34 results about "Continuous parameter" patented technology

The invention relates to the technical field of composite material forming and manufacturing, and particularly discloses a multi-stage multi-target process parameter optimization method under strong coupling in a composite material fiber placement process, which comprises the following steps: training a neural network prediction model according to fiber placement machine process parameter experimental data and formed part quality index experimental data; inputting process parameter sampling data of the fiber placement machine into the trained neural network prediction model so as to output quality index prediction data of the formed part; outputting a plurality of groups of optimal solutions according to the process parameter sampling data and the quality index prediction data through an NSGA-II algorithm; a target optimal solution is screened out from the multiple sets of optimal solutions through a TOPSIS algorithm, and technological parameter sampling data in the target optimal solution serve as final technological parameters; and verifying the final process parameters. Cooperative regulation and control of continuous parameters and discrete parameters in the fiber placement process can be achieved, defect control and mechanical property improvement of the formed part are both considered, and process support is provided for batch production of high-performance composite material components.

The invention provides a coverage-driven OSC2.0 format scene generation method and device, and the method comprises the steps: analyzing an OSC2.0 abstract scene file, extracting parameters and constraint conditions, and discretizing continuous parameters according to a preset step length; defining a coverage index based on a k-wise combination test idea, and determining a to-be-covered parameter value and a combination set P; calculating a potential value of each parameter combination com and a vulnerability hole set, preferentially selecting the com with the most vulnerabilities, and determining a parameter value of the com; updating the vulnerability set for the remaining unassigned parameters, selecting the parameter with the most vulnerabilities, determining the value of the parameter in combination with the k-wise combination coverage improvement effect, adding the value into the current com, repeating until the remaining parameters have no vulnerabilities, and completing assignment according to constraints to generate an OSC1.0 concrete scene; and adding the concrete scene into the test set, updating the coverage degree and the vulnerability set, and stopping generation until the coverage degree reaches the standard or no vulnerability exists. According to the application, the OSC2.0 and the OSC1.0 can be efficiently connected, the scale of the test set is controlled, and the coverage integrity is guaranteed.

The invention discloses a three-dimensional geological modeling and stress analysis system based on comprehensive geophysical prospecting, and relates to the field of geological geophysical prospecting analysis. The system comprises a vector acquisition unit, a mode clustering unit, a coordinate acquisition unit, a mode screening unit, a sequence construction unit and a sequence extraction unit, and is used for extracting Q geological state evolution sequences corresponding to Q stable geological clusters from N geological state evolution sequences; the sequence embedding unit is used for embedding the Q geological state evolution sequences into a pre-constructed three-dimensional geological model to generate a high-quality geological unit for stress analysis; the internal parameters of the finally generated high-quality geological units are in smooth transition along the depth direction, and the transverse adjacent units have good transverse transition due to a stable mode of underground space aggregation, so that the internal consistency of the three-dimensional geological model is remarkably improved, and sample data with stable structure and continuous parameters are provided for stress analysis.

The embodiment of the invention provides a space bounding box calculation method for a three-dimensional model. The method comprises the steps that continuous parameterized expression of the surface of a target three-dimensional model on a parameter domain is determined; according to the continuous parameterized expression, analyzing a plurality of profile curves obtained by projecting the target three-dimensional model on a plurality of mutually orthogonal coordinate planes in a three-dimensional coordinate system; according to the parameterized expressions of the plurality of profile curves, determining two-dimensional bounding boxes respectively corresponding to the plurality of profile curves on the corresponding coordinate planes; and performing extreme value selection operation on the boundary value of the two-dimensional bounding box in the coordinate axis direction to generate a three-dimensional space bounding box of the target three-dimensional model. A projection profile curve is directly analyzed through continuous parameterization expression, the geometric boundary of the model is locked in mathematics essence, the defect that extreme points are missed in discrete sampling is avoided, and the calculation precision is improved; and three-dimensional extremum search is converted into two-dimensional profile curve analysis and two-dimensional bounding box construction on a plurality of orthogonal coordinate planes, so that the calculation complexity is reduced.

The invention relates to the technical field of water treatment equipment multi-dimensional management, and discloses a water treatment equipment multi-dimensional management system based on an intelligent water affair cloud platform, and the system comprises the steps: calculating a dynamic alpha factor in real time through on-site off-gas detection data and a clear water reference mass transfer coefficient, and carrying out the real-time calculation of a dynamic alpha factor based on the monotonous relation between a bubble stagnation cap and mass transfer resistance; and performing inversion to obtain continuous parameters for quantifying the gas-liquid interface state. And further calculating the real-time sensitivity of the parameter to the air volume of the air blower, constructing an air volume and mass transfer nonlinear model, and solving the target air volume with the minimum energy consumption under the condition of meeting the technological oxygen demand constraint condition. And finally, the target air volume serves as a control instruction to be issued, meanwhile, interface recovery time is output according to the time change rate of the gas-liquid interface parameters, and dynamic energy saving and interface state collaborative management of the water treatment aeration system is achieved.

The application discloses a multiple-input multiple-output antenna decoupling performance prediction method based on deep learning, belongs to the technical field of computer-aided electromagnetic system design and prediction, is used for antenna decoupling performance prediction, and comprises the following steps: determining key size parameters of an antenna to be optimized; performing parameter scanning on the variation range of the key size parameters to obtain full-band S parameter simulation data and form an original data set; the data amount of the original data set is expanded by using an SMOTE algorithm to generate an enhanced data set; an antenna decoupling performance prediction model is constructed and trained, the antenna to be optimized is input, and corresponding full-band S parameter curve final prediction results are output. The application effectively expands high-dimensional data by introducing the SMOTE algorithm, solves the problem that small sample data in the electromagnetic simulation field cannot support deep learning model training, constructs a DNN model to establish a nonlinear mapping of a continuous S parameter curve, and can comprehensively evaluate the S parameters of the antenna, instead of being limited to discrete frequency points.

This invention proposes a safety assessment method for delaminated composite materials based on the principle of damage non-propagation, belonging to the fields of composite material delaminated damage modeling and composite structure safety assessment. The method includes: constructing ultimate load envelopes that meet the damage non-propagation requirement under different damage sizes; for readily available binary observation data in practical engineering, i.e., whether damage propagates under specific damage sizes and load conditions, connecting discrete observations with a continuous parameter space through a latent variable model; using the Probit link function to map the predicted difference between the load and the ultimate load into a damage propagation probability; and constructing a likelihood model using a Bernoulli likelihood function; and optimizing the ultimate load envelope within a Bayesian framework. This invention can effectively support safety assessment and maintenance decisions for composite material structures in aerospace and other fields.

The invention discloses a genetic algorithm-based tight gas reservoir horizontal well fracturing section automatic division method. The method comprises the following steps of S1, determining a reasonable fracturing section division principle; s2, performing quantitative expression on the key parameters of the fractured section of the tight gas reservoir horizontal well; s3, acquiring continuous parameter values of the key parameters of the target layer section of the horizontal well according to the logging curve data, and respectively constructing a porosity vector, a Young modulus vector, a Poisson's ratio vector and a minimum horizontal principal stress vector of each fracturing section; s4, taking the standard deviation minimization of the attribute values of the fracturing sections as an optimization target, and establishing a mathematical model for dividing the fracturing sections of the single well; and S5, fracturing section division is converted into an optimization problem, and a genetic algorithm (GA) is adopted for efficient solving. According to the method, automatic division of the tight gas reservoir horizontal well fracturing sections can be rapidly and efficiently achieved, balanced crack initiation of hydraulic fractures is facilitated, the fracturing effect is improved, and tight gas reservoir development is optimized.

The invention discloses a soil environment detection method based on a gun-shaped detection gun, which is used for rapid field detection of a soil environment, and is characterized in that a multi-sensor probe array is integrated through gun configuration equipment, a probe can be inserted into soil by pulling a trigger once by a user, and in the insertion process, the equipment utilizes a probe insertion kinetic model to perform self-adaptive adjustment, so that the detection accuracy is improved. The method is used for overcoming resistance of different soil firmness and synchronously acquiring parameters such as soil humidity, pH value, conductivity, temperature and nitrogen phosphorus and potassium nutrients, and the core of the method is that depth profile reconstruction and intelligent decision are realized through algorithm processing: measured data of discrete depth points are reconstructed into a continuous parameter vertical distribution curve; according to the method, multi-source data is subjected to fusion and uncertainty quantification, a comprehensive report containing multi-parameter values, profile curves, soil quality indexes and risk levels is finally output, the detection time is shortened from several hours to about 30 seconds, and the efficiency, depth and intelligent level of field detection are greatly improved.

The invention discloses a roadway rockburst risk level total-space prediction method and system, and relates to the technical field of rockburst disaster risk prediction. The method comprises the steps that rock mass mechanical parameters collected through a plurality of discrete monitoring points in a roadway target area are obtained; performing spatial interpolation processing on the rock mass mechanical parameters, and constructing a continuous parameter field covering the target area; inputting the continuous parameter field into a pre-trained rockburst prediction model; the rockburst prediction model performs feature extraction on the continuous parameter field through a Transform module, inputs the extracted features into an Adaboost module for integrated classification, and outputs a rockburst risk level prediction result; and based on the rockburst risk level prediction result, generating a rockburst risk level three-dimensional visualization graph of the target area. According to the method, the full-space continuous prediction and three-dimensional visual expression of the rockburst risk can be realized, the fragmentation limitation of traditional local monitoring is overcome, and the accuracy and decision-making efficiency of disaster early warning are remarkably improved.

The invention discloses a network element data acquisition method and device. The method comprises the following steps: receiving network element information of target network element equipment; a prediction model is adopted to analyze network element information of the target network element equipment to obtain a data acquisition strategy of the target network element equipment, the prediction model is obtained based on fusion feature training, fusion features are determined according to discrete parameter features and continuous parameter features, and the discrete parameter features comprise a manufacturer strategy inheritance relation graph, a manufacturer strategy inheritance relation graph and a manufacturer strategy inheritance relation graph; the manufacturer strategy inheritance relation graph is used for representing an acquisition strategy inheritance relation between network element devices of different manufacturers, and the continuous parameter characteristics are used for representing a time sequence dependency relation between parameters; and collecting data of the target network element equipment according to the data collection strategy of the target network element equipment. The method provided by the invention at least solves the technical problem of low acquisition success rate caused by low matching degree between the network element data acquisition strategy and the network element in the related technology.

The invention provides a simulation configuration generation method and system based on a physical-numerical constraint feasible region. The method comprises the following steps: constructing a unified simulation parameter vector according to a continuous parameter set and a discrete parameter set; calculating a key dimensionless physical quantity according to the continuous parameter set; constructing a physical consistency constraint set and a numerical stability constraint set, and constructing a feasible region according to the physical consistency constraint set and the numerical stability constraint set; generating an initial candidate parameter vector from a natural language or a semi-structured engineering demand input by a user through semantic analysis and parameter extraction; judging whether the initial candidate parameter vector belongs to a feasible region or not; if yes, generating a simulation configuration file; if not, executing a hierarchical projection correction strategy; intelligent generation of simulation configuration is achieved, the success rate of one-time starting is remarkably increased, the number of manual parameter adjustment times is reduced, dependence of simulation analysis on personal experience of engineers is reduced, and efficient and reliable technical support is provided for numerical simulation of complex engineering problems.

The invention relates to a method (100) for producing a gas analysis device (10) having at least one separating device (12) and a plurality of pneumatic modules (20). The method (100) comprises a first step (110) in which at least one target parameter (62) for the gas analysis device (10) to be produced is preset and a plurality of basic aerodynamic structures (35) are provided. In a second step (120), a plurality of continuous and discrete parameters (42, 44) of the basic aerodynamic structure (35) are provided. A basic configuration (50) is generated from the basic aerodynamic structure (35), said basic configuration being determined by changing (55) the discrete parameter (42). In a third step (130) of the method (100), candidate aerodynamic structures (60) are determined from each basic configuration (50), at least one consecutive parameter (44) of each basic configuration (50) being varied by means of an optimization algorithm (72). Furthermore, in a fourth step (140), a candidate pneumatic structure (60) is selected as a function of the desired value (64) of the target parameter (62) and output to the user and / or data interface. In addition, the gas analysis device (10) is produced on the basis of the selected candidate pneumatic structure (66). The invention also relates to a corresponding computer program product (70), a gas analysis device (10), a simulation method (200) and a simulation program product (80).

This invention discloses a method and system for 3D reconstruction and neural rendering of large-scale scenes. The method includes the following steps: S1, constructing a 3D scene representation model based on anchor points; S2, training the 3D scene representation model based on multi-view images, the training process including the following sub-steps: S21, periodically calculating the generation efficiency score of each anchor point and removing anchor points with generation efficiency scores lower than a dynamic threshold; S22, using a global codebook to perform vector quantization on the parameter set of the anchor points, mapping the continuous parameter set to discrete codebook indices; S3, rendering using the trained 3D scene representation model, adjusting the opacity of each Gaussian primitive generated by the anchor point according to the viewing distance from the anchor point to the camera through a modulation function to reduce the rendering computation of distant areas, thereby significantly reducing the model storage size and improving the rendering speed without losing visual quality.

The application provides a radar jamming decision and parameter optimization method and device based on hierarchical reinforcement learning, which comprises the following steps: constructing a radar countermeasure model; according to the radar countermeasure model, establishing state-action samples of radar working modes and jamming styles; according to the collected samples, an outer Q-Learning network obtains an optimal jamming strategy; an environment of an inner DDPG network is constructed, and the jamming strategy obtained by the outer layer is mapped to the inner DDPG network; an inner DDPG network model is constructed, the optimal jamming strategy is pulse parameter optimized by selecting actions on a continuous parameter space; an interference effect evaluation algorithm based on AHP-TOPSIS is used to evaluate the interference effect, and the interference effect evaluation result is used as environment feedback to update the radar countermeasure model; the application is suitable for cognitive jamming decision and parameter optimization in radar electronic warfare, the interference effect is evaluated by using the evaluation algorithm based on AHP-TOPSIS, and then the jamming decision and parameter optimization are carried out by using the model based on hierarchical reinforcement learning.

The present application provides a pile foundation construction surrounding stratum deformation prediction method and system, belonging to the technical field of geotechnical engineering monitoring and numerical prediction, comprising: S1: determining the calculation domain range; S2: laying monitoring equipment; S3: selecting natural circulation load source; S4: inverting sample calculation small strain constraint modulus; S5: correcting the hysteresis loop and small strain constraint modulus of low-permeability fine-grained soil layer; S6: statistical uncertainty; S7: identifying low-modulus abnormal section and high-modulus section; S8: inputting parameters into the pile foundation construction surrounding deformation prediction model to obtain the prediction result; S9: if the prediction result deviates more than the threshold value, constructing a calibration objective function, updating the parameters, and outputting rolling prediction. The method can realize low-disturbance parameter acquisition; can realize continuous parameter acquisition, overcoming the shortcomings of discrete test points; can accurately identify soft interlayers; can update and rolling predict online, reducing the inversion multi-solution; can output early warning information for engineering application; and has wide applicability.

The invention discloses an unmanned aerial vehicle communication fault diagnosis method based on Bayesian decision and man-machine cooperation, and belongs to the technical field of unmanned aerial vehicle fault prediction and health management. The method comprises the following steps: firstly, acquiring three types of multivariate operation parameters including a link state, a hardware parameter and a flight state through a real-time acquisition module, and preprocessing by adopting a multi-mode adaptive filtering and parameter discretization unit; secondly, a fault knowledge base is constructed, the prior probability and the likelihood probability are calculated based on the labeled samples, Gaussian kernel density estimation is adopted for continuous parameters, and polynomial distribution modeling is adopted for discrete parameters; further, constructing a maximum posterior probability decision engine based on the Bayesian theorem, and outputting a fault diagnosis result and a confidence level; finally, through a human-in-the-loop mechanism, expert intervention is started when the confidence coefficient is insufficient or manual rechecking errors exist, and the probability model is dynamically updated by utilizing a correction sample, so that system self-optimization is realized. According to the method, the three problems of insufficient multi-source parameter coupling analysis, poor static rule base adaptability and low manual intervention efficiency are effectively solved, and the fault diagnosis accuracy and the system adaptive capacity in a complex environment are remarkably improved.

This invention relates to the field of multi-axis CNC machining and industrial robot machining technology, specifically disclosing an asymmetric five-axis toolpath smoothing method based on path weighted synthesis. The method includes: constructing a universal corner smoothing model based on path weighting; constructing an analytical expression for a high-order continuous weighted function; Cartesian space smoothing of the tool tip position path; unit spherical space smoothing of the tool axis direction path; synchronous adjustment of C3 continuous parameters for position and direction paths; interference verification and asymmetric adjustment of adjacent corner transition intervals; continuous speed planning and real-time interpolation; and motion control of the machining equipment. This invention overcomes the shortcomings of existing five-axis path transition smoothing technologies, such as poor geometric versatility, incoordination between tool tip and tool axis movements, easy interference in short path segments, and inconsistent bidirectional machining trajectories. It improves geometric versatility and processing efficiency, achieves high-order continuity and consistent machining surfaces, and can be ported across platforms and extended to higher orders.

The invention discloses a radar interference signal classification and identification method based on multi-dimensional, multi-domain and multi-frame feature extraction. The radar interference signal classification and identification method comprises the implementation steps that a randomized interference identification training framework is built; performing multi-dimensional signal feature cross-domain joint extraction to form a high-dimensional feature vector; constructing a distance-speed dimension motion state mapping matrix; the high-dimensional feature vector and the mapping matrix are fused, and a classification recognition model is trained; and outputting the radiation source interference type label. According to the method, a randomized interference identification training framework is used, an identification sample library covering a continuous parameter space is constructed, multi-dimensional and multi-domain feature information of multi-frame data is comprehensively extracted, a high-dimensional feature vector is synthesized, and a classification model is trained in combination with a distance-speed dimension motion state mapping matrix. Accurate identification of the type of the active interference in a complex electromagnetic environment with suppressing interference and pull-off deception interference is realized.

This invention discloses a method for constructing evaluation indicators for power batteries of pure electric vehicles based on big data analysis, including the following steps: A1. Marking the vehicles; A2. During the operation and use of the P1th vehicle; A3. The P2th...P... n Repeat steps A2 and A5 for each new car, and remove the first DB. ND The parameters are based on the data before the 1% mark and the data after the 99th mark following a normal distribution. A6. Use the labeled samples as the target values. A7. Use neural network models MM1...MM. n The calculated values ​​at each sampling point are used as the dependent variable. A8. Regression analysis is used. A9. One or more regression equations with the lowest overall error rate are extracted from the multiple regression equations and used as a continuous parameter prediction model. Compared with traditional neural network model calculation methods, this invention significantly reduces the computational load. Compared with single correlation regression prediction methods, it has better accuracy and adaptability, and has broad practical application value.

This invention discloses a method and system for optimizing environmental control strategies in facility agriculture based on a large language model, relating to the fields of smart agriculture and environmental control technology. The invention constructs a four-layer closed-loop architecture: "data layer - simulation layer - evaluation layer - optimization layer," achieved through the following steps: S1 Constructing a hybrid simulator including a gray-box physical model, residual correction, and periodic state reset; S2 Establishing a six-dimensional weighted scoring function; S3 Configuring a safety verification mechanism; S4 Executing the large language model structure discovery phase, retrieving domain experience from pre-trained knowledge through structured experimental loops and transforming it into a control strategy framework; S5 Fine-tuning the continuous parameters in the strategy framework using Bayesian optimization; S6 Outputting the final controller. This invention is applicable to various facility agriculture scenarios such as poultry houses, greenhouses, plant factories, and aquaculture, achieving safe, efficient, and reproducible automatic optimization of environmental control strategies.

The invention provides an FDTD-based metasurface optimization method and device, and relates to the technical field of metasurface design regulation, and the method comprises the steps: dividing a transverse plane of a target metasurface, and obtaining a plurality of same grids; wherein each grid is provided with a corresponding continuous parameter to be optimized; on the basis of a predetermined target function, through a time domain finite difference FDTD method, obtaining a gradient of the target function to the to-be-optimized parameter; wherein the objective function is used for quantifying and optimizing the performance of the metasurface; and updating the to-be-optimized parameter based on the gradient of the to-be-optimized parameter so as to optimize the target metasurface based on the updated to-be-optimized parameter. Through the method provided by the invention, optimization of the metasurface is efficiently realized.

This invention relates to computer systems utilizing computational models, and discloses a computing device for a physics simulator and a computing system for physics simulation. A computing device for a physics simulator includes: computing resources; an agent for invoking a large language model to: receive initial simulation parameters and iteratively execute a simulation parameter generation process, including performing the following processes for each time step: receiving current simulation parameters and current simulation results from the physics simulator; generating next simulation parameters based on the current simulation parameters and historical interaction information; providing the next simulation parameters to the physics simulator; outputting all simulation parameters and corresponding simulation results; the agent calculating a performance score based on the simulation results and determining a reward function based on the performance score to train the large language model. The computing device according to the invention overcomes the problem of existing agents struggling to efficiently optimize continuous parameters under limited simulation budgets, and improves the ability of large language models to optimize simulation parameters.

The invention discloses a multi-input multi-output antenna decoupling performance prediction method based on deep learning, belongs to the technical field of computer-aided electromagnetic system design and prediction, is used for antenna decoupling performance prediction, and comprises the steps of determining key size parameters of a to-be-optimized antenna, performing parameter scanning on a change range of the key size parameters, and calculating the decoupling performance of the to-be-optimized antenna. Full-band S parameter simulation data is obtained, and an original data set is formed; expanding the data volume of the original data set by adopting an SMOTE algorithm, and generating an enhanced data set; and an antenna decoupling performance prediction model is constructed and trained, a to-be-optimized antenna is input, and a corresponding full-band S parameter curve final prediction result is output. According to the method, high-dimensional data are effectively expanded by introducing the SMOTE algorithm, the problem that small sample data in the field of electromagnetic simulation cannot support deep learning model training is solved, a DNN model is constructed to establish nonlinear mapping of a continuous S parameter curve, and the S parameter of the antenna can be comprehensively evaluated instead of only limiting discrete frequency points.

This invention provides a multi-design-constraint optimization method based on a dynamic surrogate model, comprising: constructing an engineering parameter optimization problem with a mixture of integer and continuous parameters; obtaining initial sample points; constructing a Kriging surrogate model for the objective function and Kriging surrogate models for each implicit constraint; during the Kriging surrogate model optimization process, using a batch constraint expectation criterion to search for sample points, and during the search process, searching for q sample points under the batch constraint expectation improvement criterion; utilizing parallel computing resources to calculate the response values ​​of the objective and implicit constraints for the q sample points, adding these new samples to the sample set, reconstructing the surrogate model for the objective and implicit constraints, and completing one update of the surrogate model; repeating the above steps of updating the Kriging surrogate model, and outputting the current optimal design variables and their response values. This invention can collaboratively handle explicit and implicit multiple design constraints and fully utilize parallel computing resources, which is beneficial to improving the efficiency and quality of optimization.

The invention provides an underwater robot interval type-2 fuzzy motion control method and system based on PSO optimization, and the method comprises the following steps: (1), building a simulation evaluation platform which comprises an ROV simulation model, a controller architecture and a basic fuzzy rule file; (3) configuring and executing an improved PSO algorithm: randomly initializing a position vector and a speed vector of a particle swarm in the boundary defined in the step (2); (4) fitness evaluation based on hybrid mapping: decomposing a 167-dimensional floating-point number position vector X into continuous parameters and parameters to be mapped, performing discrete mapping on the parameters to be mapped, converting the parameters to be mapped into integer vectors, updating membership function parameters and fuzzy rules in the basic fuzzy rule file through character string replacement operation, and generating a temporary configuration file; (5) iteration and convergence are carried out; according to the technical scheme, higher response speed, more accurate steady-state control and lower total accumulative error are achieved, and better control performance is achieved.

The invention relates to the technical field of parameter adaptive control, and further relates to a power grid process quantification-oriented cost parameter adaptive optimization method and system, and the method comprises the following steps: 1, obtaining a process unit set and a material price index sequence of a target power grid project, constructing a physical information neural network, and forming a cost parameter initial set; 2, on the basis of the cost parameter initial set, constructing a mixed integer programming model, setting constraint conditions of the mixed integer programming model, solving the mixed integer programming model by adopting a branch cutting method to obtain an optimal solution, and obtaining a cost parameter reference set; and step 3, constructing a hierarchical reinforcement learning agent, defining a state vector and a reward signal of the hierarchical reinforcement learning agent, and outputting a construction cost parameter after adaptive optimization. According to the method, deep fusion of data driving and cost field knowledge is realized, and the collaborative optimization problem of continuous parameter adjustment and discrete quota selection is solved.

A computer-implemented method includes obtaining a first kinematic graph that represents a first articulated object. The method includes generating a set of primitive structures by decomposing the first kinematic graph. The method includes generating a graph hierarchy by enumerating one or more valid combinations of the set of primitives. The method includes determining the production rules based on the graph hierarchy. The method includes determining kinematic graph grammar based on the set of primitive structures and the production rules. The method includes determining the continuous parameters for the discrete kinematic graph structure based on the production rules using a generative flow network. The method includes generating a second kinematic graph to represent a second articulated object based on the kinematic graph grammar and the continuous parameters.

The invention relates to the field of robot control and decision models, and discloses a robustness evaluation method of a visual language motion model in a changing scene. Aiming at the problems that an existing evaluation method is single in disturbance, cannot be physically realized and depends on white box information, the method converts object three-dimensional transformation, illumination variation and adversarial patches into a continuous parameter space, realizes worst scene search in a black box environment by adopting a CMA-ES algorithm, and constructs a simulation-evaluation integrated framework. And the model robustness system test is realized by unifying the task benchmark and the quantitative index. Gradient information is not needed, automatic execution can be achieved, and the result can be physically verified. The method is suitable for safety verification and robustness testing of various visual language action models, and provides technical support for reliable deployment of a robot system in the fields of manufacturing, medical treatment and service.