Blast furnace crude fuel granularity intelligent detection system and method based on deep learning algorithm

Abstract

The invention relates to the technical field of industrial automation and intelligent detection, and discloses a blast furnace crude fuel granularity intelligent detection system and method based on a deep learning algorithm. Firstly, an image of blast furnace crude fuel is collected in real time through the image acquisition module, and then the image is subjected to denoising and enhancement processing through the preprocessing module. Then, an improved YOLOv8 model which takes the OfficientNet as a backbone network and combines an FPN and PAN feature fusion strategy is adopted to carry out granularity detection, bounding box regression is optimized through a GIoU Loss loss function, and the detection precision and speed are remarkably improved. In addition, the system also has a particle size distribution statistical function, and can generate a particle size distribution histogram in real time. Therefore, intelligent upgrading of blast furnace crude fuel granularity detection is realized, the production efficiency and the product quality are improved, and the requirements of modern steel production on efficient, accurate and safe granularity detection are met.

Description

Technical Field

[0001] The present invention relates to the field of industrial automation and intelligent detection technology, and in particular to a blast furnace raw fuel particle size intelligent detection system and method based on a deep learning algorithm. Background Art

[0002] In the blast furnace smelting process, the particle size of the blast furnace raw fuel is one of the important factors affecting the smelting effect. At present, the particle size detection method commonly used by steel mills is mainly manual sampling. The specific method is that the staff will take samples from the raw material yard at regular intervals, and then use the screen to screen to determine whether the particle size of the blast furnace raw fuel meets the requirements. However, this detection method has the following disadvantages: low detection frequency and poor real-time performance.

[0003] With the rapid development and technological progress of the steel industry, the requirements for production efficiency and product quality are constantly increasing. The traditional manual sampling particle size detection method can no longer meet the needs of modern steel production. Therefore, it is urgent to develop a new particle size detection method to improve the real-time and accuracy of the detection. Summary of the invention

[0004] In view of the above-mentioned technical deficiencies, the purpose of the present invention is to provide a blast furnace raw fuel particle size intelligent detection system and method based on deep learning algorithm to solve the problems of low detection frequency and poor real-time performance in the prior art.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a blast furnace raw fuel particle size intelligent detection system based on a deep learning algorithm, the system comprising: An image acquisition module, which acquires real-time images of blast furnace raw fuel to be used; A preprocessing module performs preprocessing on the acquired image, including denoising and enhancement processing; Model training module, making a data set and using the data set to train the improved YOLOv8 model; Model detection module, which uses the improved YOLOv8 model to detect the preprocessed real-time image; The particle size distribution statistics module calculates the particle size of each blast furnace raw fuel particle based on the output results of the model and statistics the particle size distribution.

[0006] Preferably, in a possible implementation of the first aspect, the denoising process adopts a bilateral filtering method, combining spatial proximity and pixel value similarity to construct a weight template for filtering; the enhancement process adopts a contrast-limited adaptive histogram equalization method.

[0007] Preferably, in a possible implementation manner of the first aspect, the bilateral filtering method is implemented by the following steps: Construct a spatial distance weight template to calculate the exponential weight of the Euclidean distance between each point in the neighborhood and the center point; Construct a similarity weight template and calculate the exponential weight of the difference between the pixel value of each point and the pixel value of the center coordinate point of the template window; The above two templates are multiplied and normalized to obtain a weight template for each position, and the template is used to filter the image.

[0008] Preferably, in a possible implementation manner of the first aspect, the adaptive histogram equalization method is implemented by the following steps: Divide the image into blocks of a preset size; Histogram equalization is performed independently within each block, and contrast limitation is introduced; Use bilinear interpolation to handle the boundaries between adjacent blocks; Map the processed image values ​​into the grayscale value range.

[0009] Preferably, in a possible implementation manner of the first aspect, the data set is composed of preprocessed historical blast furnace raw fuel images, and data enhancement is performed using Mixup, CutMix and Mosaic data enhancement techniques.

[0010] Preferably, in a possible implementation manner of the first aspect, the improved YOLOv8 model includes: EfficientNet is used as the backbone network, and the FPN and PAN feature fusion strategy is introduced. The detection head is applied to the feature map after feature fusion to perform anchor box generation, bounding box regression and category prediction; Optimize the loss function, including using the GIoU Loss loss function for bounding box regression training and the CrossEntropy Loss classification loss function for category prediction training.

[0011] Preferably, in a possible implementation manner of the first aspect, the particle size distribution statistics module specifically includes: Extract the bounding box information of each particle from the output of the model; Convert pixel coordinates into actual size by combining industrial camera calibration parameters; By calculating the actual distance of multiple points on the particle boundary, and then taking the average value as the particle diameter; Traverse all detected particles, calculate the diameter of each particle, and divide the particle size range into preset intervals; Generate a particle size distribution histogram.

[0012] In a second aspect, the present invention provides a method for intelligently detecting particle size of blast furnace raw materials and fuels based on a deep learning algorithm, comprising the following steps: Step S100: Acquire a real-time image of the blast furnace raw fuel to be used; Step S200: preprocessing the acquired image, including denoising and enhancement; Step S300: Create a data set, and use the data set to train an improved YOLOv8 model; Step S400: using the improved YOLOv8 model to detect the preprocessed real-time image; Step S500: Calculate the particle size of each blast furnace raw fuel particle according to the output result of the model, and calculate the particle size distribution.

[0013] The beneficial effects of the present invention are as follows: by integrating the improved YOLOv8 model and the EfficientNetV2 backbone network, the detection accuracy and speed are improved, and the computing resource requirements are reduced; by combining the FPN and PAN feature fusion strategies, the complex and changeable blast furnace raw fuel forms in the blast furnace are effectively dealt with, and the multi-scale target detection capability is enhanced; by adopting a variety of data enhancement technologies and progressive learning strategies, it can work stably under various blast furnace working conditions, and effectively deal with complex industrial environments and data occlusion, missing problems and other problems; through image preprocessing and GIoU Loss bounding box regression training, the image quality and detection accuracy are optimized; and finally, the intelligent upgrade of blast furnace raw fuel particle size detection is realized to meet the modern steel production The demand for efficient, accurate and safe particle size detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 A system structure diagram for intelligent detection of blast furnace raw fuel particle size based on deep learning algorithm is provided for this application.

[0016] Figure 2 A flow chart of a method for intelligent detection of blast furnace raw fuel particle size based on a deep learning algorithm is provided for this application.

[0017] Explanation of the accompanying drawings: 1-image acquisition module, 2-preprocessing module, 3-model training module, 4-model detection module, 5-particle size distribution statistics module. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] Embodiment 1: Figure 1 As shown, the blast furnace raw fuel particle size intelligent detection system based on the deep learning algorithm of the present invention comprises: The image acquisition module 1 acquires the real-time image of the blast furnace raw fuel to be used.

[0020] Specifically, real-time images of blast furnace raw fuel are obtained through industrial cameras, light sources, industrial computers, etc. This embodiment uses an industrial-grade embedded industrial computer Advantech ARK-3520F, equipped with a high-performance multi-core processor to ensure data processing capabilities, and equipped with an NVMe SSD hard drive to provide high-speed data reading and writing speeds, further reducing system delays.

[0021] Basler's industrial-grade cameras are used. They can not only work stably in complex industrial environments and capture clear and delicate images, but also have excellent low-light sensitivity and strong anti-electromagnetic interference capabilities. The built-in intelligent temperature control system can effectively resist the high temperature in the blast furnace area, and achieve high-speed transmission of image data through the GigE Vision high-speed interface.

[0022] In view of the material and color characteristics of blast furnace raw fuel, an LED light source system is used. The light source system provides stable, uniform lighting conditions suitable for camera imaging, and has intelligent brightness and color temperature adjustment functions. Through preset scene modes or automatic perception of ambient light changes, the light source system can dynamically optimize lighting conditions to ensure image quality. The light source housing is made of high temperature and corrosion resistant materials, and has a built-in dust filter to extend service life and reduce maintenance costs.

[0023] The preprocessing module 2 performs preprocessing on the acquired image, including denoising and enhancement.

[0024] Specifically, the denoising process uses bilateral filtering method, combining spatial proximity and pixel value similarity to build a weight template for filtering. Bilateral filtering is a nonlinear filtering method, which is a compromise method combining the spatial proximity and pixel value similarity of the image. According to the neighborhood of each position, a different weight template is built for that position. The detailed process is as follows: First, construct a spatial distance weight template.

[0025] in Represents a point in the neighborhood With center point The Euclidean distance of is the standard deviation of the Gaussian function.

[0026] Then, construct the similarity weight template by It measures the exponential difference between the value at and its neighboring values.

[0027] in Indicates that the image is at point The pixel value at ; is the pixel value of the center coordinate point of the template window, is the standard deviation of the Gaussian function.

[0028] Finally, multiply the above two templates and then normalize them to get the weight template of the position.

[0029] The pixel value of the filtered image is,

[0030] Multiply the obtained weight template and the corresponding position in the neighborhood of the position, and then sum them up to get the output value of the position.

[0031] The enhancement process uses a contrast-limited adaptive histogram equalization method. Contrast-limited adaptive histogram equalization is an improvement on the traditional adaptive histogram equalization to solve the overexposure problem that may occur in high-contrast areas. The steps of contrast-limited adaptive histogram equalization include: The image is divided into multiple rectangular areas using a grid partitioning method, and histogram equalization is performed independently in each small area. The histogram of each area is calculated, and histogram equalization is performed to enhance the contrast.

[0032] In order to avoid excessive contrast caused by histogram equalization, contrast limitation is introduced. When the number of pixels at a certain gray level exceeds a predetermined threshold, the excess will be dispersed to other gray levels to prevent excessive local contrast.

[0033] After performing histogram equalization in each small area, the processing results are merged into a complete image. Bilinear interpolation is used to smooth the boundaries between adjacent blocks to avoid obvious block effects. The processed image values ​​are mapped to a suitable range. In this embodiment, the grayscale value range of the image is mapped to between 0 and 255.

[0034] Model training module 3: Create a data set and use the data set to train the improved YOLOv8 model.

[0035] Specifically, we first screened out clear, unobstructed, and representative images from historical blast furnace raw fuel images as the basis of the dataset, and removed blurry, repeated, or poor quality images. Then we used LabelImg to annotate the screened images. For the blast furnace raw fuel detection task, we mainly annotated the boundaries and category labels of the target objects.

[0036] Then, Mixup, CutMix and Mosaic data enhancement techniques are used to enhance the labeled data set.

[0037] Mixup is a regularization technique that mixes two images and their labels through linear interpolation to generate new training samples, which helps to increase the generalization ability of the model and prevent overfitting. In this embodiment, the specific implementation of Mixup includes the following steps: Randomly select two blast furnace raw fuel images and their corresponding labels; Generate a mixing coefficient through Beta distribution , The value range is between; Use the mixing coefficient to linearly mix the two images and their labels to generate a new image and label.

[0038] The formula of Mixup is expressed as:

[0039]

[0040] CutMix is ​​another regularization technique that cuts out a region from one image and pastes it to a random position in another image while adjusting the mixing ratio of the corresponding labels. This technique helps improve the robustness of the model to occlusion and missing data. In this embodiment, the implementation steps of CutMix are as follows: Randomly select two blast furnace raw fuel images and their corresponding labels; Randomly crop a region from the first image; Paste this region into a random position in the second image and update the label of the second image to reflect the blended data.

[0041] Mosaic is an image synthesis technology that stitches four images together to form a new training image. This technology increases the diversity of training data and helps the model better handle complex scenes. In this embodiment, the implementation of Mosaic includes the following steps: Randomly select four blast furnace raw fuel images and their corresponding labels; Stitch these four images together to form a new image; Update the labels corresponding to the new images to reflect the stitched data.

[0042] The data set after data augmentation is divided into training set, validation set and test set. The proportion of training set is 70%, the proportion of validation set is 20%, and the proportion of test set is 10%. The training set is used for model training, the validation set is used for adjusting model parameters, and the test set is used to evaluate the performance of the model.

[0043] Use the training set to train the improved YOLOv8 model. The improved YOLOv8 model network structure includes: Backbone network: EfficientNetV2 is selected as the backbone network of the model. Through scaling and efficient convolution blocks, the network significantly improves the training speed and inference efficiency while maintaining accuracy. This is particularly important for application scenarios such as processing blast furnace raw fuel that require fast response. At the same time, it reduces the demand for computing resources and facilitates deployment on edge computing devices.

[0044] Feature fusion: Based on EfficientNetV2, a feature fusion strategy combining feature pyramid network (FPN) and path aggregation network (PAN) is introduced. FPN transmits high-level semantic information to low-level layers through a top-down path, enhancing the spatial details of low-level features; while PAN further fuses information through a bottom-up path, making high-level features rich in spatial details, thereby improving the detection capability of multi-scale targets, which is particularly effective for blast furnace raw materials of different sizes and shapes in the blast furnace.

[0045] Anchor box generation and adjustment: Based on the statistical characteristics of blast furnace raw materials, a set of anchor box sizes and proportions are designed as the initial estimate of the target bounding box. During the detection process, each grid will generate multiple candidate bounding boxes based on these anchor boxes, and adjust their positions and sizes through bounding box regression technology to match the actual boundaries of the target object.

[0046] Bounding box regression: The feature map is further processed using a deep convolutional network to output the precise offset of each candidate bounding box. These offsets are used to fine-tune the position and size of the anchor box to generate the final predicted bounding box, achieving accurate positioning of the blast furnace raw fuel.

[0047] Category prediction: The detection head is also responsible for predicting the category of the object in each bounding box. The category probability output by the convolutional layer is combined with the softmax function to determine the category information of the object in each bounding box. This step ensures that the model can accurately distinguish different types of blast furnace raw materials in the blast furnace.

[0048] Non-maximum suppression (NMS): NMS technology is used to optimize the overlapping and redundant bounding boxes generated during the detection process, remove low-confidence bounding boxes, and retain the bounding boxes that are most likely to represent the real target, thereby improving the accuracy and clarity of the detection results.

[0049] Progressive learning and non-uniform scaling: The optimized backbone network uses a progressive learning method to dynamically adjust the regularization strategy according to the size of the training image, which enhances the stability and robustness of the model. At the same time, the non-uniform scaling strategy is used to optimize the different stages of the model, further improving the training speed and parameter efficiency, ensuring the efficient operation of the model in the blast furnace raw fuel detection task.

[0050] Optimize the loss function, including using the GIoU Loss loss function for bounding box regression training and the CrossEntropy Loss classification loss function for category prediction training.

[0051] GIoU Loss, as the loss function of bounding box regression, effectively solves the problem of gradient vanishing of traditional IoU Loss when bounding boxes do not intersect, and more comprehensively reflects the overlap degree and position relationship of bounding boxes. The GIoU Loss formula is:

[0052] Among them, A is the predicted box, B is the real box, and C is the minimum bounding box containing A and B. The range of GIoU is , when the two boxes completely overlap, GIoU is 1; when the two boxes have no intersection and are far apart, GIoU is close to -1.

[0053] In this embodiment, GIoU Loss is used to train the bounding box regressor. By minimizing the GIoU difference between the predicted box and the true box, the model can more accurately predict the position and size of the bounding box. This improvement not only improves the matching accuracy of the bounding box, but also enhances the model's ability to detect blast furnace raw materials of different shapes, sizes, and positional relationships.

[0054] In terms of category prediction, Cross Entropy Loss is widely used in classification tasks as a standard to measure the difference between the predicted probability distribution and the true probability distribution. Cross Entropy Loss is calculated as:

[0055] Where M is the total number of categories, is an indicator variable (1 if category c is the true category, 0 otherwise), is the probability that the model predicts class c.

[0056] In this embodiment, each bounding box is allowed to belong to multiple categories at the same time, which is suitable for the mixed or overlapping situations existing in blast furnace raw fuel.

[0057] Model detection module 4 uses the improved YOLOv8 model to detect the preprocessed real-time image.

[0058] Specifically, first, load the trained improved YOLOv8 model file into the system, and use the cache mechanism of the deep learning framework to reduce the overhead of repeated loading. According to the specific needs of real-time detection, configure the input size of the model to balance the detection accuracy and computational efficiency. At the same time, set a reasonable batch size and confidence threshold to ensure that while maintaining a high detection rate, false detection and missed detection are reduced. In this embodiment, the input size is set to 640x640, the batch size is 16, and the confidence threshold is 0.7.

[0059] The preprocessed real-time blast furnace raw fuel image is input into the improved YOLOv8 model, and the network structure and optimization algorithm of the model are used for forward propagation calculation. The multi-scale feature fusion capability of the model is used to detect targets at different scales and improve the detection accuracy of particles of different sizes. According to the actual size distribution of blast furnace raw fuel particles, the size and proportion of the anchor frame are optimized to make it easier for the model to capture the target particles. Finally, the detection results of the blast furnace raw fuel image and the confidence of each detection frame are output.

[0060] The particle size distribution statistics module 5 calculates the particle size of each blast furnace raw fuel particle according to the output result of the model, and calculates the particle size distribution.

[0061] In the particle size distribution statistics module, first extract the bounding box information of each particle from the results of the detection model. Combined with the calibration parameters of the camera, the pixel coordinates are converted to actual size. The actual distances of multiple points on the particle boundary are calculated, and then the average value is taken as the particle diameter. Traverse all detected particles, calculate the diameter of each particle, and collect these diameter data.

[0062] In this embodiment, the particle size range is divided into multiple intervals (including 0-10 mm, 10-20 mm, 20-30 mm, 30-40 mm and above 40 mm). The diameter data of all particles are traversed, and the number of particles in each interval is counted according to the particle size interval division. Matplotlib is used to generate a particle size distribution histogram. The proportion of particles in different particle size ranges is shown in a chart for easy analysis and evaluation.

[0063] Embodiment 2: Figure 2 As shown, the present invention provides a method for intelligently detecting particle size of blast furnace raw materials and fuels based on a deep learning algorithm, comprising: Step S100: Acquire a real-time image of the blast furnace raw fuel to be used; Step S200: preprocessing the acquired image, including denoising and enhancement; Step S300: Create a data set, and use the data set to train an improved YOLOv8 model; Step S400: using the improved YOLOv8 model to detect the preprocessed real-time image; Step S500: Calculate the particle size of each blast furnace raw fuel particle according to the output result of the model, and calculate the particle size distribution.

[0064] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. Intelligent detection system for blast furnace raw fuel particle size based on deep learning algorithm, characterized by: The system comprises: An image acquisition module, which acquires real-time images of blast furnace raw fuel to be used; A preprocessing module performs preprocessing on the acquired image, including denoising and enhancement processing; Model training module, making data sets and using the data sets to train the improved YOLOv8 model; Model detection module, which uses the improved YOLOv8 model to detect the preprocessed real-time image; The particle size distribution statistics module calculates the particle size of each blast furnace raw fuel particle based on the output results of the model and statistics the particle size distribution.

2. The system according to claim 1, characterized in that: The denoising process adopts a bilateral filtering method, and a weight template is constructed by combining spatial proximity and pixel value similarity for filtering; the enhancement process adopts a contrast-limited adaptive histogram equalization method.

3. The system according to claim 2, characterized in that: The bilateral filtering method is implemented by the following steps: Construct a spatial distance weight template to calculate the exponential weight of the Euclidean distance between each point in the neighborhood and the center point; Construct a similarity weight template and calculate the exponential weight of the difference between the pixel value of each point and the pixel value of the center coordinate point of the template window; The above two templates are multiplied and normalized to obtain a weight template for each position, and the template is used to filter the image.

4. The system according to claim 2, characterized in that: The adaptive histogram equalization method is implemented by the following steps: Divide the image into blocks of a preset size; Histogram equalization is performed independently in each block, and contrast limitation is introduced; Use bilinear interpolation to handle the boundaries between adjacent blocks; Map the processed image values ​​into the grayscale value range.

5. The system according to claim 1, characterized in that: The dataset consists of preprocessed historical blast furnace raw fuel images, and is enhanced using Mixup, CutMix and Mosaic data enhancement techniques.

6. The system according to claim 1, characterized in that: The improved YOLOv8 model includes: EfficientNet is used as the backbone network, and the FPN and PAN feature fusion strategy is introduced. The detection head is applied to the feature map after feature fusion to perform anchor box generation, bounding box regression and category prediction; Optimize the loss function, including using the GIoU Loss loss function for bounding box regression training and the CrossEntropy Loss classification loss function for category prediction training.

7. The system according to claim 1, characterized in that: The particle size distribution statistics module specifically includes: Extract the bounding box information of each particle from the output of the model; Convert pixel coordinates into actual size by combining industrial camera calibration parameters; By calculating the actual distance of multiple points on the particle boundary, and then taking the average value as the particle diameter; Traverse all detected particles, calculate the diameter of each particle, and divide the particle size range into preset intervals; Generate a particle size distribution histogram.

8. An intelligent detection method for blast furnace raw fuel particle size based on deep learning algorithm, characterized in that: The following steps are involved: Step S100: Acquire a real-time image of the blast furnace raw fuel to be used; Step S200: preprocessing the acquired image, including denoising and enhancement; Step S300: Create a data set, and use the data set to train an improved YOLOv8 model; Step S400: using the improved YOLOv8 model to detect the preprocessed real-time image; Step S500: Calculate the particle size of each blast furnace raw fuel particle according to the output result of the model, and calculate the particle size distribution.

Images