion tube into the molten steel is calculated. This application, through the fusion of visual feature extraction and vacuum pressure data, combined with a neural network model, accurately calculates the insertion depth of the immersion tube into the molten steel during RH production, solving the problem of inaccurate depth detection caused by molten steel disturbance and slag interference in traditional detection methods, and significantly improving the intelligence level of the RH refining process.

[0055] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 this application.

Claims

1. A method for intelligently detecting the depth of RH tube insertion into molten steel by fusing vision and pressure, characterized in that, The method comprises the following steps: Step S101, during the vacuumizing process of the RH vacuum tank, starting the built-in camera to collect image signals of the bottom area of the RH vacuum tank in real time, and obtaining an RH tank bottom preprocessed image after preprocessing the image signals; Step S102, extracting the average pixel gray value of each frame of image in the RH tank bottom preprocessed image, and calculating the local maximum brightness gradient of the image to obtain a key feature representing the brightness change of the image; Step S103, generating a two-dimensional feature image with a spatial structure based on the key feature, applying an artificial intelligence algorithm to construct a molten steel positioning intelligent model, and determining and identifying the time when the RH molten steel enters the RH vacuum tank and covers the bottom thereof; Step S104, at the time when the RH molten steel covers the bottom of the RH vacuum tank, reading the vacuum pressure in the RH vacuum tank at this time, and calculating the depth of the RH immersion tube inserted into the ladle molten steel based on the vacuum pressure, the preset length of the immersion tube, and in combination with the density of the molten steel, the density of the slag, and the thickness of the slag.

2. The method of claim 1, wherein, The step S101 comprises: During the vacuumizing process of the RH vacuum tank, starting the built-in camera and collecting continuous image signals of the bottom area of the RH vacuum tank in real time at a preset frame rate; Performing gray scale processing on the continuous image signals to obtain a first preprocessed image signal; Performing spatial filtering processing on the first preprocessed image signal to obtain a second preprocessed image signal; Performing normalization processing on the second preprocessed image signal to obtain an RH tank bottom preprocessed image.

3. The method of claim 1, wherein, The step S102 comprises: For the RH tank bottom preprocessed image, the average pixel gray value representing the overall brightness change trend of each frame of image is extracted frame by frame; At the same time, for the RH tank bottom preprocessed image, the gradient amplitude of each frame of image is calculated by using a Sobel operator, and the maximum gradient amplitude of each frame of image is selected as the local maximum brightness gradient of the frame of image; The average pixel gray value and the local maximum brightness gradient of each frame of image are integrated to obtain a key feature representing the brightness change of the RH tank bottom preprocessed image.

4. The method of claim 1, wherein, The step S103 comprises: The key feature is arranged one by one in the spatial position of the original image to generate a two-dimensional feature image with a spatial structure; The two-dimensional feature image is input into the molten steel positioning intelligent model to output the probability of the RH molten steel reaching the bottom of the RH vacuum tank; The brightness slope and the brightness amplitude increment of the average brightness between adjacent frames are calculated; When the probability of the molten steel reaching the bottom of the RH vacuum tank is greater than or equal to a preset first threshold value, the brightness slope is greater than a preset slope threshold value, and the brightness amplitude increment is less than a preset amplitude threshold value, it is preliminarily determined that the RH molten steel has entered the RH vacuum tank and covered the bottom thereof, and the time when the preliminary determination is made is recorded as the time when the RH molten steel enters the RH vacuum tank and covers the bottom thereof.

5. The method of claim 4, wherein, The molten steel positioning intelligent model in the step S103 comprises a feature extraction module, a time series analysis module, and a determination output module, and each module cooperates to realize the calculation of the probability of the RH molten steel reaching the bottom of the RH vacuum tank, and specifically as follows: The feature extraction module is constructed by using a convolutional neural network, inputs the two-dimensional feature image, and outputs a low-dimensional feature vector representing the spatial features of the molten steel region. The convolutional neural network includes an input layer, at least 4 convolutional layers, at least 4 pooling layers, and 1 fully connected layer. The time series analysis module is constructed by using a long short-term memory network, inputs the low-dimensional feature vector, and outputs a time series optimized feature vector. The long short-term memory network includes at least 2 LSTM units, 1 dropout layer, and 1 fully connected layer. The determination output module is constructed by using a sigmoid activation function to build a binary classification output layer, which maps the time series optimized feature vector to a probability value between 0 and 1. This probability value is the probability of RH molten steel reaching the bottom of the RH vacuum tank.

6. The method of claim 4, wherein, After the preliminary determination that the RH molten steel has entered the RH vacuum tank and covered its bottom in step S103, a molten steel liquid level stability verification step is added, which includes: Calculating the brightness fluctuation standard deviation and the brightness change mean value within a continuous preset number of frames after the preliminary determination time; Determining whether both conditions, i.e., the brightness fluctuation standard deviation being less than a preset fluctuation threshold and the brightness change mean value being less than a preset change threshold, are met; If yes, the determination time is confirmed as the final time when the RH molten steel enters the RH vacuum tank and covers its bottom; If no, the brightness slope and brightness amplitude increment between adjacent frames are recalculated, and the molten steel arrival probability is used for re-determination until a determination time that meets the stability requirements is obtained.

7. The method of claim 1, wherein, The step S104 includes: After the time when the RH molten steel is determined to spread on the bottom of the RH vacuum vessel, the pressure detecting device arranged in the RH vacuum vessel reads the vacuum pressure in the RH vacuum vessel at this time ; acquiring a preset length of the immersion tube , molten steel density , and slag density , and reading a current slag thickness in the ladle ; based on the vacuum pressure , the length of the immersion tube , the density of the liquid steel , the density of the slag , and the thickness of the slag , the depth H of the RH immersion tube inserted into the ladle liquid steel is calculated by the formula , where the length unit is m, the density unit is t / m³, and the pressure unit is Pa.

Images