3D printing product sorting system based on multi-cascade grabbing network framework

Abstract

The invention provides a 3D printing product sorting system based on a multi-cascade grabbing network framework. The 3D printing product sorting system comprises a control module, a sorting platform used for containing 3D printing products and a mechanical arm used for sorting the 3D printing products. A camera connected with the control module is arranged above the sorting platform; the method comprises the following steps of: acquiring an image of a 3D printing product through a camera, generating a high-resolution image through a super-resolution reconstruction technology, processing the image by using an improved YOLOv5 + E-A-SPPFCSPC + BiFPN instance segmentation algorithm, accurately positioning a target 3D printing product to be grabbed, and controlling the grabbing operation of a mechanical arm through an AFFGA grabbing detection network; the track of the mechanical arm is dynamically planned through the optimized particle swarm algorithm, it is guaranteed that the grabbing network finds the optimal path to reach the target grabbing position within limited time, and meanwhile the smoothness of the track is guaranteed; according to the method, dynamic planning optimization of the grabbing track of the mechanical arm is combined, all links of sorting of the 3D printing products can be comprehensively optimized, and high-definition imaging, accurate positioning and intelligent decision making are achieved.

Description

Technical Field

[0001] The present invention relates to the field of mechanical control technology, and in particular to a 3D printing product sorting system based on a multi-stage grabbing network framework. Background Art

[0002] The rise of additive manufacturing has brought opportunities for mass customization and rapid design changes, but when the manufacturing industry hopes to achieve robotic automation, the complex geometry of parts makes the gripper face challenges in design and planning. How to solve this problem has become a research direction. Summary of the invention

[0003] The present invention proposes a 3D printed product sorting system based on a multi-level grasping network framework. The system is a highly integrated and intelligent 3D printed product sorting system that can fundamentally cope with the multiple challenges of accuracy, efficiency, safety, etc. faced in the field of 3D printed product distribution. The core of the system design is to build a multi-level grasping network that integrates SRCNN, improved YOLOv5 instance segmentation module and enhanced AFFGA network, combined with dynamic planning optimization of the robot arm grasping trajectory, to comprehensively optimize all aspects of 3D printed product sorting, and achieve high-definition imaging, precise positioning, and intelligent decision-making.

[0004] The present invention adopts the following technical solutions.

[0005] A 3D printed product sorting system based on a multi-stage grasping network framework, the sorting system comprises a control module, a sorting platform for placing 3D printed products and a mechanical arm for sorting 3D printed products; a binocular camera connected to the control module is arranged above the sorting platform; the control module firstly collects the original 3D printed product image through the binocular camera, generates a high-resolution image through super-resolution reconstruction technology, and then uses an improved YOLOv5+EA-SPPFCSPC+BiFPN instance segmentation algorithm to process the image, accurately locates the target 3D printed product to be grasped, and controls the grasping operation of the mechanical arm through an AFFGA grasping detection network; and dynamically plans the trajectory of the mechanical arm through an optimized particle swarm algorithm, so as to ensure that the grasping network finds the best path to reach the target grasping position within a limited time, and at the same time ensures the smoothness of the trajectory, thereby minimizing the potential risks in the grasping process.

[0006] The 3D printed products are placed on the grabbing platform, and some of the 3D printed products are in a stacked state.

[0007] The sorting system improves the gripping capability of the robot arm for chaotic, overlapping and variously shaped 3D printed products by using a gripping framework model that combines a multi-level gripping network with an adaptive robot mechanism;

[0008] The working method of the grasping framework model includes: firstly, preprocessing the original image through the improved super-resolution convolutional neural network algorithm SRCNN, and then using the YOLOv5+EA-SPPFCSPC+BIFPNC instance segmentation algorithm, that is, the YOLO-EASB algorithm, to accurately segment the 3D printed products, and then determine the most suitable 3D printed products for grasping by evaluating the integrity of the segmentation mask, and then through the improved adaptive feature fusion and grasping perception network IAFFGA-Net, combined with the optimized loss function, these segmented 3D printed products are processed to ensure high-precision grasping operations even in complex environments.

[0009] The workflow of the improved super-resolution convolutional neural network algorithm SRCNN includes:

[0010] Step A1, operation of feature extraction layer Conv(5, d, 1): a convolution kernel of size 5×5 is used to replace the convolution kernel of 9×9 in SRCNN for the original low-resolution image, and a feature extraction operation is performed to achieve the effect of small information loss and high speed. The number of channels c i is 1, the number of convolution kernels is n i is d;

[0011] Step A2, the operation of the contraction layer Conv(1, s, d). A convolution kernel of size 1×1 is used, and a small number of convolution kernels s is used to reduce the LR feature dimension d to s, so as to save computational cost;

[0012] Step A3, operation of the nonlinear mapping layer Conv(3, s, s): If SRCNN lacks deep network experiments, in order to maintain the same performance as SRCNN, s convolution kernels of size 3×3 are used to replace the original wide convolution kernel, and the number of channels c is maintained. i for s;

[0013] Step A4, operation of the expansion layer Conv(1, d, s): If the contraction layer reduces the LR feature dimension, which will reduce the quality of image restoration, then add a expansion layer after the nonlinear mapping to expand the HR feature dimension; use s convolution kernels of size 1×1 for the expansion operation, which is consistent with the contraction layer;

[0014] Step A5, operation of the deconvolution layer DeConv(9, 1, s): deconvolution is used as the inverse operation process of convolution. When the step size is k, the resolution is increased by k times, and a deconvolution kernel of size 9×9 is used.

[0015] In the YOLOv5+EA-SPPFCSPC+BIFPNC instance segmentation algorithm, based on the multi-scale features generated by SPP, the channel attention mechanism ECA (ECA, Efficient Channel Attention) is further integrated, so that the model can intelligently allocate the response weight of each channel, that is, when the 3D printed product is partially occluded, the key features of the non-occluded area are highlighted, so as to facilitate the detection of the occluded 3D printed products; in addition, in order to fully explore the complementary advantages between deep-level features and shallow-level features, the Adaptive Dual Attention Feature Fusion (ADaFF) module ADaFF is adopted; the ADaFF module is designed to efficiently integrate cross-level features and suppress redundancy, so that the model can not only extract high-level abstract features from the deep network, but also make full use of the delicate texture information of the shallow input.

[0016] The ADaFF module adds the input features and the residual features channel by channel, and then guides these mixed features to undergo two sub-processes: local attention and global attention. In the local attention sub-process, the ECA mechanism is used to enhance the expressiveness of local features, while the global attention enhances the integrity of features by capturing global contextual information. The output features of the attention mechanisms of the two sub-processes interact through a feature fusion strategy to achieve adaptive integration of features. Finally, in the output stage, local and global features are weightedly merged after independent weight calculation to generate the final output features.

[0017] In the YOLOv5+EA-SPPFCSPC+BIFPNC instance segmentation algorithm, BIFPNC is a weighted bidirectional feature pyramid network model BiFPN. In this model, the strategy of enhancing feature fusion through jump connection is used to utilize shallow features and deep features to improve the perception of the details of the target 3D printed product and optimize the perception of the segmentation of occluded objects.

[0018] Assume that the feature maps of different scales are x = [x0, x1, x2], the weights [w0, w1, w2] are learnable parameters obtained through model training, and ∈ is a small constant; the steps for fusion are as follows:

[0019] Weight normalization, the formula used is: Weighted feature fusion, the formula used is y = w′0·x0+w′1·x1+w′2·x2Formula 2;

[0020] The activation after weighted feature fusion uses the formula y′=SiLU(y) Formula 3;

[0021] Feature concatenation, the formula used is z = concat (x0, y′) Formula 4;

[0022] The convolution operation uses the formula output = Conv2d (z) Formula 5.

[0023] If the proportion of effective grasping points of 3D printed products in the visual input is small, the BCE in the AFFGA grasping detection network is upgraded to FocalLoss; FocalLoss is defined by the formula:

[0024] FL(p t )=-α t (1-p t )γlog(p t ) Formula 6;

[0025] Among them, p t is the probability of the model predicting the true class label, α t is the category balance weight, which is used to adjust the ratio of the importance of positive and negative samples, and is set according to the actual situation; and γ is an adjustable focusing parameter, which smoothly determines the speed at which the loss weight of easy-to-resolve samples decreases;

[0026] During the training process, as p t The closer it is to 1, the lighter the penalty FocalLoss imposes on correctly classified samples, making the model training pay more attention to and strengthen the learning of difficult samples, so that FocalLoss can dynamically adjust the learning speed of difficult and easy samples, enhance the adaptability and stability of the model when dealing with significant changes in grasping posture, complex object shapes or severe occlusion, reduce the overfitting phenomenon caused by uneven data distribution, and thus optimize the execution of complex object grasping tasks.

[0027] The optimized particle swarm algorithm dynamically programs the trajectory of the robot arm as follows: first, a 3-5-3 piecewise polynomial interpolation function is constructed, and then the optimization objective function is constructed using three interpolation times. Subsequently, the improved particle swarm algorithm is used to optimize each interpolation time with the goal of minimizing the total running time while satisfying the constraints.

[0028] The optimized particle swarm algorithm improves the Camellia pollen picking algorithm of the standard particle swarm algorithm and uses a nonlinear decreasing strategy to adjust the value of the inertia weight ω, that is, the expression is:

[0029]

[0030] In the formula, ω max ,ω min (omega max , omega min) are the maximum and minimum inertia weights, respectively. In this paper, they are 0.86 and 0.44; n is the current iteration number; N is the maximum evolutionary generation. The dynamic method is used to assign values ​​to c1 and c2. The formula is:

[0031]

[0032]

[0033] Among them, c 11 , c 21 is a constant;

[0034] In the initial stage, IPSO randomly generates M three-dimensional particles and obtains the initial position and velocity of each particle, which leads to uneven distribution of population initialization and affects the optimization of the algorithm. Based on IPSO, a double-mixed initialization sequence based on Logitic-Tent is introduced, and Logisfic mapping (taking μ = 1) is used to generate i chaotic sequences chx η : Where t=1, 2, ...D, J=1, 2, ...m. The chaotic sequence chx is transformed into an optimization variable through the carrier wave, and the formula is:

[0035] x tj =x max，j -(x max，j -x min，j )·chx tj Formula 10;

[0036] In the formula, x max 、x min Respectively represent the maximum and minimum values ​​of the optimization variable x, D is the variable dimension, and m is the population size. The optimized variable x1 obtained by mapping is used as the initialization value of the particle. Calculate the fitness of the particle, and then perform a rough search of the particle, perform chaotic perturbations on the premature particles, and generate the sequence thx according to the Tent mapping (taking φ=0.6) v ,t=1,2,...D,J=1,2,...m. The chaotic perturbation is carried out by the following formula:

[0037] ψ + =(P d -α·thx tj )(1-α) Formula XI;

[0038] I ** =ψ * (x max，j -x min，j )+x min，j Formula 12;

[0039] In the formula, α (0≤α≤1) is the adjustment parameter, P d(d=1,2,...D) is the current optimal solution vector; where, That is, the chaotic vector after disturbance; it enables particles to search the entire empty direction quickly and locally, effectively improving the accuracy and convergence speed of the particle swarm algorithm.

[0040] The control module includes a multi-level grasping network module and a robot grasping trajectory planning module. In formula 8 and formula 9, take c 11 =c 21 =1.3.

[0041] The present invention is a highly integrated and intelligent 3D printing product sorting system, which can fundamentally cope with the multiple challenges of accuracy, efficiency, safety, etc. faced in the field of 3D printing product distribution.

[0042] The system described in the present invention can construct a multi-level grasping network integrating SRCNN, improved YOLOv5 instance segmentation module and enhanced AFFGA network, and combines the dynamic planning optimization of the robot arm grasping trajectory to comprehensively optimize each link of 3D printed product sorting and achieve high-definition imaging, precise positioning and intelligent decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0044] Attached Figure 1 is a schematic diagram of the present invention;

[0045] Attached Figure 2 is a schematic diagram of a flow chart of an embodiment of the present invention;

[0046] Attached Figure 3 It is a schematic diagram comparing the network model structure corresponding to the improved SRCNN algorithm and the original structure;

[0047] Attached Figure 4 This is a schematic diagram of the YOLOv5+EA-sppfcspc+BiFPN algorithm performing instance segmentation;

[0048] Attached Figure 5 It is a schematic diagram of the weighted bidirectional feature pyramid network model BiFPN performing operations;

[0049] Attached Figure 6 It is a flowchart of the optimized particle swarm algorithm for dynamic planning of the robot's trajectory. DETAILED DESCRIPTION

[0050] As shown in the figure, a 3D printed product sorting system based on a multi-level grasping network framework comprises a control module, a sorting platform for placing 3D printed products and a robotic arm for sorting 3D printed products; a binocular camera connected to the control module is arranged above the sorting platform; the control module first collects the original 3D printed product image through the binocular camera, and generates a high-resolution image through super-resolution reconstruction technology, and then uses the improved YOLOv5+EA-SPPFCSPC+BiFPN instance segmentation algorithm to process the image, accurately locate the target 3D printed product to be grasped, and control the grasping operation of the robotic arm through the AFFGA grasping detection network; and dynamically plans the trajectory of the robotic arm through the optimized particle swarm algorithm to ensure that the grasping network finds the best path to reach the target grasping position within a limited time, while ensuring the smoothness of the trajectory, and minimizing the potential risks in the grasping process.

[0051] The 3D printed products are placed on the grabbing platform, and some of the 3D printed products are in a stacked state.

[0052] The sorting system improves the gripping capability of the robot arm for chaotic, overlapping and variously shaped 3D printed products by using a gripping framework model that combines a multi-level gripping network with an adaptive robot mechanism;

[0053] The working method of the grasping framework model includes: firstly, preprocessing the original image through the improved super-resolution convolutional neural network algorithm SRCNN, and then using the YOLOv5+EA-SPPFCSPC+BIFPNC instance segmentation algorithm, that is, the YOLO-EASB algorithm, to accurately segment the 3D printed products, and then determine the most suitable 3D printed products for grasping by evaluating the integrity of the segmentation mask, and then through the improved adaptive feature fusion and grasping perception network IAFFGA-Net, combined with the optimized loss function, these segmented 3D printed products are processed to ensure high-precision grasping operations even in complex environments.

[0054] The workflow of the improved super-resolution convolutional neural network algorithm SRCNN includes:

[0055] Step A1, operation of feature extraction layer Conv(5, d, 1): a convolution kernel of size 5×5 is used to replace the convolution kernel of 9×9 in SRCNN for the original low-resolution image, and a feature extraction operation is performed to achieve the effect of small information loss and high speed. The number of channels c i is 1, the number of convolution kernels is n i is d;

[0056] Step A2, the operation of the contraction layer Conv(1, s, d). A convolution kernel of size 1×1 is used, and a small number of convolution kernels s is used to reduce the LR feature dimension d to s, so as to save computational cost;

[0057] Step A3, operation of the nonlinear mapping layer Conv(3, s, s): If SRCNN lacks deep network experiments, in order to maintain the same performance as SRCNN, s convolution kernels of size 3×3 are used to replace the original wide convolution kernel, and the number of channels c is maintained. i for s;

[0058] Step A4, operation of the expansion layer Conv(1, d, s): If the contraction layer reduces the LR feature dimension, which will reduce the quality of image restoration, then add a expansion layer after the nonlinear mapping to expand the HR feature dimension; use s convolution kernels of size 1×1 for the expansion operation, which is consistent with the contraction layer;

[0059] Step A5, operation of the deconvolution layer DeConv(9, 1, s): deconvolution is used as the inverse operation process of convolution. When the step size is k, the resolution is increased by k times, and a deconvolution kernel of size 9×9 is used.

[0060] In the YOLOv5+EA-SPPFCSPC+BIFPNC instance segmentation algorithm, based on the multi-scale features generated by SPP, the channel attention mechanism ECA (ECA, Efficient Channel Attention) is further integrated, so that the model can intelligently allocate the response weight of each channel, that is, when the 3D printed product is partially occluded, the key features of the non-occluded area are highlighted, so as to facilitate the detection of the occluded 3D printed products; in addition, in order to fully explore the complementary advantages between deep-level features and shallow-level features, the Adaptive Dual Attention Feature Fusion (ADaFF) module ADaFF is adopted; the ADaFF module is designed to efficiently integrate cross-level features and suppress redundancy, so that the model can not only extract high-level abstract features from the deep network, but also make full use of the delicate texture information of the shallow input.

[0061] The ADaFF module adds the input features and the residual features channel by channel, and then guides these mixed features to undergo two sub-processes: local attention and global attention. In the local attention sub-process, the ECA mechanism is used to enhance the expressiveness of local features, while the global attention enhances the integrity of features by capturing global contextual information. The output features of the attention mechanisms of the two sub-processes interact through a feature fusion strategy to achieve adaptive integration of features. Finally, in the output stage, local and global features are weightedly merged after independent weight calculation to generate the final output features.

[0062] In the YOLOv5+EA-SPPFCSPC+BIFPNC instance segmentation algorithm, BIFPNC is a weighted bidirectional feature pyramid network model BiFPN. In this model, the strategy of enhancing feature fusion through jump connection is used to utilize shallow features and deep features to improve the perception of the details of the target 3D printed product and optimize the perception of the segmentation of occluded objects.

[0063] Assume that the feature maps of different scales are x = [x0, x1, x2], the weights [w0, w1, w2] are learnable parameters obtained through model training, and ∈ is a small constant; the steps for fusion are as follows:

[0064] Weight normalization, the formula used is: Weighted feature fusion, the formula used is y = w′0·x0+w′1·x1+w′2·x2Formula 2;

[0065] The activation after weighted feature fusion uses the formula y′=SiLU(y) Formula 3;

[0066] Feature concatenation, the formula used is z = concat (x0, y′) Formula 4;

[0067] The convolution operation uses the formula output = Conv2 d(z) Formula 5.

[0068] If the proportion of effective grasping points of 3D printed products in the visual input is small, the BCE in the AFFGA grasping detection network is upgraded to FocalLoss; FocalLoss is defined by the formula:

[0069] FL(p t )=-α t (1-p t )γlog(p t ) Formula 6;

[0070] Among them, p t is the probability of the model predicting the true class label, α tis the category balance weight, which is used to adjust the ratio of the importance of positive and negative samples, and is set according to the actual situation; and γ is an adjustable focusing parameter, which smoothly determines the speed at which the loss weight of easy-to-resolve samples decreases;

[0071] During the training process, as p t The closer it is to 1, the lighter the penalty FocalLoss imposes on correctly classified samples, making the model training pay more attention to and strengthen the learning of difficult samples, so that FocalLoss can dynamically adjust the learning speed of difficult and easy samples, enhance the adaptability and stability of the model when dealing with significant changes in grasping posture, complex object shapes or severe occlusion, reduce the overfitting phenomenon caused by uneven data distribution, and thus optimize the execution of complex object grasping tasks.

[0072] The optimized particle swarm algorithm dynamically programs the trajectory of the robot arm as follows: first, a 3-5-3 piecewise polynomial interpolation function is constructed, and then the optimization objective function is constructed using three interpolation times. Subsequently, the improved particle swarm algorithm is used to optimize each interpolation time with the goal of minimizing the total running time while satisfying the constraints.

[0073] The optimized particle swarm algorithm improves the Camellia pollen picking algorithm of the standard particle swarm algorithm and uses a nonlinear decreasing strategy to adjust the value of the inertia weight ω, that is, the expression is:

[0074]

[0075] In the formula, ω max ,ω min (omega max , omega min ) are the maximum and minimum inertia weights, respectively. In this paper, they are 0.86 and 0.44; n is the current iteration number; N is the maximum evolutionary generation. The dynamic method is used to assign values ​​to c1 and c2. The formula is:

[0076]

[0077]

[0078] Among them, c 11 , c 21 is a constant;

[0079] In the initial stage, IPSO randomly generates M three-dimensional particles and obtains the initial position and velocity of each particle, which leads to uneven distribution of population initialization and affects the optimization of the algorithm. Based on IPSO, a double-mixed initialization sequence based on Logitic-Tent is introduced, and Logisfic mapping (taking μ = t) is used to generate i chaotic sequences chx η: Where t=1, 2, ...D, J=1, 2, ...m. The chaotic sequence chx is transformed into an optimization variable through the carrier wave, and the formula is:

[0080] x tj =x max，j -(x max，j -x min，j )·chx tj Formula 10;

[0081] In the formula, x max 、x min represents the maximum and minimum value of the optimization variable x, D is the variable dimension, and m is the population size. The optimized variable x1 obtained by mapping is used as the initial value of the particle. The fitness of the particle is calculated, and then a rough search of the particle is performed. The premature particles are subjected to chaotic perturbation. According to the Tent mapping (take )Generate sequence thx v ,t=1,2,...D,J=1,2,...m. The chaotic perturbation is carried out by the following formula:

[0082] ψ + =(P d -α·thx tj ) / (1-α) Formula 11;

[0083] I ** =ψ(x max，j -x min，j )+x min，j Formula 12;

[0084] In the formula, α (0≤α≤1) is the adjustment parameter, P d (d=1,2,...D) is the current optimal solution vector; where, That is, the chaotic vector after disturbance; it enables particles to search the entire empty direction quickly and locally, effectively improving the accuracy and convergence speed of the particle swarm algorithm.

[0085] The control module includes a multi-level grasping network module and a robot grasping trajectory planning module. In formula 8 and formula 9, take c 11 =c 21 =1.3.

[0086] Example:

[0087] In this example, the sorting system process is as follows: 3D printed products are placed on a 65cm×65cm grasping platform, and some products are stacked into multiple layers. First, the original 3D printed product image is collected using the D435 binocular camera, and a high-resolution image is generated using super-resolution reconstruction technology. Then, the image is processed using the improved YOLOv5+EA-SPPFCSPC+BiFPN instance segmentation algorithm to accurately determine the target 3D printed product to be grasped, and the grasping operation is performed through the AFFGA network. The trajectory of the robotic arm is dynamically planned using the optimized particle swarm algorithm to ensure that the grasping network finds the best path to reach the target grasping position within a limited time, while ensuring the smoothness of the trajectory and minimizing the potential risks in the grasping process. With the synergistic effect of each module, the system can efficiently and stably sort different 3D printed products in a complex environment.

[0088] In this example, the 3D printed product sorting system consists of two main modules: a multi-level grasping network module and a robotic arm grasping trajectory planning module.

[0089] The network structure of the multi-level grasping framework is as follows: In order to improve the grasping ability of the robot arm for chaotic, overlapping and variously shaped 3D printed products, a new grasping framework combining a multi-level grasping network with an adaptive robot mechanism is proposed. The framework first pre-processes the original image through an improved super-resolution convolutional neural network (SRCNN) algorithm, and then uses the independently proposed YOLOv5+EA-SPPFCSPC+BIFPNC (YOLO-EASB) instance segmentation algorithm to accurately segment 3D printed products. By evaluating the integrity of the segmentation mask, the most suitable 3D printed product for grasping is determined. Next, these segmented 3D printed products are processed through the improved adaptive feature fusion and grasping perception network (IAFFGA-Net) combined with the optimized loss function to ensure high-precision grasping operations even in complex environments.

[0090] Image super-resolution preprocessing based on improved SRCNN: The main reason why the SRCNN algorithm is slow is that the bicubic interpolation preprocessing takes a long time and the convolution kernel size is large. Therefore, the SRCNN algorithm was improved. The improved algorithm first uses the original low-resolution image as input, no longer performs bicubic interpolation on the input image, but introduces a deconvolution layer at the end of the network for upsampling; secondly, shrinkage, mapping and expansion are used to replace the nonlinear mapping layer in SRCNN, and a smaller convolution kernel and a deeper network structure are used to effectively improve the model convergence speed while ensuring the image reconstruction quality.

[0091] The robot grasping trajectory planning module is based on the robot trajectory planning framework. Specifically, a new trajectory planning method is proposed to ensure the efficient operation of the multi-level grasping network. The algorithm can plan the optimal path of the robot in real time according to the target grasping position provided by the multi-level grasping network. While ensuring the smoothness of the robot trajectory and the grasping stability, the optimized particle swarm algorithm can find the best path to the grasping position within a limited time and effectively reduce potential risks in the grasping process, such as misoperation. This combination enables the system to not only accurately identify and grasp 3D printed products in complex environments, but also improve the safety and stability of grasping by dynamically adjusting the motion trajectory of the robot.

[0092] The trajectory planning algorithm first constructs a 3-5-3 piecewise polynomial interpolation function, and then uses three interpolation times to construct the optimization objective function. Subsequently, the improved particle swarm algorithm optimizes each interpolation time with the goal of minimizing the total running time while satisfying the constraints. The algorithm flow chart is shown in the figure below: Figure 6 shown.

[0093] Improved particle swarm algorithm: When the standard particle swarm algorithm is used to optimize the problem, its parameters are fixed. describes the inertia of the particles. It should be larger in the early stage of optimization to ensure that each particle can fly independently and fully search the space. It should be smaller in the later stage to learn more from other particles. and respectively describe the individual and global maximum flight steps. They should be as large as possible in the early stage and as large as possible in the later stage, so that the global search ability and local search ability of the particles can be balanced. The three parameters jointly affect the flight direction of the particles. Assuming that other particles have found a better position, but the inertia of the current particle is too large to fly to a better position quickly, this may lead to problems such as premature algorithm or slow convergence of the algorithm in the later stage. In view of the problems that the standard particle swarm algorithm is prone to premature and slow convergence in the later stage, the improved particle swarm algorithm in this example proposes a nonlinear decreasing strategy in the trajectory planning of the oil tea pollen picking robot arm based on the improved particle swarm algorithm.

Claims

1. A 3D printed product sorting system based on a multi-stage grabbing network framework, characterized in that: The sorting system includes a control module, a sorting platform for placing 3D printed products, and a mechanical arm for sorting 3D printed products; a camera connected to the control module is arranged above the sorting platform; the control module first collects the original 3D printed product image through the camera, generates a high-resolution image through super-resolution reconstruction technology, and then uses the improved YOLOv5+EA-SPPFCSPC+BiFPN instance segmentation algorithm to process the image, accurately locate the target 3D printed product to be grasped, and control the grasping operation of the mechanical arm through the AFFGA grasping detection network; and dynamically plans the trajectory of the mechanical arm through the optimized particle swarm algorithm to ensure that the grasping network finds the best path to reach the target grasping position within a limited time, while ensuring the smoothness of the trajectory.

2. A 3D printed product sorting system based on a multi-stage grabbing network framework according to claim 1, characterized in that: The 3D printed products are placed on a grabbing platform, and some of the 3D printed products are in a stacked state, and the camera is a binocular camera.

3. The 3D printing product sorting system based on a multi-stage grabbing network framework according to claim 1 is characterized in that: The sorting system improves the gripping capability of the robot arm for chaotic, overlapping and variously shaped 3D printed products by using a gripping framework model that combines a multi-level gripping network with an adaptive robot mechanism; The working method of the grasping framework model includes: firstly, preprocessing the original image through the improved super-resolution convolutional neural network algorithm SRCNN, and then using the YOLOv5+EA-SPPFCSPC+BIFPNC instance segmentation algorithm, that is, the YOLO-EASB algorithm, to accurately segment the 3D printed products, and then determine the most suitable 3D printed products for grasping by evaluating the integrity of the segmentation mask, and then through the improved adaptive feature fusion and grasping perception network IAFFGA-Net, combined with the optimized loss function, these segmented 3D printed products are processed to ensure high-precision grasping operations even in complex environments.

4. The 3D printed product sorting system based on a multi-stage grabbing network framework according to claim 3 is characterized in that: The workflow of the improved super-resolution convolutional neural network algorithm SRCNN includes: Step A1, operation of feature extraction layer Conv(5, d, 1): a convolution kernel of size 5×5 is used to replace the convolution kernel of 9×9 in SRCNN for the original low-resolution image, and a feature extraction operation is performed to achieve the effect of small information loss and high speed. The number of channels c i is 1, the number of convolution kernels is n i is d; Step A2, the operation of the contraction layer Conv(1, s, d). A convolution kernel of size 1×1 is used, and a small number of convolution kernels s is used to reduce the LR feature dimension d to s, so as to save computational cost; Step A3, operation of the nonlinear mapping layer Conv(3, s, s): If SRCNN lacks deep network experiments, in order to maintain the same performance as SRCNN, s convolution kernels of size 3×3 are used to replace the original wide convolution kernel, and the number of channels c is maintained. i for s; Step A4, operation of the expansion layer Conv(1, d, s): If the contraction layer reduces the LR feature dimension, which will reduce the quality of image restoration, then add a expansion layer after the nonlinear mapping to expand the HR feature dimension; use s convolution kernels of size 1×1 for the expansion operation, which is consistent with the contraction layer; Step A5, operation of the deconvolution layer DeConv(9, 1, s): deconvolution is used as the inverse operation process of convolution. When the step size is k, the resolution is increased by k times, and a deconvolution kernel of size 9×9 is used.

5. The 3D printing product sorting system based on a multi-stage grabbing network framework according to claim 3 is characterized in that: In the YOLOv5+EA-SPPFCSPC+BIFPNC instance segmentation algorithm, based on the multi-scale features generated by SPP, the channel attention mechanism ECA is further integrated, so that the model can intelligently assign the response weight of each channel, that is, highlight the key features of the non-occluded area when the 3D printed product is partially occluded, so as to facilitate the detection of the occluded 3D printed products; in addition, in order to fully explore the complementary advantages between deep-level features and shallow-level features, the adaptive dual attention feature fusion module ADaFF is adopted; the ADaFF module is designed to efficiently integrate cross-level features and suppress redundancy, so that the model can not only extract high-level abstract features from the deep network, but also make full use of the delicate texture information of the shallow input.

6. A 3D printed product sorting system based on a multi-stage grabbing network framework according to claim 5, characterized in that: The ADaFF module adds the input features and the residual features channel by channel, and then guides these mixed features to undergo two sub-processes: local attention and global attention. In the local attention sub-process, the ECA mechanism is used to enhance the expressiveness of local features, while the global attention enhances the integrity of features by capturing global context information. The output features of the attention mechanisms of the two sub-processes interact through feature fusion strategies to achieve adaptive integration of features. Finally, in the output stage, local and global features are weighted and merged after independent weight calculation to generate the final output features.

7. The 3D printing product sorting system based on a multi-stage grabbing network framework according to claim 5 is characterized in that: In the YOLOv5+EA-SPPFCSPC+BIFPNC instance segmentation algorithm, BIFPNC is a weighted bidirectional feature pyramid network model BiFPN. In this model, the strategy of enhancing feature fusion through jump connection is used to utilize shallow features and deep features to improve the perception of the details of the target 3D printed product and optimize the perception of the segmentation of occluded objects. Assume that the feature maps of different scales are x = [x0, x1, x2], the weights [w0, w1, w2] are learnable parameters obtained through model training, and ∈ is a small constant; the steps for fusion are as follows: Weight normalization, the formula used is: Formula 1; Weighted feature fusion, the formula used is y = w′0·x0+w′1·x1+w′2·x2 Formula 2; The activation after weighted feature fusion uses the formula y′=SiLU(y) Formula 3; Feature concatenation, the formula used is z = concat (x0, y′) Formula 4; The convolution operation uses the formula output = Conv2d(z) Formula 5.

8. The 3D printing product sorting system based on a multi-stage grabbing network framework according to claim 1 is characterized in that: If the proportion of effective grasping points of 3D printed products in the visual input is small, the BCE in the AFFGA grasping detection network is upgraded to Focal Loss; Focal Loss is defined by the formula: FL(p t )=-α t (1-pt ) γ log(p t ) Formula 6; Among them, p t is the probability of the model predicting the true class label, α t is the category balance weight, which is used to adjust the ratio of the importance of positive and negative samples, and is set according to the actual situation; and γ is an adjustable focusing parameter, which smoothly determines the speed at which the loss weight of easy-to-resolve samples decreases; During the training process, as p t The closer it is to 1, the lighter the penalty Focal Loss imposes on correctly classified samples, making the model training pay more attention to and strengthen the learning of difficult samples, so that Focal Loss can dynamically adjust the learning speed of difficult and easy samples, enhance the adaptability and stability of the model when dealing with significant changes in grasping posture, complex object shapes or severe occlusion, reduce the overfitting phenomenon caused by uneven data distribution, and thus optimize the execution of complex object grasping tasks.

9. The 3D printing product sorting system based on a multi-stage grabbing network framework according to claim 1, characterized in that: The optimized particle swarm algorithm dynamically programs the trajectory of the robot arm as follows: first, a 3-5-3 piecewise polynomial interpolation function is constructed, and then the optimization objective function is constructed using three interpolation times. Subsequently, the improved particle swarm algorithm is used to optimize each interpolation time with the goal of minimizing the total running time while satisfying the constraints. The optimized particle swarm algorithm improves the Camellia pollen picking algorithm of the standard particle swarm algorithm and uses a nonlinear decreasing strategy to adjust the value of the inertia weight ω, that is, the expression is: In the formula, ω max ,ω min are the maximum and minimum inertia weights respectively; n is the current number of iterations; N is the maximum evolutionary generation, and a dynamic method is used to assign values ​​to c1 and c2. The formula is: Among them, c 11 , c 21 is a constant; In the initial stage, IPSO randomly generates M three-dimensional particles and obtains the initial position and velocity of each particle, which leads to uneven distribution of population initialization and affects the optimization of the algorithm. Based on IPSO, a dual-mixed initialization sequence based on Logitic-Tent is introduced, and i chaotic sequences chx are generated using Logisfic mapping. η : Take μ = t; where t = 1, 2, .. D, J = 1, 2, ... m. Transform the chaotic sequence chx into an optimization variable through the carrier wave, the formula is: x tj =x max，j -(x max，j -x min，j )·chx tj Formula 10; In the formula, x max 、x min Respectively represent the maximum and minimum values ​​of the optimization variable x, D is the variable dimension, and m is the population size. The optimized variable x1 obtained by mapping is used as the initialization value of the particle. Calculate the fitness of the particle, then perform a rough search of the particle, perform chaotic perturbations on the premature particles, and generate the sequence thx according to the Tent mapping. v ,t=1,2,...D,J=1,2,...m. The chaotic perturbation is carried out by the following formula: ψ + =(P d -α·thx tj ) / (1-α) Formula 11; I ** =ψ * (x max，j -x min，j )+x min，j Formula 12; In the formula, α (0≤α≤1) is the adjustment parameter, P d (d=1,2,...D) is the current optimal solution vector; where, That is, the chaotic vector after disturbance; it enables particles to search the entire empty direction quickly and locally, effectively improving the accuracy and convergence speed of the particle swarm algorithm.

10. A 3D printed product sorting system based on a multi-stage grabbing network framework according to claim 9, characterized in that: The control module includes a multi-level grasping network module and a robot grasping trajectory planning module. In formula 8 and formula 9, take c 11 =c 21 =1.3, in formula 7, ω max ,ω min Take 0.86 and 0.44 respectively; Tent mapping takes

Images