CN-121980218-A - Sequential test design method based on BDNN and critical area density weighting

CN121980218ACN 121980218 ACN121980218 ACN 121980218ACN-121980218-A

Abstract

The invention relates to a sequential test design method based on BDNN and critical area density weighting, and relates to the technical field of radars. The method comprises the steps of carrying out general feature extraction on an initial data set through a constructed dynamic self-adaptive BDNN model, outputting a continuous prediction result and a classification prediction result, adopting a multi-objective chaotic particle swarm optimization algorithm to identify a key region in a multi-dimensional factor space, carrying out iterative optimization to obtain an optimal solution set, calculating comprehensive scores, screening, outputting a high-priority key region, carrying out density weighted sequential sampling based on the high-priority key region to generate multi-source candidate points, calculating comprehensive weights of all the candidate points, screening out a final sampling point through maximizing the sum of the comprehensive weights and the space distance, verifying the final sampling point, fusing the verified sample with the initial data set, and retraining the dynamic self-adaptive BDNN model until preset convergence conditions are met. The invention can improve the performance of the detection system.

Inventors

BAO LEI
LI CHAOLONG
GAO XIANZHONG
LI DONGFANG
HE TING

Assignees

中国人民解放军国防科技大学

Dates

Publication Date: 20260505
Application Date: 20260116

Claims (10)

1. A sequential trial design method based on BDNN and critical area density weighting, the method comprising: Step 1, an initial dataset is obtained, general feature extraction is carried out on the initial dataset through a constructed dynamic self-adaptive BDNN model, and a continuous prediction result and a classified prediction result are output; Step 2, based on the continuous prediction result and the classified prediction result, a multi-objective chaotic particle swarm optimization algorithm is adopted to identify a key region in a multi-dimensional factor space, the key region is defined through multi-dimensional characteristics, a multi-response characterization index system is constructed, after an optimal solution set is obtained through iterative optimization, the comprehensive score of the optimal solution set is calculated, the key region is screened according to the comprehensive score, and a high-priority key region is output; step 3, based on the high-priority key region, adopting an improved MP-CE algorithm to implement density weighted sequential sampling, generating multi-source candidate points, calculating the comprehensive weight of each candidate point, and screening out a final sampling point by maximizing the sum of the comprehensive weight and the space distance; and 4, verifying the final sampling point, fusing the verified sample with the initial data set, updating and retraining the dynamic self-adaptive BDNN model, and repeatedly executing the step 2 and the step 3 based on the updated dynamic self-adaptive BDNN model until a preset convergence condition is met.
2. The sequential trial design method of claim 1, wherein in step 1, the dynamic adaptive BDNN model comprises an input layer, a shared feature layer, and dual output branches; the shared feature layer comprises a first shared layer and a second shared layer which are connected in series, and the first shared layer and the second shared layer have the same structure and are used for extracting the general features of the initial data set; the dual-output branch comprises a continuous response branch and a classified response branch, and is used for respectively processing the general characteristics and correspondingly outputting a continuous prediction result and a classified prediction result.
3. The sequential trial design method of claim 2, wherein in step 1, a dynamic adaptive training strategy is adopted in the dynamic adaptive BDNN model, and the dynamic adaptive training strategy comprises: presetting a first sample size threshold and a second sample size threshold, wherein the first sample size threshold is smaller than the second sample size threshold; correspondingly adjusting the network complexity parameter, regularization strength parameter and training parameter of the dynamic self-adaptive BDNN model according to the threshold interval where the current training sample size is located, wherein when the sample size is smaller than a first sample size threshold, a low-complexity network structure, a high-strength regularization parameter and a strong weight attenuation parameter are adopted; When the sample size is larger than or equal to the first sample size threshold and smaller than the second sample size threshold, adopting a medium-complexity network structure, a medium-strength regularization parameter and a medium-weight attenuation parameter; When the sample size is larger than or equal to a second sample size threshold, adopting a high-complexity network structure, a low-strength regularization parameter and a low-weight attenuation parameter; And carrying out general feature extraction on the initial dataset through the dynamic self-adaptive BDNN model trained by the dynamic self-adaptive training strategy, and outputting a continuous prediction result and a classified prediction result.
4. The sequential trial design method of claim 3, further comprising employing an uncertainty quantization mechanism in the dynamic adaptive BDNN model, the uncertainty quantization mechanism comprising: Performing multiple Monte Carlo dropouout sampling on the continuous response branch to obtain multiple groups of outputs of the continuous response branch; Based on a plurality of groups of outputs, obtaining an optimal estimated value of a continuous prediction result through prediction mean value calculation, and taking the optimal estimated value of the continuous prediction result and the classified prediction result together as the input of the step 2; And calculating the confidence coefficient of the dynamic self-adaptive BDNN model prediction through prediction variance based on the optimal estimation values of a plurality of groups of output and the continuous prediction result, and calculating the uncertainty weight of each candidate point based on the confidence coefficient.
5. The sequential test design method based on BDNN and critical area density weighting according to claim 4, wherein the optimal estimated value of the continuous prediction result is obtained by calculating the prediction mean value, and the expression is: ; In the formula, An optimal estimated value representing a continuous prediction result; Representing configuration in network topology Outputting a random variable by prediction of a sample to be predicted; Representing input samples to be predicted in an initial dataset; Representing a data set for training; representing the sampling times; represent the first When the Monte Carlo sampling is performed, weight parameter based And network topology configuration A network single forward propagation prediction value of (a); Representing the network topology configuration after the Dropout mechanism is applied; represent the first The subsampled dynamic adaptation BDNN model parameters.
6. The sequential trial design method based on BDNN and critical area density weighting of claim 4, wherein the method is based on And outputting the optimal estimated value of the continuous prediction result by a group, and calculating the confidence coefficient of model prediction by a prediction variance, wherein the expression is as follows: ; In the formula, Representing the confidence level of the dynamic adaptive BDNN model prediction; an optimal estimated value representing a continuous prediction result; Representing configuration in network topology Outputting a random variable by prediction of a sample to be predicted; Representing input samples to be predicted in an initial dataset; Representing a data set for training; representing the sampling times; represent the first When the Monte Carlo sampling is performed, weight parameter based And network topology configuration A network single forward propagation prediction value of (a); Representing the network topology configuration after the Dropout mechanism is applied; represent the first The subsampled dynamic adaptation BDNN model parameters.
7. The sequential trial design method of claim 1-6, wherein in step 2, the critical areas are defined by multi-dimensional characteristics expressed as: ; In the formula, Representing a key region; representing any one of the input sample vectors in the multi-dimensional factor space; representing a factor space; Representing local performance extremum characteristics; representing gradient sensitivity characteristics; representing global distribution anomaly characteristics; Representing extreme value dense characteristics; 、、、 Representing the characteristic threshold.
8. The sequential trial design method based on BDNN and critical area density weighting of any of claims 1-6, wherein in step 2, the multi-response characterization index system includes an information layer extremum mean, a maximum gradient change metric, a local to global distribution divergence, a local to global extremum ratio; Quantizing the multidimensional characteristics through the information layer extremum average value, the maximum gradient change measurement, the local and global distribution divergence and the local and global extremum ratio to obtain quantized characteristic parameters; according to the quantized characteristic parameters, performing iterative optimization through a multi-objective chaotic particle swarm optimization algorithm to obtain an optimal solution set; and calculating comprehensive scores of the optimal solution set by adopting an entropy weight method based on the quantized characteristic parameters, screening the key areas according to the comprehensive scores, and outputting high-priority key areas.
9. The method of claim 1 to 6, wherein in step 3, generating multiple source candidate points and calculating the comprehensive weight of each candidate point, and screening out the final sampling point by maximizing the sum of the comprehensive weight and the spatial distance comprises: Generating a first candidate point, a second candidate point and a third candidate point in the high-priority key region, the boundary preset range of the high-priority key region and the global factor space according to the candidate point occupation ratio; and after combining the first candidate point, the second candidate point and the third candidate point into multi-source candidate points, selecting a final sampling point by calculating comprehensive weights and maximizing the sum of the comprehensive weights and the space distance.
10. The sequential trial design method based on BDNN and critical area density weighting of any of claims 1-6, wherein in step 3, the calculation formula of the comprehensive weight is: ; In the formula, Representing the comprehensive weight; Representing the generated multisource candidate points to be evaluated; Representing the key region weights; representing the spatial density weights; Representing an uncertainty weight; Representing diversity weights; 、、、 Representing the weight coefficient.

Description

Sequential test design method based on BDNN and critical area density weighting Technical Field The invention relates to the technical field of radars, in particular to a sequential test design method based on BDNN and critical area density weighting. Background In a modern complex electromagnetic environment, performance of a detection system directly influences situation awareness and performance of related systems, and a large number of high-fidelity performance tests are required to be carried out in order to evaluate and improve robustness of the detection system in a complex electromagnetic interference scene. However, the conventional test design method faces many core challenges such as difficult modeling of the mixed response, one-sided identification of the key area, inefficient allocation of test resources, and the like. In the aspect of mixed response modeling, performance test response variables of the detection system usually show mixed multi-mode characteristics, and the performance test response variables comprise continuous variables such as angle tracking precision (ATA), distance tracking precision (RTA) and the like, and classified variables such as working state judgment and the like. The early method adopts a strategy of independent modeling to neglect potential association between different responses, which is easy to cause information loss and model deviation, while some joint modeling methods still do not fully capture the internal coupling relation between responses, or the original classification information is easy to be distorted because the linear mapping assumption is difficult to adapt to complex nonlinear scenes. In terms of critical area identification, critical areas of a test sample space are generally defined as areas with high response values, high uncertainties, or high rates of change, which definition makes it difficult to fully characterize the diversity and heterogeneity of a complex test space. The existing methods are characterized in that the minimum energy design only takes the height of a response value as a unique judgment standard, and defines one-sided or focuses only two types of characteristics, key dimensions such as a high response value are not considered, or the distribution characteristics of the pareto optimal solution are excessively depended, the local gradient change, the distribution difference and other inherent characteristics of the area are not fully fused, and the methods lack system quantization indexes adapting to test sample space and are difficult to directly migrate and apply to complex test design tasks. In terms of test resource allocation, the traditional space filling design method such as Latin hypercube design and maximum projection design can ensure global uniformity, but the problem of insufficient sample representativeness is easy to occur when mixed factors in the test of the detection system are processed, while some improvement methods solve the problem of coexistence of qualitative and quantitative factors to a certain extent, but the problem is still insufficient when the complicated irregular constraint space is faced, and the method does not introduce factor importance weight mechanism, so that the critical subspace filling density of the critical factors is insufficient, the sample information gain is reduced, and the methods lack the targeted exploration of the 'critical area', so that a large amount of test resources are consumed in a non-critical area with low value, the model precision is slowly improved, and the test efficiency is low. Disclosure of Invention Based on this, it is necessary to provide a sequential test design method based on BDNN and critical area density weighting, which can realize modeling, identification, sampling, iterative closed-loop optimization and improve the performance of the detection system, aiming at the technical problems. A sequential trial design method based on BDNN and critical area density weighting, the method comprising: Step 1, an initial dataset is obtained, general feature extraction is carried out on the initial dataset through a constructed dynamic self-adaptive BDNN model, and a continuous prediction result and a classified prediction result are output; Step 2, based on the continuous prediction result and the classified prediction result, a multi-objective chaotic particle swarm optimization algorithm is adopted to identify a key region in a multi-dimensional factor space, the key region is defined through multi-dimensional characteristics, a multi-response characterization index system is constructed, after an optimal solution set is obtained through iterative optimization, the comprehensive score of the optimal solution set is calculated, the key region is screened according to the comprehensive score, and a high-priority key region is output; step 3, based on the high-priority key region, adopting an improved MP-CE algorithm to implement density weighted sequential