Air target multi-source data completion method and device based on depth generation domain adaptation

Abstract

The invention discloses an air target multi-source data complementation method and device based on depth generation domain adaptation, and the method comprises the steps: obtaining the multi-source observation data of different sensors for an air target, taking the multi-source observation data as the input of a generator in a generative adversarial interpolation network, and obtaining the complemented multi-source observation data after interpolation; wherein when a generator in the generative adversarial interpolation network is trained, generator domain adversarial loss and information entropy loss are increased; the generator domain adversarial loss is calculated based on a domain label of a multi-source observation value; according to the method, a domain adaptation mechanism is introduced into a generator and discriminator structure, so that the problem of distribution offset caused by multi-attribute data missing can be effectively relieved.

Description

Technical Field

[0001] This invention belongs to the field of aerial target data mining technology, and particularly relates to a method and apparatus for multi-source data completion of aerial targets based on deep generation domain adaptation. Background Technology

[0002] In recent years, the rapid development of information technology and aircraft technology has intensified competition among countries in fields such as artificial intelligence, drones, and hypersonic vehicles. The threat characteristics of aerial targets are becoming increasingly complex, placing higher demands on the real-time monitoring, identification, and response capabilities of traditional air defense systems. At the same time, the rapid rise of the low-altitude economy further highlights the strategic significance of aerial target identification technology in airspace management.

[0003] Existing aerial target identification technologies typically rely on multi-source sensing data, including optical, infrared, radar, and acoustic information sources. Due to the small size, high speed, and susceptibility of aerial targets to weather and environmental factors, relying solely on a single sensor source often fails to achieve stable and reliable identification results. In contrast, multi-source data fusion can integrate the advantages of different sensor sources, characterizing the target from a multi-dimensional attribute space, thus improving the accuracy and robustness of identification and becoming one of the main research directions.

[0004] However, existing multi-source aerial target identification methods still have significant shortcomings in practical applications. In complex environments or under dynamic monitoring conditions, some sensors are susceptible to factors such as obstruction, signal attenuation, or communication interruptions, leading to missing or incomplete attributes in the observation data. Such omissions not only disrupt the overall structure and correlation of the multi-source data but also cause changes in its distribution. Over time, this results in inconsistencies between the distribution of the collected training set and the test set data to be detected. This dynamic change often leads to significant distribution differences between the test set and the training set, deviating from the traditional assumption of independent and identically distributed data, further exacerbating the distribution inconsistency. As the proportion of missing data increases, this distribution shift effect becomes more pronounced, severely impacting the reliability and generalization ability of the identification system.

[0005] Since current multi-source data interpolation strategies do not fully consider the fundamental problem of distribution bias, these methods still cannot effectively improve the distribution offset problem. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for multi-source data completion of aerial targets based on deep generative domain adaptation. By constructing a shared feature space, the training and test sets of data with different attributes are mapped to a unified space for relevant alignment and consistency constraints, thereby effectively alleviating the distribution offset problem caused by missing data.

[0007] This invention adopts the following technical solution: a method for multi-source data completion of aerial targets based on deep generation domain adaptation, comprising the following steps:

[0008] Acquire multi-source observation data of aerial targets from different sensors, and use the multi-source observation data as input to the generator in the generative adversarial interpolation network to obtain the interpolated and complete multi-source observation data.

[0009] The generator contains a cascaded linear transformation layer, a first self-attention mechanism module, and a second self-attention mechanism module. The first self-attention module adopts a residual structure design, and its output features and input features are added element-wise through skip connections to form a residual mapping result. The residual mapping result is then batch normalized and nonlinearly activated as the input features of the second self-attention mechanism module.

[0010] Specifically, when training the generator in the generative adversarial interpolation network, generator domain adversarial loss and information entropy loss are added; the generator domain adversarial loss is calculated based on the domain labels of multi-source observations.

[0011] The beneficial effects of this invention are as follows: By introducing a domain adaptation mechanism into the generator and discriminator structures, this invention achieves the distribution alignment of the training set and the test set, and guides the interpolation process through category priors to maintain the feature consistency of aerial targets of the same category. This effectively alleviates the distribution shift problem caused by missing multi-attribute data, improves the robustness and generalization performance of the aerial target recognition model in complex environments, and provides a reliable data processing and classification solution for accurate aerial target recognition. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the method for multi-source data completion of aerial targets based on deep generation domain adaptation in an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of the t-SNE visualization results after using different imputation methods at different missing rates in the verification embodiment of the present invention. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0015] Existing methods generally assume that the interpolated training and test sets are independent and identically distributed. However, in practical applications, due to differences in data acquisition environment, time, or equipment, the collected multi-source, multi-attribute aerial target training data and test data often have significant distributional shifts. This deviation from the assumption leads to a decrease in the generalization performance of the model during the testing phase.

[0016] Secondly, traditional imputation methods typically focus only on the accuracy of numerical reconstruction when generating missing values, ignoring prior class information. This makes it difficult to maintain feature consistency between samples of different classes, thus weakening the classifier's discriminative ability. When the proportion of missing data from multiple sources is high, the performance of existing methods will significantly degrade, and the generated results are prone to distortion or noise accumulation, making it difficult to recover target features and consequently affecting the recognition accuracy and system stability in complex environments.

[0017] To address the above problems, this invention proposes a method for multi-source data completion of aerial targets based on deep generation domain adaptation, namely, a method for missing data interpolation and identification. The technical problems solved include:

[0018] 1. In the process of imputing missing data from multiple sources and attributes, the distribution of the training set and the test set are aligned, which alleviates the inconsistency in distribution caused by missing data, thereby improving the generalization ability and robustness of the model.

[0019] 2. Effectively incorporate prior knowledge of the categories of multi-source aerial target data during the interpolation process to enhance the consistent representation of similar aerial target samples in the feature space and improve the class discriminativeness of the interpolation results.

[0020] 3. It can still generate high-quality, semantically consistent imputation samples under high missing rate conditions, avoiding noise amplification and feature distortion, and ensuring the recognition performance and stability of the system in complex environments.

[0021] This invention addresses the imputation and identification of incomplete data on multi-source, multi-attribute aerial targets (observation data for different attributes comes from different sensors). To resolve issues with existing data imputation methods, such as training and test set distribution misalignment, severe feature loss, and insufficient class discriminative power, this invention proposes a deep generative domain adaptation-based imputation method. By introducing a domain adaptation mechanism into the generator and discriminator structures, the distribution of the training and test sets is aligned. Furthermore, class priors guide the imputation process to maintain feature consistency for aerial targets of the same class. This invention effectively mitigates the distribution misalignment problem caused by missing multi-attribute data, improves the robustness and generalization performance of aerial target identification models in complex environments, and provides a reliable data processing and classification solution for accurate aerial target identification.

[0022] like Figure 1 As shown, the depth-generated domain adaptive interpolation method consists of two parts: a prior knowledge acquisition module for multi-source, multi-attribute offset correction of aerial targets (…). Figure 1 (below) and the generation domain adaptive interpolation module ( Figure 1 (Above).

[0023] First, even if the missing rate of multi-source, multi-attribute aerial target data is fixed, each data point will have a different missing rate. Data with lower missing rates, or even no missing rates, may contain more reliable latent category information. This invention obtains class prior knowledge by filtering this portion of data, thereby implementing class attribute constraints during the imputation process.

[0024] Secondly, when the distribution of multi-source, multi-attribute training and test sets is inconsistent due to missing data, they can be regarded as the source and target domains for domain adaptation, respectively. Therefore, this invention proposes a generative domain adaptation method for data imputation, minimizing the overall distribution difference between the imputed training and test sets.

[0025] The core of this invention is to use GAN-generated data to process missing values ​​and to use the minimization of distribution differences as a constraint in the final optimization function for imputation. The purpose is to improve the quality of missing value imputation by reducing the distribution differences of the imputed values.

[0026] This invention discloses a method for multi-source data completion of aerial targets based on deep generative domain adaptation, comprising the following steps: acquiring multi-source observation data of aerial targets from different sensors, and using the multi-source observation data as input to the generator in a generative adversarial interpolation network to obtain the interpolated completed multi-source observation data; the generator includes a cascaded linear transformation layer, a first self-attention mechanism module, and a second self-attention mechanism module. The first self-attention module adopts a residual structure design, and its output features and input features are added element-wise through skip connections to form a residual mapping result. The residual mapping result is batch normalized and nonlinearly activated and then used as the input features of the second self-attention mechanism module; wherein, when training the generator in the generative adversarial interpolation network, generator domain adversarial loss and information entropy loss are added; the generator domain adversarial loss is calculated based on the domain labels of the multi-source observations.

[0027] Suppose X = {x1, x2, ..., x} N} represents an incomplete multi-source, multi-attribute dataset of aerial targets, containing N multi-source observation data (also called samples or training samples), sample x i It contains d attributes, where d represents the number of attributes in the i-th training sample. It can be defined as x. i ={x i,1 ,x i,2 ,...,x i,d}, each sample x i Each has a binary mask vector M(x) i )∈{0,1} d Where 0 represents missing features and 1 represents available features. The dataset is randomly divided into training and testing sets, and class attribute labels w are constructed for each sample in the training and testing sets. i∈{0,1}, where the samples in the training set are labeled 0 and the samples in the test set are labeled 1. Finally, the dataset after imputation of missing data is represented as follows:

[0028] As a specific implementation method, a sample represents multi-source, multi-attribute data corresponding to an aerial target. For example, when the aerial target is a drone, the fifth sample could be composed of the drone's wingspan x 5,1 fuselage length x 5,2 and cruising speed x 5,3 In a multi-source, multi-attribute drone data set consisting of three attribute values, if the binary mask vector M(x5) = m5 = [0, 1, 1], then the sample is missing data, specifically the drone's wingspan x. 5,1 For the missing features, fuselage length x 5,2 and cruising speed x 5,3 These are available attribute features. When w5∈{1}, it indicates that the sample is a sample in the test set; otherwise, w5∈{0} indicates that the sample is a sample in the training set.

[0029] The generator training method in this invention includes: using multi-source observation data from the target domain and the source domain as training samples; wherein, the multi-source observation data of the target domain comes from the multi-source observation data to be interpolated; using random noise sources, training samples, and the mask vectors corresponding to the training samples as inputs to the generator to obtain the completed multi-source observation data output by the generator; using the completed multi-source observation data and cue vectors as inputs to generate a discriminator in the adversarial interpolation network to obtain the prediction domain labels and prediction mask vectors of the training samples; and training the adversarial interpolation network based on the prediction mask vectors, the completed multi-source observation data, the prediction domain labels, and the information entropy loss.

[0030] Step 1: Class attribute offset correction.

[0031] First, a classification system is defined and trained. The specific method is as follows: select multi-source observation data with low missing rate from the training samples; cluster the multi-source observation data with low missing rate to generate pseudo-labels for each multi-source observation data with low missing rate; and train the classifier based on the pseudo-labels.

[0032] Step 1-1: Filter data with low missing rate.

[0033] Although the missing rate of multi-source, multi-attribute aerial target data is fixed, the missing rate of each sample will differ. Therefore, all samples are sorted by their missing rates. For the i-th training sample x... i Missing rate mr(x) i The formula for calculating ) is:

[0034]

[0035] Where, m i,j Let d represent the mask of the j-th attribute value in the mask vector of the i-th training sample, where d is the number of attributes.

[0036] Step 1-2: Clustering to obtain pseudo-labels.

[0037] Data with low missing rates are selected as original knowledge samples with class priors, and pseudo-labels are generated for them using a clustering algorithm. For example, in a dataset consisting of 100 multi-source, multi-attribute data points, their missing rates are sorted from smallest to largest. Then, starting with the data with the smallest missing rate, 30 data points are selected as original knowledge samples with class priors, and pseudo-labels are assigned to each data point using a clustering algorithm to create a dataset.

[0038] To preserve the prior information of class attributes in the initial multi-source, multi-attribute data, no imputation or preprocessing operations are performed on these data with low missing rates. Subsequently, the obtained pseudo-labels are combined with the corresponding data to construct a complete aerial target dataset.

[0039] Steps 1-3: Pre-train classifier C.

[0040] The classifier C is constructed from simple linear layers, and the training process uses the cross-entropy loss function. The trained classifier C is used to constrain the data generated by the generator G in the subsequent domain adaptation interpolation network, so that the generator G can learn attribute features of different categories. For the constraint function, the standard information entropy loss is used, and its calculation formula is:

[0041]

[0042] in, This represents the completed multi-source observation data output by the generator from i training samples. express The output after classifier C, 1e -7 It is represented as a minimum value, usually 0.0000001.

[0043] In pre-trained classifiers, there are several feasible alternatives for obtaining pseudo-labels. For example, different clustering algorithms can be used to generate pseudo-labels; at the same time, self-training or active learning mechanisms can be combined with the framework to first generate initial pseudo-labels using clustering, and then continuously update the label set through self-iterative optimization or pseudo-active learning strategies, thereby further improving the robustness and generalization performance of the model.

[0044] Step 2: Adaptive interpolation of the adversarial domain.

[0045] Step 2-1: To adaptively interpolate missing aerial target attribute data based on existing multi-source, multi-attribute data, this invention uses GAIN as the basic model framework. In GAIN, the generator G takes the incomplete sample set X, the mask vector, and the random noise source Z as input, completes the data interpolation, and outputs complete data. The discriminator D takes the sample set interpolated by the generator G... The algorithm takes the cue vector H as input and attempts to distinguish which samples in the imputed sample set are imputed values ​​and which are observed values.

[0046] The cue vector H is the "partially visible information" of the mask vector M, used to provide the discriminator with controlled prior information about missing locations during the training phase. Therefore, H = M⊙B + 0.5(1-B), where B represents a random binary vector, B∈{0,1}. d .

[0047] Specifically, the output of G is defined as:

[0048]

[0049] in, This indicates element-wise multiplication. The interpolated sample set. Defined as:

[0050]

[0051] The output of D is a binary vector. The prediction mask vector is defined as follows:

[0052]

[0053] The optimization objective of GAIN is defined as:

[0054]

[0055] Where α is the weighting parameter, m i For x i mask vector, for The prediction mask vector.

[0056] L D The cross-entropy loss of the mask is defined as:

[0057]

[0058] L G Defined as the generator's mask adversarial loss, it is used to constrain the features generated by the generator at missing locations to be indistinguishable from the true observations in the discriminator's view. Specifically:

[0059]

[0060] L M The mean squared error loss function is defined as follows:

[0061]

[0062] Where, m i,j Let x represent the mask of the j-th attribute value in the mask vector of the i-th training sample. i,j This represents the j-th attribute value in the i-th training sample. This indicates that the j-th attribute value is to be completed in the multi-source observation data.

[0063] Step 2-2: The core motivation of the adversarial loss function lies in using a generator function that can effectively capture the data distribution, making it difficult for the discriminator to distinguish between observed and imputed values ​​in the feature vector. However, when the lack of multi-source data leads to a large difference in the distribution between the training and test sets, it may not be possible to accurately identify this difference in distribution from a microscopic perspective alone, resulting in a significant distributional difference between the imputed training and test sets. Therefore, this invention constructs a novel dual-head discriminator and loss function to identify this difference from a macroscopic perspective of the samples.

[0064] To more fully model the complex dependencies between features in multi-source, multi-attribute data of aerial targets, this invention introduces a self-attention mechanism and a residual block into the generator structure to achieve global correlation modeling and feature consistency maintenance. Specifically, the generator's input consists of concatenated sample features and a missing mask (X; M), and feature representations are obtained through linear transformation and nonlinear activation.

[0065] X0 = ReLU(W0[X;M] + b0) (11)

[0066] Where [X; M] represents the input vector obtained by concatenating the sample feature vector X and the mask vector M along the feature dimension, W0 represents the weight matrix of the first linear transformation layer (fully connected layer) of the generator, which is used to map the concatenated input to the hidden representation space, and b0 represents the bias vector corresponding to the linear transformation.

[0067] Self-attention mechanisms are used to capture global dependencies between attributes. Let the input be... Generate the query matrix Q, key matrix K, and value matrix V through linear transformation:

[0068] Q = X0W Q K = X0W K V = X0W V (12)

[0069] Among them, W Q W represents the weight matrix that linearly projects the input feature X0 into the query matrix Q. K This represents the linear projection of the input feature X0 into the weight matrix and W of the key matrix K. V This represents the weight matrix that linearly projects the input feature X0 into the value matrix V.

[0070] The self-attention weight A is calculated by scaling the dot product:

[0071]

[0072] Where, d k The characteristic dimension of the key matrix K is represented.

[0073] Finally, the globally weighted feature representation is obtained:

[0074] Y = AV (14)

[0075] The generator employs a two-layer self-attention module and a residual connection structure. The first-layer self-attention module uses a residual structure design, where its output and input features are element-wise added through skip connections to form a residual mapping result. This residual mapping result, after batch normalization and ReLU nonlinear activation, serves as the input feature for the second-layer self-attention module, obtaining the final output feature Y2 after nonlinear mapping as the imputation result. Through this design, the generator can simultaneously capture local and global feature relationships during the imputation process, effectively mitigating the feature shift problem in multi-source, multi-attribute data of aerial targets and improving the class consistency and global reasonableness of the imputed values. This global perspective allows the generator to better understand the relationships between data, thus predicting missing values ​​more accurately. Simultaneously, the self-attention mechanism helps the generator better learn class relationships from the constraints of class attributes. This means the generator can better capture feature differences between different classes and reduce the risk of class attribute shift during imputation. Therefore, by introducing the self-attention mechanism, the generator can more effectively generate imputed values ​​that conform to the data distribution and reduce the potential class attribute shift problem during imputation.

[0076] GANs can be used to distinguish between real and fake samples. In contrast, the method of this invention creates new domain labels W for the training and test sets. The generator is responsible for mitigating the distributional differences caused by missing data, while the discriminator tries its best to distinguish between training and test set data. Figure 1 As shown, the constructed discriminator has two outputs, evaluating this distributional difference from both macroscopic and microscopic perspectives, and can be defined as follows:

[0077]

[0078] in, The predicted values ​​used to distinguish between the training and test sets, This is the predicted value of the mask vector M. It's important to note that, to balance the strength of the discriminator, the cue matrix H is fed as an intermediate input to the discriminator.

[0079] In summary, the optimization objectives for the discriminator and generator can be redefined as:

[0080]

[0081] Where α, β, and γ are hyperparameters. The details have been explained in Formula 2. The cross-entropy loss function representing the domain label can be defined as:

[0082]

[0083] By using The loss function continuously enhances the discriminator's ability to distinguish between the training and test sets. Simultaneously, the generator domain employs adversarial loss. It can be defined as:

[0084]

[0085] in, Indicates generator domain adversarial loss, w i This represents the true domain label of the i-th multi-source observation data. This represents the prediction domain label for the i-th multi-source observation data.

[0086] As can be seen from the above formula, when When m approaches 0 i =1; when When m approaches 1 i =0,L GD Minimum. That is, when the discriminator's ability to recognize interpolated values ​​is weak (unable to distinguish between the training and test sets), The value will be smaller. This is achieved by introducing additional entropy loss. This encourages the model to learn more diverse classification features to correct for potential class attribute shifts during interpolation. Simultaneously, it introduces... The loss helps the model learn the sample distribution of the training and test sets, continuously reducing the distribution difference of the imputed data.

[0087] Steps 2-3: This invention uses an iterative approach to solve the optimization problem. The pseudocode is shown in Table 1. Where k G and k DThe minimum batch size represents the optimization process of the generator and discriminator, and epoch represents the number of training iterations. To more effectively optimize the performance of the Domain Adaptive Imputation Network (GADAIN) of this invention, an update strategy of training the generator for three epochs and the discriminator for one epoch is adopted. It is worth noting that index l here corresponds to the number of samples in the batch.

[0088] Table 1

[0089]

[0090]

[0091] In Table 1, x Gl This represents the generated feature vector output by generator G after modeling the missing attributes of the l-th sample; This is an operator that calculates the gradient of the discriminator D parameters, used to update the discriminator model parameters; This operator represents the gradient of the generator G parameters, used to update the generator model parameters with a fixed discriminator. During the training of the generative adversarial interpolation network, the discriminator is updated less frequently than the generator. This is to prevent the discriminator from converging too quickly, which could lead to poor generator training performance; that is, to ensure that the generator's performance gradually increases with the number of training iterations.

[0092] Step 3: Identification of multi-attribute data of aerial targets.

[0093] After imputing missing data, various classification methods can be used to identify multi-source, multi-attribute data of aerial targets. The imputed data has high integrity and distribution consistency, providing reliable input for subsequent classification models. Depending on the task requirements, traditional machine learning classifiers (such as support vector machines, random forests, K-nearest neighbors, etc.) or deep learning models (such as convolutional neural networks, fully connected neural networks, or Transformer structures based on attention mechanisms) can be used for feature learning and category discrimination.

[0094] In the specific implementation, the imputed data is first standardized and feature-selected to remove redundant features and enhance data discriminativeness; then the preprocessed samples are input into the classification model for training and prediction. The classifier achieves accurate identification of different categories of aerial targets by comprehensively learning multi-source attributes (such as acoustic, radar, infrared, and optical features).

[0095] In summary, this invention proposes a multi-source data interpolation method for aerial targets based on generative adversarial networks (GANs) and domain adaptation mechanisms. This method effectively corrects the distributional bias between the training and test sets and recovers missing data features. Specifically, this invention designs a GAN model that integrates a dual-head discriminator and a self-attention mechanism to achieve global dependency modeling and accurate interpolation of multi-source, multi-attribute features. An auxiliary classifier is introduced to guide the interpolation process, correcting class attribute biases during generation and fully utilizing prior class information to improve the discriminative power of the interpolated data.

[0096] This invention addresses the issues of distribution shift and insufficient imputation quality caused by missing values ​​in the incomplete classification of multi-source, multi-attribute aerial target data. It proposes a data distribution shift correction and missing data imputation method based on generative adversarial networks (GANs) and domain adaptation. This method introduces a distribution alignment mechanism during the data imputation process, mapping the training and test sets to a new relevant common space, achieving consistent distributions within this space, thus effectively mitigating the distribution mismatch problem caused by missing multi-source features. By designing a generative adversarial model that integrates a dual-head discriminator and a self-attention mechanism, the realism and consistency of the imputed samples are improved. Furthermore, by combining clustering pseudo-labels for low-missing-rate samples with an auxiliary classifier, bias correction of category attributes is achieved, fully utilizing prior class knowledge.

[0097] Compared with existing methods, this invention maintains high imputation accuracy and classification reliability even under conditions of multi-source, multi-attribute aerial target data with high missing rates, significantly improving the model's generalization ability and robustness. This method not only enhances the reliability of aerial target identification in multi-source missing data environments but also provides a scalable and transferable technical solution for correcting distribution shifts in complex multi-source data.

[0098] The core idea of ​​this invention is to address the distribution shift problem caused by missing data during the imputation of missing values. Besides this solution, other strategies can also achieve the same technical objective: reducing distribution shift while maintaining the integrity of data features, thereby improving classification accuracy and stability.

[0099] To address the lack and incompleteness of attribute data in aerial target recognition tasks, a simulation method for UAV attribute data based on statistical modeling of common attributes and constraint sampling is proposed. This method uses actual UAV parameters as a basis and generates a simulation dataset that conforms to physical laws through multi-parameter constraints and scaling, which is then used to train a classification model.

[0100] The raw data comes from a publicly available UAV parameter database, containing attribute information for 93 UAV types. Key parameters include: Maximum Takeoff Weight (MTOW), Payload, Wingspan, Length, Cruise Speed, Maximum Speed, Endurance, and Altitude. During the simulation, 12 UAV types with the most complete parameter sets were selected as the original models, and only common attributes were retained to ensure dimensional consistency in the generated data.

[0101] Let the original sample of the drone be:

[0102]

[0103] in, Represents an attribute vector, y i The category is represented by the drone model name. The goal is to generate 500 simulation samples for each model category, ensuring that the data attribute dimensions are consistent across all categories, satisfying both physical and statistical constraints. The overall method flow is as follows:

[0104] a. Read raw drone data from the original file;

[0105] b. Extract the numerical attributes common to all categories;

[0106] c. Simulate based on the true attribute range of each type of sample;

[0107] d. Correct outliers according to physical constraint rules;

[0108] e. Output a uniformly structured .txt simulation dataset.

[0109] For category c (c∈y) i The simulation samples are generated by sequentially extracting baseline samples from the original samples and proportionally perturbing them.

[0110]

[0111] in, The attribute parameter value of the j-th simulated sample representing the c-th type of target. The attribute parameters ε of the j-th benchmark sample selected from the c-th original samples represent the following: j Let U(0.9, 1.1) represent the proportion of random disturbance introduced into the j-th reference sample, and let U(0.9, 1.1) represent a uniform distribution following the interval [0.9, 1.1] to introduce reasonable random fluctuations. To ensure that the generated samples conform to the dynamic laws of the UAV, the following constraint functions are added:

[0112] a. Range constraint (range × cruising speed):

[0113] Range=CruiseSpeed×Endurance×σ,σ~U(0.8,1.1) (22)

[0114] Where Range represents the maximum range of the UAV, CruiseSpeed ​​represents the cruise speed, Endurance represents the endurance, and σ represents the correction coefficient, which follows a uniform distribution within an interval and is used to characterize the range fluctuation under actual flight conditions.

[0115] b. Load and takeoff weight constraints:

[0116] Payload≤MTOW×η,η~U(0.3,0.7) (23)

[0117] Where MTOW represents the maximum takeoff weight of the UAV, and η represents the ratio between the payload and the maximum takeoff weight.

[0118] c. Wing aspect ratio constraint Chord Lenth:

[0119]

[0120] Where ChordLenth represents the average chord length of the wing, Wingspan represents the wingspan, AspectRatio represents the aspect ratio of the wing, and the correction coefficient ξ follows a uniform distribution in the interval [0.9, 1.1].

[0121] After obtaining the complete simulation dataset, this invention performs missing data processing: the complete dataset is randomly shuffled, and 50% is taken as the training set, with the first half of its attributes randomly missing; the remaining 50% is taken as the test set, with the second half of its attributes randomly missing. This dataset serves as the UAV_miss dataset for imputation and recognition of multi-source, multi-attribute missing data of aerial targets. Table 2 shows a portion of this dataset, where Length represents the aircraft length and Label represents the tag.

[0122] Table 2

[0123]

[0124]

[0125] The methods compared during the experiment and their abbreviations are as follows: Mean Imputation (MI), Multiple Imputation by Chained Equations (MICE), Generative Adversarial Imputation Nets (GAIN), Pseudo-label Conditional Generative Adversarial Imputation Nets (PCGAIN), Transformed Distribution Matching (TDM), and Hyperimpute (Hyper).

[0126] For MI, 'SimpleImputer' from the 'sklearn.impute' database was used for data imputation. For other methods, the code and default parameter settings provided in the original paper were used.

[0127] Table 3. Classification results (%) on the UAV-miss dataset.

[0128]

[0129] Table 3 shows the classification results of imputed data using different imputation methods under different missing rates using different classifiers (the differences between the selected base classifiers are not within the scope of this invention). Here, 'Cla' represents different base classifiers and 'MR' represents the missing rate.

[0130] Specifically, when the missing rate is high (MR=0.8), compared with GAIN, PCGAIN, TDM and Hyper, the accuracy of the method GADAIN of this invention is improved by 7.5%, 4.58%, 5.84% and 10.62% respectively under the KNN classifier.

[0131] GAIN and PCGAIN employ a generative adversarial network framework, where the generator produces data that closely approximates the distribution of real-world aerial target multi-source data, while the discriminator distinguishes differences by determining whether target attributes in the data are imputed. However, missing data can lead to distributional discrepancies between regions within the entire dataset. This can cause the generator to suffer from pattern collapse, where generated samples tend to fall into a specific local region of the dataset, failing to adequately cover the entire distribution. This can result in imputed values ​​being overly concentrated in certain local areas, ignoring other regions.

[0132] TDM leverages the fact that "any two batches of missing data come from the same data distribution." It transforms two batches of samples into a latent space using a depthwise invertible function and matches them distributionally to impute missing values. However, when the distributions of the data subsets differ significantly, the depthwise invertible function may encounter stability issues. This is because different data distributions can lead to instability in the representation within the latent space, making the matching process difficult and affecting the stability and reliability of the imputation results.

[0133] In summary, the methods described above all consider the problem of data tending towards a uniform distribution, thus performing well with low missing rates. However, as the missing rate increases, the data distribution shifts, and these methods struggle to effectively mitigate this shift, resulting in less than ideal classification results.

[0134] Figure 2 The visualizations show the t-SNE results after applying various imputation methods at different missing rates. Orange and green represent the training set, where 'Ori' represents the original dataset without imputation after missing data, 'MICE' and 'GAIN' represent the results after applying the corresponding imputation methods, and 'Ours' represents the method proposed in this invention. These visualizations intuitively demonstrate the impact of different imputation methods on the data, helping to understand the distribution of the data after imputation.

[0135] As observed from the raw data, the training and test sets gradually tend towards two different distributions as the missing data rate increases. In this situation, a classifier trained on the training set struggles to achieve good results on the test set. Therefore, imputation preprocessing is essential for datasets with missing data. The results after imputation show that the difference between GAIN and the proposed method is small when the missing data rate is low. However, when the missing data rate reaches 60%, the data after GAIN imputation still tends towards two distributions, while the method of this invention can reduce the difference between the data distributions, making them more uniform.

[0136] The present invention also discloses an aerial target multi-source data completion device based on deep generation domain adaptation, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the above-described method when executing the computer program.

[0137] The present invention also discloses an embodiment that provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0138] The present invention also provides a computer program product that, when run on a data storage device, enables the data storage device to implement the steps in the above-described method embodiments.

[0139] If the integrated unit module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a storage device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0140] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0141] Those skilled in the art will recognize that the algorithmic steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

Claims

1. A method for multi-source data completion of aerial targets based on deep generation domain adaptation, characterized in that, Includes the following steps: Acquire multi-source observation data of aerial targets from different sensors, and use the multi-source observation data as input to the generator in the generative adversarial interpolation network to obtain interpolated and complete multi-source observation data. The generator includes a cascaded linear transformation layer, a first self-attention mechanism module, and a second self-attention mechanism module. The first self-attention module adopts a residual structure design, and its output features and input features are added element-wise through skip connections to form a residual mapping result. The residual mapping result is then batch normalized and nonlinearly activated as the input features of the second self-attention mechanism module. Specifically, when training the generator in the generative adversarial interpolation network, generator domain adversarial loss and information entropy loss are added; the generator domain adversarial loss is calculated based on the domain labels of the multi-source observations.

2. The method for multi-source data completion of aerial targets based on deep generation domain adaptation as described in claim 1, characterized in that, The generator domain adversarial loss is defined as: in, w represents the generator domain adversarial loss. i This represents the true domain label of the i-th multi-source observation data. This represents the prediction domain label for the i-th multi-source observation data.

3. The method for multi-source aerial target data completion based on deep generation domain adaptation as described in claim 2, characterized in that, The information entropy loss is defined as: in, This represents the information entropy loss. This indicates the i-th complete multi-source observation data. This represents the output of the classifier after the completed multi-source observation data has passed through it, 1e -7 It is a local minimum.

4. The method for multi-source aerial target data completion based on deep generation domain adaptation as described in claim 2 or 3, characterized in that, The training method for the generator includes: Multi-source observation data from the target domain and the source domain are used as training samples; wherein, the multi-source observation data of the target domain comes from the multi-source observation data to be interpolated. The random noise source, the training sample, and the mask vector corresponding to the training sample are used as inputs to the generator to obtain the completed multi-source observation data output by the generator. The discriminator in the generative adversarial interpolation network is input with the completed multi-source observation data and cue vector to obtain the prediction domain label and prediction mask vector of the training sample; The generative adversarial interpolation network is trained based on the predicted mask vector, the completed multi-source observation data, the predicted domain label, and the information entropy loss.

5. The method for multi-source data completion of aerial targets based on deep generation domain adaptation as described in claim 4, characterized in that, The training method for the classifier is as follows: Multi-source observation data with low missing rates are selected from the training samples; Cluster the multi-source observation data with low missing rate and generate pseudo-labels for each multi-source observation data with low missing rate. The classifier is trained based on the pseudo-labels.

6. The method for multi-source data completion of aerial targets based on deep generation domain adaptation as described in claim 4, characterized in that, During the training of the generative adversarial interpolation network, the discriminator updates at a lower frequency than the generator.

7. The method for multi-source data completion of aerial targets based on deep generation domain adaptation as described in claim 6, characterized in that, When training the generator in the generative adversarial interpolation network, the loss function also includes mean squared error loss and mask adversarial loss; in, This represents the masked adversarial loss, where M represents the mask vector corresponding to the training sample. Represents the prediction mask vector. x represents the mean squared error loss. i This represents the i-th training sample. Let m represent the completed multi-source observation data output by the generator for the i-th training sample. i,j Let x represent the mask of the j-th attribute value in the mask vector of the i-th training sample. i,j This represents the j-th attribute value in the i-th training sample. This indicates that the j-th attribute value is completed in the multi-source observation data, and d represents the number of attributes of the i-th training sample.

8. The method for multi-source data completion of aerial targets based on deep generation domain adaptation as described in claim 7, characterized in that, When training the discriminator in the generative adversarial interpolation network, the loss function includes the cross-entropy loss of the mask and the cross-entropy loss of the domain label; in, This represents the cross-entropy loss of the mask. This represents the cross-entropy loss of the domain label.

9. A device for multi-source data completion of aerial targets based on deep generation domain adaptation, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-8.

Images