Automatic Identification Method and System for Closed Harvesting in Corridors

Abstract

The present invention provides a corridor closure harvesting automatic identification method and system, which relates to the field of harvesting technology, including acquiring environmental parameters in a corridor closure area, calculating the growth status of plants in the corridor, and obtaining plant growth data; acquiring images of plants in the corridor in real time and performing segmentation processing to obtain plant area images, identifying the maturity of plants, and obtaining plant maturity data; judging whether the plants in the corridor meet the harvesting conditions according to the plant growth data and the plant maturity data, and if the plants meet the harvesting conditions, controlling a harvesting execution mechanism to perform a harvesting operation on the plants in the corridor, and acquiring harvesting volume data, and calculating a commodity rate and an inventory based on the harvesting volume data; adjusting environmental parameters in a plant growth model, and repeating the steps of calculating the commodity rate and the inventory according to the adjusted environmental parameters until the commodity rate reaches a preset commodity rate threshold and the inventory is lower than the preset inventory threshold, thereby completing the corridor closure harvesting automatic identification.

Description

Technical Field

[0001] The invention relates to harvesting technology, and in particular to an automatic identification method and system for corridor closure harvesting. Background Art

[0002] With the rapid development of artificial intelligence and automation technology, it has become a trend to apply automation systems and intelligent control technologies in agricultural production to improve production efficiency and product quality. At present, in facility agricultural environments such as greenhouses and artificial climate chambers, the manual harvesting of plants has the disadvantages of low efficiency and high labor intensity. At the same time, due to the changes in environmental parameters and the uncertainty of plant growth status, it is difficult to achieve the optimal effect through a fixed harvesting method. Therefore, there is an urgent need for an intelligent harvesting system that can automatically identify the growth status of plants and adjust the harvesting strategy according to actual conditions.

[0003] In the existing technology, some harvesting systems can only perform harvesting scheduling based on simple time thresholds or growth cycles, and lack accurate recognition and evaluation of the actual plant growth status. Although other systems try to use computer vision and other technologies to identify plant status, the recognition accuracy and robustness need to be improved. In addition, existing systems are usually unable to automatically adjust environmental parameters according to harvesting results to optimize plant growth conditions, resulting in the inability to effectively control the harvested commodity rate and inventory.

[0004] Therefore, the present invention proposes an intelligent corridor closed harvesting automatic identification method based on environmental parameters, plant growth status and maturity. The method comprehensively utilizes multiple technologies such as environmental monitoring, computer vision and artificial intelligence, and can accurately identify the growth status and maturity of plants, thereby judging the harvesting time and performing the harvesting operation. At the same time, the commodity rate and inventory are automatically evaluated according to the harvesting results. If the preset conditions are not met, the environmental parameters are adjusted to form a closed-loop control to optimize the plant growth condition and harvesting effect. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for automatically identifying closed corridor harvesting, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A method for automatically identifying closed corridor harvesting is provided, comprising:

[0008] Acquire environmental parameters in the closed area of ​​the corridor, the environmental parameters including temperature, humidity and carbon dioxide concentration, calculate the growth status of plants in the corridor according to a preset plant growth model and the environmental parameters, and obtain plant growth data; at the same time, acquire images of plants in the corridor in real time and perform segmentation processing to obtain plant area images, extract color, texture and shape features of the plant area images, input them into a pre-trained state recognition model, identify the maturity of the plants, and obtain plant maturity data;

[0009] According to the plant growth data and plant maturity data, it is judged whether the plants in the corridor have reached the harvesting conditions. If the plants have reached the harvesting conditions, the harvesting execution mechanism is controlled to harvest the plants in the corridor and obtain the harvesting quantity data. Based on the harvesting quantity data, the commodity rate and inventory are calculated;

[0010] The calculated commodity rate and inventory are compared with the preset commodity rate threshold and inventory threshold respectively. If the commodity rate is lower than the preset commodity rate threshold or the inventory is higher than the preset inventory threshold, the environmental parameters in the corridor closed area are adjusted to control the environmental parameters in the corridor closed area to optimize the plant growth conditions; according to the adjusted environmental parameters, the steps of calculating the commodity rate and inventory are repeated until the commodity rate reaches the preset commodity rate threshold and the inventory is lower than the preset inventory threshold, and the automatic identification of corridor closure and harvesting is completed.

[0011] In an optional embodiment,

[0012] The images of plants in the corridor are acquired in real time and segmented to obtain plant area images. The color, texture and shape features of the plant area images are extracted and input into the pre-trained state recognition model to identify the maturity of the plants. The obtained plant maturity data include:

[0013] Capture multi-view images of plants in the corridor, obtain plant images from multiple viewpoints, segment and preprocess them, and obtain preprocessed multi-view plant images;

[0014] Extract feature points from the preprocessed multi-view plant images, control the number of extracted feature points by adjusting the parameters of the extraction algorithm, and obtain corresponding feature point pairs between images of different viewpoints based on the extracted feature points;

[0015] Using the corresponding feature point pairs, the 3D structure of the plant is restored through a 3D reconstruction algorithm to obtain multiple 3D plant models, and the corresponding phenotypic parameters are extracted from each 3D plant model. The phenotypic parameters include plant height, canopy volume, leaf area, stem diameter, and number of fruits.

[0016] According to each plant 3D model and the corresponding phenotypic parameters, construct the time series data of plant growth, analyze the time series data of plant growth, fit the plant growth curve, establish a plant maturity model, and use a machine learning algorithm to determine the optimal feature subset related to plant maturity through feature selection based on the phenotypic parameters of the plant 3D model and the plant maturity model, construct a state recognition model, and use the training data to train the state recognition model;

[0017] For the newly acquired plant three-dimensional model, its phenotypic parameters are extracted, and the extracted phenotypic parameters are input into the trained state recognition model to identify the maturity of the plant and obtain the plant maturity data.

[0018] In an optional embodiment,

[0019] For the newly acquired plant three-dimensional model, the phenotypic parameters are extracted, and the extracted phenotypic parameters are input into the trained state recognition model to identify the maturity of the plant. The steps before obtaining the plant maturity data also include:

[0020] The phenotypic parameters of the plant three-dimensional model are normalized, and the normalized phenotypic parameters are resampled by using the farthest point sampling algorithm to obtain a resampled parameter matrix;

[0021] The resampled parameter matrix is ​​input into the state recognition model, which extracts feature maps of different scales and abstraction levels through multiple convolutional layers, downsamples the features through the pooling layer, and finally fuses the features through the fully connected layer to generate the final feature representation;

[0022] Input the obtained feature representation into a classifier, map the feature representation into a probability distribution of plant maturity states, and the dimension of the probability distribution is equal to the number of predefined maturity state categories;

[0023] The convolutional neural network is trained using the labeled plant maturity status information, and the network parameters are optimized by minimizing the cross entropy loss function to obtain a trained state recognition model.

[0024] For the newly acquired plant 3D model, its phenotypic parameters are extracted and input into the trained state recognition model. The predicted probability distribution of maturity state is obtained through forward propagation, and the maturity state with the highest probability is taken as the recognition result.

[0025] In an optional embodiment,

[0026] According to the plant growth data and plant maturity data, it is judged whether the plants in the corridor have reached the harvesting conditions. If the plants have reached the harvesting conditions, the harvesting execution mechanism is controlled to harvest the plants in the corridor and obtain the harvesting volume data. Based on the harvesting volume data, the commodity rate and inventory are calculated, including:

[0027] Input plant growth data and plant maturity data into a pre-trained mapping model between plant growth status and harvesting conditions to determine whether the plant has reached the optimal harvesting condition;

[0028] When the plant reaches the optimal harvesting condition, the harvesting decision result is output and transmitted to the harvesting execution mechanism. The harvesting execution mechanism uses the three-dimensional posture information of the target fruit, visual servo and deep reinforcement learning algorithms to accurately locate the target fruit and perform adaptive picking;

[0029] The picked fruits are transported to the intelligent sorting system via a conveyor belt. The intelligent sorting system uses computer vision algorithms to detect the appearance, size and quality of the fruits, automatically sorts them according to the preset grading standards, and counts the number and weight of fruits at each grade to generate harvest volume data and inventory data.

[0030] The commodity rate and inventory volume are calculated based on the harvest volume data and inventory data and uploaded to the cloud management platform in real time. Using big data analysis and machine learning algorithms, the harvest plan, resource scheduling and product demand are dynamically predicted and optimized, and the optimization results are fed back to the harvest decision model and harvest execution agency to form a closed-loop control.

[0031] In an optional embodiment,

[0032] The harvesting actuator uses the three-dimensional position information of the target fruit, visual servo and deep reinforcement learning algorithms to accurately locate the target fruit and perform adaptive picking, including:

[0033] The harvesting execution mechanism decomposes and optimizes the harvesting tasks according to the received harvesting instructions and the three-dimensional posture information of the target fruit, combined with the corridor environment map, and uses a distributed task allocation algorithm to generate the target area and navigation path of each harvesting robot, and sends them to the corresponding harvesting robot;

[0034] The harvesting robot moves to the designated position through the autonomous navigation algorithm according to the received target area and navigation path, and uses the three-dimensional posture error as the feedback signal to control the joint movement of the multi-degree-of-freedom robotic arm through the adaptive fuzzy control algorithm, and adjusts the movement trajectory and speed of the robotic arm in real time to accurately align the end effector with the target fruit;

[0035] The end effector adaptively adjusts the inflatable cavity inside the gripper to make the gripper fit softly with the surface of the target fruit. At the same time, a near-infrared spectrometer is used to collect the reflectance spectrum data of the target fruit surface. The maturity and quality parameters of the fruit are evaluated through a spectral analysis algorithm to obtain a quality evaluation value.

[0036] The harvesting robot uses the obtained quality evaluation value as the reward function, the robotic arm joint angle, speed and grasping force as the action space, and the three-dimensional posture information, size, maturity and environmental obstacle information of the target fruit as the state space. A deep reinforcement learning framework is constructed to enable the harvesting robot to learn to adaptively adjust the picking strategy in a dynamically changing environment through interactive exploration with the environment.

[0037] In an optional embodiment,

[0038] The calculated commodity rate and inventory are compared with the preset commodity rate threshold and inventory threshold respectively. If the commodity rate is lower than the preset commodity rate threshold or the inventory is higher than the preset inventory threshold, the environmental parameters in the closed area of ​​the corridor are adjusted. The environmental parameters in the closed area of ​​the corridor are controlled to optimize the plant growth conditions, including:

[0039] The commodity rate, inventory and energy consumption of the corridor closed harvesting environment are taken as the objective functions of the multi-objective optimization problem, and a decision variable vector including environmental parameters such as temperature, humidity and light is established to construct a multi-objective optimization model.

[0040] The multi-objective optimization model is solved by using a multi-objective optimization algorithm, and an intelligent mutation operator and an intelligent crossover operator are introduced to adaptively adjust the parameters of the multi-objective optimization algorithm to obtain an optimal solution set of corridor closure harvesting environment parameters;

[0041] Based on the optimal solution set of corridor closure harvesting environmental parameters, the decision preference information guided by agricultural production demand is introduced. By setting the target weights of commodity rate, inventory and energy consumption, as well as the defined target satisfaction threshold and fuzzy decision rules, the environmental parameter combinations in the optimal solution set are evaluated and ranked. Combined with the knowledge and practical experience of agricultural experts, the optimal combination of environmental parameters is obtained, and an intelligent tuning scheme for the environmental parameters is generated. The obtained intelligent tuning scheme is applied to the actual control process of the corridor closure harvesting environment.

[0042] In an optional embodiment,

[0043] The multi-objective optimization model is solved by using a multi-objective optimization algorithm, and an intelligent mutation operator and an intelligent crossover operator are introduced to adaptively adjust the parameters of the multi-objective optimization algorithm to obtain the optimal solution set of corridor closure harvesting environment parameters, including:

[0044] Initialize a particle swarm composed of a combination of environmental parameters, evaluate the fitness value and non-dominated relationship of each particle, and update the individual historical optimal position of the particle and the global non-dominated solution set;

[0045] The particle swarm is quickly sorted and the crowding distance is calculated. The particles with a priority higher than the threshold are selected as parent particles. An intelligent mutation operator is introduced to adaptively adjust the mutation step length and mutation probability. At the same time, an intelligent crossover operator is introduced to adaptively adjust the crossover mode and crossover probability to obtain the child particles.

[0046] The parent particles and the child particles are merged, and the target particle group is obtained by screening based on the non-dominated relationship. The adaptive weight factor and direction correction factor are introduced to update the particle velocity, generate a new position vector, and repeat the iteration until the preset termination condition is met to obtain the optimal solution set of corridor closure harvesting environment parameters.

[0047] The calculation formula of the adaptive weight factor is as follows:

[0048] ;

[0049] in, represents the adaptive weight factor, ω max represents the upper bound of the inertia weight, ω min represents the lower bound of the inertia weight, represents the fitness value of particle i at the tth iteration, Indicates the minimum fitness value of the current particle swarm, Represents the average fitness value of the current particle swarm;

[0050] The calculation formula of the direction correction factor is as follows:

[0051] ;

[0052] in, represents the direction correction factor, α max represents the upper bound of the direction correction factor, The position of particle i at iteration t, represents the position of particle k at the tth iteration, ||·||2 represents the Euclidean distance, and N represents the number of particles.

[0053] According to a second aspect of the embodiments of the present invention,

[0054] Provided is an automatic identification system for corridor closure harvesting, comprising:

[0055] The first unit is used to obtain environmental parameters in the closed area of ​​the corridor, wherein the environmental parameters include temperature, humidity and carbon dioxide concentration, and calculate the growth status of plants in the corridor according to a preset plant growth model in combination with the environmental parameters to obtain plant growth data; at the same time, images of plants in the corridor are obtained in real time and segmented to obtain plant area images, and color, texture and shape features of the plant area images are extracted and input into a pre-trained state recognition model to identify the maturity of the plants and obtain plant maturity data;

[0056] The second unit is used to judge whether the plants in the corridor meet the harvesting conditions according to the plant growth data and the plant maturity data. If the plants meet the harvesting conditions, the harvesting execution mechanism is controlled to harvest the plants in the corridor, and the harvesting quantity data is obtained. Based on the harvesting quantity data, the commodity rate and inventory are calculated;

[0057] The third unit is used to compare the calculated commodity rate and inventory with the preset commodity rate threshold and inventory threshold respectively. If the commodity rate is lower than the preset commodity rate threshold or the inventory is higher than the preset inventory threshold, the environmental parameters in the corridor closed area are adjusted to control the environmental parameters in the corridor closed area to optimize the plant growth conditions; according to the adjusted environmental parameters, the steps of calculating the commodity rate and inventory are repeated until the commodity rate reaches the preset commodity rate threshold and the inventory is lower than the preset inventory threshold, and the automatic identification of corridor closure and harvesting is completed.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] a memory for storing processor-executable instructions;

[0062] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0063] A fourth aspect of the embodiments of the present invention is:

[0064] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0065] In this embodiment, by monitoring environmental parameters, calculating plant growth conditions, and combining computer vision to identify plant maturity, automatic and accurate identification of harvesting conditions is achieved, and manual subjective judgment is no longer required, thereby improving the level of harvesting automation. According to the actual harvesting results, the environmental parameters are dynamically adjusted to form a closed-loop control, which can optimize the environmental conditions according to different planting varieties and growth cycles, so that the plant growth reaches the optimal state. By visually identifying the maturity of the plant, it can be harvested at the best time to avoid the quality degradation caused by premature or late harvesting. The preset commodity rate and inventory thresholds are used to digitize the harvesting results, which is conducive to standardized and intelligent production management, improving management efficiency and decision-making level. Personalized environmental control enables plant growth to reach the optimal state, thereby maximizing the yield per unit area. Automated harvesting recognition replaces manual operation, greatly reducing labor intensity and improving harvesting efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a schematic diagram of the process of the automatic identification method for corridor closure harvesting according to an embodiment of the present invention;

[0067] Figure 2 It is a schematic diagram of the structure of the automatic identification system for corridor closure harvesting according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0070] Figure 1 FIG. 1 is a flow chart of an automatic identification method for corridor closure harvesting according to an embodiment of the present invention, as shown in FIG. Figure 1 As shown, the method includes:

[0071] S101. Obtain environmental parameters within the closed area of ​​the corridor, wherein the environmental parameters include temperature, humidity and carbon dioxide concentration, and calculate the growth status of plants in the corridor according to a preset plant growth model in combination with the environmental parameters to obtain plant growth data; simultaneously, obtain images of plants in the corridor in real time and perform segmentation processing to obtain plant area images, extract color, texture and shape features of the plant area images, and input them into a pre-trained state recognition model to identify the maturity of the plants and obtain plant maturity data.

[0072] Exemplarily, environmental sensors are deployed in the closed area of ​​the corridor to collect environmental parameters such as temperature, humidity and carbon dioxide concentration in the corridor in real time. The collected environmental parameter data are uploaded to the data processing center through the wireless transmission module to provide data support for the subsequent calculation of plant growth conditions. According to the preset plant growth model, combined with the collected environmental parameters, the growth conditions of plants in the corridor are calculated. The plant growth model is a mathematical model established based on a large amount of experimental data and expert experience, which can accurately reflect the growth law of plants under different environmental conditions. By substituting the environmental parameters into the plant growth model, the growth condition data such as the growth rate, leaf area index, and dry matter accumulation of the plant can be obtained.

[0073] Video acquisition units are deployed in the corridor to obtain real-time image information of plants in the corridor. The acquired images are preprocessed, and plant area images are extracted through image segmentation algorithms. Feature extraction is performed on the plant area images to obtain the color, texture, and shape features of the plants. The extracted features are input into the pre-trained state recognition model to identify the maturity of the plants. The state recognition model uses a deep learning algorithm and is trained with a large amount of labeled plant image data. It can accurately identify plants at different maturity stages and obtain plant maturity data.

[0074] In an optional embodiment,

[0075] The images of plants in the corridor are acquired in real time and segmented to obtain plant area images. The color, texture and shape features of the plant area images are extracted and input into the pre-trained state recognition model to identify the maturity of the plants. The obtained plant maturity data include:

[0076] Capture multi-view images of plants in the corridor, obtain plant images from multiple viewpoints, segment and preprocess them, and obtain preprocessed multi-view plant images;

[0077] Extract feature points from the preprocessed multi-view plant images, control the number of extracted feature points by adjusting the parameters of the extraction algorithm, and obtain corresponding feature point pairs between images of different viewpoints based on the extracted feature points;

[0078] Using the corresponding feature point pairs, the 3D structure of the plant is restored through a 3D reconstruction algorithm to obtain multiple 3D plant models, and the corresponding phenotypic parameters are extracted from each 3D plant model. The phenotypic parameters include plant height, canopy volume, leaf area, stem diameter, and number of fruits.

[0079] According to each plant 3D model and the corresponding phenotypic parameters, construct the time series data of plant growth, analyze the time series data of plant growth, fit the plant growth curve, establish a plant maturity model, and use a machine learning algorithm to determine the optimal feature subset related to plant maturity through feature selection based on the phenotypic parameters of the plant 3D model and the plant maturity model, construct a state recognition model, and use the training data to train the state recognition model;

[0080] For the newly acquired plant three-dimensional model, its phenotypic parameters are extracted, and the extracted phenotypic parameters are input into the trained state recognition model to identify the maturity of the plant and obtain the plant maturity data.

[0081] For example, a multi-camera system is used to arrange multiple high-resolution digital cameras around the plant to simultaneously capture plant images from different perspectives. The number and layout of cameras need to be optimized according to the size and complexity of the plant to ensure that all parts of the plant can be covered and there is enough overlap between adjacent perspectives for subsequent 3D reconstruction.

[0082] After acquiring multi-view plant images, the images are preprocessed. Specifically, image filtering algorithms, such as Gaussian filtering and median filtering, are used to remove noise interference in the image and improve the signal-to-noise ratio of the image. Image enhancement techniques, such as histogram equalization and contrast stretching, are used to improve the contrast and dynamic range of the image, making the texture and details of the plant clearer. Finally, the camera calibration parameters are used to correct the image for distortion and chromatic aberration to ensure the geometric consistency and color authenticity of the image.

[0083] Furthermore, feature points are extracted from multi-view images, and the correspondence between different viewpoints is found. Common feature extraction algorithms include SIFT, SURF, ORB, etc. These algorithms can detect key points in images and calculate their local descriptors for feature matching. In the feature matching process, it is necessary to find the corresponding feature points between images with different viewpoints. Common feature matching algorithms include brute force matching, FLANN matching, etc.

[0084] Based on the matched feature point pairs, a 3D reconstruction algorithm can be used to restore the 3D structure of the plant. Commonly used 3D reconstruction algorithms include structured light, multi-view geometry, monocular vision, etc. The epipolar geometry relationship between multi-view images is used to calculate the 3D coordinates of the feature points through the triangulation principle. Optionally, the epipolar geometry relationship between multi-view images can be used to calculate the 3D coordinates of the feature points through the triangulation principle to obtain multiple 3D plant models, and various plant phenotypic parameters can be extracted from the reconstructed 3D models to characterize the growth status and maturity of the plant, where the plant phenotypic parameters include plant height, canopy volume, leaf area, stem thickness, and number of fruits.

[0085] In order to analyze the phenotypic changes of plants during their growth, it is necessary to perform multiple three-dimensional reconstructions and extract phenotypic parameters of the plants to construct time series data of plant growth. By analyzing the time series data, the characteristics of different growth stages of plants can be characterized, and growth curves and maturity models can be established. Specifically, the plant growth curve is fitted by performing time series analysis on phenotypic parameters such as plant height and canopy volume. The growth curve reflects the rate and trend of plants at different growth stages and can be used to compare and predict growth patterns. A plant maturity model is established by performing time series analysis on maturity-related parameters such as the number of plant fruits and fruit size. The maturity model describes the process from flowering to fruit maturity of plants, and can be used to predict harvest time and evaluate fruit quality.

[0086] Based on the three-dimensional phenotypic parameters of the plant and the results of time series analysis, the maturity of the plant can be estimated. Optionally, a maturity estimation model can be established through machine learning algorithms, such as support vector machines (SVM), random forests, etc. Specifically, features related to maturity, such as fruit size, color, hardness, etc., are selected from the plant phenotypic parameters. Feature selection algorithms, such as correlation coefficient method, recursive feature elimination method, etc., can be used to determine the optimal feature subset. The maturity estimation model is trained using the selected features and sample data of known maturity. The performance and generalization ability of the model are evaluated by methods such as cross-validation. For the newly collected three-dimensional plant data, the corresponding phenotypic parameters are extracted and input into the trained maturity estimation model to predict the maturity of the plant.

[0087] In this embodiment, through multi-view image acquisition and three-dimensional reconstruction, a full-view three-dimensional model of the plant can be obtained, and the extracted phenotypic parameter information is richer and more accurate. At the same time, based on time series data analysis and growth curve fitting, the model can adapt to changes in different growth stages of plants. The machine learning algorithm is used to automatically mine the optimal feature subset related to maturity from multi-source phenotypic parameters as the input of the maturity recognition model, thereby greatly improving the recognition accuracy. Based on the time series data and phenotypic parameters of plant growth, the entire growth process can be restored. At the same time, through feature selection, a feature subset highly correlated with maturity is mined, and the judgment basis of the model has strong interpretability and can be traced back to the physical properties of the plant, which is conducive to agricultural experts to analyze and optimize planting plans.

[0088] In an optional embodiment,

[0089] For the newly acquired plant three-dimensional model, the phenotypic parameters are extracted, and the extracted phenotypic parameters are input into the trained state recognition model to identify the maturity of the plant. The steps before obtaining the plant maturity data also include:

[0090] The phenotypic parameters of the plant three-dimensional model are normalized, and the normalized phenotypic parameters are resampled by using the farthest point sampling algorithm to obtain a resampled parameter matrix;

[0091] The resampled parameter matrix is ​​input into the state recognition model, which extracts feature maps of different scales and abstraction levels through multiple convolutional layers, downsamples the features through the pooling layer, and finally fuses the features through the fully connected layer to generate the final feature representation;

[0092] Input the obtained feature representation into a classifier, map the feature representation into a probability distribution of plant maturity states, and the dimension of the probability distribution is equal to the number of predefined maturity state categories;

[0093] The convolutional neural network is trained using the labeled plant maturity status information, and the network parameters are optimized by minimizing the cross entropy loss function to obtain a trained state recognition model.

[0094] For the newly acquired plant 3D model, its phenotypic parameters are extracted and input into the trained state recognition model. The predicted probability distribution of maturity state is obtained through forward propagation, and the maturity state with the highest probability is taken as the recognition result.

[0095] For example, the phenotypic parameters of the plant three-dimensional model are first normalized, and the parameters of different dimensions can be converted to the same numerical range, such as [0, 1] or standard normal distribution, using methods such as Min-Max normalization or Z-score normalization. Further, a point is randomly selected as a seed point, and then the point farthest from the seed point is found among the remaining points and added to the sampling set. The process is repeated until the required number of points is sampled to obtain the resampled parameter matrix.

[0096] The resampled parameter matrix is ​​used as the input of the state recognition model, which uses a convolutional neural network. The convolution layer performs convolution operations on the input through sliding convolution kernels to extract local features. Each convolution layer can extract different feature patterns, such as edges, textures, etc. The pooling layer usually uses maximum pooling or average pooling to downsample the neighborhood of the feature map, reduce the feature dimension, and improve the generalization ability of the model; the fully connected layer concatenates and transforms the pooled features to generate the final feature representation vector.

[0097] The feature representation vector is input into the fully connected layer as a classifier, which maps the feature representation to a probability distribution of predefined maturity state categories, with the dimension of the distribution equal to the number of predefined maturity state categories. A cross entropy loss function is defined to measure the difference between the model prediction and the true label. The parameters of the state recognition model are updated using an optimization algorithm (such as SGD, Adam, etc.) to minimize the loss function. During the training process, the model will gradually learn to extract feature representations that distinguish maturity states from phenotypic parameters.

[0098] For the new plant three-dimensional model, its phenotypic parameters are extracted, normalized and resampled. The preprocessed parameter matrix is ​​input into the trained state recognition model. The model performs forward propagation, calculates the predicted probability distribution of the maturity state, and takes the category with the largest probability as the recognition result.

[0099] In this embodiment, the phenotypic parameters of the plant three-dimensional model are directly input into the convolutional neural network for end-to-end learning, abandoning the traditional complex manual feature extraction and feature engineering process, greatly simplifying the processing flow of maturity identification, and improving the automation level of the system. Directly utilizing the rich phenotypic parameter data in the plant three-dimensional model, including the geometric and structural information of the plant in three-dimensional space, can better characterize the state characteristics of the plant compared to the traditional method based on two-dimensional images, thereby improving the accuracy and generalization ability of maturity identification. Automated maturity identification based on the plant three-dimensional model is realized, which combines high precision, high efficiency and good scalability, and can greatly improve the intelligence and automation level of agricultural production.

[0100] S102. Determine whether the plants in the corridor have reached the harvesting conditions based on the plant growth data and the plant maturity data. If the plants have reached the harvesting conditions, control the harvesting execution mechanism to perform harvesting operations on the plants in the corridor, and obtain the harvesting volume data. Based on the harvesting volume data, calculate the commodity rate and inventory.

[0101] Among them, the commodity rate refers to the proportion of plants that meet the preset commodity standards in the total harvest volume, reflecting the quality of the plants. The inventory volume refers to the difference between the harvest volume and the shipment volume within the preset unit time, reflecting the adequacy of plant supply.

[0102] In an optional embodiment,

[0103] According to the plant growth data and plant maturity data, it is judged whether the plants in the corridor have reached the harvesting conditions. If the plants have reached the harvesting conditions, the harvesting execution mechanism is controlled to harvest the plants in the corridor and obtain the harvesting volume data. Based on the harvesting volume data, the commodity rate and inventory are calculated, including:

[0104] Input plant growth data and plant maturity data into a pre-trained mapping model between plant growth status and harvesting conditions to determine whether the plant has reached the optimal harvesting condition;

[0105] When the plant reaches the optimal harvesting condition, the harvesting decision result is output and transmitted to the harvesting execution mechanism. The harvesting execution mechanism uses the three-dimensional posture information of the target fruit, visual servo and deep reinforcement learning algorithms to accurately locate the target fruit and perform adaptive picking;

[0106] The picked fruits are transported to the intelligent sorting system via a conveyor belt. The intelligent sorting system uses computer vision algorithms to detect the appearance, size and quality of the fruits, automatically sorts them according to the preset grading standards, and counts the number and weight of fruits at each grade to generate harvest volume data and inventory data.

[0107] The commodity rate and inventory volume are calculated based on the harvest volume data and inventory data and uploaded to the cloud management platform in real time. Using big data analysis and machine learning algorithms, the harvest plan, resource scheduling and product demand are dynamically predicted and optimized, and the optimization results are fed back to the harvest decision model and harvest execution agency to form a closed-loop control.

[0108] Exemplarily, first, a plant growth state feature vector is constructed, and key indicators (such as plant height, leaf area, fruit color, etc.) in the plant growth and maturity data set are combined into a multidimensional feature vector to represent the growth state of the plant at a certain moment. Using historical planting data, the plant growth state feature vector is annotated to mark whether it has reached the optimal harvesting condition. Samples that have reached the optimal harvesting condition are marked as positive examples, and samples that have not reached the optimal harvesting condition are marked as negative examples to construct a binary classification data set. Select an appropriate machine learning algorithm, such as support vector machine (SVM), random forest, etc., to train the annotated data set, and establish a classification model between the plant growth state feature vector and whether the harvesting condition is reached. The accuracy and generalization ability of the model are evaluated by methods such as cross-validation. Using the trained harvesting decision model, the plant growth state feature vector collected in real time is predicted to determine whether the plant has reached the harvesting condition. When the result output by the model is a positive example, it is considered that the plant has reached the optimal harvesting condition, triggering the harvesting execution mechanism for automatic harvesting.

[0109] When the harvesting decision model determines that the plant has reached the harvesting conditions, it is necessary to control the harvesting actuator to automatically harvest the plant. The harvesting actuator usually includes a harvesting robot, a conveyor belt, a sorting system, etc., which need to work together to complete tasks such as picking, transporting and sorting plants. Specifically, the harvesting robot is equipped with a high-precision visual sensor, such as an RGB-D camera, to take real-time photos of the plant and obtain the three-dimensional point cloud data of the plant. Point cloud processing algorithms, such as clustering and segmentation, are used to identify the position and posture information of the ripe fruit. According to the identified fruit position and posture information, the inverse kinematics algorithm is used to calculate the target position and posture of the end effector (such as a manipulator) of the harvesting robot to achieve accurate positioning of the fruit. A motion planning algorithm based on a potential field is used to treat the plant as an obstacle, generate an artificial potential field, guide the robot to avoid obstacles, and smoothly and safely reach the target fruit position. After the harvesting robot reaches the target fruit position, the manipulator or a special picking tool is used to separate the fruit from the plant with appropriate force and angle and place it on the conveyor belt. During the harvesting process, force sensors monitor the picking force in real time to avoid damage to the fruit. The conveyor belt transports the harvested fruit to the sorting system, which uses computer vision technology to detect the appearance and size of the fruit, automatically sorts it according to the preset grading standards, and counts the number and weight of each grade of fruit.

[0110] In the process of automated harvesting, it is necessary to obtain the harvest volume data in real time, including the number and weight of harvested fruits, so as to analyze the harvest volume data and calculate the commodity rate and inventory. Specifically, a weight sensor is set in the sorting system to weigh the fruits of each level after sorting and obtain their weight data. At the same time, the number of fruits of each level is counted through a counter or visual recognition technology. The obtained fruit weight and quantity data are uploaded to the data center and associated with information such as harvest batches and harvest time to form a structured harvest volume data set. ) The harvest volume data is cleaned and preprocessed to remove abnormal values ​​and duplicate values ​​to ensure the accuracy and consistency of the data. Calculate the commodity rate index. The commodity rate reflects the proportion of harvested fruits that reach the commodity level and is an important indicator for measuring the quality of plant production. By analyzing the grading data, the proportion of the number of fruits of each level is calculated to obtain the commodity rate. The inventory reflects the number of commodity fruits currently available for sale or processing, which is the basis for production and marketing balance and inventory management. The current inventory is calculated by comparing and analyzing the harvest volume data and sales and outbound data.

[0111] Commodity rate = total number of harvested fruits ÷ number of commercial grade fruits × 100%;

[0112] Inventory volume = cumulative harvest volume − cumulative sales volume.

[0113] In an optional embodiment,

[0114] The harvesting actuator uses the three-dimensional position information of the target fruit, visual servo and deep reinforcement learning algorithms to accurately locate the target fruit and perform adaptive picking, including:

[0115] The harvesting execution mechanism decomposes and optimizes the harvesting tasks according to the received harvesting instructions and the three-dimensional posture information of the target fruit, combined with the corridor environment map, and uses a distributed task allocation algorithm to generate the target area and navigation path of each harvesting robot, and sends them to the corresponding harvesting robot;

[0116] The harvesting robot moves to the designated position through the autonomous navigation algorithm according to the received target area and navigation path, and uses the three-dimensional posture error as the feedback signal to control the joint movement of the multi-degree-of-freedom robotic arm through the adaptive fuzzy control algorithm, and adjusts the movement trajectory and speed of the robotic arm in real time to accurately align the end effector with the target fruit;

[0117] The end effector adaptively adjusts the inflatable cavity inside the gripper to make the gripper fit softly with the surface of the target fruit. At the same time, a near-infrared spectrometer is used to collect the reflectance spectrum data of the target fruit surface. The maturity and quality parameters of the fruit are evaluated through a spectral analysis algorithm to obtain a quality evaluation value.

[0118] The harvesting robot uses the obtained quality evaluation value as the reward function, the robotic arm joint angle, speed and grasping force as the action space, and the three-dimensional posture information, size, maturity and environmental obstacle information of the target fruit as the state space. A deep reinforcement learning framework is constructed to enable the harvesting robot to learn to adaptively adjust the picking strategy in a dynamically changing environment through interactive exploration with the environment.

[0119] For example, multiple harvesting robots are deployed in the closed corridors of large-scale plant factories, and the efficient execution of harvesting tasks is achieved through a multi-robot collaborative control system. The collaborative control system is based on a distributed architecture. Each harvesting robot is equipped with an independent control unit and communication module, and communicates with the central dispatching system in real time through a wireless network. The dispatching system decomposes and optimizes the harvesting tasks based on the harvesting decision results and corridor environment information, and assigns target areas and path planning to each harvesting robot. The harvesting robot autonomously navigates to the designated area to perform harvesting operations based on its own position and task instructions.

[0120] The harvesting robot uses a computer vision algorithm based on deep learning to accurately detect and locate the target fruit. The robot is equipped with a high-definition color camera and a depth camera to obtain RGB images and three-dimensional point cloud data of the fruit. Using pre-trained convolutional neural networks (such as Faster R-CNN, YOLO, etc.), the image is processed in real time to detect the position and bounding box of the ripe fruit. At the same time, the three-dimensional spatial coordinates and posture information of the target fruit are obtained through the registration and segmentation of the point cloud data. In order to adapt to different lighting and occlusion conditions, the attention mechanism and contextual information are also introduced to improve the robustness of detection and positioning. The detection and positioning results are passed to the motion planning module of the harvesting robot in the form of the three-dimensional pose of the target fruit.

[0121] The end effector of the harvesting robot adopts a flexible design, equipped with a multi-degree-of-freedom robotic arm and a soft gripper to achieve precise operation and flexible grasping of the target fruit. The motion control of the robotic arm adopts a closed-loop control strategy based on visual servoing, using the three-dimensional position of the target fruit as a feedback signal to adjust the motion trajectory and speed of the robotic arm in real time. Visual servo control combines position control and speed control, and achieves real-time synchronization and precise alignment between the robotic arm and the target fruit through proportional-integral-differential (PID) controller and Jacobian matrix estimation.

[0122] For the fruit grabbing operation, a flexible grabbing control method based on force feedback is adopted. The soft gripper has a built-in pressure sensor array to detect the contact force and pressure distribution during the grabbing process in real time. Through the force feedback signal, the deformation and gripping force of the gripper are adaptively adjusted to achieve soft and non-destructive grabbing of the fruit. At the same time, the gripper is also integrated with a near-infrared spectral sensor to further judge the maturity and quality of the fruit by analyzing the spectral characteristics of the fruit surface, and obtain the quality evaluation value.

[0123] The harvesting robot uses the obtained quality evaluation value as the reward function, the robotic arm joint angle, speed and grasping force as the action space, and the three-dimensional posture information, size, maturity and environmental obstacle information of the target fruit as the state space. A deep reinforcement learning framework is constructed to enable the harvesting robot to learn to adaptively adjust the picking strategy in a dynamically changing environment through interactive exploration with the environment.

[0124] Select a deep reinforcement learning algorithm, where the Actor network (policy network) outputs actions, including joint angles, speeds, and grasping strength, and the Critic network (value function network) evaluates the value function of the current state and action. Build a deep neural network, where the input layer is used to splice the features of each part of the state space, the convolution layer is used to extract the spatial features of image and point cloud data, the LSTM layer is used to capture time series information, and the full connection is used to fuse different features, and output state value estimates or Q values.

[0125] Deploy the trained deep neural network model to the harvesting robot hardware system, test the model performance in a real picking scenario, collect indicators such as harvesting success rate, quality score, efficiency, etc., analyze and evaluate the results, optimize and iterate the model and reward function design, and use technologies such as transfer learning to speed up model training in new scenarios.

[0126] In this embodiment, by inputting plant growth data and maturity data into the pre-trained model, it is possible to automatically determine whether the best time for harvesting has been reached, avoiding the traditional harvesting strategy based on experience or fixed time, and improving the quality and benefits of harvesting. Combining three-dimensional posture information, visual servoing and deep reinforcement learning, the target fruit can be accurately positioned and adaptively picked, reducing picking damage and improving the level of automation. The adaptive grasping strategy can adapt to differences such as different varieties and sizes, and enhance the versatility of the system. The harvesting data is uploaded to the cloud, combined with big data analysis and machine learning prediction, the dynamic optimization of harvesting plans, resource allocation and product demand can be achieved, and the optimization results can be fed back to the decision-making model and the executive agency to form a closed-loop control, so that the entire harvesting process is dynamically adaptive and continuously optimized to ensure the best harvesting effect. The integration of automated picking, intelligent sorting and mobile navigation has greatly improved operating efficiency, reduced manpower requirements and improved economic benefits.

[0127] S103. Compare the calculated commodity rate and inventory with the preset commodity rate threshold and inventory threshold respectively. If the commodity rate is lower than the preset commodity rate threshold or the inventory is higher than the preset inventory threshold, adjust the environmental parameters in the corridor closed area, and control the environmental parameters in the corridor closed area to optimize plant growth conditions; according to the adjusted environmental parameters, repeat the steps of calculating the commodity rate and inventory until the commodity rate reaches the preset commodity rate threshold and the inventory is lower than the preset inventory threshold, and complete the automatic identification of corridor closure and harvesting.

[0128] In an optional embodiment,

[0129] The calculated commodity rate and inventory are compared with the preset commodity rate threshold and inventory threshold respectively. If the commodity rate is lower than the preset commodity rate threshold or the inventory is higher than the preset inventory threshold, the environmental parameters in the closed area of ​​the corridor are adjusted. The environmental parameters in the closed area of ​​the corridor are controlled to optimize the plant growth conditions, including:

[0130] The commodity rate, inventory and energy consumption of the corridor closed harvesting environment are taken as the objective functions of the multi-objective optimization problem, and a decision variable vector including environmental parameters such as temperature, humidity and light is established to construct a multi-objective optimization model.

[0131] The multi-objective optimization model is solved by using a multi-objective optimization algorithm, and an intelligent mutation operator and an intelligent crossover operator are introduced to adaptively adjust the parameters of the multi-objective optimization algorithm to obtain an optimal solution set of corridor closure harvesting environment parameters;

[0132] Based on the optimal solution set of corridor closure harvesting environmental parameters, the decision preference information guided by agricultural production demand is introduced. By setting the target weights of commodity rate, inventory and energy consumption, as well as the defined target satisfaction threshold and fuzzy decision rules, the environmental parameter combinations in the optimal solution set are evaluated and ranked. Combined with the knowledge and practical experience of agricultural experts, the optimal combination of environmental parameters is obtained, and an intelligent tuning scheme for the environmental parameters is generated. The obtained intelligent tuning scheme is applied to the actual control process of the corridor closure harvesting environment.

[0133] For example, in a closed corridor harvesting environment, the tuning of environmental parameters needs to comprehensively consider multiple objectives, such as commodity rate, inventory, energy consumption, etc. In order to achieve intelligent tuning of environmental parameters, it is first necessary to model it as a multi-objective optimization problem.

[0134] Aiming at the multi-objective optimization problem of intelligent tuning of environmental parameters, a solution method that combines the global search capability of particle swarm optimization algorithm and the local search capability of evolution strategy is adopted. By adaptively adjusting the flight behavior and evolution operation of particles, efficient multi-objective optimization solution is achieved. Furthermore, an adaptive coding strategy is introduced to dynamically adjust the coding accuracy of decision variables according to the convergence state of particles in the optimization process. In the initial stage of the algorithm, a lower coding accuracy is adopted to speed up the global exploration speed of particles; as the iteration proceeds, the coding accuracy is gradually improved to enhance the local search ability of particles. The adjustment of coding accuracy is achieved through adaptive functions, such as exponential functions or logarithmic functions. Intelligent mutation operators and intelligent crossover operators are also introduced. The intelligent mutation operator adaptively adjusts the mutation step length and mutation probability according to the fitness value and diversity information of particles. For particles with higher fitness values ​​and lower diversity, a smaller mutation step length and higher mutation probability are adopted to promote local search; for particles with lower fitness values ​​and higher diversity, a larger mutation step length and lower mutation probability are adopted to encourage global exploration. The intelligent crossover operator adaptively selects the crossover mode and crossover probability based on historical optimization experience by introducing a knowledge-driven mechanism. For particles near the historical optimal solution, sophisticated crossover methods such as uniform crossover or arithmetic crossover are used to increase the crossover probability; for particles far away from the historical optimal solution, simple crossover methods such as two-point crossover or multi-point crossover are used to reduce the crossover probability.

[0135] Through the solution method combining multi-objective particle swarm optimization with evolutionary strategy, the Pareto optimal solution set of environmental parameters can be obtained. According to factors such as agricultural product market demand and energy cost, the weights of commodity rate, inventory and energy consumption targets are reasonably set to comprehensively evaluate the pros and cons of each Pareto optimal solution. It is also possible to set the satisfaction threshold of each target based on the knowledge and experience of agricultural experts and select the Pareto optimal solution that meets the threshold requirements. Through the introduction and analysis of decision preferences, the environmental parameter combination that best meets actual needs can be selected from the Pareto optimal solution set, and an intelligent tuning scheme for environmental parameters can be generated, which can be applied to the actual control process of the corridor closed harvesting environment.

[0136] In an optional embodiment,

[0137] The multi-objective optimization model is solved by using a multi-objective optimization algorithm, and an intelligent mutation operator and an intelligent crossover operator are introduced to adaptively adjust the parameters of the multi-objective optimization algorithm to obtain the optimal solution set of corridor closure harvesting environment parameters, including:

[0138] Initialize a particle swarm composed of a combination of environmental parameters, evaluate the fitness value and non-dominated relationship of each particle, and update the individual historical optimal position of the particle and the global non-dominated solution set;

[0139] The particle swarm is quickly sorted and the crowding distance is calculated. The particles with a priority higher than the threshold are selected as parent particles. An intelligent mutation operator is introduced to adaptively adjust the mutation step length and mutation probability. At the same time, an intelligent crossover operator is introduced to adaptively adjust the crossover mode and crossover probability to obtain the child particles.

[0140] The parent particles and the child particles are merged, and the target particle group is obtained by screening based on the non-dominated relationship. The adaptive weight factor and direction correction factor are introduced to update the particle velocity, generate a new position vector, and repeat the iteration until the preset termination condition is met to obtain the optimal solution set of corridor closure harvesting environment parameters.

[0141] The calculation formula of the adaptive weight factor is as follows:

[0142] ;

[0143] in, represents the adaptive weight factor, ω max represents the upper bound of the inertia weight, ω min represents the lower bound of the inertia weight, represents the fitness value of particle i at the tth iteration, Indicates the minimum fitness value of the current particle swarm, Represents the average fitness value of the current particle swarm;

[0144] The calculation formula of the direction correction factor is as follows:

[0145] ;

[0146] in, represents the direction correction factor, α max represents the upper bound of the direction correction factor, The position of particle i at iteration t, represents the position of particle k at the tth iteration, ||·||2 represents the Euclidean distance, and N represents the number of particles.

[0147] In this embodiment, multiple key indicators such as commodity rate, inventory and energy consumption are used as optimization targets, and a multi-objective optimization model is constructed, so that the adjustment of environmental parameters can balance the pros and cons between various targets and achieve global optimality. Intelligent mutation and intelligent crossover operators are introduced to adaptively adjust the parameters of the multi-objective optimization algorithm, which improves the search ability and convergence speed of the algorithm and enhances the intelligence level of the system. The knowledge and practical experience of agricultural experts are introduced into the decision-making process, and the optimal solution set is evaluated and ranked in combination with quantitative target weights and fuzzy decision-making rules to make the decision more scientific and reasonable. The optimized environmental parameter tuning scheme is applied to the actual control link, forming an environmental parameter closed-loop adjustment mechanism based on multi-objective optimization and intelligent decision-making, realizing the automatic generation and dynamic optimization adjustment of the control strategy, and improving the intelligence and adaptability of the entire system.

[0148] In another optional embodiment,

[0149] After the harvesting of agricultural products is completed in the greenhouse, the harvested agricultural products are automatically transferred to the closed corridor. At the entrance of the corridor where agricultural products enter, a radio frequency identification (RFID) tag reading and writing device is set up to uniquely identify each harvesting box loaded with agricultural products, and write the relevant information of the harvesting box (such as harvesting time, harvesting personnel, variety, etc.) into the RFID tag. When the harvesting box moves in the corridor, the RFID reading equipment deployed along the line can continuously track the location and flow of the harvesting box, realizing the full traceability management of agricultural products.

[0150] Inside the corridor, computer vision and image recognition technology are used to identify the variety and grade the quality of agricultural products in the harvesting boxes. By training deep learning models such as convolutional neural networks, the appearance characteristics, color, size, etc. of agricultural products are extracted and analyzed, and the variety category and grade of agricultural products are automatically determined, providing a basis for subsequent sorting and packaging. At the same time, by associating with RFID tag information, key data such as the variety, quantity and quality of each batch of agricultural products can be counted and recorded.

[0151] In order to ensure the freshness and quality of agricultural products, intelligent weighing and counting equipment is installed in the corridor to accurately measure the agricultural products in the harvesting boxes. High-precision electronic scales and visual counting algorithms are used to automatically obtain the weight and quantity information of agricultural products, and upload the data to the central management system to facilitate inventory management and production and marketing decisions. The weighing and counting equipment works in conjunction with the RFID tag reading and writing equipment to achieve real-time collection and updating of agricultural product data.

[0152] Environmental monitoring sensors and control equipment are deployed in different areas of the corridor to monitor parameters such as temperature, humidity, and gas concentration in real time, and dynamically adjust the environmental conditions in the corridor according to the preservation needs of agricultural products. By establishing a mathematical model between agricultural product quality and environmental factors, the quality change trend of agricultural products under different environments is predicted, and the control strategy is optimized accordingly, so that agricultural products are always in the best preservation state in the corridor.

[0153] Special fresh-keeping treatment zones are set up in the corridor for different types and quality grades of agricultural products. Each zone is equipped with corresponding pre-cooling equipment and control systems. According to the optimal storage temperature and cooling rate of agricultural products, precise pre-cooling treatment is implemented to make agricultural products reach the ideal initial temperature before storage, thus extending the shelf life and shelf life. During the pre-cooling treatment process, the core temperature data of agricultural products is collected in real time through the wireless sensor network, and the pre-cooling curve is fitted and optimized in combination with thermodynamic models and intelligent algorithms to improve pre-cooling efficiency and energy utilization.

[0154] At the end of the corridor, automatic sorting and packaging equipment is set up to sort agricultural products to different packaging lines according to their variety, grade and target market. The sorting equipment accurately identifies the attributes and contents of each harvesting box through RFID tag recognition and machine vision technology, and intelligently plans the sorting route and sequence based on order information and inventory status. On the packaging line, automated packaging machines and intelligent coding equipment are used to quickly and standardizedly package and label agricultural products, improving commercial processing efficiency and product quality.

[0155] Finally, the sorted and packaged agricultural products are intelligently dispatched and distributed. Using the Internet of Things technology and geographic information system, the location and status of agricultural products are monitored in real time, the transportation routes and dispatch strategies are optimized, the time and energy costs during transportation are minimized, and the agricultural products are delivered to the target market and customers in a timely and safe manner.

[0156] In this embodiment, by combining the coordinated application of automatic identification, intelligent processing and precise control technology, the automatic identification method for corridor closure harvesting can significantly improve the efficiency, quality and traceability of post-harvest processing of agricultural products, reduce manual operation and management risks, and provide strong support for the intelligent development of modern agriculture. With the continuous advancement of cutting-edge technologies such as the Internet of Things, big data, and artificial intelligence, this method will continue to expand and deepen in intelligent perception, decision optimization, remote diagnosis, etc., and promote agricultural production to a more intelligent, refined, high-quality and efficient direction.

[0157] Figure 2 FIG. 1 is a schematic diagram of the structure of the automatic identification system for corridor closure harvesting according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0158] The first unit is used to obtain environmental parameters in the closed area of ​​the corridor, wherein the environmental parameters include temperature, humidity and carbon dioxide concentration, and calculate the growth status of plants in the corridor according to a preset plant growth model in combination with the environmental parameters to obtain plant growth data; at the same time, images of plants in the corridor are obtained in real time and segmented to obtain plant area images, and color, texture and shape features of the plant area images are extracted and input into a pre-trained state recognition model to identify the maturity of the plants and obtain plant maturity data;

[0159] The second unit is used to judge whether the plants in the corridor meet the harvesting conditions according to the plant growth data and the plant maturity data. If the plants meet the harvesting conditions, the harvesting execution mechanism is controlled to harvest the plants in the corridor, and the harvesting quantity data is obtained. Based on the harvesting quantity data, the commodity rate and inventory are calculated;

[0160] The third unit is used to compare the calculated commodity rate and inventory with the preset commodity rate threshold and inventory threshold respectively. If the commodity rate is lower than the preset commodity rate threshold or the inventory is higher than the preset inventory threshold, the environmental parameters in the corridor closed area are adjusted to control the environmental parameters in the corridor closed area to optimize the plant growth conditions; according to the adjusted environmental parameters, the steps of calculating the commodity rate and inventory are repeated until the commodity rate reaches the preset commodity rate threshold and the inventory is lower than the preset inventory threshold, and the automatic identification of corridor closure and harvesting is completed.

[0161] According to a third aspect of the embodiments of the present invention,

[0162] An electronic device is provided, comprising:

[0163] processor;

[0164] a memory for storing processor-executable instructions;

[0165] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0166] A fourth aspect of the embodiments of the present invention is:

[0167] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0168] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatically identifying closed corridor harvesting, characterized in that: include: Acquire environmental parameters in the closed area of ​​the corridor, the environmental parameters including temperature, humidity and carbon dioxide concentration, calculate the growth status of plants in the corridor according to a preset plant growth model and the environmental parameters, and obtain plant growth data; at the same time, acquire images of plants in the corridor in real time and perform segmentation processing to obtain plant area images, extract color, texture and shape features of the plant area images, input them into a pre-trained state recognition model, identify the maturity of the plants, and obtain plant maturity data; According to the plant growth data and plant maturity data, it is judged whether the plants in the corridor have reached the harvesting conditions. If the plants have reached the harvesting conditions, the harvesting execution mechanism is controlled to harvest the plants in the corridor and obtain the harvesting quantity data. Based on the harvesting quantity data, the commodity rate and inventory are calculated; The calculated commodity rate and inventory are compared with the preset commodity rate threshold and inventory threshold respectively. If the commodity rate is lower than the preset commodity rate threshold or the inventory is higher than the preset inventory threshold, the environmental parameters in the corridor closed area are adjusted to control the environmental parameters in the corridor closed area to optimize the plant growth conditions; according to the adjusted environmental parameters, the steps of calculating the commodity rate and inventory are repeated until the commodity rate reaches the preset commodity rate threshold and the inventory is lower than the preset inventory threshold, and the automatic identification of corridor closure and harvesting is completed.

2. The method according to claim 1, characterized in that The images of plants in the corridor are acquired in real time and segmented to obtain plant area images. The color, texture and shape features of the plant area images are extracted and input into the pre-trained state recognition model to identify the maturity of the plants. The obtained plant maturity data include: Capture multi-view images of plants in the corridor, obtain plant images from multiple viewpoints, segment and preprocess them, and obtain preprocessed multi-view plant images; Extract feature points from the preprocessed multi-view plant images, control the number of extracted feature points by adjusting the parameters of the extraction algorithm, and obtain corresponding feature point pairs between images of different viewpoints based on the extracted feature points; Using the corresponding feature point pairs, the 3D structure of the plant is restored through a 3D reconstruction algorithm to obtain multiple 3D plant models, and the corresponding phenotypic parameters are extracted from each 3D plant model. The phenotypic parameters include plant height, canopy volume, leaf area, stem diameter, and number of fruits. According to each plant 3D model and the corresponding phenotypic parameters, construct the time series data of plant growth, analyze the time series data of plant growth, fit the plant growth curve, establish a plant maturity model, and use a machine learning algorithm to determine the optimal feature subset related to plant maturity through feature selection based on the phenotypic parameters of the plant 3D model and the plant maturity model, construct a state recognition model, and use the training data to train the state recognition model; For the newly acquired plant three-dimensional model, its phenotypic parameters are extracted, and the extracted phenotypic parameters are input into the trained state recognition model to identify the maturity of the plant and obtain the plant maturity data.

3. The method according to claim 2, characterized in that For the newly acquired plant three-dimensional model, the phenotypic parameters are extracted, and the extracted phenotypic parameters are input into the trained state recognition model to identify the maturity of the plant. The steps before obtaining the plant maturity data also include: The phenotypic parameters of the plant three-dimensional model are normalized, and the normalized phenotypic parameters are resampled by using the farthest point sampling algorithm to obtain a resampled parameter matrix; The resampled parameter matrix is ​​input into the state recognition model, which extracts feature maps of different scales and abstraction levels through multiple convolutional layers, downsamples the features through the pooling layer, and finally fuses the features through the fully connected layer to generate the final feature representation; Input the obtained feature representation into a classifier, map the feature representation into a probability distribution of plant maturity states, and the dimension of the probability distribution is equal to the number of predefined maturity state categories; The convolutional neural network is trained using the labeled plant maturity status information, and the network parameters are optimized by minimizing the cross entropy loss function to obtain a trained state recognition model. For the newly acquired plant 3D model, its phenotypic parameters are extracted and input into the trained state recognition model. The predicted probability distribution of maturity state is obtained through forward propagation, and the maturity state with the highest probability is taken as the recognition result.

4. The method according to claim 1, characterized in that: According to the plant growth data and plant maturity data, it is judged whether the plants in the corridor have reached the harvesting conditions. If the plants have reached the harvesting conditions, the harvesting execution mechanism is controlled to harvest the plants in the corridor and obtain the harvesting volume data. Based on the harvesting volume data, the commodity rate and inventory are calculated, including: Input plant growth data and plant maturity data into a pre-trained mapping model between plant growth status and harvesting conditions to determine whether the plant has reached the optimal harvesting condition; When the plant reaches the optimal harvesting condition, the harvesting decision result is output and transmitted to the harvesting execution mechanism. The harvesting execution mechanism uses the three-dimensional posture information of the target fruit, visual servo and deep reinforcement learning algorithms to accurately locate the target fruit and perform adaptive picking; The picked fruits are transported to the intelligent sorting system via a conveyor belt. The intelligent sorting system uses computer vision algorithms to detect the appearance, size and quality of the fruits, automatically sorts them according to the preset grading standards, and counts the number and weight of fruits at each grade to generate harvest volume data and inventory data. The commodity rate and inventory volume are calculated based on the harvest volume data and inventory data and uploaded to the cloud management platform in real time. Using big data analysis and machine learning algorithms, the harvest plan, resource scheduling and product demand are dynamically predicted and optimized, and the optimization results are fed back to the harvest decision model and harvest execution agency to form a closed-loop control.

5. The method according to claim 4, characterized in that The harvesting actuator uses the three-dimensional position information of the target fruit, visual servo and deep reinforcement learning algorithms to accurately locate the target fruit and perform adaptive picking, including: The harvesting execution mechanism decomposes and optimizes the harvesting tasks according to the received harvesting instructions and the three-dimensional posture information of the target fruit, combined with the corridor environment map, and uses a distributed task allocation algorithm to generate the target area and navigation path of each harvesting robot, and sends them to the corresponding harvesting robot; The harvesting robot moves to the designated position through the autonomous navigation algorithm according to the received target area and navigation path, and uses the three-dimensional posture error as the feedback signal to control the joint movement of the multi-degree-of-freedom robotic arm through the adaptive fuzzy control algorithm, and adjusts the movement trajectory and speed of the robotic arm in real time to accurately align the end effector with the target fruit; The end effector adaptively adjusts the inflatable cavity inside the gripper to make the gripper fit softly with the surface of the target fruit. At the same time, a near-infrared spectrometer is used to collect the reflectance spectrum data of the target fruit surface. The maturity and quality parameters of the fruit are evaluated through a spectral analysis algorithm to obtain a quality evaluation value. The harvesting robot uses the obtained quality evaluation value as the reward function, the robotic arm joint angle, speed and grasping force as the action space, and the three-dimensional posture information, size, maturity and environmental obstacle information of the target fruit as the state space. A deep reinforcement learning framework is constructed to enable the harvesting robot to learn to adaptively adjust the picking strategy in a dynamically changing environment through interactive exploration with the environment.

6. The method according to claim 1, characterized in that The calculated commodity rate and inventory are compared with the preset commodity rate threshold and inventory threshold respectively. If the commodity rate is lower than the preset commodity rate threshold or the inventory is higher than the preset inventory threshold, the environmental parameters in the closed area of ​​the corridor are adjusted. The environmental parameters in the closed area of ​​the corridor are controlled to optimize the plant growth conditions, including: The commodity rate, inventory and energy consumption of the corridor closed harvesting environment are taken as the objective function of the multi-objective optimization problem, and a decision variable vector including temperature, humidity and light environment parameters is established to construct a multi-objective optimization model. The multi-objective optimization model is solved by using a multi-objective optimization algorithm, and an intelligent mutation operator and an intelligent crossover operator are introduced to adaptively adjust the parameters of the multi-objective optimization algorithm to obtain an optimal solution set of corridor closure harvesting environment parameters; Based on the optimal solution set of corridor closure harvesting environmental parameters, the decision preference information guided by agricultural production demand is introduced. By setting the target weights of commodity rate, inventory and energy consumption, as well as the defined target satisfaction threshold and fuzzy decision rules, the environmental parameter combinations in the optimal solution set are evaluated and ranked. Combined with the knowledge and practical experience of agricultural experts, the optimal combination of environmental parameters is obtained, and an intelligent tuning scheme for the environmental parameters is generated. The obtained intelligent tuning scheme is applied to the actual control process of the corridor closure harvesting environment.

7. The method according to claim 6, characterized in that The multi-objective optimization model is solved by using a multi-objective optimization algorithm, and an intelligent mutation operator and an intelligent crossover operator are introduced to adaptively adjust the parameters of the multi-objective optimization algorithm to obtain the optimal solution set of corridor closure harvesting environment parameters, including: Initialize a particle swarm composed of a combination of environmental parameters, evaluate the fitness value and non-dominated relationship of each particle, and update the individual historical optimal position of the particle and the global non-dominated solution set; The particle swarm is quickly sorted and the crowding distance is calculated. The particles with a priority higher than the threshold are selected as parent particles. An intelligent mutation operator is introduced to adaptively adjust the mutation step length and mutation probability. At the same time, an intelligent crossover operator is introduced to adaptively adjust the crossover mode and crossover probability to obtain the child particles. The parent particles and the child particles are merged, and the target particle group is obtained by screening based on the non-dominated relationship. The adaptive weight factor and direction correction factor are introduced to update the particle velocity, generate a new position vector, and repeat the iteration until the preset termination condition is met to obtain the optimal solution set of corridor closure harvesting environment parameters. The calculation formula of the adaptive weight factor is as follows: ； in, represents the adaptive weight factor, ω max represents the upper bound of the inertia weight, ω min represents the lower bound of the inertia weight, represents the fitness value of particle i at the tth iteration, Indicates the minimum fitness value of the current particle swarm, Represents the average fitness value of the current particle swarm; The calculation formula of the direction correction factor is as follows: ； in, represents the direction correction factor, α max represents the upper bound of the direction correction factor, The position of particle i at iteration t, represents the position of particle k at the tth iteration, ||·||2 represents the Euclidean distance, and N represents the number of particles.

8. An automatic identification system for closed corridor harvesting, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain environmental parameters in the closed area of ​​the corridor, wherein the environmental parameters include temperature, humidity and carbon dioxide concentration, and calculate the growth status of plants in the corridor according to a preset plant growth model in combination with the environmental parameters to obtain plant growth data; at the same time, images of plants in the corridor are obtained in real time and segmented to obtain plant area images, and color, texture and shape features of the plant area images are extracted and input into a pre-trained state recognition model to identify the maturity of the plants and obtain plant maturity data; The second unit is used to judge whether the plants in the corridor meet the harvesting conditions according to the plant growth data and the plant maturity data. If the plants meet the harvesting conditions, the harvesting execution mechanism is controlled to harvest the plants in the corridor, and the harvesting quantity data is obtained. Based on the harvesting quantity data, the commodity rate and inventory are calculated; The third unit is used to compare the calculated commodity rate and inventory with the preset commodity rate threshold and inventory threshold respectively. If the commodity rate is lower than the preset commodity rate threshold or the inventory is higher than the preset inventory threshold, the environmental parameters in the corridor closed area are adjusted to control the environmental parameters in the corridor closed area to optimize the plant growth conditions; according to the adjusted environmental parameters, the steps of calculating the commodity rate and inventory are repeated until the commodity rate reaches the preset commodity rate threshold and the inventory is lower than the preset inventory threshold, and the automatic identification of corridor closure and harvesting is completed.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images