CN-122004046-A - Rigid-flexible coupling sequence driving picking device and intelligent control method thereof

CN122004046ACN 122004046 ACN122004046 ACN 122004046ACN-122004046-A

Abstract

The invention discloses a rigid-flexible coupling sequence driving picking device and an intelligent control method thereof, belonging to the technical field of agricultural robots and intelligent control, and comprising a double-layer X-shaped telescopic driving mechanism, a driving rope and an integrated rigid-flexible coupling terminal actuator; the double-layer X-shaped mechanism automatically realizes Z-axis positioning, driving force transmission and terminal sequence action switching by means of a single motor and mechanical logic, a terminal executor completes clamping and breaking actions by means of single rope driving through rigidity gradient design, and a control method constructs a deep learning sensorless architecture, wherein the deep learning sensorless architecture comprises the steps of estimating a terminal state by a forward state estimation network, identifying breakage by a fruit stem breakage detection network millisecond level, reversely controlling a decision network to generate an optimal instruction, and adapting a time-varying factor by an online learning mechanism. The rigid-flexible coupling sequence driving picking device and the intelligent control method thereof provided by the invention have the advantages of simple structure, improved picking efficiency and reliability, and suitability for large-scale picking of various fruits.

Inventors

LI CHEN
QIN HAN
QIAO YUFEI
YU JINLONG

Assignees

西北农林科技大学

Dates

Publication Date: 20260512
Application Date: 20260210

Claims (9)

1. A rigid-flexible coupling sequence driving picking device is characterized by comprising a driving mechanism, a driving rope and an integrated rigid-flexible coupling terminal actuator, wherein the driving mechanism is fixedly arranged on a movable base, the driving rope is connected with the driving mechanism and the integrated rigid-flexible coupling terminal actuator, and the integrated rigid-flexible coupling terminal actuator is a continuous flexible structure body which is integrally formed and is arranged above the driving mechanism.
2. The rigid-flexible coupling sequence driving picking device according to claim 1 is characterized in that the driving mechanism adopts a double-layer X-shaped telescopic mechanism and comprises an outer-layer X-shaped mechanism, an inner-layer X-shaped mechanism and a driving motor, the driving motor is arranged in a reserved fixed position of a moving base, the lower end of the outer-layer X-shaped mechanism is fixed on the moving base, the upper end of the outer-layer X-shaped mechanism is a moving end and stretches in the vertical direction, the inner-layer X-shaped mechanism is nested in the outer-layer X-shaped mechanism, the size of the inner-layer X-shaped mechanism is slightly smaller than that of the outer-layer X-shaped mechanism, the lower end of the inner-layer X-shaped mechanism is fixed on the moving base, the lower limit of a limit trigger mechanism is arranged nearby, the upper end of the inner-layer X-shaped mechanism is connected with an integrated rigid-flexible coupling terminal actuator, the upper limit of the limit trigger mechanism is arranged nearby, and a damping mechanism is arranged between hinge nodes of the inner-layer X-shaped mechanism.
3. A rigid-flexible coupling sequence driven picking device as defined in claim 2 wherein the outer X-shaped mechanism and the inner X-shaped mechanism are diamond-shaped telescoping structures.
4. The rigid-flexible coupling sequence driven picking device according to claim 2, wherein the integrated rigid-flexible coupling end effector comprises a support member, a wrist elastic sheet, a paw elastic sheet and flexible fingers, wherein the lower ends of the wrist elastic sheet and the support member are connected with an upper platform of an inner layer X-shaped mechanism, the upper ends of the wrist elastic sheet and the support member are connected with the flexible fingers, and the periphery of the paw elastic sheet is connected with the flexible fingers.
5. The rigid-flexible coupling sequence driven picking device of claim 4, wherein the driving ropes are two, one end of each driving rope is connected with the elastic claw sheet, the other end of each driving rope is connected with the outer layer X-shaped mechanism, and the other end of each driving rope is respectively connected with the elastic wrist sheet and the outer layer X-shaped mechanism.
6. An intelligent control method applied to a rigid-flexible coupling sequence driven picking device as claimed in claims 1-5, comprising the steps of: s1, designing and training a forward state estimation network, a fruit stem fracture detection network and a reverse control decision network; S2, data acquisition, namely acquiring driving motor signal time sequence information and RGB-D visual images in real time, wherein the driving motor signal time sequence information is driving motor current, voltage and corner signals in a period of time window; s3, fruit identification, namely identifying the position, the category and the maturity of the fruit through a deep convolutional neural network according to the RGB-D visual image, estimating the hardness and the weight of the fruit, and determining an expected target, wherein the expected target comprises an expected paw closing degree, an expected clamping force, an expected breaking moment and an expected operation stage; S4, estimating a forward state, namely estimating the current terminal state, including the paw closure degree, the actual clamping force, the wrist bending angle, the breaking moment, the fruit stem breaking probability and the operation stage mark, according to the driving motor signal time sequence acquired in the S2 and the fruit information acquired in the S3 through a forward state estimation network; s5, reverse control decision and execution, namely generating an instruction through a reverse control decision network according to the operation stage identification by combining a forward state estimation result and an expected target, and driving a motor by an execution control module, wherein the method comprises the following specific steps of: The ascending positioning stage, namely generating a forward rotation instruction of a driving motor according to a Z-axis height target identified by fruits by a reverse control decision network, and enabling an integrated rigid-flexible coupling terminal executor to reach the lower part of the fruit trees; The clamping stage comprises the steps of adopting force closed-loop control, monitoring and estimating the clamping force in real time, and adjusting the current of a driving motor according to the error of the actual clamping force and the expected clamping force until the clamping force reaches a target value; the breaking-off stage comprises the steps that an execution control module adopts moment gradual control to gradually increase the breaking-off moment from zero, and the increasing rate is adjusted according to the fruit characteristics; The breaking response and fruit receiving stage comprises that after the fruit stem breaking detection network detects breaking, the execution control module immediately sends out a stop instruction, drives the motor to brake emergently, and then switches to the ascending fruit receiving stage; The reset stage comprises adopting constant speed control, driving a motor to rotate positively at a constant speed, enabling the outer X-shaped mechanism to return to an upper limit, enabling the inner X-shaped mechanism to reset synchronously, enabling wrists to return to normal, enabling the claws to be opened completely, enabling fruits to fall off, and waiting for the next picking cycle; S6, online learning and self-adaption, namely automatically recording complete data sequences and result evaluation after each picking cycle is finished, adding new data into a training set according to a set period, fine-tuning network parameters with a small learning rate, updating each network module, evaluating the performance of a model after fine tuning, updating the network model in a main controller after precision improvement, and enabling time-varying factors to comprise creep of self-adaptive materials of the system, mechanical abrasion and fruits of new varieties to keep long-term stable control.
7. The intelligent control method for a rigid-flexible coupling sequence driven picking device according to claim 6, wherein the designing of the forward state estimation network in step S1 specifically comprises: The method comprises the steps of 1, designing an input module, wherein the input module divides input branches according to signal types, and comprises an electric signal branch, a visual branch and a fruit characteristic branch, wherein the electric signal branch inputs a motor signal time sequence, and comprises historical current, voltage, a corner and derivatives thereof; Step 2, designing an output module, namely dividing output into three types of regression task output, probability task output and classification task output according to the data type of the output state, wherein the regression task output comprises a paw closure degree, a clamping force, a wrist bending angle and a breaking moment; Step 3, building a network architecture, namely building a multi-branch input, feature fusion and multi-task output architecture, wherein an electric signal branch processes a time sequence by using a one-dimensional convolutional neural network or a long-short-period memory network, a visual branch extracts image features by using the convolutional neural network, a fruit attribute branch processes by using a full-connection layer, each branch feature is fused through a splicing or attention mechanism, then a high-level semantic feature is extracted through a multi-layer full-connection network, and finally each target variable is respectively predicted; step 4, setting a loss function, namely regressing the mean square error or the smooth loss for the task, classifying the cross entropy loss for the task, and enabling the total loss to be a weighted sum of the losses of the tasks; And 5, training a network by adopting supervised learning, acquiring training data by temporarily installing auxiliary sensors on a picking device, recording motor signals, visual information, fruit attributes and real labels measured by the sensors, acquiring a large number of samples which cover different fruits, different environments and different mechanical states, expanding the samples by data enhancement, training on a training set by using a standard deep learning training process, monitoring performance on a verification set, evaluating generalization capability on a test set, removing all the auxiliary sensors after training is finished, and reasoning according to the motor signals and the visual information.
8. The intelligent control method for driving a picking device by a rigid-flexible coupling sequence according to claim 6, wherein designing the stem breakage detection network in step S1 specifically comprises: Step 1, designing network input, namely extracting current, power and angular acceleration of a driving motor in a short-time signal window as input data; Step 2, designing network output, wherein the network output is binary classification or continuous probability; step 3, constructing a network architecture of a bidirectional LSTM, transformer encoder or a one-dimensional convolution network, and automatically capturing mutation characteristics; step 4, designing a loss function, namely using class weighting or a special loss function to process a sample imbalance problem; setting a network evaluation standard, namely taking recall rate, accuracy rate and detection delay as core indexes as training targets and generalization capability judgment bases; Step 6, training a network, namely adopting supervised learning, wherein training data are current, power and angular acceleration of a driving motor in a short-time signal window, manually marking or assisting a sensor to match a real label for the signal window, using the loss function of the step 4 to treat the problem of sample imbalance, dividing the samples of the signal window and the real label into a training set, a verification set and a test set according to a proportion, using a standard deep learning training process, selecting an Adam optimizer, setting proper learning rate and batch size, inputting the training set into the network, training for a plurality of times until convergence, monitoring performance on the verification set, calculating the index of the step 5 on the test set after the training is completed, and completing model training after reaching standards, otherwise, reworking optimization.
9. The intelligent control method for a rigid-flexible coupling sequence driven picking device according to claim 6, wherein the designing of the reverse control decision network in step S1 specifically comprises: the method comprises the steps of 1, designing network input, wherein input data comprise a forward state estimated value obtained in the step S3, an expected target value obtained in the step S2, a state error of the forward state estimated value and the expected target value and a history control instruction; step 2, designing network output, wherein the output data comprises a current instruction, a speed instruction and a stop signal; step 3, constructing a multi-layer full-connection network or a circulating neural network; And 4, performing network training, wherein the training method comprises a supervised learning, reinforcement learning or mixing method.

Description

Rigid-flexible coupling sequence driving picking device and intelligent control method thereof Technical Field The invention relates to the technical field of agricultural robots and intelligent control, in particular to a rigid-flexible coupling sequence driving picking device and an intelligent control method thereof. Background Fruit picking is used as labor intensive agricultural operation, and China is used as the world maximum fruit producing country, and faces the outstanding problems of high picking labor occupation ratio, high labor cost and seasonal labor shortage. Picking robots are key technical approaches for breaking the dilemma, but the prior art has two major core bottlenecks, which severely restrict commercial application: the bottleneck of the driving mechanism is that the existing picking device mostly adopts a multi-driving-source scheme, and an independent motor is configured for clamping, separating, approaching and other actions, so that the tail end is large in quality, high in inertia, low in picking efficiency, complex in structure, poor in reliability, high in control coordination difficulty and high in cost, and the problems of dependence on a complex air source, slow response, low in precision, or difficult control of spring parameters, unstable in time sequence, poor in adaptability and the like exist in a part of pneumatic or underactuated scheme, so that the integrated mechanical logic driving of the lower driving source and the multi-action sequence is not realized. The control method is bottleneck in that the picking device has long transmission chain and rigid-flexible coupling, mapping relations of states such as motor signals, terminal clamping force, breaking moment and the like show high-dimensional nonlinearity, time variability and the like, traditional physical modeling errors are seriously accumulated, the end sensor scheme has the defects of high cost, poor reliability, serious drift, response delay and the like, the traditional machine learning method has limited expression capacity, and a deep learning control technology for the complex system, particularly a multi-mode fusion, transient fracture detection and online self-adaptive learning related scheme, has not formed effective breakthrough. Disclosure of Invention The invention aims to provide a rigid-flexible coupling sequence driving picking device and an intelligent control method thereof, which solve the technical problems of complex structure, large tail end inertia, high cost, poor reliability, difficult force control under the condition of a complex transmission chain and the like caused by multiple driving sources of the existing picking robot. The invention provides a rigid-flexible coupling sequence driving picking device which comprises a driving mechanism, a driving rope and an integrated rigid-flexible coupling terminal actuator, wherein the driving mechanism is fixedly arranged on a movable base, the driving rope is connected with the driving mechanism and the integrated rigid-flexible coupling terminal actuator, and the integrated rigid-flexible coupling terminal actuator is an integrally formed continuous flexible structure body and is arranged above the driving mechanism. The driving mechanism is preferably a double-layer X-shaped telescopic mechanism and comprises an outer-layer X-shaped mechanism, an inner-layer X-shaped mechanism and a driving motor, the driving motor is arranged in a reserved fixed position of the moving base, the lower end of the outer-layer X-shaped mechanism is fixed on the moving base, the upper end of the outer-layer X-shaped mechanism is a moving end and stretches in the vertical direction, the inner-layer X-shaped mechanism is nested in the outer-layer X-shaped mechanism, the size of the inner-layer X-shaped mechanism is slightly smaller than that of the outer-layer X-shaped mechanism, the lower end of the inner-layer X-shaped mechanism is fixed on the moving base and is provided with the lower limit of a limit trigger mechanism nearby, the upper end of the inner-layer X-shaped mechanism is connected with an integrated rigid-flexible coupling terminal actuator and is provided with the upper limit of the limit trigger mechanism nearby, and a damping mechanism is arranged between hinge nodes of the inner-layer X-shaped mechanism. Preferably, the outer layer X-shaped mechanism and the inner layer X-shaped mechanism are diamond-shaped telescopic structures. Preferably, the integrated rigid-flexible coupling end effector comprises a wrist elastic sheet, a paw elastic sheet and flexible fingers, wherein the lower end of the wrist elastic sheet is connected with an upper platform of the inner-layer X-shaped mechanism, the upper end of the wrist elastic sheet is connected with the flexible paws, and the periphery of the paw elastic sheet is connected with the flexible fingers. Preferably, the driving rope has two driving ropes, one end of the driving rope is connected with