Data enhancement method and device for anomaly detection of unmanned aerial vehicle
By generating synthetic signals based on relationship analysis, the problems of scarcity of drone flight feature data and insufficient generalization capabilities of model are solved, and the performance and robustness of drone anomaly detection are significantly improved.
Patent Information
- Application Number
- CN202410929177.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-11
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-07-11
AI Technical Summary
The scarcity of data in the flight characteristics of the drone and the insufficient generalization capability of model, resulting in poor results in data-driven anomaly detection methods in actual applications.
A generative network architecture based on relationship analysis is adopted, including signal vector generation network and signal analysis network, and a synthetic signal is generated through dot product operation and loss optimization strategies, which enhances the original flight data and improves data quality and quantity.
It significantly improves the performance of the data-driven drone anomaly detection system, enhances the generalization ability and robustness of the model, and improves the accuracy and efficiency of anomaly detection.
Smart Images

Figure CN119026034A_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the field of data processing technology, and in particular relates to a data enhancement method and device for drone anomaly detection. Background Art
[0002] With the rapid development of unmanned aerial vehicle (UAV) technology, its wide application in many fields such as military, commerce and scientific research is gradually becoming a reality. However, with it comes the urgent need for accurate prediction of UAV behavior and anomaly detection, especially in key areas such as public safety and airspace management. UAV anomaly detection can be divided into two main methods: model-based and data-driven.
[0003] Model-based anomaly detection methods use mathematical models and statistical techniques to detect anomalies in UAV systems. First, a mathematical model describing the normal behavior of the system is established, and then a system model is constructed based on the operating mechanism of the UAV system, historical data, or domain knowledge. Although model-based anomaly detection methods are generally effective, the established models usually do not have characteristics that are common to other aircraft types.
[0004] Data-driven anomaly detection methods mainly perform anomaly detection by utilizing and analyzing the data generated by drone sensors. Compared with model-based methods, data-driven methods are more versatile. However, data-driven methods require a large amount of data support. In the actual use of drones, due to the limitations of drone onboard equipment and sensor performance, it is usually impossible to record a large amount of high-quality flight data. Summary of the invention
[0005] In order to solve the above problems, the present application provides a data enhancement method and device for drone anomaly detection to solve the problems of scarcity of drone flight feature data and insufficient model generalization ability.
[0006] The first aspect of the present application provides a data enhancement method for drone anomaly detection, mainly comprising:
[0007] Step S1, obtaining flight time series data for drone anomaly detection as a real signal;
[0008] Step S2, constructing a generative network architecture based on relational analysis, wherein the generative network architecture based on relational analysis includes a signal vector generation network and a signal analysis network, wherein the signal vector generation network includes a first parallel relational generation layer for performing a first dot product operation on an input vector, and then outputting a synthetic signal through a first feedforward multilayer perceptron and a convolutional layer, wherein the signal analysis network includes a second parallel relational generation layer for performing a second dot product operation on two types of signals, synthetic or real, and then outputting a predicted signal type through a second feedforward multilayer perceptron and a classification layer, wherein the signal vector generation network uses a fully connected attribute pairing loss to optimize strategy parameters so that the output synthetic signal is more similar to a real signal, and wherein the signal analysis network uses an interleaved vector distortion loss to optimize strategy parameters so as to be able to better distinguish the synthetic signal from the real signal;
[0009] Step S3: The synthetic signal is continuously outputted through the signal vector generation network and inputted into the signal analysis network. The two networks are optimized alternately to obtain a trained generation network architecture based on relationship analysis.
[0010] Step S4: Output a sufficient amount of synthetic signal through the trained signal vector generation network based on the relationship analysis generation network architecture to perform data enhancement on the original real signal.
[0011] Preferably, 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 is converted into an input vector I through position embedding and normalization.
[0012] Preferably, in step S2, before performing the second dot product operation on the synthetic or real two types of signals, the method further includes:
[0013] For synthetic signals or real signals, each type of signal is classified into c channels according to its features, each channel is divided into multiple patches of specified size N, and a soft position encoding value is added to the front end of each patch. Then, the input vector for the second dot product operation is formed by position 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 is the dimension of the context vector, and PRG(I,C,W) is the output of the dot product operation.
[0017] Preferably, in step S2, the interleaved vector distortion loss LCVD (A) is:
[0018]
[0019] Where N is the number of signals, for the i-th signal, y i is the label, for the real signal, y i is 1, for the synthetic signal, y i is 0, f is A(x i ) is obtained through nonlinear transformation, A(x i ) is the signal analysis network giving the sample x i is the probability of being true data.
[0020] Preferably, in step S2, the fully connected attribute pairing loss L FCAP (G) is:
[0021]
[0022]
[0023] Among them, φ represents the activation function of the middle layer of the signal analysis network, G(z i ) is the synthetic signal output by the signal vector generation network, x i is the real signal, and N is the number of signals.
[0024] The second aspect of the present application provides a data enhancement device for drone anomaly detection, mainly comprising:
[0025] The aircraft sequence data acquisition module is used to obtain the flight time series data for drone anomaly detection as a real signal;
[0026] A generative network architecture construction module, used to construct a generative network architecture based on relational analysis, wherein the generative network architecture based on relational analysis includes a signal vector generation network and a signal analysis network, wherein the signal vector generation network includes a first parallel relation generation layer, used to perform a first dot product operation on an input vector, and then output a synthetic signal through a first feedforward multilayer perceptron and a convolution layer, wherein the signal analysis network includes a second parallel relation generation layer, used to perform a second dot product operation on two types of signals, synthetic or real, and then output a predicted signal type through a second feedforward multilayer perceptron and a classification layer, wherein the signal vector generation network uses a fully connected attribute pairing loss to optimize strategy parameters so that the output synthetic signal is more similar to the real signal, and wherein the signal analysis network uses an interleaved vector distortion loss to optimize strategy parameters so as to be able to better distinguish the synthetic signal from the real signal;
[0027] The generation network architecture training module is used to continuously output the synthetic signal through the signal vector generation network and input it into the signal analysis network. The two networks are optimized alternately to obtain a trained generation network architecture based on relationship analysis;
[0028] The synthetic signal generation module is used to output a sufficient amount of synthetic signals through a signal vector generation network of a trained relationship analysis-based generation network architecture to perform data enhancement on the original real signal.
[0029] Preferably, the generative network architecture building module includes an input vector generating unit for generating a noise sequence having the same mean and variance as the flight time series data based on the noise input unit, and converting it into an input vector I through position embedding.
[0030] Preferably, the generative network architecture building module includes a sequence data processing unit, which is used to classify synthetic signals or real signals, each type of signal is respectively classified into c channels according to features, each channel is divided into multiple patches of a specified size N, and a soft position encoding value is added to the front end of each patch to form an input vector for performing a second dot product operation.
[0031] Preferably, the parallel relationship generation layer is used to perform a dot product operation, and the dot product operation includes:
[0032]
[0033] Where I is the input vector or signal, C is the context vector, W is the weight vector, and d C is the dimension of the context vector, and PRG(I,C,W) is the output of the dot product operation.
[0034] This application uses advanced relationship analysis generation technology to enhance the UAV flight data, and significantly improves the performance of the data-driven UAV anomaly detection system by increasing the data volume and improving the data quality. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 It is a schematic diagram of a generated network architecture based on relationship analysis of a preferred embodiment of the data enhancement method for drone anomaly detection in the present application.
[0036] Figure 2 This application Figure 1 A schematic diagram of the parallel relationship generation layer of the illustrated embodiment. DETAILED DESCRIPTION
[0037] In order to make the purpose, technical scheme and advantages of the implementation of this application clearer, the technical scheme in the implementation of this application will be described in more detail in combination with the drawings in the implementation of this application. In the drawings, the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The described implementation is a part of the implementation of this application, not all of the implementations. The implementation described below with reference to the drawings is exemplary and is intended to be used to explain this application, and cannot be understood as a limitation on this application. Based on the implementation in this application, all other implementations obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. The implementation of this application is described in detail below in combination with the drawings.
[0038] The first aspect of the present application provides a data enhancement method for drone anomaly detection, such as Figure 1 As shown, it mainly includes:
[0039] Step S1, obtaining flight time series data for drone anomaly detection as a real signal;
[0040] Step S2, constructing a generative network architecture based on relational analysis, wherein the generative network architecture based on relational analysis includes a signal vector generation network and a signal analysis network, wherein the signal vector generation network includes a first parallel relational generation layer for performing a first dot product operation on an input vector, and then outputting a synthetic signal through a first feedforward multilayer perceptron and a convolutional layer, wherein the signal analysis network includes a second parallel relational generation layer for performing a second dot product operation on two types of signals, synthetic or real, and then outputting a predicted signal type through a second feedforward multilayer perceptron and a classification layer, wherein the signal vector generation network uses a fully connected attribute pairing loss to optimize strategy parameters so that the output synthetic signal is more similar to a real signal, and wherein the signal analysis network uses an interleaved vector distortion loss to optimize strategy parameters so as to be able to better distinguish the synthetic signal from the real signal;
[0041] Step S3: The synthetic signal is continuously outputted through the signal vector generation network and inputted into the signal analysis network. The two networks are optimized alternately to obtain a trained generation network architecture based on relationship analysis.
[0042] Step S4: Output a sufficient amount of synthetic signal through the trained signal vector generation network based on the relationship analysis generation network architecture to perform data enhancement on the original real signal.
[0043] The present application constructs a generative network architecture based on relational analysis in step S2, performs training in step S3, and generates a synthetic signal in step S4 to expand the original real signal. In step S3, for such adversarial learning network, it can be determined whether the network has been trained by the following methods:
[0044] (1) The loss values of the generator and the analysis network. If the losses of both networks tend to be stable and the loss of the generator network does not continue to rise or fall, it may mean that the model has reached an equilibrium state. (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 the analysis network has reached an equilibrium state. If the analysis network is too strong, the generator network cannot learn effective features; if the generator network is too strong, the analysis network cannot distinguish between real and fake samples.
[0045] Monitor the training process and analyze the loss curves of the generated network to see if they are stable.
[0046] The generative network based on relational analysis is a deep learning model composed of two key components: a signal vector generation network ( Figure 1 ), and a signal analysis network ( Figure 1 The lower half of the CNN is shown in Figure 2). They are both built on the Relational Analysis Network Encoder architecture. The encoder consists of two composite blocks. Parallel relation generation layers build the first module, and the second module is a feed-forward multi-layer perceptron (MLP) with GELU activation function. A normalization layer is applied before both blocks, and a dropout layer is added after each block. Both blocks adopt residual connections.
[0047] The core function of the signal vector generation network is to convert the input noise into a vector representation with certain characteristics by deeply analyzing and learning the complex relationships contained in the training set data, so as to generate a synthetic signal with a specific relationship. The task of the signal analysis network is to analyze the input signal to determine the signal type or characteristics it represents, and to feed this information back to the signal vector generation network. In other words, the purpose of the signal vector generation network is to generate a synthetic signal that is highly similar to the real signal after training, so as to achieve the purpose of confusing the fake with the real, while the purpose of the signal analysis network is to accurately distinguish whether the input is a synthetic signal or a real signal after training. The two networks are in confrontation with each other, and this process continues until a dynamic balance is reached, that is, the generation network can generate high-quality signal vectors that are sufficient to confuse the analysis network. At this time, the ability of the analysis network in determining the signal type or characteristics forms a certain balance with the generation ability of the generation network. Through alternating optimization, the two promote each other and gradually improve the quality of synthetic data and the discrimination ability of the analysis network.
[0048] In some optional embodiments, 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 is converted into an input vector I through position embedding and normalization.
[0049] refer to Figure 1 After collecting the flight time series in step S1, the mean and variance of the flight time series data set are calculated. The data set is normalized using the mean and variance of the obtained data set. This ensures that the data has a uniform scale and range to improve the training effect of the model, and then a set of normally distributed noise is randomly generated, and the mean and variance of the noise are the same as the mean and variance of the above data set, and then embedded by position through the fully connected layer, so that the neural network has the ability to understand the order of the parameters in the input vector and the dependencies between them.
[0050] In some optional implementations, in step S2, before performing the second dot product operation on the synthetic or real two types of signals, the method further includes:
[0051] For synthetic signals or real signals, each type of signal is classified into c channels according to its features, each channel is divided into multiple patches of specified size N, and a soft position encoding value is added to the front end of each patch. Then, the input vector for the second dot product operation is formed by position embedding.
[0052] In this embodiment, since the signal analysis network uses a parallel relationship generation layer, and the parallel relationship generation layer is mainly used to process image data. The core idea of the visual relationship analysis network is to divide the input image into a series of image blocks, and then flatten these image blocks into a one-dimensional vector sequence. Therefore, the present application pre-processes the synthetic signal or the real signal according to the sequence format required by the image data, specifically including:
[0053] The time series data is regarded as image data with a height of 1, where the number of time steps is defined as the width of the image, denoted by w. The time series can contain a single or multiple channels (i.e., multiple features, such as speed, posture, etc.), which are equivalent to the number of channels of the image (such as RGB), denoted by c. Therefore, the input sequence can be represented by a matrix with a dimension of (BatchSize, c, 1, w). For further processing, this application selects a specific patch size N to divide the sequence into w / N patches. On this basis, a soft position encoding value is added at the end of each patch, which will be learned during the model training process. Finally, the data shape input to the signal analysis network encoder block is adjusted to (BatchSize, c, 1, (w / N) + 1).
[0054] In some optional implementations, in step S2, the dot product operation includes:
[0055]
[0056] Where I is the input vector or signal, C is the context vector, W is the weight vector, and d C is the dimension of the context vector, and PRG(I,C,W) is the output of the dot product operation.
[0057] The parallel relation generation layer (PRG) is one of the core innovations of the relational analysis network. It enables the model to better capture global information and long-distance dependencies when processing sequence data without relying on the position information of the sequence. The parallel relation generation layer converts each element in the input sequence into three vectors, called the input vector (I), the context vector (C), and the weight vector (W). These three vectors are obtained by linear transformation of the input sequence. Then, by calculating the dot product between the input vector and the context vector, and then passing through a distribution balancing activation function (DistEq), the weight distribution between each input and all contexts can be obtained. Finally, these weights are multiplied and added with the corresponding numerical vectors to obtain the output of the parallel relation generation layer. It should be noted that the dimension d of the context vector C It is also the scaling factor of the dot product operation. When calculating the inner product of each row vector of matrix I and C, in order to prevent the inner product from being too large, it is divided by d C The square root of .
[0058] Those skilled in the art can understand that I is the input vector, and the process of model training is to continuously modify the context vector C and the weight vector W. After the training is completed, the context vector C and the weight vector W are solidified. Figure 2 A parallel relation generation layer network architecture is given. In fact, multiple stacked parallel relation generation layers are connected through a fully connected layer. Figure 2 In , each parallel relation generation layer outputs one data, and multiple data are reconstructed into a vector.
[0059] In some optional implementations, in step S2, the interleaved vector distortion loss L CVD (A) is:
[0060]
[0061] Where N is the number of signals, for the i-th signal, y i is the label, for the real signal, y i is 1, for the synthetic signal, y i is 0, f is A(x i) is obtained through nonlinear transformation, A(x i ) is the signal analysis network giving the sample x i is the probability of being true data.
[0062] In this embodiment, in the training process of the generative network based on relational analysis, the signal analysis network and the signal vector generation network use different loss functions to guide their learning and optimization. The signal analysis network uses the interleaved vector distortion loss, and its core purpose is to measure the ability of the signal analysis network 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 distinguish between real data and generated data.
[0063] In some optional implementations, in step S2, the fully connected attribute pairing loss L FCAP (G) is:
[0064]
[0065] Among them, φ represents the activation function of the middle layer of the signal analysis network, G(z i ) is the synthetic signal output by the signal vector generation network, x i is the real signal, and N is the number of signals.
[0066] In this embodiment, the signal vector generation network uses a fully connected attribute pairing loss, which focuses on the difference between the feature representation of the signal vector generation network output and the real data in the middle layer 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 that are more similar to the real data, thereby improving the quality of the generated image. This loss function is usually calculated by comparing the statistical properties of the activations of the middle layer, such as the 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.
[0067] The relational analysis generation network of the present application can learn the distribution and characteristics of the original data, so that new data with similar statistical characteristics can be generated in step S4. Subsequently, the generated data is merged with the original data to construct a balanced training set to ensure that the model can fully learn the characteristics of the normal flight state.
[0068] After enhancing the data for drone anomaly detection, another deep learning training model can be used for drone anomaly detection training. Commonly used models include sequence long-range dependency networks, convolutional networks, etc. Deep learning models can learn complex data features and dynamic patterns, thereby improving the accuracy and robustness of anomaly detection.
[0069] After the model training is completed, the prediction residual of the model is calculated on the validation set and filtered using an IIR filter to reduce the impact of noise on the anomaly detection results. By setting a suitable threshold, the residual can be associated with the anomaly detection result to determine whether the flight status is abnormal. For cases where the anomaly detection result is uncertain, further analysis and processing can be performed to improve the robustness and reliability of the system.
[0070] This application generates data through a relational analysis network, which can effectively increase training data and improve the generalization and robustness of the model. The anomaly detection method of relational analysis network enhanced data has broad application prospects in the fields of drones, aircraft, etc. By continuously improving and optimizing this method, the performance and efficiency of anomaly detection can be further improved, providing more reliable protection for the safe operation of aircraft and other systems.
[0071] In order to better reflect the advantages of the present application, the following is explained through experiments.
[0072] In the anomaly detection classification based on relational analysis to generate network-enhanced 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 indicators 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 by the classifier that are actually positive. Recall refers to the proportion of positive samples successfully identified by the classifier among the actual positive samples. The F1 score is the harmonic mean that comprehensively considers precision and recall, which can help evaluate the overall performance of the classifier. In the case of a fault, if the residual is always below the threshold, the false negative increases by 1; if the residual is greater than the threshold at least once, the number of true positives increases by 1. The specific experimental results are shown in Table 1.
[0073] Table 1 Test results
[0074]
[0075] It can be clearly seen from the experimental results that the method proposed in this application has obvious performance advantages. In contrast, the method that does not use relational analysis to generate network-enhanced data to expand the training set has insufficient performance. In the first three methods, neural networks are used to predict multidimensional features and then perform anomaly detection, while the latter methods use sequence long-range dependency networks to predict single features and then perform anomaly detection. Use sequence long-range dependency networks to compare the effects. First, the method of this application is trained using a data set after data enhancement, and the rest of the comparative experiments use the original training set. In the first two experiments, the output prediction content is the entire local state, that is, the output speed, posture and angular velocity total 10-dimensional features. The subsequent three experiments were trained using the original training set, outputting speed, posture and angular velocity as outputs respectively.
[0076] Experiments show that the invention of the present application has obvious performance advantages. In contrast, the method that does not use relational analysis to generate network enhanced data shows obvious insufficient prediction performance. However, the method of the present application still does not achieve 100% accuracy. This is because the noise of the original data is too large, and the data enhanced by the method of the present application also has similar noise. These noises affect the prediction performance of the model, and thus affect the accuracy of anomaly detection.
[0077] The second aspect of the present application provides a data enhancement device for drone anomaly detection corresponding to the above method, mainly comprising:
[0078] The aircraft sequence data acquisition module is used to obtain the flight time series data for drone anomaly detection as a real signal;
[0079] A generative network architecture construction module, used to construct a generative network architecture based on relational analysis, wherein the generative network architecture based on relational analysis includes a signal vector generation network and a signal analysis network, wherein the signal vector generation network includes a first parallel relation generation layer, used to perform a first dot product operation on an input vector, and then output a synthetic signal through a first feedforward multilayer perceptron and a convolution layer, wherein the signal analysis network includes a second parallel relation generation layer, used to perform a second dot product operation on two types of signals, synthetic or real, and then output a predicted signal type through a second feedforward multilayer perceptron and a classification layer, wherein the signal vector generation network uses a fully connected attribute pairing loss to optimize strategy parameters so that the output synthetic signal is more similar to the real signal, and wherein the signal analysis network uses an interleaved vector distortion loss to optimize strategy parameters so as to be able to better distinguish the synthetic signal from the real signal;
[0080] The generation network architecture training module is used to continuously output the synthetic signal through the signal vector generation network and input it into the signal analysis network. The two networks are optimized alternately to obtain a trained generation network architecture based on relationship analysis;
[0081] The synthetic signal generation module is used to output a sufficient amount of synthetic signals through a signal vector generation network of a trained relationship analysis-based generation network architecture to perform data enhancement on the original real signal.
[0082] In some optional embodiments, the generative network architecture building module includes an input vector generating unit for generating a noise sequence having the same mean and variance as the flight time series data based on a noise input unit, and converting it into an input vector I through position embedding.
[0083] In some optional embodiments, the generative network architecture building module includes a sequence data processing unit, which is used to classify synthetic signals or real signals, each type of signal is classified into c channels according to features, each channel is divided into multiple patches of a specified size N, and a soft position encoding value is added to the front end of each patch to form an input vector for performing a second dot product operation.
[0084] In some optional implementations, the parallel relationship generation layer is used to perform a dot product operation, and the dot product operation includes:
[0085]
[0086] Where I is the input vector or signal, C is the context vector, W is the weight vector, and d C is the dimension of the context vector, and PRG(I,C,W) is the output of the dot product operation.
[0087] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.
Claims
1. A data enhancement method for drone anomaly detection, characterized in that: include: Step S1, obtaining flight time series data for drone anomaly detection as a real signal; Step S2, constructing a generative network architecture based on relational analysis, wherein the generative network architecture based on relational analysis includes a signal vector generation network and a signal analysis network, wherein the signal vector generation network includes a first parallel relational generation layer for performing a first dot product operation on an input vector, and then outputting a synthetic signal through a first feedforward multilayer perceptron and a convolutional layer, wherein the signal analysis network includes a second parallel relational generation layer for performing a second dot product operation on two types of signals, synthetic or real, and then outputting a predicted signal type through a second feedforward multilayer perceptron and a classification layer, wherein the signal vector generation network uses a fully connected attribute pairing loss to optimize strategy parameters so that the output synthetic signal is more similar to a real signal, and wherein the signal analysis network uses an interleaved vector distortion loss to optimize strategy parameters so as to be able to better distinguish the synthetic signal from the real signal; Step S3: The synthetic signal is continuously outputted through the signal vector generation network and inputted into the signal analysis network. The two networks are optimized alternately to obtain a trained generation network architecture based on relationship analysis. Step S4: Output a sufficient amount of synthetic signal through the trained signal vector generation network based on the relationship analysis generation network architecture to perform data enhancement on the original real signal.
2. The data enhancement method for drone anomaly detection according to claim 1, characterized in that: 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 is converted into an input vector I through position embedding and normalization.
3. The data enhancement method for drone anomaly detection according to claim 1, characterized in that: In step S2, before performing the second dot product operation on the synthetic or real two types of signals, the method further includes: For synthetic signals or real signals, each type of signal is classified into c channels according to its features, each channel is divided into multiple patches of specified size N, and a soft position encoding value is added to the front end of each patch. Then, the input vector for the second dot product operation is formed by position embedding.
4. The data enhancement method for drone anomaly detection according to claim 3, characterized in that: 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, and d C is the dimension of the context vector, and PRG(I,C,W) is the output of the dot product operation.
5. The data enhancement method for drone anomaly detection according to claim 1, characterized in that: In step S2, the interleaved vector distortion loss L CVD (A) is: Where N is the number of signals, for the i-th signal, y i is the label, for the real signal, y i is 1, for the synthetic signal, y i is 0, f is A(x i ) is obtained through nonlinear transformation, A(x i ) is the signal analysis network giving the sample x i is the probability of being true data.
6. The data enhancement method for drone anomaly detection according to claim 1, characterized in that: In step S2, the fully connected attribute pairing loss L FCAP (G) is: Among them, φ represents the activation function of the middle layer of the signal analysis network, G(z i ) is the synthetic signal output by the signal vector generation network, x i is the real signal, and N is the number of signals.
7. A data enhancement device for drone anomaly detection, characterized in that: include: The aircraft sequence data acquisition module is used to obtain the flight time series data for drone anomaly detection as a real signal; A generative network architecture construction module, used to construct a generative network architecture based on relational analysis, wherein the generative network architecture based on relational analysis includes a signal vector generation network and a signal analysis network, wherein the signal vector generation network includes a first parallel relation generation layer, used to perform a first dot product operation on an input vector, and then output a synthetic signal through a first feedforward multilayer perceptron and a convolution layer, wherein the signal analysis network includes a second parallel relation generation layer, used to perform a second dot product operation on two types of signals, synthetic or real, and then output a predicted signal type through a second feedforward multilayer perceptron and a classification layer, wherein the signal vector generation network uses a fully connected attribute pairing loss to optimize strategy parameters so that the output synthetic signal is more similar to the real signal, and wherein the signal analysis network uses an interleaved vector distortion loss to optimize strategy parameters so as to be able to better distinguish the synthetic signal from the real signal; The generation network architecture training module is used to continuously output the synthetic signal through the signal vector generation network and input it into the signal analysis network. The two networks are optimized alternately to obtain a trained generation network architecture based on relationship analysis; The synthetic signal generation module is used to output a sufficient amount of synthetic signals through a signal vector generation network of a trained relationship analysis-based generation network architecture to perform data enhancement on the original real signal.
8. The data enhancement device for drone anomaly detection according to claim 7, characterized in that: The generative network architecture building module includes an input vector generating unit, which is used to generate a noise sequence with the same mean and variance as the flight time series data based on the noise input unit, and convert it into an input vector I through position embedding.
9. The data enhancement device for drone anomaly detection according to claim 7, characterized in that: The generative network architecture building module includes a sequence data processing unit, which is used to classify synthetic signals or real signals, each type of signal is classified into c channels according to features, each channel is divided into multiple patches of a specified size N, and a soft position encoding value is added to the front end of each patch to form an input vector for performing a second dot product operation.
10. The data enhancement device for drone anomaly detection according to claim 9, characterized in that: The parallel relationship generation layer is used to perform a dot product operation, and the dot product operation includes: Where I is the input vector or signal, C is the context vector, W is the weight vector, and d C is the dimension of the context vector, and PRG(I,C,W) is the output of the dot product operation.
Citation Information
Patent Citations
Inplanatable time series data anomaly detection method based on lightweight reconstruction network
CN117668720A
Flight special situation data enhancement method, device and equipment
CN118036677A
Deep compression power lithium battery fault diagnosis method under perceptual adversarial generation
WO2021017416A1
Cited By
Unmanned aerial vehicle anomaly detection method and device based on multiple data nonlinear combinations
CN121211248A
