An aerial mechanism active learning reliability analysis method based on dual-view competitive fusion sampling

Abstract

The present application relates to a kind of aviation mechanism active learning reliability analysis method based on double view angle competition fusion sampling.The method includes three stages: initial sample distribution stage, through sensitivity analysis, parameter is divided into sensitive and non-sensitive parameter, sensitive parameter is encrypted sampling, non-sensitive parameter takes median value fixed;Double view angle adaptive sampling stage, combined with global prediction variance maximization criterion and local U learning function, through dynamic weight fusion realizes the intelligent selection of new sample point;Convergence determination stage, based on the variation coefficient of failure probability estimate value and the double index determination convergence of U function minimum value.The present application effectively balances the precision and efficiency, significantly reduces the calculation cost of complex engineering structure small failure probability analysis, and is suitable for the reliability evaluation scene of high-dimensional nonlinear engineering system.

Description

Technical Field

[0001] This invention relates to the field of reliability analysis and active learning agent model technology for aviation organizations, and in particular to an active learning reliability analysis method for aviation organizations based on dual-view competitive fusion sampling. Background Technology

[0002] Aircraft mechanisms are critical actuators in aircraft systems, and their reliability directly impacts flight safety and passenger boarding and disembarking. With the increasing demands for structural safety in the aviation industry, reliability analysis of aircraft mechanisms under uncertainty conditions has become a core aspect of airworthiness certification. Traditional reliability analysis methods mainly include the first-order second-moment method, the second-order second-moment method, and Monte Carlo simulation. While Monte Carlo simulation offers high accuracy, it requires simulation calculations with a large number of samples.

[0003] To address the aforementioned issues, reliability analysis methods based on surrogate models have emerged. However, existing methods suffer from the following shortcomings: (1) the initial sample placement does not consider parameter sensitivity, making it difficult for limited samples to effectively cover the sensitive parameter space; (2) the sampling strategy is singular, making it difficult to balance global exploration and local exploitation; and (3) the fixed-weight hybrid strategy cannot adaptively adjust the balance between exploration and exploitation.

[0004] Therefore, there is an urgent need for an active learning reliability analysis method that can efficiently construct high-precision surrogate models with limited computing resources, adaptably balance global exploration and local development, and robustly terminate sampling through dual criteria, based on the parameter distribution characteristics of aviation agencies. Summary of the Invention

[0005] The purpose of this invention is to propose an active learning reliability analysis method for aviation agencies based on dual-view competitive fusion sampling, so as to solve the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides the following solution: A proactive learning reliability analysis method for aviation agencies based on dual-view competitive fusion sampling includes: The global sensitivity analysis method is used to determine the global sensitivity index of the input parameters of the aviation agency to the output response. The input parameters of the aviation agency are divided into sensitive parameters and non-sensitive parameters according to the preset threshold, and a training sample set is constructed. A Kriging proxy model is constructed based on the training sample set, and an active learning process is carried out by adopting a dual-view adaptive sampling strategy to iteratively update the Kriging proxy model. When the Kriging agent model satisfies the convergence condition, it outputs an estimated failure probability value for the aerospace mechanism.

[0007] Optionally, the input parameters for the aerodynamic mechanism include: multiple bushing clearances and spline clearances.

[0008] Optionally, constructing the training sample set includes: For sensitive parameters, a modified Latin hypercube and Halton sequence mixed sampling method is used to generate sample points within their distribution space; For insensitive parameters, the median value of their distribution is fixed; The training sample set is formed by combining sample points extracted based on sensitive and insensitive parameters.

[0009] Optionally, sample points are generated using a modified Latin hypercube and Halton sequence mixed sampling method, including: Generate Halton sequence seed points to determine prime number base values; The seed points are perturbed and reorganized using the hierarchical constraints of the Latin hypercube. Iterative optimization is then performed to maximize the minimum distance between sample points in the sensitive parameter space. The objective function is: in, and For sample points in the normalized space, i and j These represent the indices of any two distinct sample points in the training sample set. This represents the 2-norm, referring to the Euclidean distance.

[0010] Optionally, a dual-view adaptive sampling strategy is adopted to conduct an active learning process, which iteratively updates the Kriging proxy model, including: Based on the Kriging surrogate model, the global perspective sampling score and the local perspective sampling score are calculated in parallel. The global perspective sampling score is determined based on the prediction variance maximization criterion, and the local perspective sampling score is determined based on the U learning function. A weight coefficient that changes dynamically with the number of iterations is introduced. By merging the sampling scores of the two perspectives with weights, the newly added sample points in this round are determined, added to the training sample set, and the Kriging surrogate model is updated.

[0011] Optionally, the criterion for maximizing the prediction variance is: in, For the Kriging agent model at point The predicted variance at that location This refers to a global perspective sampling scoring function used to quantify candidate sample points. The value of exploration throughout the design space.

[0012] Optionally, the U learning function is: in, For the limit state function, For the Kriging model at points The predicted standard deviation at that location This indicates how close the predicted response value is to the failure boundary. The smaller the value, the closer the point is to the failure boundary.

[0013] Optionally, the sampled scores from the two perspectives can be weighted and combined to include: in, The weighting coefficients change dynamically with the number of iterations. and These are the normalized global score and the local score, respectively. For comprehensive sampling and scoring; in, As the initial weights, To terminate the weight, To preset the maximum number of iterations, For iteration rounds.

[0014] Optionally, when the Kriging proxy model satisfies the convergence condition, it includes: After each update, the coefficient of variation and the minimum value of the U learning function are calculated based on the Monte Carlo method. When the coefficient of variation is less than the threshold 1 and the minimum value of the U learning function is greater than the threshold 2, the sampling is automatically terminated.

[0015] Optionally, the formula for calculating the coefficient of variation is: in, This is the current estimated failure probability. This represents the number of samples in Monte Carlo.

[0016] The beneficial effects of this invention are as follows: This invention drives the initial sample placement through global sensitivity analysis, performs mixed sampling only on sensitive parameters and takes the median value of non-sensitive parameters, and captures the main response trend to the greatest extent with a limited sample size; it overcomes the limitations of a single sampling criterion by using a dual-view competitive fusion mechanism to dynamically and adaptively balance global exploration and local development with weights; and it ensures both the statistical stability of failure probability estimation and the local fitting accuracy of failure boundary through dual convergence criteria. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a method for active learning reliability analysis of aviation agencies based on dual-view competitive fusion sampling, according to an embodiment of the present invention. Figure 2 This is a graph showing the Morris sensitivity analysis results of an embodiment of the present invention; Figure 3 The diagram shows the failure probability convergence curve and the minimum value convergence curve of the U function of this method in this embodiment of the invention. Figure 4 This is a comparison chart of the total number of samples and relative error between the method of this invention and other methods in this embodiment of the invention; Figure 5 This is a functional schematic diagram of the inner handle mechanism according to an embodiment of the present invention; Figure 6 This is a diagram showing the motion relationships between the components in an embodiment of the present invention; Figure 7 This is a diagram showing the transmission relationship between the main influencing factors and the failure modes of the mechanism in this embodiment of the invention. Figure 8 This is a diagram showing the final Kriging model fitting effect of an embodiment of the present invention; Figure 9 This is a distribution diagram of the maximum position offset in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1 As shown, this embodiment proposes an active learning reliability analysis method for aviation agencies based on dual-view competitive fusion sampling, including the following steps: Step S1: Initial Sample Placement Stage The global sensitivity analysis method is used to calculate the global sensitivity index of the input parameters of the aviation agency to the output response. The input parameters are divided into a set of sensitive parameters and a set of non-sensitive parameters according to a preset threshold. For the sensitive parameters, a mixed sampling method of Latin hypercube and Halton sequence is used to generate sample points in their distribution space. For the non-sensitive parameters, the median value of their distribution is fixed, and the combination forms the initial training sample set. Step S2: Dual-view adaptive sampling stage: Based on the Kriging agent model, global and local view sampling scores are computed in parallel. The global view sampling score is based on the prediction variance maximization criterion, and the local view sampling score is based on the U-learning function. Weight coefficients that dynamically change with the number of iterations are introduced. By weighted merging the sampling scores from two perspectives, the newly added sample points in this round are determined, added to the training set, and the proxy model is updated. The construction of the Kriging proxy model based on the training sample set includes: Using the training sample set as the input-output pair, a Kriging model combining a regression function and a Gaussian stochastic process is constructed. The Gaussian correlation function is used to describe the spatial correlation between sample points, and the correlation parameters are optimized through maximum likelihood estimation to determine the surrogate model. The Kriging model outputs the predicted mean and predicted variance.

[0022] Step S3: Convergence Determination Phase After each update, the coefficient of variation and the minimum value of the U function of the failure probability estimate are calculated based on the Monte Carlo method. When the coefficient of variation is less than the threshold 1 and the minimum value of the U function is greater than the threshold 2, the sampling is automatically terminated and the final failure probability estimate of the aviation agency is output.

[0023] Furthermore, the input parameters for the aerospace mechanism include, but are not limited to, multiple bushing clearances and spline clearances.

[0024] Furthermore, the improved method for mixed sampling of Latin hypercube and Halton sequence is as follows: First, Halton sequence seed points are generated, and prime bases are respectively taken as follows: Then, the seed points are perturbed and reorganized using the hierarchical constraints of the Latin hypercube. Through iterative optimization, the minimum distance of the sample points in the sensitive parameter space is maximized. The objective function is: in and These are sample points in the normalized space.

[0025] Generating Halton sequence seed points involves: determining the prime base for each dimension, independently calculating the Van der Corput sequence value for each dimension, and finally combining them.

[0026] The U-learning function in step S2, the dual-view adaptive sampling stage, is defined as: in Let be the limit state function. For the Kriging model at points The standard deviation of the prediction at a given point; the smaller the U-function value, the better. The closer to the failure boundary, the higher the prediction uncertainty at that point; the local perspective sampling scoring function is defined as... ,in To prevent small positive numbers from being divided by zero.

[0027] The global perspective sampling score uses the mean square error maximization function: in For the Kriging model at points The prediction variance at a given point is used to select the unsampled point with the greatest prediction uncertainty in the current proxy model as the global candidate point.

[0028] Dynamic weighting coefficients With each iteration round Changes according to a linear pattern: Among them, the initial weights (A weight of 0.7 is acceptable) This approach favors global exploration and terminates the weighting. (0.3 is acceptable) This approach favors localized development. (60 can be selected) is the preset maximum number of iterations; the comprehensive sampling scoring function is: in and These are the normalized global score and the local score, respectively; the final number of newly added sample points is: .

[0029] The formula for calculating the coefficient of variation is: in This is the current estimated failure probability. This represents the Monte Carlo sample size. If the coefficient of variation is too high, the Monte Carlo sample size needs to be increased further. The minimum value of the U function is... ;when and Iteration stops when both conditions are met.

[0030] The relevant functions of the Kriging surrogate model use the Matern 5 / 2 kernel function: in The Euclidean distance between the two points is... The relevant parameters are optimized through maximum likelihood estimation.

[0031] Specifically, the active learning reliability analysis method for aviation agencies based on dual-view competitive fusion sampling proposed in this embodiment includes the following three stages: Phase 1: Initial sample placement based on global sensitivity; Step 1.1: Define the parameters and limit state function; The input parameters for aerospace mechanisms include, but are not limited to, multiple bushing clearances and spline clearances. Bushing clearances and spline clearances are major sources of uncertainty; their deviations accumulate progressively in the kinematic chain. If the maximum positional offset exceeds the design allowable threshold, it will directly lead to component interference and locking failure, jeopardizing the normal opening and locking safety of the hatch, and consequently affecting the safety of the aircraft. Therefore, the limit state function regarding the positional offset is defined as follows: in, Each represents a different input variable. The maximum allowable position offset, This represents the actual maximum positional offset during the mechanism's movement, i.e., the offset threshold; when the positional offset... Less than the allowed offset threshold If the accuracy of the mechanism's motion position is reliable, then it is considered reliable; otherwise, it is considered unreliable.

[0032] Step 1.2: Parameter sensitivity analysis; Using the parameter sensitivity analysis module, the global sensitivity index of each input parameter to the output response is calculated by adopting the Sobol total effect index or the Morris basic effect method, and the input parameters are divided into a sensitive parameter set and a non-sensitive parameter set according to a preset threshold.

[0033] The formula for calculating the Sobol total effect index is as follows: in, For the first The total effect index of the input parameters Output the response for the model. Except for the All other input parameters besides the one parameter, For output The total variance measures the overall uncertainty of the output response. To calculate the expected value of the random variable within the parentheses using the mathematical expectation operator, To calculate the variance of the random variable within the parentheses using the variance operator, For conditional variance fixed, except for the first When all parameters other than the specified parameter are used, the output is... The conditional variance of The expected value of the conditional variance. Let V be the expected variance.

[0034] The screening criteria for the Morris basic effects method are: in, The total number of input parameters. For the number of trajectories, Output the response for the model. for The starting point of the trajectory, For the first The first point of the trajectory starting point The values ​​of the parameters, The step size factor has a value of ,in The number of discrete levels.

[0035] The Morris method is a cost-effective way to quickly identify a few sensitive parameters from a large number of parameters.

[0036] The Sobol method performs a fine-grained variance decomposition on a few key parameters after screening to obtain interaction information.

[0037] Step 1.3: Mixed sampling of sensitive parameters; For sensitive parameters within their normalized distribution space, an improved Latin hypercube and Halton sequence mixed sampling method is used to generate sample points, specifically including: First, a Halton low-dissimilarity sequence is generated as the initial seed point matrix. ,in This is the initial sample size. The number of sensitive parameters; The Halton points are mapped into the hierarchical structure of the Latin hypercube by independently performing sorting perturbations on each dimension. Through iterative optimization, the minimum Euclidean distance between sample points is maximized to ensure the uniformity of filling the low-dimensional sensitive parameter space. The objective function for the improved Latin hypercube and Halton sequence mixed sampling is: Indicates the first The sample point and the first The larger the Euclidean distance between sample points, the more uniform the distribution of sample points.

[0038] Step 1.4: Assign fixed values ​​to insensitive parameters; For each insensitive parameter, the median value is taken as its constant value across all samples, based on its probability distribution type: Uniform distribution :Pick ; normal distribution :Pick ; Log-normal distribution :Pick .

[0039] Step 1.5: Initial training set construction; The sensitive parameter sample points generated in step 1.3 are combined with the non-sensitive parameter values ​​fixed in step 1.4 one by one to form a complete input sample vector. The corresponding output response is calculated through the simulation model to construct the initial training dataset, which is used to establish the initial Kriging or Gaussian process proxy model. The initial total number of samples was controlled between 50 and 100.

[0040] Phase Two: Dual-view Adaptive Sampling Strategy; Step 2.1: Global view sampling; The Kriging surrogate model is a spatial interpolation method based on known sample points, first proposed by the South African geologist Krige. This model can unbiasedly fit complex nonlinear functions and obtain mean estimates and errors for predicted points, thus finding wide application in reliability analysis, structural design optimization, and other fields. The Kriging model can be expressed as follows: In the formula, It is a basis function vector. It corresponds to The regression coefficients, Corresponding to the number of basis functions, This represents a Gaussian random process with a mean of 0, used to realize local bias in the model.

[0041] The covariance function of any two points in a random process can be expressed as: In the formula, These are known sample points; Represents the correlation function between any two known sample points; The hyperparameter is a key factor affecting the accuracy of the surrogate model; the Kriging model's prediction of unknown points is... , This represents the predicted mean. The variance of the forecast represents the uncertainty of the forecast.

[0042] Based on the current Kriging surrogate model, the prediction mean square error of each candidate point in the entire parameter space is calculated, and the point with the largest prediction variance is selected as the global candidate sample point; the global sampling scoring function is defined as: in For the proxy model at point The prediction variance at that location.

[0043] Step 2.2: Local view sampling; Based on the limit state threshold passed from the reliability analysis module, the local sampling score of each candidate point is calculated using the improved expected improvement function or the U learning function. The definition of the expected improvement function is: in: and These are the cumulative distribution function (CDF) and probability density function (PDF) of the standard normal distribution, respectively.

[0044] The definition of the U learning function is: The smaller the U-function value, the more points... The closer to the failure boundary, the higher the prediction uncertainty at that point; the local perspective sampling scoring function is defined as... ,in To prevent small positive numbers from being divided by zero.

[0045] The sampling scoring function for a local viewpoint is defined as follows: or Step 2.3: Competition and Integration Mechanism; Introducing weight coefficients that dynamically change with the number of iterations By weighted merging the sampling scores from the two perspectives, the newly added sample points in the current cycle are determined; The dynamic weighting coefficient It changes according to the following pattern: Alternatively, an exponential decay form can be used: Among them, the initial weights (e.g., 0.7) biased towards global exploration, terminating the weight. (e.g., 0.3) biased towards failure boundary development, For the current iteration round, Set the maximum number of iterations (e.g., 30-50 times). The attenuation rate; The comprehensive sampling scoring function is defined as follows: The final number of newly added sample points is: in The candidate sample pool typically consists of a large number of sample points generated by Monte Carlo simulation (MCS) according to the joint probability distribution of the input parameters. This set covers the entire possible value space of the random variable.

[0046] Phase 3: Convergence determination of the active learning process; Step 3.1: Failure probability estimation; Based on the currently updated agent model, the Monte Carlo method is used to estimate the failure probability and extract... Given a set of sample points that follow a joint distribution of input parameters, the limiting state function value of each sample point is predicted using a surrogate model. The proportion of samples falling into the failure region is then used as an estimate of the failure probability. in, This is an indicator function.

[0047] Step 3.2: Calculation of coefficient of variation; Based on the failure probability estimate Compared with Monte Carlo sample size Calculate its coefficient of variation: when At that time, it was assumed that the failure probability estimate had sufficient statistical stability.

[0048] Step 3.3: Calculate the minimum value of the U function; For each sample point in the Monte Carlo sample pool, calculate the U function and take its minimum value. .

[0049] when At that time, it was believed that the surrogate model had sufficient accuracy in identifying failure boundaries.

[0050] Step 3.4: Termination of decision under dual conditions; If and only if and When both conditions are met, the active learning sampling process is automatically terminated, and the current failure probability estimate is output as the final result. If any condition is not met, return to stage two to continue iterative sampling.

[0051] The following embodiment uses an aircraft cabin door handle mechanism as an example to further describe the method of this embodiment in detail.

[0052] The following five-dimensional nonlinear limit state function is used as a test case; I. Problem Definition: Consider the following five-dimensional nonlinear limit state function: Input parameters All follow a standard normal distribution The range of values ​​is .

[0053] Failure region defined as .

[0054] Reference failure probability: obtained through large-scale Monte Carlo simulation ( The calculation yielded: .

[0055] II. Parameter Configuration; The parameter configuration table for this embodiment is shown in Table 1.

[0056] Table 1. Parameter Configuration Table for Mathematical Verification Cases Phase 1: Initial sample placement based on global sensitivity: 3.1: Morris method for quickly filtering insensitive parameters: Using the standard Morris design, each parameter is discretized. Level, step size .generate Each trajectory contains 1 independent trajectory. 1 sample point. For each trajectory, parameters The basic effect is: in For the first A unit vector in each coordinate direction.

[0057] Calculate the average absolute basic effect of each parameter .

[0058] The calculation results are as follows Figure 2 As shown in Table 2.

[0059] Table 2 Morris Sensitivity Analysis Results According to the threshold , The parameter was determined to be non-sensitive, while the rest were sensitive parameters.

[0060] Insensitive parameters. Take the median value (i.e., the median of the standard normal distribution, 0), and keep it fixed in all subsequent analyses.

[0061] 3.2 Spatial Mixed Sampling of Sensitive Parameters; The set of sensitive parameters is Dimension An improved method using mixed sampling of Latin hypercube and Halton sequences was employed to generate... 1 initial sample point.

[0062] 1. Halton sequence generation: using prime base This generates an 80×4 normalized Halton matrix; 2. Latin hypercube perturbation: Sort each dimension independently and add random jitter so that each stratified interval contains exactly one sample; 3. Maximum-Minimum Distance Optimization: Maximize the minimum Euclidean distance by swapping sample points through 100 iterations. 4. Inverse Transformation Mapping: Mapping via inverse transformation of the standard normal distribution. Map the normalized samples to the original parameter space; 5. Combined Samples: Combining insensitive parameters The value is fixed at 0, and combined with the sensitive parameter samples to form 80 complete 5-dimensional input vectors.

[0063] Calculate the limit state function values ​​for each sample point using the real model (or the analytical function here). This yields the initial training set. The initial set contains 8 failed samples, representing an initial failure rate of 10%.

[0064] 3.3 Initial Kriging Model Construction; A kriging model was constructed based on 80 initial sample points, using the Gaussian correlation function: in For the first The relevant parameters of dimension are optimized using maximum likelihood estimation (MLE). The negative log-likelihood function is: Solving for the optimal correlation parameters The initial Kriging model had a mean relative error of 4.2% on the training set.

[0065] Phase 2: Dual-view adaptive sampling strategy; 4.1 Dynamic weight settings; Dynamic weighting coefficients with linear decay are used: Initial round Focuses on overall exploration, later stages It tends to favor localized development.

[0066] 4.2 Global View Sampling; In the candidate sample pool ( Calculate the prediction variance of the current Kriging model for each candidate point in a Monte Carlo sample. and normalize: The region with the largest prediction variance was selected for exploration.

[0067] 4.3 Local view sampling; Calculate the U-learning function for each candidate point: Define local sampling score: in To prevent division by zero for small positive numbers, points with smaller U values ​​(i.e., predicted mean close to the threshold and with high uncertainty) receive higher scores.

[0068] 4.4 Competition and Integration Mechanism; The overall sampling score is: choose The largest candidate point is used as a new sample point in this round. Its output is calculated by calling the real model, added to the training set, and the Kriging model is retrained.

[0069] Phase 3: Convergence determination of the active learning process; After each iteration, the Monte Carlo validation sample pool is predicted using the current Kriging model. (At each point, the insensitive parameter is fixed at 0), calculate the estimated failure probability. Coefficient of variation (Cov) and minimum value of U function : If and only if and The sampling process will automatically terminate when the time is reached. The complete iterative process is shown in Table 3.

[0070] Table 3 Iteration process It finally converged at the 48th position, with a total sample size of [number missing]. Failure probability estimate , and reference value The relative error is 2.69%.

[0071] VI. Convergence Process Analysis; (a) Convergence of the minimum value of the U-function: Initial stage This indicates that the surrogate model exhibits significant uncertainty near the failure boundary. As iterations proceed, new sample points continuously fill the boundary region. It gradually increased, and for the first time exceeded 2.0 at the 48th.

[0072] (b) Convergence of the coefficient of variation: Since the failure probability is approximately ,use With a Monte Carlo sample size, the Cov remains stable at around 3.30%, less than the 5% threshold, consistently meeting the statistical stability requirements. The failure probability convergence curve and the minimum U-function convergence curve of the method in this embodiment are shown below. Figure 3 As shown.

[0073] VII. Comparison and verification with existing methods, such as... Figure 4 As shown in Table 4; The superiority of this invention is verified by comparing it with existing active learning reliability analysis methods.

[0074] Comparison method: Method A: Standard AK-MCS (U learning function only); Method B: EGO algorithm (maximizing only the variance of predictions); Method C: Fixed-weight mixture (α=0.5 fixed); Table 4 Performance Comparison of Various Methods Comparative analysis: Convergence speed: This invention converges at the 48th round. AK-MCS requires 17 rounds, but the relative error calculated by AK-MCS is 3.23%. EGO requires 60 rounds, and fixed weight hybrid requires 44 rounds, but the relative error of the fixed weight hybrid method is 3.66%.

[0075] Prediction accuracy: The relative error of this invention is 2.69%, which is lower than that of other methods.

[0076] Taking all factors into consideration, although the total number of samples and convergence rounds required by AK-MCS are lower than those of this method, the resulting error is larger. EGO and fixed-weight hybrid methods require a smaller total number of samples, fewer convergence rounds, and have lower subsequent relative errors than the present invention. Therefore, the method in this embodiment has certain advantages.

[0077] Reliability analysis of the motion of the handle mechanism inside the boarding door of a certain type of aircraft: I. Problem Description and Parameter Settings; This example analyzes the internal handle mechanism of a certain type of aircraft boarding gate. The functional principle of the internal handle mechanism is as follows: Figure 5 As shown. During the boarding gate opening process, when the inner handle lever rotates upward from its initial position to a certain angle, the assist spring assembly and the inner handle lever are located at opposite ends of the shaft. Their torques acting on the shaft are in opposite directions, so the assist spring assembly will hinder the normal rotation of the inner handle lever. When the inner handle lever continues to move upward from this angle, the assist spring assembly moves to the right end of the shaft through the flight lock rocker arm. Their torques acting on the shaft are in the same direction, thus assisting the rotation of the inner handle lever.

[0078] Based on the functional principle of the internal handle mechanism, the motion relationships of each component of the internal handle mechanism were established, such as... Figure 6 As shown, a dynamic model of the inner handle mechanism was established based on the motion relationships of each component. Analysis reveals that the bushing clearance and spline clearance of the mechanism have varying degrees of influence on the motion accuracy of the inner handle mechanism. Figure 7 As shown, therefore, a function for the offset is established: In the formula, This is a function related to the position offset; 0.2 is the maximum allowed position offset. This represents the actual maximum positional offset during the mechanism's movement, i.e., the offset threshold; when the positional offset... If the offset is less than the allowable threshold of 0.2, the positional accuracy of the mechanism is considered reliable; otherwise, it is considered unreliable.

[0079] Furthermore, according to Figure 7 Determine the input parameters that affect the output response. There are a total of 5, and their distribution and statistical characteristics are shown in Table 5.

[0080] Table 5 Probability distribution of input parameters for the internal handle mechanism II. Initial sample placement based on Morris global sensitivity-driven approach; The Morris fundamental effects method was used to perform global sensitivity screening on five input parameters (bulb gaps 1-3 and spline gaps 1-2). The sensitivity analysis results are shown in Table 6, after which the initial training set was constructed.

[0081] Table 6 Morris Sensitivity Analysis Results Among them, the set of sensitive parameters is: Insensitive parameter set: .

[0082] III. The process of establishing the Kriging proxy model; Based on the initial training set, a Kriging surrogate model is established, and an active learning process is carried out using a dual-view adaptive sampling strategy. After each iteration, the current Kriging model is used to predict the Monte Carlo validation sample pool, and the failure probability estimate is calculated. Coefficient of variation (Cov) and minimum value of U function If and only if and When the iteration terminates, the final Kriging model fitting result is obtained as follows: Figure 8 As shown, the goodness of fit is 0.997, which meets the requirements.

[0083] Output the final failure probability estimate. And the distribution of the maximum position offset is as follows Figure 9 As shown.

[0084] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for active learning reliability analysis of aviation agencies based on dual-view competitive fusion sampling, characterized in that, include: The global sensitivity analysis method is used to determine the global sensitivity index of the input parameters of the aviation agency to the output response. The input parameters of the aviation agency are divided into sensitive parameters and non-sensitive parameters according to the preset threshold, and a training sample set is constructed. A Kriging proxy model is constructed based on the training sample set, and an active learning process is carried out by adopting a dual-view adaptive sampling strategy to iteratively update the Kriging proxy model. When the Kriging agent model satisfies the convergence condition, it outputs an estimated failure probability value for the aerospace mechanism.

2. The active learning reliability analysis method for aviation agencies based on dual-view competitive fusion sampling according to claim 1, characterized in that, The input parameters for the aviation mechanism include: multiple bushing clearances and spline clearances.

3. The active learning reliability analysis method for aviation agencies based on dual-view competitive fusion sampling according to claim 1, characterized in that, Constructing the training sample set includes: For sensitive parameters, a modified Latin hypercube and Halton sequence mixed sampling method is used to generate sample points within their distribution space; For insensitive parameters, the median value of their distribution is fixed; The training sample set is formed by combining sample points extracted based on sensitive and insensitive parameters.

4. The active learning reliability analysis method for aviation agencies based on dual-view competitive fusion sampling according to claim 3, characterized in that, The improved Latin hypercube and Halton sequence mixed sampling method is used to generate sample points, including: Generate Halton sequence seed points to determine prime number base values; The seed points are perturbed and reorganized using the hierarchical constraints of the Latin hypercube. Iterative optimization is then performed to maximize the minimum distance between sample points in the sensitive parameter space. The objective function is: in, and For sample points in the normalized space, i and j These represent the indices of any two distinct sample points in the training sample set. This represents the 2-norm, referring to the Euclidean distance.

5. The active learning reliability analysis method for aviation agencies based on dual-view competitive fusion sampling according to claim 1, characterized in that, The active learning process employs a dual-view adaptive sampling strategy to iteratively update the Kriging proxy model, including: Based on the Kriging surrogate model, the global perspective sampling score and the local perspective sampling score are calculated in parallel. The global perspective sampling score is determined based on the prediction variance maximization criterion, and the local perspective sampling score is determined based on the U learning function. A weight coefficient that changes dynamically with the number of iterations is introduced. By merging the sampling scores of the two perspectives with weights, the newly added sample points in this round are determined, added to the training sample set, and the Kriging surrogate model is updated.

6. The active learning reliability analysis method for aviation agencies based on dual-view competitive fusion sampling according to claim 5, characterized in that, The criterion for maximizing the prediction variance is: in, For the Kriging surrogate model at candidate sample points The predicted variance at that location This refers to a global perspective sampling scoring function used to quantify candidate sample points. The value of exploration throughout the design space.

7. The active learning reliability analysis method for aviation agencies based on dual-view competitive fusion sampling according to claim 5, characterized in that, The learning function U is: in, For the limit state function, For the Kriging model at points The predicted standard deviation at that location This indicates how close the predicted response value is to the failure boundary. The smaller the value, the closer the point is to the failure boundary.

8. The active learning reliability analysis method for aviation agencies based on dual-view competitive fusion sampling according to claim 5, characterized in that, The weighted merging of the sampled scores from the two perspectives includes: in, The weighting coefficients change dynamically with the number of iterations. and These are the normalized global score and the local score, respectively. For comprehensive sampling and scoring; in, As the initial weights, To terminate the weight, To preset the maximum number of iterations, For iteration rounds.

9. The active learning reliability analysis method for aviation agencies based on dual-view competitive fusion sampling according to claim 1, characterized in that, When the Kriging proxy model satisfies the convergence condition, it includes: After each update, the coefficient of variation and the minimum value of the U learning function are calculated based on the Monte Carlo method. When the coefficient of variation is less than the threshold 1 and the minimum value of the U learning function is greater than the threshold 2, the sampling is automatically terminated.

10. The active learning reliability analysis method for aviation agencies based on dual-view competitive fusion sampling according to claim 9, characterized in that, The formula for calculating the coefficient of variation is: in, This is the current estimated failure probability. This represents the number of samples in Monte Carlo.

Images