Data-driven multidisciplinary design optimization method for AUV considering simulation unit time consumption difference

Abstract

The application belongs to the technical field of multidisciplinary design, and discloses a data-driven AUV multidisciplinary design optimization method considering time consumption difference of simulation units. The method comprises the following steps: generating initial samples for an optimization object and dividing the initial samples into design samples of each discipline sequence; analyzing high-fidelity evaluation average time consumption of the design samples of each discipline sequence, calculating expansion multiples of the design samples of other non-most time-consuming discipline sequences based on the design samples of the most time-consuming discipline sequence as a benchmark to construct design samples according to a total time consumption consistency principle, performing high-fidelity evaluation on the constructed design samples to construct a discipline sequence library; establishing a proxy model of each discipline sequence based on the discipline sequence library; and dividing a multidisciplinary system of the optimization object into each discipline sequence and performing rapid evaluation based on the corresponding proxy model. The method of the application significantly improves optimization efficiency by solving the problem of evaluation time heterogeneity in the process of multidisciplinary design optimization.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of AUV multidisciplinary design, and particularly relates to a data-driven AUV multidisciplinary design optimization method considering time consumption differences of simulation units. BACKGROUND

[0002] In the modern engineering field, the development of high-performance systems such as aerospace, automotive engineering and energy systems is increasingly complex, and the overall performance requirements of design schemes are increasingly stringent. In the traditional design process, different disciplines are often treated in isolation, and this separated design method is extremely likely to lead to a non-optimal overall design, and even conflicts between disciplines, which cannot meet the urgent needs of complex systems for efficient and reasonable design. Multidisciplinary design optimization (MDO) as a systematic engineering design method has emerged as the times require. Through a unified optimization framework, the models and constraints of multiple disciplines are integrated, and the coupling relationship between structure, aerodynamics, thermal control, control and other disciplines is coordinated, aiming to realize the collaborative optimization of the overall performance of complex systems, and is widely used in the development of the above high-performance systems.

[0003] The core of MDO is to solve the problem of multidisciplinary collaborative optimization through reasonable architecture design, advanced optimization algorithms and efficient computing resources. In the aspect of architecture design, various strategies have been developed, such as the multidisciplinary feasible (MDF) strategy, which performs a complete multidisciplinary analysis in each iteration to ensure the feasibility of each step, but the computational cost is high; the individual discipline feasible (IDF) strategy introduces auxiliary variables to reduce the dimension of the design space, but sacrifices the autonomy of disciplines; the simultaneous optimization (AAO) strategy decouples sub-disciplines by introducing auxiliary variables, has high parallelism and discipline autonomy, but increases the dimension of the design space and the consistency convergence condition. In the field of optimization algorithms, traditional gradient methods such as Newton's method and steepest descent method have slow convergence speed and are prone to fall into local optimum when dealing with large-scale problems. In order to overcome these problems, various modern optimization algorithms have been introduced, such as genetic algorithm (GA), particle swarm optimization (PSO), ant colony algorithm (ACO), simulated annealing (SA) and the like. These algorithms have the advantages of strong global search ability and independence from gradient information, and are suitable for complex MDO problems. In recent years, data-driven optimization methods have been widely used in MDO, especially in dealing with high-dimensional, nonlinear and multi-objective complex problems. It constructs proxy models (such as Kriging model, Gaussian process regression, neural network, etc.) to approximate complex physical models, effectively reduces the computational cost and improves the optimization efficiency.

[0004] With the increasing complexity and scale of engineering systems, distributed MDO has become a research hotspot. Distributed MDO improves computational efficiency and scalability by distributing optimization tasks to multiple computing nodes for parallel computing. Typical distributed MDO methods include concurrent subspace optimization (CSSO), collaborative optimization (CO), enhanced collaborative optimization (ECO), and two-level integrated algorithm (BLISS). For example, the CSSO method decomposes the system optimization problem into several sub-optimization problems with non-intersecting design variables. The optimization objective function of each subsystem is the original system optimization problem objective function, and each subsystem only optimizes local design variables. The required coupling input variables are replaced by an approximate model. The CO method hierarchically decomposes the system, and each subsystem only exchanges data with the parent system. The system layer objective function is the original problem objective function, and each subsystem minimizes the difference with the target scheme allocated by the system. The system layer coordinates each sub-discipline for parallel optimization through consistency constraints. The ECO method reverses the system layer and subsystem layer objective functions of the CO algorithm, enhancing the "influence" of each sub-discipline on the objective function. The BLISS method divides design variables into global design variables and discipline local design variables, constructs an approximate expression of the original problem objective function using Taylor expansion, and alternately optimizes global design variables and discipline local design variables until the system optimal solution is obtained. In addition, distributed MDO also uses high-performance computing technologies such as grid computing and cloud computing to improve the solution capability of large-scale MDO problems using distributed computing resources. Grid computing achieves the sharing and collaboration of computing resources through resource management, information services, security services, data migration and management mechanisms; cloud computing provides flexible computing resources and storage capacity to support large-scale data processing and analysis.

[0005] In the overall design of underwater vehicles, multi-disciplinary coupling problems are particularly prominent, involving complex interactions between disciplines such as hydrodynamics, structure, energy, and control. Due to the significant differences in the computational time of each discipline simulation unit, for example, high-precision flow field analysis (CFD) requires several hours, structural strength simulation requires minutes to hours, and control response simulation only requires seconds. This time heterogeneity poses a serious challenge to existing distributed multi-disciplinary design optimization methods. Current mainstream architectures (such as CSSO, CO, and BLISS) support parallel computing, but their static task allocation mechanisms cannot dynamically adapt to the time differences between disciplines, causing fast disciplines (such as control simulation) to idle due to waiting for slow disciplines (such as CFD). Meanwhile, data transfer congestion further restricts the optimization efficiency of new configurations such as wing-body blends, severely limiting engineering practicality. SUMMARY

[0006] In view of the evaluation time heterogeneity problem faced in the existing AUV multidisciplinary design optimization (MDO) process, that is, the significant difference in the time required by different disciplines in the evaluation link, which further adversely affects the overall optimization efficiency and rational allocation of computing resources, the present application proposes a data-driven AUV multidisciplinary design optimization method considering the time consumption difference of simulation units. The method and system aim to break through the key restricting factors of MDO in the aspects of evaluation efficiency improvement and computing resource management optimization from the system level, and build an optimization system with high efficiency, good scalability and high demand for engineering practice.

[0007] In order to achieve the above technical purpose, the following technical solutions are specifically adopted in the present application:

[0008] In one aspect of the present application, a data-driven AUV multidisciplinary design optimization method considering the time consumption difference of simulation units is provided, comprising the following steps:

[0009] S1: generating initial samples for the optimization object and dividing them into design samples of each discipline sequence; the discipline sequence is composed of disciplines and data interaction between disciplines;

[0010] S2: analyzing the high-fidelity evaluation average time consumption of the design samples of each discipline sequence, and calculating the expansion rate of the design samples of other non-most time-consuming discipline sequences based on the total time consumption consistency principle and taking the design samples of the most time-consuming discipline sequence as the benchmark;

[0011] S3: generating expanded samples according to the corresponding expansion rate for the design samples of each non-most time-consuming discipline sequence and replacing the nearest neighbor samples;

[0012] S4: performing high-fidelity evaluation on the design samples of each discipline sequence to build a discipline sequence library;

[0013] S5: establishing Kriging surrogate models for each discipline sequence based on the discipline sequence library;

[0014] S6: dividing the multidisciplinary system of the optimization object into each discipline sequence and performing rapid evaluation based on the corresponding Kriging surrogate model.

[0015] In one embodiment, step S1 includes: generating a plurality of initial samples for the entire design space of the optimization object according to the Latin hypercube sampling criterion, and dividing the initial samples into design samples corresponding to different discipline sequences.

[0016] In one embodiment, analyzing the average time consumption of each discipline sequence in step S2 includes: generating the same number of Latin hypercube sampling samples for the design samples of each discipline sequence, performing real physical simulation, recording the evaluation time of each discipline sequence design sample and calculating the arithmetic mean.

[0017] In one embodiment, step S3 comprises: generating augmented samples for each non-most time-consuming discipline sequence design sample according to corresponding expansion factor, deleting the closest equivalent size design sample from the remaining discipline sequence design samples split from the initial sample from the augmented samples, and using the remaining augmented samples and the remaining discipline sequence design samples split from the initial sample to build the surrogate model.

[0018] In one embodiment, the AUV multidisciplinary design optimization method further comprises step S7:

[0019] generating a child population according to a given optimization algorithm according to the current population state, and performing fast evaluation on the child population using the Kriging surrogate model;

[0020] periodically decomposing the frontier sample into each discipline sequence, triggering high-fidelity evaluation and surrogate model updating.

[0021] In one embodiment, periodically decomposing the frontier sample into each discipline sequence, triggering high-fidelity evaluation and surrogate model updating comprises:

[0022] when the optimization iteration reaches a preset period, decomposing the frontier sample design variables of the current population into discipline sequence design samples;

[0023] analyzing the high-fidelity evaluation average time consumption of each discipline sequence design sample, generating augmented samples for each non-most time-consuming discipline sequence design sample according to corresponding expansion factor and replacing the nearest neighbor sample;

[0024] triggering high-fidelity evaluation and surrogate model updating using the reconstructed discipline sequence design samples.

[0025] In another aspect of the present application, a heterogeneous distributed computing system for AUV multidisciplinary design optimization is provided, comprising:

[0026] a computing node configured to perform high-fidelity evaluation on each discipline sequence in parallel and build a discipline sequence library, and analyze the high-fidelity evaluation average time consumption of each discipline sequence, calculate the sampling size of other non-most time-consuming discipline sequences based on the most time-consuming discipline sequence as a benchmark according to the total time consumption consistency principle;

[0027] a training node configured to build a Kriging surrogate model for each discipline sequence based on the discipline sequence library to form a surrogate model library;

[0028] a data node configured to store the discipline sequence library, the surrogate model library and a computing sequence; the computing sequence is used as a storage buffer to cache design samples that have been allocated to the computing node but have not been evaluated;

[0029] The optimization node comprises an optimizer and a multi-disciplinary evaluation module; the optimizer generates a sub-generation population according to an optimization algorithm; and the multi-disciplinary evaluation module calls parameters in the agent model library to perform rapid performance prediction on the sub-generation population.

[0030] The application has the following beneficial effects:

[0031] The application provides an AUV multi-disciplinary design optimization method based on a discipline sequence and a heterogeneous distributed agent model, aiming to improve optimization efficiency and solve the problems of computational resource allocation and efficiency bottleneck faced by traditional methods in engineering applications. In view of the problems of high computational overhead and low resource utilization rate in the existing multi-disciplinary design optimization process, the application introduces an agent model to accelerate system performance evaluation, and combines a heterogeneous distributed computing architecture to realize more efficient task scheduling and resource management. In the embodiments, the application is superior to the traditional method in achieving the effect of each optimization objective under the same number of iterations; under the condition of the same working time, the constructed architecture can effectively avoid idle of the computing nodes and improve the overall parallel computing efficiency. In addition, the adopted heterogeneous distributed architecture is closer to the needs of actual complex engineering systems, has good adaptability and scalability, and significantly improves the operability and practicality of the method in actual engineering. Therefore, the application realizes innovative improvement on the multi-disciplinary design optimization method, has clear technical progress, and can bring significant performance improvement and application value to related engineering fields. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 FIG. 1 is a schematic diagram of a heterogeneous distributed multi-disciplinary design optimization process in the application;

[0033] Figure 2 FIG. 4 is a parameterized definition of the outer dimension of the BWBUG in the embodiment of the application;

[0034] Figure 3 FIG. 5 is a parameterized definition of the structural dimension of the BWBUG in the embodiment of the application;

[0035] Figure 4 FIG. 6 is a population state diagram of the final result output of the method of the embodiment of the application and the traditional MDF method; the result is better as it goes to the lower left corner in the diagram, the red dot is the final optimization result of the method of the application, and the blue dot is the final optimization result of the traditional method;

[0036] Figure 5is a comparison diagram of the method of the embodiment of the application and the traditional multidisciplinary feasible method; wherein the vertical axis IGD represents the reverse generation distance of the optimization result front population, the smaller the better, the horizontal axis represents the number of real evaluations, according to the content of the application, the proxy model-based discipline evaluation is performed every 100 times, then the current front population is taken to perform real evaluation once, the number in the bracket represents the number of real evaluations, the blue data is the traditional multidisciplinary feasible method, and the red data is the method of the embodiment of the application. DETAILED DESCRIPTION

[0037] The technical solutions of the application will be described clearly and completely in combination with specific embodiments, but those skilled in the art will understand that the following described embodiments are part of the embodiments of the application, not all the embodiments, and are only used to illustrate the application, and should not be regarded as limiting the scope of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0038] In one specific embodiment of the application, a multidisciplinary design optimization method based on discipline sequence and heterogeneous distributed proxy model is provided, which specifically comprises the following steps:

[0039] S1: Construct a heterogeneous distributed computing architecture, including a data node, a computing node, a training node and an optimization node.

[0040] The application proposes the concept of discipline sequence, establishes the input-output mapping relationship of numerical variables, and the discipline sequence satisfies: composed of data interaction between disciplines; the variable data of the input and output discipline sequence is in pure numerical form; cannot be further decomposed into smaller discipline sequences.

[0041] The data node is deployed on a storage-type node with large storage space and high data throughput, including a discipline sequence library, a proxy model library and a computing sequence.

[0042] The discipline sequence library is used to centrally store the time-consuming evaluation results generated by the discipline sequence in the evaluation process. These results include performance data of each discipline sequence under different design variable samples, providing an information basis for subsequent analysis and optimization. The proxy model library is responsible for managing the parameter configuration of the proxy model corresponding to each discipline sequence. The computing sequence acts as a temporary storage buffer, caching design samples that have been allocated to the computing node but have not been evaluated. These samples are stored in the computing sequence during the waiting evaluation process, so that the computing node can timely obtain and perform the evaluation task when there is idle resource.

[0043] The computing nodes are used to execute specific subject assessment tasks. Based on the computational requirements of the assigned subject assessment task, they are equipped with corresponding hardware. Different subject sequences have specific computing nodes, and these nodes operate independently in parallel. The computing nodes and data nodes interact with each other by: the computing nodes retrieving design samples from the data nodes' computing sequences via the network. These samples are transmitted in a specific data format and typically contain the design parameters required for each subject sequence. After completing the subject assessment of the design samples, the computing nodes send the assessment results back to the data nodes' subject sequence database via the network.

[0044] The training node is used for training and updating the surrogate model and interacts with the data node. Specifically, the training node retrieves high-fidelity data of each subject sequence design sample from the subject sequence library of the data node to train the surrogate model; the training node imports the surrogate model parameters of the training result into the surrogate model library of the data node; when subject sequence expansion is required, subject sequence design samples with given requirements are generated based on the Latin hypercube sampling criterion and input into the calculation sequence of the data node.

[0045] The optimization node includes an optimizer and a multidisciplinary evaluation module. The optimizer is an independent optimization algorithm with modular characteristics, outputting a progeny population with superior performance based on the input current optimization population. The optimizer interacts with the multidisciplinary evaluation module and the computational sequences of the data nodes. The optimizer inputs the superior progeny population into the multidisciplinary evaluation module for surrogate model substitution evaluation, predicting the response of a given discipline sequence. When the optimizer's iteration count reaches a given period, the frontier of the current population is decomposed into design samples for each discipline sequence and input into the computational sequences of the data nodes. The multidisciplinary evaluation module originates from the multidisciplinary system of the optimization object. According to the definition of discipline sequences, it reconstructs the multidisciplinary system of the optimization object into a discipline sequence network, establishing a surrogate model for each output variable of each discipline sequence. The multidisciplinary evaluation module interacts with the optimizer in the optimization node and the surrogate model library in the data nodes. When performing rapid evaluation of discipline sequences, the multidisciplinary evaluation module calls the parameters of each surrogate model in the surrogate model library in the data nodes, achieving rapid input-to-output evaluation and rapid prediction of system performance. The multidisciplinary evaluation module feeds back the prediction results of a given sample to the optimizer in the optimization node for the next round of optimization iteration.

[0046] Heterogeneous Distributed Multidisciplinary Design Optimization Process Reference Figure 1 As shown, the specific implementation steps are as follows:

[0047] S2: According to the discipline sequence definition, the multidisciplinary system of the optimization object is decomposed into independent discipline sequences, each of which represents a specific aspect or discipline field in the system, such as structural discipline, fluid dynamics discipline, etc. The average time consumption of high-fidelity evaluation of each discipline sequence is analyzed, and the expansion ratio of other non-most time-consuming discipline sequences is calculated based on the total time consumption consistency principle with the most time-consuming discipline sequence as the benchmark.

[0048] wherein the average time consumption of high-fidelity evaluation of each discipline sequence is specifically: generating the same number of Latin hypercube samples for each discipline sequence, performing real physical simulation, recording the evaluation time of each discipline sequence and calculating the arithmetic mean value as the average time consumption of the discipline sequence.

[0049] The expansion ratio of other non-most time-consuming discipline sequences is calculated based on the total time consumption consistency principle with the most time-consuming discipline sequence as the benchmark. The specific expansion ratio can be obtained by dividing the average time consumption of the most time-consuming discipline sequence by the average time consumption of other non-most time-consuming discipline sequences.

[0050] S3: Initialization, including initial population sampling, discipline sequence division, sequence performance evaluation and surrogate model establishment.

[0051] S31: The optimization node generates 100 initial samples based on the Latin hypercube criterion for the entire design space, and divides the 100 initial samples into design samples of different discipline sequences according to the discipline sequence decomposition principle, and inputs them into the calculation sequence of the data node.

[0052] S32: Identify the most time-consuming discipline sequence, and the training node generates expanded samples according to the corresponding expansion ratio for the design samples of each non-most time-consuming discipline sequence, and deletes the design sample closest to the expanded sample from the design sample of the discipline sequence divided from the initial sample, and inputs the remaining expanded sample into the calculation sequence of each discipline sequence in the data node.

[0053] S33: The calculation node corresponding to each discipline sequence calls the discipline sequence design sample in the corresponding calculation sequence for high-fidelity discipline sequence evaluation, and inputs the result into the discipline sequence library of the data node.

[0054] S34: The training node calls the sample data in the discipline sequence library to establish the Kriging surrogate model of the output data of each discipline sequence, and then saves the parameter data file of each surrogate model in the surrogate model library of the data node.

[0055] S4: The optimization node performs the iteration process of the optimizer and the multidisciplinary evaluation module, and performs high-fidelity evaluation when the preset evaluation period is reached.

[0056] S41: The optimizer inputs the offspring population generated by the optimizer according to the current population state and the type of the object to be optimized into the multidisciplinary evaluation module. For example, for single-objective optimization, SQP, GA, etc. can be selected, and for multi-objective optimization algorithm, NSGAII, etc. can be selected. The selection of the algorithm only needs to match the optimization problem, and the selection is independent.

[0057] S42: The multidisciplinary evaluation module is composed of proxy models of each discipline sequence, and the proxy model parameters of the proxy model library in the data node are called to quickly evaluate the design samples given by the optimizer and feed back the results to the optimizer.

[0058] S43: When the iteration of the optimizer reaches the preset period, the front sample of the current population is decomposed into discipline sequence design samples and input into the calculation sequence of the data node. The training node generates discipline sequence design samples of a specified size based on the total time consistent principle for each discipline sequence design sample, deletes the discipline sequence design samples of the same size closest to the initial sample generated in the discipline sequence, and inputs the remaining discipline sequence design samples into the calculation sequence of the data node.

[0059] In another embodiment of the present application, a heterogeneous distributed computing system for AUV multidisciplinary design optimization is provided, comprising:

[0060] The computing node is configured to perform high-fidelity evaluation of each discipline sequence in parallel and construct a discipline sequence library, and analyze the high-fidelity evaluation average time of each discipline sequence, and calculate the sampling size of other non-most time-consuming discipline sequences based on the total time consistent principle with the most time-consuming discipline sequence as the reference;

[0061] The training node establishes Kriging proxy models of each discipline sequence based on the discipline sequence library to form the proxy model library;

[0062] The data node is used to store the discipline sequence library, the proxy model library and the calculation sequence; the calculation sequence is used as a storage buffer to cache the design samples that have been allocated to the computing node but have not been evaluated;

[0063] The optimization node includes an optimizer and a multidisciplinary evaluation module; the optimizer generates an offspring population according to an optimization algorithm; the multidisciplinary evaluation module calls the parameters in the proxy model library to quickly predict the performance of the offspring population.

[0064] Embodiments

[0065] In an embodiment, it is applied to the multidisciplinary design optimization of a Blended-Wing-Body Underwater Glider (BWBUG). BWBUG has excellent lift-drag ratio and endurance performance due to its unique wing panel design and natural transition of the fuselage wing. Like traditional toroidal autonomous underwater vehicles, BWBUG is also a complex multidisciplinary system involving multiple coupled disciplines such as shape, fluid, structure, etc. Therefore, the effective implementation of the BWBUG multidisciplinary design optimization problem can prove that the present application is effective for the design optimization of AUV multidisciplinary systems.

[0066] A BWBUG multidisciplinary design optimization problem involving shape, fluid and structure disciplines is proposed in this embodiment. The optimization problem contains 18 design variables, which are 14 parameters of the BWBUG shape parameterized model and 4 parameters of the structure parameterized model. The optimization problem contains 3 objective variables, which are the shape volume obtained by maximizing the shape discipline evaluation, the lift-drag ratio obtained by maximizing the fluid discipline evaluation, and the skeleton mass obtained by minimizing the structure discipline evaluation. The optimization problem contains one constraint, which requires that the maximum skeleton stress is not greater than 10% of the material yield strength. Specifically, the optimization problem can be described as follows:

[0067]

[0068] wherein, V represents the volume of the BWBUG, C ld represents the lift-drag ratio of the BWBUG, M represents the skeleton mass of the BWBUG, x represents the design variable, represents the maximum skeleton stress, represents the material yield strength.

[0069] In this embodiment, the length of the BWBUG is specified as 1 meter, and the wingspan is 1.75 meters. The design variable ranges are as follows:

[0070]

[0071] wherein, the first to fourteenth variables belong to the shape discipline, and the fifteenth to eighteenth variables belong to the structure discipline. The following is the specific application process of this embodiment.

[0072] 1) Complete the hardware environment construction of the heterogeneous distributed computing platform

[0073] 11) Select different threads of a personal computer with a central processing unit (CPU) of Intel i7-10700 and a memory of 32GB as optimization nodes and training nodes.

[0074] 12) Select a commercial network attached storage (NAS) device with a storage capacity of 8 TB as the data node;

[0075] 13) Select a workstation with a dual EPYC7502 CPU and 256 GB of memory as the computing node.

[0076] 14) Use Category 6 network cables to connect the personal computer and the workstation to the data node, forming a local area network.

[0077] 2) Based on the file sharing folder function of the Windows operating system, the software environment of the heterogeneous distributed computing platform is built.

[0078] 21) In the data node, folders ComputingSequence are established to store the design variables of the computing sequence to be solved, folders DisciplinarySequenceLibrary are established to store the high-fidelity evaluation results of each disciplinary sequence, and folders SurrogateModelLibrary are established to store or update the computing results of the surrogate model.

[0079] 22) In the optimization node, folders DistributedMDOPlatform_NodeA are established to store the optimization program and the training program of the surrogate model.

[0080] 23) In the computing node, folders DistributedMDOPlatform_NodeB are established to store the high-fidelity evaluation program of each disciplinary sequence.

[0081] 24) Map the folders in the data node to the optimization node and the computing node, respectively, to ensure that the optimization node and the computing node can normally write, read, and delete files in each folder of the data node.

[0082] 3) The multidisciplinary analysis environment of the blended wing body underwater glider (BWBUG) is built on the computing node.

[0083] 31) The analysis environment of the shape discipline is built.

[0084] 311) In this embodiment, the shape generation logic of the BWBUG is to loft the parameter-controlled airfoil section along the parameter-controlled leading and trailing edge contour lines.

[0085] 312) The airfoil section is defined using CST, which is controlled by five parameters and only determines the shape, not the size.

[0086] 313) The leading edge profile starts from the center of the nose of the vehicle and consists of a third-order Bézier curve and a straight line segment. The trailing edge profile is similar. The leading and trailing edge profiles are controlled by nine parameters.

[0087] 314) Using the NACA airfoil as the reference and the CST rule as the parameterization requirement, generate the normalized airfoil section shape value points from five parameters and save them as a data file.

[0088] 315) Then, generate the leading and trailing edge contours based on the last nine parameters. For example... Figure 2 As shown, the nine variable parameters are, in order, D2, D3, D4, D5, Z1, Z2, Z3, Z4, and Z5. The control point of the leading edge Bézier curve is P. 11 and P 12 , where P 11 The z-coordinate is Z5, the x-coordinate is a constant 0, and P 12 The z-coordinate is Z4, and the perpendicular coordinate is taken as the intersection of this point and the extension of the leading edge straight segment. The coordinates of the leading edge straight segment are (Z3, D4) and (L, D5). The control point of the trailing edge Bézier curve is P. 21 and P 22 , where P 21 The z-coordinate is Z1, the x-coordinate is a constant D1, and P... 22 The z-coordinate is Z2, and the perpendicular coordinate is the intersection of Z2 and the extension of the trailing edge line segment. The coordinates of the trailing edge line segment are (Z3, D2+D4) and (L, D5+D3), respectively. The program outputs the coordinate values ​​of the leading and trailing edges based on the nine input parameters.

[0089] 316) Take 51 points evenly along the z-direction on the leading and trailing edge coordinate model points, and calculate the length of each point to determine the chord length of the airfoil under the current z coordinate. Then, the program calls the normalized airfoil model points generated in step 314), enlarges them proportionally to form 51 airfoil section model points, and saves them as data files.

[0090] 317) Write a script in the COMSOL software using Java language to read 51 airfoil section profile points and generate 51 solid surfaces. Then read the profile points of the leading and trailing edge contours and make the 51 solid surfaces loft along the contours to form a geometric solid. Use the software's built-in evaluation tool to evaluate its volume as an output parameter. Finally, save the geometric solid as a standard geometric file.

[0091] 32) After completing the analysis environment setup for the external morphology discipline, proceed with the setup of the analysis environment for the fluid dynamics discipline.

[0092] 321) Numerical simulation was performed on BWBUG underwater at a speed of 2 knots and an angle of attack of 2 degrees. The work was divided into two parts: mesh drawing and numerical simulation.

[0093] 322)Meshing is completed in ICEM by using automation script. Considering the symmetry of the evaluation results, a geometric model is established for half of the BWBUG and its flow domain. A long rectangle with a length of ten meters and a width of five meters is drawn with the symmetric plane of the BWBUG as a reference plane, and a cuboid is stretched five meters on one side of the BWBUG model. The remaining geometric entity is defined as the flow domain by using Boolean operation to subtract the solid geometry of the BWBUG.

[0094] 323)A boundary layer is established on the contact surface between the BWBUG and the flow domain, and the number of layers and the thickness of each layer of the boundary layer are automatically specified. Meanwhile, the flow domain grid is divided according to the requirement of from near to far and from fine to coarse, and the results are output as a grid file.

[0095] 324)The Fluent software calls the script to automatically read the grid file exported by ICEM, sets the fluid material to water, and solves the k-Ω model with a second-order SIMPLEC format.

[0096] 325)In the boundary conditions, the reference plane defined in step 322 is set as a symmetric condition, and the opposite side is set as a zero shear condition. The contact surface between the BWBUG and the flow domain is set as a no-slip condition. The front and lower surfaces of the flow domain are defined as velocity inlets with input components of 2cos2 and 2sin2, respectively. The back and upper surfaces of the flow domain are defined as pressure outlets.

[0097] 326)The surface integral of the BWBUG is calculated to define the lift and drag, and the lift and drag during the solution process are exported as data files. The upper limit of the iteration number is set to 500 times.

[0098] 327)After the CFD evaluation is completed, the computer thread continues to execute the main program, reads the lift and drag calculated in step 327, and calculates the lift-drag ratio by taking the ratio after adding the angle of attack condition.

[0099] 33)After completing the analysis environment setup of the shape and fluid dynamics disciplines, the analysis environment of the structure discipline is continued to be set up

[0100] 331)The structure discipline analysis mainly evaluates the stress of the designed BWBUG skeleton under the self-weight load. The skeleton cross-section is defined as a rectangle as shown in Figure 3 .

[0101] 332)In this embodiment, there are four input parameters for the structure discipline, which are , , and t. In step 315, the leading and trailing edge profiles of BWBUG are defined as a Bezier curve and a straight line, respectively. The intersection of the two lines, A and B, divides BWBUG into a fuselage and a wing. In the wing, five ribs are distributed along the x direction and two ribs are distributed along the wing direction. In the fuselage, three ribs are distributed along the x direction and four ribs are distributed along the z direction.

[0102] 333) The x direction distance from the head to point B is defined as d1, the x direction distance from the tail to point A is defined as d2, and the z direction distance from the symmetric end face of BWBUG to line segment AB is defined as d3. The x direction ribs on both sides of the fuselage are fixed, each being half the thickness t. The position of the middle rib is determined by the parameter control, which represents the ratio of the position of the middle rib to the total width of the fuselage. The actual distance is . The z direction rib on the right side of the fuselage center is connected to the two ribs of the wing, and the z direction ribs on both sides of the fuselage are controlled by and , where the z direction distance from the z direction rib near the head to the head is , and the z direction distance from the z direction rib near the tail to the tail is .

[0103] 334) The skeleton model of BWBUG is automatically drawn by COMSOL. First, the shape parameters are read in to determine the positions and sizes of the fuselage end face, the fuselage-wing interface, and the wing end face. Then the structure parameters , and are read in to determine the positions of the two endpoints of each rib. Next, the structure parameter t is read in to draw rectangles and parallelograms, and then stretch them along the y direction to form a solid geometric model. Finally, the geometric model generated in step 317 is read in, and a Boolean intersection operation is performed with the original geometric model to form the skeleton geometric model, and the skeleton mass is simultaneously exported.

[0104] 335) The skeleton geometric model drawn in step 334) is automatically meshed using COMSOL.

[0105] 336) Using the structural mechanics module in COMSOL, set the boundary condition as the fixed constraint of the fuselage section, and the load condition as the volume force equal to the buoyancy force applied to the entire skeleton. Automatically perform finite element solving to evaluate the stress distribution in the skeleton.

[0106] 337) After the calculation is completed, the maximum stress in the skeleton is automatically output.

[0107] 4) According to prior knowledge, after the completion of the multidisciplinary automatic analysis environment setup in step 3), three main analysis programs are written to solve the shape discipline sequence according to step 31), to solve the shape-fluid discipline sequence according to steps 31) and 32), and to solve the shape-structure discipline sequence according to steps 31) and 33).

[0108] 5) Optimization framework initialization. This step mainly performs the initialization process of the wing-body blended underwater glider multidisciplinary optimization work, including initial population sampling, discipline sequence division, sequence performance evaluation, and proxy model establishment, which specifically includes the following steps:

[0109] 51) The optimization node generates initial samples based on the Latin hypercube sampling criterion, which is divided into shape discipline sequence samples, shape-fluid discipline sequence samples, and shape-structure discipline sequence samples.

[0110] 52) In the design variable space of the shape-fluid discipline sequence and the shape-structure discipline sequence, generate four times the population size of the additional discipline sequence samples according to the Latin hypercube sampling criterion. The above three sets of discipline sequence samples are input into the calculation sequence.

[0111] 53) The calculation node performs the automatic evaluation process of each discipline sequence in step 4) in parallel, and stores the calculation results in the discipline sequence library of the corresponding data node.

[0112] 54) The training node retrieves the shape discipline sequence samples, shape-fluid discipline sequence samples, and shape-structure sequence sample data from the library, establishes Kriging proxy models for the BWBUG shape volume, lift-drag ratio, skeleton mass, and maximum skeleton stress with respect to the design variable space corresponding to the discipline sequence, and then saves the parameter data files of each proxy model in the proxy model library of the data node.

[0113] 6) Optimization node optimization. The RVEA in the optimization node generates a child population according to the current population state.

[0114] 7) Proxy model assisted BWBUG multidisciplinary system performance evaluation. The multidisciplinary analysis process established by the proxy model in the optimization node reads the child population generated by the RVEA to quickly predict the system performance.

[0115] 8) Discipline sequence evaluation. This step is continuously performed independently after step 5 is completed. The calculation node monitors the calculation sequence and starts the automatic evaluation process of each discipline sequence in parallel when there are discipline sequence samples to be evaluated, and stores the calculation results in the discipline sequence library of the corresponding data node.

[0116] 9) Agent model updating. This step is performed independently after step 5) is completed. When the training node monitors that a certain amount of new disciplinary sequence samples are generated in the disciplinary sequence library, it invokes the shape disciplinary sequence samples, shape-fluid disciplinary sequence samples and shape-structure disciplinary sequence samples data in the library, establishes the Kriging agent models of BWBUG shape volume, drag-diffusion ratio, skeleton mass and maximum stress of the skeleton with respect to the design variable space of the corresponding disciplinary sequence, and then saves the parameter data files of each agent model in the agent model library of the data node.

[0117] 10) It is judged whether the number of times of multi-disciplinary analysis process evaluation of agent model replacement in the optimization module reaches the upper limit of 500 times of function evaluation. If it is reached, the optimization process is stopped, and the current optimization population of the front sample population is output, which is decomposed into shape disciplinary sequence samples, shape-fluid disciplinary sequence samples and shape-structure disciplinary sequence samples, and is input into the calculation sequence. If it is not reached, it is judged whether the number of times of evaluation reaches the result output period of 100 times. If it is reached, the optimization continues, and the current optimization population of the front sample population is output, which is decomposed into shape disciplinary sequence samples, shape-fluid disciplinary sequence samples and shape-structure disciplinary sequence samples, and four times the size of the front sample population is generated in the design variable space of the shape-fluid disciplinary sequence and the shape-structure disciplinary sequence according to the Latin hypercube sampling criterion. Finally, the above three sets of samples are input into the calculation sequence, otherwise the iterative optimization is continued.

[0118] The embodiment method of the present application is compared with the traditional single-disciplinary feasible (IDF) method. In the IDF method, each sample is decomposed into shape disciplinary sequence, shape-fluid disciplinary sequence and shape-structure disciplinary sequence, which are input into each calculation node for high-fidelity evaluation, without considering the sample size expansion operation considering disciplinary time consumption. The embodiment method is represented as the experimental group (Experimental MHDCF), and the IDF method is represented as the control group (Control MHDCF).

[0119] After 50,000 iterations of the optimizer, the final frontiers of the experimental and control MHDCF are shown in Figure 4 It is obvious that the MHDCF considering the difference in time consumption performs better. The evolution process generates six non-dominated solution sets, and the convergence is quantified by the inverse generational distance (IGD) in Figure 5 It is obvious that the IGD of the MHDCF considering the difference in time consumption decreases faster and converges to a lower value compared with the control MHDCF. In addition, the numbers in the brackets in each point in Figure 5 represent the number of shape-fluid disciplinary sequence evaluations performed when obtaining the non-dominated set. It can be seen that the experimental MHDCF achieves better results with less consumption of expensive evaluation time.

[0120] Only the IGD performance of MHDCF considering time consumption difference is concerned, and the decline of this indicator proves the effectiveness of this framework. In addition, by comparing the evaluation difference of experimental MHDCF and control MHDCF in the experiment, the former achieves about 25% improvement, while the cost is only 198 times of the most expensive evaluation time consumption, while the latter needs 307 times of the most expensive evaluation time consumption. This shows that MHDCF considering time consumption difference has a significant advantage in the strategy of supplementing and completing expensive evaluation through the evaluation gap of each node.

[0121] Under the same evaluation conditions, the traditional multidisciplinary design optimization (MDO) method, whether it is integral or distributed, has a certain degree of resource waste due to not considering time difference. In contrast, this framework, with the help of proxy model and heterogeneous deployment, to a certain extent, alleviates the impact of time difference. The time-consuming less discipline sequence can obtain more effective samples. In addition to the advantages of MHDCF in computational efficiency, the implementation mode of this framework also has great research value. This framework decomposes and sets the distributed MDO workflow on different nodes, and realizes the interconnection between nodes through offline data rather than online instructions, thereby enhancing stability and ease of implementation. In addition, the fact that MHDCF is based on an integral architecture also increases its importance. It strictly separates the optimizer from the discipline analysis, allowing the optimizer and the framework to optimize independently. This function conforms to the design mode of complex equipment in practical applications and is suitable for practical use.

[0122] Although the embodiments of the present application are described above in combination with the drawings, the present application is not limited to the above specific embodiments and application fields, and the above specific embodiments are only illustrative and guiding, but not limiting. Those skilled in the art can make many forms under the inspiration of the present application and without departing from the scope protected by the claims of the present application, which all belong to the protection of the present application.

Claims

1. A data-driven AUV multidisciplinary design optimization method considering simulation unit time consumption difference, characterized in that, Comprising: S1: generating initial samples for optimization objects and dividing into discipline sequence design samples; The discipline sequence is composed of disciplines and data interaction between disciplines; S2: analyzing the high-fidelity evaluation average time of each discipline sequence design sample, calculating the expansion ratio of other non-most time-consuming discipline sequence design samples based on the total time consistency principle with the most time-consuming discipline sequence design sample as the benchmark; S3: generating expanded samples according to the corresponding expansion ratio for each non-most time-consuming discipline sequence design sample and replacing the nearest neighbor sample; S4: constructing a discipline sequence library by high-fidelity evaluation of each discipline sequence design sample; S5: establishing a Kriging surrogate model for each discipline sequence based on the discipline sequence library; S6: dividing the multi-disciplinary system of the optimization object into each discipline sequence, and performing rapid evaluation based on the corresponding Kriging surrogate model.

2. The AUV multidisciplinary design optimization method of claim 1, wherein, Step S1 includes: according to the Latin hypercube sampling criterion, generating a plurality of initial samples for the entire design space of the optimization object, and dividing the initial samples into design samples corresponding to different discipline sequences.

3. The AUV multidisciplinary design optimization method of claim 1, wherein, In step S2, analyzing the average time of each discipline sequence includes: generating the same number of Latin hypercube sampling samples for the design samples of each discipline sequence, performing real physical simulation, recording the evaluation time of each discipline sequence design sample and calculating the arithmetic mean.

4. The AUV multidisciplinary design optimization method of claim 1, wherein, Step S3 includes: generating expanded samples according to the corresponding expansion ratio for each non-most time-consuming discipline sequence design sample, deleting the same size design sample closest to the corresponding expanded sample from the discipline sequence design sample divided from the initial sample, and establishing a surrogate model according to the remaining expanded sample and the remaining discipline sequence design sample divided from the initial sample.

5. The AUV multidisciplinary design optimization method of claim 1, wherein, The multi-disciplinary design optimization method further comprises step S7: According to the given optimization algorithm, a child population is generated according to the current population state, and the Kriging surrogate model is used to perform rapid evaluation on the child population; Periodically decompose the frontier sample into each discipline sequence to trigger high-fidelity evaluation and surrogate model updating.

6. The AUV multidisciplinary design optimization method of claim 5, wherein, Periodically decomposing the frontier sample into each discipline sequence to trigger high-fidelity evaluation and surrogate model updating includes: When the optimization iteration reaches the preset period, the frontier sample design variables of the current population are decomposed into discipline sequence design samples; Analyzing the high-fidelity evaluation average time of each discipline sequence design sample, generating expanded samples according to the corresponding expansion ratio for each non-most time-consuming discipline sequence design sample and replacing the nearest neighbor sample; Triggering high-fidelity evaluation and surrogate model updating using the reconstructed discipline sequence design sample.

7. A heterogeneous distributed computing system for data-driven AUV multidisciplinary design optimization considering simulation unit time consumption difference, characterized in that, Comprising: A computing node configured to perform high-fidelity evaluation of each discipline sequence and construct a discipline sequence library in parallel; And analyzing the high-fidelity evaluation average time of each discipline sequence, calculating the sampling size of other non-most time-consuming discipline sequences based on the total time consistency principle with the most time-consuming discipline sequence as the benchmark; A training node establishes a Kriging surrogate model for each discipline sequence based on the discipline sequence library to form a surrogate model library; a data node for storing a discipline sequence library, a surrogate model library and a calculation sequence; the calculation sequence is used as a storage buffer to cache the design samples which have been assigned to the calculation node but not evaluated; an optimization node comprising an optimizer and a multidisciplinary evaluation module; the optimizer generates a child population according to an optimization algorithm; the multidisciplinary evaluation module calls the parameters in the surrogate model library to perform rapid performance prediction on the child population.

Images