An intelligent calculation method for carbide gas generation based on computer vision

Abstract

The present invention proposes an intelligent carbide gas generation amount calculation method based on computer vision, which includes the steps of: using a carbide furnace tapping recognition model to automatically recognize the sampling rod in the camera video stream and capture pictures containing carbide; using a picture quality determination model to determine the quality of the pictures containing carbide and perform screening; segmenting the screened pictures through a segmentation model to remove interfering pixels around the carbide; using a gas generation amount calculation model to calculate the gas generation amount value of the segmented carbide pictures; using a gas generation amount correction model to detect the carbide grade of the segmented carbide pictures and give a gas generation amount correction value for correcting the gas generation amount value corresponding to different regions of the carbide; using a gas generation amount weighted calculation model to calculate the final gas generation amount value by combining the current carbide furnace batching ratio and the calculated gas generation amount value and gas generation amount correction value. The present invention can calculate the gas generation amount of the carbide in the tapping sampling rod in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of carbide gas generation measurement, and specifically relates to an intelligent carbide gas generation calculation model based on computer vision. Background Art

[0002] Calcium carbide, chemically named calcium carbide, is produced by the reaction of quicklime and coke in an electric furnace. It is an important raw material in the chemical industry, and its gas generation is an important indicator to measure the quality of calcium carbide. The level of gas generation directly affects the enterprise's operating costs and production quality.

[0003] At present, the measurement method of domestic calcium carbide gas generation mainly relies on national standards. Due to the unevenness of industrial calcium carbide, especially poor-quality calcium carbide, the reproducibility of gas generation data is not good. By observing the cooling state of the calcium carbide out of the furnace, it is found that the appearance of calcium carbide with different gas generation amounts is not the same. Some experienced quality inspectors can obtain the corresponding gas generation amount only based on its appearance, and the error from the gas generation amount measured by the national standard is within 1-5, which basically meets the requirements of workshop production. In addition, calcium carbide production is listed as a hazardous chemical production process and requires major supervision. Therefore, in calcium carbide production, enterprises should minimize manual operations as much as possible to improve production safety, and there is an urgent need to adopt a vision-based method for calculating calcium carbide gas generation.

[0004] With the rapid development of computer hardware technology in recent years, deep learning technology, which originally required a large amount of calculations, has also made great progress. Especially in the field of computer vision, applications such as face recognition and object detection have been widely used. By analyzing the characteristics of the calcium carbide surface, it is found that different calcium carbide gas generation levels have obvious characteristics. For example, calcium carbide with a gas generation amount of about 270 L / kg is brown and has a relatively rough surface; calcium carbide with a gas generation amount of about 280 L / kg is black and darker; calcium carbide with a gas generation amount of about 290 L / kg is black and shiny with a certain number of white dots distributed; calcium carbide with a gas generation amount of about 300 L / kg is black and shiny with many white dots distributed. At the same time, convolutional neural networks have a strong ability to extract these characteristics. Therefore, it is theoretically feasible to estimate the calcium carbide gas generation through a vision-based solution. For this reason, the present invention proposes a method for calculating calcium carbide gas generation based on computer vision. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an intelligent carbide gas generation calculation method based on computer vision for real-time calculation of the gas generation amount of the calcium carbide on the sampling rod out of the furnace in view of the deficiencies of the prior art.

[0006] To solve the above technical problem, the content of the present invention includes:

[0007] An intelligent carbide gas generation calculation method based on computer vision, comprising the following steps:

[0008] S1: Use the calcium carbide furnace recognition model to automatically identify the proofing rods on the camera video stream and capture pictures containing calcium carbide;

[0009] S2: Use the image quality judgment model to judge the quality of the images containing calcium carbide and screen them, and eliminate those images that contain too much light source interference;

[0010] S3: Segment the filtered images through a segmentation model to remove interference pixels around calcium carbide;

[0011] S4: using a gas generation calculation model to calculate the gas generation value of the segmented calcium carbide image;

[0012] S5: using the gas generation correction model to detect the calcium carbide grade on the segmented calcium carbide image, and providing a gas generation correction value to correct the gas generation value for the corresponding grade of different calcium carbide areas;

[0013] S6: Using a gas generation weighted calculation model, the final gas generation value is calculated by combining the current calcium carbide furnace batching ratio and the calculated gas generation value and gas generation correction value.

[0014] Furthermore, in step S1, the calcium carbide furnace identification model adopts the s model in the Yolov5 series, the model optimizer is SGD, the model input image size is img=(640,384,3), and capturing pictures containing calcium carbide from the camera video stream specifically includes the following steps:

[0015] S11: Convert the camera video stream to a picture, collect a large number of pictures of proofing rods with and without calcium carbide, and prepare the training set, verification set and test set required for the calcium carbide furnace recognition model;

[0016] S12: Model training, observe the training loss and validation set loss, analyze the convergence of the model, and after this round of training, test the generalization of the model on the test set;

[0017] S13: Test the model's false detection and missed detection of calcium carbide images. If the model fails to meet expectations, increase the training set, adjust parameters, and iterate training to eventually meet all detection requirements.

[0018] S14: During the model inference stage, a picture is collected from the camera video stream every 3-5 seconds for calcium carbide detection and identification. When calcium carbide is detected in 5 consecutive frames, the capture of calcium carbide pictures is started, and the last frame is stored until the calcium carbide in the furnace is completely removed from the furnace.

[0019] Furthermore, in step S2, the picture quality determination model is used to determine and screen the quality of pictures containing calcium carbide, which specifically includes the following steps:

[0020] S21: Perform secondary screening on the collected calcium carbide pictures, select the calcium carbide pictures without light source interference and cracks, and make the training set, validation set, and test set required for the picture quality judgment model;

[0021] S22: Select a classification network based on ResNet50, determine the training optimizer and learning rate, and use the gradient descent method to train and optimize the model on the training set until the model is trained to the global optimum;

[0022] S23: Conduct tests on the trained model on the test set to test the generalization effect of each model on various incomplete calcium carbide and calcium carbide with excessive light interference. If the judgment is not good, add such data specifically and retrain the model according to the above steps until all calcium carbide that does not meet the on-site requirements can be accurately judged;

[0023] S24: In the model inference stage, use the model to perform quality judgment on the pictures containing calcium carbide, and screen out the pictures containing red light, invalid calcium carbide, many cracks, incomplete calcium carbide, and uncooled calcium carbide.

[0024] Further, in step S3, the segmentation model uses the Mask R-CNN instance segmentation model. The specific steps for segmenting the screened pictures are as follows:

[0025] S31: Select parameters before training, use the gradient descent method to train the Mask R-CNN model. The optimizer used is the SGD optimizer, and the input image size of the model is img=(512,512,3);

[0026] S32: Model training. By observing the training loss and the validation set loss, determine several relatively optimal models, and then recheck the stability of the model on the test set to determine the final model. If the effect does not meet the expectation, change the training parameters and continue training until the expectation is met; if it still does not reach the prefetch, specifically increase the training set, then adjust the parameters and perform iterative training until the segmentation requirements are finally met;

[0027] S33: In the model inference stage, the segmentation model outputs the boding box and segmentation mask of the calcium carbide. Then, use the box to crop the calcium carbide from the 4K original picture, and use the mask value to replace the non-calcium carbide area with 128 pixels.

[0028] 5. The intelligent calcium carbide gas generation amount calculation method based on computer vision according to claim 1, wherein in step S4, the gas generation amount calculation model uses a regression and classification model to calculate the gas generation amount of the segmented calcium carbide pictures, and the specific steps are as follows:

[0029] S41: Collect large-scale photos containing calcium carbide and annotate the gas emission volume. The corresponding calcium carbide gas emission volume value is obtained according to the traditional statistical method. Finally, the training set, verification set and test set required for the gas emission volume calculation model are prepared;

[0030] S42: Build a regression and classification synchronous convolutional neural network based on residual network, determine the data preprocessing parameters, determine the model optimization parameters, and train the network until the global optimum. The regression optimization function is the following MSE function, and the classification optimization function is the cross entropy loss function;

[0031] S43: Perform a generalization test on the test set to test the model robustness. If the requirements are not met, add a data set and repeat the above steps to train the model. MAE is used as the evaluation index for calcium carbide during the model training stage, and the final test index is the accuracy of the difference between the predicted value and the true value within 5 as the final index for screening the model.

[0032] Furthermore, in step S5, a gas volume correction model is used to provide a gas volume correction value, which specifically includes the following steps:

[0033] S51: according to the visual features presented by different calcium carbide gas emission levels, the calcium carbide gas emission levels are divided into 8 levels, and the calcium carbide images are labeled with gas emission area levels according to the divided calcium carbide gas emission levels, so as to prepare the training set, validation set and test set required for the gas emission correction model;

[0034] S52: The Yolov5 detection model is used to train the model until the global optimal model is reached; the parameters are selected before training, and the model is trained using the gradient descent method. The optimizer used is the SGD optimizer, and the model input image size is img = (640, 640, 3); in the model training stage, the final training of the model is completed through training, testing and fine-tuning;

[0035] S53: In the model reasoning stage, according to the detection results of the calcium carbide regional grade detection network, the box frame of the largest category and the box frame of the smallest category are calculated, and the gas emission correction value E is calculated by the following formula:

[0036]

[0037]

[0038] E={ij,P i -Q j |}

[0039] Among them, P8 represents the area ratio of all boxes of level 8 to the original image, and i and j represent the gas emission levels.

[0040] Further, in the step S6, the calculation method of the final gas generation value V is as follows:

[0041] S61: Calculate the influence coefficient of the calcium carbide furnace ratio according to the current calcium carbide furnace ratio Q;

[0042] S62: Multiply the calculated influence coefficient of the calcium carbide furnace ratio by 5 to obtain the ratio influence value M;

[0043]

[0044] S63: Add the gas generation value P r calculated by the gas generation calculation model, the gas generation correction value E and the ratio influence value M to obtain the final gas generation value V:

[0045] V = P r + E + M.

[0046] The beneficial effects of the present invention are as follows:

[0047] The present invention first uses a calcium carbide out-of-furnace recognition model to automatically recognize the sampling rod, and recognizes calcium carbide and the sampling rod from the camera video stream; secondly, uses a picture quality determination model to determine the current picture quality and eliminates those pictures containing excessive light source interference; furthermore, segments the selected pictures through a Mask R-CNN instance segmentation model to remove the interference pixels around the calcium carbide; finally, uses a regression and classification model to calculate the gas generation model for the segmented calcium carbide; when outputting the final gas generation, correct the gas generation value through a gas generation correction model, which identifies the corresponding levels for different regions of the calcium carbide to give a correction value, and then combines the current ingredient ratio of the calcium carbide furnace, and finally performs weighted summation on the calculated value of the gas generation model, the influence value of the ingredient ratio, and the correction value to obtain the final calcium carbide gas generation value.

[0048] The present invention can not only calculate the real-time gas generation of the out-of-furnace sampling rod calcium carbide, avoiding the time-consuming problem brought by the traditional metering gas generation calculation, but also calculate whether the surface of the sampling rod calcium carbide is uniform, thereby indirectly judging the distribution of various raw materials in the calcium carbide furnace, providing objective data support for calcium carbide production enterprises. Description of the Drawings

[0049] Figure 1 is the flow schematic diagram of the present invention;

[0050] Figure 2 is the schematic diagram of automatic calcium carbide out-of-furnace recognition of the present invention;

[0051] Figure 3 is the schematic diagram of calcium carbide picture quality determination of the present invention;

[0052] Figure 4It is the effect diagram of removing the interference area around calcium carbide by Mask R-CNN of the present invention;

[0053] Figure 5 It is the schematic diagram of the regression classification synchronous convolutional neural network based on the residual network of the present invention;

[0054] Figure 6 It is the schematic diagram of correcting the correction model of the present invention. Detailed implementation manners

[0055] For the convenience of understanding the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners. Those skilled in the art should understand that the described embodiments are only to help understand the present invention and should not be regarded as specific limitations to the present invention.

[0056] The present invention calculates the gas generation amount of calcium carbide by cascading multiple vision-based deep learning models, which mainly consists of video acquisition, algorithm service, and gas generation amount calculation.

[0057] As Figure 1 shown, the present invention provides an intelligent calcium carbide gas generation amount calculation method based on computer vision, including the following steps:

[0058] S1: Use the calcium carbide tapping recognition model to automatically recognize the sampling rod in the camera video stream and capture the pictures containing calcium carbide;

[0059] S2: Use the picture quality determination model to determine and screen the pictures containing calcium carbide, and eliminate those pictures containing excessive light source interference;

[0060] S3: Segment the screened pictures through the segmentation model to remove the interference pixels around calcium carbide;

[0061] S4: Use the gas generation amount calculation model to calculate the gas generation amount of the segmented calcium carbide pictures;

[0062] S5: Use the gas generation amount correction model to detect the calcium carbide grade of the segmented calcium carbide pictures. By analyzing the calcium carbide grade in the pictures and whether the calcium carbide is incomplete, obtain the gas generation amount correction value for correcting the gas generation amount value;

[0063] S6: Use the gas generation amount weighted calculation model to calculate the final gas generation amount value by combining the current charge ratio of the calcium carbide furnace and the calculated gas generation amount value and gas generation amount correction value.

[0064] As Figure 2As shown in the figure, in step S1, according to the time limit for determining the gas evolution of calcium carbide by calcium carbide production enterprises, the calcium carbide tapping recognition model uses the s model in the Yolov5 series, the model optimizer is SGD, the input image size of the model is img=(640,384,3), and the specific steps for capturing pictures containing calcium carbide from the camera video stream are as follows:

[0065] S11: The video stream collected by the camera is transmitted to the NVR for video storage; the original video stored in the NVR is converted into pictures, and a large number of proofing rods pictures with and without calcium carbide are collected to produce the training set, validation set, and test set required for the calcium carbide tapping recognition model;

[0066] S12: Model training, observing the training loss and validation set loss, analyzing the convergence of the model. After this round of training ends, the generalization of the model is tested on the test set;

[0067] S13: Testing the misdetection and missed detection of the calcium carbide pictures by the model. If the expected results are not achieved, the training set is increased accordingly, the parameters are adjusted, and iterative training is carried out until all detection requirements are met;

[0068] S14: In the model inference stage, a picture is collected from the camera video stream every 3 - 5 seconds for calcium carbide detection and recognition. When calcium carbide is detected for 5 consecutive frames, that is, within 15 - 25 seconds, the capture of calcium carbide pictures is started, and the last frame is saved until the calcium carbide tapping of this furnace ends.

[0069] As Figure 3 shown in the figure, in step S2, a picture quality determination model is used to determine and screen the pictures containing calcium carbide, and the specific steps are as follows:

[0070] S21: Making a targeted data set according to the calcium carbide lighting conditions of the proofing rod and the possible light sources and incomplete calcium carbide around: The collected calcium carbide pictures are screened again to select the calcium carbide pictures without light source interference and cracks, and the training set, validation set, and test set required for the picture quality determination model are produced;

[0071] S22: Selecting a classification network based on ResNet50, determining the training optimizer and learning rate, and using the gradient descent method to train and optimize the model on the training set until the model is trained to the global optimum;

[0072] S23: Testing the trained model on the test set to test the generalization effect of each model on various incomplete calcium carbide and calcium carbide with excessive light interference. If the determination is not good, the corresponding data is added specifically, and the model is trained again according to the above steps until all calcium carbide that does not meet the on-site requirements is accurately determined;

[0073] S24: In the model inference stage, use the model to determine the quality of pictures containing calcium carbide, and screen out pictures containing red light, invalid calcium carbide, many cracks, incomplete calcium carbide, and uncooled calcium carbide.

[0074] As Figure 4 shown, the segmentation model uses the Mask R-CNN instance segmentation model, and the specific steps for segmenting the screened pictures are as follows:

[0075] S31: Select parameters before training, use the gradient descent method to train the Mask R-CNN model, the optimizer used is the SGD optimizer, and the input image size of the model is img=(512,512,3);

[0076] S32: Model training, by observing the training loss and the validation set loss, determine several relatively optimal models, and then verify the stability of the model on the test set again to determine the final model. If the effect does not meet the expectations, change the training parameters and continue training until the expectations are met; if the expectations are still not met, specifically increase the training set, then adjust the parameters and perform iterative training until the segmentation requirements are finally met;

[0077] S33: In the model inference stage, the segmentation model outputs the boding box and segmentation mask of calcium carbide, then uses the box to crop calcium carbide from the 4K original image, and uses the mask value to replace 128 pixels in the non-calcium carbide area.

[0078] The mask calculation formula is as follows:

[0079]

[0080] where q ij is the model prediction confidence.

[0081] As Figure 5 shown, in step S4, the gas generation calculation model uses a regression and classification model to calculate the gas generation value of the segmented calcium carbide pictures, and the specific steps are as follows:

[0082] S41: Collect a large number of pictures containing calcium carbide and perform gas generation annotation. The corresponding calcium carbide gas generation values are obtained according to traditional statistical methods, and finally the training set, validation set, and test set required for the gas generation calculation model are made;

[0083] In the calcium carbide production link of calcium carbide production enterprises, the gas generation of calcium carbide in the current calcium carbide furnace is recorded by sampling every day for each furnace eye. The gas generation of the sampled calcium carbide for each furnace eye is also different, and the appearance is also different, so it can be used to make a gas generation data set;

[0084] S42: Build a regression and classification synchronous convolutional neural network based on residual network, determine the data preprocessing parameters, determine the model optimization parameters, and train the network until the global optimum. The regression optimization function is the MSE function, and the classification optimization function is the cross entropy loss function;

[0085] The formulas for the MSE function and the cross entropy loss function are as follows:

[0086]

[0087]

[0088] S43: Perform a generalization test on the test set to test the robustness of the model, such as MAE and PC. If the requirements are not met, add a data set in a targeted manner and repeat the above steps to train the model. MAE is used as the evaluation index for calcium carbide during the model training phase. The formula is as follows. The final test index uses the accuracy rate of the difference between the predicted value and the true value within 5 as the final index for the screening model. The formula is as follows:

[0089]

[0090]

[0091] like Figure 6 As shown, in step S5, the gas emission correction model is used to detect the carbide grade of the segmented calcium carbide, and the gas emission correction value is calculated by analyzing the carbide grade in the image and whether the calcium carbide is incomplete. The process is as follows:

[0092] S51: according to the visual features presented by different calcium carbide gas emission levels, the calcium carbide gas emission levels are divided into 8 levels, and the calcium carbide images are labeled with gas emission area levels according to the divided calcium carbide gas emission levels, so as to prepare the training set, validation set and test set required for the gas emission correction model;

[0093] S52: The Yolov5 detection model is used to train the model until the global optimal model is reached; the parameters are selected before training, and the model is trained using the gradient descent method. The optimizer used is the SGD optimizer, and the model input image size is img = (640, 640, 3); in the model training stage, the final training of the model is completed through training, testing and fine-tuning;

[0094] S53: In the model reasoning stage, according to the detection results of the calcium carbide regional grade detection network, the box frame of the largest category and the box frame of the smallest category are calculated, and the gas emission correction value E is calculated by the following formula:

[0095]

[0096]

[0097] E = {i - j, |P i - Q j |}

[0098] Among them, P8 represents the proportion of the area of all boxes with grade 8 in the original image, and i and j represent the gas generation level.

[0099] In step S6, the calculation method of the final gas generation value V is as follows:

[0100] S61: Calculate the influence coefficient of the calcium carbide furnace ratio according to the current calcium carbide furnace ratio Q;

[0101] S62: Multiply the calculated influence coefficient of the calcium carbide furnace ratio by 5 to obtain the ratio influence value M;

[0102]

[0103] S63: Add the gas generation value P calculated by the gas generation calculation model r , the gas generation correction value E and the ratio influence value M to obtain the final gas generation value V:

[0104] V = P r + E + M

[0105] The present invention breaks the traditional way that the measurement of calcium carbide gas generation must rely on chemical reactions, avoids personal safety problems, and also saves a large amount of labor costs, playing a crucial role in the intelligent production of calcium carbide enterprises.

[0106] The present invention uses the Yolov5 detection model to perform real-time detection on the camera video stream, performs real-time high-speed detection on calcium carbide rods for 24 hours, automatically identifies the sampled calcium carbide, locates the tapping time, cooling time, end time of each furnace of calcium carbide, etc., and completes the interception of daily tapped calcium carbide pictures and the statistics of the daily tapping number; when calcium carbide is detected, a network based on residual is used to judge the quality of calcium carbide pictures, and calcium carbide pictures without excessive light source interference are screened out; then these pictures are segmented by the Mask R-CNN instance segmentation model to remove surrounding interfering pixels, and finally calcium carbide pictures without interference items are obtained; the segmented pictures are sent to the gas generation calculation model for gas generation calculation and the gas generation correction model, and finally the final gas generation is calculated.

[0107] The regression classification synchronous convolutional neural network based on the residual network is adopted to solve the problems of gradient explosion and network non-convergence caused by too large loss during the training of direct classification or regression networks, enabling the network to perform fine-grained regression under large classification, and making the predicted value of the model closer to the true value. Since in the calcium carbide production process, calcium carbide is not evenly distributed, resulting in different local gas generation amounts of calcium carbide, an error correction model and a calcium carbide furnace ratio parameter are introduced to finally correct the gas generation value.

[0108] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent calculation method for carbide gas generation based on computer vision, characterized in that, The method comprises the following steps: S1: Use the calcium carbide furnace recognition model to automatically identify the proofing rods on the camera video stream and capture pictures containing calcium carbide; S2: Use the image quality judgment model to judge the quality of the images containing calcium carbide and screen them, and eliminate those images that contain too much light source interference; S3: Segment the filtered images through a segmentation model to remove interference pixels around calcium carbide; S4: using a gas generation calculation model to calculate the gas generation value of the segmented calcium carbide image; S5: using the gas generation correction model to detect the calcium carbide grade on the segmented calcium carbide image, and providing a gas generation correction value to correct the gas generation value for the corresponding grade of different calcium carbide areas; S6: using a gas generation weighted calculation model, by combining the current calcium carbide furnace batching ratio and the calculated gas generation value and gas generation correction value, the final gas generation value is calculated; In the step S5, a gas volume correction model is used to give a gas volume correction value, which specifically includes the following steps: S51: according to the visual features presented by different calcium carbide gas emission levels, the calcium carbide gas emission levels are divided into 8 levels, and the calcium carbide images are labeled with gas emission area levels according to the divided calcium carbide gas emission levels, so as to prepare the training set, validation set and test set required for the gas emission correction model; S52: Train the model using the Yolov5 detection model until the global optimum of the model is reached; select the parameters before training, use the gradient descent method to train the model, and the optimizer used is the SGD optimizer. The input image size of the model is ; In the model training stage, complete the final training of the model through training, testing, and re-fine-tuning; S53: During the model inference stage, based on the detection results of the calcium carbide area level detection network, calculate the box frames of the class with the largest proportion and the box frames of the class with the smallest proportion, and calculate the correction value of the gas evolution amount through the following formula : ， ， ， Among them, represents the proportion of the area of all boxes with a grade of 8 to the original image, and represents the gas evolution grade; In the step S6, the final gas generation value is calculated as follows: S61: According to the current carbide furnace ratio Calculate the influence coefficient of the carbide furnace ratio; S62: Multiply the calculated influence coefficient of the calcium carbide furnace ratio by 5 to obtain the ratio influence value ; ； S63: The gas generation value calculated by the gas generation calculation model , the gas generation correction value and the ratio influence value are added to obtain the final gas generation value : 。 2. The intelligent carbide gas generation amount calculation method based on computer vision according to claim 1, wherein In the step S1, the calcium carbide tapping recognition model uses the s model in the Yolov5 series, the model optimizer is SGD, and the model input image size is , and the specific steps for capturing pictures containing calcium carbide from the camera video stream are as follows: S11: Convert the camera video stream to a picture, collect a large number of pictures of proofing rods with and without calcium carbide, and prepare the training set, verification set and test set required for the calcium carbide furnace recognition model; S12: Model training, observe the training loss and validation set loss, analyze the convergence of the model, and after this round of training, test the generalization of the model on the test set; S13: Test the model's false detection and missed detection of calcium carbide images. If the model fails to meet expectations, increase the training set, adjust parameters, and iterate training to eventually meet all detection requirements. S14: During the model inference stage, a picture is collected from the camera video stream every 3-5 seconds for calcium carbide detection and identification. When calcium carbide is detected in 5 consecutive frames, the capture of calcium carbide pictures is started, and the last frame is stored until the calcium carbide is removed from the furnace.

3. The intelligent carbide gas generation amount calculation method based on computer vision according to claim 1, wherein, In step S2, the picture quality determination model is used to determine and screen the quality of the pictures containing calcium carbide, which specifically includes the following steps: S21: Perform secondary screening on the collected calcium carbide images, select calcium carbide images without light source interference and cracks, and prepare the training set, verification set and test set required for the image quality judgment model; S22: Select a classification network based on ResNet50, determine the training optimizer and learning rate, use the gradient descent method to train the optimization model on the training set, and train the model to the global optimum; S23: Testing the trained models on the test set to test the generalization effect of each model on various types of incomplete calcium carbide and calcium carbide with excessive light interference. If the judgment is not good, the data of the test set is added in a targeted manner, and the model is retrained according to the above steps until all calcium carbide that does not meet the on-site needs is accurately judged; S24: In the model inference stage, use the model to judge the quality of pictures containing calcium carbide, and screen out pictures containing red light, invalid calcium carbide, many cracks, incomplete calcium carbide, and uncooled calcium carbide.

4. The intelligent calcium carbide gas generation amount calculation method based on computer vision according to claim 1, characterized in that In step S3, the segmentation model uses the Mask R-CNN instance segmentation model. The specific steps for segmenting the screened pictures are as follows: S31: Select parameters before training, use the gradient descent method to train the Mask R-CNN model, and the optimizer used is the SGD optimizer. The input image size of the model is ; S32: Model training. By observing the training loss and the validation set loss, determine several relatively optimal models. Then, verify the stability of the model again on the test set to determine the final model. If the effect does not meet the expectation, change the training parameters and continue training until the expectation is met; if it still does not reach the prefetch, add the training set specifically, then adjust the parameters and perform iterative training until the segmentation requirements are finally met. S33: In the model inference stage, the segmentation model outputs the bounding box and segmentation mask of calcium carbide. Then, use the box to crop calcium carbide from the 4K original image, and use the mask value to replace the non-calcium carbide area with 128 pixels.

5. The intelligent carbide gas generation amount calculation method based on computer vision according to claim 1, wherein, In step S4, the gas generation calculation model uses a regression and classification model to calculate the gas generation of the segmented calcium carbide pictures. The specific steps are as follows: S41: Collect a large number of photos containing calcium carbide and perform gas generation annotation. The corresponding calcium carbide gas generation values are obtained according to traditional statistical methods. Finally, prepare the training set, validation set, and test set required for the gas generation calculation model. S42: Build a regression and classification synchronous convolutional neural network based on the residual network, determine the data preprocessing parameters, determine the model optimization parameters, and train the network until the global optimum. The regression optimization function is the following MSE function, and the classification optimization function is the cross-entropy loss function. S43: Conduct model generalization testing on the test set to test the robustness of the model. If the requirements are not met, add the dataset specifically and repeat the above steps to train the model; the evaluation index for calcium carbide uses MAE in the model training stage, and the final test index is the accuracy rate that the difference between the predicted value and the true value is within 5 as the final index for screening the model.

Images