A deep learning-based chemical fertilizer granulation particle size detection method and device

Abstract

The application provides a kind of based on deep learning's chemical fertilizer granulation granularity detection method and device.The method includes controlling industrial camera timing shooting image in the process of chemical fertilizer granulation, image is pretreated, utilizes deep learning to carry out detection model training and optimization, the detection model of training is deployed to WEB Internet of Things platform, industrial computer real-time shooting granulation image is uploaded to platform and carries out granulation detection.In addition, the application provides a supporting device for granulation detection, including: chemical fertilizer conveyor belt, fixed support, industrial camera, light supplement lamp, industrial computer, WEB Internet of Things cloud platform.Fixed support is arranged on the chemical fertilizer conveyor belt, industrial camera is erected, light supplement lamp is arranged on the both sides of support, industrial computer is connected with camera through optical fiber, obtains granulation image, then is uploaded to WEB Internet of Things cloud platform through HTTP, obtains granulation detection result, further gives early warning.The application realizes the intellectualization of chemical fertilizer granulation granularity detection by deep learning, reduces cost, improves safety.

Description

Technical Field

[0001] This invention belongs to the field of compound fertilizer granulation particle size detection and analysis technology, specifically involving a fertilizer granulation particle size detection method and device based on deep learning. Background Technology

[0002] Currently, compound fertilizer production methods in fertilizer plants include slurry method, melt method, and granulation method. The amino acid method, due to its advantages of high yield, low application cost, and low energy consumption, has become an important development direction in compound fertilizer production technology and is the mainstream granulation technology for medium and large-scale fertilizer production plants. In amino acid granulation technology, various raw materials are metered, mixed evenly, and fed into a granulator. The granulator rotates to form a material bed, and then liquid ammonia (gaseous ammonia), concentrated sulfuric acid, washing liquid, and steam are added through a buried pipe to wet and heat the material. The reacted material rolls and agglomerates into granules within the granulator. After proper reaction, the material in the granulator has a uniform color, full granules, a dense and glossy surface, and high strength, laying the foundation for preventing subsequent product caking.

[0003] Traditional amino acid granulation processes require skilled workers in fertilizer plants to visually monitor the material's condition inside the granulator in real time, relying on past granulation experience to assess its performance. This method, reliance on skilled workers for granulation inspection, is time-consuming and labor-intensive. Furthermore, the granulation process releases strong chemical gases; long-term inhalation of these gases, mixed with amino acids, can lead to ammonia poisoning and damage to the respiratory system. Therefore, unmanned inspection of amino acid granulation processes is imperative. With the development of science and technology, factory production is trending towards automation and intelligence. Computer vision technology, due to its high recognition rate and non-contact characteristics, meets the requirements of factory production and is being increasingly widely used. Summary of the Invention

[0004] Based on the above background description, the purpose of this invention is to provide a fertilizer granulation particle size detection method and device based on deep learning. By deploying corresponding image acquisition devices in the fertilizer production workshop and using precise fertilizer image processing methods, the particle size of fertilizer granules can be detected remotely and without contact, providing timely early warnings and adjusting the material ratio and production environment during granulation to improve fertilizer granulation quality. At the same time, it reduces labor costs, frees workers from real-time detection at the granulation site, avoids ammonia poisoning, and improves the working environment.

[0005] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0006] A deep learning-based method for detecting particle size in fertilizer granulation, the method comprising the following steps:

[0007] S1. The industrial computer controls the industrial camera to acquire granulation images;

[0008] S2. Preprocess the acquired images to make the features of fertilizer granules more obvious;

[0009] S3. Image annotation: Annotate the preprocessed dataset to distinguish between normal and abnormal particle images; specifically, annotate the collected dataset to distinguish between images of fertilizer particles within the normal range, images of fertilizer particles with fine powdery particle size, and images of fertilizer granulation with large-area agglomeration.

[0010] S4. Train the granulation detection network model using deep learning methods; randomly divide the images in the dataset into training, validation, and test sets. Input the training set into a CNN neural network for training to obtain a model that can be used for image recognition;

[0011] S5. Optimize the model to improve recognition accuracy and speed; optimize the model by adjusting the parameters of each layer of the CNN neural network model.

[0012] S6. Deploy the trained granulation detection network model to the WEB Internet of Things cloud platform.

[0013] S7. The industrial control computer acquires granulation images in real time and uploads them to the WEB Internet of Things cloud platform for particle size detection. The detection results are then returned to the industrial control computer to provide granulation assistance and guidance to on-site personnel.

[0014] Furthermore, in step S1, the image acquisition device is a CCD industrial camera, which is vertically mounted on the conveyor belt transporting fertilizer granules and periodically acquires images of the fertilizer granules produced by the drum granulation process. Considering the granulation object and the characteristics of the industrial camera, a CCD industrial camera was chosen to acquire the images, and a camera control program was developed on the industrial control computer. In addition, since the fertilizer production plant operates 24 hours a day with two shifts rotating, and nighttime lighting is weak, supplementary lighting is placed on both sides of the image acquisition device, i.e., the fixed bracket.

[0015] Furthermore, in step S2, image preprocessing is performed through the following operations:

[0016] 2.1) The image is pixel-wise summed to increase the pixel features of the fertilizer particles. A bilateral filtering algorithm is then used to filter out noise and retain the edge information of the fertilizer particles.

[0017] 2.2) Enhance different local areas or pixels using the mean and variance of local images: Let f(x,y) represent the image to be enhanced, and g(x,y) represent the output enhanced image, i.e., the locally enhanced image. The enhanced local area or pixel is the region in the image with relatively low gray level and low contrast. The following formula is used for calculation:

[0018]

[0019] where \(m(x,y)\) and \(\sigma(x,y)\) represent the mean and variance of the pixels within the neighborhood centered on the pixel at the position \((x,y)\) respectively; \(M\) represents the mean of the image \(f(x,y)\) to be enhanced, \(S\) represents the standard deviation of the enhanced image \(f(x,y)\), \(a\), \(b\), \(k\) and \(E\) are all coefficients, and \(a < b < 0.5\), \(k < 0.5\), \(E\) represents the enhancement coefficient, \(2 < E < 5\), \(M\leq km(x,y)\) is used to select the regions that are darker compared to the whole image; \(aS\leq\sigma(x,y)\leq bS\) is used to select the regions with relatively small variance (relatively small contrast) compared to the whole image.

[0020] Furthermore, in the step S3, the image is labeled according to the morphology of the fertilizer particles in the image, including the labeling of three morphologies: powdery particles, normal particles, and large particles. When the CCD industrial camera is first fixed to take images, qualified fertilizer particles are placed on the conveyor belt and photographed to obtain standard fertilizer particle images. Then, consult and learn from the staff engaged in granulation detection to label the three morphologies of powdery particles, normal particles, and large particles for subsequent images.

[0021] Furthermore, in the step S4, the images in the dataset are randomly divided into a training set, a validation set, and a test set in a ratio of 7:2:1; the input image pixels are uniformly adjusted to a size of 512*512, and trained on a computer with 1 3070 graphics card using the programming language Python in the Python 3.7 environment. The required abnormal particle regions are represented by the recognized bounding boxes, and the large particle regions and powdery particle regions are distinguished by bounding boxes of different colors. The regions outlined by the red bounding boxes are large particles, and the regions outlined by the blue bounding boxes are powdery particles.

[0022] Furthermore, in the step S5, during the training process, the dataset is iterated 200 times, and 8 samples are selected for weight update each time, that is, the batch_size is 8. The training is visualized through TensorBoard, and the final optimal weights of the training are retained.

[0023] Furthermore, in the step S7, an industrial control computer is used to control the industrial camera to collect granulation images at regular intervals and automatically upload them to the WEB Internet of Things cloud platform. The platform passes through the granulation detection network model, gives the detected images, calculates the proportion of abnormal particles, and feeds it back to the warning system on the industrial control computer, and then adjusts the fertilizer granulation elements.

[0024] The present invention also provides a supporting device for fertilizer granulation particle size detection:

[0025] A fertilizer granulation detection device for implementing the fertilizer granulation particle size detection method, the device comprising: a fertilizer conveyor belt (generally provided by the factory granulation unit), a fixed bracket, an industrial camera, a supplementary light, an industrial control computer, and a WEB Internet of Things cloud platform.

[0026] The fertilizer conveyor belt is installed below the outlet of the fertilizer granulation drum. A fixed bracket is positioned above the conveyor belt and vertically mounts an industrial camera. Supplemental lighting is provided on both sides of the bracket. An industrial control computer connects to the industrial camera via fiber optic cable to acquire granulation images, which are then uploaded to a web-based IoT cloud platform via HTTP protocol to obtain granulation detection results. The fertilizer conveyor belt is an existing component of the fertilizer plant, carrying a stream of fertilizer granules freshly granulated from the drum granulator. A CCD industrial camera bracket is fixed above the conveyor belt to capture images of the fertilizer granules.

[0027] Furthermore, two supplementary lights are provided, which are placed on the left and right sides of the fixed bracket respectively, symmetrically distributed, to illuminate the flow of fertilizer particles on the fertilizer conveyor belt with LED light.

[0028] Furthermore, the fixed bracket is a stainless steel bracket, including: an overall support frame, horizontal crossbars, and vertical bars; the overall support frame consists of left and right vertical bars, fixed beside the fertilizer conveyor belt; the support frame has horizontal crossbars, and the vertical bars are fixed to the crossbars by horizontal fasteners; the industrial camera is fixed to the vertical bars by fasteners. Adjusting the crossbar fasteners allows the industrial camera to move left and right, and adjusting the vertical bar fasteners allows the industrial camera to move up and down. This allows for flexible adjustment of the industrial camera to capture clear and bright granulation images.

[0029] The CCD industrial camera is connected to the router via the Gige interface protocol to transmit images to the industrial control computer.

[0030] The industrial control computer uses the WinForm framework in the Visual Studio integrated development environment to write the industrial control computer software, which can control the exposure time and gain value of the industrial camera to read clear and bright images.

[0031] The WEB IoT cloud platform can receive granulation images transmitted from industrial control computers via HTTP protocol, perform online detection of granulation particle size using deep learning algorithms, and realize early warning of granulation particle size and data visualization.

[0032] The present invention has the following beneficial effects:

[0033] This invention discloses a deep learning-based method and apparatus for detecting fertilizer granulation particle size. By setting industrial camera parameters on an industrial control computer, the CCD industrial camera is controlled to periodically capture images of fertilizer granules during the production process, and the images are transmitted to a web-based IoT cloud platform via HTTP. A neural network model detects abnormal areas in the fertilizer granulation process and calculates the particle size distribution. Finally, the detection results are returned to the industrial control computer to provide warnings to granulation workers. This invention achieves unmanned and intelligent particle size detection in compound fertilizer granulation through this non-contact detection method, reducing labor costs and avoiding the potential hazard of ammonia poisoning caused by prolonged inhalation of air containing amino acids, thus improving safety. Attached Figure Description

[0034] Figure 1 This is a flowchart of the granulation image detection method of the present invention;

[0035] Figure 2 This is a schematic diagram of the detection device of the present invention;

[0036] Figure 3 This is a topology diagram of the data transmission in the detection system of the present invention.

[0037] In the picture: 1-Fertilizer conveyor belt; 2-Fixed bracket; 3-CCD industrial camera; 4-LED fill light. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0039] like Figure 1 As shown, this invention is a method for detecting the particle size of fertilizer granulation based on deep learning, specifically including the following steps:

[0040] Step S1 involves acquiring images of fertilizer granulation. Industrial cameras typically consist of a camera with a CCD or CMOS image sensor and a lens. CCD cameras, due to their single-point output, offer better signal consistency, facilitating high-quality image acquisition. Furthermore, digital signals effectively avoid attenuation and external interference during transmission, making them better suited for capturing real-time moving objects. CMOS cameras, on the other hand, are lower in cost, consume less power, have higher integration, and offer faster linear array output, resulting in relatively better image quality when capturing static objects. In this invention, the primary images acquired are those of fertilizer granules moving on a conveyor belt. Therefore, a FLIR BFLY-PGE-13E4C-CS camera (CCD industrial camera) from FLIR Systems, Inc. (USA) is used for data acquisition. Because fertilizer plants operate on a 24-hour two-shift system, and the angle of natural light varies at different times, resulting in different image shadows, this significantly impacts subsequent model training and recognition accuracy. Therefore, two LED supplementary lights (4) are symmetrically mounted on the fixed bracket of the industrial camera to provide stable illumination for the fertilizer granules under the camera.

[0041] The CCD industrial camera 3 is a high-speed camera; in this embodiment, the industrial camera has a resolution of 1280*1024. Since the conveyor belt carrying the fertilizer granules operates at a speed of 0.6 m / s, to ensure clear imaging, the exposure time in the camera control software on the industrial computer is generally set to 1 ms. Considering the pressure on network transmission, the camera is set to take a picture and upload it to the backend server every 10 seconds.

[0042] Step S2 involves preprocessing the acquired granulation images. Pixel summation is performed to enhance the pixel features of the fertilizer granules. A bilateral filtering algorithm is then used to remove noise while preserving the edge information of the fertilizer granules. Furthermore, since the fertilizer granules accumulate on the conveyor belt to form small hills, the edge contrast of the granules is not obvious. Therefore, the mean and variance of the local image are used to enhance different local areas or pixels. The gray-level mean is a measure of image brightness, and the gray-level variance is a measure of contrast. Therefore, the gray-level mean and gray-level variance of all pixels in the neighborhood of a given pixel can characterize the information of the local image.

[0043] Let f(x,y) represent the image to be enhanced, and g(x,y) represent the output enhanced image, i.e., the locally enhanced image. To enhance areas of the image with low grayscale and low contrast, the following formula can be used for calculation:

[0044]

[0045] Here, m(x, y) and σ(x, y) represent the mean and variance of the pixels within the neighborhood centered on the pixel at the position (x, y) respectively; M represents the mean of the image f(x, y) to be enhanced, S represents the mean square deviation of the enhanced image f(x, y), and a, b, k, and E are all coefficients. Generally, a < b < 0.5, k < 0.5, and E represents the enhancement coefficient, where 2 < E < 5. M ≤ km(x, y) is used to select the regions that are relatively darker compared to the entire image; aS ≤ σ(x, y) ≤ bS is used to select the regions with relatively smaller variance (lower contrast) compared to the entire image.

[0046] In step S3, when the CCD industrial camera is fixed to take images, qualified fertilizer particles are placed on the conveyor belt and photographed to obtain comparison fertilizer particle images. Refer to the relevant standards for the particle size of fertilizer particles, consult and learn from the experienced workers in granulation detection, and use the labellmg software to label the 2000 sets of granulation image datasets collected into three forms: powdery particles, normal particles, and large particles.

[0047] In step S4, the images in the dataset are randomly divided into a training set, a validation set, and a test set in a ratio of 7:2:1. The input image pixels are uniformly adjusted to a size of 512 * 512, and the training is carried out on a computer with 1 3070 graphics card, using the programming language Python in the Python 3.7 environment. The recognized target boxes are the required abnormal particle regions. Among them, the regions framed by the red boxes are large particles, and the regions framed by the blue boxes are powdery particles.

[0048] In step S5, during the training process, the dataset is iterated 200 times, and 8 samples are selected for weight update each time, that is, the batch_size is 8. Visualize the training through TensorBoard and retain the final optimal weights of the training. Compare multiple CNN network models and select the model with high accuracy and fast recognition speed to meet the requirements of actual detection.

[0049] In step S6, package the trained deep learning model and deploy it to the WEB Internet of Things cloud platform. Enable all industrial control computers in the granulation workshops to be connected for image transmission and receipt of detection results.

[0050] In step S7, use the industrial control computer to control the industrial camera to collect granulation images regularly and automatically upload them to the WEB Internet of Things cloud platform. The platform passes through the CNN detection model, outputs the detection image results, calculates the proportion of abnormal particles, and feeds it back to the warning system on the industrial control computer, and then the staff adjusts the fertilizer granulation elements.

[0051] On the other hand, the present invention also provides a fertilizer granulation particle size detection device based on deep learning. As Figure 2 And Figure 3As shown, the device includes: fertilizer conveyor belt 1, fixed bracket 2, CCD industrial camera 3, LED fill light 4, industrial control computer, and WEB Internet of Things cloud platform.

[0052] The fertilizer conveyor belt is an existing feature of the fertilizer plant, carrying a stream of freshly granulated fertilizer particles. A CCD industrial camera bracket is fixed above the conveyor belt, allowing the camera to capture images of the fertilizer particles.

[0053] A horizontal crossbar and a vertical bar are installed on the fixed bracket to secure the CCD industrial camera and the LED supplementary light. The CCD industrial camera is mounted on the vertical bar of the bracket, with its shooting angle perpendicular to the transport direction of the fertilizer conveyor belt.

[0054] There are two supplementary lights, positioned symmetrically on the left and right sides of the fixed bracket. These LED lights illuminate the flow of fertilizer granules on the conveyor belt.

[0055] The CCD industrial camera connects to the router via the Gige interface protocol to transmit images to the industrial control computer.

[0056] The industrial control computer uses the WinForm framework in the Visual Studio integrated development environment to write the industrial control computer software, which can control the exposure time and gain value of the industrial camera to read clear and bright images.

[0057] The WEB IoT cloud platform can receive granulation images transmitted from industrial control computers via HTTPS protocol, perform online granulation particle size detection using deep learning algorithms, and provide early warnings and data visualization of granulation particle size.

[0058] The fixed support frame is made of stainless steel and includes: an overall support frame, horizontal crossbars, and vertical bars. The support frame consists of left and right vertical bars, fixed beside the fertilizer conveyor belt. Symmetrical supplementary lights are installed on the support frame, and the height and direction of the light can be adjusted.

[0059] The support frame has a horizontal crossbar, and a vertical bar is fixed to the crossbar by clips. A CCD industrial camera is fixed to the vertical bar by clips. This allows the industrial camera to be moved left and right by adjusting the crossbar clips, and up and down by adjusting the vertical bar clips. This flexible adjustment of the industrial camera allows for the capture of clear and bright granulation images.

[0060] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting particle size in fertilizer granulation based on deep learning, characterized in that, The method includes the following steps: S1. The industrial computer controls the industrial camera to acquire granulation images; S2. Preprocess the acquired images to make the features of fertilizer granules more obvious; the image preprocessing is performed through the following operations: 2.1) Perform pixel addition on the image to increase the pixel features of the fertilizer particles, and use a bilateral filtering algorithm to filter out noise and retain the edge information of the fertilizer particles; 2.2) Enhance different local areas or pixels using the mean and variance of local images: using To represent the image to be enhanced, use This represents the output enhanced image, i.e., the image after local enhancement. The enhanced local area or pixel is a region in the image with relatively low grayscale and low contrast, calculated using the following formula: where m and represent the mean and variance of the pixels in the neighborhood centered on the pixel at position respectively; M represents the mean of the image to be enhanced , S represents the mean square deviation of the enhanced image , a, b, k and E are all coefficients, and a < b < 0.5, k < 0.5, E represents the enhancement coefficient, 2 < E < 5, is used to select the area darker relative to the whole image; is used to select the area with a smaller variance relative to the whole image; S3. Image annotation: Annotate the preprocessed dataset to distinguish between normal and abnormal particle images. S4. Use deep learning methods to train the granulation detection network model; S5. Model Optimization: Optimize the model to improve recognition accuracy and speed; S6. Deploy the trained granulation detection network model to the WEB Internet of Things cloud platform. S7. The industrial control computer acquires granulation images in real time and uploads them to the WEB Internet of Things cloud platform for particle size detection.

2. The method for detecting the particle size of fertilizer granulation based on deep learning according to claim 1, characterized in that: In step S1, the image acquisition device is a CCD industrial camera, which is vertically mounted on the conveyor belt transporting fertilizer granules and periodically acquires images of the fertilizer granules produced by the drum granulation.

3. The method for detecting the particle size of fertilizer granulation based on deep learning according to claim 1, characterized in that: In step S3, the image is labeled according to the morphology of fertilizer particles in the image, including labeling three types of morphology: powdery particles, normal particles, and large particles; the powdery particles and large particles are abnormal particles.

4. The method for detecting the particle size of fertilizer granulation based on deep learning according to claim 3, characterized in that: In step S4, the images in the dataset are randomly divided into a training set, a validation set, and a test set in a 7:2:1 ratio; the input image pixels are uniformly adjusted to 512. The size is 512, and the identified target boxes represent the required abnormal particle areas, with different colored target boxes distinguishing between large particle areas and powdery particle areas.

5. The method for detecting the particle size of fertilizer granulation based on deep learning according to claim 1, characterized in that: In step S5, the dataset is iterated 200 times during the training process. Each time, 8 samples are selected for weight updates, and the training is visualized, retaining the final optimal training weights.

6. The method for detecting the particle size of fertilizer granulation based on deep learning according to claim 1, characterized in that: In step S7, an industrial computer is used to control an industrial camera to periodically collect granulation images, which are then automatically uploaded to a WEB Internet of Things cloud platform. The platform, through the granulation detection network model, outputs the detection images and calculates the proportion of abnormal particles, which is then fed back to the early warning system of the industrial computer for adjustment of fertilizer granulation elements.

7. A fertilizer granulation particle size detection device for implementing the fertilizer granulation particle size detection method according to any one of claims 1-6, characterized in that, The device includes: a fertilizer conveyor belt, a fixed bracket, an industrial camera, a supplementary light, an industrial control computer, and a WEB Internet of Things cloud platform; The fertilizer conveyor belt is installed below the outlet of the fertilizer amino acid granulation roller. The fixed bracket is set above the fertilizer conveyor belt and vertically supports an industrial camera. The fixed bracket is equipped with supplementary lights on both sides. The industrial control computer is connected to the industrial camera through optical fiber to acquire granulation images and then uploads them to the WEB Internet of Things cloud platform to obtain granulation detection results.

8. The fertilizer granulation particle size detection device according to claim 7, characterized in that, Two supplementary lights are provided, which are placed on the left and right sides of the fixed bracket respectively, symmetrically distributed, to illuminate the flow of fertilizer particles on the fertilizer conveyor belt with LED light.

9. The fertilizer granulation particle size detection device according to claim 7, characterized in that, The fixed bracket is a stainless steel bracket, including: an overall support frame, a horizontal crossbar, and a vertical bar; the overall support frame consists of left and right vertical bars, fixed next to the fertilizer conveyor belt; the support frame is provided with a horizontal crossbar, and the vertical bar is fixed to the horizontal crossbar by a buckle, and the industrial camera is fixed to the vertical bar by a buckle. Adjusting the buckles on the crossbar allows the industrial camera to move left and right, and adjusting the buckles on the vertical bar allows the industrial camera to move up and down.

Images