Tomato maturity grading and picking integrated machine based on deep learning

Abstract

The application provides a tomato maturity grading and picking integrated machine based on deep learning, and belongs to the technical field of agricultural machinery automation. The machine comprises a vehicle body, a picking assembly installed on the top of the vehicle body, a storage mechanism arranged on the top of the vehicle body, a conveying mechanism arranged on the outer side of the storage mechanism, a control system arranged in the vehicle body, and a visual processing module included in the control system. The visual processing module runs a deep learning model to obtain the spatial coordinates and maturity characteristics of target tomatoes. The control system controls the picking assembly to pick tomatoes according to the spatial coordinates, and controls the conveying mechanism and the storage mechanism to perform damping conveying and grading collection according to the maturity characteristics. The application extracts continuous characteristic variables of tomatoes through a visual processing module, and maps the continuous characteristic variables into target air pressure values by a main control module to control a miniature air pump to inflate an inner air bag in a conveying pipe.

Description

Technical Field

[0001] This invention relates to the field of agricultural machinery automation technology, and in particular to a tomato ripeness grading and harvesting integrated machine based on deep learning. Background Technology

[0002] With the development of agricultural automation, agricultural harvesting robots are widely used in the harvesting of fruits and vegetables such as tomatoes. Existing tomato harvesting equipment typically uses a robotic arm in conjunction with an end effector for target identification and grasping. In the conventional workflow, after the robotic arm grasps the target fruit and separates it from the stem, it needs to carry the fruit back to the storage area on the machine for release, and then move to the next target location. This back-and-forth transport mode consumes a significant amount of machine operating time, resulting in a long single harvesting cycle and low overall continuous operation efficiency when dealing with densely fruited tomato plants.

[0003] To reduce the unnecessary back-and-forth travel of the robotic arm, some harvesting equipment attempts to connect a collection hose directly below the arm, allowing the fruit to slide directly into the collection box by gravity after separation. However, tomatoes are berries, and their flesh firmness decreases as they ripen. When falling directly by gravity or sliding through a pipe with no indiscriminate friction, the fruit's falling speed is difficult to control, making it prone to collisions with the pipe's interior or the bottom of the collection box. Especially when harvesting highly ripe tomatoes, this unguided transport method can cause skin damage or internal bruising, reducing the fruit's commercial value.

[0004] Furthermore, tomatoes mature at different times on the same plant, and fruits collected in a single harvest typically contain different maturity grades. Most current harvesting equipment only has a mixed collection function; harvested tomatoes are stored uniformly in the same container and need to be transported to a processing workshop for secondary grading via a specialized sorting line. Integrating a traditional optical sorting line directly into existing harvesting vehicles would significantly increase the physical size and manufacturing cost of the equipment, resulting in a bulky structure that is difficult to adapt to the narrow furrows of greenhouses. Current equipment struggles to simultaneously complete the lossless transport and maturity grading of fruits without introducing a large, independent sorting mechanism. Therefore, this application provides a deep learning-based integrated tomato maturity grading and harvesting machine to meet this need. Summary of the Invention

[0005] The purpose of this invention is to provide a tomato ripeness grading and harvesting integrated machine based on deep learning to solve the above-mentioned problems. It extracts ripeness features through a visual model to dynamically adjust the air pressure damping in the conveying pipe, and combines in-situ release with a synchronous track-changing mechanism based on slip time prediction. This solves the problems mentioned in the background technology, such as low efficiency of robotic arm reciprocating transport, lack of targeted buffering leading to easy damage to the fruit, and difficulty in simultaneously completing compact classification and collection during the harvesting and transfer process.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A deep learning-based tomato ripeness grading and harvesting integrated machine includes a vehicle body, a harvesting component installed on the top of the vehicle body, a storage mechanism installed on the top of the vehicle body, a conveying mechanism installed on the outside of the storage mechanism, and a control system installed inside the vehicle body. The harvesting assembly is mounted on the vehicle body and includes a base and a robotic arm; the storage mechanism is located at the rear of the vehicle body. The conveying mechanism is located between the harvesting component and the storage mechanism, and includes a micro air pump and a conveying pipe. The micro air pump is connected to the conveying pipe, and one end of the conveying pipe is connected to a funnel, while the other end is connected to a diversion pipe. The control system includes a vision processing module, which runs a deep learning model to obtain the spatial coordinates and maturity characteristics of the target tomatoes. The control system controls the picking component to pick the tomatoes based on the spatial coordinates, and controls the conveying mechanism and the storage mechanism to perform damped conveying and graded collection operations based on the maturity characteristics.

[0007] Optionally, the delivery pipe has a double-layer structure, the delivery pipe includes an outer air bladder, an inner air bladder is provided inside the outer air bladder, and a buffer ring is provided on the inner wall of the inner air bladder.

[0008] Optionally, the storage mechanism includes a housing, a top cover movably connected to the top of the housing, multiple handles mounted on the top of the top cover, a pivot rotatably connected to the rear of the housing, a tailgate fixedly connected to the outer side of the pivot, multiple telescopic rods installed at the port of the housing, a stop post fixedly connected to the top of each telescopic rod, and multiple inclined plates provided inside the housing.

[0009] Optionally, a flexible claw is rotatably connected to the end of the robotic arm, and a camera is mounted on the top of the flexible claw; a motor is provided at the front end of the vehicle body.

[0010] Optionally, the front end of the housing has multiple openings, and the diversion pipe communicates with the interior of the housing through the openings.

[0011] Optionally, the control system further includes a main control module, a motion control module, and a pneumatic adjustment module; the main control module is electrically connected to the vision processing module, the motion control module, and the pneumatic adjustment module, respectively.

[0012] Optionally, the visual processing module runs the deep learning model to extract the three-dimensional spatial coordinates of the target tomato and outputs continuous feature variables corresponding to the current maturity of the target tomato as the maturity feature; The visual processing module multiplies the confidence probability of the target tomato belonging to each maturity category predicted by the deep learning model with the feature mapping weights corresponding to each maturity category, and calculates the sum of all products, using the sum as the continuous feature variable. The feature mapping weights are obtained by taking the reciprocal of the average critical pressure data for tomato skin rupture under each maturity category and then normalizing the mapping.

[0013] Optionally, the main control module is configured with air pressure mapping logic. The main control module generates a target air pressure value based on the continuous feature variable and sends the target air pressure value to the pneumatic adjustment module. The target air pressure value increases as the continuous feature variable increases. The pneumatic adjustment module drives the micro air pump to inflate the delivery pipe according to the target air pressure value, thereby changing the positive pressure applied by the delivery pipe to the surface of the target tomato and converting the positive pressure into a sliding friction force that prevents the target tomato from sliding down the delivery pipe.

[0014] Optionally, after the motion control module controls the robotic arm and the flexible claw to complete the separation action of the fruit stem of the target tomato, it maintains the current spatial extension posture of the robotic arm unchanged and controls the flexible claw to perform an opening action in place, so that the target tomato is released from the gripper and falls into the funnel by gravity. The main control module is equipped with parallel control logic. During the period when the motion control module drives the robotic arm to move from the current position to the next gripping point, the main control module controls the pneumatic adjustment module in parallel to perform air pressure damping conveying operation on the target tomato that falls into the conveying pipe.

[0015] Optionally, the main control module records the system time as the start time at the instant the target tomato falls into the funnel, estimates the time it takes for the target tomato to slide down based on the target air pressure value and the physical lengths of the delivery pipe and the diversion pipe, and adds the sliding time to the start time to generate the time node when the target tomato reaches the outlet of the diversion pipe. When the current system running time reaches the specified time node, the main control module outputs an execution command to the storage mechanism, causing the storage mechanism to block the sliding channel of the diversion tube, causing the target tomato to deflect its trajectory and enter the corresponding classification area inside the storage mechanism.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects: 1. This invention extracts continuous characteristic variables of tomatoes through a visual processing module, and the main control module maps these variables to a target air pressure value, controlling a micro-pump to inflate the inner air bladder inside the delivery pipe. This technical feature dynamically changes the positive pressure applied to the tomato surface by the inner wall of the delivery pipe according to the different ripeness levels of the tomatoes, thus converting it into an adjustable sliding friction force. The control system provides greater frictional damping for highly ripe tomatoes with softer flesh to reduce their falling speed, and provides less damping for less ripe tomatoes, avoiding skin damage caused by collisions when the tomatoes fall through the pipe and into the storage mechanism, achieving adaptive buffered delivery.

[0017] 2. This invention decouples the picking and conveying actions by incorporating a moving funnel and conveying pipe beneath the robotic arm, along with the parallel control logic of the main control module. After separating the fruit stem, the robotic arm directly opens its flexible claws in place, allowing the tomato to fall into the funnel under gravity, and then moves directly to the next grasping point to retrieve the fruit. Simultaneously, the main control module controls the pneumatic mechanism in parallel to perform damped conveying of the tomato that has fallen into the pipe. This in-situ release mechanism eliminates the physical travel of the robotic arm to and from the machine body to store the fruit, significantly shortening the mechanical operation time for a single picking and improving the continuous operation efficiency of the entire machine.

[0018] 3. This invention utilizes the main control module to record the starting time of the fruit falling into the funnel, and combines the target air pressure value and the physical length of the conveying pipeline to estimate the time it takes for the tomato to slide down, thereby accurately calculating the time node when the target tomato reaches the outlet of the diversion pipe. When the system clock reaches this time node, the main control module triggers the telescopic rod corresponding to the storage mechanism to raise the baffle, physically blocking the sliding tomato and causing it to deflect its trajectory, rolling it along the inclined plate into the corresponding ripeness classification area. This dynamic trajectory-changing mechanism based on the prediction of the slide time eliminates the need for secondary visual recognition or a separate sorting line at the collection end, and completes the ripeness grading simultaneously during the physical transfer of the fruit.

[0019] The above-mentioned solution is a tomato ripeness grading and harvesting integrated machine based on deep learning provided in this application. Attached Figure Description

[0020] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate embodiments of the invention and, together with the specification, further serve to explain the principles of the invention and enable those skilled in the art to practice and use the invention.

[0021] Figure 1 This is a perspective view of the present invention; Figure 2 This is a schematic diagram of the internal structure of the conveying mechanism of the present invention; Figure 3 This is a schematic diagram of the internal structure of the storage mechanism of the present invention; Figure 4 For the present invention Figure 3 Enlarged view of point A in the middle; Figure 5 This is a cross-sectional view of the external airbag of the present invention; Figure 6 This is a flowchart illustrating the working process of the present invention.

[0022] Figure label: 1. Body; 2. Wheel; 3. Harvesting component; 31. Base; 32. Robotic arm; 33. Flexible claw; 34. Camera; 35. Motor; 4. Conveying mechanism; 41. Miniature air pump; 42. Conveying pipe; 421. External airbag; 422. Internal airbag; 423. Buffer ring; 43. Funnel; 44. Fixing plate; 45. Diverter pipe; 5. Storage mechanism; 51. Shell; 52. Top cover; 53. Handle; 54. Shaft; 55. Tailgate; 56. Telescopic rod; 57. Stop post; 58. Sloping plate.

[0023] As shown in the figure, specific structures and devices are marked in the figure to clearly illustrate the structure of the embodiments of the present invention. However, this is only for illustrative purposes and is not intended to limit the present invention to this specific structure, device and environment. Those skilled in the art can adjust or modify these devices and environments according to specific needs. Detailed Implementation

[0024] The following is a detailed description of a deep learning-based tomato ripeness grading and harvesting integrated machine provided by the present invention, with reference to the accompanying drawings and specific embodiments. It should be noted that, to make the embodiments more detailed, the following embodiments are the best and preferred embodiments; those skilled in the art can also use other alternative methods to implement some known technologies; and the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0026] See attached document Figure 1 To be continued Figure 5 The present invention provides a tomato ripeness grading and harvesting integrated machine based on deep learning, which may include a body 1, wheels 2, harvesting components 3, conveying mechanism 4 and storage mechanism 5.

[0027] The harvesting machine also includes a control system installed inside the vehicle body 1. The control system specifically comprises a main control module, a vision processing module, a motion control module, and a pneumatic adjustment module. The main control module is electrically connected to the vision processing module, motion control module, and pneumatic adjustment module, respectively, to achieve data interaction and timing synchronization control of the entire machine.

[0028] Multiple wheels 2 are installed at the bottom of the vehicle body 1, and a motor 35 for driving the wheels 2 is installed at the front end of the vehicle body 1. A harvesting assembly 3 is fixedly installed on the top plane of the vehicle body 1, and a storage mechanism 5 is installed at the rear of the vehicle body 1. A conveying mechanism 4 is installed between the harvesting assembly 3 and the storage mechanism 5, establishing a physical transmission channel from the end of the harvesting process to the rear end of the multi-compartment storage.

[0029] The storage mechanism 5 includes a housing 51, a top cover 52 movably connected to the top of the housing 51, a plurality of handles 53 mounted on the top of the top cover 52, a rotating shaft 54 ​​rotatably connected to the tail of the housing 51, a tailgate 55 fixedly connected to the outside of the rotating shaft 54, a plurality of telescopic rods 56 installed at the port of the housing 51, a stop post 57 fixedly connected to the top of each telescopic rod 56, and a plurality of inclined plates 58 provided inside the housing 51.

[0030] See attached document Figure 6 This invention achieves tomato harvesting, buffer conveying, and track-changing grading operations through the coordinated operation of mechanical structure and control system. The overall coordinated workflow of the machine consists of the following steps: S1, Visual Data Acquisition and Feature Extraction. During the movement of the vehicle body 1, a camera 34 mounted on top of the flexible claw 33 acquires image data of the tomato plant and transmits it to the visual processing module. The visual processing module runs a preset deep learning model to process the image, extracting the three-dimensional spatial coordinates of the target tomato, and simultaneously outputting continuous feature variables corresponding to the current tomato maturity. The coordinate data and continuous feature variables are then packaged and sent to the main control module.

[0031] S2, In-situ picking and gravity-induced fruit drop trigger. The main control module receives the three-dimensional spatial coordinates and sends pose control commands to the motion control module. The motion control module drives the base 31 to rotate and the robotic arm 32 to extend, causing the flexible claw 33 at the end of the robotic arm 32 to reach the target coordinates and close to grasp the tomato. After the picking action is completed, the flexible claw 33 opens in place to release the tomato. The tomato detaches from the flexible claw 33 by gravity and falls directly into the funnel 43 fixed below the robotic arm 32, entering the internal channel of the conveying mechanism 4.

[0032] S3, Feedforward Variable Damping Conveying. The main control module reads continuous characteristic variables, generates a target pressure signal through a built-in mapping algorithm, and sends this signal to the pneumatic adjustment module. The pneumatic adjustment module drives the micro air pump 41 to operate according to the target pressure signal. The micro air pump 41 fills the double-layer conveying pipe 42 with gas at a set pressure through a flexible tube, changing the degree of expansion of the inner wall of the conveying pipe 42. When the tomato slides down inside the conveying pipe 42, it rubs against the expanded pipe wall, achieving deceleration and buffering to match the hardness of the tomato skin through physical damping.

[0033] S4, Spatiotemporal Synchronous Interception and Dynamic Trajectory Change. Based on the target pressure signal in S3 and the physical length of the delivery pipe 42, the main control module estimates the actual time it takes for the tomato to slide down within the delivery pipe 42. The main control module calculates the trigger timestamp based on this sliding time and outputs an execution command to the storage mechanism 5 at that timestamp. The corresponding telescopic rod 56 located at the port of the housing 51 rises, causing the top baffle 57 to block the sliding channel of the diversion pipe 45. After sliding out of the diversion pipe 45 and hitting the baffle 57, the tomato's trajectory deflects, and it rolls along a specific inclined plate 58 into the corresponding classification area inside the housing 51.

[0034] This invention provides a deep learning-based tomato ripeness grading and harvesting integrated machine, whose visual data acquisition and processing process may include the following steps: S101, Real-time image sequence acquisition. While the motor 35 drives the vehicle body 1 forward, the camera 34 mounted on top of the flexible claw 33 continuously acquires images of the work scene. To support spatial positioning, the camera 34 specifically adopts an RGB-D depth camera, simultaneously acquiring a two-dimensional color image stream and a matching depth map sequence.

[0035] S102, Image preprocessing. Camera 34 transmits its generated raw image data to the vision processing module in the control system via the data bus. The vision processing module performs de-mosaicing, color space conversion, and white balance calibration on the incoming raw data, outputting a standard format RGB digital image. For hardware-level data processing such as camera low-level signal decoding, those skilled in the art can employ conventional image signal processing techniques, which are well-known in the field and will not be elaborated upon here.

[0036] S103, Data Matrix Normalization Processing. The visual processing module extracts a single-frame RGB digital image and scales its spatial resolution to a preset size using an interpolation algorithm. To reduce the interference of ambient lighting changes on feature extraction, the visual processing module independently calculates the normalization result of pixels for each color channel. Specifically, the scaling method is as follows: subtract the global pixel mean of the pre-training dataset in that channel from the scaled current pixel value, and then divide by the global pixel standard deviation of that channel, thereby standardizing the pixel distribution.

[0037] S104, Network Input Tensor Construction. The vision processing module stacks the three normalized two-dimensional channel matrices along the depth direction to construct a three-dimensional tensor, and then combines it with an independent batch dimension to form a four-dimensional input tensor, which is stored in the cache of the computation unit for use by the deep learning model.

[0038] S105, Deep Learning Model Construction and Training. The deep learning model used is built on a convolutional neural network architecture, specifically including a convolutional backbone network for feature extraction, a feature pyramid fusion network, and a multi-task prediction head. The prediction head includes a bounding box regression branch that outputs 3D coordinate offsets, and a classification probability branch that outputs continuous maturity features. Before device deployment, the model was trained using an offline image dataset containing tomatoes at different growth stages. During training, the sample images were cropped and rotated to increase the data volume. The manually labeled ground truth bounding box positions and maturity levels were used as labels. The cross-entropy function was used to calculate the classification loss, and the intersection-over-union (IoU) correlation function was used to calculate the bounding box loss. The network weight parameters were updated using the backpropagation algorithm until the loss value decreased to a set range, completing the training.

[0039] This invention provides a tomato ripeness grading and harvesting integrated machine based on deep learning. After acquiring the output of the deep learning model, the vision processing module performs a three-dimensional spatial coordinate positioning operation. The specific process includes the following steps: S106, Object Detection Box Parsing and Center Point Extraction. The vision processing module extracts the 2D bounding box data of all detected targets from the bounding box regression branch of the deep learning model. For scenarios where multiple tomatoes exist within the field of view, the vision processing module reads the classification confidence score associated with each bounding box and extracts the bounding box with the highest confidence score exceeding a preset threshold as the current target. The vision processing module extracts the diagonal pixel coordinates of this bounding box and calculates its average value. This average value is used as the center pixel coordinate of the target tomato in the 2D RGB image, denoted as... .

[0040] S107, Depth Map Alignment and Depth Information Extraction. Based on the synchronization triggering mechanism of the camera's underlying hardware, the vision processing module extracts a depth map aligned with the current RGB image timestamp. The vision processing module then extracts coordinates from the depth map. The corresponding depth pixel value. To eliminate noise interference, the vision processing module establishes a local pixel window based on this center coordinate. The vision processing module traverses the depth values ​​within the window, removes invalid depth points with values ​​of 0, calculates the median of the remaining valid depth values ​​within the local pixel window, and uses this median result as the actual physical depth value of the target tomato relative to camera 34, denoted as . .

[0041] S108, the two-dimensional pixel position only represents the projection relationship of the tomato on the photosensitive plane. To guide the movement of the mechanical actuator, the two-dimensional information needs to be restored to three-dimensional physical information. Based on the pinhole imaging principle, the vision processing module calls the camera intrinsic parameter data obtained from the offline calibration before the device leaves the factory. This intrinsic parameter data includes the camera's physical focal length parameters in the horizontal and vertical directions, as well as the principal point coordinate parameters of the imaging plane. Using the aforementioned intrinsic parameter data, the vision processing module determines the center pixel coordinates... Combined with actual physical depth values The coordinates are transformed into three-dimensional physical coordinates in the camera coordinate system through geometric mapping calculations.

[0042] S109, Base Coordinate System Transformation and Data Output. Since camera 34 is mounted on the flexible gripper 33 at the end of robotic arm 32, its spatial position is dynamically changing. The vision processing module needs to transform the target coordinates from the camera coordinate system to the universal base coordinate system of the entire machine. The vision processing module reads the angle encoder data of each joint of the robotic arm in real time, and combines it with the inherent link parameter table of the robotic arm to calculate the real-time pose matrix from the robotic arm base to the end flexible gripper through forward kinematics.

[0043] The vision processing module multiplies the real-time pose matrix with the pre-stored hand-eye calibration matrix from the flexible gripper to the camera to obtain the extrinsic transformation matrix of the entire machine. Subsequently, the vision processing module uses this extrinsic transformation matrix to perform coordinate system transformation operations to obtain the absolute physical coordinates of the target tomato in the base coordinate system. The coordinate data is then sent as the positioning result to the motion control module inside the control system, serving as the position input signal for the subsequent grasping action of the robotic arm.

[0044] This invention provides a tomato ripeness grading and harvesting integrated machine based on deep learning. To achieve stepless damping control of the pneumatic actuator, the vision processing module performs continuous ripeness feature extraction after acquiring the network output. The specific process includes the following steps: S110, Obtain the classification layer probability output. After locating and capturing the target, the visual processing module reads the data from the end of the classification probability branch in the deep learning model. To transform the output feature tensor into a format conforming to the probability distribution, the visual processing module calls the Softmax activation function to perform exponential normalization on the original values ​​of the classification branch. The visual processing module extracts the probability values ​​of each maturity category predicted for the current target. Among them, the... The prediction confidence probability corresponding to each maturity category is denoted as . And the sum of the probabilities of each category is 1.

[0045] S111, Set the physical characteristic mapping weights. Before deploying the hardware, the system pre-establishes a correspondence between physical hardness and algorithm weights for each predefined maturity category. The category index is arranged in order of tomato growth stage to full ripeness. Those skilled in the art can statistically obtain the average critical pressure data for tomato skin rupture under various categories through conventional texture analyzer puncture experiments. Since the higher the tomato's maturity, the lower its critical pressure, the system takes the reciprocal of the critical pressure data measured for each category and maps the values ​​to the interval (0,1) using a maximum value normalization algorithm. The system uses the mapped value as the... The feature mapping weights for each category are denoted as... The weight increases in ascending order of category level. The value exhibits a monotonically increasing property.

[0046] S112, Calculation of continuous feature variables. The visual processing module performs a weighted calculation using the acquired probability distribution data and feature mapping weights. The biological ripening of tomatoes is a non-linear, gradual process; this is achieved by calculating the confidence probability for each category. With corresponding weights The summation of these products yields a smoothed estimate of the physical state. The visual processing module defines this summation as a continuous feature variable, denoted as... The specific formula for calculating continuous features is as follows: ; In the formula, These are the continuous feature variables of the output; For the first Feature mapping weights for each category; The model predicts that the target belongs to the first... The confidence probability of each category. Continuous feature variables derived from this calculation. This is a floating-point number with a value in the range (0,1). The closer the value is to 1, the lower the hardness of the target tomato skin, and the more easily it will be damaged by compression during subsequent physical transport.

[0047] S113, Data Packaging and Communication Transmission. After completing the above calculations, the vision processing module extracts the absolute physical coordinates obtained during the positioning calculation stage. and the continuous characteristic variables calculated in this stage The vision processing module packages the coordinate data and feature variables according to the system's internal communication protocol to generate a target feature data frame, which is then sent to the main control module via an electrically connected data bus. This data frame serves as the feedforward input signal for the control system to execute subsequent robotic arm grasping actions and micro-pump damping adjustment actions.

[0048] This invention provides a tomato ripeness grading and harvesting integrated machine based on deep learning. After acquiring the target coordinates, the motion control module executes low-level motion driving and physical grasping actions. The specific process includes the following steps: S201, Target Pose Data Analysis and Trajectory Planning. The main control module receives target feature data frames transmitted by the vision processing module and parses the absolute physical coordinates of the target tomato in the base coordinate system. The main control module sends this coordinate data to the motion control module as the target point for the robotic arm's end effector. The motion control module internally runs an interpolation algorithm to plan the absolute physical coordinates from the current spatial position of the flexible gripper 33 to the target. The movement trajectory. For the spatial trajectory interpolation calculation of the robotic arm, those skilled in the art can use conventional cubic polynomial interpolation or B-spline curve techniques. The trajectory planning algorithm is a well-known technology in this field and will not be described in detail here.

[0049] S202, Inverse Kinematics Calculation. Since three-dimensional spatial coordinates only represent the end position, and the underlying drive motor is controlled by rotation angle, the system needs to perform a coordinate-to-angle transformation. The motion control module performs inverse kinematics calculation based on the device's inherent link geometry model and the DH parameter table. The motion control module converts the absolute physical coordinates in three-dimensional space... Using the preset end-effector gripping posture matrix as input, the target angle variables that the base 31 rotary joint and each extension joint of the robotic arm 32 need to reach are calculated in reverse.

[0050] S203, underlying motor servo control. The motion control module reads the calculated target angle variables of each joint and converts them into underlying drive commands for the servo motors. The pulse generation circuit inside the motion control module compares the difference between the current actual angle and the target angle of each joint and generates a control signal with the corresponding duty cycle. This control signal is synchronously sent to the drive motors of each joint of the base 31 and the robotic arm 32 through hardware circuitry. The drive motors rotate, causing the corresponding mechanical links to move spatially, guiding the flexible claw 33 mounted at the end effector to move along the planned trajectory to the absolute physical coordinates. The corresponding position. For the specific algorithm implementation of motor servo closed-loop control, those skilled in the art can use conventional PID control theory. The adjustment mechanisms of its internal current loop and speed loop are well-known technologies in this field and will not be elaborated upon here.

[0051] S204, End effector closing operation. During the extension operation of the robotic arm 32, the motion control module reads the encoder feedback data of each joint motor in real time to determine whether the flexible gripper 33 has reached the target position. When the system error between the actual pose and the target pose is within the set tolerance range, the motion control module determines that the spatial movement action has been completed. Subsequently, the motion control module outputs a closing command to the opening and closing drive mechanism inside the flexible gripper 33. The physical grippers of the flexible gripper 33 execute an inward closing action according to the command, completing the physical gripping of the target tomato.

[0052] This invention provides a tomato ripeness grading and harvesting integrated machine based on deep learning. After the motion control module controls the flexible claw 33 to complete the physical clamping of the target, it performs fruit separation and collection control. The specific process includes the following steps: S205, Fruit picking and physical separation. After confirming that the flexible claw 33 has reached the closed state, the motion control module sends a separation command to the base 31 and the robotic arm 32. The robotic arm 32 and the flexible claw 33 perform a pull-back displacement, causing the target tomato in the clamped state to physically break off from the connected vine stem, completing the fruit picking and separation operation.

[0053] S206, Handle-free in-situ release. The system does not control the robotic arm 32 to return to the vehicle body 1 for storage while carrying the fruit. After the fruit stem separation action is completed, the motion control module maintains the current spatial extension posture of the robotic arm 32 and directly issues an opening command to the drive mechanism of the flexible claw 33. The picked tomato is released from the gripper under the action of gravity and falls into the inlet of the conveyor channel fixedly installed below the flexible claw 33. Through this operation logic, the system achieves handle-free release of the fruit.

[0054] S207, Decoupling and Distribution of Motion Commands. The main control module is internally configured with parallel control logic, separating the robotic arm's spatial picking action from the fruit pneumatic conveying action into two execution flows. When the motion control module sends a status signal to the main control module indicating that the release action has been completed, the main control module extracts the next target's physical coordinates from the vision processing module's cache queue and sends them to the motion control module, driving the robotic arm 32 to move from its current position to the next grasping point.

[0055] S208, Parallel conveying operation. While the robotic arm 32 is spatially displacing towards the next target, the main control module takes over the tomatoes that have fallen into the conveying channel. The main control module extracts the continuous feature variables associated with that tomato. Based on this characteristic variable, a damping adjustment command is generated and sent to the micro air pump and pneumatic actuator. The pneumatic actuator adjusts the air pressure inside the pipe according to the command, driving the tomatoes to move towards the collection device. The fruit-finding displacement of the front-end robotic arm and the conveying action of the rear-end pneumatic actuator are executed in parallel without interference, realizing decoupled control of the harvesting system.

[0056] To avoid pressure regulation conflicts caused by multiple tomatoes entering the conveyor pipe 42 simultaneously, the main control module is equipped with single-fruit conveying cycle control logic. Before the current tomato passes through the damped conveying area of ​​the conveyor pipe 42 and completes the corresponding pressure reset, the motion control module can drive the robotic arm 32 to find the next target, but does not control the flexible claw 33 to release the next tomato into the funnel 43. Only after the main control module receives the reset completion signal of the conveyor mechanism 4 can it execute the release action of the next tomato.

[0057] This invention provides a tomato ripeness grading and harvesting integrated machine based on deep learning. The system is equipped with a funnel 43 to connect the fruit harvesting and pneumatic conveying process. The specific process includes the following steps: S209, Follow-up mechanism for positioning. The funnel 43 is positioned below the flexible claw 33 and fixedly connected to one end of the conveying pipe 42. The conveying pipe 42 is positioned below the robotic arm 32 via a fixing plate 44, allowing the funnel 43 to move synchronously with the robotic arm 32 and the flexible claw 33. During the process of the motion control module driving the robotic arm 32 to perform fruit-finding and picking actions, the funnel 43 moves synchronously with the flexible claw 33, ensuring that the opening of the funnel 43 is below the path of the tomato after release.

[0058] S210, Gravity-driven transition. After the flexible claw 33 opens, the tomato detaches from the gripper and falls into the funnel 43 under the influence of gravity. The funnel 43 catches the falling tomato and restricts its lateral displacement.

[0059] S211, Pipeline transition and sealed guidance. The bottom outlet of funnel 43 is connected to the inlet of conveying pipe 42, and the inner diameter of conveying pipe 42 is larger than the maximum outer diameter of the tomato to be harvested. The tomato is guided through the inner wall of funnel 43 and slides into conveying pipe 42. Conveying pipe 42 provides a closed channel environment for subsequent air pressure regulation. For the mechanical seal connection between funnel 43 and conveying pipe 42, those skilled in the art can use conventional clamp locking techniques, etc., and the connection and fixing method is well known in the art and will not be described in detail here.

[0060] This invention provides a tomato ripeness grading and harvesting integrated machine based on deep learning. The conveying pipe 42 is internally equipped with an outer airbag 421, an inner airbag 422, and a buffer ring 423. The specific structure and mechanical mechanism include the following steps: S301, Airway Connection and Chamber Layout. The delivery tube 42 has a double-layer structure, including an outer airbag 421 and an inner airbag 422. The outer airbag 421 forms the outer support and sealing structure of the delivery tube 42. The inner airbag 422 is located inside the outer airbag 421 and forms a tomato-like sliding channel. A buffer ring 423 is located on the inner wall of the inner airbag 422. A vent is provided on the side wall of the delivery tube 42. One end of the vent connects to the sealed internal chamber of the inner airbag 422, and the other end connects to the micro air pump 41 controlled by the main control module through an external airway.

[0061] S302, Inflation and Deformation. A miniature air pump 41 inflates the internal chamber of the inner air bladder 422, causing it to expand into the internal space of the delivery pipe 42. When a harvested tomato falls along the delivery pipe 42 and comes into contact with the inflated inner air bladder 422, the inner air bladder 422 is compressed and deformed. This deformation causes the surface of the inner air bladder 422 to conform to the outer contour of the tomato, forming a wrap around the tomato skin. A buffer ring 423 provides flexible contact with the tomato surface and reduces localized impact during the tomato's fall.

[0062] S303, Damping Conversion Mechanics Principle. The pneumatic damping mechanism provided by the inner airbag 422 changes the normal pressure acting on the tomato surface by controlling the internal gas pressure, thus converting it into sliding friction that resists falling. With the tomato encased in the inner airbag 422, the micro-air pump 41 adjusts the inflation pressure inside the inner airbag 422, directly changing the radial pressure exerted by the inner airbag 422 on the tomato skin, thereby synchronously controlling the contact friction force experienced by the tomato as it slides downwards within the conveying tube 42. This friction force is in the opposite direction to the tomato's downward motion due to gravity, constituting a damping force during the conveying process. This counteracts part of the tomato's gravitational acceleration, limiting the tomato's falling speed within the conveying tube 42, thus achieving lossless conveying.

[0063] This invention provides a tomato ripeness grading and harvesting integrated machine based on deep learning. The main control module dynamically adjusts the inflation pressure of the inner air bladder 422 inside the delivery pipe 42 according to the ripeness characteristics output by the vision system. The specific process includes the following steps: S304, Maturity Feature Data Analysis. The main control module reads continuous feature variables associated with the currently picked tomatoes from the cache queue of the vision processing module. This continuous characteristic variable After normalization, its value ranges from [0,1]. This value is positively correlated with the actual ripeness of the tomato; the closer the value is to 1, the higher the ripeness of the tomato. The riper the tomato, the lower the firmness of the flesh. To prevent it from being damaged when it falls into the storage mechanism 5, the system needs to provide greater air pressure damping during the conveying stage in the conveying pipe 42 to reduce its falling speed.

[0064] S305, Target Pressure Mapping Calculation. The main control module is internally configured with pressure mapping logic to establish continuous feature variables. and target air pressure value The continuous mapping relationship between them. Based on this mapping relationship, the main control module processes the input continuous feature variables... The target air pressure value to be established when converting to internal airbag 422 In this mapping relationship, the target air pressure value With continuous characteristic variables The pressure increases with the increase of the target pressure. Using a continuous mapping mechanism instead of segmented threshold determination can avoid sudden pressure changes within the delivery pipe 42. Setting an upper limit ensures that the radial pressure exerted by the inner airbag 422 on the tomato surface is lower than the safe compression threshold for tomatoes of corresponding maturity, thereby increasing frictional damping while avoiding damage to the tomato skin.

[0065] S306, Pressure Follow-up Control. After acquiring the target air pressure value Pt, the main control module converts it into a control command and sends it to the motor driver of the micro air pump 41. The micro air pump 41 adjusts the motor speed and output air volume according to the control command, changing the inlet pressure of the vent hole on the side wall of the delivery pipe 42, so that the actual pressure inside the inner airbag 422 rises to the target air pressure value. And maintain stability. For the closed-loop feedback and following adjustment of the actual air pressure in the delivery pipe 42, those skilled in the art can use conventional PID control algorithms combined with air pressure sensors arranged in the pipe components to achieve this. Its pressure closed-loop servo adjustment mechanism is a well-known technology in the field and will not be described in detail here.

[0066] This invention provides a tomato ripeness grading and harvesting integrated machine based on deep learning. The main control module controls the micro air pump 41 to adjust the pressure of the inner air bladder 422 inside the delivery pipe 42 according to the calculated target air pressure. The specific process includes the following steps: S307, Air circuit connection and inflation. The main control module outputs a control signal to the micro air pump 41 based on the target air pressure value. The micro air pump 41 starts and outputs pressurized gas, which is introduced into the sealed chamber of the inner air bladder 422 distributed on the side wall of the delivery pipe 42.

[0067] S308, Damping Establishment and Deceleration. The inner air bladder 422 receives gas and expands towards the center of the delivery pipe 42, reducing the cross-sectional area of ​​the channel inside the delivery pipe 42. When the tomato falls along the delivery pipe 42 through this area, its skin compresses the expanded inner air bladder 422. Based on its internal air pressure, the inner air bladder 422 applies a radial positive pressure to the tomato surface. This positive pressure is converted into a frictional force that resists the tomato's descent. The main control module adjusts the inflation pressure to change the magnitude of this frictional force. Under the action of friction, the tomato's falling speed decreases, reducing the impact force when it falls into the storage mechanism 5.

[0068] S309, Exhaust and Depressurization and Channel Reset. After the tomato completes the buffer transport and falls into the storage mechanism 5, the main control module outputs a reset command. The micro air pump 41 stops supplying air and performs a depressurization operation. The gas inside the inner air bladder 422 is discharged outward, the inner air bladder 422 shrinks in volume and retracts to the inner wall edge of the conveying pipe 42. The inner cavity of the conveying pipe 42 returns to its initial straight-through state, completing the buffer transport operation of the current fruit and providing space for the next tomato to fall. For the inflation and deflation drive of the micro air pump 41, those skilled in the art can use a conventional pneumatic control system, and its air circuit opening and closing and depressurization control are well-known technologies in the field, and will not be described in detail here.

[0069] This invention provides a tomato ripeness grading and harvesting integrated machine based on deep learning. To achieve time-synchronous control of the storage mechanism 5, the main control module estimates the slippage time of the tomato based on the force state of the tomato in the conveying pipe 42 and the diversion pipe 45, and calculates the time node when the fruit reaches the outlet of the diversion pipe 45. The specific process includes the following steps: S401, Start Time Recording. At the instant the motion control module drives the flexible claw 33 to open and release the tomato, the main control module reads the current system clock and records the system time at the time of the release action, using it as the start time of the tomato's fall.

[0070] S402, Force Analysis and Slippage Time Estimation. When the tomato slides down the conveyor pipe 42, it is subjected to the combined effects of downward gravity and upward frictional resistance applied by the inner air bladder 422. As described in the previous steps, because the system applies different levels of aerodynamic frictional resistance to fruits of different ripeness, the sliding speed of highly ripe fruits is slower than that of less ripe fruits. The main control module determines the sliding speed based on the current target air pressure value. The physical lengths of the conveying pipe 42 and the diversion pipe 45, along with the pre-established sliding time calibration relationship, are used to estimate the sliding time required for the tomato to slide from the funnel 43 into the conveying pipe 42 and out of the diversion pipe 45.

[0071] S403, Arrival Time Prediction and Command Distribution. After acquiring the descent time, the main control module adds it to the descent start time to generate the time node when the tomato arrives at the outlet of the diversion pipe 45. The main control module uses continuous characteristic variables... The system determines the current maturity level of the tomato by establishing a correspondence with preset maturity ranges, and writes this maturity level, target classification region, and arrival time into the control queue of storage mechanism 5. This time node provides a time reference for triggering the actions of telescopic rod 56 and stop post 57, ensuring that the execution sequence of grading actions matches the time node when the corresponding fruit arrives at the outlet of diversion pipe 45. For time synchronization and task queue management within the main control module, those skilled in the art can employ conventional real-time operating system task scheduling mechanisms. Multi-threaded data synchronization is a well-known technology in the field and will not be elaborated upon here.

[0072] This invention provides a tomato ripeness grading and harvesting integrated machine based on deep learning. The main control module controls the storage mechanism 5 to perform corresponding sorting actions based on the predicted fruit arrival time. The specific process includes the following steps: S404, Task Scheduling and Clock Polling. The main control module sorts the sorting tasks in the queue according to their arrival time nodes. The main control module reads the system's current running time in real time and compares it with the earliest arrival time node in the queue. When the current running time reaches that time node, the main control module determines that the corresponding tomato has reached the outlet position of the diversion pipe 45 and activates the hierarchical control command corresponding to that task.

[0073] S405, Signal Conversion and Flow Guidance Response. The main control module analyzes the maturity level data in the activated command, determines the target classification area based on the level data, and outputs an extension control signal to the corresponding telescopic rod 56. After the telescopic rod 56 rises, it drives the baffle 57 into the sliding path after the outlet of the diversion pipe 45, causing the baffle 57 to block the sliding tomatoes. After the tomatoes slide out of the diversion pipe 45, they hit the baffle 57 and their trajectory deflects. Then, they roll along the corresponding inclined plate 58 into the corresponding classification area inside the housing 51.

[0074] S406, Fruit Collection and Status Update. Tomatoes are guided by the diversion pipe 45, baffle 57, and inclined plate 58, falling into the corresponding sorting area inside the housing 51. After maintaining the telescopic rod 56 and baffle 57 in their preset positions for a predetermined time, the main control module controls the telescopic rod 56 to reset, causing the baffle 57 to exit the sliding path. The currently executed sorting task is removed from the control queue, the queue status is updated, and the time node of the next task is polled for triggering. For the drive control of the telescopic rod 56, those skilled in the art can use conventional electric or pneumatic telescopic mechanisms. Its telescopic drive and mechanical support structure are well-known technologies in the field and will not be described further here.

[0075] This invention provides a tomato ripeness grading and harvesting integrated machine based on deep learning. To achieve the classified collection of fruits at different ripeness levels, the integrated machine has multiple openings at the front end of the housing 51, through which the diversion pipe 45 communicates with the interior of the housing 51. Each opening has a telescopic rod 56, a stop post 57, and an inclined plate 58 correspondingly arranged on its inner side. The storage mechanism 5 uses the telescopic rod 56 to drive the stop post 57 to rise and fall, and works with the inclined plate 58 to change the path of the fruit's descent. The specific process includes the following steps: S407, Stopper Drive and Position Adjustment. The main control module outputs control signals to the corresponding telescopic rod 56 based on the arrival time and maturity level instructions. When the telescopic rod 56 extends, it raises the stopper 57, causing it to enter the tomato's sliding path after the corresponding opening; when the telescopic rod 56 retracts, it lowers the stopper 57, causing it to exit the corresponding sliding path.

[0076] S408, Path Diversion. The tomato slides out of the diversion pipe 45 and enters the interior of the housing 51 through the corresponding opening at the front end of the housing 51. If the corresponding baffle 57 is in the raised state, the tomato will deflect its trajectory after hitting the baffle 57 and roll along the corresponding ramp 58, entering the corresponding classification area inside the housing 51; if the corresponding baffle 57 is not in the blocking position, the tomato enters the corresponding classification area inside the housing 51 along its original sliding path.

[0077] S409, Categorized Collection and Buffer Protection. Multiple independent categorization zones are formed inside the shell 51, each corresponding to a different maturity level. Tomatoes, after being diverted via the aforementioned path, fall into their corresponding categorization zones inside the shell 51. The inclined plate 58 guides the tomatoes to roll into designated areas and reduces the impact caused by direct fall. For the drive control of the telescopic rod 56, those skilled in the art can use conventional electric or pneumatic telescopic mechanisms; its telescopic drive and mechanical support structure are well-known technologies in the field and will not be described further here.

[0078] The working principle of this invention is as follows: At the start of operation, motor 35 drives wheel 2 to rotate, causing vehicle body 1 to move along the tomato planting area. A harvesting assembly 3 is installed on the top of vehicle body 1. The base 31 in the harvesting assembly 3 drives the robotic arm 32 to adjust its orientation. A camera 34 mounted on top of the flexible claw 33 continuously acquires images of the tomato plants and transmits the acquired image data to the control system. The control system uses a deep learning model to identify the tomatoes in the images and obtain the spatial location and maturity information of the target tomatoes.

[0079] Once the control system determines the location of the tomato to be harvested, the base 31 rotates and the robotic arm 32 extends, causing the flexible claw 33 at the end of the robotic arm 32 to move to the location of the target tomato. After the flexible claw 33 closes and grips the tomato, the robotic arm 32 performs a pull-back action, separating the tomato from the vine and stem, completing the harvest. After harvesting, the flexible claw 33 opens in place, and the harvested tomato falls into the funnel 43 located below the flexible claw 33 under the influence of gravity.

[0080] A funnel 43 is fixedly connected to one end of a conveying pipe 42, which is positioned below the robotic arm 32 via a fixing plate 44. After the tomato enters the funnel 43, it is guided into the conveying pipe 42. The conveying pipe 42 has a double-layer structure, including an outer air bladder 421, inside which is an inner air bladder 422. A buffer ring 423 is installed on the inner wall of the inner air bladder 422. A miniature air pump 41 is connected to the conveying pipe 42 via a flexible hose. The control system controls the miniature air pump 41 to operate based on the tomato ripeness information identified by the camera 34, allowing gas to enter the inner air bladder 422 within the conveying pipe 42.

[0081] As the inner air bladder 422 inflates, the cross-sectional area of ​​the conveying pipe 42 changes. When the tomato slides down the conveying pipe 42, its surface comes into contact with the inner air bladder 422 and the buffer ring 423. The inner air bladder 422 provides flexible support and frictional damping to the tomato surface based on its internal air pressure, while the buffer ring 423 provides auxiliary cushioning, thereby reducing the tomato's sliding speed within the conveying pipe 42. For tomatoes with higher maturity and lower skin hardness, the control system increases the inflation pressure of the inner air bladder 422, allowing the conveying pipe 42 to provide greater cushioning damping; for tomatoes with lower maturity and higher skin hardness, the control system decreases the inflation pressure of the inner air bladder 422, allowing the tomato to pass smoothly through the conveying pipe 42. This achieves adaptive buffering conveying of tomatoes at different maturity levels.

[0082] After passing through the conveying pipe 42, the tomatoes enter the diversion pipe 45, which is fixedly connected to the other end of the conveying pipe 42. Multiple openings are provided at the front end of the housing 51, through which the diversion pipe 45 conveys the tomatoes into the storage mechanism 5. The control system predicts the time it takes for the tomatoes to reach the outlet of the diversion pipe 45 based on the time it takes for the tomatoes to enter the conveying pipe 42 from the funnel 43, the lengths of the conveying pipe 42 and the diversion pipe 45, and the damping state of the inner air bladder 422, and determines the corresponding classification area based on the tomato's maturity.

[0083] As the tomatoes are about to enter the housing 51 from the diversion pipe 45, the control system sends a control signal to the corresponding telescopic rod 56. After the telescopic rod 56 extends, it raises the baffle 57, causing the baffle 57 to enter the tomato's sliding path after the corresponding opening. After the tomato slides out of the diversion pipe 45 and enters the housing 51, it impacts the baffle 57, causing its trajectory to deflect. It then rolls along the corresponding inclined plate 58 and enters the corresponding classification area inside the housing 51. The coordinated action of the telescopic rod 56, baffle 57, and inclined plate 58 at different positions ensures that tomatoes of different ripeness levels enter different storage areas within the housing 51, thus achieving graded storage of tomatoes based on their ripeness.

[0084] After a single tomato completes buffering, conveying, and grading storage, the control system stops supplying air to the micro-pump 41 and depressurizes it, causing the inner airbag 422 to contract and reset, and the conveying pipe 42 to return to its initial channel state. Simultaneously, the telescopic rod 56 retracts, and the baffle 57 exits the tomato's sliding path, preparing for the conveying and grading of the next tomato. As the front-end robotic arm 32 continues to harvest the next tomato, the rear-end conveying mechanism 4 and storage mechanism 5 repeatedly perform inflation buffering, diversion guidance, and classification collection actions based on the corresponding tomato's ripeness information, thereby achieving integrated continuous operation of tomato harvesting, lossless conveying, and ripeness-graded storage.

[0085] When tomatoes need to be removed from storage unit 5, operators can open the top cover 52 using handle 53 to inspect or remove tomatoes from inside housing 51; alternatively, the tailgate 55 can be rotated around the pivot 54 to open and remove the sorted and stored tomatoes from the tail of housing 51. After fruit removal, the top cover 52 and tailgate 55 are closed, and the equipment returns to normal harvesting, conveying, and grading operation.

[0086] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0087] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A tomato ripeness grading and harvesting integrated machine based on deep learning, characterized in that, The vehicle includes a vehicle body (1), a picking assembly (3) is installed on the top of the vehicle body (1), a storage mechanism (5) is provided on the top of the vehicle body (1), a conveying mechanism (4) is provided on the outside of the storage mechanism (5), and a control system is provided inside the vehicle body (1). The harvesting component (3) is mounted on the vehicle body (1) and includes a base (31) and a robotic arm (32); the storage mechanism (5) is mounted at the rear of the vehicle body (1); The conveying mechanism (4) is located between the harvesting component (3) and the storage mechanism (5), and includes a micro air pump (41) and a conveying pipe (42). The micro air pump (41) is connected to the conveying pipe (42). One end of the conveying pipe (42) is connected to the funnel (43), and the other end is connected to the diversion pipe (45). The control system includes a vision processing module, which runs a deep learning model to obtain the spatial coordinates and ripeness characteristics of the target tomato. The control system controls the picking component (3) to pick according to the spatial coordinates, and controls the conveying mechanism (4) and the storage mechanism (5) to perform damped conveying and graded collection operations according to the maturity characteristics.

2. The tomato ripeness grading and harvesting integrated machine based on deep learning according to claim 1, characterized in that, The delivery pipe (42) has a double-layer structure. The delivery pipe (42) includes an outer airbag (421), and an inner airbag (422) is provided inside the outer airbag (421). A buffer ring (423) is provided on the inner wall of the inner airbag (422).

3. The tomato ripeness grading and harvesting integrated machine based on deep learning according to claim 1, characterized in that, The storage mechanism (5) includes a housing (51), a top cover (52) is movably connected to the top of the housing (51), a plurality of handles (53) are installed on the top of the top cover (52), a rotating shaft (54) is rotatably connected to the tail of the housing (51), a tailgate (55) is fixedly connected to the outside of the rotating shaft (54), a plurality of telescopic rods (56) are installed at the port of the housing (51), a stop post (57) is fixedly connected to the top of each telescopic rod (56), and a plurality of inclined plates (58) are provided inside the housing (51).

4. The tomato ripeness grading and harvesting integrated machine based on deep learning according to claim 1, characterized in that, The end of the robotic arm (32) is rotatably connected to a flexible claw (33), and a camera (34) is mounted on the top of the flexible claw (33); a motor (35) is provided at the front end of the vehicle body (1).

5. The tomato ripeness grading and harvesting integrated machine based on deep learning according to claim 3, characterized in that, The front end of the housing (51) has multiple openings, and the diversion pipe (45) communicates with the interior of the housing (51) through the openings.

6. The tomato ripeness grading and harvesting integrated machine based on deep learning according to claim 4, characterized in that, The control system also includes a main control module, a motion control module, and a pneumatic adjustment module. The main control module is electrically connected to the vision processing module, the motion control module, and the pneumatic adjustment module, respectively.

7. The tomato ripeness grading and harvesting integrated machine based on deep learning according to claim 6, characterized in that, The visual processing module runs the deep learning model, extracts the three-dimensional spatial coordinates of the target tomato, and outputs continuous feature variables corresponding to the current maturity of the target tomato as the maturity feature. The visual processing module multiplies the confidence probability of the target tomato belonging to each maturity category predicted by the deep learning model with the feature mapping weights corresponding to each maturity category, and calculates the sum of all products, using the sum as the continuous feature variable. The feature mapping weights are obtained by taking the reciprocal of the average critical pressure data for tomato skin rupture under each maturity category and then normalizing the mapping.

8. The tomato ripeness grading and harvesting integrated machine based on deep learning according to claim 7, characterized in that, The main control module is configured with air pressure mapping logic. The main control module generates a target air pressure value based on the continuous feature variable and sends the target air pressure value to the pneumatic adjustment module. The target air pressure value increases as the continuous feature variable increases. The pneumatic adjustment module drives the micro air pump (41) to inflate the inside of the delivery pipe (42) according to the target air pressure value, thereby changing the positive pressure applied by the delivery pipe (42) to the surface of the target tomato and converting the positive pressure into a sliding friction force that prevents the target tomato from sliding down the delivery pipe (42).

9. The tomato ripeness grading and harvesting integrated machine based on deep learning according to claim 6, characterized in that, After the motion control module controls the robotic arm (32) and the flexible claw (33) to complete the separation action of the fruit stem of the target tomato, it maintains the current spatial extension posture of the robotic arm (32) unchanged, and controls the flexible claw (33) to perform an opening action in place, so that the target tomato is released from the grip and falls into the funnel (43) by gravity. The main control module is configured with parallel control logic. During the period when the motion control module drives the robotic arm (32) to move from the current position to the next gripping point, the main control module controls the pneumatic adjustment module in parallel to perform pneumatic damping conveying operation on the target tomato that falls into the conveying pipe (42).

10. The tomato ripeness grading and harvesting integrated machine based on deep learning according to claim 8, characterized in that, The main control module records the system time as the start time the target tomato falls into the funnel (43) at the instant. Based on the target air pressure value and the physical length of the delivery pipe (42) and the diversion pipe (45), it estimates the time it takes for the target tomato to fall. The time it takes to fall is added to the start time to generate the time node when the target tomato reaches the outlet of the diversion pipe (45). When the current running time of the system reaches the time node, the main control module outputs an execution command to the storage mechanism (5), causing the storage mechanism (5) to block the sliding channel of the diversion tube (45), so that the target tomato deflects its trajectory and enters the corresponding classification area inside the storage mechanism (5).

Images