Fresh soybean air intelligent detection and sorting operation system

Abstract

The invention discloses a fresh soybean air intelligent detection and sorting operation system which is characterized in that an image acquisition device and an air blowing execution device are arranged at the conveying tail end of conveying equipment, and a control device is electrically connected with the conveying equipment, the image acquisition device and the air blowing execution device; the image acquisition device comprises a light source, a light source controller and a high-speed linear array industrial camera, the light source and the high-speed linear array industrial camera are arranged at the oblique upper side position and the oblique lower side position of the material throwing channel and are connected with the control device, the control device controls the light source to stroboflash, and the stroboflash time is consistent with the exposure time of the high-speed linear array industrial camera; an intelligent sorting model is arranged in the control device, the soybean quality is classified by adopting an improved YOLOv8 algorithm, and the air blowing execution device executes blowing selection after receiving a classification signal. By optimizing the image processing and parallel computing method and combining the detection result and the high-speed positioning algorithm, the sorting accuracy and efficiency are improved, the manual operation dependence is reduced, and the large-scale stable production requirement can be met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fresh soybean sorting, and more particularly to an intelligent fresh soybean air detection and sorting system. BACKGROUND

[0002] Fresh soybean, also known as green soybean, is an important agricultural product, and its quality directly affects the eating experience of consumers and market value. Traditional fresh soybean sorting methods mainly rely on manual screening, which is not only time-consuming and labor-intensive, but also has low sorting precision and efficiency, making it difficult to meet the demand for high-quality soybeans in large-scale production and the market. With the development of modern agriculture and intelligent technology, automated and intelligent sorting technology has gradually gained attention. Existing automatic sorting systems are mostly applied to regular-shaped, easily transported, and large-grained agricultural products such as citrus, apples, potatoes, and kiwis. There is relatively little research on the sorting technology of fresh soybeans. These systems often use image recognition, sensor detection, and other technical means to analyze the appearance, color, and other characteristics of the products to achieve automatic sorting. However, the sorting of fresh soybeans requires more detailed and comprehensive judgment of multiple parameters such as appearance, color, shape, and size to ensure that the sorted fresh soybeans meet market demand.

[0003] Traditional mechanical sorting equipment mainly relies on the color of fresh soybeans for classification, but these devices cannot effectively identify minor external damage to fresh soybeans. For detecting small defects such as damage, insect eyes, and rust spots, the accuracy and reliability of mechanical sorting systems are poor. At the same time, traditional sorting equipment is also susceptible to external environmental and physical characteristics of soybeans, resulting in low sorting precision.

[0004] Some sorting systems based on spectral technology, such as those based on near-infrared spectroscopy or hyperspectral imaging, can accurately detect soybean maturity, internal defects, and insect infestations, but these technologies have the problem of slow processing speed. In addition, spectral technology may experience response time lag and reduced sorting accuracy when processing high-density, large-scale production lines, making it difficult to achieve real-time and efficient sorting.

[0005] Although some fresh soybean sorting equipment has been optimized using machine vision and artificial intelligence technologies, most existing systems lack adaptive capabilities and cannot automatically adjust sorting parameters based on different batches and sources of soybeans. In particular, when faced with the diversity of soybean varieties, sizes, colors, and appearances, existing systems have poor adaptability and cannot provide high-precision, high-efficiency personalized sorting services.

[0006] In addition, the air blowing valve, as a common rejection execution unit in existing fruit and vegetable sorting equipment, also has some defects on the market. The response time and blowing precision of the air blowing valve are affected by the control system and the air path structure, and are prone to delay or deviation, resulting in that the target fresh soybeans are not accurately rejected or normal fresh soybeans are mistakenly blown, reducing the overall sorting accuracy. The blowing intensity of the air blowing valve is difficult to adaptively adjust for fresh soybeans of different sizes, weights and surface states, and is prone to incomplete sorting or secondary damage to the fresh soybeans.

[0007] In view of the characteristics of fresh soybeans, it is particularly necessary to develop an efficient and accurate intelligent sorting system. The system should have the following characteristics: (1) capable of detecting soybeans in all directions, identifying and rejecting unqualified products; (2) fast sorting speed, suitable for large-scale production needs; (3) high system stability, capable of long-term continuous operation; (4) easy to operate, easy to maintain and manage. SUMMARY

[0008] In view of the technical problems of low sorting accuracy, incomplete sorting and secondary damage in the prior art, the present application provides an intelligent air detection and sorting system for fresh soybeans, which improves the accuracy and efficiency of sorting by optimizing image processing and parallel computing methods and combining detection results with high-speed positioning algorithms, reduces the dependence on manual operation, and can meet the needs of large-scale stable production, promoting the development of the fresh soybean processing industry towards intelligence, automation and high efficiency.

[0009] The technical solutions adopted are as follows:

[0010] An intelligent air detection and sorting system for fresh soybeans, comprising a conveying device, an image acquisition device, an air blowing execution device and a control device, the image acquisition device and the air blowing execution device are arranged at the conveying end of the conveying device, the control device is electrically connected with the conveying device, the image acquisition device and the air blowing execution device, the image acquisition device comprises a light source, a light source controller and a high-speed linear array industrial camera, the light source and the high-speed linear array industrial camera are arranged on the upper side and the lower side of the material throwing channel, and are connected with the control device, the light source controller is connected with the light source, the control device controls the stroboscopic of the light source, and the stroboscopic time is consistent with the exposure time of the high-speed linear array industrial camera; the control device is provided with an intelligent sorting model, which receives the soybean image data transmitted by the image acquisition device in real time, classifies the soybean quality by using an improved YOLOv8 algorithm, and executes the blowing selection action after receiving the classification signal.

[0011] Preferably, an SE attention module is introduced in the improved YOLOv8 algorithm, which is located at the end of the backbone network of the improved YOLOv8 algorithm and after the SPPF module, which performs multi-scale pooling on the feature map to generate a feature representation with multiple scale information, and then adjusts the weight of the global feature through the SE attention module to distinguish the defective soybeans from foreign matters.

[0012] Further, the neck of the improved YOLOv8 algorithm introduces an MSDA attention module, which is arranged at the end of the C2f module in the YOLOv8 algorithm and directly outputs to the detection head, which adaptively allocates weights to different scale features, so that the intelligent sorting model can dynamically adjust the attention degree of different scale features.

[0013] Further, the control device (4) automatically adjusts the blowing timing of the blowing execution device (3) according to the conveying speed v of the conveying device (1) and the blowing speed v air of the blowing execution device (3), and the specific algorithm is as follows:

[0014] t valve-trigger = t current + Δt delay -t blow

[0015] d target = p predicted -p initial

[0016]

[0017] p predicted = p current +v×Δt

[0018] wherein: t blow : the time required for the nozzle in the blowing execution device to release airflow to impact the defective beans;

[0019] Δt delay : the time delay of the control device for image analysis of the defective beans;

[0020] t curent : the current time when the control device analyzes the image of the defective beans;

[0021] d target : the distance from the nozzle in the blowing execution device to the defective beans;

[0022] t valve-trigger : the blowing timing of the blowing execution device;

[0023] p predicted: Predicted defect bean position

[0024] p initial : Initial position of the nozzle of the air blowing execution device

[0025] p current : Position of the defect bean when the control device performs image analysis on the defect bean

[0026] Further, the control device automatically controls the air blowing pressure value and the spray valve response time of the air blowing execution device according to the size of the soybean kernels collected by the high-speed linear array industrial camera.

[0027] Further, the feeding port, the detection area and the sorting port of the conveying device are respectively provided with photoelectric sensors for blockage detection, and each photoelectric sensor is electrically connected with the control device, for controlling the speed of the conveying device.

[0028] Preferably, the conveying device comprises a vibrating feeder and a belt conveyor connected in series, an incremental encoder is installed on the driving roller of the belt conveyor, the incremental encoder is connected with the control device, and the pulse signal of the incremental encoder is input as the "row trigger" signal of the high-speed linear array industrial camera, for synchronous equidistant sampling with the light source and the high-speed linear array industrial camera.

[0029] Further, a negative pressure dust collection device is arranged at the throwing channel at the end of the conveying device, for sucking away dust and pod hair.

[0030] Further, the feeding port, the detection area and the sorting port of the conveying device are respectively provided with photoelectric sensors for blockage detection, and each photoelectric sensor is electrically connected with the control device, for controlling the speed of the conveying device.

[0031] Further, the system further comprises a remote monitoring platform, the control device is connected with the remote monitoring platform through a wireless data transmission module, for uploading detection data to the remote monitoring platform in real time, performing data analysis through a real-time data processing module in the remote monitoring platform, quickly identifying the quality defects of soybeans, and recording and storing the analysis results.

[0032] The technical scheme of the present application has the following advantages:

[0033] A.The system of the present application integrates machine vision, deep learning, image processing, precise control execution and other technologies, obtains image information from the side upper and side lower during the free falling process of fresh soybeans, avoids defects and blind areas caused by a single perspective, synchronizes the strobe time of the light source with the exposure time of the high-speed linear array industrial camera to ensure the consistency of image brightness under short exposure, and uses an improved YOLOv8 algorithm in the intelligent sorting model in the control device to classify soybean quality, improve the generalization ability and classification accuracy of detection.

[0034] B.The present application introduces an SE attention module in the improved YOLOv8 algorithm, effectively improves the attention to key features, reduces the interference of unimportant or redundant features on the detection result, and thus helps to improve the detection accuracy of the intelligent sorting model for external defects and foreign matter of fresh soybeans, adapts to more complex background appearance features, can more clearly distinguish defective soybeans and foreign matter, and the overall detection result is more accurate.

[0035] C.The present application introduces an MSDA attention module in the improved YOLOv8 algorithm, adaptively allocates weights to feature maps of different scales, so that the intelligent sorting model can dynamically adjust the attention to features of different scales, ensures that the model can focus on the most representative features in the current detection task, ignores unimportant or redundant feature information, thereby improves the accuracy and robustness of detection, and is more accurate, which is of great significance for improving the detection accuracy of the model in complex scenes, especially identifying small targets and fuzzy targets.

[0036] D.The present application realizes real-time detection and sorting decision under high-speed conveying conditions through an optimized image processing and parallel computing method, accurately identifies and sorts five kinds of defective pods of fresh soybeans, including broken pods, abnormal maturity pods, worm eye pods, rust spot pods, single kernel fruit pods and various foreign matters, combines the detection result with a high-speed positioning algorithm, controls the air blowing valve to complete accurate rejection in a low response time, reduces the error blowing rate and the missing blowing rate, can greatly improve the accuracy and efficiency of sorting, and reduces the dependence on manual operation.

[0037] E.The system of the present application not only can adjust the sorting parameters in real time, adapt to soybeans of different seasons, different varieties and different production areas, but also can maintain efficient and stable sorting operation in a high-speed production line, meet the demand of large-scale production.

[0038] F.The system of the present application also provides a remote monitoring platform, which can upload operation related data in real time through a wireless data transmission module, and the remote monitoring platform records and traces key operation parameters for a long time to complete traceable management of the process; through accurate automatic sorting and operation control, the system improves the overall efficiency of the production line, reduces production cost, ensures stable and consistent quality of each batch of soybeans, and promotes the development of fresh soybean processing industry towards intelligence, automation and high efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the specific embodiments of the present application, the drawings required in the specific embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0040] Figure 1 is the overall structure view of the system provided by the present application;

[0041] Figure 2 is the control device structure architecture diagram provided by the present application;

[0042] Figure 3 is the improved YOLOv8 network structure diagram provided by the present application;

[0043] Figure 4 is the working flow chart of the fresh soybean sorting system provided by the present application;

[0044] Figure 5 is the terminal control and monitoring interface design diagram provided by the present application.

[0045] The identifier symbols provided in the drawings are explained as follows:

[0046] 1-conveying equipment

[0047] 11-vibrating feeder, 12-belt conveyor

[0048] 2-image acquisition device

[0049] 21-light source, 22-high-speed linear array industrial camera

[0050] 211-upper auxiliary light source, 212-lower auxiliary light source

[0051] 3-air blowing execution device; 4-control device; 5-negative pressure dust collection device; 6-vibration rack

[0052] 7-defective product box; 8-good product box; 9-good product outlet; 10-defective product outlet

[0053] 20-shading plate; 30-touch screen.

[0054] a-throwing channel. DETAILED DESCRIPTION

[0055] The technical solutions of the present application will be described clearly and completely below in combination with the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0056] As Figures 1 to 5 shown, the present application provides a fresh soybean air intelligent detection and sorting operation system, which comprises a conveying device 1, an image acquisition device 2, a gas blowing execution device 3, a control device 4 and a touch screen 30, and the touch screen 30 can realize man-machine interaction. The conveying device 1 further preferably comprises a vibrating feeder 11 and a belt conveyor 12, the vibrating feeder 11 is installed on a vibrating rack 6, and the conveying end thereof is overlapped above the feeding end of the belt conveyor 12, the image acquisition device 2 and the gas blowing execution device 3 are arranged at the conveying end of the conveying device 1, the control device 4 is electrically connected with the conveying device 1, the image acquisition device 2 and the gas blowing execution device 3, respectively, forming a closed-loop operation structure of "single particle feeding-air imaging-real-time identification-gas blowing removal-classification collection". The image acquisition device 2 comprises a light source 21, a light source controller (not shown in the figure) and a high-speed linear array industrial camera 22, the light source 21 and the high-speed linear array industrial camera 22 are arranged at the obliquely upper side and the lower side of the material throwing channel a, and are connected with the control device 4, the light source controller is connected with the light source 21, and is used for controlling the opening and closing of the light source 21, the control device 4 controls the stroboscopic of the light source 21, and the stroboscopic time is consistent with the exposure time of the high-speed linear array industrial camera 22; the control device 4 is provided with an intelligent sorting model, the intelligent sorting model receives the soybean image data transmitted by the image acquisition device 2 in real time, classifies the soybean quality by using an improved YOLOv8 algorithm, and the gas blowing execution device 3 executes the blowing selection action after receiving the classification signal.

[0057] The vibrating frequency and amplitude of the vibrating feeder 11 are adjustable, the vibrating feeder 11 adopts a plurality of guide groove structures, and under the driving of the vibrating motor, the fresh soybeans are dispersed and enter the detection area in a stable posture, a bridge prevention rib and a material level sensor are arranged in the hopper of the vibrating feeder 11 to ensure continuous feeding.

[0058] The conveying belt of the belt conveyor 12 is preferably provided with a food-grade concave-convex anti-skid surface structure, and the belt speed can be continuously adjusted in the range of 0.5-5 m / s. As a further preferred embodiment of the application, an incremental encoder is installed on the driving roller of the belt conveyor 12, and the incremental encoder is connected with the control device 4 and inputs the pulse signal of the incremental encoder as a “row trigger” signal of the high-speed linear array industrial camera 22, so that the equal-distance sampling is synchronously performed with the light source 21 and the high-speed linear array industrial camera 22. The incremental encoder installed at the driving end is used for speed and displacement measurement. An adjustable drop-off edge is arranged at the end of the conveying belt, so that the material enters the air detection and sorting channel in a predictable trajectory, and the good product outlet 9 and the defective product outlet 10 are arranged below the conveying end, and the good product box 8 and the defective product box 7 are correspondingly installed, for receiving the classified material.

[0059] A matching fixed-focus lens is arranged on the high-speed linear array industrial camera 22, and an upper auxiliary light source 211 and a lower auxiliary light source 212 are arranged on both sides of the throwing channel. The high-speed linear array industrial camera 22 is preferably a 2K color three-line linear array industrial camera with a GigE interface, an image element of 7 μm, a maximum line frequency of 80 kHz, support for RGB true color and multiple trigger (line trigger / frame trigger / line+frame) modes, and is arranged at an oblique upper side position and a lower side position to obtain multi-angle information. The lens can be equipped with a polarization and short-pass filter to suppress reflection. The light source 21 is preferably a high-brightness strip LED and supports stroboscopic, and is synchronized with the camera exposure hardware. The detection area is provided with a light shield and a dustproof observation window to ensure the contrast and stability of imaging.

[0060] In order to better shield external stray light and suppress reflected glare, the application is provided with a negative pressure dust collection device 5 at the throwing channel a of the conveying equipment 1, so as to form a micro-negative pressure in the imaging area to suck away dust and hair, and reduce the interference of floating objects on recognition and blowing trajectory.

[0061] The air blowing execution device 3 used in the application mainly comprises a high-speed air blowing valve array, a gas collecting manifold, a pressure stabilizing and filtering unit, a nozzle assembly and an adjustable mounting beam. The valve array is arranged according to the channel spacing, supports millisecond-level response and independent addressing, and the working gas source is 0.4-0.8 MPa. After three-stage filtering and pressure stabilization, the working gas enters the gas collecting manifold. The transverse and longitudinal distances of the nozzle to the throwing trajectory can be adjusted to obtain the best jet flow coverage and energy utilization. The valve array is powered by shunt power supply on the logic side and the coil side, and has overcurrent, short circuit and temperature rise protection.

[0062] The application is also provided with photoelectric sensors (not shown in the figure) for blockage detection at the feeding port, the detection area and the sorting port of the conveying equipment 1. Each photoelectric sensor is electrically connected with the control device 4, and is used for controlling the speed of the conveying equipment 1 to prevent blockage caused by too much material.

[0063] In addition, the control device in the present application mainly includes an integrated industrial computer (containing GPU / AI acceleration), a programmable logic controller (PLC), an image acquisition card, an isolated DI, a valve drive board, a light source controller and a power module in an electrical cabinet. The industrial computer completes image processing, deep learning inference, trajectory prediction and valve scheduling, the PLC is responsible for high-speed I / O and safety interlocking, both of which communicate through RS-485 and take the encoder as a unified time base to realize phase alignment of the camera, the light source and the valve. The power supply adopts 220VAC input, which is converted into 24VDC control power and 24VDC execution power in the cabinet, and is configured with emergency stop and access control interlocking, surge, overvoltage and leakage protection.

[0064] The present application sets an intelligent sorting model in the control device, adopts a deep learning algorithm, receives image data transmitted by the image acquisition device 2 in real time, and classifies the quality of soybeans by using an improved YOLOv8 algorithm. In order to improve the accuracy and recognition efficiency of the image acquisition device in soybean sorting, the YOLOv8 algorithm is improved, and SE (Squeeze-and-Excitation) attention module and MSDA (Multi-Scale Dilated Attention) attention module are added. The introduction of these two kinds of attention mechanisms enables the model to focus better on key features, thereby improving detection accuracy in complex backgrounds and diversified target morphologies. The improved overall network structure diagram is as shown in Figure 3 The SE attention module is a channel attention mechanism that adaptively adjusts the weight of each channel in the feature map, so that the intelligent sorting model can pay more attention to the channel features that are more meaningful for classification or detection. Its main process is as follows: a global average pooling operation is performed on each channel of the feature map, the feature information of each channel is compressed into a global feature value, and a description vector with the same number of channels is generated. This step aggregates spatial information into global channel information, which is convenient for subsequent weighting. The compressed description vector is processed through a series of fully connected layers and activation functions to generate channel weight coefficients, which represent the importance of each channel in the current task. Finally, the channel weight coefficients are used to weight the channels in the original feature map, enhancing important features and suppressing unimportant features, so that the model pays more attention to channel features that are useful for detection. The introduction of the SE attention module enables YOLOv8 to achieve adaptive feature selection on different channels, focusing more on important features in soybean images such as defects or discoloration areas, improving detection accuracy and robustness.

[0065] The design and introduction of the SE attention module follows the following three steps:

[0066] The first step, Squeeze, compresses the spatial information of each channel into a scalar through global average pooling, capturing the global information of each channel. This step reduces the spatial dimension while preserving the global features between channels.

[0067] Let the input feature map be: where H represents the height of the feature map, W represents the width of the feature map, and C represents the number of channels of the feature map.

[0068] The formula is as follows:

[0069]

[0070] where Xc represents the feature map of the Cth channel, and H and W are the height and width of the feature map.

[0071] The second step, Excitation, uses two fully connected layers to perform nonlinear transformation on the squeezed features to generate weights for each channel. The channel description vector is passed through a two-layer fully connected neural network, first compressed in dimension through the ReLU activation function, and then recalibrated through the Sigmoid activation function to generate the weight coefficient of each channel. The formula is:

[0072] s = σ(W2·ReLU(W1·Z))

[0073] where W1 and W2 are learnable weight matrices, and σ is the Sigmoid function.

[0074] The third step, Recalibration, adjusts each channel of the input feature map by weighting to obtain new features. The s value activated by the Sigmoid function is used to recalibrate the original input Xc. The purpose of this step is to adjust the feature response of each channel to highlight important features and suppress less important information. Recalibration is achieved through element multiplication, and the specific formula is:

[0075] x c ′ = s c · x c

[0076] where s c is the scale factor activated by the Sigmoid function, and x c ′ is the calibrated feature channel. This step helps the model more effectively utilize important features during subsequent processing, improving the model's expression ability and performance. By normalizing these weights through the activation function and redistributing them to each channel, the feature expression of important channels is dynamically enhanced.

[0077] In the present application, the SE attention module is inserted at the end of the backbone network of YOLOv8, immediately after the SPPF (Spatial Pyramid Pooling-Fast) module. The choice of this position aims to ensure that the last stage of the feature extraction process can make full use of the SE attention module to make sufficient weight adjustment on the global features. The SPPF module generates feature representations with multiple scale information by performing multi-scale pooling on the feature maps, while the introduction of the SE attention module further enhances the model's ability to adaptively adjust the importance of different channels. By compressing and exciting the global features of each channel, the SE attention module effectively enhances the focus on key features and reduces the interference of unimportant or redundant features on the detection results. This design not only helps to enhance the model's detection accuracy of fresh soybean external defects and foreign objects, especially in complex backgrounds or when faced with similar appearance features, it can more clearly distinguish between defective soybeans and foreign objects. In addition, due to the lightweight nature of the SE attention module, it hardly introduces additional computational overhead, yet it can significantly improve the model's performance in target detection, especially when dealing with small targets and complex defects, the SE attention module helps to improve the model's sensitivity to details and generalization ability. By adding the SE attention module at the end of the backbone network, the model can further strengthen the important channel information before performing the final target classification and bounding box prediction, making the overall detection result more accurate.

[0078] In the action system provided by the application, in order to further improve the performance of YOLOv8 in the task of detecting external defects and foreign matters of fresh soybeans, especially the performance in processing complex background and multi-scale targets, a multi-scale dilated attention (MSDA, Multi-Scale Dilated Attention) mechanism is introduced into the model. The main reasons and motivations for adding the MSDA attention module are as follows: the external defects and foreign matters of fresh soybeans have diversified feature representations, including defect regions of different sizes, foreign matters of different shapes, and complex and variable background environments. In traditional convolutional neural networks, single-scale feature extraction often fails to capture these diverse features simultaneously, resulting in suboptimal performance of the model when facing targets of different scales. The MSDA attention module can extract features at different scales through a multi-scale feature extraction strategy, allowing the model to focus on both global information at large scales and local details at small scales, thereby improving the detection capability of diverse targets. The surface defects and foreign matters of fresh soybeans are usually distributed in different directions of the background, making it difficult to extract target features. Traditional convolution operations are limited by fixed receptive fields, making it difficult to obtain sufficient context information without increasing computational complexity. By introducing dilated convolution, the MSDA attention module can effectively expand the receptive field, allowing the model to capture more extensive context information while maintaining the original resolution, thereby enhancing the model's ability to perceive long-distance dependent features. This is of great significance for improving the detection accuracy of the model in complex scenes, especially for identifying small targets and blurred targets. In addition, the features of the surface defects and foreign matters of fresh soybeans may vary greatly in the image, with some features being significant at certain scales and not obvious at other scales. The MSDA attention module introduces an attention mechanism to adaptively assign weights to feature maps of different scales, allowing the model to dynamically adjust the attention to different scale features. This mechanism ensures that the model can focus on the most representative features in the current detection task and ignore unimportant or redundant feature information, thereby improving the accuracy and robustness of the detection.

[0079] The MSDA attention module is a multi-scale dilated convolution attention mechanism designed to enhance the model's ability to capture features at different scales. In soybean image sorting, soybeans vary in size, defects, and morphology, making multi-scale feature extraction crucial. The MSDA attention module works as follows: The MSDA attention module extracts multi-scale features through convolution operations with different dilation rates. Dilated convolution expands the receptive field by introducing holes between convolution kernels, allowing it to capture a larger range of contextual information without increasing computational complexity. Different dilation rates are used to obtain features at different scales, enabling the model to recognize targets of varying sizes and morphologies. After obtaining multi-scale features, the MSDA attention module weights and aggregates features at different scales to highlight features that are beneficial for detection. This attention focusing process is similar to adaptively selecting the optimal scale of features for comprehensive processing, allowing the model to accurately recognize soybean features in complex backgrounds and at different scales. The MSDA attention module ultimately fuses different scale features together to form attention feature maps with multi-scale information. These fused feature maps will be input into the subsequent YOLOv8 detection module to further improve the model's detection capabilities. By introducing the MSDA attention module, YOLOv8 can effectively adapt to the various scale changes that soybeans may exhibit in images, enabling the system to accurately identify targets and improve the detection's generalization ability when dealing with complex and diverse image backgrounds. When the system detects unqualified soybeans, the air blowing execution device will automatically start, using the impact force of the air flow to accurately blow the unqualified products into the defective product outlet. The air blowing execution device receives signals from the intelligent sorting module and makes corresponding operations based on the real-time state of each soybean to ensure sorting accuracy. For qualified soybeans, they continue to pass through the conveyor belt into the good product outlet, realizing automatic and interference-free diversion operation. This module achieves efficient diversion of soybeans without increasing mechanical friction, greatly improving the durability and stability of the sorting system.

[0080] The application realizes intelligent sorting of fresh soybeans through image acquisition and analysis, defect identification, and coordination between sorting operations. There is a time difference between image acquisition and analysis of fresh soybean pictures and blowing sorting of defective beans. The image acquisition unit performs high-frequency image acquisition of fresh soybeans through a high-definition camera. The acquired image information is transmitted to the central processing unit in the control device for image processing, feature extraction, and defect identification. Through deep learning algorithms, the appearance features of each soybean are analyzed in real time, and defective beans are identified, including features such as damage, foreign matter, insect damage, single-pod, abnormal maturity, etc. Then, the image analysis results are synchronized with the movement of the soybeans. During the flow of the soybeans, a speed sensor is used to detect the movement speed of the sorting operation system in real time, and the actual flow distance of the soybeans is calculated to accurately calculate the position of the soybeans and the corresponding defect analysis data, ensuring that the position information of each soybean accurately corresponds to its defect state. There is a spatial interval between the air blowing execution device and the image acquisition device, so it is necessary to accurately control the blowing operation based on the real-time processing results and the current position of the soybeans. When the system calculates the position of the defective beans, the air blowing execution device accurately controls the release of the air flow through the electromagnetic valve, and blows and separates the defective beans in time when they pass; when normal beans pass, the system stops the air blowing operation to ensure that only defective beans are blown out. The working flowchart of the fresh soybean sorting operation system is shown in Figure 4 .

[0081] (1) Real-time calculation of fresh soybean position

[0082] Due to the time delay in the image acquisition and analysis process, the control device needs to use the speed information of the soybean flow to predict the position of the defective beans. Through the speed sensor installed on the system (not shown in the figure), the speed of the soybean flow (such as meters per second) is measured in real time, and combined with the current position of the soybeans, the predicted position of the defective beans is calculated. Assuming that the flow rate of each soybean is v, the image analysis position of the defective beans is p current , and the current time is t curent , the predicted position of the defective beans is:

[0083] p predicted＝ p current +v×Δt

[0084] Where Δt is the time delay after image analysis, and v is the flow rate of fresh soybeans.

[0085] (2) Accurate positioning of blowing time

[0086] The air blowing execution device needs to accurately release air flow when the defective beans reach the predetermined position. Considering the speed of the soybean flow and the response time of the air blowing execution device, the system accurately controls the blowing time through the following steps:

[0087] a. Calculation of nozzle position

[0088] The jet position of the blowing nozzle needs to be calculated according to the current position, speed and predicted delay of the soybean. The initial position of the blowing nozzle is defined as p initial , the blowing nozzle should release the air flow when the defective soybean arrives. The distance d target from the blowing nozzle to the defective soybean is calculated as follows:

[0089] d target = p predicted -p initial

[0090] According to the flow speed of the soybean and the blowing response time, the triggering time of the blowing nozzle is calculated.

[0091] b. Blowing time delay

[0092] In order to ensure that the air flow can accurately blow the defective soybean, the control system needs to calculate the time t blow required from the release of the air flow from the blowing nozzle to the impact of the defective soybean by the air flow. air This time depends on the distance from the blowing nozzle to the pod and the speed v air of the air flow. The control device needs to start the air flow release operation in advance before the defective soybean passes through the blowing nozzle. That is, the blowing nozzle should trigger the air jet t blow in advance. The calculation formula of the time is as follows:

[0093]

[0094] (3) Control of the response of the electromagnetic valve

[0095] The opening and closing of the electromagnetic valve is the key to the regulation of the air flow. The response time of the electromagnetic valve and the blowing time delay need to be optimized synchronously. The electromagnetic valve needs to be accurately controlled according to the time point calculated by the control device to ensure that the air flow is released at the appropriate time. The opening time of the electromagnetic valve should be set according to the predicted position of the defective soybean, the relative position of the blowing nozzle and the defective soybean, and the propagation time of the air flow. The calculation formula is as follows:

[0096] t valve-trigger = t current + Δt delay -t blow

[0097] Where Δt delay is the time delay of image analysis and processing, t blow is the time of air flow propagation, and t valve-trigger is the triggering time of the electromagnetic valve.

[0098] Through the prediction of the position of the defective soybean, the regulation of the air flow intensity, the calculation of the blowing time and the accurate operation of the electromagnetic valve, the present application can accurately identify the defective soybean and timely perform the blowing operation during the flow of the soybean.

[0099] (4) Air-blow solenoid valve driving module design

[0100] The system contains 16 solenoid valves, each of which is precisely controlled through an independent control channel. The control card is responsible for receiving signals from the outside and transmitting them to the solenoid valve driving module, ensuring the speed and stability of signal transmission to meet the demand for fast response. To improve the visual control effect of the system, an indicator light panel is designed to provide real-time feedback on the working status of the solenoid valves, such as eye and kernel defects. In terms of solenoid valve driving module, a single-reversing pneumatic solenoid valve with fast response speed, low power consumption, and durability is chosen. Each solenoid valve is controlled by a current signal to ensure precise airflow regulation. The driving circuit uses a power amplifier circuit to ensure that the solenoid valve receives sufficient driving current. The driving chip uses TB6612 to carry the current load required by the solenoid valve. The power management design uses a stabilized power supply to ensure the stable operation of the driving module. In terms of control card, the main control chip uses an STM32 single-chip microcomputer, which has multiple IO ports to support the independent control of 16 solenoid valves and communicates with the upper computer and sensor system through serial ports and Modbus protocols. The switching timing of each solenoid valve is controlled by a PWM signal or a digital control signal. The control card also includes an input signal interface for receiving external signals such as airflow detection and sorting requirements, and is equipped with a power supply and protection circuit to ensure stable operation within the voltage and current fluctuation range. In the design of the indicator light panel, each solenoid valve corresponds to an indicator light, and there are a total of 16 indicator lights to display the working status of the solenoid valves. Green indicates that the solenoid valve is normally open, and red indicates that the solenoid valve is closed or has failed. The status of the indicator lights is fed back in real time by the control card through digital signals, improving the operability and diagnostic efficiency of the system. In terms of fine airflow regulation, the air valve device uses an adjustable airflow pressure system. To establish the relationship between driving frequency and air flow, the device obtains the following fitting curve:

[0101]

[0102] where f is the driving frequency in HZ, and p is the air pressure in MPa. During system integration and debugging, all hardware (including the driving module, control card, solenoid valve, and indicator light panel) is integrated into a case to ensure the stability of electrical connections. During the debugging phase, the start and stop states of each solenoid valve are tested to ensure the synchronization of control signals and solenoid valve actions, and the control program is optimized to meet the accuracy requirements of sorting operations. In addition, the system undergoes environmental testing, functional testing, and long-term operation testing to verify the stability and reliability of the equipment under different working environments.

[0103] In addition, the system of the present application is also provided with a remote monitoring platform, which has data storage and remote monitoring functions, ensuring the safe, reliable storage and real-time monitoring of data in the soybean sorting process. The system adopts a distributed data storage architecture, combines real-time data acquisition and processing, supports efficient storage, fast retrieval and accurate analysis of big data, and provides remote monitoring and control functions, facilitating operators to monitor and adjust the running state of the system in real time. Based on the distributed storage architecture, the present application adopts a combination of solid state drives (SSD) and high-performance disk array systems (RAID) to provide high bandwidth and large capacity storage capabilities, supporting fast access and efficient management of data. The system of the present application ensures the safety and integrity of data by adopting a multi-level backup strategy. The terminal control and monitoring interface design is shown in Figure 5

[0104] The specific storage scheme is that the main storage unit adopts NVMe SSD supporting high-speed reading and writing, with a data reading speed of up to 3GB / s, which is used to store real-time sorting data, image data and sensor acquisition data. The backup storage unit is set in a separate disk array RAID 1 or RAID 5 configuration, providing data redundancy backup to ensure data recoverability in any hardware failure. The remote monitoring platform supports real-time data uploading and remote storage, ensuring long-term storage and global access of big data. All stored data are analyzed by the real-time data processing module, which is based on an efficient distributed computing framework and can process and analyze large-scale data streams from the image acquisition system, fluorescence detection system and sensor network, and generate sorting reports in real time. The system of the present application supports real-time data stream analysis, quickly identifies the quality defects of soybeans, and records and stores the analysis results. The historical data query and report generation data storage module supports historical data query and statistical analysis, and users can select time interval, sorting conditions and other parameters through a graphical interface to quickly query the detection results of soybeans of a specific batch. The system automatically generates analysis reports, including quality evaluation, sorting efficiency, abnormal detection rate and other indicators of soybeans, and supports export to PDF, Excel and other formats, facilitating production managers to further control and analyze quality. The remote monitoring platform supports real-time monitoring through PC and mobile terminals. The remote monitoring platform is linked with the data acquisition and processing system, which can display the running state, working parameters such as valve control state, image acquisition frequency, fluorescence detection results and system performance, sorting efficiency, processing capacity and fault alarm of each sorting module in real time. Operators can view the detection data of each soybean in the sorting process in real time, including appearance image, fluorescence spectrum and defect identification information. By setting up a cloud platform, system data can be accessed remotely anywhere for remote diagnosis and adjustment. Through the remote monitoring platform, users can dynamically adjust the equipment settings according to production needs to ensure sorting accuracy and efficiency.

[0105] ​The system of the present application has an intelligent alarm and early warning mechanism, which can automatically identify abnormal conditions (such as airflow abnormalities, image acquisition errors, fluorescence detection failures, etc.) based on real-time data and issue an alarm. Alarm information will be transmitted to the operator in real time through SMS, email or push notification, ensuring timely response and problem solving. The alarm system supports multi-level alarm and response strategies, automatically taking appropriate measures according to the severity of the problem to avoid the impact of faults on the sorting process. The remote monitoring platform not only provides real-time running data, but also has fault diagnosis function. The system of the present application can identify potential fault risks of the equipment through self-learning algorithm and historical data analysis, and automatically diagnose. The operator can view fault logs and equipment status through the remote platform to troubleshoot and handle problems. When hardware failure occurs, the platform can guide maintenance personnel to perform remote repair or dispatch on-site maintenance services. All fault and maintenance records will be archived in detail for subsequent analysis and management.

[0106] The system of the present application has data storage and big data analysis capabilities, which can predict production trends and optimize production scheduling through long-term data accumulation and analysis. For example, the platform can analyze production bottlenecks based on sorting efficiency, failure rate and other indicators, and predict potential future problems to take preventive measures in advance to avoid production interruptions or quality fluctuations. The data storage and backup mechanism of the system of the present application ensures data security and high availability, preventing data loss and production interruptions caused by system failures.

[0107] The unmentioned parts of the present application are applicable to the prior art.

[0108] Obviously, the above embodiments are only examples for clarity and do not limit the embodiments. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, all embodiments need not and cannot be exhausted. The obvious changes or variations derived therefrom are still within the protection scope of the present application.

Claims

1. An intelligent detecting and sorting system for fresh soybean in air, comprising a conveying device (1), an image acquisition device (2), a gas blowing execution device (3) and a control device (4), the image acquisition device (2) and the gas blowing execution device (3) are arranged at the conveying end of the conveying device (1), the control device (4) is electrically connected with the conveying device (1), the image acquisition device (2) and the gas blowing execution device (3) respectively, characterized in that, The image acquisition device (2) comprises a light source (21), a light source controller and a high-speed linear array industrial camera (22), the light source (21) and the high-speed linear array industrial camera (22) are arranged at the obliquely upper side and the lower side of the material throwing channel (a) and are connected with the control device (4), the light source controller is connected with the light source (21), the control device (4) controls the stroboscopic of the light source (21), and the stroboscopic time is consistent with the exposure time of the high-speed linear array industrial camera (22); the control device (4) is provided with an intelligent sorting model, which receives soybean image data transmitted by the image acquisition device (2) in real time, classifies soybean quality by using an improved YOLOv8 algorithm, and executes the blowing action after receiving the classification signal.

2. The fresh market soybean aerial intelligent detection and sorting system of claim 1, wherein, The SE attention module is introduced in the improved YOLOv8 algorithm, the SE attention module is located at the end of the backbone network of the improved YOLOv8 algorithm and after the SPPF module, the SPPF module performs multi-scale pooling on the feature map to generate a feature representation with multiple scale information, and then the SE attention module adjusts the weight of the global feature to distinguish the defective soybeans and foreign matters.

3. The fresh market soybean aerial intelligent detection and sorting system of claim 2, wherein, The MSDA attention module is introduced in the neck of the improved YOLOv8 algorithm, the MSDA attention module is arranged at the end of the C2f module in the YOLOv8 algorithm and is directly output to the detection head, which adaptively allocates weights to features of different scales, so that the intelligent sorting model can dynamically adjust the attention degree of features of different scales.

4. The fresh market soybean aerial intelligent detection and sorting system of claim 1, wherein, The control device (4) automatically adjusts the blowing timing of the blowing execution device (3) according to the conveying speed v of the conveying device (1) and the blowing speed v of the blowing execution device (3) air , and the specific algorithm is as follows: t valve-trigger = t current + Δt delay - t blow d target = p predicted - p initial p predicted = p current + v x Δt Wherein: t blow : Time required for the release of the air flow from the nozzle in the air blowing execution device to the air flow impact defect beans; Δt delay : time delay of the control device for image analysis of defective beans; t curent : current time when the control device performs image analysis on the defective beans; d target : distance from the nozzle in the air blowing execution device to the defective beans; t valve-trigger : blow selection timing of the air blow executing device; p predicted : predicted defective bean position; p initial : initial position of the nozzle of the blow executing device; p current : Position of the control device when performing image analysis of the defective beans.

5. The fresh market soybean aerial intelligent detection and sorting system of claim 4, wherein, The control device (4) automatically controls the air blowing pressure value and the spray valve response time of the air blowing execution device (3) according to the size of the soybean kernels collected by the high-speed linear array industrial camera (22).

6. The fresh market soybean aerial intelligent detection and sorting operation system of claim 1, wherein, The conveying equipment (1) comprises a vibrating feeder (11) and a belt conveyor (12) connected in series, an incremental encoder is installed on a driving roller of the belt conveyor (12), the incremental encoder is connected with the control device (4) and inputs the pulse signal of the incremental encoder as a "row trigger" signal of the high-speed linear array industrial camera (22), and the light source (21) and the high-speed linear array industrial camera (22) are synchronously sampled at equal distances.

7. The fresh market soybean aerial intelligence detection and sorting system of any of claims 1-6, wherein, A negative pressure dust collection device (5) is arranged at the throwing channel (a) at the end of the conveying equipment (1) and is used for sucking away dust and pod hair.

8. The fresh food soybean aerial intelligent detection and sorting operation system according to claim 7, characterized in that, Photoelectric sensors for blocking detection are arranged at the feeding port, the detection area and the sorting port of the conveying equipment (1) respectively, each photoelectric sensor is electrically connected with the control device (4) and is used for controlling the speed of the conveying equipment (1).

9. The fresh market soybean aerial intelligent detection and sorting operation system of claim 1, wherein, The system further comprises a remote monitoring platform, the control device is connected with the remote monitoring platform through a wireless data transmission module, is used for uploading detection data to the remote monitoring platform in real time, performs data analysis through a real-time data processing module in the remote monitoring platform, quickly identifies the quality defects of soybeans, and records and stores the analysis results.

Images