Converter slagging period end point intelligent judgment method and system based on computer vision

Abstract

The invention discloses a method and a system for intelligently judging an end point of a slagging period of a converter based on computer vision, and the method comprises the following steps: collecting image data of a historical copper slag sample, preprocessing, and establishing a data set; constructing an intelligent prediction model, and training the intelligent prediction model by taking the data set as input to obtain a slagging end point intelligent prediction model; and inputting the image data of the copper slag sample collected in real time into the intelligent slagging end point prediction model for prediction to obtain a prediction result, and performing logic post-processing on the obtained prediction result to obtain the end point value of the converter in the slagging period. According to the intelligent judgment method for the end point of the slagging period of the converter, sampling-intelligent identification-display automation is realized, the labor cost is greatly reduced, and the production safety is improved; the prediction recall rate and accuracy of the slagging end point reach the level of copper smelting experts; the algorithm and the model are simple and the operation speed is fast.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision, and in particular to a method and system for intelligently judging the end point of a converter slagging period based on computer vision. Background Art

[0002] Currently, converter copper smelting operations are divided into two phases: slagging and copper production. During the slagging phase, the slagging endpoint must be determined to determine when to stop blowing. The accuracy of slagging endpoint determination determines the safety of blowing operations and the quality of blister copper, thus impacting both operational safety and production costs at the copper smelter. Improving the accuracy of slagging endpoint determination is crucial for reducing costs, improving quality, and enhancing operational safety in converter copper smelting.

[0003] Currently, the determination of the slag formation endpoint relies entirely on manual experience. Due to various factors, there are no successful cases of developing an intelligent slag formation endpoint determination system in the domestic industry. The accuracy of slag formation endpoint determination cannot be effectively guaranteed based on manual experience, which leads to the following industry pain points:

[0004] 1) Inaccurate judgment of the slag end point during the slag making period will cause over-blowing of slag and spraying of the furnace, seriously affecting production, and even burning equipment and threatening personal safety.

[0005] 2) If the slag is not blown thoroughly, the slag making will continue during the copper making period. There will be too much slag in the furnace during the copper making period, the liquid level will be high, and large slag will be splashed and trapped. In severe cases, the furnace will be sprayed, threatening the safety of people and equipment.

[0006] 3) There is a lot of bottom slag in the furnace, and it is necessary to arrange for screening and washing the furnace before normal blowing operations can be carried out. This process directly takes up a lot of normal operating time, resulting in excessive copper in the upstream flash furnace and reduced material.

[0007] 4) Excessive slag during the copper making period will also cause crude copper to be easily slag-carrying, affecting the smooth operation of subsequent anode furnace oxidation and slag removal. This will cause the anode furnace to require multiple oxidations or extend the oxidation time before the slag can be removed. This will affect the anode furnace operation rhythm and even cause furnace blockage.

[0008] 5) Excessive bottom ash in the furnace will also cause the matte to react violently with the oxidized bottom ash during the next furnace feeding process, generating a large amount of flue gas that enters the ring-collected desulfurization system, causing frequent peaks in the desulfurization tail emissions and even exceeding the instantaneous value, increasing environmental risks. The PS converter is rated as a Class B kiln by the domestic industry, and uncontrollable flue gas emission is also the main reason.

[0009] Therefore, it is necessary to propose an "intelligent judgment system for the end point of the converter slagging period" to address the above industry pain points. Summary of the Invention

[0010] The present invention discloses a method and system for intelligently judging the end point of the converter slagging period based on computer vision, so as to solve any of the above and other potential problems in the prior art.

[0011] This invention aims to achieve intelligent judgment of the slagging phase of the converter copper smelting process through advanced visual recognition technology and AI algorithms, replacing manual judgment based on experience. This approach addresses the issue of endpoint judgment relying on human experience, thereby improving copper smelting efficiency and reducing production costs, while also contributing to environmental protection and resource utilization. It also enhances the intelligence level of converter operations, achieving digitalization and intelligentization of the entire converter blowing production cycle.

[0012] The technical solution adopted by the present invention is: a method for intelligently judging the end point of the converter slag-making period based on computer vision, and the method specifically comprises the following steps:

[0013] S1) collecting image data of historical copper slag samples, preprocessing, and establishing a data set;

[0014] S2) constructing an intelligent prediction model, and using the data set of S1) as input to train the intelligent prediction model, thereby obtaining an intelligent prediction model for slag forming end point;

[0015] S3) inputting the image data of the copper slag sample collected in real time into the slag forming end point intelligent prediction model for prediction to obtain a prediction result, and performing logical post-processing on the obtained prediction result to obtain the converter slag forming period end point value.

[0016] Furthermore, the pre-processing steps in S1) are:

[0017] S1.1) First, the video is segmented and the effective area of ​​the video data containing the drill rod is segmented and screened;

[0018] S1.2) Extraction of the drill rod area;

[0019] S1.3) Correcting the drill rod region to obtain a drill rod image with a height of H and a width of W;

[0020] S1.4) Label the corrected drill rod image;

[0021] S1.5) Finally, the dataset is divided into training set, test set and validation set.

[0022] Furthermore, the specific steps of S1.2) are:

[0023] S1.21) converting the collected RGB three-channel image into a grayscale image;

[0024] S1.22) performing threshold segmentation on the grayscale image to convert the grayscale image into a binary image;

[0025] S1.23) Perform an erosion operation on the segmented area and mark different connected areas. Based on the characteristics of the original image of the brazing rod, the upper and lower connected areas are considered as the background area, and the remaining area is the brazing rod area. All pixel values ​​in the background area are marked as 0.

[0026] Furthermore, the step S1.3) is as follows:

[0027] S1.31) Compress the ordinates of the drill rod region into one dimension. That is, for each abscissa x in the drill rod region, find the mean ordinate of the drill rod pixels at that ordinate to obtain the coordinates of a set of points, i.e., each abscissa corresponds to a ordinate.

[0028] S1.32) For this set of points, use the least squares method to fit a straight line, which is the center line of the drill rod area; find the midpoint of this center line, which is the center point of the drill rod area;

[0029] S1.33) Determine the angle between the centerline of the drill rod and the horizontal line; this angle is the deflection angle of the drill rod.

[0030] S1.34) After the deflection angle and center point are obtained, the image is rotated and cropped according to the center point and the pre-set length and width.

[0031] Furthermore, the step S1.4) is as follows: Assume that there are m samplings in a slag-making period, and the sampling times are S1, S2, ..., S m-1 , the actual end time of the slag making period is S final , then the label of the drill rod image obtained by the method of claim 4 in the i-th sampling video is d i =S final -S i , the unit is minutes; if the number of drill rod images obtained as described above is greater than N, randomly select N drill rod images and put them into the data set.

[0032] Furthermore, the network structure of S2) works in collaboration with the backbone network and the output layer. The backbone network extracts the feature map of the input image through a convolutional neural network, and the softmax function of the output layer normalizes the feature map so that the prediction result has probabilistic interpretation. The specific steps are:

[0033] S2.1) Construct an artificial intelligence visual network model with a softmax activation function as the last layer, where the backbone network adopts a general network architecture and the output layer is a classification head with n categories, where the n categories correspond to a preset set of candidate time intervals {t0, t1, ..., t n-1}, output the probability distribution of each time interval category through the softmax activation function;

[0034] S2.2) Calculate the mathematical expectation value of the current time from the end point based on the probability distribution As the prediction time interval, where p i is the predicted probability of the i-th time interval, and the inverse of the variance of the probability distribution is used as the confidence C of the prediction result;

[0035] S2.3) The prediction time interval T output by the network is supervisedly trained using the mean square error loss function to optimize the model parameters, thereby obtaining an intelligent prediction model for the slag endpoint.

[0036] Further, the specific steps of S3) are:

[0037] The real-time collected video data is preprocessed to obtain several drill rod images. These images are used as input into the intelligent prediction model for slag formation endpoint, and the predicted value of the slag formation endpoint time interval corresponding to each frame of the drill rod image is output; the predicted values ​​of all drill rod images are analyzed through the post-processing module to generate the final distance to the slag formation endpoint time of the video sample.

[0038] Furthermore, the specific steps of the logic post-processing are:

[0039] The prediction results are grouped into N binary groups (T0, C0), (T1, C2), ..., (T N-1 , C N-1 ), the binary tuples are arranged from large to small according to the confidence, and the first G results are taken as the credible tuples; the prediction results in the credible tuples are averaged as the final prediction result of the sampled video.

[0040] Another object of the present invention is to provide a system for implementing the above-mentioned prediction method, the system comprising:

[0041] The data acquisition and processing unit is used to collect image data of historical copper slag samples, pre-process them, and establish a data set, while also collecting image data of copper slag samples in real time and pre-processing them;

[0042] The model training unit is used to build an intelligent prediction model and use the data set as input to train the intelligent prediction model to obtain an intelligent prediction model for slag forming end point;

[0043] The prediction unit is used to input the image data of the copper slag sample collected in real time into the slag end point intelligent prediction model to obtain the prediction result, and perform logical post-processing on the obtained prediction result to obtain the end point value of the converter slag forming period.

[0044] Furthermore, the system also includes an automatic sampling module for controlling the robotic arm holding the drill rod to automatically sample from the PS converter tuyere.

[0045] The beneficial effects of the present invention are as follows: due to the adoption of the above technical solution, the method of the present invention realizes the automation of sampling-intelligent identification-display, greatly reducing labor costs and improving production safety; the recall rate and accuracy of slag endpoint prediction reach the level of copper smelting experts; the algorithm and model are simple and run fast. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of an intelligent algorithm for determining the endpoint of the converter copper smelting slag forming period involved in an embodiment;

[0047] Figure 2 This is the logic block diagram of the intelligent judgment system for the end point of the converter slagging period based on computer vision. DETAILED DESCRIPTION

[0048] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0049] like Figure 1 As shown, the present invention

[0050] The intelligent judgment method for the end point of the converter slag-making period based on computer vision specifically comprises the following steps:

[0051] S1) collecting image data of historical copper slag samples, preprocessing, and establishing a data set;

[0052] S2) constructing an intelligent prediction model, and using the data set of S1) as input to train the intelligent prediction model, thereby obtaining an intelligent prediction model for slag forming end point;

[0053] S3) inputting the image data of the copper slag sample collected in real time into the slag forming end point intelligent prediction model for prediction to obtain a prediction result, and performing logical post-processing on the obtained prediction result to obtain the converter slag forming period end point value.

[0054] Furthermore, the pre-processing steps in S1) are:

[0055] S1.1) First, the video is segmented and the effective area of ​​the video data containing the drill rod is segmented and screened;

[0056] S1.2) Extraction of the drill rod area;

[0057] S1.3) Correcting the drill rod region to obtain a drill rod image with a height of H and a width of W;

[0058] S1.4) Label the corrected drill rod image;

[0059] S1.5) Finally, the dataset is divided into training set, test set and validation set.

[0060] Furthermore, the specific steps of S1.2) are:

[0061] S1.21) converting the collected RGB three-channel image into a grayscale image;

[0062] S1.22) performing threshold segmentation on the grayscale image to convert the grayscale image into a binary image;

[0063] S1.23) Perform an erosion operation on the segmented area and mark different connected areas. Based on the characteristics of the original image of the brazing rod, the upper and lower connected areas are considered as the background area, and the remaining area is the brazing rod area. All pixel values ​​in the background area are marked as 0.

[0064] Furthermore, the step S1.3) is as follows:

[0065] S1.31) Compress the ordinates of the drill rod region into one dimension. That is, for each abscissa x in the drill rod region, find the mean ordinate of the drill rod pixels at that ordinate to obtain the coordinates of a set of points, i.e., each abscissa corresponds to a ordinate.

[0066] S1.32) For this set of points, use the least squares method to fit a straight line, which is the center line of the drill rod area; find the midpoint of this center line, which is the center point of the drill rod area;

[0067] S1.33) Determine the angle between the centerline of the drill rod and the horizontal line; this angle is the deflection angle of the drill rod.

[0068] S1.34) After the deflection angle and center point are obtained, the image is rotated and cropped according to the center point and the pre-set length and width.

[0069] Furthermore, the step S1.4) is as follows: Assume that there are m samplings in a slag-making period, and the sampling times are S1, S2, ..., S m-1 , the actual end time of the slag making period is S final , then the label of the drill rod image obtained by the method of claim 4 in the i-th sampling video is d i =S final -S i , the unit is minutes; if the number of drill rod images obtained as described above is greater than N, randomly select N drill rod images and put them into the data set.

[0070] Furthermore, the network structure of S2) works in collaboration with the backbone network and the output layer. The backbone network extracts the feature map of the input image through a convolutional neural network, and the softmax function of the output layer normalizes the feature map so that the prediction result has probabilistic interpretation. The specific steps are:

[0071] S2.1) Construct an artificial intelligence visual network model with a softmax activation function as the last layer, where the backbone network adopts a general network architecture and the output layer is a classification head with n categories, where the n categories correspond to a preset set of candidate time intervals {t0, t1, ..., t n-1}, output the probability distribution of each time interval category through the softmax activation function;

[0072] S2.2) Calculate the mathematical expectation value of the current time from the end point based on the probability distribution As the prediction time interval, where p i is the predicted probability of the i-th time interval, and the inverse of the variance of the probability distribution is used as the confidence C of the prediction result;

[0073] S2.3) The prediction time interval T output by the network is supervisedly trained using the mean square error loss function to optimize the model parameters, thereby obtaining an intelligent prediction model for the slag endpoint.

[0074] Further, the specific steps of S3) are:

[0075] The real-time collected video data is preprocessed to obtain several drill rod images. These images are used as input into the intelligent prediction model for slag formation endpoint, and the predicted value of the slag formation endpoint time interval corresponding to each frame of the drill rod image is output; the predicted values ​​of all drill rod images are analyzed through the post-processing module to generate the final distance to the slag formation endpoint time of the video sample.

[0076] Furthermore, the specific steps of the logic post-processing are:

[0077] The prediction results are grouped into N binary groups (T0, C0), (T1, C2), ..., (T N-1 , C N-1 ), the binary tuples are arranged from large to small according to the confidence, and the first G results are taken as the credible tuples; the prediction results in the credible tuples are averaged as the final prediction result of the sampled video.

[0078] like Figure 2 As shown, another object of the present invention is to provide a system for implementing the above-mentioned prediction method, the system comprising:

[0079] The data acquisition and processing unit is used to collect image data of historical copper slag samples, pre-process them, and establish a data set, while also collecting image data of copper slag samples in real time and pre-processing them;

[0080] The model training unit is used to build an intelligent prediction model and use the data set as input to train the intelligent prediction model to obtain an intelligent prediction model for slag forming end point;

[0081] The prediction unit is used to input the image data of the copper slag sample collected in real time into the slag end point intelligent prediction model to obtain the prediction result, and perform logical post-processing on the obtained prediction result to obtain the end point value of the converter slag forming period.

[0082] Furthermore, the system also includes an automatic sampling module for controlling the robotic arm holding the drill rod to automatically sample from the PS converter tuyere.

[0083] The data collection uses an industrial camera to collect visual features of the sample on the brazing rod. The specific operation process is as follows: an operator or a robotic arm uses the brazing rod to collect the copper slag sample, inserts the brazing rod into the shooting area of ​​the industrial camera, and then rotates it at a constant speed for at least one circle to ensure that the surface slag sample image of the brazing rod is collected throughout the circle.

[0084] The dataset is established, and the process includes:

[0085] S2.1) intercepting images containing the drill rod from the video frame by frame as a drill rod image sample set;

[0086] S2.2) Manually screen and remove unqualified drill rod image samples, such as images of incomplete drill rod parts, blurred drill rods, images without drill rods, and images of drill rods without slag samples collected;

[0087] S2.3) Drill rod area extraction and rotation correction. Since only the middle part of the sampled image is the drill rod area, most of the area is background. Due to the complex on-site working environment, the background area contains a lot of noise. Therefore, the drill rod area needs to be extracted to remove the background and eliminate interference factors. The drill rod area extraction algorithm is as follows:

[0088] S2.31) converting the collected RGB three-channel image into a grayscale image;

[0089] S2.32) performing threshold segmentation on the grayscale image to convert the grayscale image into a binary image;

[0090] S2.33) Perform an erosion operation on the segmented area and mark different connected areas. Based on the characteristics of the original drill rod image, the upper and lower connected areas can be regarded as the background area, and the remaining area is the drill rod area. All pixel values ​​in the background area are marked as 0.

[0091] When sampling manually, due to manual operation errors, the drill rod is not placed horizontally. In order to facilitate observation and maintain data consistency, the drill rod needs to be horizontally corrected. The drill rod area horizontal rotation correction algorithm is as follows:

[0092] S2.4) Compress the vertical coordinates of the drill rod region into one dimension (for each horizontal coordinate x in the drill rod region, find the mean vertical coordinate of the drill rod pixels at that vertical coordinate) to obtain the coordinates of a set of points (each horizontal coordinate corresponds to a vertical coordinate);

[0093] S2.41) For this set of points, use the least squares method to fit a straight line, which is the center line of the drill rod area; find the midpoint of this center line, which is the center point of the drill rod area;

[0094] S2.42) Determine the angle between the centerline of the drill rod and the horizontal line; this angle is the deflection angle of the drill rod.

[0095] S2.43) After the deflection angle and center point are determined, the image is rotated and cropped based on the center point and the pre-set length and width.

[0096] S2.5) Drill rod image clustering: When the drill rod is inserted into the detection cabinet for intelligent identification, the converter slag forming period end point prediction system will save the sampled video at a speed of 20 frames per second, with a duration of about 10 seconds, totaling about 200 drill rod images. These images have a lot of repetition and similarity. The model cannot be improved from overly similar images during training, but it will waste training resources and cause data redundancy problems. Therefore, a data clustering method is used to screen and remove redundancy from the collected data to ensure that each sampled drill rod image contains more unique feature information as much as possible, which can improve model performance and prediction accuracy. The perceptual hash value (pHash) of the image is used as the image feature vector, and then the similarity of the feature vectors of different images is calculated to measure the similarity of different images. Then, the images are clustered according to the similarity between the images. The clustering algorithm is the classic K-Means algorithm, and the current number of clusters is 8.

[0097] S2.6) Data labeling: The last video in a collection time period is the slag formation endpoint video, and its time from the slag formation endpoint is marked as 0; the time from all other videos in this time period to the slag formation endpoint is the difference between the collection time of this video and the collection time of the slag formation endpoint video.

[0098] This embodiment uses Vision Transformer (ViT) as the base model of the artificial intelligence vision model. In this embodiment, n = 6, t0 = 2.5, t1 = 2.5, t2 = 7.5, t3 = 12.5, t4 = 17.5, t5 = 22.5, and the probability values ​​are p0, p1, p2, p3, p4, and p5, respectively, where these probability values ​​satisfy:

[0099] Then the final distance to the end point time T is the mathematical expectation of these 6 time nodes:

[0100] Each endpoint prediction time corresponds to a probability distribution. We calculate the variance of this probability distribution and use it as the confidence C of the endpoint prediction time:

[0101] Then for each image I i , the output of the model is a binary (T i , C i ), where T i is the model’s prediction value of the time from the image to the slag forming end point, C i is the confidence level of the predicted value.

[0102] S3.11) Loss function: The loss function uses the Mean Squared Error (MSE) loss function commonly used in regression models.

[0103] S3.12) Model training, validation, and testing;

[0104] S3.121) Continue training the model on the training set until the model loss converges or the number of training rounds reaches a preset upper limit;

[0105] S3.122) Evaluate the trained model using classic machine learning metrics: accuracy, recall, and mean absolute error (MAE). Based on the evaluation results of the model on the validation set, optimize the hyperparameters of the trained model.

[0106] The fourth step is the post-processing of the prediction results. This step fuses the prediction results of multiple images in the video and outputs the final prediction result of the video. The fusion step is to assume that a sampled video contains N images, and then reason about these images separately. We get N binary pairs (T0, C0), (T0, C0), ..., (T N-1 , C N-1 ), sort these binary tuples from large to small according to the confidence level, and take the results of the top G (here we set it to 0.2, that is, the top 20%) as the credible tuples. Then we take the average of the prediction results in the credible tuples as the final prediction result of the sampled video.

[0107] The above describes in detail the computer vision-based intelligent method and system for determining the end point of the converter slagging period, as provided in the embodiments of this application. The description of the above embodiments is intended only to facilitate understanding of the method and core concept of this application. Furthermore, those skilled in the art will appreciate that variations in the specific implementation and scope of application may occur based on the concepts of this application. Therefore, the contents of this specification should not be construed as limiting this application.

[0108] For example, certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that hardware manufacturers may use different nouns to refer to the same component. This specification and claims do not use differences in names as a way to distinguish components, but use differences in the functions of components as the criteria for distinction. For example, "including" and "comprising" mentioned throughout the specification and claims are open-ended terms, so they should be interpreted as "including / including but not limited to". "Approximately" means that within an acceptable error range, those skilled in the art can solve the technical problems within a certain error range and basically achieve the technical effects. The subsequent description in the specification is a preferred embodiment of the present application, but the description is for the purpose of illustrating the general principles of the present application, and is not used to limit the scope of the present application. The scope of protection of the present application shall be as defined in the attached claims.

[0109] It should also be noted that the terms "include," "comprises," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a product or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such product or system. In the absence of further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the product or system comprising the element.

[0110] It should be understood that the term "and / or" as used herein is merely a description of the relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the associated objects are in an "or" relationship.

[0111] The above description shows and describes several preferred embodiments of the present application. However, as previously mentioned, it should be understood that the present application is not limited to the form disclosed herein and should not be construed as excluding other embodiments. Instead, the present application can be used in various other combinations, modifications, and environments and can be modified within the scope of the application concept described herein through the above teachings or technology or knowledge in the relevant field. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present application should be protected by the claims appended hereto.

Claims

1. An intelligent method for determining the end point of converter slagging period based on computer vision, characterized in that: The intelligent determination method for the end point of the converter slagging period specifically comprises the following steps: S1) collecting image data of historical copper slag samples, preprocessing, and establishing a data set; S2) constructing an intelligent prediction model, and using the data set of S1) as input to train the intelligent prediction model, thereby obtaining an intelligent prediction model for slag forming end point; S3) inputting the image data of the copper slag sample collected in real time into the slag forming end point intelligent prediction model for prediction to obtain a prediction result, and performing logical post-processing on the obtained prediction result to obtain the converter slag forming period end point value.

2. The intelligent method for determining the end point of the converter slag-making period according to claim 1, characterized in that: The pre-processing steps in S1) are: S1.1) First, the video is segmented and the effective area of ​​the video data containing the drill rod is segmented and screened; S1.2) Extraction of the drill rod area; S1.3) Correcting the drill rod region to obtain a drill rod image with a height of H and a width of W; S1.4) Label the corrected drill rod image; S1.5) Finally, the dataset is divided into training set, test set and validation set.

3. The intelligent determination method for the end point of converter slag forming period according to claim 2, characterized in that: The specific steps of S1.2) are: S1.21) converting the collected RGB three-channel image into a grayscale image; S1.22) performing threshold segmentation on the grayscale image to convert the grayscale image into a binary image; S1.23) Perform an erosion operation on the segmented area and mark different connected areas. Based on the characteristics of the original image of the brazing rod, the upper and lower connected areas are considered as the background area, and the remaining area is the brazing rod area. All pixel values ​​in the background area are marked as 0.

4. The intelligent method for determining the end point of converter slag forming period according to claim 2, characterized in that: The steps of S1.3) are: S1.31) Compress the ordinates of the drill rod region into one dimension. That is, for each abscissa x in the drill rod region, find the mean ordinate of the drill rod pixels at that ordinate to obtain the coordinates of a set of points, i.e., each abscissa corresponds to a ordinate. S1.32) For this set of points, use the least squares method to fit a straight line, which is the center line of the drill rod area; find the midpoint of this center line, which is the center point of the drill rod area; S1.33) Determine the angle between the centerline of the drill rod and the horizontal line; this angle is the deflection angle of the drill rod. S1.34) After the deflection angle and center point are obtained, the image is rotated and cropped according to the center point and the pre-set length and width.

5. The intelligent method for determining the end point of converter slag forming period according to claim 2, characterized in that: The steps of S1.4) are as follows: Assume that there are m samplings in a slag-making period, and the sampling times are S1, S2, ..., S m-1 , the actual end time of the slag making period is S final , then the label of the drill rod image obtained by the method of claim 4 in the i-th sampling video is d i =S final -S i , the unit is minutes; if the number of drill rod images obtained as described above is greater than N, randomly select N drill rod images and put them into the data set.

6. The intelligent method for determining the end point of converter slag forming period according to claim 1, characterized in that: The network structure of S2) works in tandem with the backbone network and the output layer. The backbone network extracts the feature map of the input image through a convolutional neural network, and the softmax function of the output layer normalizes the feature map so that the prediction result has probabilistic interpretation. The specific steps are as follows: S2.1) Construct an artificial intelligence visual network model with a softmax activation function as the last layer, where the backbone network adopts a general network architecture and the output layer is a classification head with n categories, where the n categories correspond to a preset set of candidate time intervals {t0, t1, ..., t n-1 }, output the probability distribution of each time interval category through the softmax activation function; S2.2) Calculate the mathematical expectation value of the current time from the end point based on the probability distribution As the prediction time interval, where p i is the predicted probability of the i-th time interval, and the inverse of the variance of the probability distribution is used as the confidence C of the prediction result; S2.3) The prediction time interval T output by the network is supervisedly trained using the mean square error loss function to optimize the model parameters, thereby obtaining an intelligent prediction model for the slag endpoint.

7. The intelligent method for determining the end point of converter slag forming period according to claim 1, characterized in that: The specific steps of S3) are: The real-time collected video data is preprocessed to obtain several drill rod images. These images are used as input into the intelligent prediction model for slag formation endpoint, and the predicted value of the slag formation endpoint time interval corresponding to each frame of the drill rod image is output; the predicted values ​​of all drill rod images are analyzed through the post-processing module to generate the final distance to the slag formation endpoint time of the video sample.

8. The intelligent method for determining the end point of converter slag forming period according to claim 1, characterized in that: The specific steps of the logic post-processing are: The prediction results are grouped into N binary groups (T0, C0), (T1, C2), ..., (T N-1 , C N-1 ), the binary tuples are arranged from large to small according to the confidence, and the first G results are taken as the credible tuples; the prediction results in the credible tuples are averaged as the final prediction result of the sampled video.

9. A system for implementing the prediction method according to any one of claims 1 to 8, characterized in that: The system comprises: The data acquisition and processing unit is used to collect image data of historical copper slag samples, pre-process them, and establish a data set, while also collecting image data of copper slag samples in real time and pre-processing them; The model training unit is used to build an intelligent prediction model and use the data set as input to train the intelligent prediction model to obtain an intelligent prediction model for slag forming end point; The prediction unit is used to input the image data of the copper slag sample collected in real time into the slag end point intelligent prediction model to obtain the prediction result, and perform logical post-processing on the obtained prediction result to obtain the end point value of the converter slag forming period.

10. The system according to claim 9, characterized in that The system also includes an automatic sampling module for controlling a robotic arm holding a drill rod to automatically sample from the PS converter tuyere.

Images