A multi-objective performance optimization method for engineering forklifts based on deep agent model
Through the deep proxy model-based method, using fully connected deep neural network and multi-task learning technology, the problem of high-dimensional nonlinear relationship representation of multi-design variables and multi-performance responses in multi-objective performance optimization of engineering forklifts is solved, achieving high precision and efficient Optimized design.
Patent Information
- Application Number
- CN202111308958.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-05
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2041-11-05
AI Technical Summary
The prior art is difficult to effectively optimize the multi-objective performance of engineering forklifts, especially in the high-dimensional and high-order nonlinear change characteristic relationship representation between multiple design variables and multi-performance responses, resulting in a large deviation from the design points from reality.
Using a deep agent model based method, a fully connected deep neural network is used as the agent model carrier. Through pre-training and fine-tuning, combined with real vehicle test data and working condition simulation data, an accurate deep agent model is constructed to achieve multiple goals for engineering forklifts. Performance optimization.
Effectively utilize the high-order nonlinear relationship representation ability between multiple inputs and multiple outputs of deep neural networks, solve the problem of difficult to construct strong generalization capabilities under a small amount of simulation data, and realize the high-precision and efficient simulation of engineering forklifts. optimization.
Smart Images

Figure CN114048564B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of engineering forklift development, and specifically to the application of global optimization and deep learning technology in engineering forklift development, and more specifically to a multi-objective performance optimization method for engineering forklifts based on a deep proxy model. Background Art
[0002] Engineering forklift is an industrial vehicle widely used in ports, mines, gardens, etc. for cargo transportation. Since its design optimization needs to comprehensively consider multiple performances, such as safety, comfort, economy, automation, intelligence, etc., these performances involve many influencing factors, and the various factors are coupled with each other, so that the simulation model often presents significant characteristics such as multidisciplinary and strong nonlinearity, and the joint simulation solution time is very long. To address this problem, academia and industry often use surrogate models (SurrogateModel) or response surface (Response Surface) methods, treating the simulation model as a black box function, and through computer experimental design, planning a small number of design points for simulation and performance response value approximation, obtaining a surrogate model that replaces the black box function, so as to minimize the number of calls to the black box function during the optimization process. It has now become an indispensable and important technical means in the forklift development process.
[0003] At present, academia or industry has proposed a variety of surrogate model methods, such as polynomial regression (PRS), spline function (MARS), artificial neural network (ANN), radial basis function (RBF), Kriging, support vector regression (SVR), sparse response surface (SRS), etc. However, the existing methods mainly focus on the characterization of the relationship between low-dimensional, few variables and single performance responses. When optimizing multiple target performances such as comfort and safety of forklifts at the same time, it is usually necessary to construct multiple surrogate models. In addition, the existing methods mainly focus on constructing surrogate models based on a small amount of simulation data planned on the simulation model of the current product, and pay less attention to the large amount of simulation data generated in the previous research and development process. There is a lack of reuse of the knowledge of the change characteristics between the design variables and performance responses contained in them. It may be difficult to support the characterization of the high-dimensional, high-order nonlinear change characteristic relationship between multiple design variables and multiple performance responses of forklifts, resulting in a large deviation between the optimized design point of the forklift and the actual situation.
[0004] The Chinese invention application with a publication date of 2021.05.28 and publication number of CN112861255A: In a scheme of an adaptive sampling method based on rocket stage inter-stage proxy optimization, an attempt is made to cope with optimization problems with multiple design variables, multiple extreme values, and strong nonlinearity through an adaptive sampling method based on fuzzy clustering and variable design space, in order to improve the local exploration and global search capabilities of sample refinement during the proxy optimization process, and to improve the efficiency of the optimization process through parallelized high-precision analysis of samples. However, this scheme is mainly aimed at the optimization of rocket aerodynamic shape and structural design, and is not suitable for applications in the development of engineering forklifts. Summary of the invention
[0005] In view of the limitations of the prior art, the present invention proposes a multi-objective performance optimization method for an engineering forklift based on a deep proxy model. The technical solution adopted by the present invention is:
[0006] A multi-objective performance optimization method for an engineering forklift based on a deep agent model comprises the following steps:
[0007] S1, according to the preset optimization goal and design variables, select and obtain the existing data set from the existing simulation data of the same type of engineering forklift as the engineering forklift to be optimized;
[0008] S2, using a fully connected deep neural network as a proxy model carrier of the engineering forklift to be optimized, and using the existing data set to pre-train the fully connected deep neural network to obtain a pre-trained deep proxy model;
[0009] S3, obtaining the actual vehicle test data of the engineering forklift to be optimized with respect to the optimization target, the virtual model of the engineering forklift to be optimized, and the working condition simulation data obtained by the virtual model; combining the actual vehicle test data and the working condition simulation data to obtain a mixed precision data set;
[0010] S4, fine-tuning the pre-trained deep proxy model using the mixed precision dataset to obtain an accurate deep proxy model;
[0011] S5, optimizing the performance of the engineering forklift to be optimized by using the precise deep proxy model to obtain an optimized design point of the engineering forklift to be optimized.
[0012] Compared with the prior art, the present invention makes full use of the deep neural network's strong ability to accurately characterize high-order nonlinear relationships between multiple inputs and multiple outputs and its good migration adaptability characteristics, and can effectively utilize a large amount of simulation data accumulated in the previous research and development process of existing forklifts of the same type; it solves the problem that it is difficult to construct a strong generalization proxy model for the current forklift under a small amount of simulation data under complex simulation optimization problems; for the virtual model of an engineering forklift with complex nonlinear change characteristics of multiple inputs and multiple outputs, only one proxy model needs to be constructed to achieve high-precision and efficient simulation optimization of the engineering forklift.
[0013] As a preferred solution, the step S2 iteratively obtains the pre-trained deep proxy model through the following process: using Bayesian optimization and Hyperband hybrid hyperparameter optimization method to screen and update the hyperparameter configuration of the fully connected deep neural network from a preset hyperparameter search space; according to the hyperparameter configuration, using the existing data set, and using the homoscedastic uncertainty multi-task learning method combined with the Adam algorithm to pre-train the fully connected deep neural network to obtain the proxy model; after reaching the preset number of pre-training iterations, the proxy model with the smallest verification loss in each round of iteration is used as the pre-trained deep proxy model.
[0014] As a preferred solution, step S3 includes the following process:
[0015] S31, constructing a whole vehicle virtual model, a control system virtual model and a three-dimensional model of a preset road condition of the engineering forklift to be optimized;
[0016] S32, performing driving condition simulation on the whole vehicle virtual model and the control system virtual model of the engineering forklift to be optimized in a three-dimensional model of preset road conditions, and obtaining a comparison error between the driving condition simulation result of step S32 and the actual vehicle test data;
[0017] S33, determining whether the comparison error is within a preset data error range, if so, executing step S34; otherwise, adjusting the parameters of the whole vehicle virtual model and the control system virtual model of the engineering forklift to be optimized according to the comparison error, and returning to step S32;
[0018] S34, according to the design variables, a number of design points are uniformly and randomly generated using a Latin hypercube design method; according to the design points generated in step S34, a driving condition simulation is performed on the whole vehicle virtual model and the control system virtual model of the engineering forklift to be optimized in a three-dimensional model of a preset road condition to obtain driving condition simulation data;
[0019] S35, combining the actual vehicle test data and the working condition simulation data to obtain a mixed precision data set.
[0020] As a preferred solution, in step S5, a decomposition-based multi-objective evolutionary algorithm is used to optimize the performance of the engineering forklift to be optimized.
[0021] As a preferred solution, the optimization target includes reducing the maximum roll angle of the frame and the amplitude of the driver's seat when the forklift to be optimized drives over a bump.
[0022] Furthermore, the design variable X=[x 1 ,x 2 ,x 3 ,x4 ,x 5 ] T , X∈Ω; where x 1 is the stiffness of the rubber shock absorber at the frame base, x 2 is the damping, x 3 is the cylinder hydraulic control system parameter, x 4 is the vehicle weight, x 5 is the centroid; Ω is the design space of the forklift to be optimized; Ω=[[1269,3269],[9,9.3],[100,750],[11478,13478],[850,950]] T .
[0023] Furthermore, the modeling object of the fully connected deep neural network is expressed by the following formula:
[0024] f(X)=[f1(X),f2(X)] T ;
[0025] Among them, f1(X) and f2(X) represent the maximum roll angle α of the frame and the amplitude z of the driver's seat, respectively.
[0026] The present invention also provides the following contents:
[0027] A multi-objective performance optimization system for an engineering forklift based on a deep proxy model, including an existing data set acquisition module, a pre-training module, a mixed precision data set acquisition module, a fine-tuning module and an optimized design point acquisition module, wherein the existing data set acquisition module is connected to the pre-training module, and the fine-tuning module is respectively connected to the pre-training module, the mixed precision data set acquisition module and the optimized design point acquisition module, wherein:
[0028] The existing data set acquisition module is used to select and obtain the existing data set from the existing simulation data of the same type of engineering forklift as the engineering forklift to be optimized according to the preset optimization target and design variables;
[0029] The pre-training module is used to use a fully connected deep neural network as a proxy model carrier of the engineering forklift to be optimized, and use the existing data set to pre-train the fully connected deep neural network to obtain a pre-trained deep proxy model;
[0030] The mixed precision data set acquisition module is used to acquire the real vehicle test data of the engineering forklift to be optimized with respect to the optimization target, the virtual model of the engineering forklift to be optimized, and the working condition simulation data obtained by the virtual model; and to combine the real vehicle test data and the working condition simulation data to obtain the mixed precision data set;
[0031] The fine-tuning module is used to fine-tune the pre-trained deep proxy model using the mixed precision dataset to obtain an accurate deep proxy model;
[0032] The optimization design point acquisition module is used to optimize the performance of the engineering forklift to be optimized by using the precise deep proxy model to obtain the optimization design point of the engineering forklift to be optimized.
[0033] A medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the aforementioned multi-objective performance optimization method for an engineering forklift based on a deep proxy model are implemented.
[0034] A computer device includes a medium, a processor, and a computer program stored in the medium and executable by the processor. When the computer program is executed by the processor, the steps of the aforementioned multi-objective performance optimization method for engineering forklifts based on a deep proxy model are implemented. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 A schematic diagram of the steps of a multi-objective performance optimization method for an engineering forklift based on a deep proxy model provided in Example 1 of the present invention;
[0036] Figure 2 This is a schematic diagram of the operating principle of step S2 of embodiment 1 of the present invention;
[0037] Figure 3 This is a schematic diagram of step S3 of Example 1 of the present invention;
[0038] Figure 4 Schematic diagram of a multi-objective performance optimization system for an engineering forklift based on a deep proxy model provided in Example 2 of the present invention. DETAILED DESCRIPTION
[0039] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;
[0040] It should be clear that the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present application.
[0041] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the embodiments of the present application. The singular forms of "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.
[0042] When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the attached claims. In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and do not have to be used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.
[0043] In addition, in the description of this application, unless otherwise specified, "multiple" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. The present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0044] In order to solve the limitations of the prior art, this embodiment provides a technical solution, and the technical solution of the present invention is further described below in conjunction with the drawings and embodiments.
[0045] Example 1
[0046] Please refer to Figure 1 , a multi-objective performance optimization method for engineering forklifts based on a deep agent model, comprising the following steps:
[0047] S1, according to the preset optimization goal and design variables, select and obtain the existing data set from the existing simulation data of the same type of engineering forklift as the engineering forklift to be optimized;
[0048] S2, using a fully connected deep neural network as a proxy model carrier of the engineering forklift to be optimized, and using the existing data set to pre-train the fully connected deep neural network to obtain a pre-trained deep proxy model;
[0049] S3, obtaining the actual vehicle test data of the engineering forklift to be optimized with respect to the optimization target, the virtual model of the engineering forklift to be optimized, and the working condition simulation data obtained by the virtual model; combining the actual vehicle test data and the working condition simulation data to obtain a mixed precision data set;
[0050] S4, fine-tuning the pre-trained deep proxy model using the mixed precision dataset to obtain an accurate deep proxy model;
[0051] S5, optimizing the performance of the engineering forklift to be optimized by using the precise deep proxy model to obtain an optimized design point of the engineering forklift to be optimized.
[0052] Compared with the prior art, the present invention makes full use of the deep neural network's strong ability to accurately characterize high-order nonlinear relationships between multiple inputs and multiple outputs and its good migration adaptability characteristics, and can effectively utilize a large amount of simulation data accumulated in the previous research and development process of existing forklifts of the same type; it solves the problem that it is difficult to construct a strong generalization proxy model for the current forklift under a small amount of simulation data under complex simulation optimization problems; for the virtual model of an engineering forklift with complex nonlinear change characteristics of multiple inputs and multiple outputs, only one proxy model needs to be constructed to achieve high-precision and efficient simulation optimization of the engineering forklift.
[0053] Specifically, this embodiment takes the safety performance and comfort performance of the engineering forklift to be optimized under driving conditions as optimization goals. More specifically, as a preferred embodiment, the optimization goals include reducing the maximum roll angle α of the frame and the driver's seat amplitude z when the engineering forklift to be optimized drives over a bump.
[0054] Furthermore, the design variable X=[x 1 ,x 2 ,x 3 ,x 4 ,x 5 ] T , X∈Ω; where x 1 is the stiffness of the rubber shock absorber at the frame base, x 2 is the damping, x 3 is the cylinder hydraulic control system parameter, x 4 is the vehicle weight, x 5 is the centroid; Ω is the design space of the forklift to be optimized; Ω=[[1269,3269],[9,9.3],[100,750],[11478,13478],[850,950]] T .
[0055] Specifically, the design parameters associated with the optimization target may also include tire parameters x 6 , and since the tires used in construction forklifts are generally already finalized, the tire parameters x 6 can be regarded as a constant. Among them, the stiffness of the rubber shock absorber at the frame base is x 1 The unit is (N / mm); damping x 2 The unit is (Ns / mm); the cylinder hydraulic control system parameter x 3 The unit is (mA); vehicle weight x 4 The unit is (kg); center of mass x 5 The unit is (mm);
[0056] Furthermore, the modeling object of the fully connected deep neural network is expressed by the following formula:
[0057] f(X)=[f1(X),f2(X)] T ;
[0058] Among them, f1(X) and f2(X) represent the maximum roll angle α of the frame and the amplitude z of the driver's seat, respectively.
[0059] The existing data set and the mixed precision data set need to be divided into training sets and validation sets respectively, which can be done using the K-Flod method.
[0060] As a preferred embodiment, the step S2 iteratively obtains the pre-trained deep proxy model through the following process: using Bayesian optimization and Hyperband hybrid hyperparameter optimization method (BOHB algorithm) to screen and update the hyperparameter configuration of the fully connected deep neural network from a preset hyperparameter search space; according to the hyperparameter configuration, using the existing data set, and using the homoscedastic uncertainty multi-task learning method combined with the Adam algorithm to pre-train the fully connected deep neural network to obtain the proxy model; after reaching the preset number of pre-training iterations, the proxy model with the smallest verification loss in each round of iteration is used as the pre-trained deep proxy model.
[0061] Specifically, the hyperparameter configuration of the fully connected deep neural network includes the network structure parameters of the fully connected deep neural network and the Adam algorithm parameters of the fully connected deep neural network during the pre-training process. Figure 2 In an optional embodiment, the specific content of step S2 is as follows:
[0062] S21, setting the search space of the hyperparameters of the fully connected deep neural network Contains the structural parameters of the fully connected deep neural network: the number of hidden layers of the neural network n∈[3,7], the number of neurons in each hidden layer k i ∈[2e 4 ,2e 11 ], i = 1, 2, ..., n; and the Adam algorithm parameters of the fully connected deep neural network in the pre-training process: learning rate l∈[1e -6 ,1e -2 ], training data batch size Learning step size s∈[0.9,1], learning rate reduction factor τ∈[0.8,1], first-order coefficient β1∈[0.8,1], second-order coefficient β2∈[0.9,1];
[0063] S22, using the uniform sampling (UD) method to generate n sets of hyperparameter configurations:
[0064] λ=[λ (1) ,…,λ (m) ] T ,
[0065] in, is the hyperparameter configuration, j=1,2,…,m,
[0066] S23, initialize the deep neural network parameters θ={w,b} corresponding to the hyperparameter configuration λ, where w is the vector form of the hidden layer neuron connection weight w, and b is the vector form of the hidden layer neuron bias b;
[0067] S24, training adjustment parameters w and b, and calculating the validation loss of the deep neural network model corresponding to the hyperparameter configuration λ;
[0068] S25, constructing hyperparameter configuration λ and validation loss based on Gaussian process (GP) Gaussian prior model of mapping relationship
[0069] S26, from the hyperparameter search space N sets of hyperparameter configurations are randomly generated in , N = 1000, and based on Calculate the expected improvement (EI) of N groups of configurations. The larger the EI value, the greater the probability that the validation loss of the proxy model constructed by the corresponding hyperparameter configuration is improved (in the present invention, a smaller validation loss indicates that the proxy model is improved);
[0070] S27, according to the EI value, retain the best M groups of N groups of hyperparameter configurations;
[0071] S28, use the Hyperband algorithm to further screen the M groups of hyperparameter configurations, and retain the I group with the best validation loss;
[0072] S29, merge group I of hyperparameter configurations and their corresponding validation losses into the existing hyperparameter configuration λ and validation loss in step 3.1.5 and execute step S25;
[0073] In step S2, after reaching the preset number of pre-training iterations, the pre-trained deep proxy model of the engineering forklift to be optimized is verified to have the smallest proxy model loss.
[0074] More specifically, in this embodiment, the verification loss can be calculated in the following manner:
[0075] Step 1: Define the joint loss function of network training as:
[0076]
[0077] Where K is the number of surrogate model outputs, y and f θ (X) are the simulation value corresponding to the design variable X and the predicted value of the surrogate model, v i is the learning weight of the ith output, is the training loss for the ith output.
[0078] Step 2: Use homoscedastic uncertainty to construct a Gaussian likelihood estimate of the model output:
[0079]
[0080] Where σ is the output loss weight or the observation noise parameter of the model.
[0081] Step 3: Use Adam to adjust θ and σ, maximize the likelihood estimation to construct the surrogate model, that is, minimize
[0082]
[0083] in,
[0084]
[0085]
[0086] Step 4: Input the validation set into the trained proxy model and calculate the validation loss of the model corresponding to the hyperparameter configuration λ And it is represented by a vector as:
[0087]
[0088] As a preferred embodiment, please refer to Figure 3 , the step S3 includes the following process:
[0089] S31, constructing a whole vehicle virtual model, a control system virtual model and a three-dimensional model of a preset road condition of the engineering forklift to be optimized;
[0090] S32, performing driving condition simulation on the whole vehicle virtual model and the control system virtual model of the engineering forklift to be optimized in a three-dimensional model of preset road conditions, and obtaining a comparison error between the driving condition simulation result of step S32 and the actual vehicle test data;
[0091] S33, determining whether the comparison error is within a preset data error range, if so, executing step S34; otherwise, adjusting the parameters of the whole vehicle virtual model and the control system virtual model of the engineering forklift to be optimized according to the comparison error, and returning to step S32;
[0092] S34, according to the design variables, a number of design points are uniformly and randomly generated using a Latin hypercube design method; according to the design points generated in step S34, a driving condition simulation is performed on the whole vehicle virtual model and the control system virtual model of the engineering forklift to be optimized in a three-dimensional model of a preset road condition to obtain driving condition simulation data;
[0093] S35, combining the actual vehicle test data and the working condition simulation data to obtain a mixed precision data set.
[0094] Specifically, step S3 can be implemented by applying software such as Pro / E, ADAMS and AMESim; Pro / E can be used to first construct a whole vehicle virtual model of the engineering forklift to be optimized and a three-dimensional model of the preset road condition, and the whole vehicle virtual model of the engineering forklift to be optimized is imported into ADAMS for configuration, and AMESim is used to construct a control system virtual model of the engineering forklift to be optimized, including virtual models of the hydraulic control system and the circuit control system; the simulation of this step is implemented by joint simulation of ADAMS and AMESim.
[0095] In step S34, a Latin hypercube design method is used to uniformly and randomly generate a small number of design points from the design space Ω and a driving condition simulation is performed in ADAMS to obtain the condition simulation data.
[0096] In step S4, the mixed precision data set is input into the pre-trained deep proxy model, and the deep neural network parameters θ are updated by fine-tuning the parameters of the entire model.
[0097] The specific contents of step S5 in this embodiment include:
[0098] S51, initializing MOEA / D parameters, including: optimization problem function f(X) = [f1(X), f2(X)] T , size is N P The population P, N P uniformly distributed weight vectors: The number of neighbors of the weight vector N B (used to construct the neighborhood set B i , i=1…N P ), maximum number of iterations MaxIT, ideal point Z * =(Z1,Z2) T ,in Select the penalty-based boundary intersection method (PBI) as the fitness function:
[0099] minimize g(X|μ,Z * )=d1+γd2,subject toX∈Ω,
[0100] in,
[0101]
[0102] d2=‖f(X)-(Z * +d1μ)‖,
[0103] γ>0 is the penalty parameter;
[0104] S52, uniformly and randomly obtaining an initial population P from the design space, and calculating an ideal point Z using the precise deep proxy model * , and further calculate the fitness (aggregation) function value g(X);
[0105] S53, with 90% probability and 10% probability, respectively, from the neighborhood set B i Select parents from the population P and construct a mating pool;
[0106] S54, for each design point X in the mating pool j Perform binary coding crossover and polynomial mutation to obtain offspring XX j ;
[0107] S55, using the precise deep proxy model to calculate the response value f(XX) of the offspring j ), and use the fitness function to calculate the response value g(XX j );
[0108] S56, environmental selection, that is, in all g(XX j ) <g(X j ), randomly select C design points to replace the design points in population P;
[0109] S57, update reference point Z * :
[0110]
[0111] S58, update neighborhood set B i ;
[0112] S59, if the preset maximum number of optimization iterations MaxIT has not been reached, iterate steps S53 to S58;
[0113] S510, select an optimal design point from P.
[0114] Example 2
[0115] A multi-objective performance optimization system for engineering forklifts based on deep agent model, see Figure 4, including an existing data set acquisition module 1, a pre-training module 2, a mixed precision data set acquisition module 3, a fine-tuning module 4 and an optimized design point acquisition module 5, wherein the existing data set acquisition module 1 is connected to the pre-training module 2, and the fine-tuning module 4 is respectively connected to the pre-training module 2, the mixed precision data set acquisition module 3 and the optimized design point acquisition module 5, wherein:
[0116] The existing data set acquisition module 1 is used to select and obtain the existing data set from the existing simulation data of the same type of engineering forklift as the engineering forklift to be optimized according to the preset optimization target and design variables;
[0117] The pre-training module 2 is used to use a fully connected deep neural network as a proxy model carrier of the engineering forklift to be optimized, and use the existing data set to pre-train the fully connected deep neural network to obtain a pre-trained deep proxy model;
[0118] The mixed precision data set acquisition module 3 is used to acquire the real vehicle test data of the engineering forklift to be optimized with respect to the optimization target, the virtual model of the engineering forklift to be optimized, and the working condition simulation data obtained by the virtual model; and to combine the real vehicle test data and the working condition simulation data to obtain a mixed precision data set;
[0119] The fine-tuning module 4 is used to use the mixed precision data set to fine-tune the pre-trained deep proxy model to obtain an accurate deep proxy model;
[0120] The optimized design point acquisition module 5 is used to optimize the performance of the engineering forklift to be optimized by using the precise deep proxy model to obtain the optimized design point of the engineering forklift to be optimized.
[0121] Example 3
[0122] A medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the multi-objective performance optimization method for an engineering forklift based on a deep proxy model in Example 1 are implemented.
[0123] Example 4
[0124] A computer device includes a medium, a processor, and a computer program stored in the medium and executable by the processor. When the computer program is executed by the processor, the steps of the multi-objective performance optimization method for engineering forklifts based on a deep proxy model in Example 1 are implemented.
[0125] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.
Claims
1. A multi-objective performance optimization method for engineering forklifts based on a deep proxy model, characterized in that: The following steps are involved: S1, according to the preset optimization goal and design variables, the optimization goal includes reducing the maximum roll angle of the frame and the vibration amplitude of the driver's seat when the optimized engineering forklift is driving over a bump, and the design variable X=[x 1 ,x 2 ,x 3 ,x 4 ,x 5 ] T , X∈Ω; where x 1 is the stiffness of the rubber shock absorber at the frame base, x 2 is the damping, x 3 is the cylinder hydraulic control system parameter, x 4 is the vehicle weight, x 5 is the centroid; Ω is the design space of the forklift to be optimized; Ω=[[1269,3269],[9,9.3],[100,750],[11478,13478],[850,950]] T ; Filter and obtain an existing data set from existing simulation data of the same type of engineering forklift as the engineering forklift to be optimized; S2, using a fully connected deep neural network as a proxy model carrier of the engineering forklift to be optimized, and using the existing data set to pre-train the fully connected deep neural network to obtain a pre-trained deep proxy model; S3, obtaining the actual vehicle test data of the engineering forklift to be optimized with respect to the optimization target, the virtual model of the engineering forklift to be optimized, and the working condition simulation data obtained by the virtual model; combining the actual vehicle test data and the working condition simulation data to obtain a mixed precision data set; S4, fine-tuning the pre-trained deep proxy model using the mixed precision dataset to obtain an accurate deep proxy model; S5, optimizing the performance of the engineering forklift to be optimized by using the precise deep proxy model to obtain an optimized design point of the engineering forklift to be optimized.
2. The multi-objective performance optimization method for engineering forklifts based on a deep proxy model according to claim 1 is characterized in that: The step S2 iterates to obtain the pre-trained deep proxy model through the following process: using Bayesian optimization and Hyperband hybrid hyperparameter optimization method to screen and update the hyperparameter configuration of the fully connected deep neural network from a preset hyperparameter search space; according to the hyperparameter configuration, using the existing data set, and using the homoscedastic uncertainty multi-task learning method combined with the Adam algorithm to pre-train the fully connected deep neural network to obtain the proxy model; after reaching the preset number of pre-training iterations, the proxy model with the smallest verification loss in each round of iteration is used as the pre-trained deep proxy model.
3. The multi-objective performance optimization method for engineering forklifts based on a deep proxy model according to claim 1 is characterized in that: The step S3 includes the following process: S31, constructing a whole vehicle virtual model, a control system virtual model and a three-dimensional model of a preset road condition of the engineering forklift to be optimized; S32, performing driving condition simulation on the whole vehicle virtual model and the control system virtual model of the engineering forklift to be optimized in a three-dimensional model of preset road conditions, and obtaining a comparison error between the driving condition simulation result of step S32 and the actual vehicle test data; S33, determining whether the comparison error is within a preset data error range, if so, executing step S34; otherwise, adjusting the parameters of the whole vehicle virtual model and the control system virtual model of the engineering forklift to be optimized according to the comparison error, and returning to step S32; S34, according to the design variables, a number of design points are uniformly and randomly generated using a Latin hypercube design method; according to the design points generated in step S34, a driving condition simulation is performed on the whole vehicle virtual model and the control system virtual model of the engineering forklift to be optimized in a three-dimensional model of a preset road condition to obtain driving condition simulation data; S35, combining the actual vehicle test data and the working condition simulation data to obtain a mixed precision data set.
4. The multi-objective performance optimization method for engineering forklifts based on a deep proxy model according to claim 1 is characterized in that: In the step S5, a decomposition-based multi-objective evolutionary algorithm is used to optimize the performance of the engineering forklift to be optimized.
5. The multi-objective performance optimization method for engineering forklifts based on a deep proxy model according to claim 1 is characterized in that: The modeling object of the fully connected deep neural network is expressed by the following formula: f(X)=[f1(X),f2(X)] T ; Among them, f1(X) and f2(X) represent the maximum roll angle α of the frame and the amplitude z of the driver's seat, respectively.
6. A multi-objective performance optimization system for engineering forklifts based on a deep proxy model, characterized in that: The method comprises an existing data set acquisition module (1), a pre-training module (2), a mixed precision data set acquisition module (3), a fine-tuning module (4) and an optimized design point acquisition module (5), wherein the existing data set acquisition module (1) is connected to the pre-training module (2), the fine-tuning module (4) is respectively connected to the pre-training module (2), the mixed precision data set acquisition module (3) and the optimized design point acquisition module (5), wherein: The existing data set acquisition module (1) is used to obtain the data according to the preset optimization target and design variables, wherein the optimization target includes reducing the maximum roll angle of the frame and the vibration amplitude of the driver's seat when the forklift to be optimized is driving over a bump, and the design variable X=[x 1 ,x 2 ,x 3 ,x 4 ,x 5 ] T , X∈Ω; where x 1 is the stiffness of the rubber shock absorber at the frame base, x 2 is the damping, x 3 is the cylinder hydraulic control system parameter, x 4 is the vehicle weight, x 5 is the centroid; Ω is the design space of the forklift to be optimized; Ω=[[1269,3269],[9,9.3],[100,750],[11478,13478],[850,950]] T ; Filter and obtain an existing data set from existing simulation data of the same type of engineering forklift as the engineering forklift to be optimized; The pre-training module (2) is used to use a fully connected deep neural network as a proxy model carrier of the engineering forklift to be optimized, and use the existing data set to pre-train the fully connected deep neural network to obtain a pre-trained deep proxy model; The mixed precision data set acquisition module (3) is used to acquire the real vehicle test data of the engineering forklift to be optimized with respect to the optimization target, the virtual model of the engineering forklift to be optimized and the working condition simulation data obtained by the virtual model; and to combine the real vehicle test data and the working condition simulation data to obtain the mixed precision data set; The fine-tuning module (4) is used to use the mixed precision data set to fine-tune the pre-trained deep proxy model to obtain an accurate deep proxy model; The optimized design point acquisition module (5) is used to optimize the performance of the engineering forklift to be optimized using the precise deep proxy model to obtain the optimized design point of the engineering forklift to be optimized.
7. A medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the multi-objective performance optimization method for an engineering forklift based on a deep proxy model as described in any one of claims 1 to 5 are implemented.
8. A computer device, characterized in that: The method comprises a medium, a processor, and a computer program stored in the medium and executable by the processor, wherein when the computer program is executed by the processor, the steps of the multi-objective performance optimization method for an engineering forklift based on a deep proxy model are implemented as described in any one of claims 1 to 5.
