Robot structure generative design and optimization method based on digital twinning

Abstract

The application provides a robot structure generative design and optimization method based on digital twinning, and comprises the following steps: step 1, a mechanical arm structure generation model is established based on a deep neural network; the input of the mechanical arm structure generation model is task demand, and the output is the length of the arm of the mechanical arm and the number of joints; step 2, the output of the mechanical arm structure generation model is optimized based on an improved particle swarm algorithm; in the improved particle swarm algorithm, the inertia weight is an adaptive weight; step 3, a robot is designed and built in combination with the optimal solution output in step 2. The application can generate a robot based on a specific scene, and the performance evaluation indexes are diversified.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of robots, and particularly relates to a robot structure generative design and optimization method based on digital twinning. BACKGROUND

[0002] Robot structure design is one of the core fields of robot technology development, involving multiple disciplines such as mechanical engineering, control engineering, and computer science. Its main goal is to design a robot structure that can meet the requirements of specific tasks, ensuring its stability, flexibility, and efficiency in various environments.

[0003] The current technical development of robot structure design and its shortcomings mainly include:

[0004] 1) Computer-aided design (CAD)

[0005] With the development of computer technology, CAD software is widely used in robot structure design. CAD tools can help designers quickly create and modify design models, improving design efficiency and accuracy. However, CAD tools are mainly used in the design stage and lack automated optimization functions.

[0006] 2) Finite element analysis (FEA)

[0007] FEA technology is used to simulate and analyze the performance of robot structures under different loads and conditions. Through finite element analysis, designers can predict the stress, deformation, and vibration characteristics of the structure, thereby optimizing the design. However, FEA usually requires a large amount of computing resources and time, and relies on the professional knowledge of designers.

[0008] 3) Generative design

[0009] Generative design is an emerging design method that uses algorithms to generate multiple design schemes and selects the best scheme through simulation and optimization. Generative design can explore a wider design space and discover innovative structures that traditional methods cannot think of. However, the application of this technology in robot structure design is still in its early stages, facing challenges such as algorithm complexity and computational cost.

[0010] 4) Digital twinning technology

[0011] Digital twinning technology creates a virtual model of a physical object, enabling real-time monitoring and optimization of robot structures. This technology can perform multiple simulations and tests in a virtual environment, reducing the need for physical prototypes. However, existing digital twinning technology still needs to be improved in terms of intelligence and multi-physical field coupling simulation.

[0012] In addition, the current robot structure design also has the following problems:

[0013] 1) Reliance on manual design and labor consumption. Current robot structural design processes largely rely on human experience and trial-and-error methods, requiring designers to manually adjust and optimize structural parameters. This approach is not only time-consuming, but also requires more manpower as design complexity increases, resulting in a waste of resources.

[0014] 2) Low design consistency. Because the design process relies on the designer's personal experience and judgment, the consistency of the design results is low. Different designers may come up with different design solutions, making it difficult to ensure product consistency and quality in large-scale production.

[0015] 3) The optimization dimension is relatively single. Current design optimization focuses primarily on a single performance indicator, such as strength or weight, while less consideration is given to comprehensive optimization across multiple dimensions. Existing technologies struggle to optimize multiple performance indicators simultaneously, resulting in designs that may not meet diverse requirements in practical applications.

[0016] 4) Application scenarios (task requirements) are relatively limited. The designed commercial robots have fixed structures and parameters, making it difficult to fully guarantee optimal efficiency, cost, and other comprehensive performance in specific working conditions or application scenarios. Summary of the Invention

[0017] The purpose of this invention is to provide a generative design and optimization method for robot structures based on digital twins to solve the above technical problems. The technical solutions adopted are:

[0018] A generative design and optimization method for robot structures based on digital twins, comprising the following steps:

[0019] Step 1: Establish a robotic arm structure generation model based on a deep neural network;

[0020] The input of the manipulator structure generation model is the task requirement, and its output is the length of the manipulator rod and the number of joints;

[0021] Step 2: Based on the improved particle swarm algorithm, optimize the output of the robotic arm structure generation model;

[0022] Among them, in the improved particle swarm algorithm, the inertia weight is an adaptive weight;

[0023] Step 3: Based on the optimal solution output in step 2, design the robot and complete its construction.

[0024] Preferably, in step 1, the specific process of establishing the robotic arm structure generation model includes:

[0025] Step 1A, data collection and pretreatment: the length of the rod and the number of joints of different mechanical arms are collected, and the working space of each mechanical arm is calculated; according to different task requirements, the length of the rod and the number of joints are marked to form a structured data set;

[0026] Step 1B, deep learning model training: based on the data set, a deep learning neural network model is constructed and trained until the model reaches the preset accuracy requirement;

[0027] Step 1C, model saving and application: when the network model reaches the preset accuracy, the system automatically saves the model.

[0028] Preferably, in step 2, the expression of the inertia weight is:

[0029]

[0030] Wherein, t is the current iteration number of the algorithm;

[0031] T is the total iteration number of the algorithm;

[0032] ω max is the inertia weight at the beginning of the algorithm;

[0033] ω min is the inertia weight at the end of the algorithm.

[0034] Preferably, the weight of the self information referred to in the iteration number t search The expression is:

[0035]

[0036] The weight of the group information referred to in the iteration number t search The expression is:

[0037]

[0038] The learning factor at the beginning of the algorithm;

[0039] The learning factor at the end of the algorithm; i = 1, 2.

[0040] Compared with the prior art, the advantages of the present application are:

[0041] 1. Combination of digital twin and generative design: The application of generative design for robots in actual scenarios is often unsatisfactory, and often faces problems such as collision, interference, or deviation between theory and actual effect. By introducing digital twin technology (integrated into the control system) for collision detection and obstacle avoidance path planning, the running mode and effect of pose estimation and path planning can be simulated and optimized, many problems that may be encountered in the hardware environment are predicted and solved in advance, the accuracy, reliability, stability and safety are improved, and "virtual simulation, virtual reality, virtual control, virtual prediction, virtual optimization, virtual and real coexistence" are realized.

[0042] 2. Innovation of structure parameter generation method: Compared with the conventional method of designing robot structure parameters relying on assumptions or artificial experience, a large number of robot structure parameter design samples are used, a neural network is used as a prediction model, and the neural network is trained to generate a prediction of robot structure parameters, avoiding subjective speculation and human intervention, saving manpower and time, and realizing the "self-organization" and "decentralization" of the design process.

[0043] 3. Universality and popularization of application scenarios: Currently, commercial robots are generally only suitable for single application scenarios, and it is difficult to completely guarantee the optimal comprehensive performance of efficiency, cost, etc. in specific working conditions or application scenarios. This robot structure generative design and optimization method based on digital twin can generate a robot based on a specific scenario to achieve optimal comprehensive performance in the current scenario, while meeting the characteristics of lightweight, low energy consumption, high precision, etc.

[0044] 4. Diversification of performance evaluation indicators: The optimization indicators for robot structure are usually single, causing problems such as serious energy consumption of robots and low motion trajectory tracking accuracy. This method performs quantitative analysis on each key performance indicator (robot arm length, number of joints), determines a new optimization objective function (fitness function), further optimizes the robot arm layout, avoids energy resource waste, and improves production efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 Flowchart of the robot structure generative design and optimization method based on digital twin;

[0046] Figure 2 Structure diagram of the four-axis robot arm finally generated according to the optimal solution;

[0047] Figure 3 Optimization result graph.

[0048] 1-First joint of the robot arm moving pair; 2-Second joint of the robot arm rotating pair; 3-Third joint of the robot arm rotating pair; 4-Fourth joint of the robot arm rotating pair; 5-End gripper of the robot arm. DETAILED DESCRIPTION

[0049] The digital-twin-based robot structure generative design and optimization method of the present application will be described in more detail below in conjunction with the schematic diagram, which represents the preferred embodiment of the present application, it should be understood that the present application described herein can be modified by those skilled in the art, while still achieving the advantageous effects of the present application. Therefore, the following description should be understood as extensive knowledge for those skilled in the art, and not as a limitation of the present application.

[0050] As shown in Figure 1 A digital-twin-based robot structure generative design and optimization method includes the following steps:

[0051] 1) Digital-twin-based robot structure generative design

[0052] This technical route realizes the automatic generation and optimization of mechanical arm structure parameters through deep learning neural network.

[0053] That is, step 1, based on deep neural network, establish mechanical arm structure generation model.

[0054] The input of the mechanical arm structure generation model is the task demand, and the output is the length of the arm and the number of joints of the mechanical arm; the robot is the mechanical arm.

[0055] Step 1 specifically includes the following steps:

[0056] Step 1A, data acquisition and preprocessing: collect the length of the arm and the number of joints of different mechanical arms, and calculate the workspace of each mechanical arm; similarly, one workspace may correspond to different task demands. According to the different task demands of the workspace, mark the length of the arm and the number of joints to form a structured data set.

[0057] Among them, the calculation of the workspace is to evaluate the reachability of the mechanical arm and verify whether the mechanical arm workspace meets the task demand of the work scene, that is, whether it can reach the required farthest distance of work.

[0058] Step 1B, deep learning model training.

[0059] Based on the data set, construct and train the deep learning neural network model. The input of the model is the task demand, and the output is the length of the arm and the number of joints of the mechanical arm. Through continuous iteration training, the model parameters are optimized until the model reaches the preset accuracy requirement.

[0060] Step 1C, model saving and application.

[0061] When the network model reaches the preset accuracy, the system automatically saves the model.

[0062] The model can be used to predict and generate mechanical arm structure parameters that meet specific task requirements.

[0063] The pseudo code is as follows:

[0064]

[0065] Among them, one task requirement corresponds to one work label.

[0066] 2) Robot structure optimization based on digital twinning

[0067] That is, step 2, based on the improved particle swarm algorithm, the output of the mechanical arm structure generation model is optimized.

[0068] Step 2 specifically includes the following steps:

[0069] (1) Initialization: Randomly generate the same number of particles as the predetermined dimension in the search space, and determine their initial position and speed.

[0070] Assume that in a D-dimensional search space, a particle swarm is composed of N particles, and the current iteration number of the algorithm is t, then the position vector of the i-th particle is:

[0071]

[0072] The velocity vector of the i-th particle is:

[0073]

[0074] (2) Evaluation: Calculate the fitness value of each particle to determine the quality of the solution.

[0075] For this study, the time taken by the mechanical arm from the initial position to the target position, the overall mass of the mechanical arm, and the energy consumed by the mechanical arm movement are used as evaluation parameters.

[0076] Among them, the fitness function is:

[0077]

[0078] T represents time, M represents mass, and E represents energy. The denominator represents the maximum time, mass, and energy before optimization, and the optimized values are divided by normalization. γ1-γ3 are weights, and the weights are set based on the importance of the three indicators. The specific values are set to 6, 2, and 2.

[0079] That is, the fitness function includes and These three parts.

[0080] (3) Update individual extreme value: If the current fitness of the particle is higher than its individual extreme value, set the individual extreme value of the particle to the current position;

[0081] The velocity update of the i-th particle is as follows:

[0082]

[0083] is the inertia part, representing the tendency of the particle to maintain the current motion state;

[0084] is the individual cognition part, representing the tendency of the particle to move to its best position according to its own previous search results;

[0085] is the group cognition part, representing the tendency of the particle to move to the global best position according to the group's previous search experience.

[0086] The calculation is as follows: After that, the position update of the i-th particle is as follows:

[0087]

[0088] wherein r 1i and r 2i are uniformly distributed in the range of [0, 1], are random numbers related to the particle i, and are used to increase the randomness of the search.

[0089] ω t is the inertia weight of the particle corresponding to the current iteration number t, which determines the degree of inheritance of the particle velocity, and gradually decreases with the increase of the iteration number t, so as to accelerate the iteration speed of the algorithm.

[0090]

[0091] and is the learning factor, representing the weight of the particle's reference to its own information and group information when searching at the iteration number t, the individual learning factor gradually decreases with the increase of the iteration number t, and the social learning factor gradually increases with the increase of the iteration number t:

[0092]

[0093] wherein ω max and ω min are the inertia weights at the beginning and end of the algorithm, respectively, and

[0094] (i = 1, 2) are the learning factors at the beginning and end of the algorithm, and T is the total iteration number of the algorithm. Obviously, with the increase of the iteration number, the above parameters increase or decrease linearly.

[0095] (4) Update the global extreme value: if the individual extreme value of the current particle is better than the global extreme value of the entire particle swarm, update the global extreme value to the individual extreme value of the particle.

[0096] (5) Update the inertia weight, speed and position: first update the inertia weight of the particle, and then adjust the speed and position of each particle according to the individual extreme value and the global extreme value.

[0097] By repeating steps (2) to (5) until the number of iterations is reached, the global optimal solution can be finally obtained.

[0098] The pseudo code is as follows:

[0099]

[0100] In summary, in step 2, the length of the robot arm and the number of joints are used as the optimization target of the algorithm, and the running time, energy consumption and overall mass of the robot are used as the algorithm evaluation function (fitness function) combined with the robot evaluation index, so as to achieve the purpose of optimizing the performance of the generated robot.

[0101] As shown in Figure 3 , after optimization by the algorithm, the running time, total mass and energy consumption of the robot are all well optimized, and the optimization results meet the actual scene requirements, ensuring that the robot shortens the working time, reduces the mass and reduces the energy consumption under the condition of completing the task.

[0102] Figure 3 In the figure, the horizontal coordinate is the number of iterations, and the vertical coordinate is the value of each part of the fitness function.

[0103] Step 3, combine the optimal solution output in step 2 to design the robot and complete its construction.

[0104] This robot is generated by the above generative system according to the scene requirements of the sorting task and finally completes the construction and debugging of the robot, and is finally applied to related scenes. The specific implementation method is as follows:

[0105] I. Target detection and pose estimation of sorting target:

[0106] (1) Use YOLOv8 algorithm for target detection, including target recognition and position acquisition.

[0107] (2) Based on depth camera and geometric features for target recognition and pose estimation.

[0108] (3) Use image processing techniques (such as Canny edge detection, Hough transform) to assist in completing pose estimation.

[0109] II. Robot structure generation and optimization:

[0110] (1) Use a deep learning model to input the scene requirements and obtain a basic structure of a sorting robot that can complete a radius of 700-800 mm and a vertical height difference range of 800 mm. The robot can grasp different shaped targets, including cylinders and cubes, with a load of no more than 1 kg.

[0111] (2) When the model reaches the preset accuracy, automatically save the model. This model is the robot structure parameters that meet the specific task requirements.

[0112] (3) Optimize the physical model of the robot basic structure by improving the algorithm. The algorithm adjusts the robot structure by reorganizing the joints, adjusting the rod length, and simulating the robot running scenario to obtain the final robot structure model and parameters.

[0113] III. Construction of generative robot digital twin system:

[0114] (1) Design and build the robot based on the generated robot structure and its parameters.

[0115] (2) Integrate target detection and pose estimation to build a complete fully automatic robot system.

[0116] (3) Implement robot control and combine with the digital twin system (integrated into the control system).

[0117] (4) Run the robot and the digital twin system to complete the scheduled sorting task.

[0118] As shown in Figure 2 , the robot includes:

[0119] a robot arm moving pair first joint 1;

[0120] a robot arm rotating pair second joint 2 connected to and movable with the robot arm moving pair first joint 1;

[0121] a robot arm rotating pair third joint 3 connected to the robot arm rotating pair fourth joint 4 and the robot arm rotating pair second joint 2 through a rod.

[0122] a robot arm end gripper 5 connected to the robot arm rotating pair fourth joint 4.

[0123] a controller connected to the robot arm moving pair first joint 1.

[0124] a digital twin system integrated with target detection and pose estimation connected to the controller.

[0125] The above merely describes the preferred embodiments of the present application and does not limit the present application in any way. Any person skilled in the art can make any form of equivalent replacement or modification to the technical solutions and technical contents disclosed by the present application without departing from the scope of the technical solutions of the present application, and such changes still belong to the protection scope of the present application.

Claims

1. A generative design and optimization method for robot structure based on digital twin, characterized by: The following steps are involved: Step 1: Establish a robotic arm structure generation model based on a deep neural network; The input of the manipulator structure generation model is the task requirement, and its output is the length of the manipulator rod and the number of joints; Step 2: Based on the improved particle swarm algorithm, optimize the output of the robotic arm structure generation model; Among them, in the improved particle swarm algorithm, the inertia weight is an adaptive weight; The expression of the inertia weight is: ； in, -The current iteration number of the algorithm; - is the total number of iterations of the algorithm; - Inertia weight at the beginning of the algorithm; - Inertia weight at the end of the algorithm; In the number of iterations The weight of your own information is used when searching , whose expression is: ； In the number of iterations The weight of reference group information when searching , whose expression is: ； is the learning factor at the beginning of the algorithm; is the learning factor at the beginning and end of the algorithm; i =1,2; Step 3: Based on the optimal solution output in step 2, design the robot and complete its construction.

2. The robot structure generative design and optimization method based on digital twin according to claim 1 is characterized in that: In step 1, the specific process of establishing the robotic arm structure generation model includes: Step 1A, Data Collection and Preprocessing: Collect the rod lengths and number of joints of different robotic arms and calculate the workspace of each robotic arm. Label the rod lengths and number of joints according to different task requirements to form a structured dataset. Step 1B, deep learning model training: Based on the data set, build and train a deep learning neural network model until the model reaches the preset accuracy requirement; Step 1C: Model saving and application: When the network model reaches the preset accuracy, the system automatically saves the model.

Images