Automatic continuous casting pouring method and system based on intelligent visual detection

Abstract

The application provides a continuous casting automatic casting method and system based on intelligent visual detection, comprising the following steps: S1, collecting video data of molten steel flowing to a tundish through an industrial camera; S2, cutting and positioning a target region to be identified, and then preprocessing video data of the target region; S3, extracting molten steel splashing features based on the video data of the target region, judging whether the target region appears a molten steel splashing probability through a deep separable convolution, if the result is yes, executing step S4, and if the result is no, not processing; S4, detecting the molten steel splashing in a preset time window, and judging whether the molten steel splashing appearance probability and duration reach a preset threshold value at the same time, if the result is yes, triggering a casting signal, and if the result is no, re-executing step S1. The application realizes intelligent upgrading of the continuous casting machine casting operation, avoids molten steel temperature loss caused by delayed casting, and guarantees smooth continuous casting production of special steel.

Description

Technical Field

[0001] This invention belongs to the field of intelligent visual recognition technology, specifically relating to an automatic casting start method and system based on intelligent visual detection. Background Technology

[0002] The start-up operation of the continuous casting machine is a crucial step in the entire continuous casting process. Specifically, it refers to the process by which molten steel travels from the ladle to the pouring platform and is injected into the crystallizer. During this process, each step must be completed quickly and stably to reduce the temperature loss of the molten steel and prevent equipment failure during start-up.

[0003] Currently, the commonly used method in China, which uses ladle weighing to assist manual casting, suffers from limited detection accuracy and poor real-time performance due to the high-temperature environment. This can easily cause a ton-level delay in determining the casting start time, seriously affecting the accuracy of billet separation for different steel grades and the quality of the cast billets.

[0004] In the context of pursuing higher production efficiency of high-quality steel, accurately and timely determining the timing of pouring from the ladle is of paramount importance.

[0005] Patent document CN 110340322 B introduces a complex machine learning model and multi-dimensional data input, integrating multiple key parameters such as molten steel level, tundish nozzle opening, steel type, molten steel temperature, and casting speed. This solution optimizes multi-parameter collaborative control, incorporates on-site equipment information to reduce false alarms caused by other brightness changes outside the casting start time, and is simple and low-cost to implement, achieving an accuracy rate of 99.9% for casting start time. It improves upon the traditional method of relying solely on weight changes to determine casting start time in specific scenarios, and is easier to modify and deploy on existing facilities, making it particularly suitable for steel plant automation upgrades. However, this solution requires the simultaneous integration of data from five types of sensors, including molten steel level, nozzle opening, and temperature; failure of any sensor will lead to system misjudgment. The model needs to be retrained for different steel types, limiting its generalization ability; and the existing weighing system needs to be modified and a temperature sensor array added, resulting in high implementation costs.

[0006] Patent document CN117830904A discloses a method for accurately determining the ladle pouring start time based on image recognition. This method involves installing camera hardware in the casting area to acquire video streams, extracting single-frame images from the acquired video streams for recognition, cropping an image of the ladle impact zone directly below the long nozzle, and applying Gaussian blur to the selected area to remove noise. The selected area is then binarized, and the steel spatter and brightness in the binarized image are used for target detection. The brightness and steel spatter of the current frame are compared with the previous frame; if the increase exceeds a threshold and persists for 0.5 seconds, the ladle pouring start time is determined. This method can accurately and promptly determine the ladle pouring start time during continuous casting, significantly improving the accuracy of billet separation in multi-furnace continuous casting conditions, bringing substantial economic benefits to enterprises. However, this method cannot significantly reduce the dependence on manual experience parameters, has low accuracy, and poor anti-interference capabilities. This problem urgently needs to be solved. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the purpose of this invention is to provide an automatic casting start method and system for continuous casting based on intelligent visual inspection.

[0008] An automatic casting start method based on intelligent vision detection according to the present invention includes:

[0009] Step S1: Collect video data of the molten steel flowing to the ladle using an industrial camera;

[0010] Step S2: Cropping and locating the target region to be identified, and then preprocessing the video data of the target region;

[0011] Step S3: Based on the video data of the target area, extract the molten steel splash feature, and determine the probability of whether steel flowers appear in the target area through depthwise separable convolution. If the result is yes, proceed to step S4; if the result is no, do not process.

[0012] Step S4: Detect steel sparks within a preset time window and determine whether the probability of steel sparks appearing and the duration of steel sparks simultaneously reach a preset threshold. If the result is yes, trigger the pouring signal; if the result is no, repeat step S1.

[0013] Preferably, it further includes: step S0: real-time monitoring and judgment of the ladle, i.e. whether the ladle has rotated to the target position and the sleeve of the continuous casting machine has been installed. If the result is yes, the ladle sliding gate is opened to allow the molten steel to flow out, and then step S1 is executed. If the result is no, step S0 is executed again.

[0014] In step S2, the target region, i.e. the region of interest, is abbreviated as ROI;

[0015] Let (ROI) xROI y ) are the x and y coordinates of the top left corner of the ROI, respectively. w ROI h Given the width and height of the ROI, respectively, the mathematical expression for ROI clipping is:

[0016] I ROI (x′,y′)=I(x+ROI x ,y+ROI y ),x′∈[0,ROI w ],y′∈[0,ROI h ]

[0017] Among them, I ROI (x′,y′) represents the cropping result, i.e., the target area; x′ represents the x-coordinate of the target area, y′ represents the y-coordinate of the target area, I represents the input original image, x represents the x-coordinate of the original image, and y represents the y-coordinate of the original image.

[0018] Preferably, in step S2, the preprocessing includes noise reduction and sharpening.

[0019] The noise reduction process is essentially median filtering of the video data.

[0020] The mathematical expression for the gray value g′(x,y) of the pixel at point (x,y) after median filtering is:

[0021] g′(x,y)=Median{f(xw,yh),...,f(x+w,y+h)}

[0022] Where f(x,y) represents the gray value of the original image at point (x,y), the Median function calculates the median of the gray values ​​of all pixels in the neighborhood around point (x,y), and (w,h) represents the width and height of the neighborhood, i.e. the size of the filtering window;

[0023] The image sharpening mentioned above refers to image sharpening using the Laplacian operator; the mathematical expression for the Laplacian operator is:

[0024]

[0025] in, Let I(x,y) represent the Laplacian operator, where I(x,y) is the grayscale value or color channel value of the image at coordinate position (x,y), and... and Let represent the second-order partial derivatives of the image in the x and y directions, respectively.

[0026] Preferably, step S3 includes:

[0027] Step S3.1: Extract and obtain the multidimensional feature map of the video data of the target region using a convolutional neural network;

[0028] Step S3.2: Extract one-dimensional feature vectors from the multi-dimensional feature map through the fully connected layer of the MobileNetV2 model, and then map them to the class space;

[0029] Step S3.3: Convert the one-dimensional feature vector into a normalized probability distribution through the Softmax layer of the MobileNetV2 model; output the probability of 'splashing' and the probability of 'no splashing';

[0030] Step S3.4: Determine whether the probability of sparks flying, i.e., the probability of 'sparks flying', exceeds a preset threshold. If the result is yes, proceed to step S4; if the result is no, do not process.

[0031] In step S3.1, the convolutional neural network includes: depthwise convolution and pointwise convolution;

[0032] The mathematical expression for the depthwise convolution is:

[0033]

[0034] Where I represents the input feature map, C in w is the number of channels in the input feature map. i It is the i-th depth convolution kernel, O d It is the output feature map after depthwise convolution. DepthwiseConv represents channel-wise spatial convolution. i b represents the i-th channel slice of the input feature map I; b represents the bias term;

[0035] The mathematical expression for pointwise convolution is:

[0036]

[0037] Among them, C dw C is the number of output channels after depthwise convolution. out w is the number of channels output by pointwise convolution. ij It is a pointwise convolution kernel, O di It is the i-th channel of the feature map output by the depthwise convolution, b j It is the bias term of the j-th output channel, O p This is the final output feature map after pointwise convolution; PointwiseConv represents pointwise convolution; the symbol * represents multiplication.

[0038] In step S3.2, the mathematical expression for the one-dimensional feature vector is:

[0039] z = W fc ·X flatten +b fc

[0040] Where z represents a one-dimensional eigenvector, W fc Represents the weight matrix. b fc Indicates the bias term. symbol X flatten This represents the flattened feature input. in, Let c be the set of real numbers, c be the number of categories, and d be the dimension of the flattened vector.

[0041] In step S3.3, the mathematical expression for the normalized probability distribution is:

[0042]

[0043] Where, p i Let represent the probability that the input image belongs to the i-th class, i.e., the probability of 'splashing' and the probability of 'no splashing', satisfying . That is, p i ∈[0,1], i=1,2, where i=1 corresponds to the probability of 'splashing' and i=2 corresponds to the probability of 'no splashing'; This indicates that the score of the i-th category is subjected to an exponential operation; This indicates that the score of the j-th category is subjected to an exponential operation.

[0044] Preferably, in step S4, it is determined whether there is a probability of molten steel splashing greater than a preset probability threshold within a consecutive preset number of seconds, and whether the duration of molten steel splashing reaches a preset time threshold; if the result is yes, a pouring start signal is issued and the event timestamp is recorded synchronously; if the result is no, step S1 is executed again.

[0045] The probability of sparks flying is 0.9;

[0046] The preset time threshold is 1 second.

[0047] An automatic casting start system based on intelligent vision detection, according to the present invention, includes:

[0048] Imaging module: Collects video data of the molten steel flowing to the ladle using an industrial camera;

[0049] Image processing module: Crops and locates the target area to be identified, and then preprocesses the video data of the target area;

[0050] Image recognition module: Based on the video data of the target area, extract the molten steel splash feature, and determine the probability of whether steel flowers appear in the target area through depthwise separable convolution. If the result is yes, then execute step S4; if the result is no, then do not process.

[0051] Control unit module: Detects steel sparks within a preset time window and determines whether the probability and duration of steel spark appearance simultaneously reach a preset threshold. If the result is yes, it triggers the pouring start signal; if the result is no, it re-triggers the imaging module to work.

[0052] Preferably, it also includes: a preparatory module: which monitors and determines in real time whether the ladle has rotated to the target position and whether the sleeve of the continuous casting machine has been installed. If the result is yes, the ladle sliding gate is opened to allow molten steel to flow out, thereby triggering the imaging module to work. If the result is no, the preparatory module is triggered again.

[0053] In the image recognition module, the target region, i.e. the region of interest, is abbreviated as ROI;

[0054] Let (ROI) x ROI y ) are the x and y coordinates of the top left corner of the ROI, respectively. w ROI h Given the width and height of the ROI, respectively, the mathematical expression for ROI clipping is:

[0055] I ROI (x′,y′)=I(x+ROI x ,y+ROI y ),x′∈[0,ROI w ],y′∈[0,ROI h ]

[0056] Among them, I ROI (x′,y′) represents the cropping result, i.e., the target area; x′ represents the x-coordinate of the target area, y′ represents the y-coordinate of the target area, I represents the input original image, x represents the x-coordinate of the original image, and y represents the y-coordinate of the original image.

[0057] Preferably, in the image processing module, the preprocessing includes noise reduction and sharpening.

[0058] The noise reduction process is essentially median filtering of the video data.

[0059] The mathematical expression for the gray value g′(x,y) of the pixel at point (x,y) after median filtering is:

[0060] g′(x,y)=Median{f(xw,yh),...,f(x+w,y+h)}

[0061] Where f(x,y) represents the gray value of the original image at point (x,y), the Median function calculates the median of the gray values ​​of all pixels in the neighborhood around point (x,y), and (w,h) represents the width and height of the neighborhood, i.e. the size of the filtering window;

[0062] The image sharpening mentioned above refers to image sharpening using the Laplacian operator; the mathematical expression for the Laplacian operator is:

[0063]

[0064] in, Let I(x,y) represent the Laplacian operator, where I(x,y) is the grayscale value or color channel value of the image at coordinate position (x,y), and... and Let represent the second-order partial derivatives of the image in the x and y directions, respectively.

[0065] Preferably, the image recognition module includes:

[0066] Image recognition unit 1: Extracts and obtains a multi-dimensional feature map of the video data of the target region through a convolutional neural network;

[0067] Image recognition unit 2: Extracts one-dimensional feature vectors from the multi-dimensional feature map through the fully connected layer of the MobileNetV2 model, and then maps them to the category space;

[0068] Image recognition unit three: Converts the one-dimensional feature vector into a normalized probability distribution through the Softmax layer of the MobileNetV2 model; outputs the probability of 'splashing' and the probability of 'no splashing';

[0069] Image recognition unit four: Determine whether the probability of sparks flying, i.e., the probability of 'sparks flying', exceeds a preset threshold. If the result is yes, then execute step S4; if the result is no, then do not process.

[0070] In the image recognition unit one, the convolutional neural network includes: depthwise convolution and pointwise convolution;

[0071] The mathematical expression for the depthwise convolution is:

[0072]

[0073] Where I represents the input feature map, C in w is the number of channels in the input feature map. i It is the i-th depth convolution kernel, O d It is the output feature map after depthwise convolution. DepthwiseConv represents channel-wise spatial convolution.i b represents the i-th channel slice of the input feature map I; b represents the bias term;

[0074] The mathematical expression for pointwise convolution is:

[0075]

[0076] Among them, C dw C is the number of output channels after depthwise convolution. out w is the number of channels output by pointwise convolution. ij It is a pointwise convolution kernel, O di It is the i-th channel of the feature map output by the depthwise convolution, b j It is the bias term of the j-th output channel, O p This is the final output feature map after pointwise convolution; PointwiseConv represents pointwise convolution; the symbol * represents multiplication.

[0077] In the image recognition unit two, the mathematical expression for the one-dimensional feature vector is:

[0078] z = W fc ·X flatten +b fc

[0079] Where z represents a one-dimensional eigenvector, W fc Represents the weight matrix. b fc Indicates the bias term. symbol X flatten This represents the flattened feature input. in, Let c be the set of real numbers, c be the number of categories, and d be the dimension of the flattened vector.

[0080] In the image recognition unit three, the mathematical expression for the normalized probability distribution is:

[0081]

[0082] Where, p i Let represent the probability that the input image belongs to the i-th class, i.e., the probability of 'splashing' and the probability of 'no splashing', satisfying . That is, p i ∈[0,1], i=1,2, where i=1 corresponds to the probability of 'splashing' and i=2 corresponds to the probability of 'no splashing'; This indicates that the score of the i-th category is subjected to an exponential operation; This indicates that the score of the j-th category is subjected to an exponential operation.

[0083] Preferably, in the control unit module, it is determined whether the probability of all steel sparks splashing is greater than a preset probability threshold within a continuous preset number of seconds, and whether the duration of steel sparks splashing reaches a preset time threshold; if the result is yes, a pouring start signal is issued and the event timestamp is recorded synchronously; if the result is no, the imaging module is re-triggered to work.

[0084] The probability of sparks flying is 0.9;

[0085] The preset time threshold is 1 second.

[0086] Compared with the prior art, the present invention has the following beneficial effects:

[0087] 1. This invention uses intelligent visual recognition technology to replace the traditional method of determining the start of casting based on changes in the weight of the ladle, achieving accurate and real-time detection of the start of casting on the continuous casting machine. This significantly improves the automation level and safety of continuous casting production and solves the problems of billet separation confusion and billet quality fluctuation caused by weighing delay.

[0088] 2. This invention replaces the traditional weighing method with intelligent visual recognition technology, which solves the problem of misjudging the start-up time caused by the low accuracy and poor real-time performance of weighing equipment under high temperature conditions, and improves the success rate of the start-up operation and the accuracy of billet separation.

[0089] 3. This invention realizes the intelligent upgrade of the continuous casting machine's start-up operation, avoids the loss of molten steel temperature caused by start-up delay, and ensures the smooth progress of continuous casting production of special steel grades. Attached Figure Description

[0090] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0091] Figure 1 This is a schematic diagram of the continuous casting machine's start-up operation provided by the present invention;

[0092] Figure 2 This is a schematic diagram of the workflow provided by the present invention. Detailed Implementation

[0093] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0094] The present invention aims to provide a novel automatic casting method based on intelligent visual detection, so as to make up for the error of traditional weighing to determine the casting time, and improve the success rate of one-time casting and the accuracy of blank division.

[0095] According to the present invention, an automatic casting start method based on intelligent vision detection is provided, the software system flow includes:

[0096] Step 0: Monitor and determine whether the ladle has rotated to the target position and the sleeve of the continuous casting machine has been installed. If the result is yes, open the ladle sliding gate to allow the molten steel to flow out, and then execute Step 1. If the result is no, then execute Step 0 again.

[0097] Step 1: Collect video data of the molten steel flowing to the ladle using an industrial camera;

[0098] Step 2: Cropping and locating the target area to be identified, and then preprocessing the video data of the target area;

[0099] Step 3: Analyze the video data of the target area based on the convolutional neural network to determine whether there is a spark splatter phenomenon. If the result is yes, proceed to step 4; if the result is no, do not process it.

[0100] Step 4: Determine whether there is a continuous splattering event in the target area. If yes, issue a pouring start signal; if no, repeat step 2.

[0101] Step S1: The control module monitors the signal of the large package rotating into position in real time and confirms that the sleeve installation is complete. If the result is yes, then proceed to step S2; if the result is no, then re-execute step S1.

[0102] Step S2: The control module receives the command to open the sliding gate of the large package. If the result is yes, then step S3 is executed. If the result is no, then step S2 is executed again.

[0103] Step S3: The industrial camera starts video acquisition and continuously captures the dynamic process of molten steel flowing to the tundish;

[0104] Step S4: Locate the boundary of the bag opening, crop and locate the target area to be identified, and then preprocess the video data of the target area;

[0105] Step S5: Extract molten steel splash features based on the MobileNetV2 model, and determine the probability of steel flowers appearing in the target region through depthwise separable convolution;

[0106] Step S6: Detect steel sparks within a preset time window. When the probability and duration of steel sparks both reach the threshold, trigger the pouring signal and record the event timestamp and detection parameters simultaneously. If the result is negative, repeat step S3.

[0107] In step S2, the target area to be identified is cropped and located, that is, the boundary of the middle bag opening is accurately identified by the edge detection Canny algorithm, and then the region of interest (ROI) is determined.

[0108] Let (ROI) x ROI y ) are the x and y coordinates of the top left corner of the ROI, respectively. w ROI h Let ) represent the width and height of the ROI, respectively. Then the ROI clipping formula is:

[0109] I ROI (x′,y′)=I(x+ROI x ,y+ROI y ), for x′∈[0,ROI w ],y′∈[0,ROI h ]

[0110] Among them, I ROI (x′,y′) represents the cropping result, i.e., the target area; x′ represents the x-coordinate of the target area, y′ represents the y-coordinate of the target area, I represents the input original image, x represents the x-coordinate of the original image, and y represents the y-coordinate of the original image;

[0111] Specifically, in step S2, the preprocessing includes noise reduction and sharpening.

[0112] The noise reduction process involves using a two-dimensional median filter. Median filtering is a non-linear filtering method that can effectively remove salt-and-pepper noise and speckle noise without affecting image edge details.

[0113] The original image is often accompanied by noise, including quantum noise and readout noise.

[0114] The mathematical expression for the gray value g′(x,y) of the pixel at point (x,y) after median filtering is:

[0115] g′(x,y)=Median{f(xw,yh),...,f(x+w,y+h)}

[0116] Where f(x,y) represents the gray value of the original image at point (x,y), the Median function calculates the median of the gray values ​​of all pixels in the neighborhood around point (x,y), and (w,h) represents the width and height of the neighborhood, i.e. the size of the filtering window.

[0117] The image sharpening and detail enhancement employs the Laplacian operator to increase the edge sharpness of the steel sparks in the spark splatter area. The Laplacian operator is expressed as follows:

[0118]

[0119] in, Let I(x,y) represent the Laplacian operator, where I(x,y) is the grayscale value or color channel value of the image at coordinate position (x,y), and... and These represent the second-order partial derivatives of the image in the x and y directions, respectively. To ensure sensitivity to changes in the diagonal direction, an extended convolution kernel is used:

[0120]

[0121] Specifically, in step S3, the MobileNetV2 deep learning model is used as the core algorithm for image feature extraction and recognition. The MobileNetV2 model is built on the basis of depthwise separable convolution, which is a lightweight and computationally efficient form of convolution, designed to reduce the number of model parameters while maintaining high recognition accuracy.

[0122] Depthwise separable convolution consists of two parts: depthwise convolution and pointwise convolution. Depthwise convolution performs the convolution operation independently on each input channel, and its mathematical expression is:

[0123]

[0124] Where I represents the input feature map, C in w is the number of channels in the input feature map. i It is the i-th depth convolution kernel, O d It is the output feature map after depthwise convolution. DepthwiseConv represents channel-wise spatial convolution. i b represents the i-th channel slice of the input feature map I; b represents the bias term;

[0125] Then, pointwise convolution is performed, which integrates the information from all channels. The formula is expressed as:

[0126]

[0127] Among them, C dw C is the number of output channels after depthwise convolution. out w is the number of channels output by pointwise convolution. ij It is a pointwise convolution kernel, O di It is the i-th channel of the feature map output by the depthwise convolution, b j It is the bias term of the j-th output channel, Op It is the final output feature map after pointwise convolution, where PointwiseConv represents pointwise convolution.

[0128] The MobileNetV2 model further introduces learnable shortcut connections, achieving effective extraction and recognition of splatter features through a series of deep separable convolutions, bottleneck layers, and shortcut connection layers. The intelligent recognition module inputs the processed image features into the model, and the model finally outputs probability distributions for two categories—with and without splatter—through fully connected layers and Softmax layers to determine whether splatter phenomena exist and trigger the pouring command.

[0129] Specifically, a fully connected layer and a softmax layer are added after the convolutional neural network to obtain the probability distribution of whether sparks will fly.

[0130] Specifically, it is converted into a normalized probability distribution. The presence of sparks is then determined by combining this with the statement in the detailed description below that "the probability of sparks splashing in each model output exceeds the threshold."

[0131] In the example, the probability of having sparks is 0.8, and the probability of not having sparks is 0.2.

[0132] Fully connected layers flatten the multidimensional feature maps extracted by convolutional networks into one-dimensional feature vectors, and then map them to the class space through a linear transformation:

[0133] z = W fc ·X flatten +b fc

[0134] Where z represents a one-dimensional eigenvector. Represents the weight matrix. Indicates the bias term, symbol This represents the flattened feature input, where c is the number of categories and d is the dimension of the flattened vector; the symbol · represents the dot product.

[0135] The Softmax layer transforms the output z of the fully connected layer into a normalized probability distribution:

[0136]

[0137] Wherein, the output p i ∈[0,1] represents the probability that the input image belongs to the i-th class, satisfying

[0138] Specifically, the preset threshold for the probability of steel sparks splashing is the probability of the 'sparks present' category, which is set to 0.9; when the probability of 'sparks present' is continuously detected to be greater than or equal to 0.9, it is considered a valid splashing event;

[0139] The duration of steel spatter is usually within 1 second. If the duration of a valid spatter event is continuously detected within this range, it is determined that the pouring has started.

[0140] The two thresholds mentioned above are manually adjusted based on the on-site process and imaging conditions.

[0141] The control unit first monitors the ladle's rotation into position and confirms the casing is complete, then receives the ladle nozzle opening command. The camera begins capturing video footage of the molten steel flowing into the tundish. The image recognition module analyzes the video stream in real time according to preset logic. When it detects a continuous bright flash in the target area that matches the characteristics of pouring commencement, the control module immediately responds, issues a pouring commencement signal, and records this event in the designated business database.

[0142] The intelligent vision inspection system includes an imaging module, an image processing module, an image recognition module, and a control unit.

[0143] Imaging module: Utilizes existing industrial cameras in the casting area to capture image data of the moment the ladle opens for casting at key nodes during the continuous casting process.

[0144] Image processing module: performs real-time analysis on the received continuous video stream, performs precise cropping of the central port area, and locates the target area to be identified.

[0145] Image recognition module: By analyzing cropped image segments using a convolutional neural network, it identifies the presence of continuous spark splatter, serving as a basis for determining when to start casting. Specifically, in the intelligent vision recognition automatic casting system for continuous casting machines based on image recognition technology, the image recognition module employs the MobileNetV2 deep learning model as the core algorithm for image feature extraction and recognition. The MobileNetV2 model is built upon depthwise separable convolution, a lightweight and computationally efficient form of convolution designed to reduce the number of model parameters while maintaining high recognition accuracy.

[0146] Control Unit: Receives visual recognition results. When the model output changes from no bright flame to a state of continuous bright flame, the system is triggered to automatically determine the time for the ladle to start pouring and sends relevant information to the continuous casting production line control system to guide the stable start of subsequent processes.

[0147] The imaging module, in the intelligent vision recognition automatic casting system for continuous casting machines based on image recognition technology, utilizes existing industrial cameras in the casting area to capture image data of the moment the ladle begins casting during the continuous casting process. The first challenge for the camera module is hardware selection. Considering the harsh environment of the continuous casting workshop—high temperature, dust, and strong electromagnetic interference—the selected camera should possess the following characteristics:

[0148] The camera should be heat resistant; specifically, the camera housing material should be able to withstand long-term high-temperature environments, usually made of metal or special plastic, to ensure normal operation even at high temperatures.

[0149] The camera must meet optical performance requirements; specifically, the lens must have high light transmittance and resolution, with a resolution of no less than 1080P, to ensure that the captured image clarity is sufficient for subsequent image processing and recognition.

[0150] The camera is designed to meet low latency requirements; specifically, the camera design supports a low latency mode, ensuring that the latency from data acquisition to network output does not exceed 40 milliseconds, thus ensuring the immediacy and accuracy of the judgment of the start time of continuous casting.

[0151] The camera is dustproof and waterproof; specifically, the camera should have an IP65 protection rating to ensure it is not damaged in dusty environments.

[0152] In the intelligent vision recognition automatic casting system for continuous casting machines based on image recognition technology, the image preprocessing module employs a variety of key technologies to perform a series of processing steps on the raw images captured by the industrial camera in the imaging module, in order to optimize image quality, highlight feature information, and prepare for subsequent intelligent recognition.

[0153] To better extract the features of splattered steel sparks, the image preprocessing module converts the original RGB color space to the HSV color space. The brightness information in the HSV color space better reflects the difference between the splattered steel sparks and the background. The following are the key formulas in the conversion process:

[0154] The mathematical expression for standardizing RGB values ​​from the range [0,255] to [0,1] is:

[0155]

[0156] Calculate C max C min The mathematical expressions for Δ are as follows, from top to bottom:

[0157] C max =max(R′,G′,B′)

[0158] C min =min(R′,G′,B′)

[0159] Δ=C max -C min

[0160] Among them, C max C represents the maximum color value. minΔ represents the minimum color value, and Δ represents the chromaticity span between the maximum and minimum values ​​of the three channels (R', G', B') in the RGB color space.

[0161] Hue is calculated based on the maximum color value, and the mathematical expression is:

[0162]

[0163] Here, H represents hue.

[0164] The mathematical expression for calculating saturation is:

[0165]

[0166] Where S represents saturation;

[0167] Brightness is calculated based on the maximum color value, and the mathematical expression is:

[0168] V = C max

[0169] Where V represents brightness.

[0170] In the intelligent vision recognition automatic casting start system of the continuous casting machine, the control unit adopts the following strategy to determine the casting start time:

[0171] 1. Continuous monitoring and recognition results: The control unit continuously receives and processes the probability sequence of steel sparks splashing output by the image recognition module for continuous image frames, denoted as: P 钢花 (t), t=1,2,…

[0172] 2. Set a recognition threshold: Set a preset threshold θ. Only when the probability of steel spark splashing detected by the model exceeds this threshold is the model considered to have detected a possible steel spark splashing phenomenon: P 钢花 (t)>θ

[0173] 3. Time Window Judgment: Set a time window T. If the probability of sparks splashing in each model output exceeds the threshold θ within a consecutive T seconds, then it is determined that there is a continuous sparks splashing phenomenon during this period. P 钢花 (t)>θ; where, middle, and These represent existence and randomness, respectively.

[0174] 4. Determining the start time of casting: When the probability of molten steel splashing continuously exceeds the threshold for a given period of time, the time point t0 at which the condition is first met is defined as the start time of casting. At this time, the control unit uses this time point as the key time node to trigger the start operation of the ladle casting on the continuous casting machine and records it in the real-time database on site.

[0175] The present invention provides an automatic continuous casting start system based on intelligent vision detection, comprising: a control unit that monitors the ladle rotation position signal in real time and confirms that the sleeve is completed, monitors the ladle nozzle opening signal, and starts the automatic continuous casting start system after receiving the ladle nozzle opening signal.

[0176] After receiving the start signal from the control unit, the imaging module uses the existing industrial camera in the casting area to continuously capture image data of the continuous casting process.

[0177] The image processing module performs real-time analysis on the continuous video stream captured by the imaging module, accurately crops the central port area, locates the target area to be identified, and performs image processing to improve image quality.

[0178] The image recognition module uses a convolutional neural network to analyze the processed image fragments to identify whether there is a spark splash phenomenon, and outputs the category information and probability of the judgment.

[0179] The control unit determines whether the conditions for triggering the pouring have been met based on the steel splash feature information and its duration fed back by the identification module, and records the pouring time into the on-site real-time database.

[0180] The present invention also provides an automatic continuous casting start system based on intelligent vision detection. The automatic continuous casting start system based on intelligent vision detection can be implemented by executing the process steps of the automatic continuous casting start method based on intelligent vision detection. That is, those skilled in the art can understand the automatic continuous casting start method based on intelligent vision detection as a preferred embodiment of the automatic continuous casting start system based on intelligent vision detection.

[0181] An automatic casting start system based on intelligent vision detection, according to the present invention, includes:

[0182] Imaging module: Acquires video data of the molten steel flowing to the tundish;

[0183] Image processing module: Crops and locates the target area to be identified, and then preprocesses the video data of the target area;

[0184] Image recognition module: Based on the video data of the target area, extract the molten steel splash feature, and determine the probability of whether steel flowers appear in the target area through depthwise separable convolution. If the result is yes, then execute step S4; if the result is no, then do not process.

[0185] Control unit module: Detects steel sparks within a preset time window and determines whether the probability and duration of steel spark appearance simultaneously reach a preset threshold. If the result is yes, it triggers the pouring start signal; if the result is no, it re-triggers the imaging module to work.

[0186] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, enabling the system and its various devices, modules, and units to function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered both software modules implementing the method and structures within the hardware component.

[0187] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. An automated casting start method based on intelligent vision inspection, characterized in that, include: Step S1: Collect video data of the molten steel flowing to the tundish; Step S2: Crop and locate the target area to be identified, and then preprocess the video data of the target area; Step S3: Based on the video data of the target area, extract the molten steel splash feature, and determine the probability of whether steel flowers appear in the target area through depthwise separable convolution. If the result is yes, proceed to step S4; if the result is no, do not process. Step S4: Detect steel sparks within a preset time window and determine whether the probability of steel sparks appearing and the duration of steel sparks simultaneously reach a preset threshold. If the result is yes, trigger the pouring signal; if the result is no, repeat step S1. It also includes: Step S0: Real-time monitoring and judgment of the ladle, i.e. whether the ladle has rotated to the target position and the sleeve of the continuous casting machine has been installed. If the result is yes, the ladle sliding gate is opened to allow the molten steel to flow out, and then step S1 is executed. If the result is no, step S0 is executed again. In step S2, the target region, i.e. the region of interest, is abbreviated as ROI; make These are the x and y coordinates of the top left corner of the ROI, respectively. Let the width and height of the ROI be the values, respectively. Then, the mathematical expression for ROI clipping is: in, This indicates the cropping result, i.e., the target area; Represents the x-coordinate of the target region. Represents the ordinate of the target region. This represents the original image input. This represents the x-coordinate of the original graph. Represents the ordinate of the original graph; In step S2, the preprocessing includes noise reduction and sharpening. The denoising process is a median filter applied to the video data. point The grayscale value of the pixel after median filtering The mathematical expression is: in, Indicates the original image at points grayscale value at that location Functions represent points Calculate the median gray value of all pixels in the surrounding neighborhood. This represents the width and height of the neighborhood, i.e., the size of the filtering window; The image sharpening mentioned above refers to image sharpening using the Laplacian operator; the mathematical expression for the Laplacian operator is: in, Represents the Laplace operator. The image is at coordinate position The grayscale value or color channel value at that location, and and These represent the images in direction and Second-order partial derivatives in the direction; Step S3 includes: Step S3.1: Extract and obtain the multidimensional feature map of the video data of the target region using a convolutional neural network; Step S3.2: Extract one-dimensional feature vectors from the multi-dimensional feature map through the fully connected layer of the MobileNetV2 model, and then map them to the class space; Step S3.3: Convert the one-dimensional feature vector into a normalized probability distribution through the Softmax layer of the MobileNetV2 model; output the probability of splashing and the probability of no splashing; Step S3.4: Determine whether the probability of splashing exceeds a preset threshold. If the result is yes, proceed to step S4; if the result is no, do not process. In step S3.1, the convolutional neural network includes: depthwise convolution and pointwise convolution; The mathematical expression for the depthwise convolution is: in, Represents the input feature map, It is the number of channels in the input feature map. It is the first One deep convolutional kernel, It is the output feature map after depthwise convolution. This represents channel-wise spatial convolution. Represents the input feature map The One channel slice; Indicates the bias term; The mathematical expression for pointwise convolution is: in, It is the number of output channels after depthwise convolution. It is the number of channels output by pointwise convolution. It is a pointwise convolution kernel. It is the first feature map output by the depthwise convolution. One channel, It is the first The bias term for each output channel. It is the final output feature map after pointwise convolution. Represents pointwise convolution; symbol Indicates product; In step S3.2, the mathematical expression for the one-dimensional feature vector is: in, Represents a one-dimensional eigenvector. Represents the weight matrix. , Indicates the bias term. ,symbol This represents the flattened feature input. ,in, It is the set of real numbers; For the number of categories, The dimension of the flattened vector; In step S3.3, the mathematical expression for the normalized probability distribution is: in, Indicates that the input image belongs to the first... The probability of a class, i.e., the probability of splashing and the probability of not splashing, satisfies ,Right now , ,in, The probability of splashing is corresponding to the probability of splashing. The probability of no splashing; This indicates that the score of the i-th category is subjected to an exponential operation; This indicates that the score of the j-th category is subjected to an exponential operation.

2. The automatic casting start method based on intelligent vision detection according to claim 1, characterized in that, In step S4, it is determined whether there is a probability of molten steel splashing greater than a preset probability threshold within a consecutive preset number of seconds, and whether the duration of molten steel splashing reaches a preset time threshold; if the result is yes, a pouring start signal is issued and the event timestamp is recorded synchronously; if the result is no, step S1 is executed again. The probability of sparks flying is 0.9; The preset time threshold is 1 second.

3. An automatic casting start system based on intelligent vision inspection, characterized in that, include: Imaging module: Acquires video data of the molten steel flowing to the tundish; Image processing module: Crops and locates the target area to be identified, and then preprocesses the video data of the target area; Image recognition module: Based on the video data of the target area, extract the molten steel splash feature, and determine the probability of whether steel flowers appear in the target area through depthwise separable convolution. If the result is yes, then execute step S4; if the result is no, then do not process. Control unit module: Detects steel sparks within a preset time window and determines whether the probability of steel sparks appearing and the duration of steel sparks simultaneously reach a preset threshold. If the result is yes, the pouring start signal is triggered; if the result is no, the imaging module is re-triggered to work. It also includes: a preparatory module: real-time monitoring and judgment of the ladle, i.e. whether the ladle has rotated to the target position and the continuous casting machine sleeve has been installed. If the result is yes, the ladle sliding gate is opened to allow molten steel to flow out, thereby triggering the imaging module to work. If the result is no, the preparatory module is re-triggered. In the image recognition module, the target region, i.e. the region of interest, is abbreviated as ROI; make These are the x and y coordinates of the top left corner of the ROI, respectively. Let the width and height of the ROI be the values, respectively. Then, the mathematical expression for ROI clipping is: in, This indicates the cropping result, i.e., the target area; Represents the x-coordinate of the target region. Represents the ordinate of the target region. This represents the original image input. This represents the x-coordinate of the original graph. Represents the ordinate of the original graph; In the image processing module, the preprocessing includes noise reduction and sharpening. The denoising process is a median filter applied to the video data. point The grayscale value of the pixel after median filtering The mathematical expression is: in, Indicates the original image at points grayscale value at that location Functions represent points Calculate the median gray value of all pixels in the surrounding neighborhood. This represents the width and height of the neighborhood, i.e., the size of the filtering window; The image sharpening mentioned above refers to image sharpening using the Laplacian operator; the mathematical expression for the Laplacian operator is: in, Represents the Laplace operator. The image is at coordinate position The grayscale value or color channel value at that location, and and These represent the images in direction and Second-order partial derivatives in the direction; The image recognition module includes: Image recognition unit 1: Extracts and obtains a multi-dimensional feature map of the video data of the target region through a convolutional neural network; Image recognition unit 2: Extracts one-dimensional feature vectors from the multi-dimensional feature map through the fully connected layer of the MobileNetV2 model, and then maps them to the category space; Image recognition unit three: The one-dimensional feature vector is converted into a normalized probability distribution through the Softmax layer of the MobileNetV2 model; the probability of splashing and the probability of no splashing are output. Image recognition unit four: Determine whether the probability of splashing exceeds a preset threshold. If the result is yes, then execute step S4; if the result is no, then do not process. In the image recognition unit one, the convolutional neural network includes: depthwise convolution and pointwise convolution; The mathematical expression for the depthwise convolution is: in, Represents the input feature map, It is the number of channels in the input feature map. It is the first One deep convolutional kernel, It is the output feature map after depthwise convolution. This represents channel-wise spatial convolution. Represents the input feature map The One channel slice; Indicates the bias term; The mathematical expression for pointwise convolution is: in, It is the number of output channels after depthwise convolution. It is the number of channels output by pointwise convolution. It is a pointwise convolution kernel. It is the first feature map output by the depthwise convolution. One channel, It is the first The bias term for each output channel. It is the final output feature map after pointwise convolution. Represents pointwise convolution; symbol Indicates product; In the image recognition unit two, the mathematical expression for the one-dimensional feature vector is: in, Represents a one-dimensional eigenvector. Represents the weight matrix. , Indicates the bias term. ,symbol This represents the flattened feature input. ,in, It is the set of real numbers; For the number of categories, The dimension of the flattened vector; In the image recognition unit three, the mathematical expression for the normalized probability distribution is: in, Indicates that the input image belongs to the first... The probability of a class, i.e., the probability of splashing and the probability of not splashing, satisfies ,Right now , ,in, The probability of splashing is corresponding to the probability of splashing. The probability of no splashing; This indicates that the score of the i-th category is subjected to an exponential operation; This indicates that the score of the j-th category is subjected to an exponential operation.

4. The automatic casting start system based on intelligent vision inspection according to claim 3, characterized in that, In the control unit module, it is determined whether the probability of all steel sparks splashing is greater than a preset probability threshold within a continuous preset number of seconds, and whether the duration of steel sparks splashing reaches a preset time threshold; if the result is yes, a pouring start signal is issued and the event timestamp is recorded synchronously; if the result is no, the imaging module is re-triggered to work. The probability of sparks flying is 0.9; The preset time threshold is 1 second.

Images