Data-driven seedling cutting position intelligent recognition and adjustment system

Abstract

This invention discloses a data-driven intelligent recognition and adjustment system for seedling cutting pose, belonging to the field of seedling cultivation automation technology. The system includes modules for multi-source heterogeneous data acquisition, preprocessing and feature fusion, cross-modal pose recognition and decision-making, flexible adaptive execution, and end-to-end data closed-loop and iteration. Seedling physiological data is collected through a microelectrode array and mechanical sensors, combined with visual, 3D, and environmental data. After fusion via an attention mechanism, the optimal cutting pose is identified by a CNN-Transformer+GNN model, and the execution parameters are generated by an improved DDPG model. A magnetorheological flexible joint robotic arm and a force-controlled tool work together with a fuzzy PID algorithm to achieve flexible cutting. Closed-loop optimization is constructed based on edge storage and cloud-based federated learning. This invention achieves collaborative decision-making based on physiology, morphology, and environment, reducing seedling damage, improving post-cutting survival rate, adapting to the needs of multiple seedling varieties, and meeting the requirements of large-scale automated production.

Description

Technical Field

[0001] This invention relates to the field of automated seedling cultivation technology, and in particular to a data-driven intelligent recognition and adjustment system for seedling cutting posture. Background Technology

[0002] In the field of automated cutting operations for seedling cultivation, existing technologies generally suffer from a single data acquisition dimension. Most rely solely on visual images or simple three-dimensional morphological data for pose recognition, lacking the ability to deeply perceive the physiological characteristics of seedlings. They are unable to accurately capture key information such as vascular bundle distribution and lignification boundaries, resulting in recognition that only stays at the surface morphological level. Due to insufficient data support, existing systems often rely on preset fixed parameter libraries to formulate cutting schemes, making it difficult to adapt to individual differences in seedlings with different growth states and physiological characteristics. This often leads to problems such as cutting position deviations and unreasonable angles, resulting in low operational targeting and accuracy, which seriously affects the subsequent growth of seedlings.

[0003] Meanwhile, the actuators of existing cutting equipment mostly adopt a rigid design, with fixed joint stiffness and cutting force, lacking dynamic adaptability. Due to the significant differences in the mechanical properties of seedling stems of different varieties and growth stages, rigid actuators cannot flexibly adjust their working posture and cutting force. During the cutting process, they are prone to damage such as stem splitting and lodging, which not only destroys the smoothness of the cutting surface but also affects the formation of callus tissue in seedlings, resulting in a low survival rate of seedlings after cutting. This has become a key bottleneck restricting the improvement of the quality of automated cutting operations.

[0004] Furthermore, existing systems generally lack a complete data closed-loop mechanism, only completing a single operation process of data collection, decision-making, and execution. They do not correlate the growth feedback data of the seedlings after cutting with the operation parameters. Once the system model parameters are set, they remain fixed and cannot achieve self-optimization through historical operation data. They are always in an experience-driven operation mode. When it is necessary to switch seedling varieties or deal with changes in seedling growth status, the parameters need to be manually readjusted. The operation is cumbersome and the debugging cost is high. The system has poor adaptability and cannot meet the industrialization needs of large-scale, multi-variety seedling cultivation. This seriously limits the operating efficiency and application scope of automated production lines.

[0005] Therefore, this invention proposes a data-driven intelligent recognition and adjustment system for seedling cutting posture. Summary of the Invention

[0006] One objective of this invention is to propose a data-driven intelligent recognition and adjustment system for seedling cutting posture. This invention can achieve collaborative decision-making based on physiological, morphological, and environmental factors through a CNN-Transformer+GNN model, accurately locate vascular bundles and lignification boundaries, and output customized cutting schemes. Relying on magnetorheological joints, force-controlled tools, and fuzzy PID algorithms, it dynamically adapts to the mechanical properties of seedlings to avoid damage and improve survival rates. By constructing a data closed loop through edge storage and cloud-based federated learning, it achieves model self-optimization, adapts to the needs of multiple seedling varieties, and reduces debugging costs, providing an efficient and reliable technical solution for automated production lines.

[0007] According to an embodiment of the present invention, a data-driven intelligent recognition and adjustment system for seedling cutting posture includes a multi-source heterogeneous data acquisition module, a data preprocessing and feature fusion module, a cross-modal posture cognition and decision-making module, a flexible adaptive execution module, and a full-link data closed-loop and iteration module.

[0008] The multi-source heterogeneous data acquisition module simultaneously collects seedling biological characteristics, visual and three-dimensional morphology, and environmental and execution status data. The biological characteristic data includes stem bioelectrical signals and compressive strength data at different depths, providing physiological basis for accurate identification.

[0009] The data preprocessing and feature fusion module assigns feature weights through an attention mechanism and outputs a fused feature vector.

[0010] The cross-modal pose recognition and decision-making module uses a CNN-Transformer+GNN model to identify the optimal cutting pose and combines it with an improved DDPG model to generate execution parameters.

[0011] The flexible adaptive execution module completes cutting through a magnetorheological flexible joint robotic arm and a force-controlled tool, and the dynamic control unit uses a fuzzy PID algorithm to stabilize the cutting force.

[0012] The end-to-end data closed-loop and iterative module collects seedling growth data after cutting and optimizes model parameters through federated learning to improve long-term operational accuracy.

[0013] Furthermore, the multi-source heterogeneous data acquisition module adopts a modular integrated design, including a biometric acquisition unit, a microelectrode array, and a vision and 3D acquisition unit;

[0014] The aforementioned biometric acquisition unit is a contact-type structure, containing a 20-channel microelectrode array and a miniature puncture-type mechanical sensor with adjustable puncture depth;

[0015] The microelectrode array is attached to the base of the stem 5cm away to collect bioelectric signals. The vascular bundles are dense in this area. The micro-puncture mechanical sensor collects compressive strength data at a depth of 0-5mm in 0.1mm increments.

[0016] The vision and 3D acquisition unit is a non-contact multi-view structure. The hyperspectral camera captures images of the blade at a 45° tilt angle. The binocular structured light sensor simultaneously scans from the top and side to acquire 3D point cloud data. The high-speed industrial camera focuses on the cutting area to capture dynamic shaking images.

[0017] The environment and status acquisition units are distributed in a distributed manner. Environmental parameter sensors collect temperature, humidity and gas concentration, tool force feedback sensors are integrated into the tool holder, and robotic arm posture sensors are built into each joint. All units work together through a unified controller, and the data is synchronized by timestamps to ensure spatiotemporal consistency.

[0018] Furthermore, the data preprocessing and feature fusion module is deployed on edge computing nodes. It first removes bioelectrical signal noise using the Isolation Forest algorithm, smooths the mechanical data using Kalman filtering, removes environmental interference points from the 3D point cloud, and then extracts features such as bioelectrical energy entropy, stem mechanical gradient, and chlorophyll distribution. The weight calculation formula for the attention mechanism is:

[0019]

[0020] in, is the weight value of the i-th type of feature, with a value ranging from 0 to 1. The higher the weight value, the greater the influence of this type of feature on the recognition result.

[0021] The importance score for the i-th type of feature is obtained by calculating the correlation between the feature and the cutting pose recognition result using the Pearson correlation coefficient. The higher the correlation, the higher the score.

[0022] The total number of feature categories participating in the fusion, n, is fixed at 3, corresponding to biological features, visual and 3D features, and environmental and state features, respectively.

[0023] This serves as the feature category index, used to iterate through all feature categories involved in the fusion to complete the summation operation;

[0024] After being weighted, the various features are fused into a 128-dimensional fixed-dimensional feature vector through a feature concatenation algorithm. This vector can be directly input into the subsequent cognitive model for inference calculation.

[0025] Furthermore, the cross-modal pose cognition and decision-making module includes a pose recognition and adjustment decision-making sub-module, which interact through a data buffer. In the CNN-Transformer model of the pose recognition sub-module, the CNN part extracts local features such as vascular bundle position and leaf attachment point, while the Transformer part captures global features such as stem bending trend through a self-attention mechanism. The GNN association layer uses the extracted features as nodes to construct a mapping relationship between bioelectric signals and cross-dimensional features such as three-dimensional coordinates, and outputs the coordinates of the cutting center point, angle, and force threshold parameters. The pose recognition sub-module also integrates a dynamic standard library unit, constructs a basic parameter library based on seedling category, and dynamically corrects parameters through a linear interpolation algorithm by combining real-time data such as stem diameter, plant height, and growth rate.

[0026] Furthermore, the improved DDPG model of the adjustment decision submodule introduces a seedling physiological characteristic constraint term, and the reward function takes into account both growth incentives and damage penalties, and satisfies the formula:

[0027]

[0028] in, This is the model's single-step reward value, used to guide the model in learning the optimal decision-making strategy. A higher value indicates a better current decision;

[0029] This is the growth effect reward coefficient, with a value ranging from 1.0 to 1.5. It can be dynamically adjusted according to the growth sensitivity of the seedling variety; the higher the growth sensitivity, the larger the coefficient.

[0030] The growth status of seedlings after cutting is scored, with a value ranging from 0 to 1. The score is obtained by weighted summation of multiple growth feedback data such as callus formation rate, transpiration rate, and root growth activity.

[0031] The damage penalty coefficient is fixed at 2.0, which is higher than the reward coefficient to highlight the importance of damage control.

[0032] The degree of cutting damage to seedlings is scored, with a value range of 0 to 1, and is determined by indicators such as the degree of stem splitting, the damage to the growing point, and the number of vascular bundle breaks.

[0033] Through this reward function, the model can achieve an optimal balance between improving seedling growth and reducing cutting damage.

[0034] Furthermore, the flexible adaptive execution module consists of a 7-DOF collaborative robotic arm, an adaptive cutting tool, and a dynamic control unit. The three components transmit control commands via a CAN bus. The robotic arm has force sensing and collision protection functions. The end effector magnetorheological flexible joint changes stiffness by adjusting the magnetic field strength to adapt to stems with different mechanical properties. The adaptive cutting tool uses a double-edged arc blade, with a built-in force control sensor to collect contact force, and a micro cylinder to adjust the feed pressure. The dynamic control unit is the core of the execution module. The fuzzy PID algorithm combines the advantages of fuzzy control dynamic response and PID steady-state accuracy. It adjusts the cylinder intake and exhaust volume through PWM signals to stabilize the cutting force at the optimal threshold, avoiding damage to the seedlings or incomplete cutting. When the high-speed industrial camera detects that the seedling sway exceeds 2mm, the adjustment decision submodule immediately triggers a pause command, and the adjustment is performed again after the pose data is reacquired.

[0035] Furthermore, the formula for calculating the output of the fuzzy PID algorithm is as follows:

[0036]

[0037] in, The control output at time t is specifically represented by the duty cycle of the PWM signal, and its value directly determines the pressure adjustment range of the cylinder.

[0038] This is a proportionality coefficient, which mainly affects the system's response speed;

[0039] The cutting force deviation value at time t is the difference between the optimal cutting force threshold output by the cross-modal decision module and the actual cutting force collected by the force control sensor, which is the core input of the control algorithm.

[0040] This is the integral time constant, used to eliminate the steady-state error of the system;

[0041] The differential time constant is used to suppress system overshoot and oscillation;

[0042] The proportionality coefficient Integral time constant and differential time constant It is not a fixed value, but rather determined by a pre-defined fuzzy rule base, based on... The magnitude and rate of change are dynamically corrected in real time, so that the system can maintain excellent control performance in different cutting stages.

[0043] Furthermore, the end-to-end data closed-loop and iteration module includes a data storage unit, a growth feedback acquisition unit, and a model iteration unit. The data storage unit adopts an edge and cloud distributed architecture, with edge nodes storing nearly 30 days of real-time operation data and the cloud storing historical data and growth feedback data for a long period of no less than one year. It supports retrieval by seedling type and operation time. The growth feedback acquisition unit collects growth data within 72 hours after cutting using root growth sensors, transpiration rate sensors, and chlorophyll meters. This period is critical for wound healing and has high data feedback value. The model iteration unit, while protecting data privacy, aggregates multi-source data through federated learning and uses a gradient descent algorithm to update the parameters of the cross-modal cognitive and decision-making model, so that the operation accuracy continues to improve with data accumulation.

[0044] Furthermore, the microelectrode array of the biofeedback acquisition unit has an electrode spacing of 1 mm. The acquired bioelectric signals are processed by db4 wavelet base wavelet transform to extract energy entropy features. The feature peak position can locate the vascular bundle distribution area. The compressive strength data acquired by the micro-puncture mechanical sensor is used to generate a change curve through gradient calculation. The point where the slope of the curve changes abruptly is the boundary between the lignified and fleshy regions. Based on this, the optimal cutting depth is determined to ensure that the cutting surface is flat and does not damage the xylem. The hyperspectral images of the visual and three-dimensional acquisition units are segmented into leaf regions using U-Net to extract chlorophyll mean features. The three-dimensional point cloud is fitted to the stem central axis using the RANSAC algorithm to calculate the curvature features. The environmental data is normalized to environmental factors in the 0-1 interval.

[0045] Furthermore, the basic parameter library of the dynamic standard library unit covers common seedling categories such as vegetables, trees and flowers, and the standard parameters include cutting height, angle range and force threshold range.

[0046] The beneficial effects of this invention are:

[0047] 1. This invention captures deep data such as seedling bioelectric signals and stem mechanical gradients through microelectrode arrays and mechanical sensors, changing the superficial perception mode of existing technologies that rely solely on visual data. It achieves precise positioning of seedling vascular bundle distribution and lignification boundaries. Combined with a cross-modal cognitive model of CNN-Transformer+GNN, the system can establish a correlation mapping between physiological characteristics and three-dimensional morphology and environmental factors, realizing collaborative decision-making among physiology, morphology, and environment. This completely eliminates the dependence of traditional systems on fixed parameter libraries, outputting customized cutting schemes for seedlings in different growth states, and improving the targeting and recognition accuracy of operations from the source.

[0048] 2. In this invention, by dynamically adjusting the stiffness of the magnetorheological flexible joint, correcting the pressure of the force-controlled adaptive tool, and combining it with real-time control using a fuzzy PID algorithm, the system can flexibly adjust the execution posture and cutting force according to the mechanical characteristics of the seedling stem. This adapts to the mechanical differences of seedlings of different varieties and growth stages, while avoiding the stem splitting and lodging problems caused by traditional rigid cutting. This combination ensures the flatness of the cutting surface, provides a good foundation for the subsequent callus growth of the seedling, and significantly improves the survival rate of the seedling after cutting, demonstrating outstanding creativity at the execution level.

[0049] 3. Through the edge storage and cloud federated learning architecture, the system can deeply correlate operation data with seedling growth feedback after cutting, forming a virtuous cycle of operation, feedback and model optimization. It can dynamically improve recognition accuracy and decision rationality without human intervention, making it a fundamental shift from experience-driven to data-driven. It can not only adapt to the needs of multiple types of seedlings such as vegetables and trees, but also reduce the debugging cost when switching categories. It provides a more adaptable technical solution for automated production lines, highlighting its creative value in industrial applications. Attached Figure Description

[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0051] Figure 1 This is a schematic diagram of the overall framework structure of a data-driven intelligent recognition and adjustment system for seedling cutting posture proposed in this invention.

[0052] Figure 2 This is a schematic diagram of the operation process of a data-driven intelligent recognition and adjustment system for seedling cutting posture proposed in this invention. Detailed Implementation

[0053] To make the technical means and objectives and effects of the present invention easier to understand, the embodiments of the present invention will be described in detail below with reference to specific illustrations.

[0054] like Figure 1-2As shown, this invention discloses a data-driven intelligent recognition and adjustment system for seedling cutting posture. Supported by multi-dimensional data fusion and cross-modal intelligent decision-making, it aims to solve the problems of low precision, high damage rate, and poor adaptability in existing seedling cutting operations. The system consists of a multi-source heterogeneous data acquisition module, a data preprocessing and feature fusion module, a cross-modal posture recognition and decision-making module, a flexible adaptive execution module, and a full-link data closed-loop and iteration module. Each module establishes a real-time communication link via industrial Ethernet, forming a complete technical chain of acquisition, processing, decision-making, execution, and optimization. It can be directly integrated into existing seedling cultivation production lines and is suitable for batch cutting operations of various types of seedlings such as vegetables, trees, and flowers.

[0055] The system's hardware deployment needs to be modularly integrated with the seedling operation process. The overall layout adopts a circular conveyor line, with data acquisition stations, cutting execution stations, and temporary storage and detection stations set up sequentially along the conveyor line. Each station achieves coordinated linkage through a PLC controller.

[0056] The multi-source heterogeneous data acquisition module adopts a modular integrated design based on the data acquisition station as a carrier, including a biometric acquisition unit, a microelectrode array, and a vision and 3D acquisition unit;

[0057] The biometric acquisition unit is a contact structure equipped with a 20-channel microelectrode array and a micro-puncture mechanical sensor with adjustable puncture depth. The electrode spacing of the microelectrode array is set to 1mm. During acquisition, it is driven by a mechanical gripper to adhere to the base of the seedling stem at 5cm, which is the area with the densest distribution of vascular bundles. The micro-puncture mechanical sensor is installed on the fine-tuning robotic arm of the acquisition station and can collect stem compressive strength data in 0-5mm depths in 0.1mm increments.

[0058] Secondly, the vision and 3D acquisition unit is a non-contact multi-view structure. The hyperspectral camera is fixed at a 45° angle above the acquisition station, with the lens facing the seedling leaf area to obtain chlorophyll distribution images. The binocular structured light sensors are installed on the top and side of the acquisition station to achieve synchronous scanning in the top and side views to generate 3D point cloud data. The high-speed industrial camera focuses on the preset cutting area of ​​the seedling stem and captures dynamic shaking images at a frame rate of 200fps.

[0059] In addition, the environment and status acquisition unit adopts a distributed layout. Temperature and humidity sensors and CO2 concentration sensors are installed on brackets around the acquisition station. Tool force feedback sensors are integrated into the tool holder of the cutting execution station. Robotic arm posture sensors are built into the joints of the robotic arm in the cutting execution station. All acquisition units work together through a unified controller. Data transmission adopts timestamp synchronization technology to ensure the spatiotemporal consistency of multi-dimensional data with errors controlled within ±1ms.

[0060] The flexible adaptive execution module is mounted on a cutting execution station and consists of a 7-DOF collaborative robotic arm, an adaptive tool, and a dynamic control unit. The 7-DOF collaborative robotic arm has force sensing and collision protection functions. A magnetorheological flexible joint is installed at its end flange. By adjusting the magnetic field strength of the magnetorheological fluid built into the joint, it can achieve… Dynamic stiffness adjustment within a range to adapt to seedling stems with different mechanical properties;

[0061] The robotic arm is also equipped with an adaptive tool. The tool features a double-edged arc blade design and is made of food-grade stainless steel to prevent seedling infection. The tool integrates a force control sensor with a range of 0-10N and a micro cylinder. The force control sensor collects the cutting contact force in real time, while the micro cylinder can accurately adjust the cutting pressure according to the control command.

[0062] Secondly, the dynamic control unit, as the core control component of the execution module, is deployed in the control cabinet of the cutting execution station. It uses a fuzzy PID control algorithm to achieve stable control of the cutting force, and its control signal is transmitted to the drive units of the robotic arm and the tool via a CAN bus. The temporary storage and testing station is equipped with a vacuum suction cup fixing mechanism to fix the base of the seedlings during data acquisition and cutting, preventing seedling displacement from affecting the accuracy of the operation.

[0063] The data preprocessing and feature fusion module is deployed on an edge computing node, which uses an NVIDIA Jetson AGX Orin computing board with a computing power of 200 TOPS to meet real-time processing requirements.

[0064] Once the seedlings are transported to the data acquisition station and secured by a vacuum suction cup, the multi-source heterogeneous data acquisition module begins to synchronously acquire data. The raw data is first transmitted to the edge computing node for preprocessing.

[0065] The bioelectrical signal is processed by removing noise points using the isolated forest algorithm, and then the energy entropy feature is extracted after wavelet transform processing using the db4 wavelet basis. The peak position of this feature can directly locate the vascular bundle distribution area of ​​the seedling stem. The compressive strength data collected by the miniature puncture mechanical sensor is smoothed by Kalman filtering, and a mechanical gradient curve is generated by gradient calculation. The position of the slope change in the curve is the boundary between the lignified region and the fleshy region. The three-dimensional point cloud data is processed by the RANSAC algorithm to remove environmental interference points such as the conveyor platform, and then fitted to generate the central axis of the seedling stem to calculate the curvature feature.

[0066] After segmenting the leaf region from the hyperspectral image using the U-Net network, the mean chlorophyll value features within the region are extracted.

[0067] Environmental data are then normalized and converted into environmental factors in the 0-1 range.

[0068] After preprocessing, the system assigns weights to various features using an attention mechanism. The weight calculation formula is as follows:

[0069]

[0070] in, is the weight value of the i-th type of feature, with a value ranging from 0 to 1. The higher the weight value, the greater the influence of this type of feature on the recognition result.

[0071] The importance score for the i-th type of feature is obtained by calculating the correlation between the feature and the cutting pose recognition result using the Pearson correlation coefficient. The higher the correlation, the higher the score.

[0072] The total number of feature categories participating in the fusion, n, is fixed at 3, corresponding to biological features, visual and 3D features, and environmental and state features, respectively.

[0073] This serves as the feature category index, used to iterate through all feature categories involved in the fusion to complete the summation operation;

[0074] After being weighted, the various features are fused into a 128-dimensional fixed-dimensional feature vector through a feature concatenation algorithm. This vector can be directly input into the subsequent cognitive model for inference calculation.

[0075] The cross-modal pose cognition and decision-making module adopts an edge-cloud collaborative architecture, which includes a pose recognition and adjustment decision-making sub-module. The pose recognition and adjustment decision-making sub-module is deployed on edge computing nodes to achieve real-time inference, while the model training and update unit is deployed in the cloud.

[0076] The pose recognition submodule adopts a fusion model structure of CNN-Transformer backbone network and GNN association layer. The CNN part consists of 4 convolutional layers and 2 pooling layers. It extracts local detail features from the fusion features through convolution operation, such as the vascular bundle position of seedling stem and leaf attachment point.

[0077] The Transformer part contains 6 encoder layers, which capture global distribution features in the features through a self-attention mechanism, such as the overall bending trend of the stem and the spatial posture of the seedling.

[0078] The GNN association layer uses the features output by the CNN-Transformer as nodes to construct the mapping relationship between bioelectrical signal features and three-dimensional coordinate features, chlorophyll features and environmental factors, and finally outputs the optimal cutting pose parameters of the seedling, including the spatial coordinates (X,Y,Z) of the cutting center point, the cutting angle and the cutting force threshold F.

[0079] Secondly, the pose recognition submodule also integrates a dynamic standard library unit. This unit builds a basic cutting pose parameter library based on common seedling categories such as vegetables, trees, and flowers. The library contains standard cutting heights, angle ranges, and force threshold ranges for various seedlings. When real-time growth data of the seedlings, such as stem diameter, plant height, and growth rate, is obtained, the parameters in the library are dynamically corrected through a linear interpolation algorithm to make the standard parameters more consistent with the actual growth state of a single seedling.

[0080] Secondly, the decision-making submodule adopts an improved Deep Deterministic Policy Gradient (DDPG) model. This model introduces a seedling physiological characteristic constraint term into the traditional DDPG model. Specifically, when the vascular bundle activity reflected by the bioelectrical signal exceeds a preset threshold, the model automatically reduces the upper limit of the cutting force output to avoid damaging the seedling growth point. The model's reward function is designed to balance growth incentives with damage risk penalties, satisfying the formula:

[0081]

[0082] in, This is the model's single-step reward value, used to guide the model in learning the optimal decision-making strategy. A higher value indicates a better current decision;

[0083] This is the growth effect reward coefficient, with a value ranging from 1.0 to 1.5. It can be dynamically adjusted according to the growth sensitivity of the seedling variety; the higher the growth sensitivity, the larger the coefficient.

[0084] The growth status of seedlings after cutting is scored, with a value ranging from 0 to 1. The score is obtained by weighted summation of multiple growth feedback data such as callus formation rate, transpiration rate, and root growth activity.

[0085] The damage penalty coefficient is fixed at 2.0, which is higher than the reward coefficient to highlight the importance of damage control.

[0086] The degree of cutting damage to seedlings is scored, with a value range of 0 to 1, and is determined by indicators such as the degree of stem splitting, the damage to the growing point, and the number of vascular bundle breaks.

[0087] The adjustment decision submodule outputs the adjustment parameters of the robotic arm and the control parameters of the cutting tool based on the pose recognition results and real-time acquired data. When the high-speed industrial camera detects that the dynamic shaking amplitude of the seedling exceeds the preset threshold of 2mm, a pause command is immediately triggered to control the robotic arm to stop moving. After the vision and 3D acquisition unit re-acquires and updates the seedling pose data, the adjustment strategy is generated and executed based on the new data.

[0088] It is important to note that the flexible adaptive execution module executes the cutting operation based on the parameters output by the adjustment decision submodule. The dynamic control unit uses a fuzzy PID algorithm that combines the dynamic response advantages of fuzzy control with the steady-state accuracy advantages of PID control. Its output calculation formula is as follows:

[0089]

[0090] in, The control output at time t is specifically represented by the duty cycle of the PWM signal, and its value directly determines the pressure adjustment range of the cylinder.

[0091] This is a proportionality coefficient, which mainly affects the system's response speed;

[0092] The cutting force deviation value at time t is the difference between the optimal cutting force threshold output by the cross-modal decision module and the actual cutting force collected by the force control sensor, which is the core input of the control algorithm.

[0093] This is the integral time constant, used to eliminate the steady-state error of the system;

[0094] The differential time constant is used to suppress system overshoot and oscillation;

[0095] proportionality coefficient Integral time constant and differential time constant It is not a fixed value, but rather determined by a pre-defined fuzzy rule base, based on... The magnitude and rate of change are dynamically corrected in real time, so that the system can maintain excellent control performance in different cutting stages.

[0096] For example when and When, increase To speed up the response time, when When, increase To eliminate steady-state error.

[0097] After receiving control commands, the robotic arm adjusts its stiffness through magnetorheological flexible joints. For example, for seedlings with stem diameters <2mm, the joint stiffness is reduced to... To avoid rigid contact that could cause seedlings to fall over, the adaptive tool adjusts the feed pressure under the drive of a micro cylinder to ensure that the cutting force remains stable within the optimal threshold range.

[0098] The end-to-end data closed-loop and iteration module includes data storage, growth feedback collection, and model iteration units, realizing a virtuous cycle of data accumulation, effect feedback, and model optimization.

[0099] The data storage unit adopts a distributed database architecture combining edge and cloud. The local database of the edge node stores nearly 30 days of real-time operation data, including raw data, feature data, decision parameters and execution data. The cloud database receives this data synchronously through the 5G network and stores historical operation data and growth feedback data for a long time. The data retention period is no less than 1 year, and it supports fast retrieval by seedling type, operation time, growth effect and other dimensions.

[0100] Secondly, the growth feedback acquisition unit consists of a root growth sensor, a transpiration rate sensor, and a chlorophyll meter. These sensors are deployed in the seedling environment after the seedling is cut to collect key growth data within 72 hours after the cutting. This period is the core stage of seedling wound healing, and the data has extremely high feedback value.

[0101] In addition, the model iteration unit adopts a federated learning approach, which aggregates multi-source historical operation data and growth feedback data while protecting the data privacy of each production end. It updates the parameters of the cross-modal pose cognition model and adjusts the decision model through the gradient descent algorithm. After each model update, the system's operation accuracy can be improved by more than 3%, achieving a self-optimization effect that becomes more accurate with use.

[0102] Taking the cutting operation of cucumber seedlings as an example, the complete workflow of the system is as follows:

[0103] First, cucumber seedlings with a height of 15cm and a stem diameter of 3mm are brought into the data collection station via a circular conveyor line. The vacuum suction cup of the temporary storage and testing station immediately activates to fix the seedlings at the base, ensuring that the seedlings do not shake during the collection process.

[0104] Subsequently, the microelectrode array of the biometric acquisition unit is attached to the base of the stem 5cm away to begin collecting bioelectrical signals. A miniature puncture-type mechanical sensor punctures the stem surface in 0.1mm increments to collect compressive strength data at a depth of 0-5mm. Simultaneously, the hyperspectral camera, binocular structured light sensor, and high-speed industrial camera of the vision-3D acquisition unit are activated to acquire leaf chlorophyll images, seedling 3D point cloud data, and stem dynamic images, respectively. Environmental parameter sensors collect data on the current operating environment, including temperature (25℃), humidity (65%), and other parameters. The concentration data was 500 ppm. All collected data were transmitted to the edge computing node after being timestamped.

[0105] The edge computing node preprocesses the received data. After the bioelectric signal is denoised by the isolated forest algorithm, the energy entropy feature is extracted by the db4 wavelet transform to determine that the vascular bundle distribution area is within 0.8 mm of the stem center.

[0106] After the compressive strength data was smoothed by Kalman filtering, the mechanical gradient curve was calculated, and the boundary between the lignified region and the fleshy region was determined to be located at a stem depth of 1.2 mm.

[0107] After denoising the 3D point cloud data, the central axis of the stem was fitted and calculated. The hyperspectral images were segmented using U-Net, and the mean chlorophyll content of the leaves was extracted to be 48 SPAD. Environmental data were normalized to obtain environmental factors, including a temperature factor of 0.6 and a humidity factor of 0.7. Concentration factor 0.5;

[0108] Next, the system calculates the weights of various features through an attention mechanism, with biological features having a weight of 0.4, visual and 3D features having a weight of 0.35, and environmental-state features having a weight of 0.25. After weighted fusion, a 128-dimensional feature vector is output.

[0109] The cross-modal pose recognition model performs inference calculations on this feature vector, and combines the parameters corrected based on cucumber seedling category by the dynamic standard library unit to output the optimal cutting pose parameters:

[0110] Cutting center point coordinates (X=200mm, Y=150mm, Z=50mm), cutting angle =30° =0°, cutting force threshold F=1.5N;

[0111] The improved DDPG model in the decision-making submodule is adjusted by inputting this parameter and real-time environmental data to generate the adjustment parameters for execution.

[0112] The joint rotation angles of the robotic arm are J1=10°, J2=30°, J3=45°, J4=0°, J5=20°, J6=15°, and J7=5°, respectively. The moving speed is 50 mm / s, and the cutting speed of the tool is 100 mm / s. The initial parameters of the fuzzy PID algorithm are set to... =5、 =0.1、 =0.05.

[0113] After receiving the adjustment parameters, the 7-DOF collaborative robotic arm of the cutting execution station immediately initiates its action, adjusting its stiffness through magnetorheological flexible joints. Then, the adaptive tool is moved to above the cutting center point, and the tool angle is adjusted to... =30° =0°;

[0114] When the cutting tool contacts the cucumber seedling stem, the force control sensor collects cutting force data in real time and feeds it back to the dynamic control unit. The dynamic control unit calculates the control output using a fuzzy PID algorithm. Adjust the air intake of the micro cylinder to stabilize the actual cutting force at around 1.5N;

[0115] During the cutting process, the high-speed industrial camera continuously monitored the seedling's shaking amplitude, which remained within 1mm. No pause command was triggered, and the cutting tool successfully completed the cutting operation, resulting in a smooth cutting surface without damaging the vascular bundles and xylem.

[0116] After cutting, the seedlings entered the seedling area via the conveyor line, and the growth feedback acquisition unit started working. After 72 hours, the callus formation rate of the seedlings was monitored to be 0.6 mm / h, and the transpiration rate was [missing information]. The root growth vitality was 0.8, and the growth status score was calculated. =0.9, damage severity score =0, calculated according to the reward function formula. =1.2×0.9 - 2.0×0=1.08, the data from this operation was marked as "high-quality sample" and synchronized to the cloud database;

[0117] During non-operational hours at night, the model iteration unit updates the parameters of the cross-modal cognitive model and decision-making model based on the high-quality sample and other historical data. The updated model reduces the pose recognition error from 0.1mm to 0.08mm in the cucumber seedling cutting operation the next day, further improving the operation accuracy.

[0118] It is worth mentioning that the system requires regular maintenance and calibration once a month during daily use. Specifically, the microelectrode array and micro-puncture mechanical sensor of the biometric acquisition unit are calibrated using standard test blocks to ensure acquisition accuracy. The binocular structured light sensor is calibrated to correct the ranging error of the three-dimensional point cloud data. The motion accuracy of each joint of the robotic arm is checked. If deviations occur, they are compensated and adjusted by the controller.

[0119] When it is necessary to switch seedling varieties, simply select the target variety in the system's human-computer interaction interface, and the dynamic standard library unit will automatically load the corresponding basic parameter library. There is no need to perform complex parameter resets, which greatly improves the system's ease of operation and variety adaptability.

[0120] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data-driven intelligent recognition and adjustment system for seedling cutting posture, characterized in that, This includes a multi-source heterogeneous data acquisition module, a data preprocessing and feature fusion module, a cross-modal pose cognition and decision-making module, a flexible adaptive execution module, and a full-link data closed-loop and iteration module; The multi-source heterogeneous data acquisition module simultaneously collects seedling biological characteristics, visual and three-dimensional morphology, and environmental and execution status data. The biological characteristic data includes stem bioelectrical signals and compressive strength data at different depths, providing physiological basis for accurate identification. The data preprocessing and feature fusion module assigns feature weights through an attention mechanism and outputs a fused feature vector. The cross-modal pose recognition and decision-making module uses a CNN-Transformer+GNN model to identify the optimal cutting pose and combines it with an improved DDPG model to generate execution parameters. The flexible adaptive execution module completes cutting through a magnetorheological flexible joint robotic arm and a force-controlled tool, and the dynamic control unit uses a fuzzy PID algorithm to stabilize the cutting force. The end-to-end data closed-loop and iterative module collects seedling growth data after cutting and optimizes model parameters through federated learning to improve long-term operational accuracy.

2. The data-driven intelligent recognition and adjustment system for seedling cutting posture according to claim 1, characterized in that, The multi-source heterogeneous data acquisition module adopts a modular integrated design, including a biometric acquisition unit, a microelectrode array, and a vision and 3D acquisition unit; The aforementioned biometric acquisition unit is a contact-type structure, containing a 20-channel microelectrode array and a miniature puncture-type mechanical sensor with adjustable puncture depth; The microelectrode array is attached to the base of the stem 5cm away to collect bioelectric signals. The vascular bundles are dense in this area. The micro-puncture mechanical sensor collects compressive strength data at a depth of 0-5mm in 0.1mm increments. The vision and 3D acquisition unit is a non-contact multi-view structure. The hyperspectral camera captures images of the blade at a 45° tilt angle. The binocular structured light sensor simultaneously scans from the top and side to acquire 3D point cloud data. The high-speed industrial camera focuses on the cutting area to capture dynamic shaking images. The environment and status acquisition units are distributed in a distributed manner. Environmental parameter sensors collect temperature, humidity and gas concentration, tool force feedback sensors are integrated into the tool holder, and robotic arm posture sensors are built into each joint. All units work together through a unified controller, and the data is synchronized by timestamps to ensure spatiotemporal consistency.

3. The data-driven intelligent recognition and adjustment system for seedling cutting posture according to claim 1, characterized in that, The data preprocessing and feature fusion module is deployed on edge computing nodes. It first removes bioelectrical signal noise using the isolated forest algorithm, smooths the mechanical data using Kalman filtering, removes environmental interference points from the 3D point cloud, and then extracts features such as bioelectrical energy entropy, stem mechanical gradient, and chlorophyll distribution. The weight calculation formula for the attention mechanism is as follows: in, is the weight value of the i-th type of feature, with a value ranging from 0 to 1. The higher the weight value, the greater the influence of this type of feature on the recognition result. The importance score for the i-th type of feature is obtained by calculating the correlation between the feature and the cutting pose recognition result using the Pearson correlation coefficient. The higher the correlation, the higher the score. The total number of feature categories participating in the fusion, n, is fixed at 3, corresponding to biological features, visual and 3D features, and environmental and state features, respectively. This serves as the feature category index, used to iterate through all feature categories involved in the fusion to complete the summation operation; After being weighted, the various features are fused into a 128-dimensional fixed-dimensional feature vector through a feature concatenation algorithm. This vector can be directly input into the subsequent cognitive model for inference calculation.

4. The data-driven intelligent recognition and adjustment system for seedling cutting posture according to claim 1, characterized in that, The cross-modal pose cognition and decision-making module includes a pose recognition and adjustment decision-making sub-module, which interact through a data buffer. In the CNN-Transformer model of the pose recognition sub-module, the CNN part extracts local features such as vascular bundle position and leaf attachment point, while the Transformer part captures global features such as stem bending trend through a self-attention mechanism. The GNN association layer uses the extracted features as nodes to construct a mapping relationship between bioelectrical signals and cross-dimensional features such as three-dimensional coordinates, and outputs the coordinates of the cutting center point, angle, and force threshold parameters. The pose recognition sub-module also integrates a dynamic standard library unit, constructs a basic parameter library based on seedling category, and dynamically corrects parameters through a linear interpolation algorithm by combining real-time data such as stem diameter, plant height, and growth rate.

5. The data-driven intelligent recognition and adjustment system for seedling cutting posture according to claim 1, characterized in that, The improved DDPG model of the adjustment decision submodule introduces a seedling physiological characteristic constraint term, and the reward function takes into account both growth incentives and damage penalties, and satisfies the formula: in, This is the model's single-step reward value, used to guide the model in learning the optimal decision-making strategy. A higher value indicates a better current decision; This is the growth effect reward coefficient, with a value ranging from 1.0 to 1.

5. It can be dynamically adjusted according to the growth sensitivity of the seedling variety; the higher the growth sensitivity, the larger the coefficient. The growth status of seedlings after cutting is scored, with a value ranging from 0 to 1. The score is obtained by weighted summation of multiple growth feedback data such as callus formation rate, transpiration rate, and root growth activity. The damage penalty coefficient is fixed at 2.0, which is higher than the reward coefficient to highlight the importance of damage control. The degree of cutting damage to seedlings is scored, with a value range of 0 to 1, and is determined by indicators such as the degree of stem splitting, the damage to the growing point, and the number of vascular bundle breaks. Through this reward function, the model can achieve an optimal balance between improving seedling growth and reducing cutting damage.

6. The data-driven intelligent recognition and adjustment system for seedling cutting posture according to claim 1, characterized in that, The flexible adaptive execution module consists of a 7-DOF collaborative robotic arm, an adaptive cutting tool, and a dynamic control unit. The three components transmit control commands via a CAN bus. The robotic arm has force sensing and collision protection functions. The magnetorheological flexible joint at the end adjusts the stiffness by regulating the magnetic field strength to adapt to stems with different mechanical properties. The adaptive cutting tool uses a double-edged arc blade, with a built-in force control sensor to collect contact force, and a micro cylinder to adjust the feed pressure. The dynamic control unit is the core of the execution module. It combines the advantages of fuzzy PID algorithm with the dynamic response of fuzzy control and the steady-state accuracy of PID. It adjusts the intake and exhaust volume of the cylinder through PWM signal to stabilize the cutting force at the optimal threshold, avoiding damage to the seedlings or incomplete cutting. When the high-speed industrial camera detects that the seedling sway exceeds 2mm, the adjustment decision submodule immediately triggers a pause command, and the adjustment is performed again after the pose data is reacquired.

7. The data-driven intelligent recognition and adjustment system for seedling cutting posture according to claim 6, characterized in that, The formula for calculating the output of the fuzzy PID algorithm is as follows: in, The control output at time t is specifically represented by the duty cycle of the PWM signal, and its value directly determines the pressure adjustment range of the cylinder. This is a proportionality coefficient, which mainly affects the system's response speed; The cutting force deviation value at time t is the difference between the optimal cutting force threshold output by the cross-modal decision module and the actual cutting force collected by the force control sensor, which is the core input of the control algorithm. This is the integral time constant, used to eliminate the steady-state error of the system; The differential time constant is used to suppress system overshoot and oscillation; The proportionality coefficient Integral time constant and differential time constant It is not a fixed value, but rather determined by a pre-defined fuzzy rule base, based on... The magnitude and rate of change are dynamically corrected in real time, so that the system can maintain excellent control performance in different cutting stages.

8. The data-driven intelligent recognition and adjustment system for seedling cutting posture according to claim 1, characterized in that, The end-to-end data closed-loop and iteration module includes a data storage unit, a growth feedback acquisition unit, and a model iteration unit. The data storage unit adopts an edge and cloud distributed architecture, with edge nodes storing nearly 30 days of real-time operation data and the cloud storing historical data and growth feedback data for a long period of no less than one year. It supports retrieval by seedling type and operation time. The growth feedback acquisition unit collects growth data within 72 hours after cutting using root growth sensors, transpiration rate sensors, and chlorophyll meters. This period is critical for wound healing and has high data feedback value. The model iteration unit, while protecting data privacy, aggregates multi-source data through federated learning and uses a gradient descent algorithm to update the parameters of the cross-modal cognitive and decision-making model, so that the operation accuracy continues to improve with data accumulation.

9. The data-driven intelligent recognition and adjustment system for seedling cutting posture according to claim 2, characterized in that, The microelectrode array of the biofeedback acquisition unit has an electrode spacing of 1 mm. The acquired bioelectric signals are processed by db4 wavelet basis wavelet transform to extract energy entropy features. The feature peak position can locate the vascular bundle distribution area. The compressive strength data acquired by the micro-puncture mechanical sensor is used to generate a change curve through gradient calculation. The point where the slope of the curve changes abruptly is the boundary between the lignified and fleshy regions. Based on this, the optimal cutting depth is determined to ensure that the cutting surface is flat and does not damage the xylem. The hyperspectral images of the vision and three-dimensional acquisition unit are segmented into leaf regions by U-Net to extract the chlorophyll mean features. The three-dimensional point cloud is fitted to the stem central axis by the RANSAC algorithm to calculate the curvature features. The environmental data is normalized to environmental factors in the 0-1 interval.

10. The data-driven intelligent recognition and adjustment system for seedling cutting posture according to claim 4, characterized in that, The basic parameter library of the dynamic standard library unit covers common seedling categories such as vegetables, trees and flowers, and the standard parameters include cutting height, angle range and force threshold range.

Images