A data enhancement method and device for unmanned aerial vehicle anomaly detection
By constructing a generative network architecture based on relational analysis, synthetic signals similar to real signals are generated, which solves the problems of data scarcity and insufficient model generalization in UAV anomaly detection, and improves the accuracy and robustness of UAV anomaly detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENYANG AIRCRAFT DESIGN INST AVIATION IND CORP OF CHINA
- Filing Date
- 2024-07-11
- Publication Date
- 2026-06-12
AI Technical Summary
The detection of drone anomalies suffers from the scarcity of flight feature data and insufficient model generalization ability, especially in key areas such as public safety and airspace management, where existing technologies struggle to effectively and accurately detect anomalies.
A relation-based generative network architecture is adopted, including a signal vector generation network and a signal analysis network. By alternately optimizing the generation and analysis processes, a synthetic signal similar to the real signal is generated, increasing the amount of data and improving the data quality. The generalization ability of the model is improved by using fully connected attribute pairing loss and staggered vector distortion loss optimization strategies.
It significantly improves the performance of the UAV anomaly detection system, enhances the accuracy and robustness of the data-driven method, strengthens the model's generalization ability and data quality, and ensures the reliability of anomaly detection.
Smart Images
Figure CN119026034B_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of data processing technology, and specifically relates to a data augmentation method and apparatus for anomaly detection of unmanned aerial vehicles. Background Technology
[0002] With the rapid development of unmanned aerial vehicle (UAV) technology, its widespread application in various fields such as military, commerce, and scientific research is gradually becoming a reality. However, this has led to an urgent need for accurate prediction of UAV behavior and anomaly detection, especially in critical areas such as public safety and airspace management. UAV anomaly detection is mainly divided into two methods: model-based and data-driven.
[0003] Model-based anomaly detection methods utilize mathematical models and statistical techniques to detect anomalies in unmanned aerial vehicle (UAV) systems. First, a mathematical model describing the system's normal behavior is established. Then, a system model is constructed based on the UAV system's operating mechanism, historical data, or domain knowledge. While model-based anomaly detection methods generally perform well, the established models typically lack universality with other aircraft types.
[0004] Data-driven anomaly detection methods primarily utilize and analyze data generated by UAV sensors for anomaly detection. Compared to model-based methods, data-driven methods offer greater versatility. However, they require substantial amounts of data. In practical UAV use, limitations imposed by the onboard equipment and sensor performance typically prevent the recording of large volumes of high-quality flight data. Summary of the Invention
[0005] To address the aforementioned issues, this application provides a data augmentation method and apparatus for anomaly detection in unmanned aerial vehicles (UAVs), thereby resolving the problems of scarce UAV flight feature data and insufficient model generalization ability.
[0006] The first aspect of this application provides a data augmentation method for anomaly detection in unmanned aerial vehicles, mainly including:
[0007] Step S1: Obtain flight time series data for UAV anomaly detection as the real signal;
[0008] Step S2: Construct a generative network architecture based on relation analysis. The generative network architecture based on relation analysis includes a signal vector generation network and a signal analysis network. The signal vector generation network includes a first parallel relation generation layer, which performs a first dot product operation on the input vector, and then outputs a synthesized signal after passing through a first feedforward multilayer perceptron and a convolutional layer. The signal analysis network includes a second parallel relation generation layer, which performs a second dot product operation on the synthesized or real signals, and then outputs the predicted signal type after passing through a second feedforward multilayer perceptron and a classification layer. The signal vector generation network uses a fully connected attribute pairing loss to optimize the strategy parameters so that the output synthesized signal is more similar to the real signal. The signal analysis network uses an interleaved vector distortion loss to optimize the strategy parameters so that it can better distinguish between the synthesized signal and the real signal.
[0009] Step S3: The signal vector generation network continuously outputs the synthesized signal and inputs it into the signal analysis network. The two networks alternately optimize each other to obtain a trained generative network architecture based on relation analysis.
[0010] Step S4: The signal vectors of the trained relation-based generative network architecture are used to generate a sufficient amount of synthetic signals to augment the original real signals.
[0011] Preferably, in step S2, the signal vector generation network includes a noise input unit, which generates a normally distributed noise sequence with the same mean and variance as the flight time series data, and converts it into an input vector I through position embedding and normalization.
[0012] Preferably, step S2, before performing the second dot product operation on the two types of signals (synthetic or real), further includes:
[0013] For both synthetic and real signals, each signal class is classified into c channels according to its features. Each channel is divided into multiple patches of a specified size N. A soft positional encoding value is added to the front of each patch. Then, the input vector for the second dot product operation is formed by positional embedding.
[0014] Preferably, in step S2, the dot product operation includes:
[0015]
[0016] Where I is the input vector or signal, C is the context vector, W is the weight vector, and d C The dimension of the context vector is given by PRG(I,C,W), which is the output of the dot product operation.
[0017] Preferably, in step S2, the interlaced vector distortion loss LCVD (A) is:
[0018]
[0019] Where N is the number of signals, and for the i-th signal, y i It's a label; for the real signal, y i For the synthesized signal, y is 1. i If f is 0, then f is A(x) i A(x) is obtained through a nonlinear transformation operation. i The signal analysis network provides the sample x. i The probability of the actual data.
[0020] Preferably, in step S2, the fully connected attribute pairing loss L FCAP (G) is:
[0021]
[0022] Where φ represents the activation function of the intermediate layer of the signal analysis network, G(z) i ) is the synthesized signal output by the signal vector generation network, x i N represents the actual signal, and N is the number of signals.
[0023] A second aspect of this application provides a data augmentation device for anomaly detection in unmanned aerial vehicles, mainly comprising:
[0024] The aircraft sequence data acquisition module is used to acquire flight time series data for UAV anomaly detection as a real signal;
[0025] A generative network architecture building module is used to construct a generative network architecture based on relation analysis. The generative network architecture based on relation analysis includes a signal vector generation network and a signal analysis network. The signal vector generation network includes a first parallel relation generation layer, which performs a first dot product operation on the input vector, and then outputs a synthesized signal after passing through a first feedforward multilayer perceptron and a convolutional layer. The signal analysis network includes a second parallel relation generation layer, which performs a second dot product operation on the synthesized or real signals, and then outputs the predicted signal type after passing through a second feedforward multilayer perceptron and a classification layer. The signal vector generation network uses a fully connected attribute pairing loss to optimize the strategy parameters so that the output synthesized signal is more similar to the real signal. The signal analysis network uses an interleaved vector distortion loss to optimize the strategy parameters so that it can better distinguish between the synthesized signal and the real signal.
[0026] The generative network architecture training module is used to continuously output synthesized signals through the signal vector generation network and input them into the signal analysis network. The two networks alternately optimize each other to obtain a trained generative network architecture based on relation analysis.
[0027] The synthetic signal generation module is used to generate a sufficient amount of synthetic signals from the signal vectors of a pre-trained relation-analysis-based generative network architecture to augment the original real signals.
[0028] Preferably, the generative network architecture building module includes an input vector generation unit, which generates a noise sequence with the same mean and variance as the time-of-flight data based on the noise input unit, and converts it into an input vector I through position embedding.
[0029] Preferably, the generative network architecture building module includes a sequence data processing unit, which classifies each type of synthetic signal or real signal into c channels according to features, divides each channel into multiple patches of a specified size N, and adds a soft positional encoding value to the front end of each patch to form an input vector for performing the second dot product operation.
[0030] Preferably, the parallel relation generation layer is used to perform a dot product operation, the dot product operation including:
[0031]
[0032] Where I is the input vector or signal, C is the context vector, W is the weight vector, and d C The dimension of the context vector is given by PRG(I,C,W), which is the output of the dot product operation.
[0033] This application utilizes advanced relational analysis and generation technology to augment UAV flight data, significantly improving the performance of a data-driven UAV anomaly detection system by increasing data volume and enhancing data quality. Attached Figure Description
[0034] Figure 1 This is a schematic diagram of a generative network architecture based on relational analysis, representing a preferred embodiment of the data augmentation method for anomaly detection of unmanned aerial vehicles (UAVs) according to this application.
[0035] Figure 2 This application Figure 1 The schematic diagram of the parallel relationship generation layer in the embodiment shown. Detailed Implementation
[0036] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are only some, not all, of the embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.
[0037] The first aspect of this application provides a data augmentation method for anomaly detection in unmanned aerial vehicles, such as... Figure 1 As shown, it mainly includes:
[0038] Step S1: Obtain flight time series data for UAV anomaly detection as the real signal;
[0039] Step S2: Construct a generative network architecture based on relation analysis. The generative network architecture based on relation analysis includes a signal vector generation network and a signal analysis network. The signal vector generation network includes a first parallel relation generation layer, which performs a first dot product operation on the input vector, and then outputs a synthesized signal after passing through a first feedforward multilayer perceptron and a convolutional layer. The signal analysis network includes a second parallel relation generation layer, which performs a second dot product operation on the synthesized or real signals, and then outputs the predicted signal type after passing through a second feedforward multilayer perceptron and a classification layer. The signal vector generation network uses a fully connected attribute pairing loss to optimize the strategy parameters so that the output synthesized signal is more similar to the real signal. The signal analysis network uses an interleaved vector distortion loss to optimize the strategy parameters so that it can better distinguish between the synthesized signal and the real signal.
[0040] Step S3: The signal vector generation network continuously outputs the synthesized signal and inputs it into the signal analysis network. The two networks alternately optimize each other to obtain a trained generative network architecture based on relation analysis.
[0041] Step S4: The signal vectors of the trained relation-based generative network architecture are used to generate a sufficient amount of synthetic signals to augment the original real signals.
[0042] This application constructs a generative network architecture based on relational analysis in step S2, trains it in step S3, and generates a synthetic signal in step S4 to augment the original real signal. Specifically, in step S3, for this adversarial learning network, the completion of training can be determined using the following method:
[0043] (1) Loss values of the generator and analytics networks. If the losses of both tend to stabilize, and the loss of the generator network does not continue to rise or fall, it may mean that the model has reached an equilibrium. (2) Visually check the quality of the generated data. (3) Compare the generated samples with the training data. Check whether the generated samples are too similar to the training data to avoid overfitting. (4) Ensure that the game between the generator and analytics networks is in balance. If the analytics network is too strong, the generator network cannot learn effective features; if the generator network is too strong, the analytics network cannot distinguish between real and fake samples.
[0044] Monitor the generated network and analyze its loss curves during training to observe whether they tend to plateau.
[0045] Generative networks based on relational analysis are deep learning models consisting of two key components: a signal vector generation network (…). Figure 1 The upper part), and a signal analysis network ( Figure 1 (The lower half). They are all built on a relation analysis network encoder architecture. The encoder consists of two composite blocks. A parallel relation generation layer builds the first module, and the second module is a feedforward multilayer perceptron (MLP) with a GELU activation function. A normalization layer is applied before both blocks, and a random deactivation layer is added after each block. Both blocks use residual connections.
[0046] The core function of a signal vector generation network (SVR) is to transform input noise into a vector representation with specific characteristics by deeply analyzing and learning the complex relationships inherent in the training data, thereby generating a synthetic signal with specific relationships. The task of a signal analysis network (SAN) is to analyze the input signal to determine its signal type or characteristics and feed this information back to the SVR. In other words, the goal of the SVR is, after training, to generate synthetic signals that are highly similar to real signals, achieving a level of realism that is difficult to distinguish from genuine signals. The goal of the SAN, after training, is to accurately distinguish between synthetic and real input signals. The two networks work against each other, a process that continues until a dynamic equilibrium is reached. At this point, the SVR can produce high-quality signal vectors sufficient to confuse the SAN. The SAN's ability to determine signal type or characteristics balances the SVR's generation ability. Through alternating optimization, the two networks mutually promote each other, gradually improving the quality of the synthetic data and the SAN's discriminative ability.
[0047] In some alternative implementations, the signal vector generation network includes a noise input unit for generating a normally distributed noise sequence with the same mean and variance as the time-of-flight data, and converting it into an input vector I through position embedding and normalization.
[0048] refer to Figure 1 After collecting flight time series data in step S1, the mean and variance of the flight time series dataset are calculated. The dataset is then normalized using the obtained mean and variance. This ensures the data has a uniform scale and range, improving the model's training performance. Next, a set of normally distributed noise is randomly generated, with the same mean and variance as the aforementioned dataset. This noise is then passed through a fully connected layer for positional embedding, enabling the neural network to understand the order of parameters in the input vector and their dependencies.
[0049] In some alternative implementations, step S2, prior to performing the second dot product operation on the two types of signals (synthetic or real), further includes:
[0050] For both synthetic and real signals, each signal class is classified into c channels according to its features. Each channel is divided into multiple patches of a specified size N. A soft positional encoding value is added to the front of each patch. Then, the input vector for the second dot product operation is formed by positional embedding.
[0051] In this embodiment, the signal analysis network uses a parallel relation generation layer, which primarily processes image data. The core idea of the visual relation analysis network is to segment the input image into a series of image patches and then flatten these image patches into a one-dimensional vector sequence. Therefore, this application preprocesses the synthesized signal or real signal according to the sequence format required by the image data, specifically including:
[0052] The time series data is viewed as image data with a height of 1, where the number of time steps is defined as the width of the image, denoted as w. The time series can contain one or more channels (i.e., multiple features, such as velocity, pose, etc.), which are equivalent to the number of channels in the image (e.g., RGB), denoted as c. Therefore, the input sequence can be represented by a matrix of dimension (BatchSize, c, 1, w). For further processing, this application selects a specific patch size N and divides the sequence into w / N patches. Based on this, a soft positional encoding value is added to the end of each patch, which will be learned during model training. Finally, the shape of the data input to the encoder block of the signal analysis network is adjusted to (BatchSize, c, 1, (w / N) + 1).
[0053] In some alternative implementations, step S2 includes the dot product operation as follows:
[0054]
[0055] Where I is the input vector or signal, C is the context vector, W is the weight vector, and d C The dimension of the context vector is given by PRG(I,C,W), which is the output of the dot product operation.
[0056] The Parallel Relation Generation (PRG) layer is one of the core innovations of relation analysis networks. It enables the model to better capture global information and long-range dependencies when processing sequential data, without relying on the positional information of the sequence. The PRG layer transforms each element in the input sequence into three vectors: the input vector (I), the context vector (C), and the weight vector (W). These three vectors are obtained through a linear transformation of the input sequence. Next, by calculating the dot product between the input and context vectors and then applying a distributed balanced activation function (DistEq), the weight distribution between each input and all contexts is obtained. Finally, multiplying and summing these weights with their corresponding numerical vectors yields the output of the PRG layer. It should be noted that the dimension d of the context vector... C It is also a scaling factor for the dot product operation. When calculating the inner product of each row vector of matrices I and C, it is divided by d to prevent the inner product from becoming too large. C The square root of.
[0057] As those skilled in the art will understand, I is the input vector, and the model training process involves continuously modifying the context vector C and the weight vector W. After training is complete, the context vector C and the weight vector W are fixed. Furthermore... Figure 2 The paper presents a parallel relation generation layer network architecture. In fact, multiple stacked parallel relation generation layers are connected by fully connected layers. Figure 2 In this process, each parallel relation generation layer outputs one data point, and multiple data points are reconstructed into a vector.
[0058] In some alternative implementations, in step S2, the interlaced vector distortion loss L CVD (A) is:
[0059]
[0060] Where N is the number of signals, and for the i-th signal, y i It's a label; for the real signal, y i For the synthesized signal, y is 1. i If f is 0, then f is A(x) iA(x) is obtained through a nonlinear transformation operation. i The signal analysis network provides the sample x. i The probability of the actual data.
[0061] In this embodiment, during the training of the relation-based generative network, the signal analysis network and the signal vector generation network employ different loss functions to guide their learning and optimization. The signal analysis network uses interleaved vector distortion loss, the core purpose of which is to measure the signal analysis network's ability to distinguish between real data and data generated by the signal vector generation network. By minimizing the interleaved vector distortion loss, the signal analysis network is trained to better differentiate between real and generated data.
[0062] In some alternative implementations, in step S2, the fully connected attribute pairing loss L FCAP (G) is:
[0063]
[0064] Where φ represents the activation function of the intermediate layer of the signal analysis network, G(z) i ) is the synthesized signal output by the signal vector generation network, x i N represents the actual signal, and N is the number of signals.
[0065] In this embodiment, the signal vector generation network uses a fully connected attribute pairing loss. This loss function focuses on the difference between the signal vector generation network's output and the feature representations of the real data in the intermediate layers of the signal analysis network. The purpose of the fully connected attribute pairing loss is to force the signal vector generation network to learn to generate data with features more similar to the real data, thereby improving the quality of the generated images. This loss function is typically calculated by comparing the statistical properties of the intermediate layer activations, such as mean and variance. By minimizing the fully connected attribute pairing loss, the signal vector generation network is forced to generate samples that are more indistinguishable from the real data in the feature space of the signal analysis network.
[0066] The relational analysis generative network of this application can learn the distribution and characteristics of the original data, thereby generating new data with similar statistical properties in step S4. Subsequently, the generated data is merged with the original data to construct a balanced training set, ensuring that the model can fully learn the characteristics of normal flight conditions.
[0067] After augmenting the data for drone anomaly detection, another deep learning model can be trained for drone anomaly detection. Commonly used models include sequence long-range dependency networks and convolutional networks. Deep learning models can learn complex data features and dynamic patterns, thereby improving the accuracy and robustness of anomaly detection.
[0068] After model training, the impact of noise on anomaly detection results can be reduced by calculating the model's prediction residuals on the validation set and filtering them using an IIR filter. By setting an appropriate threshold, the residuals can be correlated with the anomaly detection results to determine whether the flight status is abnormal. For cases where the anomaly detection results are uncertain, further analysis and processing can be performed to improve the system's robustness and reliability.
[0069] This application utilizes relation analysis to generate data via a generative network, effectively increasing training data and improving the model's generalization ability and robustness. This anomaly detection method, which uses relation analysis to generate data to enhance anomaly detection, has broad application prospects in fields such as drones and aircraft. Through continuous improvement and optimization, the performance and efficiency of anomaly detection can be further enhanced, providing more reliable assurance for the safe operation of systems such as aircraft.
[0070] To better demonstrate the advantages of this application, the following experiment illustrates the points.
[0071] In anomaly detection and classification based on relational analysis-generated network augmentation data, in addition to considering the number of true positives, false positives, true negatives, and false negatives, it is also necessary to pay attention to metrics such as accuracy, precision, recall, and F1 score. Accuracy measures the proportion of samples correctly classified by the classifier, while precision refers to the proportion of samples predicted as positive that are actually positive. Recall is the proportion of positive samples successfully identified by the classifier among the actual positive samples. The F1 score is the harmonic mean of precision and recall, which helps to evaluate the overall performance of the classifier. In fault conditions, if the residual is consistently below the threshold, the false negative count increases by 1; if the residual is greater than the threshold at least once, the true positive count increases by 1. Specific experimental results are shown in Table 1.
[0072] Table 1. Experimental Results
[0073]
[0074] The experimental results clearly demonstrate that the proposed method exhibits a significant performance advantage. In contrast, methods that do not utilize relational analysis to augment the training set with a data-enhanced dataset show performance deficiencies. The first three methods use neural networks to predict multi-dimensional features for anomaly detection, while the latter methods use long-range sequence dependency networks to predict single features for anomaly detection. A comparison of performance using long-range sequence dependency networks is conducted. First, the proposed method is trained using the augmented dataset; the remaining comparative experiments use the original training set. In the first two experiments, the output predictions represent the entire local state, i.e., outputting 10-dimensional features: velocity, attitude, and angular velocity. The subsequent three experiments use the original training set for training, outputting velocity, attitude, and angular velocity as outputs, respectively.
[0075] Experiments show that the invention presented in this application has significant performance advantages. In contrast, methods that do not use relational analysis to generate network augmentation data exhibit significant deficiencies in predictive performance. However, the method presented in this application still does not achieve 100% accuracy. This is because the original data has excessive noise, and the data augmented using the method presented in this application also has similar noise. This noise affects the model's predictive performance, and consequently, the accuracy of anomaly detection.
[0076] A second aspect of this application provides a data augmentation device for UAV anomaly detection corresponding to the above method, mainly comprising:
[0077] The aircraft sequence data acquisition module is used to acquire flight time series data for UAV anomaly detection as a real signal;
[0078] A generative network architecture building module is used to construct a generative network architecture based on relation analysis. The generative network architecture based on relation analysis includes a signal vector generation network and a signal analysis network. The signal vector generation network includes a first parallel relation generation layer, which performs a first dot product operation on the input vector, and then outputs a synthesized signal after passing through a first feedforward multilayer perceptron and a convolutional layer. The signal analysis network includes a second parallel relation generation layer, which performs a second dot product operation on the synthesized or real signals, and then outputs the predicted signal type after passing through a second feedforward multilayer perceptron and a classification layer. The signal vector generation network uses a fully connected attribute pairing loss to optimize the strategy parameters so that the output synthesized signal is more similar to the real signal. The signal analysis network uses an interleaved vector distortion loss to optimize the strategy parameters so that it can better distinguish between the synthesized signal and the real signal.
[0079] The generative network architecture training module is used to continuously output synthesized signals through the signal vector generation network and input them into the signal analysis network. The two networks alternately optimize each other to obtain a trained generative network architecture based on relation analysis.
[0080] The synthetic signal generation module is used to generate a sufficient amount of synthetic signals from the signal vectors of a pre-trained relation-analysis-based generative network architecture to augment the original real signals.
[0081] In some alternative implementations, the generative network architecture building module includes an input vector generation unit for generating a noise sequence with the same mean and variance as the time-of-flight data based on the noise input unit, and converting it into an input vector I through position embedding.
[0082] In some optional implementations, the generative network architecture building module includes a sequence data processing unit for classifying synthetic or real signals into c channels according to features, dividing each channel into multiple patches of a specified size N, and adding a soft positional encoding value to the front of each patch to form an input vector for performing the second dot product operation.
[0083] In some alternative implementations, the parallel relation generation layer is used to perform a dot product operation, the dot product operation including:
[0084]
[0085] Where I is the input vector or signal, C is the context vector, W is the weight vector, and d C The dimension of the context vector is given by PRG(I,C,W), which is the output of the dot product operation.
[0086] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A data augmentation method for anomaly detection in unmanned aerial vehicles, characterized in that, include: Step S1: Obtain flight time series data for UAV anomaly detection as the real signal; Step S2: Construct a generative network architecture based on relation analysis. The generative network architecture based on relation analysis includes a signal vector generation network and a signal analysis network. The signal vector generation network includes a first parallel relation generation layer, which performs a first dot product operation on the input vector, and then outputs a synthesized signal after passing through a first feedforward multilayer perceptron and a convolutional layer. The signal analysis network includes a second parallel relation generation layer, which performs a second dot product operation on the synthesized or real signals, and then outputs the predicted signal type after passing through a second feedforward multilayer perceptron and a classification layer. The signal vector generation network uses a fully connected attribute pairing loss to optimize the strategy parameters so that the output synthesized signal is more similar to the real signal. The signal analysis network uses an interleaved vector distortion loss to optimize the strategy parameters so that it can better distinguish between the synthesized signal and the real signal. Step S3: The signal vector generation network continuously outputs the synthesized signal and inputs it into the signal analysis network. The two networks alternately optimize each other to obtain a trained generative network architecture based on relation analysis. Step S4: The signal vectors of the trained relation-based generative network architecture are used to generate a sufficient amount of synthetic signals to augment the original real signals. In step S2, the signal vector generation network includes a noise input unit, which is used to generate a normally distributed noise sequence with the same mean and variance as the flight time series data, and convert it into an input vector I through position embedding and normalization; In step S2, before performing the second dot product operation on the two types of signals (synthetic or real), the following is further included: For both synthetic and real signals, each signal class is classified into c channels according to its features. Each channel is divided into multiple patches of a specified size N. A soft positional encoding value is added to the front of each patch. Then, the input vector for the second dot product operation is formed by positional embedding. In step S2, the dot product operation includes: ； Where I is the input vector or signal, C is the context vector, W is the weight vector, dC is the dimension of the context vector, PRG(I,C,W) is the output of the dot product operation, and DistEq is the distributed balanced activation function.
2. The data augmentation method for anomaly detection of unmanned aerial vehicles as described in claim 1, characterized in that, In step S2, the interlaced vector distortion loss for: ； in, It is the number of signals, for the i-th signal, It's a label, for the real signal. For a synthesized signal, the value is 1. =0, yes Obtained through nonlinear transformation operation The signal analysis network provides the samples. The probability of the actual data.
3. The data augmentation method for anomaly detection of unmanned aerial vehicles as described in claim 1, characterized in that, In step S2, the fully connected attribute pairing loss for: ； in, This represents the activation function of the intermediate layer of the signal analysis network. It is the synthesized signal output by the signal vector generation network. It's a real signal. It refers to the number of signals.
4. A data augmentation device for anomaly detection in unmanned aerial vehicles, characterized in that, include: The aircraft sequence data acquisition module is used to acquire flight time series data for UAV anomaly detection as a real signal; A generative network architecture building module is used to construct a generative network architecture based on relation analysis. The generative network architecture based on relation analysis includes a signal vector generation network and a signal analysis network. The signal vector generation network includes a first parallel relation generation layer, which performs a first dot product operation on the input vector, and then outputs a synthesized signal after passing through a first feedforward multilayer perceptron and a convolutional layer. The signal analysis network includes a second parallel relation generation layer, which performs a second dot product operation on the synthesized or real signals, and then outputs the predicted signal type after passing through a second feedforward multilayer perceptron and a classification layer. The signal vector generation network uses a fully connected attribute pairing loss to optimize the strategy parameters so that the output synthesized signal is more similar to the real signal. The signal analysis network uses an interleaved vector distortion loss to optimize the strategy parameters so that it can better distinguish between the synthesized signal and the real signal. The generative network architecture training module is used to continuously output synthesized signals through the signal vector generation network and input them into the signal analysis network. The two networks alternately optimize each other to obtain a trained generative network architecture based on relation analysis. The synthetic signal generation module is used to generate a sufficient amount of synthetic signals from the signal vectors of a pre-trained relation-analysis-based generative network architecture to augment the original real signals. The generative network architecture construction module includes an input vector generation unit, which generates a noise sequence with the same mean and variance as the flight time series data based on the noise input unit, and converts it into an input vector I through position embedding; The generative network architecture building module includes a sequence data processing unit, which classifies each type of signal into c channels according to its features, and divides each channel into multiple patches of a specified size N. A soft position code value is added to the front end of each patch to form an input vector for performing the second dot product operation. The parallel relation generation layer is used to perform dot product operations, which include: ； Where I is the input vector or signal, C is the context vector, W is the weight vector, dC is the dimension of the context vector, PRG(I,C,W) is the output of the dot product operation, and DistEq is the distributed balanced activation function.