A shape optimization method for voice robot based on SVR-PSO

Abstract

This invention discloses a voice robot appearance optimization method based on SVR-PSO. The method comprises the following steps: Step 1: Modeling the outer shell and internal parts of the intelligent voice robot to obtain an outer shell model and a parts assembly model; Step 2: Parameterizing the outer shell model to obtain an outer shape fitting function, and parameterizing the assembly model to obtain the projection heights of multiple parts; Step 3: Constructing an appearance optimization model; Step 4: Optimizing the appearance optimization model using a particle swarm optimization algorithm combined with a support vector regression algorithm to find the optimal solution of the current appearance optimization model and adjust the appearance of the intelligent voice robot to its minimum volume. This method shortens optimization time and improves calculation accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of robot appearance optimization, and more particularly to an appearance optimization method for a voice robot based on SVR-PSO. Background Art

[0002] From small robots like household cleaning robots, tutoring robots, and service robots for hotels and shopping malls to large ones like medical rescue robots, post-disaster debris search robots, and even aerospace robots, the structural design of intelligent robots places even higher demands on space optimization and size reduction than traditional industrial robots. Excessive size results in low internal space utilization, a large footprint, and a somewhat bulky appearance. In some applications, this can also lead to mobility issues and transportation difficulties. Therefore, optimizing the appearance of intelligent robots is particularly important in their structural design.

[0003] However, for the optimal design of the appearance of the intelligent voice robot, some necessary design variables such as the size of its internal components and the thickness of the shell are discrete variables, the constraints are complex and the degree of nonlinearity is high. Therefore, how to establish a clear objective function expression becomes a major difficulty in dealing with such problems. The gradient method, mathematical optimization method and other methods used in traditional structural optimization design not only need to use the objective function value when solving, but also rely on the derivative value of the objective function to determine the search direction. It is difficult to find a clear functional relationship between the structural appearance and component constraints of the intelligent voice robot. Therefore, these traditional structural optimization algorithms cannot be used to solve complex intelligent voice robot structural optimization problems.

[0004] In recent years, swarm intelligence optimization algorithms have been increasingly applied to robot structural optimization problems. As a stochastic optimization algorithm, the particle swarm optimization (PSO) algorithm only uses a fitness function to complete optimization, and it has excellent parallel computing and global search capabilities. Therefore, it is often used to solve optimization problems of large and complex structures. However, PSO also has its drawbacks. For example, particles tend to aggregate in a small area in the late stages of iteration, causing the algorithm to fall into a local optimal solution. Therefore, finding an optimization algorithm suitable for the structural shape of speech recognition robots is of great significance to the research field of robot shape optimization. Summary of the Invention

[0005] The purpose of this invention is to design and develop a voice robot appearance optimization method based on SVR-PSO. Through the particle swarm optimization algorithm based on support vector regression algorithm, the number of fitness calculations is reduced, the optimization time is shortened, and the calculation accuracy is improved.

[0006] The technical solution provided by the present invention is:

[0007] A voice robot shape optimization method based on SVR-PSO includes the following steps:

[0008] Step 1: Model the outer shell and internal parts of the intelligent voice robot to obtain a shell model and a parts assembly model;

[0009] Step 2: parametrically represent the shell model to obtain an outer shape fitting function, and parametrically represent the assembly model to obtain projection heights of multiple parts;

[0010] Step 3: Build an optimized shape model:

[0011] Taking the minimum volume of the shell model as the optimization target, the equation coefficients in the shape fitting function as the optimization variables, and the part assembly model as the constraint condition;

[0012] Step 4: Optimize the shape optimization model using the particle swarm optimization algorithm. After reaching the number of iterations, the minimum value of the fitness function of the particle is the optimal solution of the shape optimization model, and the shape of the intelligent voice robot is adjusted to the minimum volume.

[0013] Among them, the operation time of the particle swarm optimization algorithm is compared with the support vector regression algorithm;

[0014] The fitness function of the particle is:

[0015]

[0016] In the formula, fitness is the fitness function, v0 is the volume of the initial original shape, v is the volume corresponding to the new individual, and fit is the fitness factor.

[0017] Preferably, the parameterized representation of the shell model comprises the following steps:

[0018] Step 1: Export the point cloud of the shell model;

[0019] Step 2: Centralize the point cloud:

[0020] Calculate the mean of all coordinate points on the X-axis and Y-axis in the point cloud, and subtract the mean from all coordinate points one by one to obtain the processed coordinate points;

[0021] Step 3: Perform polynomial fitting on the robot's outer surface based on the processed coordinate points:

[0022] The shell model coordinate system is established with the intersection of the central axis of the shell model and the upper and lower boundaries as the coordinate origin, the upper and lower dividing interfaces are used as the XOY plane, the central axis of the shell model is the v axis, and the upward direction is positive, and the coordinate points with z values ​​greater than zero are used for fitting the upper surface function, and the coordinate points with z values ​​less than zero are used for fitting the lower surface function to obtain the shape fitting function:

[0023] f up (x, y) = W 0_0 +W 0_1 x+W 0_2 y+W 0_3 x 2 +W 0_4 x*y+W 0_5 y 2 +W 0_6 x 3 +W 0_7 y*x 2 +W 0_8 x*y 2 +W 0_9 y 3 +W 0_10 x 4 +W 0_11 x 3 *y+W 0_12 x 2 *y 2 +W 0_13 x*y 3 +W 0_14 y 4 +W 0_15 x 5 +W 0_16 x 4 *y+W 0_17 x 3 *y 2 +W 0_18 x 2 *y 3 +W 0_19 x*y 4 +W 0_20 y 5 +W 0_21 x 6 +W 0_22 x 5 *y+W 0_23 x 4 *y 2 +W 0_24 x 3 *y 3 +W 0_25 x 2 *y 4 +W 0_26 x*y 5 +W0_27 y 6

[0024] f down f(x, y) = W 1_0 + W 1_1 x + W 1_2 y + W 1_3 x 2 + W 1_4 x * y + W 1_5 y 2 + W 1_6 x 3 + W 1_7 y * x 2 + W 1_8 x * y 2 + W 1_9 y 3 + W 1_10 x 4 + W 1_11 x 3 * y + W 1_12 x 2 * y 2 + W 1_13 x * y 3 + W 1_14 y 4 + W 1_15 x 5 + W 1_16 x 4 * y + W 1_17 x 3 * y 2 + W 1_18 x 2 * y 3 + W 1_ 19 x * y 4 + W 1_20 y 5 + W 1_21 x 6 + W 1_22 x 5 * y + W 1_23 x 4 * y 2 + W 1_24 x 3 * y 3 + W 1_25 x 2 * y 4 + W 1_26 x * y 5 + W 1_27 y 6

[0025] In the formula, f up(x, y) is the upper surface function, W 0_0 、W 0_1 、W 0_2 、W 0_3 …W 0_21 、W 0_22 、W 0_23 …W 0_27 are the coefficients of the equation of the upper surface function, f down (x, y) is the lower surface function, W 1_0 、W 1_1 、W 1_2 、W 1_3 …W 1_21 ,…W 1_27 are the coefficients of the equation of the lower surface function.

[0026] Preferably, the outer shell model determined by the shape fitting function has an X-axis and a Y-axis value range of [-12, 12] in the outer shell model coordinate system.

[0027] Preferably, the parametric representation of the assembly model includes:

[0028] Project multiple parts in the assembly onto the Z axis in the shell model coordinate system to obtain the projection height of the parts on the Z axis. If multiple parts intersect in the same projection range and the projection of the parts is in the positive direction of the Z axis, the maximum value of the part height is taken as the parametric representation of the part; if multiple parts intersect in the same projection range and the projection of the parts is in the negative direction of the Z axis, the minimum value of the part height is taken as the parametric representation of the part.

[0029] Preferably, the volume calculation of the shell model includes:

[0030] The bottom surface of the shell model is divided into small squares of 1mm*1mm, and the coordinates of the center point of each small square are obtained using the shell model coordinate system. The coordinates of the center point of each small square are substituted into the shape fitting function to obtain the upper surface function value and the lower surface function value. If the difference between the upper surface function value and the lower surface function value is greater than 0, then the difference is the height value of the corresponding cube, and then the volume of all cubes is calculated to obtain the volume of the shell model; if the difference between the upper surface function value and the lower surface function value is not greater than 0, then the volume of the corresponding cube is 0, and then the volume of all small squares is calculated to obtain the volume of the shell model.

[0031] Preferably, the constraints include:

[0032] When the x value of the part assembly model is within the x value range of the shell model, and the y value of the part assembly model is within the y value range of the shell model:

[0033] If the coordinate point after shape optimization is in the positive direction of the Z axis in the shell model coordinate system, the upper surface function value is calculated. If the difference between the upper surface function value and the part parameterization in the positive direction of the Z axis is greater than 0, the coordinate point after shape optimization is a feasible solution;

[0034] If the coordinate point after shape optimization is in the negative Z-axis direction in the shell model coordinate system, calculate the lower surface function value. If the difference between it and the part parameterization in the negative Z-axis direction is greater than 0, then the coordinate point after shape optimization is a feasible solution.

[0035] Preferably, the value of the optimization variable is {-8e-10, 30}.

[0036] Preferably, the initial scale of the particle swarm optimization algorithm is 200 particles, the number of iterations is 900 generations, and the particle dimension is 56.

[0037] Preferably, the support vector regression algorithm specifically includes the following steps:

[0038] Step 1: Create a local training set:

[0039] The size of the local training set is 200, and the local training set is initially empty. In the first iteration of the particle swarm optimization algorithm, the particle swarm is added to the local training set until the local training set reaches the size. At the beginning of the second iteration of the particle swarm optimization algorithm, 1 particle is randomly selected from every 4 particles, and 1 / 4 of the particles in each iteration of the particle swarm are selected to join the local training set, and the FIFO principle is implemented to maintain the size of the training set.

[0040] Step 2: Construct a support vector regression algorithm and update it with the iteration of the particle swarm optimization algorithm;

[0041] The penalty factor is set to 300, the gamma value is set to 0.003, and the updated local training set is input into the support vector regression algorithm at each iteration of the particle swarm optimization algorithm. Starting from the second iteration of the particle swarm optimization algorithm, the fitness function is calculated for the particles added to the local training set to train the support vector regression algorithm.

[0042] Step 3: Use the support vector regression algorithm to predict the fitness of the remaining 3 / 4 particles in each iteration of the particle swarm.

[0043] Preferably, the value of the fitness factor satisfies:

[0044] When the parameterized representation of any part exceeds the value of the shape fitting function under a certain shape optimization scheme, fit = 0;

[0045] When, under a certain shape optimization scheme, the parameterized representations of all parts are within the shape fitting function value, fit = 1.

[0046] The beneficial effects of the present invention are:

[0047] (1) The present invention designs and develops a voice robot appearance optimization method based on SVR-PSO, and establishes a new fitness evaluation strategy using the support vector regression algorithm. Compared with the standard particle swarm algorithm, this method greatly reduces the number of fitness calculations, can shorten the optimization iteration time, and improve the calculation accuracy, providing a reference idea for the processing of high-dimensional structure optimization problems.

[0048] (2) The SVR-PSO-based appearance optimization method for the voice robot designed and developed by the present invention establishes a design framework for optimizing the overall parameters of the voice recognition robot appearance. Solidworks software is used to perform 3D modeling on the outer shell and internal components of the voice recognition robot, and the parametric representation of the outer shell and the parametric representation of the parts assembly are realized, which can help solve the problem of difficult design of complex constraint functions in robot structure optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 The figure is a flow chart of the appearance optimization method of the voice robot based on SVR-PSO according to the present invention.

[0050] Figure 2 It is a structural schematic diagram of the shell model of the present invention.

[0051] Figure 3 It is a structural schematic diagram of the parts assembly model described in the present invention. DETAILED DESCRIPTION

[0052] The present invention is described in further detail below so that those skilled in the art can implement the invention with reference to the description.

[0053] like Figure 1 As shown in the figure, the present invention provides a voice robot appearance optimization method based on SVR-PSO, which takes the particle swarm optimization algorithm (PSO) as the basis and uses the support vector regression algorithm (SVR) to establish a new fitness evaluation strategy. The SVR model is used to predict the fitness values ​​of some individuals, aiming to reduce the number of fitness calculations, shorten the optimization iteration time, and improve the calculation accuracy, providing a reference idea for the processing of high-dimensional structure optimization problems.

[0054] The specific steps include:

[0055] Step 1: Model the outer shell and internal parts of the intelligent voice robot, including:

[0056] (1) Figure 2As shown, a picture of an existing intelligent voice robot is imported into Solidworks software (2016 version). By drawing its outer edge and adjusting the size, the outer curve corresponding to the intelligent voice robot can be obtained. The outer curve can be rotated to obtain the shell model of the intelligent voice robot.

[0057] (2) Figure 3 As shown, first complete the design of each part in Solidworks, modify the part size to meet the requirements, and then combine all the parts together to obtain the part assembly model;

[0058] The parts assembly model specifically includes: a first pickup microphone 110, a laser radar 120, an amplifier 130, a main board and hub 140, a middle fixing plate 150, a first speaker 161, a second speaker 162, a battery 170, a processor 180 and a chassis 190.

[0059] Step 2: parametrically represent the shell model and parts assembly model of the intelligent voice robot, specifically including:

[0060] 1. Parameterize the shell model of the intelligent voice robot:

[0061] (1) Use the Point Cloud Library (PCL) to export the point cloud of the shell model, that is, the set of coordinate points evenly distributed on the surface of the intelligent voice robot;

[0062] (2) Since coordinate point offsets may occur when PCL processes 3D models, the coordinate point set needs to be processed:

[0063] Considering that PCL obtains a uniformly distributed point cloud, we can calculate the mean of all coordinate points on the X-axis and Y-axis, and then subtract the mean from all coordinate points one by one to achieve the centering of the coordinate point set on the X-axis and Y-axis;

[0064] Among them, the mean of all coordinate points on the X-axis and Y-axis is calculated as:

[0065] The mean of all the coordinate points on the X axis is obtained by adding up all the X-axis values ​​in the coordinate point set and dividing it by the number of coordinates in the coordinate point set. The mean of all the coordinate points on the Y axis is obtained by adding up all the Y-axis values ​​in the coordinate point set and dividing it by the number of coordinates in the coordinate point set.

[0066] (3) Step 3: Perform polynomial fitting on the robot's outer surface based on the processed coordinate points:

[0067] The shell model coordinate system is established with the intersection of the central axis of the shell model and the upper and lower boundaries as the coordinate origin, the upper and lower dividing interfaces are used as the XOY plane, the central axis of the shell model is the Z axis, and upward is positive, and the coordinate points with z values ​​greater than zero are used for fitting the upper surface function, and the coordinate points with z values ​​less than zero are used for fitting the lower surface function to obtain the shape fitting function:

[0068] Because the shell model is a rotationally symmetrical figure, Figure 2 As shown in FIG, there is an obvious upper and lower dividing line in the shell model, so the shell model coordinate system is established with the upper and lower dividing lines of the shell model, the upper and lower dividing interfaces are taken as the XOY plane, the central axis of the shell model is the Z axis, and the vertical upward is the positive value of the Z axis, and the vertical downward is the negative value of the Z axis. Then, the height of the shape of the intelligent voice robot is divided, and the coordinate points with z values ​​greater than zero are used for fitting the upper surface function, and the coordinate points with z values ​​less than zero are used for fitting the lower surface function. Polynomial fitting is performed on these two parts of the coordinate points respectively, and R is used to calculate the shape of the intelligent voice robot. 2 The values ​​of are used to evaluate the polynomial regression models of different degrees. It can be found that the best fitting effect is achieved when both the upper and lower surface functions are sixth-order polynomials. The effect of the function fitting graph is R 2 The values ​​are 0.999785135463 and 0.971197127999 respectively, and the shape fitting function is obtained as follows:

[0069] f up (x, y) = W 0_0 +W 0_1 x+W 0_2 y+W 0_3 x 2 +W 0_4 x*y+W 0_5 y 2 +W 0_6 x 3 +W 0_7 y*x 2 +W 0_8 x*y 2 +W 0_9 y 3 +W 0_10 x 4 +W 0_11 x 3 *y+W 0_12 x 2 *y 2 +W 0_13 x*y 3 +W 0_14 y 4 +W 0_15 x 5 +W 0_16 x 4 *y+W0_17 x 3 *y 2 +W 0_18 x 2 *y 3 +W 0_19 x*y 4 +W 0_20 y 5 +W 0_21 x 6 +W 0_22 x 5 *y+W 0_23 x 4 *y 2 +W 0_24 x 3 *y 3 +W 0_25 x 2 *y 4 +W 0_26 x*y 5 +W 0_27 y 6

[0070] f down (x,y)=W 1_0 +W 1_1 x+W 1_2 y+W 1_3 x 2 +W 1_4 x*y+W 1_5 y 2 +W 1_6 x 3 +W 1_7 y*x 2 +W 1_8 x*y 2 +W 1_9 y 3 +W 1_10 x 4 +W 1_11 x 3 *y+W 1_12 x 2 *y 2 +W 1_13 x*y 3 +W 1_14 y 4 +W 1_15 x 5 +W 1_16 x 4 *y+W 1_17 x 3 *y 2 +W 1_18 x 2 *y 3 +W1_ 19 x*y 4 +W 1_20 y 5 +W 1_21 x 6 +W 1_22 x 5 *y+W 1_23 x 4 *y 2 +W 1_24 x 3 *y 3 +W 1_25 x 2 *y 4 +W 1_26 x*y 5 +W 1_27 y 6

[0071] Where, f up (x, y) is the upper surface function, W 0_0 、W 0_1 、W 0_2 、W 0_3 …W 0_21 、W 0_22 、W 0_23 …W 0_27 are the coefficients of the equation of the upper surface function, f down (x, y) is the lower surface function, W 1_0 、W 1_1 、W 1_2 、W 1_3 …W 1_21 ,…W 1_27 are the coefficients of the equation of the lower surface function.

[0072] 2. Parametric representation of the parts assembly model of the intelligent voice robot:

[0073] Considering that the position and shape of the parts are fixed and do not need to be changed like the outer shell of the intelligent voice robot, multiple parts in the parts assembly can be directly projected onto the Z axis in the outer shell model coordinate system, and the projection height of the parts on the Z axis can be used as the parametric representation of the parts;

[0074] Among them, if multiple parts intersect in the same projection range and the projection of the parts is in the positive direction of the Z axis, the maximum value of the part height is taken as the parametric representation of the part; if multiple parts intersect in the same projection range and the projection of the parts is in the negative direction of the Z axis, the minimum value of the part height is taken as the parametric representation of the part.

[0075] Step 3: Build an optimized shape model:

[0076] The minimum volume of the shell model of the intelligent voice robot is taken as the optimization target, the equation coefficients of the upper surface function and the lower surface function fitted by the three-dimensional coordinate points of the intelligent voice robot's outer surface are used as optimization variables, and the part assembly model is used as the constraint condition, that is, to judge whether there is a part passing through the upper surface or the lower surface (shell model) within the projection interval of the shell model coordinate system.

[0077] The present invention uses the idea of ​​the infinitesimal method to approximate the volume of the intelligent voice robot, that is, the area enclosed by the curved surface (upper curved surface and lower curved surface) and the plane (bottom surface) is divided into several small cubes, and the bottom area of ​​each small cube is recorded as Δx i , and record the maximum value of the bottom area as δ, the corresponding surface height as h, and then use the local "replacing the curve with a straight line" to get the volume of each small cube, and add up the volume of each cube to approximate the actual surface volume.

[0078] In the present invention, the bottom surface range is set to [-12, 12], and then the bottom surface is divided into small squares of 1mm*1mm. The coordinates of the center point of each small square are obtained using the shell model coordinate system, and the coordinates of the center point of each small square are substituted into the shape fitting function to obtain the upper surface function value and the lower surface function value. If the difference between the upper surface function value and the lower surface function value is greater than 0, then the value is retained as the height value of the cube, and then the volume of all cubes is calculated; if the difference between the upper surface function value and the lower surface function value is not greater than 0, it means that the lower surface of the cube is higher than the upper surface, which is obviously unreasonable. The volume of the corresponding cube is 0, and then the volume of all cubes is calculated to obtain the volume of the shell.

[0079] The specific constraint conditions for the part assembly model are:

[0080] Through the shape fitting function, we can know that the values ​​of x and y are in the interval [-12, 12]. Therefore, we only need to judge whether the optimization variables meet the part constraints within the interval [-12, 12]. That is, at a certain coordinate point, if the height difference between the upper surface function and the upper half of the part or the height difference between the lower surface function and the lower half of the part is less than 0, it means that the upper half of the part passes through the upper surface or the lower half of the part passes through the lower surface. Only when both differences are greater than 0, the obtained solution is a feasible solution to the optimization problem.

[0081] The optimization variables satisfy:

[0082] The shape of the intelligent voice robot is divided into two parts, the upper and lower parts, and the functional equations of the upper and lower surfaces are derived respectively by using polynomial function fitting. The coefficients of the functional equations are the parameters that need to be adjusted in the optimization process of the voice recognition robot to find the minimum volume, that is, the optimization variables. Each optimization variable has its own value range. After multiple experimental adjustments to each variable value, the maximum and minimum values ​​of each variable value are determined with the constraint of exceeding the part constraints. Then, the optimal value is searched by randomly adjusting the value of the optimization variable within its value range. The value range of the optimization variable is {-8e-10,30}.

[0083] Step 4: Optimize the shape optimization model using the particle swarm optimization algorithm to find the optimal solution for the current shape optimization model, specifically including:

[0084] Step 1: Initialize the particle swarm:

[0085] In the particle swarm optimization algorithm, particles are represented by optimization variables. The algorithm starts with 200 particles, and the position and velocity of each particle are randomly initialized. The number of iterations is set to 900 generations, and the particle dimension is the number of optimization variables, that is, 56.

[0086] Step 2: Determine the fitness function of the particle and make the fitness function of the particle as large as possible:

[0087]

[0088] Where fitness is the fitness function, v0 is the volume of the initial original shape, v is the volume corresponding to the new individual, and fit is the fitness factor;

[0089] When the projection height value of the upper part of the part is higher than the corresponding upper half of the individual surface or the projection height value of the lower part is lower than the height of the corresponding lower half of the individual surface within the projection range of the part, the part passes through the shell, then fit = 1, otherwise fit = 0.

[0090] Step 3: Reduce the number of objective function calculations using the support vector regression algorithm:

[0091] (1) Establish and maintain a local training set:

[0092] The size of the local training set is 200 and is initially empty. At the first iteration of the particle swarm optimization algorithm, the particle swarm is added to the local training set until the local training set reaches the size. At the beginning of the second iteration of the particle swarm optimization algorithm, the average sampling selection strategy is adopted. One particle is randomly selected from every four particles. The sample selected in this way is more representative. 1 / 4 of the particles in each iteration of the particle swarm are selected to join the local training set, and the FIFO principle is implemented. As many particles enter, as many are eliminated to maintain the size of the training set. Because after continuous iterative updates, the newly generated population is very different from the initial population. The use of the FIFO principle can ensure that the training set is similar to the latest population, thereby better describing the current particle distribution and making the regression results more accurate.

[0093] (2) Construct a support vector regression algorithm and train the SVR model through a local training set with the iteration of the particle swarm optimization algorithm. The size of the local training set is 200 and the acceleration constant is 2;

[0094] The essence of support vector regression algorithm is to solve the regression problem with the idea of ​​support vector machine, that is, given the training data D = {(x1, y1), (x2, y2), ..., (x m ,y m )}, we hope that f(x)=w T The solution obtained by x+b makes f(x) and y as equal as possible. w and b are model parameters. The support vector regression algorithm designs a buffer threshold ε for this calculation process, also called a slack variable. The loss is calculated only when the absolute value of the difference between the model output value f(x) and the true value y is greater than ε.

[0095] The SVR model of the present invention adopts Gaussian radial basis kernel function:

[0096]

[0097] As the kernel function in the support vector regression model, the penalty factor is set to 300 and the gamma value is set to 0.003.

[0098] Among them, in each iteration of the particle swarm optimization algorithm, the updated local training set is input into the support vector regression algorithm. Starting from the second iteration of the particle swarm optimization algorithm, the fitness function is calculated for the particles added to the local training set to train the support vector regression algorithm.

[0099] (3) By using the getmodel method to call the interface of the SVR model, the fitness of the remaining 3 / 4 particles in each iteration of the particle swarm can be predicted according to the SVR model, thereby reducing the number of fitness calculations of the particle swarm algorithm and shortening the running time of the algorithm.

[0100] Step 4: Update the individual optimal position pbest and the global optimal position gbest according to the fitness value of each particle.

[0101] Step 5: Determine whether the termination condition is met. If not, recalculate the fitness function value of each particle in the population and repeat the above operation. If the termination condition is met, determine the optimal particle based on the fitness function value of each particle in the current population, which is the optimal solution of the current shape optimization model.

[0102] The present invention designs and develops a voice robot appearance optimization method based on SVR-PSO. It combines the support vector regression algorithm with the particle swarm optimization algorithm and establishes a new fitness evaluation strategy using the support vector regression algorithm. The trained SVR model is used to predict the fitness of the individual to be evaluated, which reduces the number of fitness calculations in the particle swarm optimization algorithm, shortens the optimization iteration time, and improves the calculation accuracy. It provides a reference idea for the treatment of modern high-dimensional structural optimization problems.

[0103] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and embodiments shown and described herein.

Claims

1. A voice robot shape optimization method based on SVR-PSO, characterized in that: The steps include: Step 1: Model the outer shell and internal parts of the intelligent voice robot to obtain a shell model and a parts assembly model; Step 2: parametrically represent the shell model to obtain an outer shape fitting function, and parametrically represent the assembly model to obtain projection heights of multiple parts; Step 3: Build an optimized shape model: Taking the minimum volume of the shell model as the optimization target, the equation coefficients in the shape fitting function as the optimization variables, and the part assembly model as the constraint condition; Step 4: Optimize the shape optimization model using the particle swarm optimization algorithm. After reaching the number of iterations, the maximum value of the fitness function of the particle is the optimal solution of the shape optimization model, and the shape of the intelligent voice robot is adjusted to the minimum volume; Among them, the operation time of the particle swarm optimization algorithm is compared with the support vector regression algorithm; The fitness function of the particle is: In the formula, fitness is the fitness function, v0 is the volume of the initial original shape, v is the volume corresponding to the new individual, and fit is the fitness factor.

2. The SVR-PSO-based voice robot shape optimization method according to claim 1, characterized in that: The parameterized representation of the shell model comprises the following steps: Step 1: Export the point cloud of the shell model; Step 2: Centralize the point cloud: Calculate the mean of all coordinate points on the X-axis and Y-axis in the point cloud, and subtract the mean from all coordinate points one by one to obtain the processed coordinate points; Step 3: Perform polynomial fitting on the robot's outer surface based on the processed coordinate points: The shell model coordinate system is established with the intersection of the central axis of the shell model and the upper and lower boundaries as the coordinate origin, the upper and lower dividing interfaces are used as the XOY plane, the central axis of the shell model is the Z axis, and upward is positive, and the coordinate points with z values ​​greater than zero are used for fitting the upper surface function, and the coordinate points with z values ​​less than zero are used for fitting the lower surface function to obtain the shape fitting function: f up (x,y)=W 0_0 +W 0_1 x+W 0_2 y+W 0_3 x 2 +W 0_4 x*y+W 0_5 y 2 +W 0_6 x 3 +W 0_7 y*x 2 +W 0_8 x*y 2 +W 0_9 y 3 +W 0_10 x 4 +W 0_11 x 3 *y+W 0_12 x 2 *y 2 +W 0_13 x*y 3 +W 0_14 y 4 +W 0_15 x 5 +W 0_16 x 4 *y+W 0_17 x 3 *y 2 +W 0_18 x 2 *y 3 +W 0_19 x*y 4 +W 0_20 y 5 +W 0_21 x 6 +W 0_22 x 5 *y+W 0_23 x 4 *y 2 +W 0_24 x 3 *y 3 +W 0_25 x 2 *y 4 +W 0_26 x*y 5 +W 0_27 y 6 f down (x,y)=W 1_0 +W 1_1 x+W 1_2 y+W 1_3 x 2 +W 1_4 x*y+W 1_5 y 2 +W 1_6 x 3 +W 1_7 y*x 2 +W 1_8 x*y 2 +W 1_9 y 3 +W 1_ 10 x 4 +W 1_11 x 3 *y+W 1_12 x 2 *y 2 +W 1_13 x*y 3 +W 1_14 y 4 +W 1_15 x 5 +W 1_16 x 4 *y+W 1_17 x 3 *y 2 +W 1_18 x 2 *y 3 +W 1_19 x*y 4 +W 1_20 y 5 +W 1_21 x 6 +W 1_22 x 5 *y+W 1_23 x 4 *y 2 +W 1_24 x 3 *y 3 +W 1_25 x 2 *y 4 +W 1_26 x*y 5 +W 1_27 y 6 Where, f up (x, y) is the upper surface function, W 0_0 、W 0_1 、W 0_2 、W 0_3 …W 0_21 、W 0_22 、W 0_23 …W 0_27 are the coefficients of the equation of the upper surface function, f down (x, y) is the lower surface function, W 1_0 、W 1_1 、W 1_2 、W 1_3 …W 1_21 ,…W 1_27 are the coefficients of the equation of the lower surface function.

3. The SVR-PSO-based voice robot shape optimization method according to claim 2, characterized in that: The shape fitting function is used to determine that the value ranges of the X-axis and the Y-axis of the shell model in the shell model coordinate system are both [-12, 12].

4. The SVR-PSO-based voice robot shape optimization method according to claim 3, characterized in that: The parametric representation of the assembly model comprises: Project multiple parts in the assembly onto the Z axis in the shell model coordinate system to obtain the projection height of the parts on the Z axis. If multiple parts intersect in the same projection range and the projection of the parts is in the positive direction of the Z axis, the maximum value of the part height is taken as the parametric representation of the part; if multiple parts intersect in the same projection range and the projection of the parts is in the negative direction of the Z axis, the minimum value of the part height is taken as the parametric representation of the part.

5. The SVR-PSO-based voice robot shape optimization method according to claim 4, characterized in that: The volume calculation of the shell model includes: The bottom surface of the shell model is divided into small squares of 1mm*1mm, and the coordinates of the center point of each small square are obtained using the shell model coordinate system. The coordinates of the center point of each small square are substituted into the shape fitting function to obtain the upper surface function value and the lower surface function value. If the difference between the upper surface function value and the lower surface function value is greater than 0, then the difference is the height value of the corresponding cube, and then the volume of all cubes is calculated to obtain the volume of the shell model; if the difference between the upper surface function value and the lower surface function value is not greater than 0, then the volume of the corresponding cube is 0, and then the volume of all small squares is calculated to obtain the volume of the shell model.

6. The SVR-PSO-based voice robot shape optimization method according to claim 5, characterized in that: The constraints include: When the x value of the part assembly model is within the x value range of the shell model, and the y value of the part assembly model is within the y value range of the shell model: If the coordinate point after shape optimization is in the positive direction of the Z axis in the shell model coordinate system, the upper surface function value is calculated. If the difference between the upper surface function value and the part parameterization in the positive direction of the Z axis is greater than 0, the coordinate point after shape optimization is a feasible solution; If the coordinate point after shape optimization is in the negative Z-axis direction in the shell model coordinate system, calculate the lower surface function value. If the difference between it and the part parameterization in the negative Z-axis direction is greater than 0, then the coordinate point after shape optimization is a feasible solution.

7. The SVR-PSO-based voice robot shape optimization method according to claim 6, characterized in that: The value of the optimization variable is {-8e-10, 30}.

8. The SVR-PSO-based voice robot shape optimization method according to claim 7, characterized in that: The initial size of the particle swarm optimization algorithm is 200 particles, the number of iterations is 900 generations, and the particle dimension is 56.

9. The SVR-PSO-based voice robot shape optimization method according to claim 8, characterized in that: The support vector regression algorithm specifically includes the following steps: Step 1: Create a local training set: The size of the local training set is 200, and the local training set is initially empty. In the first iteration of the particle swarm optimization algorithm, the particle swarm is added to the local training set until the local training set reaches the size. At the beginning of the second iteration of the particle swarm optimization algorithm, 1 particle is randomly selected from every 4 particles, and 1 / 4 of the particles in each iteration of the particle swarm are selected to join the local training set, and the FIFO principle is implemented to maintain the size of the training set. Step 2: Construct a support vector regression algorithm and update it with the iteration of the particle swarm optimization algorithm; The penalty factor is set to 300, the gamma value is set to 0.003, and the updated local training set is input into the support vector regression algorithm at each iteration of the particle swarm optimization algorithm. Starting from the second iteration of the particle swarm optimization algorithm, the fitness function is calculated for the particles added to the local training set to train the support vector regression algorithm. Step 3: Use the support vector regression algorithm to predict the fitness of the remaining 3 / 4 particles in each iteration of the particle swarm.

10. The SVR-PSO based voice robot shape optimization method according to claim 9, characterized in that: The value of the fitness factor satisfies: When the parameterized representation of any part exceeds the value of the shape fitting function under a certain shape optimization scheme, fit = 0; When, under a certain shape optimization scheme, the parameterized representations of all parts are within the shape fitting function value, fit = 1.

Images