A non-destructive mango picking device and method

Abstract

This invention discloses a non-destructive mango separation intelligent agricultural harvesting device and method. The harvesting device includes a walking mechanism, a lifting mechanism, a horizontal sliding mechanism, a robotic arm, an energy-induced end effector, and a flexible buffer collection mechanism. The energy-induced end effector includes a flexible fruit clamping component, an end recognition camera, a micro-vibration spectral generation and analysis module, and a multi-frequency resonant energy-induced module. The micro-vibration spectral generation and analysis module is used to perform non-contact detection of the fruit stalk, obtain the spectral response of the fruit stalk cells, and compare it with the in vitro difficulty fingerprint spectrum to determine whether the optimal harvesting period has been reached. If so, a decision command containing the inherent resonant frequency and phase characteristics of the fruit stalk is output. The multi-frequency resonant energy-induced module emits a resonant wave with the same frequency but opposite phase as the inherent vibration frequency of the fruit stalk, and uses the phase cancellation interference principle to instantaneously weaken the cell binding force at the molecular level to achieve non-destructive separation.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural machinery technology, specifically to an intelligent agricultural harvesting device and method for non-destructive mango separation. Background Technology

[0002] As a globally important tropical fruit, mangoes have seen continuous growth in both planting area and yield, especially in major producing regions such as Southeast Asia and South China, where they have become a vital pillar of the regional agricultural economy. According to agricultural census data, China's mango production has long ranked among the world's top, and the industry scale continues to expand.

[0003] However, in stark contrast, mango harvesting in my country still relies heavily on manual labor. This traditional method is facing increasingly severe challenges: a structural shortage of rural labor has led to a sharp rise in labor costs, with harvesting expenses accounting for an excessively high proportion of total production costs; at the same time, the efficiency and quality of manual harvesting are extremely unstable, easily affected by workers' physical strength, weather conditions, and operational skills, resulting not only in low efficiency but also making it difficult to avoid physical damage to the fruit during the harvesting process.

[0004] To overcome this bottleneck, the industry has attempted to introduce some electric or semi-automated harvesting devices, but their technological efficiency still falls far short of actual needs. At the perception level, these existing devices mostly rely solely on traditional vision systems, failing to accurately judge the internal ripeness and hidden defects of mangoes. This leads to frequent instances of mistakenly harvesting "falsely ripe" or overripe fruit. Furthermore, their multi-sensor information is often processed in isolation, failing to deeply integrate and form a quantitative assessment of fruit harvesting risks. At the execution level, their mechanical gripping and cutting mechanisms lack the ability to perceive and adapt to individual fruits, resulting in fixed and rough operating actions. Particularly noteworthy is that existing devices completely ignore the critical damage pathway of sap seepage after stem cutting causing chemical burns to the peel. They also fail to effectively suppress physical collisions caused by inertial swaying of the fruit during cutting and transport. The fragile peel and easily browned flesh of mangoes make these rigid operations highly susceptible to causing explicit and implicit damage throughout the entire grasping, separation, and collection chain, severely restricting the improvement of the fruit's commercial value.

[0005] On the other hand, mango trees have dense canopies and complex fruit distribution. Existing harvesting equipment is limited by the working range and flexibility of robotic arms, making it difficult to safely and effectively reach all ripe fruits, resulting in a low overall harvest rate. A deeper problem lies in the fact that existing devices, as automated units, lack the intelligent decision-making and learning capabilities of experienced fruit growers. Their system architecture is relatively fragmented, with perception, decision-making, and execution processes isolated. They cannot generate refined operational instructions based on real-time environment and fruit condition, such as precise grasping posture, safe force control curves, and optimal cutting parameters. Furthermore, they cannot dynamically evaluate the efficiency of the work area and plan paths based on historical harvesting data to achieve continuous global optimization of harvesting efficiency and quality. Therefore, existing technology is still far from truly "precise, non-destructive, and intelligent" harvesting.

[0006] However, in-depth analysis reveals that even if the aforementioned technologies such as multimodal fusion sensing, active damage prevention, and intelligent scheduling optimization are implemented, all existing automated harvesting solutions still fall within the fundamental framework of "external intervention"—that is, no matter how precise the sensing system or how flexible the actuator, the final action still involves applying mechanical clamping and shearing forces to the fruit stalk. This process itself determines that the vascular bundles of the fruit stalk will inevitably rupture, juice will inevitably flow out, and thus chemical burns will occur. This damage risk, determined by the technological path itself and which cannot be eradicated by optimizing the "cutting" action, constitutes a fundamental bottleneck that existing technologies cannot overcome.

[0007] In conclusion, there is an urgent need to break through the traditional paradigm of "external intervention" harvesting and explore a new technical approach that can establish a physiological dialogue with fruit trees and induce spontaneous and damage-free fruit separation. This has significant economic value and urgent application demand for fundamentally solving the problem of fruit damage and achieving truly high-quality intelligent harvesting. Summary of the Invention

[0008] The purpose of this invention is to overcome the aforementioned problems and provide a non-destructive intelligent agricultural harvesting device and method for mangoes. This intelligent agricultural harvesting device and method achieves accurate judgment of the detachment state of the fruit stalk through micro-vibration spectral analysis, and emits a resonant wave with the same frequency and opposite phase as the inherent vibration frequency of the fruit stalk through a multi-frequency resonant energy induction module. This instantly weakens the cell binding force at the molecular level, allowing the fruit to separate from the mother plant in a manner that is almost self-detaching, with almost no juice seepage, fundamentally eliminating chemical burns. At the same time, since there is no need for traditional cutting actions, the physical impact at the moment of cutting is completely eliminated, achieving truly non-destructive, efficient, and intelligent harvesting.

[0009] The objective of this invention is achieved through the following technical solution: A non-destructive mango-picking intelligent agricultural device includes a walking mechanism, a lifting mechanism, a horizontal sliding mechanism, a robotic arm, an energy-induced end effector, and a flexible buffer collection mechanism. The walking mechanism is equipped with a walking recognition camera; the lifting mechanism is mounted on the walking mechanism; the horizontal sliding mechanism and the flexible buffer collection mechanism are mounted on the lifting mechanism; the robotic arm is mounted on the horizontal sliding mechanism; and the energy-induced end effector is mounted on the end of the robotic arm. The energy-induced end effector includes a flexible fruit-clamping component, an end-of-life recognition camera, a micro-vibration spectroscopy generation and analysis module, and a multi-frequency resonance energy-inducing module. Both the micro-vibration spectroscopy generation and analysis module and the multi-frequency resonance energy-inducing module are mounted on the flexible fruit-clamping component. The micro-vibration spectroscopy generation and analysis module is used for non-contact detection of the fruit stalk, acquiring the spectral response of the stalk cells at a specific frequency, and comparing it with an in vitro difficulty fingerprint to determine whether the optimal harvest period has been reached. If so, a decision command containing the stalk's inherent resonant frequency and phase characteristics is output. The multi-frequency resonance energy-inducing module emits a resonant wave with the same frequency but opposite phase as the stalk's inherent vibration frequency according to the decision command. Utilizing the phase cancellation interference principle, it instantaneously weakens cell binding forces at the molecular level, allowing the fruit to separate from the parent plant without damage.

[0010] In a preferred embodiment of the present invention, two sets of the horizontal sliding mechanism, the robotic arm, the energy-induced end effector, and the flexible buffer collection mechanism are each provided. By setting two sets of the above-mentioned mechanisms, the walking mechanism moves between two mango trees, enabling two sets of harvesting operations to be performed simultaneously, thereby improving harvesting efficiency.

[0011] In a preferred embodiment of the present invention, the flexible fruit clamping assembly includes a pair of flexible clamps and a fruit clamping drive motor for driving the flexible clamps to open and close, wherein the micro-vibration spectrum generation and analysis module and the multi-frequency resonance energy induction module are both disposed on the flexible clamps.

[0012] In a preferred embodiment of the present invention, the walking mechanism includes a track assembly and a frame, with the lifting mechanism mounted on the frame. Compared to ordinary wheeled walking mechanisms, the track assembly has a larger ground contact area and lower ground pressure, enabling stable movement in soft, muddy, or uneven terrain in orchards without easily sinking or slipping. Simultaneously, the track structure provides stronger traction and climbing ability, ensuring the equipment can move flexibly and turn between rows of trees in the orchard, thus improving the equipment's adaptability to complex orchard environments.

[0013] In a preferred embodiment of the present invention, the lifting mechanism is a scissor lift mechanism.

[0014] In a preferred embodiment of the present invention, the horizontal sliding mechanism includes an X-axis sliding mechanism and a Y-axis sliding mechanism. The X-axis sliding mechanism is mounted on the lifting mechanism, and the Y-axis sliding mechanism is mounted on the X-axis sliding mechanism. Specifically, both the X-axis and Y-axis sliding mechanisms can employ existing linear drive mechanisms. The combination of the X-axis and Y-axis sliding mechanisms enables two-dimensional translational motion of the end effector in the horizontal plane. Combined with the vertical motion of the lifting mechanism, this constructs a three-dimensional working area covering the canopy, allowing the robotic arm to quickly reach the target fruit location via the translation of the slide rails without significant adjustments to its posture. This significantly expands the working range of a single machine and improves harvesting efficiency.

[0015] In a preferred embodiment of the present invention, the flexible buffer collection mechanism includes a fruit collection basket, a flexible flow-guiding funnel, and a collection and identification camera. The inner cavity of the fruit collection basket is provided with multiple layers of elastic buffer rope netting. The flexible flow-guiding funnel is located at the top of the fruit collection basket, and the collection and identification camera is located at the top of the flexible flow-guiding funnel for real-time damage detection and grading of the fruit skin. Through this structure, the multiple layers of elastic buffer rope netting in the inner cavity of the fruit collection basket can absorb the kinetic energy of the falling fruit step by step, achieving a "soft landing" of the fruit and effectively avoiding rigid collision damage between the fruit and the bottom or side walls of the basket. The flexible flow-guiding funnel can passively or actively adjust its shape according to the falling posture of the fruit, guiding the fruit to fall smoothly into the center area of ​​the basket and preventing the fruit from tilting and impacting. The collection and identification camera located at the top of the funnel can instantly photograph the fruit skin condition before the fruit enters the basket, completing damage detection and grading, and realizing online feedback of harvest quality.

[0016] A non-destructive mango harvesting method for intelligent agriculture includes the following steps: The walking mechanism moves autonomously within the orchard, acquiring environmental information through a walking recognition camera. By linking the lifting mechanism and the horizontal sliding mechanism, the energy-induced end effector is moved to the vicinity of the target fruit in the three-dimensional space of the tree canopy; the end recognition camera scans the fruit group to obtain the spatial distribution and preliminary appearance information of the mangoes; The micro-vibration spectrum generation and analysis module emits low-energy micro-vibration waves of a specific frequency to the pedicel of the target mango and receives the spectral response of the reflected signal. The spectral response is compared with a pre-established in vitro difficulty fingerprint spectrum to quantitatively assess the degree of degradation of the cellulose and hemicellulose structures of the pedicel cells and determine whether the pedicel has reached the mutation state of the optimal harvest period. If so, a decision command containing the pedicel's inherent resonant frequency, phase characteristics, and energy-induced parameters is output. Upon receiving the decision command, the lifting mechanism drives the end effector to a designated height, the horizontal sliding mechanism adjusts the relative horizontal position of the end effector, and the posture of the robotic arm is adjusted so that the end effector moves to the mango to be picked. The multi-frequency resonant energy induction module emits a resonant wave with the same frequency but opposite phase as the natural vibration frequency of the fruit stalk, and uses the phase cancellation interference principle to instantly weaken the cell binding force at the molecular level. At the same time as the resonant wave is applied, the flexible fruit clamping component applies a preset micro-pulling force to the fruit stalk, so that the fruit is separated from the energy weakening point of the fruit stalk, achieving damage-free harvesting without juice seepage and avoiding the juice flowing down the broken fruit stalk from burning the mango peel. After being separated, the fruit is caught by a flexible fruit-clamping component; the robotic arm moves along the planned retrieval trajectory and moves towards the flexible buffer collection mechanism via a horizontal sliding mechanism; after reaching the collection area, the robotic arm adjusts its posture so that the fruit is directly above the flexible guide funnel; the flexible fruit-clamping component releases, and the fruit is guided by the guide funnel into the fruit basket with a built-in multi-layer elastic buffer net, achieving a soft landing; at the same time, the collection and recognition camera integrated at the entrance performs real-time damage detection and grading of the fruit skin.

[0017] In a preferred embodiment of the present invention, during the harvesting process, a dynamic digital twin is established for each fruit to be harvested in the work area. The digital twin includes the three-dimensional position of the fruit, the detachment preparation state index, the estimated weight, the peel strength, and a damage risk-benefit game function calculated based on historical data. When the harvesting equipment enters a work area, the scheduling engine sends a harvesting request to the digital twins of all fruits in that area; the twins of fruits with high maturity and low risk of damage offer higher prices and are required to be harvested first; the twins of fruits located deep in the canopy and with high harvesting risk offer lower prices and are temporarily rejected for harvesting, waiting for a better opportunity or a gentler harvesting strategy later. The scheduling engine dynamically determines the next harvesting target and plans the optimal movement path based on the received bidding results, the robot's own position, and energy consumption status, thereby maximizing group collaboration and overall harvesting efficiency. This avoids the high damage rate that can result from blindly harvesting in a fixed order, achieving dynamic optimization of the harvesting sequence and maximizing overall harvesting efficiency while ensuring fruit quality.

[0018] Furthermore, the entire harvesting process's data is uploaded to a database, including micro-vibration spectral response, energy-induced parameters, game theory bidding results, and actual damage detection results output by the high-speed visual re-inspection unit. This data is used to iteratively optimize the detached stem difficulty fingerprint, the parameters of the classification and regression models, and the game function and negotiation strategy of the digital twin, enabling the harvesting equipment to possess self-learning and continuous evolution capabilities. As the amount of data increases, the system's judgment of the detached stem state becomes increasingly accurate, the prediction of energy-induced parameters becomes increasingly optimized, and the game theory scheduling strategy becomes increasingly rational. This allows the harvesting equipment to possess self-learning and continuous evolution capabilities, enabling it to adapt to the harvesting needs of different mango varieties, orchards, and years after long-term operation.

[0019] Compared with the prior art, the present invention has the following advantages: 1. This invention abandons the traditional "external force intervention" mechanical clamping and cutting method. It uses a micro-vibration spectrum generation and analysis module to perform non-contact precise detection of the physiological state of the fruit stalk cells, and uses a multi-frequency resonance energy induction module to emit a resonant wave with the same frequency and opposite phase as the natural vibration of the fruit stalk. This instantly weakens the cell binding force at the molecular level. Then, through a preset micro-pulling force (which is much smaller than the force of traditional picking methods), the fruit is separated from the mother fruit in a manner that is almost self-severing, with almost no juice seepage. This fundamentally eliminates the chemical burns to the fruit peel caused by the juice of the fruit stalk. At the same time, no cutting action is required, which completely eliminates the risk of physical impact and achieves truly damage-free harvesting.

[0020] 2. This invention introduces micro-vibration spectral analysis technology, which can non-contactly acquire the spectral response of fruit stalk cells at specific frequencies and compare it with a pre-established "in vitro difficulty" fingerprint spectrum to quantitatively assess the degree of degradation of the cellulose and hemicellulose structure of the fruit stalk, accurately determine whether the fruit has reached the mutation state of the optimal harvest period. This judgment is based on physiological information at the cellular level, rather than the traditional method that relies solely on visual appearance, effectively avoiding the misharvesting of "falsely ripe" or overripe fruits and ensuring harvest quality.

[0021] 3. The present invention is equipped with a flexible buffer collection mechanism. After the fruit is broken off, it is received by a flexible fruit clamping component, transferred by a robotic arm to a flexible guide funnel, and then falls into a fruit basket with a built-in multi-layer elastic buffer net to achieve a soft landing. At the same time, the collection and identification camera at the entrance performs real-time damage detection and grading of the fruit skin, effectively avoiding secondary collision damage in the collection process and realizing online feedback of harvesting quality.

[0022] 4. This invention establishes a dynamic digital twin for each fruit to be harvested, including three-dimensional position, detached preparation state index, estimated weight, peel strength, and damage risk-reward game function. The scheduling engine is based on a "bidding auction" mechanism, prioritizing the harvesting of fruits with high maturity and low damage risk, while temporarily abandoning the harvesting of fruits with high risk, waiting for a better opportunity, thereby maximizing group collaboration and overall harvesting efficiency. Moreover, by actively analyzing whether the mango meets the harvesting conditions, if so, the robotic arm is notified to harvest, forming an interactive harvesting mechanism.

[0023] 5. This invention uploads end-to-end data from each harvesting operation (including micro-vibration spectrum response, energy-induced parameters, game bidding results, and actual damage detection results) to a database for iterative optimization of the ex vivo difficulty fingerprint spectrum, model parameters, and the game function and negotiation strategy of the digital twin. As the number of operations increases, the system can continuously improve itself, constantly enhancing the accuracy and adaptability of harvesting decisions. Attached Figure Description

[0024] Figure 1 This is a three-dimensional structural diagram of the intelligent agricultural harvesting equipment for non-destructive mango separation according to the present invention.

[0025] Figure 2 This is a three-dimensional structural diagram of the energy-induced end effector of the present invention.

[0026] Figure 3 This is a three-dimensional structural diagram of the flexible buffer collection mechanism of the present invention. Detailed Implementation

[0027] To enable those skilled in the art to fully understand the technical solutions of the present invention, the present invention will be further described below in conjunction with embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0028] Example 1 Combination Figure 1 The non-destructive mango harvesting intelligent agricultural equipment of this embodiment includes a walking mechanism, a lifting mechanism, a horizontal sliding mechanism, a robotic arm 1, an energy-induced end effector, and a flexible buffer collection mechanism. The walking mechanism is equipped with a walking recognition camera. The lifting mechanism is mounted on the walking mechanism, as are the horizontal sliding mechanism and the flexible buffer collection mechanism. The robotic arm 1 is mounted on the horizontal sliding mechanism, and the energy-induced end effector is mounted on the end of the robotic arm 1. Each of the horizontal sliding mechanism, robotic arm 1, energy-induced end effector, and flexible buffer collection mechanism is provided in two sets. By setting two sets of these mechanisms, the walking mechanism moves between two mango trees, allowing for two sets of harvesting operations to be performed simultaneously, thus improving harvesting efficiency.

[0029] Combination Figure 1 The walking mechanism includes a track assembly 2 and a frame 3, with the lifting mechanism mounted on the frame 3. Compared to ordinary wheeled walking mechanisms, the track assembly 2 has a larger ground contact area and lower ground pressure, enabling it to walk stably on soft ground, mud, or uneven terrain in orchards without easily sinking or slipping. Simultaneously, the track structure provides stronger traction and climbing ability, ensuring the equipment can move flexibly and turn between rows of trees in the orchard, improving its adaptability to complex orchard environments.

[0030] Combination Figure 1 The lifting mechanism is composed of a scissor lift 4, and the specific structure can be referred to the prior art.

[0031] Combination Figure 1 The horizontal sliding mechanism includes an X-axis sliding mechanism 5 and a Y-axis sliding mechanism 6. The X-axis sliding mechanism 5 is mounted on the lifting mechanism, and the Y-axis sliding mechanism 6 is mounted on the X-axis sliding mechanism 5. Specifically, both the X-axis sliding mechanism 5 and the Y-axis sliding mechanism 6 can adopt existing linear drive mechanisms. The combination of the X-axis sliding mechanism 5 and the Y-axis sliding mechanism 6 enables the end effector to perform two-dimensional translational motion in the horizontal plane. Combined with the vertical motion of the lifting mechanism, this creates a three-dimensional working area covering the canopy, allowing the robotic arm 1 to quickly reach the target fruit location via the translation of the slide rails without significant adjustments to its posture. This significantly expands the working range of the single machine and improves harvesting efficiency.

[0032] Combination Figure 2 The energy-induced end effector includes a flexible fruit clamping component, an end recognition camera 7, a micro-vibration spectral generation and analysis module 8, and a multi-frequency resonance energy induction module 9. Both the micro-vibration spectral generation and analysis module 8 and the multi-frequency resonance energy induction module 9 are mounted on the flexible fruit clamping component. The micro-vibration spectral generation and analysis module 8 is used for non-contact detection of the fruit stalk, acquiring the spectral response of the fruit stalk cells at a specific frequency, and comparing it with the in vitro difficulty fingerprint spectrum to determine whether the optimal harvest period has been reached. If so, a decision command containing the inherent resonant frequency and phase characteristics of the fruit stalk is output. The multi-frequency resonance energy induction module 9 is used to emit a resonant wave with the same frequency but opposite phase as the inherent vibration frequency of the fruit stalk according to the decision command. Utilizing the phase cancellation interference principle, it instantaneously weakens the cell binding force at the molecular level, allowing the fruit to separate from the parent plant without damage.

[0033] Combination Figure 2 The flexible fruit clamping assembly includes a pair of flexible clamps 10 and a fruit clamping drive motor 11 for driving the flexible clamps 10 to open and close. The micro-vibration spectrum generation and analysis module 8 and the multi-frequency resonance energy induction module 9 are both disposed on the flexible clamps 10.

[0034] Combination Figure 3 The flexible buffer collection mechanism includes a fruit collection basket 12, a flexible flow-guiding funnel 13, and a collection and identification camera 14. The inner cavity of the fruit collection basket 12 is equipped with multiple layers of elastic buffer rope netting. The flexible flow-guiding funnel 13 is located on top of the fruit collection basket 12, and the collection and identification camera 14 is located on top of the flexible flow-guiding funnel 13, used for real-time damage detection and grading of the fruit skin. Through the above structure, the multiple layers of elastic buffer rope netting in the inner cavity of the fruit collection basket 12 can absorb the kinetic energy of the falling fruit step by step, achieving a "soft landing" of the fruit and effectively avoiding rigid collision damage between the fruit and the bottom or side wall of the basket. The flexible flow-guiding funnel 13 can passively or actively adjust its shape according to the falling posture of the fruit, guiding the fruit to fall smoothly into the center area of ​​the basket and preventing the fruit from tilting and impacting. The collection and identification camera 14 located on top of the funnel can take real-time pictures of the fruit skin condition before the fruit enters the basket, complete damage detection and grading, and realize online feedback of harvest quality.

[0035] Example 2 Combination Figure 1-3 The non-destructive mango harvesting method in this embodiment includes the following steps: (1) The walking mechanism walks autonomously in the orchard and obtains environmental information through the walking recognition camera.

[0036] (2) By linking the lifting mechanism and the horizontal sliding mechanism, the energy-induced end effector is moved to the vicinity of the target fruit in the three-dimensional space of the canopy; the end recognition camera 7 scans the fruit group to obtain the spatial distribution and preliminary appearance information of the mango.

[0037] (3) The micro-vibration spectrum generation and analysis module 8 emits a low-energy micro-vibration wave of a specific frequency (the specific frequency detected by the fruit stalk is concentrated in 20–50 Hz (refer to the resonant frequency of the harvestable period of Manila mango, which is 27.5–41.8 Hz) to the fruit stalk of the target mango. This energy will not damage the fruit, but will excite the fruit stalk cells to produce characteristic resonance. The spectral response of the reflected signal is received. The spectral response is compared with the pre-established in vitro difficulty fingerprint spectrum to quantitatively evaluate the degree of degradation of the cellulose and hemicellulose structure of the fruit stalk cells and determine whether the fruit stalk has reached the mutation state of the optimal harvest period. If so, a decision command containing the inherent resonant frequency, phase characteristics and energy-induced parameters of the fruit stalk is output.

[0038] Specifically, the core of the micro-vibration spectrum generation and analysis module is a laser Doppler vibration measurement system, which consists of five hardware parts: a micro-vibration excitation unit (such as a piezoelectric exciter), an optical detection unit (laser source + interference optical path), a signal processing unit (amplification and filtering), a spectrum analysis unit (FFT), and a control and communication unit. This technology is existing, with commercially available products (such as Polytec vibration meters) and research precedents (such as using resonant frequencies to measure avocado ripeness).

[0039] Specifically, the ex vivo difficulty fingerprint spectrum is a benchmark database that correlates the resonant frequency characteristics of pedicel cells with their ex vivo difficulty. It records the spectral response characteristics of the pedicel-fruit junction of the target mango variety (such as Golden Mango) at different ripening stages under specific micro-vibration excitation, including parameters such as resonant frequency, resonant peak width (reflecting the strength of cell binding force), phase characteristics, and damping ratio. It also establishes a correlation model between these parameters and physiological indicators such as the formation state of the pedicel abscission layer and the lignin and cellulose content of the pedicel.

[0040] From a biological perspective, the ease with which a fruit can be detached depends on the developmental state of the abscission layer. The abscission layer is a special cellular structure formed at the junction of the pedicel and the fruit. During fruit ripening, the cells in the abscission layer secrete hydrolytic enzymes that break down cellulose and pectin, causing the cell walls to swell and the intercellular bonds to weaken, ultimately leading to the natural detachment of the fruit. The degree of abscission development varies at different stages of ripening, resulting in regular changes in the mechanical properties of the pedicel (such as its inherent resonant frequency and damping ratio). For example, the pedicel tensile strength of mature seedless white grapes is only (2.1±0.5) N, while that of Red Globe grapes is (6.7±3.2) N, a significant difference. Furthermore, grapes with higher pedicel tensile strength tend to have higher lignin and cellulose content in their pedicels. Therefore, by detecting the spectral characteristics of the pedicel, such as its resonant frequency, the developmental state of the abscission layer can be inferred, thus determining the ease with which the fruit can be detached.

[0041] Furthermore, the method for constructing the map includes the following steps: Step 1: Sample collection and maturity assessment.

[0042] In the mango orchard, for the target variety (such as Jin Huang mango), a certain number (it is recommended to collect no less than 30-50 mangoes with stems) of mangoes at different ripening stages, such as green ripe, harvestable, and overripe, are collected, and the ripening status of each fruit is recorded.

[0043] Step 2: Micro-vibration frequency sweep test.

[0044] Using a micro-vibration spectral generation and analysis module (i.e., a laser Doppler vibration measurement system), a sweep frequency excitation (e.g., 20-2000Hz) is applied to the stem-fruit junction of each sample, while simultaneously receiving the spectral response of the reflected signal. The principle of this technology is: a vibrator applies a prescribed vibration to the stem, and a vibration sensor detects the vibration frequency generated by the applied vibration, thereby detecting the fruit's weight or weight change. During the excitation process, changes in fruit maturity cause a regular change in its resonant frequency—for example, studies using laser Doppler vibration measurement to determine the maturity of avocados have shown that the resonant frequency of the fruit decreases by 2-4 times during the ripening process, corresponding to a decrease in firmness.

[0045] Step 3: Feature parameter extraction.

[0046] The obtained spectral response was subjected to FFT transformation to extract the following key feature parameters: the natural resonant frequency of the peduncle, the width of the resonant peak (Q value, reflecting damping characteristics), phase characteristics, and damping ratio.

[0047] Step 4: Classification and correlation modeling of the difficulty of ex vivo experiments.

[0048] A correlation was established between the extracted spectral features and the ease of fruit detachment. The ease of detachment can be quantified using the following indicators: pedicel tensile strength (positively correlated with lignin and cellulose content), abscission layer thickness and cell morphology (the mature abscission layer consists of 2-3 layers of loosely arranged, large, elliptical cells with relatively large intercellular spaces and a higher degree of lignification than surrounding tissue cells), and actual breaking force. These physiological indicators were then correlated with the spectral features to establish a classification / regression model.

[0049] Step 5: Database construction.

[0050] The calibration results are stored in the control system to form an "ex vivo difficulty fingerprint" database for the target variety. This database should include: variety identification, maturity stage label, corresponding resonance frequency range, phase characteristics, and recommended harvesting decision threshold.

[0051] Laser Doppler vibration measurement technology has a solid research foundation, and its application in the field of non-destructive testing of fruit maturity is supported by public reports and patents, including but not limited to: Laser Doppler vibration measurement technology has been used for non-destructive assessment of fruit ripeness. Cranfield University in the UK uses laser and vibration testing to determine the resonant frequency of individual fruits, providing reliable ripeness assessments without damaging the avocado. This technology has been shown to accurately predict the ready-to-eat stage of avocados and has been published in the journal *Biosystems Engineering*.

[0052] Fruit ripeness can be non-destructively determined using resonant frequencies. The resonant frequencies obtained by vibrating fruit with sinusoidal waves in the 100-3000 Hz range are closely related to fruit firmness, which is generally associated with ripeness. This method has been successfully applied to various fruits, including apples, pears, kiwifruit, and melons. In kiwifruit, the correlation coefficient between the elasticity index and firmness measured by the traditional penetration method exceeds 0.9; in melons, the correlation coefficient between the elasticity index and sensory scores is as high as 0.96.

[0053] A method and apparatus for online non-destructive testing of fruit firmness. Zhejiang University has disclosed a method and apparatus for online non-destructive testing of fruit firmness. The method involves placing fruit on a conveyor belt, emitting sound waves from a loudspeaker to excite the fruit to vibrate, and projecting a laser Doppler vibrometer onto the vibrating fruit surface to collect vibration velocity information, thereby achieving online non-destructive testing of fruit firmness and quality grading.

[0054] Specifically, the abrupt change at the optimal harvest time refers to the qualitative inflection point from the formation of the abscission layer to its functional activation, which is manifested as follows: At the molecular level: Ethylene burst → pectinase / cellulase activation → rapid cell wall degradation.

[0055] From a mechanical perspective: the fruit stalk separation force drops drastically (e.g., Manila mangoes: unripe ones require 56 Hz excitation, but harvestable ones only require 27.5–41.8 Hz).

[0056] At the signal level: the resonant frequency shifts significantly to lower frequencies, damping increases, and the spectral fingerprint undergoes abrupt changes.

[0057] For example, the resonant frequency of the fruit stalk of Manila mangoes decreases by about 25%–50% during the harvest period compared to when they are unripe. At this time, the abscission layer has fully developed, and the fruit stalk is very easy to break off.

[0058] (4) Upon receiving the decision command, the end effector is driven to a specified height via the lifting mechanism, and its horizontal relative position is adjusted via the horizontal sliding mechanism. Then, the posture of the robotic arm is adjusted so that the end effector moves to the mango to be picked. The multi-frequency resonant energy induction module 9 emits a resonant wave with the same frequency and opposite phase as the natural vibration frequency of the fruit stalk. The phase cancellation interference principle is used to instantly weaken the cell binding force at the molecular level. At the same time as the resonant wave is applied, the flexible fruit clamping component applies a preset minimum tension to the fruit stalk. At this time, the fruit stalk is cleanly separated at the energy weakening point, just like natural delamination, with almost no juice seepage. This closed-loop collaboration of "detection-induction-separation" eliminates the risk of juice burn and physical impact caused by traditional cutting in principle, realizes damage-free harvesting without juice seepage, and avoids the juice flowing down the broken fruit stalk burning the mango peel.

[0059] Specifically, the multi-frequency resonant energy induction module 9 consists of four sub-modules: Signal generation and frequency synthesis unit (such as DDS chip): accurately generates electrical signals of the required frequency and phase according to decision parameters.

[0060] Power amplifier unit (such as Class D power amplifier): amplifies weak electrical signals to a level sufficient to drive the transducer.

[0061] Energy conversion unit (multi-frequency transducer array): converts electrical signals into mechanical vibration waves and emits them directionally to the fruit stalk.

[0062] Control and communication unit (MCU + communication interface): coordinates timing, receives decision parameters, and realizes closed-loop fine-tuning.

[0063] Specifically, the preset minimum tensile force is a dynamic minimum threshold force, which can be calibrated through experiments: first measure the natural breaking force of the fruit stem, then set the preset tensile force to 1 / 10 to 1 / 5 of that value (for example, breaking force 5N → tensile force 0.5–1N), and ensure that it is lower than the damage threshold of the mango peel (refer to the visible bruise energy threshold of immature mango, which is about 0.25Nm).

[0064] (5) The fruit after separation is received by the flexible fruit clamping component; the robotic arm 1 moves along the planned recovery trajectory and moves towards the flexible buffer collection mechanism through the horizontal sliding mechanism; after reaching the collection area, the robotic arm 1 adjusts its position so that the fruit is directly above the flexible guide funnel 13; the flexible fruit clamping component is released, and the fruit is guided by the guide funnel to fall into the fruit basket with built-in multi-layer elastic buffer rope net, achieving a soft landing; at the same time, the collection and identification camera 14 integrated at the entrance performs real-time damage detection and grading of the fruit skin.

[0065] Furthermore, during the harvesting process, a dynamic digital twin is established for each fruit to be harvested within the work area. The digital twin includes the fruit's three-dimensional position, detachment preparation state index, estimated weight, peel strength, and a damage risk-benefit game function calculated based on historical data.

[0066] When the harvesting equipment enters a work area, the scheduling engine sends a harvesting request to the digital twins of all fruits in that area; the twins of fruits with high maturity and low risk of damage offer higher prices and are required to be harvested first; the twins of fruits located deep in the canopy and with high harvesting risk offer lower prices and are temporarily rejected for harvesting, waiting for a better opportunity or a gentler harvesting strategy later.

[0067] The scheduling engine dynamically determines the next harvesting target and plans the optimal movement path based on the received bidding results, the robot's own position, and energy consumption status, thereby maximizing group collaboration and overall harvesting efficiency. This avoids the high damage rate that can result from blindly harvesting in a fixed order, achieving dynamic optimization of the harvesting sequence and maximizing overall harvesting efficiency while ensuring fruit quality.

[0068] Specifically, the damage risk-reward game function is a composite evaluation function defined by combining multi-agent game theory, risk decision-making theory, and agricultural digital twin technology. Essentially, it is a utility function or payoff function that, within the framework of multi-agent game theory, dynamically calculates a comprehensive decision score for each fruit to be harvested, used to evaluate the overall "cost-effectiveness" of harvesting that fruit at the current moment.

[0069] In game theory and auction theory, multi-agent task allocation problems are typically solved using auction mechanisms—treating tasks as "commodities" and having robots act as "bidders" to submit bids, or prioritizing tasks based on their utility value, thereby achieving dynamic task allocation and global optimization. The game function in this invention is the core algorithm that provides the "bidding" basis for this auction mechanism.

[0070] The data foundation for a digital twin includes the fruit's three-dimensional location, detached readiness index, estimated weight, and peel strength. The game function integrates and calculates this foundational data, typically consisting of two core components: (a) Income Components The positive value that can be obtained from harvesting this fruit includes, but is not limited to: Maturity Benefits: The higher the detachment readiness index, the greater the yield of the fruit within the optimal harvest window.

[0071] Weight gain: The larger the estimated weight of the fruit, the higher the output value per unit of time / cost.

[0072] Quality premium: Fruits with moderate peel strength and no surface defects can obtain a higher market grade premium.

[0073] (ii) Damage risk components Quantify the expected losses that may result from harvesting this fruit, including: Harvesting risks: The fruit is located deep in the tree canopy and surrounded by dense branches and leaves, making it difficult to operate the robotic arm and posing a high risk of collision.

[0074] Fruit peel damage sensitivity: Fruits with low peel strength (i.e. thin-skinned varieties) are more sensitive to mechanical contact and have a higher probability of being damaged.

[0075] Risk of separation failure: When the separation readiness index is too low, energy-induced separation may fail, requiring a second operation or causing the fruit stalk to tear.

[0076] The damage risk-benefit game function can be expressed as: U = Reward - λ × Risk; Where U is the game function value (i.e. the “bidding” score of the fruit), Reward is the normalized expected return (0-1 interval), Risk is the normalized harvest damage risk probability (0-1 interval), and λ is the risk aversion coefficient, which can be dynamically adjusted according to harvesting strategy preferences (λ>1 is a risk-averse strategy, λ<1 is a risk-seeking strategy).

[0077] Within the framework of multi-agent game theory, this function reflects the trade-off between gains and risks, which is consistent with the classic expected utility theory in game theory—when choosing an action, an agent maximizes its expected utility, and expected utility = gain × probability of success.

[0078] The "auction bidding" mechanism driven by the damage risk-reward game function is as follows: Bid Calculation: For each fruit in the work area, a digital twin is used to calculate a "bid" (i.e., U-value) based on data such as the fruit's current detachment preparation status, peel strength, and location accessibility.

[0079] Priority ranking: Fruits with higher bids should be harvested first (higher maturity, lower risk → higher profit, lower risk → higher U value).

[0080] Low bid decision: Fruits with low bids (such as immature fruits, fruits located deep in the canopy, and fruits with high harvesting risk) can be "rejected" for harvesting at the present time and wait for a better opportunity later (for example, as immature fruits gradually enter the ripening period over time, the detachment readiness index increases, and the U value also increases accordingly).

[0081] Global scheduling: The scheduling engine comprehensively considers the bidding results of each fruit, the robot's own position and energy consumption status, dynamically decides the next harvesting target, and plans the optimal movement path.

[0082] This mechanism is consistent with the principle of the "resource-based auction algorithm" proposed in existing research—modeling multi-robot tasks, analyzing task-related models and task energy indicators, and comprehensively considering factors such as distance constraints, task execution capability constraints, and resource consumption when allocating tasks in the auction algorithm, so as to reduce the execution cost of the entire system and increase the task completion rate.

[0083] Furthermore, the entire harvesting process's data is uploaded to a database, including micro-vibration spectral response, energy-induced parameters, game theory bidding results, and actual damage detection results output by the high-speed visual re-inspection unit. This data is used to iteratively optimize the detached stem difficulty fingerprint, the parameters of the classification and regression models, and the game function and negotiation strategy of the digital twin, enabling the harvesting equipment to possess self-learning and continuous evolution capabilities. As the amount of data increases, the system's judgment of the detached stem state becomes increasingly accurate, the prediction of energy-induced parameters becomes increasingly optimized, and the game theory scheduling strategy becomes increasingly rational. This allows the harvesting equipment to possess self-learning and continuous evolution capabilities, enabling it to adapt to the harvesting needs of different mango varieties, orchards, and years after long-term operation.

[0084] Specifically, the classification and regression model is a multi-output machine learning model of the present invention. Its input is a fusion feature vector of micro-vibration spectrum data and visual positioning information, and its output is a structured decision instruction containing binary classification results and multiple regression parameters.

[0085] Classification and Regression (CART) is a classic term in machine learning—classification is used to predict discrete class labels (such as "harvestable / unharvestable"), while regression is used to predict continuous values ​​(such as resonance frequencies and energy parameters). The CART algorithm was proposed by Breiman et al. in 1984. Its core principle is: when the dependent variable of the dataset is discrete, a classification tree is built; when the dependent variable is continuous, a regression tree is built. Each non-leaf node can only derive two branches, forming a binary tree structure. Random forests build upon CART by randomly sampling samples and explanatory variables to create N independent CART models. Finally, the results of the N models are averaged or combined to obtain the final predicted value.

[0086] The core of this invention lies in the fact that traditional CART models output a single type of result after a given input, while the classification and regression model of this invention simultaneously outputs a binary classification result (whether it is harvestable) and multiple regression parameters (resonance frequency, phase shift, and duration of action), belonging to a multi-task learning framework. The basic idea of ​​multi-task learning is to allow multiple related tasks to share a feature extraction layer, establishing multiple task branches on the shared representation, each outputting different types of prediction results. For example, a multi-task geometric regression network in maize-soybean intercropping can simultaneously perform multiple tasks such as crop segmentation, row direction prediction, and missing seedling detection, demonstrating the feasibility of multi-task networks in agricultural scenarios.

[0087] The core components of the classification and regression model: 1. Input layer.

[0088] Receive the fused sensor data feature vector, including: Micro-vibration spectral response characteristics (resonance frequency, resonance peak width, phase characteristics, damping ratio).

[0089] Visual positioning and appearance information (three-dimensional position of fruit, morphological characteristics of fruit stalk).

[0090] 2. Shared feature extraction layer.

[0091] Deep neural networks or ensemble learning structures (such as random forests, support vector machines, and deep convolutional networks) are used to extract shared feature representations related to the off-site state from the original input.

[0092] 3. Classification branches.

[0093] Output a binary classification result of "harvestable" and "unharvestable". This branch has a rich application foundation in existing technologies: in the non-destructive spectral prediction of Fuji apple harvest maturity, the classification model based on SIQI+SVR achieved a classification accuracy of 85.71%; in the classification of winter jujube maturity, the DSAF-ResNet fusion model achieved a test accuracy of 97.24%; in the prediction of strawberry ripening quality based on hyperspectral technology, the SVM model showed high accuracy in predicting anthocyanins, firmness, and soluble solids.

[0094] 4. Regression branch.

[0095] The output consists of continuous numerical parameters, including: the pedicel's natural resonant frequency (Hz), phase shift (degrees or radians), energy induction duration (ms), and recommended resonant frequency (Hz). This branch also has mature applications in agricultural monitoring: the AgRegNet deep regression network is used to estimate the density, quantity, and location of orchard flowers and fruits; deep learning-based fruit weight estimation models use convolutional neural networks to predict the weight of apples, pears, oranges, and bananas from a single RGB photograph.

[0096] The workflow of the classification and regression model: 1. Feature Fusion: The visual data from the integrated camera and the spectral response data from the micro-vibration spectral generation and analysis module are aligned in time and space to construct a unified feature vector.

[0097] 2. Classification and Determination: The classification branch outputs a binary result of "recoverable / unrecoverable". If the result is unrecoverable, the current task is terminated; if the result is recoverable, the regression prediction stage begins.

[0098] 3. Regression Prediction: The regression branch outputs a structured decision vector of induced parameters, including the pedicel's intrinsic resonance frequency, phase characteristics, and recommended energy induced parameters.

[0099] 4. Command execution: The system sends the decision vector to the robotic arm (4) and the multi-frequency resonant energy induction module to perform the harvesting action.

[0100] The above are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above content. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A non-destructive mango plucking device for smart agriculture, characterized in that, It includes a walking mechanism, a lifting mechanism, a horizontal sliding mechanism, a robotic arm, an energy-induced end effector, and a flexible buffer collection mechanism; The walking mechanism is equipped with a walking recognition camera; the lifting mechanism is mounted on the walking mechanism; the horizontal sliding mechanism and the flexible buffer collection mechanism are mounted on the lifting mechanism; the robotic arm is mounted on the horizontal sliding mechanism; and the energy-induced end effector is mounted on the end of the robotic arm. The energy-induced end effector includes a flexible fruit-clamping component, an end-of-life recognition camera, a micro-vibration spectroscopy generation and analysis module, and a multi-frequency resonance energy-inducing module. Both the micro-vibration spectroscopy generation and analysis module and the multi-frequency resonance energy-inducing module are mounted on the flexible fruit-clamping component. The micro-vibration spectroscopy generation and analysis module is used for non-contact detection of the fruit stalk, acquiring the spectral response of the stalk cells at a specific frequency, and comparing it with an in vitro difficulty fingerprint to determine whether the optimal harvest period has been reached. If so, a decision command containing the stalk's inherent resonant frequency and phase characteristics is output. The multi-frequency resonance energy-inducing module emits a resonant wave with the same frequency but opposite phase as the stalk's inherent vibration frequency according to the decision command. Utilizing the phase cancellation interference principle, it instantaneously weakens cell binding forces at the molecular level, allowing the fruit to separate from the parent plant without damage.

2. The intelligent agricultural harvesting equipment for non-destructive mango separation according to claim 1, characterized in that, The horizontal sliding mechanism, robotic arm, energy-induced end effector, and flexible buffer collection mechanism are each provided in two sets.

3. The intelligent agricultural harvesting equipment for non-destructive mango separation according to claim 1, characterized in that, The flexible fruit clamping assembly includes a pair of flexible clamps and a fruit clamping drive motor for driving the flexible clamps to open and close. The micro-vibration spectrum generation and analysis module and the multi-frequency resonance energy induction module are both mounted on the flexible clamps.

4. The intelligent agricultural harvesting equipment for non-destructive mango separation according to claim 1, characterized in that, The traveling mechanism includes a track assembly and a frame, and the lifting mechanism is mounted on the frame.

5. The intelligent agricultural harvesting equipment for non-destructive mango separation according to claim 1, characterized in that, The lifting mechanism is composed of a scissor lift mechanism.

6. The intelligent agricultural harvesting equipment for non-destructive mango separation according to claim 1, characterized in that, The horizontal sliding mechanism includes an X-axis sliding mechanism and a Y-axis sliding mechanism. The X-axis sliding mechanism is mounted on the lifting mechanism, and the Y-axis sliding mechanism is mounted on the X-axis sliding mechanism.

7. The intelligent agricultural harvesting equipment for non-destructive mango separation according to claim 1, characterized in that, The flexible buffer collection mechanism includes a fruit collection basket, a flexible flow guide funnel, and a collection identification camera. The inner cavity of the fruit collection basket is equipped with multiple layers of elastic buffer rope netting. The flexible flow guide funnel is located on top of the fruit collection basket, and the collection identification camera is located on top of the flexible flow guide funnel for real-time damage detection and grading of the fruit skin.

8. A non-destructive, intelligent agricultural harvesting method for mangoes, characterized in that, Includes the following steps: The walking mechanism moves autonomously within the orchard, acquiring environmental information through a walking recognition camera. By linking the lifting mechanism and the horizontal sliding mechanism, the energy-induced end effector is moved to the vicinity of the target fruit in the three-dimensional space of the tree canopy; the end recognition camera scans the fruit group to obtain the spatial distribution and preliminary appearance information of the mangoes; The micro-vibration spectrum generation and analysis module emits low-energy micro-vibration waves of a specific frequency to the pedicel of the target mango and receives the spectral response of the reflected signal. The spectral response is compared with a pre-established in vitro difficulty fingerprint spectrum to quantitatively assess the degree of degradation of the cellulose and hemicellulose structures of the pedicel cells and determine whether the pedicel has reached the mutation state of the optimal harvest period. If so, a decision command containing the pedicel's inherent resonant frequency, phase characteristics, and energy-induced parameters is output. Upon receiving the decision command, the lifting mechanism drives the end effector to a designated height, the horizontal sliding mechanism adjusts the relative horizontal position of the end effector, and the posture of the robotic arm is adjusted to move the end effector to the mango to be picked. The multi-frequency resonant energy induction module emits a resonant wave with the same frequency but opposite phase as the natural vibration frequency of the fruit stalk, using the phase cancellation interference principle to instantaneously weaken the cell binding force at the molecular level. At the same time as the resonant wave is applied, the flexible fruit clamping component applies a preset micro-pulling force to the fruit stalk, causing the fruit to detach from the energy weakening point of the stalk, achieving damage-free harvesting without sap leakage. The separated fruit is received by the flexible fruit clamping component; the robotic arm moves along the planned recycling trajectory and moves towards the flexible buffer collection mechanism through the horizontal sliding mechanism; after reaching the collection area, the robotic arm adjusts its posture so that the fruit is directly above the flexible guide funnel. The flexible fruit clamping component releases, and the fruit is guided by the guide funnel to fall into the fruit basket with built-in multi-layer elastic buffer rope net, achieving a soft landing; at the same time, the collection and identification camera integrated at the entrance performs real-time damage detection and grading of the fruit skin.

9. The intelligent agricultural harvesting method for non-destructive mango separation according to claim 8, characterized in that, During the harvesting process, a dynamic digital twin is created for each fruit to be harvested in the work area. The digital twin includes the three-dimensional position of the fruit, the detachment preparation state index, the estimated weight, the peel strength, and the damage risk-benefit game function calculated based on historical data. When the harvesting equipment enters a work area, the scheduling engine sends a harvesting request to the digital twins of all fruits in that area; the twins of fruits with high maturity and low risk of damage offer higher prices and are required to be harvested first; the twins of fruits located deep in the canopy and with high harvesting risk offer lower prices and are temporarily rejected for harvesting, waiting for a better opportunity or a gentler harvesting strategy later. The scheduling engine dynamically decides the next harvesting target and plans the optimal movement path based on the received bidding results, the robot's own position and energy consumption status, thereby maximizing group collaboration and overall harvesting efficiency.

10. The intelligent agricultural harvesting method for non-destructive mango separation according to claim 9, characterized in that, The entire data of each harvesting operation is uploaded to the database. The entire data includes micro-vibration spectrum response, energy-induced parameters, game bidding results, and actual damage detection results output by the high-speed visual re-inspection unit. The entire data is used to iteratively optimize the parameters of the ex vivo difficulty fingerprint spectrum, classification and regression model, and the game function and negotiation strategy of the digital twin, so that the harvesting equipment has the ability to learn and continuously evolve.

Images