Multimodal perception and dynamic co-decision making for wheat milling method, apparatus, and media

Abstract

This application belongs to the field of grain processing technology and discloses a wheat milling method, equipment, and medium based on multimodal perception and dynamic collaborative decision-making. The method deploys a high-speed camera and a near-infrared multi-point online analyzer at the mill discharge point to acquire real-time image data of wheat fragments and wheat flour quality index data, respectively. A hybrid neural network (CNN-LSTM) is used to process the image data and quality index data to extract spatial-temporal features, generating real-time quality prediction values, quality change trend warning signals, and morphology-quality correlation coefficients. Based on a Markov decision process framework, the real-time quality prediction values, warning signals, and correlation coefficients are used as state inputs. An optimal strategy is solved through a reinforcement learning algorithm to generate process control instructions for real-time adjustment of continuous actions, discrete actions, and constraints during wheat milling, achieving closed-loop control of the entire chain from "real-time detection of wheat flour status to adaptive process control and dynamic optimization of the milling path."

Description

Technical Field

[0001] This invention relates to the field of grain processing technology, and in particular to a wheat milling method, equipment and medium based on multimodal perception and dynamic collaborative decision-making. Background Technology

[0002] Wheat, as my country's second largest staple food, is an important source of carbohydrates, trace elements, and vitamins for the human body. Wheat milling is the core link in grain processing, and its technological precision directly affects flour quality, flour yield, and economic benefits. Traditional milling processes mainly rely on manual experience to adjust key parameters, resulting in problems such as slow response, poor consistency, and high energy consumption. In recent years, the country has attached great importance to the intelligent and green transformation of grain processing equipment, providing policy support for new mills with intelligent sensing and feedback functions. Therefore, it is urgent to develop precise and intelligent wheat milling systems and equipment for real-time monitoring and dynamic adjustment of key steps and parameters in the wheat grain milling process.

[0003] In existing technologies, such as the Chinese invention patent CN114636646B which discloses an automatic online detection and feedback system for broken particle size, this system relies on multi-stage vibratory screening and weighing linkage, resulting in a complex process and a response delay of minutes, making it difficult to meet the requirements of high-precision milling. Another Chinese invention patent CN112317010B discloses a smart wheat milling system – an electronic milling master system. This system uses an online near-infrared analyzer to detect quality and establishes a mathematical model to optimize the wheat conditioning process. However, it is limited to the conditioning step, lacks the ability to collaboratively optimize multiple indicators, and has a low level of intelligence, lacking a feedback and decision-making mechanism driven by deep AI algorithms. The above methods generally suffer from minute-level response delays due to the use of screening and weighing to measure particle size, making it impossible to provide real-time feedback on the state of wheat flour in each process.

[0004] In summary, while existing technologies have made progress in rapid detection and automated control, they have failed to deeply integrate AI algorithms (such as deep learning and reinforcement learning) to achieve millisecond-level linkage between material crushing status and parameter decisions. The industry urgently needs a closed-loop control technology that can achieve "real-time detection of wheat flour status - adaptive process control - dynamic optimization of flour path" to solve the problems of response lag, lack of feedback and decision-making mechanisms driven by deep AI algorithms, and difficulty in achieving closed-loop control of the entire chain in existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a wheat milling method, equipment, and medium based on multimodal perception and dynamic collaborative decision-making, to solve the problems of slow response, lack of deep AI algorithm-driven feedback and decision-making mechanisms, and difficulty in achieving closed-loop control throughout the entire chain in existing technologies. This method uses a high-speed camera and a near-infrared multi-point online analyzer to rapidly detect the condition of wheat flour during the milling process. A hybrid neural network composed of convolutional neural networks and long short-term memory networks is used to accurately extract and predict the physical characteristics and quality indicators of wheat flour in each milling step. Then, a reinforcement learning algorithm based on Markov decision-making is used to solve for the optimal strategy, dynamically deciding and adjusting the process parameters of wheat milling, ultimately achieving precision and intelligent wheat milling.

[0006] In a first aspect, the present invention provides a wheat milling method based on multimodal sensing and dynamic collaborative decision-making, comprising the following steps: At the discharge point of the mill, images of wheat fragments are captured in real time by a high-speed camera to obtain image data, and the quality index data of wheat flour from multiple grinding processes are simultaneously and rapidly measured by a near-infrared multi-point online analyzer equipped with integrated sensors. A hybrid neural network consisting of a convolutional neural network and a long short-term memory network is used to process the image data and the quality index data. The convolutional neural network is used to extract the spatial features of the visual morphology of wheat flour from the image data, and the long short-term memory network is used to capture the dynamic changes in the particle size distribution of wheat flour and the long-term dependence of quality indicators, thereby generating real-time quality prediction values, quality change trend warning signals and morphology-quality correlation coefficients as the basis for quantitative decision-making. Using the Markov decision process framework, the wheat milling process is modeled as a decision problem involving changes in the state of wheat flour and dynamic process parameters. The real-time quality prediction value, the quality change trend warning signal, and the morphology-quality correlation coefficient are used as part of the current state vector. A reward function is constructed with the quality compliance rate as the core. The optimal strategy is solved by reinforcement learning algorithm to generate process control instructions for real-time adjustment of continuous actions, discrete actions, and constraints in the wheat milling process.

[0007] As an optional implementation of the first aspect of this application, the step of acquiring image data by capturing wheat fragment images in real time using a high-speed camera specifically includes: arranging a high-speed camera consisting of a pinhole sensor, a lens, and a computer under the rollers of each grinding process; the pinhole sensor is selected with an IP66 or higher protection rating; the high-speed camera has a resolution of 2560*1920 or higher, an acquisition speed of more than 1000fps, and a minimum exposure time of less than 100ns, so as to capture fragment images in real time and dynamically track changes in the particle size of the fragments.

[0008] As an optional implementation of the first aspect of this application, after acquiring image data, an improved YOLOv11 model is used to perform real-time segmentation of wheat fragment images. The improvements made to the standard YOLOv11 architecture include: embedding cross-stage local context modules after the 3rd and 4th CSPBlocks in the backbone network; enhancing the model's global context awareness of adherent particles and its ability to capture local details of edge cracks through a dual-path design of parallel dilated convolutions and standard convolutions; introducing an adaptive weighted feature pyramid at the feature fusion neck and dynamically optimizing the feature fusion process using a collaborative attention mechanism, mathematically expressed as: F fused = Conv(CA(F high ) ⊙ F high + SA(F low )⊙ F low ); where F fused This represents the fused feature map, CA() and SA() represent the channel attention and spatial attention functions respectively, ⊙ represents element-wise multiplication, and F high and F low represents the feature maps from high-level semantic features and low-level detail features, respectively, and Conv represents the convolutional fusion operation; the detection head is extended into a decoupled instance segmentation head, while retaining the original classification and regression branches, a mask prediction branch is added in parallel, RoI Align is used to extract target features from high-resolution feature maps, and a pixel-level mask is output through a lightweight decoder, and the segmentation accuracy is optimized by combining segmentation loss during training.

[0009] As an optional embodiment of the first aspect of this application, the step of simultaneously and rapidly measuring wheat flour through multiple milling processes using a near-infrared multi-point online analyzer equipped with integrated sensors to obtain quality index data specifically includes: the near-infrared multi-point online analyzer simultaneously tests four or more measurement points; the integrated sensors are arranged below the rollers of each milling process for rapidly measuring color, ash content, protein content, moisture content, bran content, and temperature; the integrated sensors are selected with an IP66 or higher protection rating; the analyzer has a wavelength range of 780-2500nm and a resolution greater than or equal to 4cm. -1 The scanning speed is greater than 1 scan per 20ms, and it is equipped with a high-purity nitrogen or dry air purging system.

[0010] As an optional implementation of the first aspect of this application, the step of processing the image data and the quality index data using a hybrid neural network composed of a convolutional neural network and a long short-term memory network specifically includes: taking the obtained real-time wheat fragment segmentation mask image sequence and its derived morphological vectors as multi-channel inputs; the convolutional neural network adopts an encoder structure with four convolutional layers, each layer containing 3×3 convolution, batch normalization and ReLU activation, and downsampling through 2×2 max pooling to encode each frame image into a 256-dimensional feature vector; the long short-term memory network receives the feature vector sequence arranged in time steps, its core consisting of two layers of bidirectional long short-term memory network units, each layer having 128 hidden units, and capturing long-term dependencies through its input gate, forget gate and output gate gating mechanism; feeding the output of the long short-term memory network at the final time step into a fully connected prediction head containing two branches, one branch outputting the real-time quality prediction value and the trend warning signal through softmax, and the other branch outputting the morphology-quality correlation coefficient through a linear layer.

[0011] As an optional implementation of the first aspect of this application, the wheat milling process is modeled as a decision problem involving changes in wheat flour state and dynamic process parameters using a Markov decision process framework. Specifically, this includes: constructing a state vector that integrates online image segmentation and morphology quantization results, key operating parameters of each grinding and grading device, and real-time predicted values ​​and recent historical trends; defining an action space that includes continuous micrometer-level fine-tuning of the milling gap and discrete selection of air-separation flow rate, with built-in equipment safety constraints; and constructing a reward function, which is a weighted score integrating multiple objective quality indicators such as ash content control accuracy, flour yield, and energy consumption, mathematically expressed as: r t =w1 R 品质 +w2 R 效率 +w3 R 能耗 Among them, R 品质 R is used to reward high ash content control precision and increased proportion of premium powder. 效率 R is used to reward increased production of qualified powder per unit time. 能耗 Used to penalize an increase in energy consumption per unit output; a deep neural network trained with a proximal policy optimization algorithm is used as the policy function πθ(a t |s t The network is in its current state s t Given the input, output the optimal action a. t The probability distribution is used to generate the process control instructions.

[0012] As an optional implementation of the first aspect of this application, the state vector specifically includes the area distribution of fragments of each particle size, average profile complexity, current rolling gap, feed flow rate, motor load, current ash content prediction value, estimated powder yield, and recent historical trends of quality indicators.

[0013] As an optional implementation of the first aspect of this application, each sub-reward in the reward function is normalized, and the weights w1, w2, and w3 are configured according to the production focus; by training, the agent learns a strategy network that maximizes the long-term cumulative discount reward, which not only focuses on immediate quality but also takes into account the stability and efficiency of the entire production cycle, thereby achieving adaptive, multi-objective dynamic optimization of the powder making process.

[0014] In a second aspect, embodiments of this application provide an electronic device, which includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of the method described in the first aspect.

[0015] Thirdly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention is based on a high-speed camera and a near-infrared multi-point online analyzer to quickly detect the morphology and quality of wheat flour. By using multimodal perception of visual morphology and multispectral characteristics, it replaces the traditional multi-stage sieving method, greatly reducing the response delay and achieving millisecond-level feedback on wheat flour production status.

[0017] (2) This invention adopts an AI-driven method of deep learning and reinforcement learning, and a hybrid neural network composed of convolutional neural network and long short-term memory network is deeply integrated to accurately extract, analyze and predict the physical characteristics and quality indicators of wheat flour in each grinding process. The optimal strategy is solved by reinforcement learning algorithm of Markov decision and the wheat milling process is adjusted in real time. This method is more scientific and intelligent.

[0018] (3) This invention uses multimodal real-time sensing of the state of wheat flour in each grinding process, dynamically coordinates and decides the process parameters of the entire wheat milling process, and performs closed-loop control of the entire chain of "real-time detection of wheat flour state - adaptive process control - dynamic optimization of flour path", thereby realizing the precision and intelligence of wheat milling. Attached Figure Description

[0019] Figure 1 This is a flowchart of a wheat milling method based on multimodal perception and dynamic collaborative decision-making according to an embodiment of the present invention; Figure 2This is a structural diagram of the backbone network in the improved YOLOv11 model in this embodiment of the invention; Figure 3 This is a structural diagram of the Cross-Stage Local Context Module (CSLC) in the backbone network in an embodiment of the present invention; Figure 4 This is a structural diagram of the feature fusion neck in the improved YOLOv11 model in this embodiment of the invention; Figure 5 This is a structural diagram of the Adaptive Weighted Feature Pyramid (AWF-FPN) in the feature fusion neckline of this invention embodiment; Figure 6 This is a structural diagram of the decoupled instance segmentation head in the improved YOLOv11 model in this embodiment of the invention; Figure 7 This is a flowchart of the dynamic coordination decision-making process optimization sequence within the Markov Decision Process (MDP) framework in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0022] Please see Figure 1 This is a flowchart illustrating a wheat milling method based on multimodal perception and dynamic collaborative decision-making, provided by an embodiment of the present invention. The method may include the following steps: S1: At the discharge point of the mill, images of wheat fragments are captured in real time by a high-speed camera to obtain image data, and the quality index data of wheat flour from multiple grinding processes are simultaneously and rapidly measured by a near-infrared multi-point online analyzer equipped with integrated sensors.

[0023] In this step, the high-speed camera consists of a pinhole sensor, a lens, and a computer. The pinhole sensor is positioned below the rollers in each grinding process to capture fragment images in real time and dynamically track changes in the particle size of the broken grains. The pinhole sensor must be an IP66 or higher protection rating device to prevent dust ingress and adapt to the high-dust, high-load environment of a flour mill. Furthermore, the high-speed camera has a resolution of 2560*1920 or higher, a data acquisition speed greater than 1000 fps, and a minimum exposure time of less than 100 ns.

[0024] Near-infrared multi-point online analyzers require simultaneous testing of four or more measurement points. Integrated sensors are positioned beneath the rollers at each milling stage, enabling rapid determination of key quality indicators for wheat flour, including color (whiteness), ash content, protein content, moisture content, bran content, and temperature. The integrated sensors must be of IP66 or higher protection rating. Furthermore, the near-infrared multi-point online analyzer has a measurable wavelength range of 780-2500 nm and a resolution greater than or equal to 4 cm⁻¹. -1 It has a scanning speed of more than 1 scan per 20 ms, high accuracy and repeatability, maintains constant ambient temperature and humidity, and is equipped with a high-purity nitrogen or dry air purging system.

[0025] Furthermore, after acquiring the image data, an improved YOLOv11 model is used to perform real-time segmentation of the wheat fragment images. The improvements made to the standard YOLOv11 architecture include: First, such as Figure 2 As shown, after the 3rd and 4th CSPBlocks in the backbone network, there is an embedded cross-stage local context module (CSLC), as follows: Figure 3 As shown, by using a dual-path design of parallel dilated convolution and standard convolution, the model's ability to simultaneously enhance its global context awareness of adhering particles and capture local details of edge cracks is improved. Second, such as Figure 4 As shown, an adaptive weighted feature pyramid (AWF-FPN) is introduced at the neck of the feature fusion process, as follows: Figure 5 As shown, the collaborative attention mechanism is used to dynamically optimize the feature fusion process and highlight key regions. Its mathematical expression can be simplified to: F fused = Conv(CA(F high ) ⊙ F high + SA(F low ) ⊙ F low ), where F fused This represents the fused feature map, CA() and SA() represent the channel attention and spatial attention functions respectively, ⊙ represents element-wise multiplication, and F high and F lowThese represent feature maps from high-level semantic features and low-level detail features, respectively, and Conv represents the convolutional fusion operation; Thirdly, such as Figure 6 As shown, the detection head is extended into a decoupled instance segmentation head. While retaining the original classification and regression branches, a mask prediction branch is added in parallel. RoI Align is used to extract target features from high-resolution feature maps, and a lightweight decoder outputs pixel-level masks. During training, a segmentation loss (Dice Loss) is incorporated to optimize segmentation accuracy under foreground-background imbalance. After training on a professional dataset and optimization with TensorRT deployment, the model can achieve real-time processing of more than 200 fps on an embedded platform, with an average segmentation mask accuracy (Mask AP@50:95) of 92.1%. The quantization error of the calculated fragmented area ratio is less than 0.5%, providing an accurate and real-time data foundation for subsequent morphological analysis.

[0026] S2: A hybrid neural network consisting of a convolutional neural network (CNN) and a long short-term memory network (LSTM) is used to process the image data and the quality index data. The CNN is used to extract the spatial features of the visual morphology of wheat flour from the image data, and the LSTM is used to capture the dynamic changes in the particle size distribution of wheat flour and the long-term dependence of quality indicators, thereby generating real-time quality prediction values, quality change trend warning signals, and morphology-quality correlation coefficients as the basis for quantitative decision-making.

[0027] In this step, a hybrid neural network (CNN-LSTM) composed of a convolutional neural network (CNN) and a long short-term memory network (LSTM) is used to deeply couple and analyze the visual morphology and time-series characteristics of wheat flour through an end-to-end multimodal data processing flow. The specific implementation is as follows: First, the real-time wheat fragment segmentation mask image (sequence) obtained in step S1 and its derived morphological vectors (such as grain size distribution histogram and contour complexity index) are used as multi-channel inputs.

[0028] The CNN module employs an encoder structure with four convolutional layers. Each layer contains 3×3 convolutions, batch normalization, and ReLU activation, and downsampling is performed using 2×2 max pooling. This design automatically extracts spatial features from local textures (such as bran cracks and starch exposure) to global structures (such as fragment aggregation) through hierarchical convolutional kernels, ultimately encoding each frame of image into a 256-dimensional feature vector.

[0029] The LSTM module receives a sequence of CNN feature vectors arranged in time steps. Its core consists of a two-layer bidirectional LSTM unit with 128 hidden units in each layer. Through its gating mechanism of input gate, forget gate and output gate, the network can selectively remember and forget, effectively capturing long-term dependencies such as gradual changes in particle size distribution and cumulative effects of quality indicators during the fragmentation process.

[0030] Finally, the output of the final time step of the LSTM is fed into a fully connected prediction head, which contains two branches: one branch outputs real-time quality levels (such as gray content and accuracy prediction values) and trend warning signals through softmax; the other branch outputs a set of morphology-quality correlation coefficients through a linear layer to quantify the influence weight of each morphological feature (such as the proportion of a specific granularity range and the average crack length) on key quality indicators.

[0031] The entire model uses mean squared error and cross-entropy as loss functions and is trained on a dataset containing tens of thousands of milling process sequences, ensuring the robustness and interpretability of its predictions and ultimately providing a quantitative basis for process decision-making.

[0032] S3: Using the Markov Decision Process (MDP) framework, the wheat milling process is modeled as a decision problem involving changes in the state of wheat flour and dynamic process parameters. The real-time quality prediction value, the quality change trend warning signal, and the morphology-quality correlation coefficient are used as part of the current state vector. A reward function is constructed with the quality compliance rate as the core. The optimal strategy is solved through a reinforcement learning algorithm to generate process control instructions for real-time adjustment of continuous actions, discrete actions, and constraints in the wheat milling process.

[0033] It should be noted that the state of wheat flour encompasses aspects such as its morphology and quality indicators; the process parameters include raw material characteristics (hardness, protein content, etc.), moisture content of the wheat flour, grinding pressure (roller gap), and flour distribution ratio. The quality compliance rate covers the basic composition of wheat flour, particle size, ash content control precision, yield of premium flour, and energy consumption.

[0034] In this step, such as Figure 7 As shown, a Markov Decision Process (MDP) framework is constructed: The state vector integrates the online image segmentation and morphological quantization results from step S1 (such as the area distribution of fragments of each particle size and the average contour complexity), the key operating parameters of each grinding and grading equipment (such as the current rolling gap, feed flow rate, and motor load), and the real-time predicted values ​​(such as the current ash content and estimated powder output) and recent historical trends based on the CNN-LSTM model output from step S2. The operating space includes continuous micron-level fine-tuning of the roll gap, discrete selection of air separation flow rate, etc., and has built-in equipment safety constraints; The reward function integrates weighted scores from multiple quality indicators such as ash content control accuracy, powder yield, and energy consumption. It can be simplified into a weighted comprehensive function based on the quality indicators output in step S2: r t =w1 R 品质 +w2 R 效率 +w3 R 能耗 Among them, R 品质 The reward is high precision in ash content control and an increased proportion of premium powder; R 效率 The reward increases the output of qualified powder per unit time; R 能耗 Punish the increase in energy consumption per unit of output.

[0035] Each sub-reward is normalized, and the weight w is configured according to the production focus. The system uses the Proximal Policy Optimization (PPO) algorithm to train a deep neural network as the policy function πθ(a t |s t The network is in its current state. t Given the input, output the optimal action a. t The probability distribution is determined. Training is conducted in a simulated environment and on historical production data, enabling the agent to learn to maximize long-term cumulative discount rewards, that is, to focus not only on immediate quality but also on the stability and efficiency of the entire production cycle. The agent is trained to learn a policy network that maximizes long-term cumulative rewards, thereby automatically generating and executing optimal process adjustment instructions based on real-time production status, ultimately achieving adaptive, multi-objective dynamic optimization of the milling process.

[0036] Example 1 The present invention provides a fully intelligent upgrade for a production line at Henan Keming Flour Industry. Specifically, integrated sensors of a near-infrared multi-point online analyzer are installed in the raw material silo and wheat preparation silo. Furthermore, pinhole sensors from high-speed cameras and integrated sensors of the near-infrared multi-point online analyzer are installed in each grinding process. A large-scale numerical simulation super-graphics workstation equipped with CNN / LTSM / MDP deep learning AI algorithms drives the entire wheat milling process through AI, enabling closed-loop control of the entire chain from "real-time detection of wheat flour status to adaptive process control and dynamic optimization of flour path." Using strong gluten wheat Zhengmai 7698 (hard wheat) as raw material, the specific operation is as follows: (1) Install pinhole sensors of high-speed cameras in each grinding process, and combine them with the improved YOLOv11 model to segment wheat fragment images in real time, so as to accurately identify the shape and particle size of each small fragment (particle) in the wheat flour. (2) Install the integrated sensor of the near-infrared multi-point online analyzer in the raw material warehouse, the wheat rinsing warehouse and each grinding process to quickly measure and analyze multiple quality indicators of wheat flour obtained in each process. (3) Using the Alpha760 super graphics workstation, the visual morphology and quality indicators of wheat flour are analyzed by hierarchical feature extraction analysis of convolutional kernels based on convolutional neural network (CNN); based on long short-term memory network (LSTM), the dynamic changes of wheat flour particle size distribution and the long-term dependence of wheat flour quality indicators are captured by gating mechanism, and real-time quality prediction value, trend warning and morphology-quality correlation coefficient are output to provide quantitative basis for decision-making. (4) The reward function is constructed with the quality compliance rate of wheat flour composition (ash content, protein content, etc.), particle size, and flour yield as the core. The wheat milling process is modeled as a decision problem of the state change of wheat flour and dynamic process parameters obtained from each milling process through the Markov decision process (MDP) framework. The optimal strategy is solved by reinforcement learning algorithm, and the continuous action, discrete action and constraint conditions are adjusted in real time to optimize the decision-making process of wheat milling.

[0037] Example 2 A production line of Wudeli Flour Group underwent a fully intelligent upgrade and trial operation according to the method of this invention. Yangmai 13 (soft wheat) was selected as the raw material, and the specific operation is as follows: (1) Install pinhole sensors of high-speed cameras in each grinding process, and combine them with the improved YOLOv11 model to segment wheat fragment images in real time, so as to accurately identify the shape and particle size of each small fragment (particle) in the wheat flour. (2) Install the integrated sensor of the near-infrared multi-point online analyzer in the raw material warehouse, the wheat rinsing warehouse and each grinding process to quickly measure and analyze multiple quality indicators of wheat flour obtained in each process. (3) Using the UltraLAB A350 high-performance multi-purpose graphics workstation, the visual morphology and quality indicators of wheat flour are analyzed by hierarchical feature extraction analysis of convolutional kernels based on convolutional neural network (CNN); based on long short-term memory network (LSTM), the dynamic changes of wheat flour particle size distribution and the long-term dependence of wheat flour quality indicators are captured by gating mechanism, and real-time quality prediction value, trend warning and morphology-quality correlation coefficient are output to provide quantitative basis for decision-making. (4) The reward function is constructed with the quality compliance rate of wheat flour composition (ash content, protein content, etc.), particle size, and flour yield as the core. The wheat milling process is modeled as a decision problem of the state change of wheat flour and dynamic process parameters obtained from each milling process through the Markov decision process (MDP) framework. The optimal strategy is solved by reinforcement learning algorithm, and the continuous action, discrete action and constraint conditions are adjusted in real time to optimize the decision-making process of wheat milling.

[0038] Comparative Example 1 A smart wheat milling system - electronic miller: Using Zhengmai 7698 (hard wheat), a high-gluten wheat variety, as raw material, the specific operation is as follows: (1) The amount of water added to the flour is automatically controlled based on the flour moisture content by real-time monitoring of flour moisture content through online near-infrared spectroscopy; (2) Based on the requirements of different grinding mills, different measuring units are used to automatically sieve and weigh the obtained samples, and measure the scraping rate, powder extraction rate and wheat flour particle size distribution of the grinding mill; (3) Based on the basic requirements of the target flour, the flour path is switched according to the online near-infrared detection indicators such as ash content, whiteness, and protein content. After switching, the flour path is redistributed according to the online near-infrared detection results to obtain the target flour. (4) Based on the interrelationships between data collected on raw materials, equipment, weather, time, and processes, the impact on the quality and yield of the final product is studied, and mathematical modeling is performed. Through the model and current production data, the key control points of the production process are adjusted, the production process is dynamically optimized, the production process is precisely controlled, and the product powder yield is improved.

[0039] Comparative Example 2 Kemin Flour Processing Production Line – A production line capable of processing 600 tons of wheat flour per day: Using Zhengmai 7698 (hard wheat), a high-gluten wheat variety, as raw material, the specific operation is as follows: The production line most widely used by Henan Kemin Flour Industry is used as the comparative example of this invention. The production line with a daily processing capacity of 600 tons of wheat flour is selected from Suiping Kemin Flour Industry Co., Ltd., a subsidiary of Kemin Flour Industry.

[0040] The basic composition and particle size results of two representative grinding processes (bread mill 3B and heart mill 8M) in the wheat flour preparation process of Examples 1 and 2, and Comparative Examples 1 and 2 were measured and analyzed according to the above method; at the same time, other key indicators were compared, including: particle size detection speed, component detection speed, ash control accuracy, response delay, special grade flour yield, and power consumption per ton of flour.

[0041] Table 1. Results of determination of basic composition of wheat flour obtained from the 3B and 8M grinding processes of the hull mill and the core mill. Table 2. Results of particle size determination of wheat flour obtained from the 3B and 8M grinding processes of the hull mill and the core mill. Table 3 Comparison of the technical advantages of other indicators (1) As shown in Table 1, the upgraded wheat milling system of this invention exhibits excellent performance in processing both hard and soft wheat. Compared to existing flour processing production lines and other existing invention patents, the wheat flour (8M) prepared by the method in Examples 1-2 of this invention has lower moisture content, higher protein content, and lower bran and ash content, indicating that the flour processed by the method of this invention is purer and has higher nutritional value. In addition, the moisture content of the bran-milled and endosperm-milled flours prepared by the method of this invention is lower than that of Comparative Examples 1-2, indicating that this system has more sensitive control over wheat moistening and saves more water. Furthermore, flour with lower moisture content is more conducive to storage and extends the shelf life of the product.

[0042] (2) As shown in Table 2, the flour produced in Examples 1-2 is finer than that in Comparative Examples 1-2 because it uses an AI-driven precision wheat milling system. The particle size distribution is significantly shifted to the left, and D10, D50 and D90 are significantly reduced. This shows that the system of the present invention can more accurately control the entire process of wheat milling, which is conducive to more precise processing and preparation of flour.

[0043] (3) As shown in Table 3, compared with Comparative Examples 1-2, Examples 1-2 use high-speed detection equipment such as high-speed cameras and near-infrared multi-point online analyzers, resulting in shorter detection time for wheat flour particle size and composition, reaching the millisecond level. In addition, due to the precise control of the entire wheat milling process through AI deep algorithms, Examples 1-2 have higher ash content control accuracy and shorter response delay time. Based on the AI-driven wheat precision milling system, the resulting premium flour has a better yield and is more energy-efficient and environmentally friendly.

[0044] As can be seen from the examples and comparative examples, the method of the present invention can achieve precise wheat milling, real-time monitoring of the milling effect, real-time detection of wheat flour composition, and dynamic coordinated adjustment of process parameters, which significantly improves the quality and yield of wheat flour, saves water and electricity, and is more scientific, intelligent, green and environmentally friendly, with a very broad prospect for industry application.

[0045] Optionally, embodiments of this application also provide an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the various processes of the above-described embodiment of a wheat milling method based on multimodal perception and dynamic collaborative decision-making, and can achieve the same technical effect. To avoid repetition, they will not be described again here.

[0046] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described embodiment of a wheat milling method based on multimodal perception and dynamic collaborative decision-making, and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0047] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0048] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0049] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0050] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A wheat milling method based on multimodal sensing and dynamic collaborative decision-making, characterized in that, Includes the following steps: At the discharge point of the mill, images of wheat fragments are captured in real time by a high-speed camera to obtain image data, and the quality index data of wheat flour from multiple grinding processes are simultaneously and rapidly measured by a near-infrared multi-point online analyzer equipped with integrated sensors. A hybrid neural network consisting of a convolutional neural network and a long short-term memory network is used to process the image data and the quality index data. The convolutional neural network is used to extract the spatial features of the visual morphology of wheat flour from the image data, and the long short-term memory network is used to capture the dynamic changes in the particle size distribution of wheat flour and the long-term dependence of quality indicators, thereby generating real-time quality prediction values, quality change trend warning signals and morphology-quality correlation coefficients as the basis for quantitative decision-making. Using the Markov decision process framework, the wheat milling process is modeled as a decision problem involving changes in the state of wheat flour and dynamic process parameters. The real-time quality prediction value, the quality change trend warning signal, and the morphology-quality correlation coefficient are used as part of the current state vector. A reward function is constructed with the quality compliance rate as the core. The optimal strategy is solved by reinforcement learning algorithm to generate process control instructions for real-time adjustment of continuous actions, discrete actions, and constraints in the wheat milling process.

2. The method according to claim 1, characterized in that, The steps for acquiring image data by capturing real-time images of wheat fragments using a high-speed camera specifically include: A high-speed camera consisting of a pinhole sensor, a lens, and a computer is placed under the rollers of each grinding process. The pinhole sensor is selected with an IP66 or higher protection rating. The high-speed camera has a resolution of 2560*1920 or higher, an acquisition speed of more than 1000fps, and a minimum exposure time of less than 100ns, in order to capture fragment images in real time and dynamically track changes in the particle size of the fragments.

3. The method according to claim 2, characterized in that, After acquiring image data, an improved YOLOv11 model is used to perform real-time segmentation of wheat fragment images. The improvements made to the standard YOLOv11 architecture include: After the 3rd and 4th CSPBlock in the backbone network, cross-stage local context modules are embedded respectively. Through the dual-path design of parallel dilated convolution and standard convolution, the model's ability to perceive the global context of the adhering particles and capture the local details of edge cracks is enhanced. An adaptive weighted feature pyramid is introduced at the neck of the feature fusion process, and a collaborative attention mechanism is used to dynamically optimize the feature fusion process. Its mathematical expression is as follows: F fused = Conv(CA(F high ) ⊙ F high + SA(F low ) ⊙ F low ); Where F fused This represents the fused feature map, CA() and SA() represent the channel attention and spatial attention functions respectively, ⊙ represents element-wise multiplication, and F high and F low These represent feature maps from high-level semantic features and low-level detail features, respectively, and Conv represents the convolutional fusion operation; The detection head is extended into a decoupled instance segmentation head. While retaining the original classification and regression branches, a mask prediction branch is added in parallel. RoI Align is used to extract target features from high-resolution feature maps and a pixel-level mask is output through a lightweight decoder. During training, the segmentation accuracy is optimized by combining segmentation loss.

4. The method according to claim 1, characterized in that, The steps for obtaining quality index data by simultaneously and rapidly measuring wheat flour through multiple milling processes using a near-infrared multi-point online analyzer equipped with integrated sensors include: The near-infrared multi-point online analyzer simultaneously tests four or more measurement points. The integrated sensor is arranged below the rollers in each grinding process to quickly determine color, ash content, protein content, moisture content, bran content, and temperature. The integrated sensor has an IP66 or higher protection rating, and the analyzer has a wavelength range of 780-2500nm and a resolution of 4cm or greater. -1 The scanning speed is greater than 1 scan per 20ms, and it is equipped with a high-purity nitrogen or dry air purging system.

5. The method according to claim 1, characterized in that, The steps of processing the image data and the quality index data using a hybrid neural network composed of a convolutional neural network and a long short-term memory network specifically include: The obtained real-time wheat fragment segmentation mask image sequence and its derived morphological vectors are used as multi-channel inputs; The convolutional neural network adopts an encoder structure with four convolutional layers. Each layer contains 3×3 convolution, batch normalization and ReLU activation, and downsampling is performed through 2×2 max pooling to encode each frame of image into a 256-dimensional feature vector. The Long Short-Term Memory (LSTM) network receives the feature vector sequence arranged in time steps. Its core consists of two layers of bidirectional LSM network units, with 128 hidden units in each layer. Through the gating mechanism of its input gate, forget gate, and output gate, it captures long-term dependencies. The output of the final time step of the Long Short-Term Memory network is fed into a fully connected prediction head containing two branches. One branch outputs the real-time quality prediction value and the trend warning signal through softmax, while the other branch outputs the morphology-quality correlation coefficient through a linear layer.

6. The method according to claim 1, characterized in that, Using the Markov decision process framework, the wheat milling process is modeled as a decision-making problem involving changes in the state of wheat flour and dynamic process parameters, specifically including: We construct a system that integrates online image segmentation and morphological quantization results, key operating parameters of various grinding and grading equipment, and state vectors based on real-time predicted values ​​and recent historical trends. The definition includes the action space for continuous micrometer-level fine-tuning of the roll gap and discrete selection of the air classifier flow rate, and incorporates built-in equipment safety constraints; A reward function is constructed, which is a weighted score that integrates multiple quality indicators such as ash content control accuracy, powder yield, and energy consumption. Its mathematical expression is as follows: r t =w1 R 品质 +w2 R 效率 +w3 R 能耗 ; Among them, R 品质 R is used to reward high ash content control precision and increased proportion of premium powder. 效率 R is used to reward increased production of qualified powder per unit time. 能耗 Used to punish an increase in energy consumption per unit of output; A deep neural network is trained using a near-end policy optimization algorithm as the policy function πθ(a) t |s t The network is in its current state s t Given the input, output the optimal action a. t The probability distribution is used to generate the process control instructions.

7. The method according to claim 6, characterized in that, The state vector specifically includes the area distribution of fragments of each particle size, average profile complexity, current rolling gap, feed flow rate, motor load, current ash content prediction value, estimated powder yield, and recent historical trends of quality indicators.

8. The method according to claim 6, characterized in that, Each sub-reward in the reward function is normalized, and the weights w1, w2, and w3 are configured according to the production focus. Through training, the agent learns a strategy network that maximizes long-term cumulative discount rewards, which not only focuses on immediate quality but also takes into account the stability and efficiency of the entire production cycle, thereby achieving adaptive and multi-objective dynamic optimization of the powder making process.

9. An electronic device, characterized in that, The method includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of a wheat milling method based on multimodal perception and dynamic collaborative decision-making as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions, which, when executed by a processor, implement the steps of a wheat milling method based on multimodal perception and dynamic collaborative decision-making as described in any one of claims 1-8.

Images