An adaptive control method for an intelligent grinding system for aleurone layers

Abstract

This invention relates to the field of grinding, specifically providing an adaptive control method for an intelligent aleurone layer grinding system. The method includes acquiring raw image data of wheat bran entering the grinding system in real time using an optical imaging device; preprocessing the acquired raw images; extracting features from the preprocessed images using a trained deep learning model, outputting the aleurone layer area ratio A, average thickness T, and surface texture roughness R; dynamically calculating the spindle rotor speed, feed rate, and airflow pressure of the grinding system based on the output data; generating control commands based on the calculated parameters; and sending the control commands to a PLC control system to adjust the operating status of the grinding system. This solution addresses the technical problem of existing grinding equipment lacking intelligent control and real-time optimization.

Description

Technical Field

[0001] This invention relates to the field of grinding, and specifically provides an adaptive control method for an intelligent grinding system for aleurone layers. Background Technology

[0002] The aleurone layer of wheat is the most nutritious part of wheat bran, rich in dietary fiber, protein, minerals, and various vitamins, possessing significant nutritional and health benefits. However, due to the tight bond between the aleurone layer and other bran structures, traditional physical separation methods suffer from the following problems:

[0003] Uneven material distribution during grinding results in low aleurone layer extraction rate; traditional mechanical equipment lacks intelligent control methods, making it impossible to adjust process parameters in real time and meet the needs of different raw material characteristics; severe mechanical wear affects equipment lifespan and product safety; high energy consumption during production and difficulty in achieving fine grading.

[0004] While existing patents for related equipment, such as multi-stage vortex mills and planetary mills, have improved grinding efficiency to some extent, they have not solved the problems of intelligent control and real-time optimization. Furthermore, current technologies do not fully utilize modern digital technologies, resulting in a lack of flexibility and adaptability in the production process. Summary of the Invention

[0005] To overcome the above-mentioned shortcomings, this invention is proposed to provide a solution to the technical problem of the lack of intelligent control and real-time optimization in existing grinding equipment.

[0006] This invention provides an adaptive control method for an intelligent aleurone layer grinding system, comprising the following steps:

[0007] The raw image data of wheat bran entering the grinding system is acquired in real time using optical imaging equipment, and the acquired raw images are preprocessed.

[0008] The trained deep learning model is used to extract features from the preprocessed original image, and outputs the aleurone layer area ratio A, average thickness T, and surface texture roughness R.

[0009] Based on the aleurone layer area ratio A, average thickness T, and surface texture roughness R, the spindle rotor speed, feed rate, and airflow pressure of the grinding system are dynamically calculated.

[0010] Based on the calculated spindle rotor speed, feed rate, and airflow pressure, control commands are generated and sent to the PLC control system to adjust the operating status of the grinding system.

[0011] Furthermore, the preprocessing of the acquired raw image includes denoising, brightness normalization, contrast enhancement, color space conversion, edge detection, and cropping / scaling.

[0012] further,

[0013] The denoising process includes using median filtering to remove salt and pepper noise from the original image;

[0014] The brightness normalization includes unifying the pixel values ​​of the original image after denoising to the range of [0, 1], that is... ,in, This represents the pixel value of a single pixel in the original image after denoising. , This represents the minimum and maximum values ​​of all pixels in the original image after denoising. This represents the normalized pixel value of the original image, ranging from [0, 1].

[0015] The contrast enhancement includes using the CLAHE algorithm to enhance local contrast, i.e. ,in, CLAHE represents the pixel value after contrast enhancement of the original image, and it represents the contrast-limited adaptive histogram equalization algorithm.

[0016] The color space conversion includes converting the contrast-enhanced original image from the RGB color space to the HSV color space, i.e. ,in, Indicates the RGB color channel. Indicates the HSV color channel;

[0017] The edge detection includes extracting material edges for images that have undergone color space conversion using the Canny operator, i.e. Where Canny represents an edge detector;

[0018] The cropping and scaling process includes cropping and scaling the image to a specified size according to the model input requirements.

[0019] Furthermore, the deep learning model uses a pre-trained ResNet50 as the feature extraction backbone network, removes the original classification head, retains the convolutional layers and residual block structure, and connects the CBAM module after the last residual block of ResNet50 to perform channel attention and spatial attention calculations on the output feature map in sequence.

[0020] Furthermore, the output of the deep learning model is configured with classification and segmentation branches.

[0021] The classification branch involves pooling the global feature map output by CBAM into a 512-dimensional vector and outputting a binary classification result of aleurone layer or non-leurone layer through a fully connected layer.

[0022] The segmentation branch uses a U-Net-structured decoder, which fuses low-level features through two upsampling operations and skip connections to output a single-channel pixel-level segmentation mask.

[0023] Furthermore, the loss function for training deep learning models includes:

[0024] Define the segmentation loss ,in, Spred This represents the segmentation scalar value predicted by the model. Strue Represents the actual segmentation mask. The values ​​represent weight coefficients. BCE Loss() refers to the binary cross-entropy loss, and Dice Loss() refers to the Dic coefficient loss.

[0025] Define classification loss ,in, This represents the classification scalar value predicted by the model. The actual classification mask;

[0026] Total loss ,in, Indicates an indicator function, hour, It is 1 if it is 1, otherwise it is 0. The weighting coefficients represent the segmentation loss.

[0027] Furthermore, the dynamic calculation of the spindle rotor speed, feed rate, and airflow pressure of the grinding system based on the aleurone layer area ratio A, average thickness T, and surface texture roughness R includes:

[0028] When the aleurone layer area ratio is less than 40%, the main shaft rotor speed is set to 1200 rpm; when the aleurone layer area ratio is between 40% and 70%, the main shaft rotor speed is set to 1400 rpm; when the aleurone layer area ratio is greater than 70%, the main shaft rotor speed is set to 1600 rpm.

[0029] Calculate feed rate ,in, and These are the weighting coefficients;

[0030] Calculate airflow pressure ,in, and These are the weighting coefficients.

[0031] Furthermore, it also includes using a 3D scanner to collect the particle size distribution, density, and moisture content of wheat bran raw materials, using sensors to monitor the operating status and environmental parameters of the grinding system in real time, and using machine learning algorithms to construct a 3D virtual model of the grinding system based on the collected data.

[0032] Furthermore, it also includes uploading operational data from the production process to the cloud in real time, and using a cloud-based data analysis platform to store and analyze the production data.

[0033] Furthermore, the grinding system includes a grinding device, a grading and sieving device, and an energy management device.

[0034] The working principle and beneficial effects of this invention:

[0035] In implementing the technical solution of this invention, based on the real-time acquired original images of wheat bran entering the grinding system, a trained deep learning model (such as ResNet50+CBAM attention mechanism) is used to extract features from the preprocessed original images, outputting the aleurone layer area ratio, average thickness, and surface texture roughness. The rotor speed, feed rate, and airflow pressure of the grinding system are dynamically adjusted based on the output parameters. This ensures that the equipment always operates under optimal conditions, avoiding over-grinding or under-grinding. Detailed Implementation

[0036] Some embodiments of the present invention are described below. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0037] Example 1

[0038] The adaptive control method of the aleurone layer intelligent grinding system in this embodiment mainly includes the following steps.

[0039] 1. Construct a digital twin model of the grinding system.

[0040] The physical properties of wheat bran raw materials (such as particle size distribution, density, and humidity) are collected using a 3D scanner, and the operating status of the equipment (speed, temperature, humidity, etc.) and environmental parameters (air pressure, workshop temperature and humidity) are monitored in real time using sensors. Based on the collected data, a virtual model of the grinding system is constructed using machine learning algorithms. This model can simulate various working conditions in the actual production process and generate optimal process parameters.

[0041] 2. Obtain raw image data of wheat bran in real time using optical imaging equipment, preprocess the acquired raw images, and extract aleurone layer features.

[0042] Specifically, preprocessing includes noise reduction, brightness normalization, contrast enhancement, color space conversion, edge detection, and cropping / scaling.

[0043] ① Denoising: Use median filtering to remove salt and pepper noise from the original image.

[0044] Specifically, the convolution kernel parameters are determined, and the sliding window size is determined based on these parameters. Assuming the convolution kernel is defined as 5×5, the sliding window is a square neighborhood window centered at pixel (x, y), where 5×5 means the window's width and height are both 5 pixels. This window moves across the image, processing each pixel one by one. For each pixel (x, y) at the center of the window, the brightness values ​​of all pixels within this window are taken, sorted from smallest to largest, and the median is used as the new pixel value for that center point.

[0045] In this embodiment, median filtering effectively eliminates salt-and-pepper noise (random black and white noise spots in the image), which may be generated during industrial camera acquisition. It is a non-linear filter that, while removing noise, better preserves the image's edge information, which is crucial for subsequent edge detection.

[0046] ② Brightness normalization: Unify the pixel values ​​of the original image after noise reduction to the range of [0, 1] to eliminate the influence of lighting differences.

[0047]

[0048] in, This represents the pixel value of a single pixel in the original image after denoising. , This represents the minimum and maximum values ​​of all pixels in the original image after denoising. This represents the normalized pixel value of the original image, ranging from [0,1].

[0049] In this embodiment, the pixel value range of an image is linearly mapped to a fixed interval [0,1], ensuring that each image is independently stretched to its own full dynamic range. This eliminates brightness differences between images acquired at different times and under different lighting conditions. For example, images taken in the morning and afternoon may have different overall brightness; after normalization, their contrast becomes consistent. Providing standardized input for deep learning models contributes to the stability and convergence speed of model training.

[0050] ③ Contrast Enhancement: The CLAHE algorithm is used to enhance local contrast and improve image clarity.

[0051]

[0052] in, Represents the pixel value after contrast enhancement of the original image, and CLAHE represents the contrast-limited adaptive histogram equalization algorithm.

[0053] In this embodiment, CLAHE divides the image into multiple small tiles and performs histogram equalization on each tile individually, thereby enhancing the contrast of local areas. The tile size needs to be defined to be small enough to enhance local details, yet large enough to ensure statistical significance, and is fine-tuned based on actual results. After calculating the histogram of each tile, a threshold is set, and portions of the histogram exceeding this threshold are "cropped" and evenly distributed across the entire histogram interval, effectively suppressing the amplification of background noise. Through contrast enhancement, the local contrast between the aleurone layer and the background in the image can be significantly enhanced, making features more prominent and facilitating model learning.

[0054] ④ Color space conversion: Convert the original image with enhanced contrast from RGB color space to HSV color space, retaining the H and S channels and reducing the impact of lighting changes on color recognition.

[0055]

[0056] in, Indicates the RGB color channel. This indicates the HSV color channel.

[0057] In this embodiment, the RGB color space, composed of the three primary colors of red, green, and blue, is highly sensitive to changes in lighting. In the HSV color space, H (hue) represents the type of color (e.g., red, yellow, green), is least affected by lighting, and is a stable feature for recognizing object colors; S (saturation) represents the purity of the color; and V (value) represents the brightness of the color. Converting to HSV separates color information from brightness information, making the algorithm more robust to changes in lighting. Regardless of light intensity, the "hue" of the aleurone layer is relatively stable.

[0058] The H and S channels are retained because the goal is to identify the aleurone layer (whose color is a stable feature), while the V channel represents brightness, which has already been processed in previous steps and is still susceptible to light, so it can be discarded or used only as an auxiliary channel.

[0059] ⑤ Edge detection: For images that have undergone color space conversion, the Canny operator is used to extract material edges to assist in subsequent feature extraction and segmentation.

[0060]

[0061] Here, Canny represents the edge detector. By setting a high threshold and a low threshold, pixels with gradient values ​​higher than the high threshold are identified as strong edges and are definitely retained. Pixels with gradient values ​​lower than the low threshold are discarded. Pixels with gradient values ​​between the two thresholds are considered weak edges. Only when a weak edge is connected to a strong edge will it be retained as a true edge. Edge detection can highlight the contour and texture information of aleurone layer regions. These edge features can be combined with color features (HSV) to provide richer and more discriminative input information for deep learning models, improving segmentation accuracy.

[0062] ⑥ Cropping and scaling: Based on the model input requirements, crop and scale the image to the specified size (e.g., 512×512).

[0063]

[0064] in, This represents the pixel values ​​after cropping and scaling the original image. A predefined rectangular region ensures the image only contains the material on the conveyor belt, eliminating unnecessary background (such as equipment borders and shadows) and reducing interference. Scaling to 512×512 to adapt to deep learning models typically requires a fixed-size input, ensuring each image input to the model has the same size and aspect ratio.

[0065] 3. Use a trained deep learning model (such as ResNet50+CBAM attention mechanism) to extract features from the preprocessed image and output the aleurone layer area ratio A, average thickness T, and surface texture roughness R.

[0066] In one implementation, the deep learning model uses a pre-trained ResNet50 as the backbone for feature extraction. The original classification head is removed, while the convolutional layers and residual block structure are retained. A CBAM (Convolutional Block Attention Module) is connected after the last residual block of the ResNet50 to sequentially calculate channel attention and spatial attention on the output feature map, enhancing the saliency of key features. Channel attention generates channel weights (0~1) through global average pooling and fully connected layers, strengthening key channel features. Spatial attention generates spatial weight masks through convolutional layers, focusing on the edges and texture features of the aleurone layer region.

[0067] In one implementation, the output of the deep learning model is configured with a classification branch and a segmentation branch. The classification branch determines whether the layer is an aleurone layer, and the segmentation branch generates a pixel-level segmentation mask. Specifically:

[0068] Classification branch: The global feature map output by CBAM is averaged and pooled into a 512-dimensional vector, and then passed through a fully connected layer (nn.Linear(512)). 7 Output the binary classification results (aleurone layer / non-aleurone layer) for 7, 1).

[0069] Segmentation Branch: The decoder, which adopts the U-Net structure, fuses low-level features through two upsampling (UpBlock) and skip connections, and finally outputs a single-channel pixel-level segmentation mask.

[0070] In one implementation, deep learning model training includes:

[0071] ①Acquire training images and perform preprocessing and data augmentation on the training images;

[0072] Preprocessing: The image is normalized to the [0,1] range in brightness and converted to the HSV color space to reduce lighting interference.

[0073] Data augmentation: Apply random flipping, rotation (±15°), and lighting changes (randomly adjust brightness / contrast) to improve generalization and avoid overfitting.

[0074] ② The segmentation loss is a weighted combination of Dice Loss and BCE Loss: ,in, Spred This represents the segmentation scalar value predicted by the model. Strue Represents the actual segmentation mask (0 or 1). This represents the weighting coefficient, used to balance the two types of loss, and is usually set to 0.5.

[0075] The classification loss uses binary cross-entropy loss: ,in, This represents the classification scalar value predicted by the model. The mask represents the true classification (0 or 1). BCE Loss (Binary CrossEntropy Loss) is suitable for binary classification problems and measures the distance between the probability output by the sigmoid function and the true label. Dice Loss refers to the Dic coefficient loss, which is widely used in image segmentation tasks and measures the degree of overlap between the predicted segmented region and the true segmented region.

[0076] Total loss ,in, Indicates an indicator function, hour, It is 1 if it is 1, otherwise it is 0. This represents the weighting coefficient of the segmentation loss, used to balance the two loss terms, and is usually set between 0.2 and 0.5.

[0077] 4. Based on the model output and preset target values, dynamically calculate the operating parameters of the grinding system, including spindle rotor speed, feed rate, and airflow pressure. Specifically:

[0078] The spindle rotor speed is dynamically adjusted according to the aleurone layer area ratio. When the aleurone layer area ratio is less than 40%, the spindle rotor speed is set to 1200 rpm; when the aleurone layer area ratio is between 40% and 70%, the spindle rotor speed is set to 1400 rpm; and when the aleurone layer area ratio is greater than 70%, the spindle rotor speed is set to 1600 rpm.

[0079] The feed rate is calculated based on the aleurone layer area ratio A and the average thickness T. ,in, and These are the weighting coefficients. The value range is 0.8 to 1.2. The value range is 0.5 to 0.8.

[0080] Calculate airflow pressure based on average thickness T and surface roughness R. ,in, and These are the weighting coefficients. The value range is 0.6 to 1.0. The value range is 0.3 to 0.6.

[0081] 5. Generate control commands based on the operating parameters of the grinding system, and send the control commands to the PLC control system via Modbus TCP / OPC UA protocol to adjust the operating status of the grinding system; at the same time, feed real-time data back to the digital twin model for status monitoring and visualization.

[0082] 6. Operational data throughout the entire production process (including energy consumption, output, quality indicators, etc.) is uploaded to the cloud in real time to form a complete production record. This record is used for historical data analysis, trend prediction, and process optimization. The cloud data analysis platform is used to store and analyze the production data. Through historical data analysis and trend prediction, a scientific basis is provided for process improvement and new product development.

[0083] Based on the above steps, the combination of digital twin technology and the grinding system enables real-time monitoring and dynamic optimization of the entire grinding process. Using real-time acquired images of wheat bran entering the grinding system, a trained deep learning model (such as ResNet50+CBAM attention mechanism) extracts features from the preprocessed images, outputting the aleurone layer area ratio, average thickness, and surface texture roughness. The rotor speed, feed rate, and airflow pressure of the grinding system are dynamically adjusted based on the output parameters. This ensures the equipment always operates under optimal conditions, avoiding over-grinding or under-grinding. The application of digital twin technology allows for seamless integration between the virtual model and actual production, achieving a prediction accuracy of over 95% and significantly reducing human error. Data on energy consumption, output, and quality throughout the production process are uploaded to the cloud in real time, forming a complete production record for easy traceability and optimization. The digital twin platform supports historical data analysis and trend prediction, providing a scientific basis for process improvement and new product development.

[0084] It should be noted that although the steps in the above embodiments are described in a specific order, those skilled in the art will understand that in order to achieve the effects of the present invention, different steps do not necessarily have to be executed in such an order. They can be executed simultaneously (in parallel) or in other orders, and these variations are all within the scope of protection of the present invention.

[0085] Example 2

[0086] An adaptive control method for an intelligent aleurone layer grinding system proposed in Example 1, wherein the grinding system includes:

[0087] 1. Grinding device

[0088] The multi-stage stepped grinding chamber design, combined with a reverse-rotating vortex shearing assembly, improves grinding efficiency and product uniformity. The rotor blades are made of high-strength, wear-resistant ceramic material, effectively extending the equipment's service life and reducing the risk of product contamination due to wear. An integrated cooling system uses circulating water to lower the grinding chamber temperature, preventing product quality degradation caused by overheating.

[0089] 2. Grading and screening device

[0090] A grading and sieving device is added after grinding to separate materials of different particle sizes through centrifugal force and airflow, meeting the needs of various application scenarios. The grading device adopts a double-layer structure design, with an upper screen of 30-50 mesh and a lower screen of 100-110 mesh. The ground material undergoes initial separation by centrifugal force and airflow. Coarse particles are returned to the grinding device for secondary processing, while fine particles are sieved and packaged to ensure that the final product has uniform particle size and meets standards. The sieving device has a built-in vibrating motor, and the amplitude and frequency can be automatically adjusted according to the material characteristics to further improve the separation accuracy.

[0091] 3. Energy management device

[0092] The heat and kinetic energy generated during the grinding process are converted into usable energy through a heat exchanger to heat the feed hopper or other auxiliary equipment. A dust collection and recycling system is also included to reduce resource waste and environmental pollution.

[0093] The technical solutions of the present invention have now been described. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions resulting from these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. An adaptive control method for an intelligent aleurone layer grinding system, characterized in that, Includes the following steps: The raw image data of wheat bran entering the grinding system is acquired in real time using optical imaging equipment, and the acquired raw images are preprocessed. The trained deep learning model is used to extract features from the preprocessed original image, and outputs the aleurone layer area ratio A, average thickness T, and surface texture roughness R. Based on the aleurone layer area ratio A, average thickness T, and surface texture roughness R, the spindle rotor speed, feed rate, and airflow pressure of the grinding system are dynamically calculated. Based on the calculated spindle rotor speed, feed rate and airflow pressure, control commands are generated and sent to the PLC control system to adjust the operating status of the grinding system. The deep learning model uses a pre-trained ResNet50 as the feature extraction backbone network, removes the original classification head, retains the convolutional layers and residual block structure, and connects the CBAM module after the last residual block of ResNet50 to perform channel attention and spatial attention calculations on the output feature map in sequence. The output of the deep learning model is configured with classification and segmentation branches. The classification branch involves pooling the global feature map output by CBAM into a 512-dimensional vector and outputting a binary classification result of aleurone layer or non-leurone layer through a fully connected layer. The segmentation branch uses a U-Net-structured decoder, which fuses low-level features through two upsampling operations and skip connections to output a single-channel pixel-level segmentation mask. The dynamic calculation of the spindle rotor speed, feed rate, and airflow pressure of the grinding system based on the aleurone layer area ratio A, average thickness T, and surface roughness R includes: When the aleurone layer area ratio is less than 40%, the main shaft rotor speed is set to 1200 rpm; when the aleurone layer area ratio is between 40% and 70%, the main shaft rotor speed is set to 1400 rpm; when the aleurone layer area ratio is greater than 70%, the main shaft rotor speed is set to 1600 rpm. Calculate feed rate ,in, and These are the weighting coefficients. The value range is 0.8 to 1.

2. The value range is 0.5 to 0.8; Calculate airflow pressure ,in, and These are the weighting coefficients. The value range is 0.6 to 1.

0. The value range is 0.3 to 0.

6.

2. The adaptive control method for an intelligent aleurone layer grinding system according to claim 1, characterized in that, The preprocessing of the acquired raw images includes noise reduction, brightness normalization, contrast enhancement, color space conversion, edge detection, and cropping / scaling.

3. The adaptive control method for an intelligent aleurone layer grinding system according to claim 2, characterized in that, The denoising process includes using median filtering to remove salt and pepper noise from the original image; The brightness normalization includes unifying the pixel values ​​of the original image after denoising to the range of [0, 1], that is... ,in, This represents the pixel value of a single pixel in the original image after denoising. , This represents the minimum and maximum values ​​of all pixels in the original image after denoising. This represents the normalized pixel value of the original image, ranging from [0, 1]. The contrast enhancement includes using the CLAHE algorithm to enhance local contrast, i.e. ,in, CLAHE represents the pixel value after contrast enhancement of the original image, and it represents the contrast-limited adaptive histogram equalization algorithm. The color space conversion includes converting the contrast-enhanced original image from the RGB color space to the HSV color space, i.e. ,in, Indicates the RGB color channel. Indicates the HSV color channel; The edge detection includes extracting material edges for images that have undergone color space conversion using the Canny operator, i.e. Where Canny represents an edge detector; The cropping and scaling process includes cropping and scaling the image to a specified size according to the model input requirements.

4. The adaptive control method for an intelligent aleurone layer grinding system according to claim 1, characterized in that, The loss functions for training deep learning models include: Define the segmentation loss ,in, Spred This represents the segmentation scalar value predicted by the model. Strue Represents the actual segmentation mask. The values ​​represent weight coefficients. BCE Loss() refers to the binary cross-entropy loss, and Dice Loss() refers to the Dic coefficient loss. Define classification loss ,in, This represents the classification scalar value predicted by the model. The actual classification mask; Total loss ,in, Indicates an indicator function, hour, It is 1 if it is true, otherwise it is 0. The weighting coefficients represent the segmentation loss.

5. The adaptive control method for an intelligent aleurone layer grinding system according to claim 1, characterized in that, It also includes using a 3D scanner to collect the particle size distribution, density, and moisture content of wheat bran raw materials, using sensors to monitor the operating status and environmental parameters of the grinding system in real time, and using machine learning algorithms to construct a 3D virtual model of the grinding system based on the collected data.

6. The adaptive control method for an intelligent aleurone layer grinding system according to claim 1, characterized in that, It also includes uploading operational data from the production process to the cloud in real time, and using a cloud-based data analysis platform to store and analyze the production data.

7. The adaptive control method for an intelligent aleurone layer grinding system according to claim 1, characterized in that, The grinding system includes a grinding device, a grading and sieving device, and an energy management device.